1 00:00:05,839 --> 00:00:07,808 >> OKAY, WELL LET'S KICK THIS 2 00:00:07,808 --> 00:00:08,242 OFF. 3 00:00:08,242 --> 00:00:09,176 GOOD MORNING EVERYONE AND THANK 4 00:00:09,176 --> 00:00:10,978 YOU SO MUCH FOR COMING IN PERSON 5 00:00:10,978 --> 00:00:13,513 OR ATTENDING IN THE VIRTUAL 6 00:00:13,513 --> 00:00:15,349 AUDIENCE TO THE BRAIN NEUROAI WORKSHOP. 7 00:00:15,349 --> 00:00:17,951 WE PUT TOGETHER A GREAT PROGRAM 8 00:00:17,951 --> 00:00:18,385 HERE. 9 00:00:18,385 --> 00:00:20,354 AND YOU'LL HEAR A LOT MORE ABOUT 10 00:00:20,354 --> 00:00:22,623 IT IN THIS OPENING SESSION BUT 11 00:00:22,623 --> 00:00:25,192 FIRST OFF I WANT TO PASS THE 12 00:00:25,192 --> 00:00:29,296 MIC TO ANDREA BECKEL-MITCHENER, THE DEPUTY DIRECTOR OF 13 00:00:29,296 --> 00:00:32,666 THE NIH BRAIN INITIATIVE. 14 00:00:32,666 --> 00:00:38,005 WELCOME, ANDREA. 15 00:00:38,005 --> 00:00:40,107 >> THANKS, JOE. 16 00:00:40,107 --> 00:00:43,076 THANKS EVERYONE FOR BEING HERE. 17 00:00:43,076 --> 00:00:46,947 I UNDERSTAND WE HAVE ABOUT 300 18 00:00:46,947 --> 00:00:50,050 PLUS IN PERSON REGISTRATIONS AND 19 00:00:50,050 --> 00:00:52,786 OVER 1,000 JOINING US VIRTUALLY. 20 00:00:52,786 --> 00:00:54,421 SO WELCOME TO ONE AND ALL. 21 00:00:54,421 --> 00:00:55,389 WE'RE REALLY EXCITED ABOUT THIS. 22 00:00:55,389 --> 00:00:59,126 I WANTED TO THANK THE ORGANIZING 23 00:00:59,126 --> 00:01:03,697 COMMITTEE INCLUDING JOE AS WELL 24 00:01:03,697 --> 00:01:04,364 AS THE SCIENTIFIC PROGRAM 25 00:01:04,364 --> 00:01:05,866 COMMITTEE WHO HAS PUT TOGETHER 26 00:01:05,866 --> 00:01:07,267 WHAT LOOKS LIKE A VERY INTERESTING 27 00:01:07,267 --> 00:01:07,834 WORKSHOP. 28 00:01:07,834 --> 00:01:08,702 IT'S VERY IMPORTANT TO THE BRAIN 29 00:01:08,702 --> 00:01:10,304 INITIATIVE AND WE'RE EXCITED TO 30 00:01:10,304 --> 00:01:11,271 LEARN MORE. 31 00:01:11,271 --> 00:01:14,641 SO AS JOE SAID I AM ANDREA THE 32 00:01:14,641 --> 00:01:16,810 DEPUTY DIRECTOR OF THE BRAIN 33 00:01:16,810 --> 00:01:17,577 INITIATIVE HERE AT NIH. 34 00:01:17,577 --> 00:01:21,281 I THOUGHT I'D TAKE A FEW MINUTES 35 00:01:21,281 --> 00:01:22,716 TO DESCRIBE THE BRAIN INITIATIVE 36 00:01:22,716 --> 00:01:24,351 AT DIFFERENT LEVELS AND WALK YOU 37 00:01:24,351 --> 00:01:25,953 THROUGH A FEW THINGS THAT WE 38 00:01:25,953 --> 00:01:28,488 THINK ARE EXTRA IMPORTANT TO PAY 39 00:01:28,488 --> 00:01:30,757 ATTENTION TO OVER THESE NEXT TWO 40 00:01:30,757 --> 00:01:31,558 DAYS. 41 00:01:31,558 --> 00:01:32,960 SO THE BRAIN INITIATIVE IS 42 00:01:32,960 --> 00:01:35,529 BROADER THAN NIH AND WAS ANNOUNCED 43 00:01:35,529 --> 00:01:36,697 BACK IN 2013. 44 00:01:36,697 --> 00:01:40,167 THE NIH MADE OUR FIRST AWARDS IN 45 00:01:40,167 --> 00:01:42,102 2014 AND REALLY THE BRAIN 46 00:01:42,102 --> 00:01:47,474 INITIATIVE IS A PARTNERSHIP. 47 00:01:47,474 --> 00:01:50,210 A PUBLIC-PRIVATE PARTNERSHIP 48 00:01:50,210 --> 00:01:52,346 AMONG SEVERAL U.S. AGENCIES AND 49 00:01:52,346 --> 00:01:53,880 FOUNDATIONS AND YOU'LL SEE THEIR ICONS 50 00:01:53,880 --> 00:01:55,349 AT THE BOTTOM OF THE SLIDE. 51 00:01:55,349 --> 00:01:58,285 WE'VE ENJOYED A LOT OF SUPPORT 52 00:01:58,285 --> 00:01:59,753 FROM CONGRESS SINCE ITS 53 00:01:59,753 --> 00:02:01,088 INCEPTION IN 2014. 54 00:02:01,088 --> 00:02:07,761 WE FUNDED OVER 1600 PROJECTS AT 55 00:02:07,761 --> 00:02:10,664 A BUDGET OF $3.5 BILLION. 56 00:02:10,664 --> 00:02:12,232 THE NIH BRAIN INITIATIVE IS WHO 57 00:02:12,232 --> 00:02:14,101 WE ARE AND WE ARE SPONSORING THIS 58 00:02:14,101 --> 00:02:16,336 WORKSHOP. WE ARE MADE UP OF 10 59 00:02:16,336 --> 00:02:18,171 INSTITUTES AND CENTERS HERE AT 60 00:02:18,171 --> 00:02:18,538 THE NIH. 61 00:02:18,538 --> 00:02:21,208 THE TWO LEAD INSTITUTES BEING 62 00:02:21,208 --> 00:02:22,843 NINDS THE NATIONAL INSTITUTE OF 63 00:02:22,843 --> 00:02:24,344 NEUROLOGICAL DISORDERS AND 64 00:02:24,344 --> 00:02:25,645 STROKE AND MY HOME INSTITUTE THE 65 00:02:25,645 --> 00:02:26,847 NATIONAL INSTITUTE OF MENTAL 66 00:02:26,847 --> 00:02:27,080 HEALTH. 67 00:02:27,080 --> 00:02:31,318 BUT WE RELY ON THE WORK AND 68 00:02:31,318 --> 00:02:33,787 EXPERTISE AND SUPPORT OF EIGHT 69 00:02:33,787 --> 00:02:36,356 OTHER INSTITUTES AND WE'VE BEEN 70 00:02:36,356 --> 00:02:38,291 FAIRLY SUCCESSFUL WE THINK USING 71 00:02:38,291 --> 00:02:39,559 THIS MODEL OVER THE LAST 10 72 00:02:39,559 --> 00:02:40,293 YEARS. 73 00:02:40,293 --> 00:02:41,595 OUR MISSION IS TO REVOLUTIONIZE 74 00:02:41,595 --> 00:02:43,697 OUR UNDERSTANDING OF THE HUMAN 75 00:02:43,697 --> 00:02:45,165 BRAIN BY ACCELERATING 76 00:02:45,165 --> 00:02:46,633 DEVELOPMENT AND APPLICATION OF 77 00:02:46,633 --> 00:02:47,534 INNOVATIVE TECHNOLOGIES. 78 00:02:47,934 --> 00:02:49,836 AGAIN, SO 10 NIH INSTITUTES AND 79 00:02:49,836 --> 00:02:50,837 CENTERS CONTRIBUTING BUT THIS IS 80 00:02:50,837 --> 00:02:52,105 THE REAL ENGINE. 81 00:02:52,105 --> 00:02:54,307 I KNOW YOU CAN'T READ THESE 82 00:02:54,307 --> 00:02:56,443 NAMES BUT THIS IS A LIST OF ALL 83 00:02:56,443 --> 00:03:00,013 THE STAFF MEMBERS ACROSS THE NIH 84 00:03:00,013 --> 00:03:01,481 INSTITUTE AS WELL AS BEYOND THE 85 00:03:01,481 --> 00:03:05,118 INSTITUTES THAT ARE 86 00:03:05,118 --> 00:03:07,187 CONTRIBUTORS. 87 00:03:07,187 --> 00:03:08,922 WE COULDN'T DO WHAT WE'RE DOING 88 00:03:08,922 --> 00:03:10,457 WITHOUT THEIR EXPERTISE AND 89 00:03:10,457 --> 00:03:11,525 DEDICATION. WE'RE GRATEFUL FOR IT. 90 00:03:11,525 --> 00:03:13,193 NOW A WALK THROUGH HISTORY OF 91 00:03:13,193 --> 00:03:13,727 OUR BUDGET. 92 00:03:13,727 --> 00:03:15,896 I'M NOT GOING TO SPEND A LOT OF 93 00:03:15,896 --> 00:03:20,200 TIME ON THIS BUT IT COMES UP IN 94 00:03:20,200 --> 00:03:20,934 CONVERSATION A LOT. THE BRAIN 95 00:03:20,934 --> 00:03:25,806 INITIATIVE IS FUNDED THROUGH TWO 96 00:03:25,806 --> 00:03:27,741 STREAMS. OUR BASE APPROPRIATION 97 00:03:27,741 --> 00:03:29,476 THE BASE BUDGET AT NIH ALL 98 00:03:29,476 --> 00:03:31,144 INSTITUTES AND CENTERS GET FROM 99 00:03:31,144 --> 00:03:34,414 CONGRESS. WE ALSO HAVE A CHUNK OF 100 00:03:34,414 --> 00:03:35,782 MONEY FROM ANOTHER AUTHORIZED 101 00:03:35,782 --> 00:03:37,984 LEGISLATION CALLED THE CURES 102 00:03:37,984 --> 00:03:38,151 ACT. 103 00:03:38,151 --> 00:03:40,454 AND YOU'LL SEE OUR BASE BUDGET 104 00:03:40,454 --> 00:03:43,757 HAS GROWN STEADILY FOR THE MOST 105 00:03:43,757 --> 00:03:47,794 PART FROM FY14 THROUGH FY22. 106 00:03:47,794 --> 00:03:51,731 AND THE CURES ACT WHICH WAS 107 00:03:51,731 --> 00:03:54,968 ENACTED IN FY17 IS ANOTHER TYPE 108 00:03:54,968 --> 00:03:58,305 OF FUNDING THAT FLUCTUATES YEAR 109 00:03:58,305 --> 00:03:58,605 TO YEAR. 110 00:03:58,605 --> 00:04:01,174 IT HAS TO BE APPROPRIATED EVERY 111 00:04:01,174 --> 00:04:02,742 YEAR AND I'M USING A LOT OF 112 00:04:02,742 --> 00:04:04,578 TECHNICAL TERMS BUT WE ARE 113 00:04:04,578 --> 00:04:05,946 AUTHORIZED THROUGH 2026 WITH 114 00:04:05,946 --> 00:04:08,849 ADDITIONAL CURES MONEY. 115 00:04:08,849 --> 00:04:12,986 SO YOU'LL NOTE -- THIS 116 00:04:12,986 --> 00:04:15,755 IS THE WAY THIS BUDGET LOOKED 117 00:04:15,755 --> 00:04:17,190 THROUGH 2022 WITH THE 118 00:04:17,190 --> 00:04:18,225 COMBINATION OF THE BLUE BAR, THE 119 00:04:18,225 --> 00:04:23,763 BASE MONEY AND THE REDDISH-PINK 120 00:04:23,763 --> 00:04:24,831 BAR THE CURES MONEY AND WHAT 121 00:04:24,831 --> 00:04:26,900 HAPPENED IN 2023 IS OUR BASE MONEY 122 00:04:26,900 --> 00:04:27,734 DROPPED BECAUSE OUR CURES MONEY 123 00:04:27,734 --> 00:04:30,437 WAS SO HIGH THAT YEAR. 124 00:04:30,437 --> 00:04:34,341 AND IF YOU RECALL ANCIENT 125 00:04:34,341 --> 00:04:39,613 HISTORY WHEN FY24 BUDGET WAS 126 00:04:39,613 --> 00:04:43,750 PASSED THEY KEPT THE NIH 127 00:04:43,750 --> 00:04:44,351 FLAT SO THE BRAIN BUDGET WAS 128 00:04:44,351 --> 00:04:46,953 ALSO FLAT. BUT THAT YEAR WHEN 129 00:04:46,953 --> 00:04:48,955 THE CURES MONEY DIPPED, WE ENDED 130 00:04:48,955 --> 00:04:52,192 UP HAVING ABOUT A $278 MILLION 131 00:04:52,192 --> 00:04:53,927 DECREASE IN OUR OVER ALL BUDGET. 132 00:04:53,927 --> 00:04:55,495 WE PLANNED FOR THIS. 133 00:04:55,495 --> 00:04:58,398 WE WEREN'T SURE IT WAS GOING TO 134 00:04:58,398 --> 00:04:59,666 HAPPEN BUT MODELLED OUR BUDGET 135 00:04:59,666 --> 00:05:00,267 IN DIFFERENT WAYS. 136 00:05:00,267 --> 00:05:01,601 IT WAS UNFORTUNATE WE HAD TO 137 00:05:01,601 --> 00:05:05,172 LOSE SOME OF OUR FUNDING 138 00:05:05,172 --> 00:05:07,073 ANNOUNCEMENTS AT LEAST 139 00:05:07,073 --> 00:05:08,508 TEMPORARILY AND HAD HOPE SOME 140 00:05:08,508 --> 00:05:10,443 MIGHT BE RESTORED IN THE FUTURE 141 00:05:10,443 --> 00:05:12,279 BUT THAT'S STILL $400 MILLION 142 00:05:12,279 --> 00:05:13,213 OF A BUDGET. 143 00:05:13,213 --> 00:05:14,948 THAT'S STILL A PRETTY HEALTHY 144 00:05:14,948 --> 00:05:16,516 BUDGET TO MAKE INVESTMENTS IN 145 00:05:16,516 --> 00:05:19,452 THE KINDS OF SCIENCE WE THINK IS IMPORTANT. 146 00:05:19,452 --> 00:05:21,221 SO HERE WE ARE, THE BRAIN 147 00:05:21,221 --> 00:05:22,756 INITIATIVE AT NIH AND HERE'S A 148 00:05:22,756 --> 00:05:24,591 LIST OF SOME OF THE THINGS WE 149 00:05:24,591 --> 00:05:27,327 SUPPORT IN BROAD STROKES. 150 00:05:27,327 --> 00:05:30,197 CELL AND CIRCUIT TECHNOLOGIES, 151 00:05:30,197 --> 00:05:31,998 RECORDING AND NEURAL MODULATION, 152 00:05:31,998 --> 00:05:33,533 IMAGING TECHNOLOGIES OF ALL 153 00:05:33,533 --> 00:05:35,402 KINDS AND DIFFERENT RESOLUTIONS, 154 00:05:35,402 --> 00:05:37,137 SYSTEMS NEUROSCIENCE AND WHAT WE 155 00:05:37,137 --> 00:05:37,737 CALL HUMAN NEUROSCIENCE. 156 00:05:37,737 --> 00:05:41,274 ON THE RIGHT-HAND SIDE OF THIS 157 00:05:41,274 --> 00:05:43,176 SLIDE IS SOME OVERARCHING THEMES 158 00:05:43,176 --> 00:05:45,011 WE THINK ARE VERY IMPORTANT. 159 00:05:45,011 --> 00:05:47,414 THE DATA SCIENCE AND INFORMATICS 160 00:05:47,414 --> 00:05:47,681 COMPONENT. 161 00:05:47,681 --> 00:05:50,250 WE HAVE A STRONG PROGRAM AND 162 00:05:50,250 --> 00:05:52,953 TRAINING, FOCUSSING ON INCLUSION 163 00:05:52,953 --> 00:05:54,087 AND EQUITY. 164 00:05:54,087 --> 00:06:00,026 WE ALSO SUPPORT NEUROETHICS 165 00:06:00,026 --> 00:06:01,561 DISSEMINATION AND COMMERCIALIZATION 166 00:06:01,561 --> 00:06:02,696 THROUGH DIFFERENT AVENUES. 167 00:06:02,696 --> 00:06:06,066 I'LL TALK LATER IN THE TALK ON 168 00:06:06,066 --> 00:06:07,500 THE DATA SCIENCE AND NEUROETHICS 169 00:06:07,500 --> 00:06:09,302 PIECES JUST TO INTRODUCE THOSE 170 00:06:09,302 --> 00:06:11,471 IN MORE DETAIL. 171 00:06:11,471 --> 00:06:14,507 WE THINK BRAIN -- WE LIKE TO 172 00:06:14,507 --> 00:06:16,243 THINK OF OURSELVES AS UNIQUE AND 173 00:06:16,243 --> 00:06:16,576 FOUNDATIONAL. 174 00:06:16,576 --> 00:06:19,479 THE TYPES OF PROJECTS WE SUPPORT 175 00:06:19,479 --> 00:06:21,414 ARE INNOVATIVE. 176 00:06:21,414 --> 00:06:23,283 WE SUPPORT A LOT OF TEAM SCIENCE IN 177 00:06:23,283 --> 00:06:25,785 ALL DIFFERENT AREAS. AGAIN, 178 00:06:25,785 --> 00:06:30,257 BEING INCLUSIVE, GROWING THE NEW 179 00:06:30,257 --> 00:06:31,891 GENERATION OF SCIENTISTS ENGINEERS 180 00:06:31,891 --> 00:06:34,527 COMPUTATIONAL BIOLOGISTS AND 181 00:06:34,527 --> 00:06:35,762 MATHEMATICIANS AND PHYSICISTS TO 182 00:06:35,762 --> 00:06:36,930 SUPPORT THE SCIENCE OF THE 183 00:06:36,930 --> 00:06:38,365 BRAIN. 184 00:06:38,365 --> 00:06:43,937 WE HAVE A VERY OPEN ASSERTIVE 185 00:06:43,937 --> 00:06:45,972 SHARING TRADITION, I GUESS 186 00:06:45,972 --> 00:06:47,407 AND WE CONTINUE TO DO THAT WHERE 187 00:06:47,407 --> 00:06:48,808 WE THINK THE SCIENCE SHOULD BE 188 00:06:48,808 --> 00:06:50,043 OPEN AND ACCESSIBLE AND 189 00:06:50,043 --> 00:06:50,877 AVAILABLE TO ALL. 190 00:06:50,877 --> 00:06:55,282 AND OF COURSE AS I MENTIONED 191 00:06:55,282 --> 00:06:59,085 NEUROETHICS AND ETHICS IN 192 00:06:59,085 --> 00:06:59,452 GENERAL IS VERY IMPORTANT. 193 00:06:59,452 --> 00:07:00,553 THERE'S A PUBLICATION THAT CAME 194 00:07:00,553 --> 00:07:05,492 OUT A MONTH OR TWO AGO. 195 00:07:05,492 --> 00:07:09,396 JOHN NGAI OUR DIRECTOR IN NEURON 196 00:07:09,396 --> 00:07:10,797 IMPORTANT TO REFLECT ON THE LAST 197 00:07:10,797 --> 00:07:12,065 10 YEARS OF INNOVATION. 198 00:07:12,065 --> 00:07:14,134 I RECOMMEND YOU ACCESS THAT. 199 00:07:14,134 --> 00:07:15,368 IT WILL TELL US MORE WHERE WE'VE 200 00:07:15,368 --> 00:07:17,871 BEEN AND WHERE WE HOPE TO GO. 201 00:07:17,871 --> 00:07:19,472 AND OF THOSE 1600 PROJECTS THAT 202 00:07:19,472 --> 00:07:23,076 I MENTIONED THAT WE FUNDED, WE 203 00:07:23,076 --> 00:07:25,011 HAVE AN EMPHASIS ON FOUR MAIN 204 00:07:25,011 --> 00:07:27,013 PROJECTS THAT WE'VE SUPPORTED IN 205 00:07:27,013 --> 00:07:27,881 GENERAL WE'RE CALLING THESE THE 206 00:07:27,881 --> 00:07:29,382 PARTS LIST. ON THE LEFT HAND 207 00:07:29,382 --> 00:07:32,886 SIDE IS THE CELL ATLASING PROGRAM 208 00:07:32,886 --> 00:07:36,022 LOOKING AT THE DIFFERENT CELLS IN 209 00:07:36,022 --> 00:07:38,358 THE BRAIN OF DIFFERENT SPECIES. 210 00:07:38,358 --> 00:07:39,859 WE'RE SUPPORTING WIRING DIAGRAMS 211 00:07:39,859 --> 00:07:41,027 AGAIN OF DIFFERENT SPECIES. 212 00:07:41,027 --> 00:07:44,364 WE CALL THIS OUR BRAIN CONNECTS 213 00:07:44,364 --> 00:07:44,597 NETWORK. 214 00:07:44,597 --> 00:07:46,066 THAT'S ANOTHER LARGE PROGRAM 215 00:07:46,066 --> 00:07:46,966 WE'RE SUPPORTING. 216 00:07:46,966 --> 00:07:49,903 WE ALSO HAVE AN INTEREST IN 217 00:07:49,903 --> 00:07:53,573 USING THE DATA FROM THE CELL 218 00:07:53,573 --> 00:07:54,507 ATLASING AND WIRING PROGRAMS TO 219 00:07:54,507 --> 00:07:57,510 UNDERSTAND HOW WE CAN INTERVENE. 220 00:07:57,510 --> 00:08:01,348 SO PRECISION CELL ACCESS WE CALL 221 00:08:01,348 --> 00:08:09,956 THAT THIS ARMAMENTARIUM AND OUR 222 00:08:09,956 --> 00:08:14,060 NEWEST PROGRAM BRAIN BEHAVIOR 223 00:08:14,060 --> 00:08:15,428 QUANTIFICATION AND SYNCHRONIZATION 224 00:08:15,428 --> 00:08:16,963 LOOKING AT REALTIME SIGNALS, BRAIN ACTIVITY 225 00:08:16,963 --> 00:08:21,034 AND HOW THOSE RELATE TEMPORALLY TO BEHAVIORS OF DIFFERENT organisms including humans 226 00:08:21,034 --> 00:08:22,068 I'M GOING TO TALK A LITTLE BIT, 227 00:08:22,068 --> 00:08:24,037 THAT I HOPE ALL OF YOU WERE 228 00:08:24,037 --> 00:08:25,372 ACCESSING IN THE LAST MONTH OR 229 00:08:25,372 --> 00:08:26,940 SO THIS WAS FLYWIRE. 230 00:08:26,940 --> 00:08:30,877 A HUGE UNDERTAKING, A REMARKABLE 231 00:08:30,877 --> 00:08:32,245 UNDERTAKING EXAMINING AND 232 00:08:32,245 --> 00:08:34,581 DEVELOPING THE TOOLS AND FINALLY 233 00:08:34,581 --> 00:08:36,316 THE OUTPUT OF THIS PROJECT THAT 234 00:08:36,316 --> 00:08:38,284 WAS PUBLISHED IN NATURE A COUPLE 235 00:08:38,284 --> 00:08:40,987 MONTHS AGO AND IT'S THE COMPLETE 236 00:08:40,987 --> 00:08:43,757 WIRING DIAGRAM OF THE 237 00:08:43,757 --> 00:08:44,057 DROSOPHILA MELANOGASTER COMMON 238 00:08:44,057 --> 00:08:45,592 FRUITFLY - IT'S AMAZING. 239 00:08:45,592 --> 00:08:48,928 I WISH THERE WAS SOME ORCHESTRAL 240 00:08:48,928 --> 00:08:50,563 MUSIC THAT WOULD PLAY WHILE THIS 241 00:08:50,563 --> 00:08:52,465 VIDEO PLAYS BECAUSE IT'S 242 00:08:52,465 --> 00:08:53,166 ABSOLUTELY GORGEOUS. 243 00:08:53,166 --> 00:08:56,069 SO WHAT WAS SUPPORTED HERE OR 244 00:08:56,069 --> 00:08:59,806 WAS DISCOVERED, I SHOULD SAY, IS 245 00:08:59,806 --> 00:09:01,975 CONNECTIVITY AT SYNAPSE LEVEL IN 246 00:09:01,975 --> 00:09:03,476 A FLY BRAIN. S0 139000 NEURONS, 247 00:09:03,476 --> 00:09:07,046 4500 OR SO UNIQUE CELL TYPES AND 248 00:09:07,046 --> 00:09:09,215 50 MILLION APPROXIMATELY, GIVE 249 00:09:09,215 --> 00:09:11,851 OR TAKE, SYNAPSES ACROSS ALL 250 00:09:11,851 --> 00:09:12,118 THOSE. 251 00:09:12,118 --> 00:09:15,055 THERE WERE NINE PUBLICATIONS IN 252 00:09:15,055 --> 00:09:16,489 NATURE, JOINT PUBLICATIONS AND 253 00:09:16,489 --> 00:09:19,426 WHAT WE'RE REALLY EXCITED ABOUT 254 00:09:19,426 --> 00:09:24,431 NOT ONLY WILL THIS ENABLE ADDITIONAL 255 00:09:24,431 --> 00:09:27,467 STUDIES AND SETS THE FOUNDATION 256 00:09:27,467 --> 00:09:29,836 FOR DOING THIS IN LARGER 257 00:09:29,836 --> 00:09:31,471 ORGANISMS LIKE THE MOUSE AND NON 258 00:09:31,471 --> 00:09:34,741 HUMAN PRIMATES AND HUMANS WHICH 259 00:09:34,741 --> 00:09:37,710 ARE PART OF THE CONNECTS PROGRAM I 260 00:09:37,710 --> 00:09:38,344 I MENTIONED EARLIER. 261 00:09:38,344 --> 00:09:40,980 SO THE OTHER PART I WANT TO 262 00:09:40,980 --> 00:09:43,283 EMPHASIZE IS THIS WAS AN 263 00:09:43,283 --> 00:09:43,850 INTERNATIONAL EFFORT. 264 00:09:43,850 --> 00:09:45,485 OVER 50 INSTITUTIONS FUNDED BY 265 00:09:45,485 --> 00:09:46,152 MULTIPLE AGENCIES. 266 00:09:46,152 --> 00:09:46,686 MANY AGENCIES I MENTIONED 267 00:09:46,686 --> 00:09:49,589 EARLIER. 268 00:09:49,589 --> 00:09:52,192 AND SOME NONPROFITS. 269 00:09:52,192 --> 00:09:53,560 THE POWER OF CITIZEN SCIENCE. 270 00:09:53,560 --> 00:09:55,261 ONE OF THE COOL THINGS ABOUT THE 271 00:09:55,261 --> 00:09:58,965 PROJECT WAS IT WAS CROWD SOURCED 272 00:09:58,965 --> 00:10:01,868 PROOF READING THE CONNECTIONS 273 00:10:01,868 --> 00:10:03,436 WOULD HAVE TAKEN 33 PERSON 274 00:10:03,436 --> 00:10:04,237 YEARS. 275 00:10:04,237 --> 00:10:05,638 BECAUSE OF THIS INTERESTING 276 00:10:05,638 --> 00:10:06,940 APPROACH OF CROWD SOURCING IT 277 00:10:06,940 --> 00:10:09,108 TOOK ONLY A FEW YEARS. 278 00:10:09,108 --> 00:10:11,411 AND FINALLY AS I MENTIONED THE 279 00:10:11,411 --> 00:10:13,246 POWER OF OPEN SCIENCE. 280 00:10:13,246 --> 00:10:15,415 ALL THESE DATA ARE AVAILABLE 281 00:10:15,415 --> 00:10:18,518 THROUGH FREE WEB TOOLS TO ENABLE 282 00:10:18,518 --> 00:10:22,422 USERS TO DELVE INTO THE DATA SET 283 00:10:22,422 --> 00:10:24,390 AND WE THINK IT WILL BE A REAL 284 00:10:24,390 --> 00:10:27,227 ENGINE FOR DISCOVERY. 285 00:10:27,227 --> 00:10:28,595 JUST A PEEK INTO ANOTHER 286 00:10:28,595 --> 00:10:33,399 IMPORTANT PAPER THAT CAME OUT 287 00:10:33,399 --> 00:10:38,404 THIS YEAR FROM THE LICHTMAN LAB 288 00:10:38,404 --> 00:10:41,107 RECONSTRUCTED A 1 CUBIC MILLIMETER 289 00:10:41,107 --> 00:10:45,044 HUMAN TEMPORAL CORTEX ACROSS ALL 290 00:10:45,044 --> 00:10:46,713 LAYERS. AGAIN, A HUGE EFFORT AND 291 00:10:46,713 --> 00:10:48,014 UNDERTAKING BUT THE SCALE AND 292 00:10:48,014 --> 00:10:49,949 RESOLUTION NEEDED TO LOOK AT 293 00:10:49,949 --> 00:10:50,850 THESE DIFFERENT CELL TYPES 294 00:10:50,850 --> 00:10:55,788 ACROSS THESE LAYERS. SO AT ULTRA 295 00:10:55,788 --> 00:10:57,223 STRUCTURAL RESOLUTION ONE CUBIC 296 00:10:57,223 --> 00:11:00,260 MILLIMETER (MM), THEY WERE ABLE TO 297 00:11:00,260 --> 00:11:04,430 RECONSTRUCT 57,000 CELLS, 230 MM OF 298 00:11:04,430 --> 00:11:07,834 BLOOD VESSELS, 150 MILLION SYNAPSES 299 00:11:07,834 --> 00:11:16,476 AT JUST REMARKABLE DETAIL. ALL 300 00:11:16,476 --> 00:11:19,045 THESE ARE AVAILABLE FOR RESEARCHERS 301 00:11:19,045 --> 00:11:20,813 TO EXAMINE AND ASK OWN QUESTIONS. 302 00:11:20,813 --> 00:11:22,015 THOSE ARE A COUPLE OF EXAMPLES 303 00:11:22,015 --> 00:11:23,149 OF THE SCIENCE WE SUPPORT. 304 00:11:23,149 --> 00:11:24,350 I WANT TO SPEND A COUPLE MINUTES 305 00:11:24,350 --> 00:11:26,686 NOW AT THE END OF THIS INTRO TO 306 00:11:26,686 --> 00:11:28,021 TALK ABOUT TWO OF THE CORE 307 00:11:28,021 --> 00:11:29,122 PRINCIPLES THAT ARE VERY 308 00:11:29,122 --> 00:11:30,323 IMPORTANT TO THE BRAIN 309 00:11:30,323 --> 00:11:30,723 INITIATIVE. 310 00:11:30,723 --> 00:11:32,225 THIS IS DATA AND DATA SHARING 311 00:11:32,225 --> 00:11:35,228 AND THE PLATFORMS TO ENABLE THAT 312 00:11:35,228 --> 00:11:37,664 AS WELL AS THE ETHICAL 313 00:11:37,664 --> 00:11:38,331 IMPLICATIONS AND CONSIDERATIONS 314 00:11:38,331 --> 00:11:39,532 OF THE KIND OF WORK THE BRAIN 315 00:11:39,532 --> 00:11:42,569 INITIATIVE SUPPORTS. 316 00:11:42,569 --> 00:11:44,370 SO ONE OF OUR PROGRAMS IS 317 00:11:44,370 --> 00:11:47,473 ACTUALLY TO SUPPORT A LIBRARY OF 318 00:11:47,473 --> 00:11:48,942 DATA ARCHIVES AT DIFFERENT DATA 319 00:11:48,942 --> 00:11:49,309 TYPES. 320 00:11:49,309 --> 00:11:52,078 WE HAVE EIGHT NOW AND A COUPLE 321 00:11:52,078 --> 00:11:53,746 OR FIVE ARE REPRESENTED HERE. 322 00:11:53,746 --> 00:11:56,983 THREE ARE NOT ON THIS SLIDE. 323 00:11:56,983 --> 00:11:59,452 YOU'LL BE HEARING A BIT ABOUT 324 00:11:59,452 --> 00:12:01,421 DABI, ANOTHER ONE FROM 325 00:12:01,421 --> 00:12:02,388 DOMINIQUE DUNCAN A LITTLE LATER. 326 00:12:02,388 --> 00:12:04,123 WE SET THEM UP A FEW YEARS BACK 327 00:12:04,123 --> 00:12:08,595 AND ALL THE DATA FROM THE BRAIN 328 00:12:08,595 --> 00:12:10,797 INITIATIVE PROJECTS ARE 329 00:12:10,797 --> 00:12:13,900 SUBMITTED TO THE DATA ARCHIVES 330 00:12:13,900 --> 00:12:15,468 THEY ARE CLEANED AND AVAILABLE 331 00:12:15,468 --> 00:12:19,439 AND FOR USE BY THE 332 00:12:19,439 --> 00:12:21,674 COMMUNITY AND IT'S BEEN AN 333 00:12:21,674 --> 00:12:22,842 INCREDIBLE UNDERTAKING BUT 334 00:12:22,842 --> 00:12:25,345 ALSO VERY IMPORTANT TO THE BRAIN 335 00:12:25,345 --> 00:12:27,380 INITIATIVE. 336 00:12:27,380 --> 00:12:28,915 NEUROETHICS IS ANOTHER PIECE 337 00:12:28,915 --> 00:12:32,585 YOU'LL HEAR MORE ABOUT. DOCTOR 338 00:12:32,585 --> 00:12:33,786 KAREN ROMMELFANGER WILL TALK IN 339 00:12:33,786 --> 00:12:35,855 SESSION 2 AND GIVE YOU INTRO ON 340 00:12:35,855 --> 00:12:36,155 NEUROETHICS. 341 00:12:36,155 --> 00:12:38,157 I WANT TO MENTION IT HERE BECAUSE 342 00:12:38,157 --> 00:12:39,626 IT'S NOT SOMETHING THAT GETS 343 00:12:39,626 --> 00:12:42,128 A LOT OF AIR TIME OUTSIDE THE 344 00:12:42,128 --> 00:12:43,196 BRAIN INITIATIVE. IT IS IMPORTANT 345 00:12:43,196 --> 00:12:45,565 AND BRAIN IS LEADING THE WAY IN 346 00:12:45,565 --> 00:12:45,765 THIS. 347 00:12:45,765 --> 00:12:47,300 THE STUDY OF ETHICAL ISSUES 348 00:12:47,300 --> 00:12:50,370 RAISED BY THE TECHNOLOGIES THAT 349 00:12:50,370 --> 00:12:52,005 ARE BEING DISCOVERED AND 350 00:12:52,005 --> 00:12:53,239 POTENTIALLY USED BOTH IN 351 00:12:53,239 --> 00:12:58,177 RESEARCH LABS AS WELL AS 352 00:12:58,177 --> 00:13:02,181 POTENTIALLY IN THE CLINIC. WE ARE 353 00:13:02,181 --> 00:13:04,384 PROACTIVE. ONGOING ASSESSMENT OF 354 00:13:04,384 --> 00:13:04,917 THE NEUROETHICAL IMPLICATIONS OF 355 00:13:04,917 --> 00:13:07,687 THE RESEARCH THAT'S DONE. 356 00:13:07,687 --> 00:13:14,027 WE INTEGRATE NEUROETHICS INTO OUR 357 00:13:14,027 --> 00:13:16,429 DECISION MAKING. A WORKING GROUP 358 00:13:16,429 --> 00:13:18,131 DR ROMMELFANGER IS PART OF HAVE 359 00:13:18,131 --> 00:13:18,798 PUBLISHED GUIDING PRINCIPLES 360 00:13:18,798 --> 00:13:20,400 FOR SCIENTISTS AND THE PUBLIC TO 361 00:13:20,400 --> 00:13:21,301 UNDERSTAND THE KINDS OF 362 00:13:21,301 --> 00:13:22,468 QUESTIONS THAT SHOULD BE ASKED 363 00:13:22,468 --> 00:13:24,637 AND THE THINKING THAT GOES 364 00:13:24,637 --> 00:13:26,639 AROUND THE ETHICAL 365 00:13:26,639 --> 00:13:27,306 CONSIDERATIONS. 366 00:13:27,306 --> 00:13:28,808 WE HOLD TOPICAL WORKSHOPS ABOUT 367 00:13:28,808 --> 00:13:30,276 ONE A YEAR ON SOME OF THESE 368 00:13:30,276 --> 00:13:31,811 ISSUES THAT ARE RAISED WHERE WE 369 00:13:31,811 --> 00:13:32,779 GET BROADER INPUT AND BROADER 370 00:13:32,779 --> 00:13:35,381 DISCUSSION. 371 00:13:35,381 --> 00:13:41,254 INTERNALLY WE HAVE A TEAM OF NIH 372 00:13:41,254 --> 00:13:46,259 PROGRAM OFFICERS THAT MEET ONCE A 373 00:13:46,259 --> 00:13:47,860 MONTH TO THINK OF THE NEUROETHICAL 374 00:13:47,860 --> 00:13:48,928 CONSIDERATIONS AND SCIENCE OF UPCOMING 375 00:13:48,928 --> 00:13:51,464 AND PROJECTS ALREADY FUNDED. AND 376 00:13:51,464 --> 00:13:53,032 FINALLY WE HAVE RELEASED FUNDING 377 00:13:53,032 --> 00:13:54,267 ANNOUNCEMENTS WHERE WE FUND 378 00:13:54,267 --> 00:13:55,301 NEUROETHICS RESEARCH IN AREAS 379 00:13:55,301 --> 00:13:57,336 THAT ARE IMPORTANT TO THE BRAIN. 380 00:13:57,336 --> 00:13:59,072 AND I MENTION THESE HERE. WE HAVE 381 00:13:59,072 --> 00:14:01,407 AN RO1 AND NEWLY ADDED R21 PROGRAM 382 00:14:01,407 --> 00:14:11,551 TO EXAMINE ETHICAL IMPLICATIONS 383 00:14:16,322 --> 00:14:17,590 OF NEUROTECHNOLOGIES. A.I. AND 384 00:14:17,590 --> 00:14:20,993 NEUROAI WILL BE COMPONENT OF THIS. 385 00:14:20,993 --> 00:14:23,429 WE ENVISION SEEING APPLICATIONS 386 00:14:23,429 --> 00:14:23,696 ON TOPICS WE WILL DISUCSS TODAY 387 00:14:23,696 --> 00:14:27,166 AND TOMORROW. SO AGAIN, ETHICAL 388 00:14:27,166 --> 00:14:28,735 IMPERATIVES ABOUT A.I. 389 00:14:28,735 --> 00:14:29,936 I WON'T SPEND A LOT OF TIME. 390 00:14:29,936 --> 00:14:32,004 KAREN WILL LEAD THIS DISCUSSION 391 00:14:32,004 --> 00:14:39,479 ON OUR BEHALF OF THE BRAIN 392 00:14:39,479 --> 00:14:42,749 INITIATIVE. THE A.I. TOOLS THROUGH 393 00:14:42,749 --> 00:14:45,885 THE LENS OF NEUROETHIC. 394 00:14:45,885 --> 00:14:49,188 THIS IDEA IT CAN BE IMPEDED BY BLACK 395 00:14:49,188 --> 00:14:55,461 BOX MODELS. ROBUST TRACKING OF 396 00:14:55,461 --> 00:14:57,864 METADATA INCLUDING DATA PROVIDENCE 397 00:14:57,864 --> 00:15:03,002 CONSENT AND WE WANT NEUROETHICS TO 398 00:15:03,002 --> 00:15:03,469 BE PERSON CENTRIC. WITH THAT 399 00:15:03,469 --> 00:15:07,073 FOUNDATION WE HOPE EVERY SESSION HAS 400 00:15:07,073 --> 00:15:08,274 SOME DISCUSSION AROUND NEUROETHICAL 401 00:15:08,274 --> 00:15:08,608 IMPLICATIONS. 402 00:15:08,608 --> 00:15:10,276 AND SO FINALLY FOR ALL OF YOU, 403 00:15:10,276 --> 00:15:14,413 BOTH IN PERSON AND ONLINE, WE 404 00:15:14,413 --> 00:15:16,482 INVITE YOU TO STAY UP TO DATE ON 405 00:15:16,482 --> 00:15:18,451 WHAT'S HAPPENING IN BRAIN 406 00:15:18,451 --> 00:15:18,751 INITIATIVE. 407 00:15:18,751 --> 00:15:24,724 YOU CAN SUBSCRIBE TO OUR BLOG, 408 00:15:24,724 --> 00:15:30,596 WEBSITE IS FULL OF INFORMATION. 409 00:15:30,596 --> 00:15:33,800 MONITOR THE NIH GUIDE NOTICES 410 00:15:33,800 --> 00:15:40,673 CONNECT WITH PROGRAM OFFICERS 411 00:15:40,673 --> 00:15:46,779 IN I.C. OF INTEREST 412 00:15:46,779 --> 00:15:49,849 PAUSE ON THE FUNDING OPPORTUNITIES 413 00:15:49,849 --> 00:15:50,683 PAGE. TAKE PICTURES. WE HAVE ABOUT 414 00:15:50,683 --> 00:15:53,719 20 OR SO ACTIVE FUNDING 415 00:15:53,719 --> 00:15:54,921 OPPORTUNITIES BETWEEN 20 AND 30 416 00:15:54,921 --> 00:16:01,260 USUALLY AT ANY GIVEN TIME. 417 00:16:01,260 --> 00:16:03,763 AND YOU CAN SUBSCRIBE TO THE 418 00:16:03,763 --> 00:16:05,665 GUIDE OR BLOG POST TO GET 419 00:16:05,665 --> 00:16:07,466 NOTIFICATIONS OF THE FUNDING 420 00:16:07,466 --> 00:16:07,800 OPPORTUNITIES. 421 00:16:07,800 --> 00:16:08,401 ALL THIS CAN BE FOUND ON OUR 422 00:16:08,401 --> 00:16:16,976 WEBSITE BRAIN.GOV 423 00:16:16,976 --> 00:16:27,453 AND I WILL TURN IT BACK OVER TO 424 00:16:27,453 --> 00:16:28,154 JOE. 425 00:16:28,154 --> 00:16:28,487 >> THANK YOU ANDREA 426 00:16:28,487 --> 00:16:31,123 THAT WAS EXCELLENT. 427 00:16:31,691 --> 00:16:33,025 >> I'M JOE MONACO A SCIENTIFIC 428 00:16:33,025 --> 00:16:33,926 MANAGER WITH THE BRAIN 429 00:16:33,926 --> 00:16:34,827 INITIATIVE. 430 00:16:35,795 --> 00:16:37,663 BEFORE WE GET TO A COUPLE KEY 431 00:16:37,663 --> 00:16:38,931 NOTES THIS MORNING I WANT TO 432 00:16:38,931 --> 00:16:41,467 GIVE A BRIEF INTRODUCTION TO THE 433 00:16:41,467 --> 00:16:46,305 WORKSHOP, GIVE YOU POINTERS AND 434 00:16:46,305 --> 00:16:51,143 MATERIALS WE PRODUCED THAT 435 00:16:51,143 --> 00:16:52,044 INCORPORATE, MOTIVATE THE 436 00:16:52,044 --> 00:16:53,279 QUESTIONS WE LOOK TO DISCUSS 437 00:16:53,279 --> 00:16:54,647 AND EXPLORE AT THE WORKSHOP. 438 00:16:54,647 --> 00:16:56,215 I WANT TO START WITH THIS HAS 439 00:16:56,215 --> 00:16:58,885 BEEN A LARGE TEAM EFFORT. 440 00:16:58,885 --> 00:17:01,053 THIS IS ALL THE INTERNAL WORKING 441 00:17:01,053 --> 00:17:03,055 GROUP MEMBERS THAT ATTRIBUTED TO 442 00:17:03,055 --> 00:17:05,057 PLANNING THIS WORKSHOP. 443 00:17:05,057 --> 00:17:07,727 I'VE CO-ORGANIZED THIS WITH GRACE 444 00:17:07,727 --> 00:17:09,862 HWANG A PROGRAM DIRECTOR AT NINDS. 445 00:17:09,862 --> 00:17:12,732 I PARTICULARLY WANT TO POINT OUT 446 00:17:12,732 --> 00:17:15,234 COURTNEY PINARD AND JESSICA MOLLICK 447 00:17:15,234 --> 00:17:20,506 WHO DID EXCELLENT WORK LEADING OUR 448 00:17:20,506 --> 00:17:23,609 EARLY-CAREER SCHOLAR EFFORTS AND THE 449 00:17:23,609 --> 00:17:24,343 POSTER SESSION AND COMPETITION. 450 00:17:24,343 --> 00:17:27,747 IT'S BEEN A BIG TEAM EFFORT 451 00:17:27,747 --> 00:17:30,483 WORKING WITH A SET OF EXCELLENT 452 00:17:30,483 --> 00:17:35,488 EXTERNAL ADVISERS AND COMMITTEE 453 00:17:35,488 --> 00:17:40,259 MEMBERS TONY ZADOR, BRAD AIMONE, 454 00:17:40,259 --> 00:17:46,966 DORIS TSAO, GINA ADAM, BLAKE RICHARDS 455 00:17:46,966 --> 00:17:48,134 WE'VE BEEN WORKING TOGETHER FOR 456 00:17:48,134 --> 00:17:50,236 QUITE A WHILE TO PUT TOGETHER THIS 457 00:17:50,236 --> 00:17:51,804 PROGRAM FOR YOU. 458 00:17:51,804 --> 00:17:54,974 IN THE LONGER RUN AS ANDREA 459 00:17:54,974 --> 00:17:57,276 TALKED ABOUT, THE BRAIN INITIATIVE 460 00:17:57,276 --> 00:17:58,711 IS CELEBRATING ITS FIRST DECADE 461 00:17:58,711 --> 00:18:00,980 WE ARE 10 YEARS INTO THE BRAIN 462 00:18:00,980 --> 00:18:01,447 INITIATIVE. 463 00:18:01,447 --> 00:18:04,050 YOU CAN ALSO THINK WE ARE 10 464 00:18:04,050 --> 00:18:08,020 YEARS INTO THE NEW MODERN ERA OF 465 00:18:08,020 --> 00:18:14,026 A.I. AROUND 2012, 2014 AND A.I. 466 00:18:14,026 --> 00:18:16,462 COMPUTING TECHNOLOGY REVOLUTION 467 00:18:16,462 --> 00:18:18,264 HIT ITS INFLECTION POINT AND 468 00:18:18,264 --> 00:18:22,668 ENABLED NEW KINDS OF SCALE AND 469 00:18:22,668 --> 00:18:25,871 DATA DRIVEN COMPUTATION AND SO 470 00:18:25,871 --> 00:18:28,541 THIS IS A MAJOR POINT OF 471 00:18:28,541 --> 00:18:30,810 CONVERGENCE WITH THE BRAIN 472 00:18:30,810 --> 00:18:31,877 INITIATIVE AND A.I. AND IT'S A 473 00:18:31,877 --> 00:18:35,581 GREAT TIME TO THINK BROADLY ABOUT 474 00:18:35,581 --> 00:18:37,850 WHAT WE CAN LEARN FROM THESE DOMAINS 475 00:18:37,850 --> 00:18:41,620 I WANT TO POINT OUT ON A MORE 476 00:18:41,620 --> 00:18:43,723 PERSONAL STORY BRIEFLY HOW I AND 477 00:18:43,723 --> 00:18:49,462 GRACE GOT HERE, THE US BRAIN 478 00:18:49,462 --> 00:18:51,664 INITIATIVE WAS LAUNCHED. NSF 479 00:18:51,664 --> 00:18:52,531 HAD A PORTFOLIO OF PROGRAMS 480 00:18:52,531 --> 00:18:53,866 INCLUDING THE NEURAL AND COGNITIVE 481 00:18:53,866 --> 00:18:59,939 SYSTEMS OR NCS. AND SO A FEW YEARS BACK 482 00:18:59,939 --> 00:19:01,707 WE WERE PART OF A PROJECT TEAM 483 00:19:01,707 --> 00:19:06,278 AND WE TOOK SOME IDEAS FROM THEORETICAL 484 00:19:06,278 --> 00:19:09,448 NEUROSCIENCE AND HIPPOCAMPAL FUNCTION 485 00:19:09,448 --> 00:19:11,083 APPLIED THEM IN AN INTERESTING WAY TO KEY 486 00:19:11,083 --> 00:19:13,285 PROBLEMS IN MULTIAGENT CONTROL AND 487 00:19:13,285 --> 00:19:18,457 AUTONOMOUS SYSTEMS. SO WE FORMED THIS 488 00:19:18,457 --> 00:19:19,925 INTERDISCIPLINARY COLLABORATION WITH 489 00:19:19,925 --> 00:19:24,864 JOHNS HOPKINS APPLIED PHYSICS LAB AND 490 00:19:24,864 --> 00:19:27,967 SAW THE POWER OF RESEARCH THERE. 491 00:19:27,967 --> 00:19:30,202 AND THAT LED TO US MOSTLY GRACE 492 00:19:30,202 --> 00:19:33,472 BUT I HELPED OUT A LITTLE BIT 493 00:19:33,472 --> 00:19:36,976 CO-ORGANIZING A SYMPOSIUM AT THE 2020 494 00:19:36,976 --> 00:19:41,514 BRAIN INITIATIVE INVESTIGATORS MEETING 495 00:19:41,514 --> 00:19:46,652 AROUND THESE ISSUES. HOW DYNAMICAL SYSTEMS IDEAS IN NEUROSCIENCE 496 00:19:46,652 --> 00:19:47,953 CONTRIBUTE TO MACHINE LEARNING. 497 00:19:47,953 --> 00:19:49,488 A.I. WASN'T THE BUZZWORD BACK THEN BUT THAT'S WHAT WE WERE TALKING ABOUT. 498 00:19:49,488 --> 00:19:54,794 THERE'S A QR CODE THAT LEADS TO VIDEO. 499 00:19:54,794 --> 00:20:00,199 HERE'S A SCREEN SHOT OF MY TALK 500 00:20:00,199 --> 00:20:02,635 AND YOU CAN SEE HERE'S TWO COLUMNS 501 00:20:02,635 --> 00:20:03,402 WE ARE ABLE TO SEE ARTIFICIAL 502 00:20:03,402 --> 00:20:05,805 INTELLIGENCE NOW AND START TO COMPARE 503 00:20:05,805 --> 00:20:08,140 THE CAPABILITIES, THE SIMILARITIES 504 00:20:08,140 --> 00:20:09,809 AND DIFFERENCES IN BIOLOGY AND 505 00:20:09,809 --> 00:20:11,577 ABOUT WHAT WE'RE SEEING IN A.I. 506 00:20:11,577 --> 00:20:12,278 AND NETWORK MODELS. 507 00:20:12,278 --> 00:20:13,979 IT'S EASY TO COME UP WITH A LIST 508 00:20:13,979 --> 00:20:22,621 LIKE THIS. MODULARITY, RECURRENCE, TEMPORAL COMPLEXITY 509 00:20:22,621 --> 00:20:27,927 AND THERE'S ALL SORTS OF WAYS 510 00:20:27,927 --> 00:20:30,029 THESE SYSTEMS ARE DIFFERENT. 511 00:20:30,029 --> 00:20:32,064 THERE'S A LOT OF SCIENTIFIC 512 00:20:32,064 --> 00:20:34,133 OPPORTUNITIES AND SO THIS JUNE 513 00:20:34,133 --> 00:20:35,768 AT THE BRAIN CONFERENCE WE HAD A 514 00:20:35,768 --> 00:20:39,205 SPECIALTY SESSION ON EMBODIED 515 00:20:39,205 --> 00:20:39,905 NEUROAI 516 00:20:39,905 --> 00:20:42,508 A SHORT ONE-HOUR PANEL TO GET THE 517 00:20:42,508 --> 00:20:44,176 CONVERSATION STARTED. THERE'S THE 518 00:20:44,176 --> 00:20:49,782 QR CODE THERE TO WATCH THE VIDEO 519 00:20:49,782 --> 00:20:54,920 OF THAT SESSION. SO WE'RE EXCITED 520 00:20:54,920 --> 00:20:56,822 THAT A LOT OF THE FOLKS WE'VE MET 521 00:20:56,822 --> 00:20:58,390 IN THE LAST YEAR IN THE ROOM AND 522 00:20:58,390 --> 00:21:01,460 IN THE VIRTUAL AUDIENCE ARE 523 00:21:01,460 --> 00:21:06,065 HAVING THESE CONVERSATIONS. I DO WANT 524 00:21:06,065 --> 00:21:10,703 TO POINT OUT THIS GAP 2020-2024. 525 00:21:10,703 --> 00:21:11,804 LOT OF INTEREST IN THE FIELD EXPLORING 526 00:21:11,804 --> 00:21:13,973 THESE QUESTIONS. WHY ARE WE SEEING 527 00:21:13,973 --> 00:21:15,241 AMAZING PROGRESS IN THE A.I. 528 00:21:15,241 --> 00:21:16,175 COMPUTING TECHNOLOGY? WHAT DOES 529 00:21:16,175 --> 00:21:18,244 IT MEAN FOR UNDERSTANDING THE BRAIN? 530 00:21:18,244 --> 00:21:20,045 HOW MUCH PROGRESS IS DUE TO WHAT 531 00:21:20,045 --> 00:21:22,715 WE KNOW ABOUT THE BRAIN? 532 00:21:22,715 --> 00:21:24,783 IS THERE A DISCOVERY LOOP THAT 533 00:21:24,783 --> 00:21:27,319 WE SHOULD BE EMPHASIZING HERE IN 534 00:21:27,319 --> 00:21:28,721 ORDER TO DRIVE HOME THE SCORE? 535 00:21:28,721 --> 00:21:32,992 WE WANT TO BRING THESE QUESTIONS 536 00:21:32,992 --> 00:21:34,126 TOGETHER AND A LOT OF THE PEOPLE 537 00:21:34,126 --> 00:21:35,661 PUTTING THESE IDEAS OUT THERE 538 00:21:35,661 --> 00:21:40,966 ARE IN THE ROOM AND IN OUR PANELS. 539 00:21:40,966 --> 00:21:48,507 A FIGURE FROM A PAPER BY BRAD AIMONE 540 00:21:48,507 --> 00:21:52,511 LAID OUT BASICALLY POST BEHAVIORISM 541 00:21:52,511 --> 00:21:54,713 LOOK AT THE TOP TIME LINE. THESE 542 00:21:54,713 --> 00:21:56,248 ARE MAJOR INNOVATIONS IN NEUROSCIENCE 543 00:21:56,248 --> 00:21:59,585 COMING IN THAT HAVE IMPACTED OR 544 00:21:59,585 --> 00:22:02,555 INFLUENCED THE EVOLUTION OF WHAT 545 00:22:02,555 --> 00:22:13,232 WE NOW CALL A.I.: REPERTOIRE 546 00:22:13,232 --> 00:22:16,969 OF APPROACHES OF LEARNING FROM DATA. 547 00:22:16,969 --> 00:22:19,939 SOME OF THIS HAS LED TO A RECENT 548 00:22:19,939 --> 00:22:25,344 NOBEL PRIZE OF COURSE. 549 00:22:25,344 --> 00:22:26,846 BUT HOW DO THESE LINES CONVERGE? 550 00:22:26,846 --> 00:22:27,780 THAT'S THE QUESTION FOR THE WORKSHOP. 551 00:22:27,780 --> 00:22:30,015 WHERE'S IT GO FOR NEUROSCIENCE? 552 00:22:30,015 --> 00:22:33,452 AND HOW CAN THEY BOTH DRIVE EACH OTHER 553 00:22:33,452 --> 00:22:36,488 FORWARD? 554 00:22:36,488 --> 00:22:39,692 ONE OF THE PLACES IT HAS GONE IS THIS NEW SET OF AMAZING A.I. BASED TOOLS THAT WE NOW HAVE IN 555 00:22:39,692 --> 00:22:46,999 NEUROSCIENCE FOR TRACKING BEHAVIOR AND INFORMING LARGE-SCALE POPULATION ANALYSIS, GEOMETRY, 556 00:22:50,202 --> 00:22:52,605 TOPOLOGICAL DATA ANALYSIS OR THE 557 00:22:52,605 --> 00:22:53,706 USE OF RNNS TO UNPACK TIME-SERIES. 558 00:22:53,706 --> 00:22:54,640 THERE'S A LOT OF GREAT TOOLS. 559 00:22:54,640 --> 00:22:56,709 THAT'S NOT WHAT THE WORKSHOP IS 560 00:22:56,709 --> 00:22:56,909 ABOUT. 561 00:22:56,909 --> 00:22:58,077 PEOPLE ARE DOING THIS KIND OF 562 00:22:58,077 --> 00:23:00,679 WORK AND IT WILL CONTINUE TO DO 563 00:23:00,679 --> 00:23:02,047 AMAZING THINGS TO HELP US 564 00:23:02,047 --> 00:23:04,383 UNDERSTAND OUR DATA BUT THIS 565 00:23:04,383 --> 00:23:06,151 WORKSHOP IS ABOUT MORE 566 00:23:06,151 --> 00:23:07,453 FUNDAMENTAL SCIENTIFIC 567 00:23:07,453 --> 00:23:08,420 QUESTIONS. 568 00:23:08,420 --> 00:23:10,055 ARTIFICIAL NEURAL NETWORKS ARE TRAINED. 569 00:23:10,055 --> 00:23:11,824 YOU SPECIFY ARCHITECTURE AND HAVE AN 570 00:23:11,824 --> 00:23:14,426 OBJECTIVE FUNCTION AND YOU HAVE 571 00:23:14,426 --> 00:23:17,963 SOME PROCESS THAT INSTANTIATES THE 572 00:23:17,963 --> 00:23:19,665 LEARNING RULE. THIS IS ALSO A 573 00:23:19,665 --> 00:23:22,835 REVIEW FROM OUR COMMITTEE MEMBER 574 00:23:22,835 --> 00:23:25,137 BLAKE RICHARDS. THOSE THREE 575 00:23:25,137 --> 00:23:27,840 COMPONENTS WE GET ALL THIS DATA 576 00:23:27,840 --> 00:23:29,441 DRIVEN POWER FOR LEARNING AS COMPUTING. 577 00:23:29,441 --> 00:23:30,743 BUT IT'S A DOUBLE-EDGED SWORD 578 00:23:30,743 --> 00:23:34,780 BECAUSE THERE'S NO GUARDRAILS 579 00:23:34,780 --> 00:23:39,118 BUILT INTO THAT. YOU CAN'T PROGRAM 580 00:23:39,118 --> 00:23:40,653 THEM AND EVERYTHING BECOMES DATA 581 00:23:40,653 --> 00:23:41,520 CURATION AND THEN DEALING WITH 582 00:23:41,520 --> 00:23:45,057 THINGS IN THE BACK END OF THE 583 00:23:45,057 --> 00:23:46,025 SYSTEMS PERSPECTIVE. 584 00:23:46,025 --> 00:23:48,661 THIS IS WHERE THE ETHICAL 585 00:23:48,661 --> 00:23:53,432 CONCERN IS AND WHAT WE'LL LEARN 586 00:23:53,432 --> 00:23:54,967 ABOUT IN THE WORKSHOP AND I JUST 587 00:23:54,967 --> 00:23:59,571 WANT TO CLARIFY WE HAVE A LOT OF 588 00:23:59,571 --> 00:24:00,673 AMAZING A.I. TOOLS FOR NEUROSCIENCE. 589 00:24:00,673 --> 00:24:01,974 THE FOCUS OF THE MEETING IS 590 00:24:01,974 --> 00:24:05,144 LOOKING FORWARD TO A SET OF WIDE 591 00:24:05,144 --> 00:24:06,912 RANGING SCIENTIFIC OPPORTUNITIES 592 00:24:06,912 --> 00:24:09,715 TO ADVANCE NEW THEORIES AND 593 00:24:09,715 --> 00:24:11,283 THINK ABOUT NEW COMPLEMENTARY 594 00:24:11,283 --> 00:24:12,584 MODELS AND NEW DISCOVERY LOOPS 595 00:24:12,584 --> 00:24:16,655 TO PUSH ALL THIS FORWARD. THIS 596 00:24:16,655 --> 00:24:19,024 COULD BE A SIMULTANEOUS ADVANCE 597 00:24:19,024 --> 00:24:20,526 IN COMPUTING TECHNOLOGY AND TOOLS 598 00:24:20,526 --> 00:24:21,060 FOR UNDERSTANDING THE BRAIN. 599 00:24:21,060 --> 00:24:24,430 THAT'S REALLY EXCITING. 600 00:24:24,430 --> 00:24:27,299 SO HOW CAN WE BRING THE TROVES OF 601 00:24:27,299 --> 00:24:30,202 DATA BRAIN HAS COLLECTED, LIKE 602 00:24:30,202 --> 00:24:32,971 FLYWIRE, BICAN, CELL ATLASES, AND 603 00:24:32,971 --> 00:24:37,876 BRING THAT TO BEAR INTO THIS PROCESS? 604 00:24:37,876 --> 00:24:38,077 I'M NOT GOING TO GO THROUGH ALL THIS. 605 00:24:38,077 --> 00:24:43,449 THIS IS A SUMMARY FROM THE POSITION 606 00:24:43,449 --> 00:24:45,117 PAPER WHICH IS PART OF THE PROGRAM 607 00:24:45,117 --> 00:24:50,389 BOOK. I ENCOURAGE EVERYONE TO CHECK 608 00:24:50,389 --> 00:24:57,262 OUT THE PROGRAM BOOK. THERE ARE 609 00:24:57,262 --> 00:24:58,797 TRADEOFFS AND WE THINK OF 610 00:24:58,797 --> 00:24:59,465 COMPUTATIONAL NEUROSCIENCE AND HOW 611 00:24:59,465 --> 00:25:01,633 TO SCALE UP. THINK ABOUT THE 612 00:25:01,633 --> 00:25:04,203 LARGE SCALE TRANSFORMERS GETTING 613 00:25:04,203 --> 00:25:06,038 RID OF BIOLOGICAL DETAIL TO SCALE 614 00:25:06,038 --> 00:25:10,376 UP MODEL SIZES BUT WHAT DO YOU LOSE? 615 00:25:10,376 --> 00:25:12,244 SO LET'S THINK OF FUNDAMENTAL 616 00:25:12,244 --> 00:25:13,779 TRADEOFFS. HOW DO WE BREAK THIS 617 00:25:13,779 --> 00:25:17,182 TRIANGLE OF TRADE-OFFS OF SPACE, 618 00:25:17,182 --> 00:25:21,487 TIME AND BIOLOGICAL REALISM. HOW 619 00:25:21,487 --> 00:25:24,056 CAN WE USE BRAIN DATA TO BREAK 620 00:25:24,056 --> 00:25:27,159 THESE BARRIERS TO SCALE, LIKE A.I. 621 00:25:27,159 --> 00:25:29,128 TECHNOLOGIES HAVE BROKEN SCALE 622 00:25:29,128 --> 00:25:31,463 BARRIERS? THAT'S THE BASIS OF THE 623 00:25:31,463 --> 00:25:33,899 DISCUSSIONS WE WANT TO HAVE HERE. 624 00:25:33,899 --> 00:25:35,901 SO TO HEAR MORE ABOUT THE 625 00:25:35,901 --> 00:25:37,669 RECIPROCITY AT THE HEART OF 626 00:25:37,669 --> 00:25:48,213 NEUROAI, LET'S TURN IT OVER TO 627 00:25:53,519 --> 00:26:03,562 TONY ZADOR. 628 00:26:19,545 --> 00:26:23,916 >> HELLO. 629 00:26:23,916 --> 00:26:25,751 FIRST, LET ME THANK THE 630 00:26:25,751 --> 00:26:27,119 ORGANIZERS AND ESPECIALLY JOE 631 00:26:27,119 --> 00:26:28,987 AND GRACE FOR PUTTING TOGETHER 632 00:26:28,987 --> 00:26:31,590 WHAT I THINK IS A REALLY 633 00:26:31,590 --> 00:26:35,027 EXCITING TWO-DAY WORKSHOP. 634 00:26:35,027 --> 00:26:37,663 I'M GOING TO SORT OF TAKE THE 635 00:26:37,663 --> 00:26:40,065 VIEW FROM 30,000 FEET AND GIVE 636 00:26:40,065 --> 00:26:44,136 AN OVERVIEW OF WHAT I THINK THE 637 00:26:44,136 --> 00:26:46,972 WHOLE FIELD OF NEUROAI IS ABOUT 638 00:26:46,972 --> 00:26:52,277 AND IT'S BASICALLY I THINK 639 00:26:52,277 --> 00:26:54,646 SHARING THE INSIGHT FROM 640 00:26:54,646 --> 00:26:57,416 NEUROSCIENCE TO AFFECT A.I. AND 641 00:26:57,416 --> 00:26:59,117 A.I. TO INFLUENCE NEUROSCIENCE. 642 00:26:59,117 --> 00:27:03,188 I THINK HISTORICALLY THAT'S BEEN 643 00:27:03,188 --> 00:27:05,224 A REALLY FRUITFUL TWO-WAY STREET 644 00:27:05,224 --> 00:27:08,961 THAT IS SOMETIMES IN THE MODERN 645 00:27:08,961 --> 00:27:10,662 ERA FORGOTTEN ABOUT. 646 00:27:10,662 --> 00:27:12,564 I THINK WE'RE ALL FAMILIAR WITH 647 00:27:12,564 --> 00:27:15,667 THE TREMENDOUS ADVANCES OF A.I. 648 00:27:15,667 --> 00:27:20,739 IN THE LAST DECADE AND 649 00:27:20,739 --> 00:27:21,807 ESPECIALLY THE LAST FEW YEARS. 650 00:27:21,807 --> 00:27:23,642 WHEN I WAS A GRADUATE STUDENT 651 00:27:23,642 --> 00:27:26,512 THE IDEA WE WOULD EVER HAVE AN 652 00:27:26,512 --> 00:27:30,415 A.I. SYSTEM THAT PASSED THE TEST 653 00:27:30,415 --> 00:27:32,618 SEEMED FAR FETCHED. 654 00:27:32,618 --> 00:27:36,021 AND THE FACT IS WE EFFECTIVELY 655 00:27:36,021 --> 00:27:38,023 DO HAVE ONE THAT PASSES THE TURING 656 00:27:38,023 --> 00:27:41,360 TEST. CHATBOTS CAN MIMIC THE 657 00:27:41,360 --> 00:27:44,096 ANSWERS THAT A HUMAN WOULD GIVE. 658 00:27:44,096 --> 00:27:46,031 SO I ASKED CHATGPT WHETHER IT WOULD 659 00:27:46,031 --> 00:27:47,666 PASS THE TURING TEST AND IT GAVE 660 00:27:47,666 --> 00:27:51,503 A LONG ANSWER BUT BASICALLY IT 661 00:27:51,503 --> 00:27:56,275 WAS BEING MODEST. 662 00:27:56,275 --> 00:27:57,843 IT'S CLEAR IT CAN AND DOES. 663 00:27:57,843 --> 00:27:59,845 WE'RE FAMILIAR WITH THE TREMENDOUS 664 00:27:59,845 --> 00:28:02,881 ADVANCES OF A.I. THIS IS 665 00:28:02,881 --> 00:28:05,183 PROTEIN FOLDING WHICH WAS AWARDED 666 00:28:05,183 --> 00:28:13,492 THE NOBEL PRIZE A MONTH AGO. 667 00:28:13,492 --> 00:28:16,094 AS WE LOOK AT THE HISTORY OF 668 00:28:16,094 --> 00:28:17,496 NEUROSCIENCE AND A.I., YOU CAN 669 00:28:17,496 --> 00:28:18,797 TAKE IT BACK TO THE BEGINNING OF 670 00:28:18,797 --> 00:28:20,432 THE COMPUTING ERA. I DON'T HAVE 671 00:28:20,432 --> 00:28:22,167 TIME TO GO THROUGH ALL OF THESE 672 00:28:22,167 --> 00:28:25,270 BUT I WILL SAY THE FIRST PAPER 673 00:28:25,270 --> 00:28:27,072 OUTLINING THE DESIGN OF THE MODERN 674 00:28:27,072 --> 00:28:34,680 COMPUTER BY VON NEUMANN CITED A 675 00:28:34,680 --> 00:28:41,987 PAPER BY MCCULLOCH AND PITTS WHICH 676 00:28:41,987 --> 00:28:42,588 LAYS THE FOUNDATION OF NEURAL 677 00:28:42,588 --> 00:28:44,323 NETWORKS. THE TWO FIELDS, 678 00:28:44,323 --> 00:28:45,757 COMPUTATION AND NEUROSCIENCE HAVE 679 00:28:45,757 --> 00:28:48,994 BEEN INTERTWINED SINCE THEIR 680 00:28:48,994 --> 00:28:52,264 INCEPTION. 681 00:28:52,264 --> 00:28:55,033 AND OF COURSE, THIS YEAR THE 682 00:28:55,033 --> 00:28:56,268 NOBEL PRIZE WAS A DIFFERENT 683 00:28:56,268 --> 00:28:57,336 NOBEL PRIZE WAS AWARDED FOR 684 00:28:57,336 --> 00:29:02,641 NEURAL NETWORKS. 685 00:29:02,641 --> 00:29:05,677 SO NEUROSCIENCE AND A.I. HAVE 686 00:29:05,677 --> 00:29:09,381 HAD A VIRTUOUS CIRCLE WHERE 687 00:29:09,381 --> 00:29:11,249 NEUROSCIENCE INFLUENCES A.I. AND 688 00:29:11,249 --> 00:29:12,517 THE A.I. PROVIDES PROBABLY THE 689 00:29:12,517 --> 00:29:15,120 BEST MODELS FOR HOW TO THINK 690 00:29:15,120 --> 00:29:17,322 ABOUT HOW BRAINS COMPUTE. 691 00:29:17,322 --> 00:29:19,992 SO THAT VIRTUOUS CIRCLE HAS BEEN 692 00:29:19,992 --> 00:29:23,295 GOING ON FOR DECADES AND HAS 693 00:29:23,295 --> 00:29:27,599 BEEN SELF-CONSUMER AND USEFUL 694 00:29:27,599 --> 00:29:32,004 PRODUCTS LIKE CHATGPT AND DALLE 695 00:29:32,004 --> 00:29:33,472 AND SOMETIMES THE TOOLS CAN BE 696 00:29:33,472 --> 00:29:36,508 USED FOR SCIENTIFIC RESEARCH. 697 00:29:36,508 --> 00:29:39,044 SO THE WAY THAT I THINK ABOUT 698 00:29:39,044 --> 00:29:40,946 THE MODELS THAT WE HAVE IN 699 00:29:40,946 --> 00:29:42,948 NEURAL NETWORKS IN A.I. CAN BE 700 00:29:42,948 --> 00:29:47,352 TRACED BACK TO THIS QUOTE BY 701 00:29:47,352 --> 00:29:50,889 RICHARD FEYNMAN WHO WROTE WHAT I 702 00:29:50,889 --> 00:29:52,224 CANNOT CREATE I DO NOT 703 00:29:52,224 --> 00:29:52,624 UNDERSTAND. 704 00:29:52,624 --> 00:29:54,192 IT TURNS OUT WHEN WE TRY TO 705 00:29:54,192 --> 00:29:55,961 UNDERSTAND HOW BRAINS COMPUTE 706 00:29:55,961 --> 00:30:00,999 AND SOLVE PROBLEMS. WE CAN HAVE 707 00:30:00,999 --> 00:30:05,003 ALL SORTS OF IDEAS BUT 708 00:30:05,003 --> 00:30:06,872 IMPLEMENTING THEM IS WHERE THE 709 00:30:06,872 --> 00:30:08,473 RUBBER MEETS THE ROADS. WHERE 710 00:30:08,473 --> 00:30:13,645 WE DISCOVER WHETHER THE IDEAS WE 711 00:30:13,645 --> 00:30:14,646 HAVE FOR HOW BRAINS COMPUTE. THAT'S 712 00:30:14,646 --> 00:30:17,182 WHERE WE DISCOVER IF THEY WORK 713 00:30:17,182 --> 00:30:17,749 WHETHER THEY SCALE OR MAKE 714 00:30:17,749 --> 00:30:27,859 SENSE. OFTEN WE'RE CONFRONTED 715 00:30:28,493 --> 00:30:30,829 WITH MORAVEC'S PARADOX 716 00:30:30,829 --> 00:30:37,636 WRITTEN 40 YEARS AGO. I WON'T 717 00:30:37,636 --> 00:30:48,146 READ THE WHOLE QUOTE. WHAT ANIMALS 718 00:30:48,980 --> 00:30:51,083 DO IS HARD. WHAT WE THINK AS THE 719 00:30:51,083 --> 00:30:53,618 HARDEST PART, COGNITION AND 720 00:30:53,618 --> 00:30:55,053 REASONING ARE THE EASIEST PART. 721 00:30:55,053 --> 00:30:58,790 THE THINGS A.I. ARE GOOD AT ARE 722 00:30:58,790 --> 00:31:01,093 UNIQUELY HUMAN OR THINGS WE 723 00:31:01,093 --> 00:31:02,627 PRIDE OURSELVES AT BEING 724 00:31:02,627 --> 00:31:04,096 PARTICULARLY GOOD AT. THINGS WE 725 00:31:04,096 --> 00:31:08,767 THINK SMART PEOPLE ARE GOOD AT 726 00:31:08,767 --> 00:31:11,136 THESE THINGS AND I'M SURE YOU 727 00:31:11,136 --> 00:31:15,373 ALL KNOW A.I. SYSTEMS BEAT THE 728 00:31:15,373 --> 00:31:16,842 BEST CHEST AND GO PLAYERS. THEY'RE 729 00:31:16,842 --> 00:31:26,985 GETTING BETTER AT MATH AND ABLE TO 730 00:31:26,985 --> 00:31:28,687 PRODUCE GREAT SENTENCES. BUT THE 731 00:31:28,687 --> 00:31:32,724 THING THEY CAN'T DO ARE SENSING, 732 00:31:32,724 --> 00:31:36,027 INTERACTING, PLANNING, AND USING A 733 00:31:36,027 --> 00:31:38,497 WORLD MODEL, ACTING AUTONOMOUSLY 734 00:31:38,497 --> 00:31:41,466 FOR AN ENTIRE LIFE TIME. 735 00:31:41,466 --> 00:31:42,868 MOST A.I. SYSTEMS CANNOT DO 736 00:31:42,868 --> 00:31:46,638 THESE THINGS OR DO THEM WELL. 737 00:31:46,638 --> 00:31:49,074 AND SO THAT LED US A FEW YEARS 738 00:31:49,074 --> 00:31:55,013 AGO TO PROPOSE A NEW CHALLENGE 739 00:31:55,013 --> 00:31:57,816 WHICH IS THE EMBODIED TURING TEST 740 00:31:57,816 --> 00:31:59,785 THAT REQUIRES NOT ONLY THE A.I. 741 00:31:59,785 --> 00:32:01,119 SYSTEMS BE ABLE TO INTERACT WITH 742 00:32:01,119 --> 00:32:05,390 THE SENSORIMOTOR WORLD BUT BEYOND 743 00:32:05,390 --> 00:32:07,726 THAT THEY SHOULD BE ABLE 744 00:32:07,726 --> 00:32:14,766 TO ACT AUTONOMOUSLY AND PLAN AND 745 00:32:14,766 --> 00:32:19,738 THESE ARE EXAMPLES OF SCIENCE 746 00:32:19,738 --> 00:32:20,338 FICTION DEPICTIONS OF A.I. 747 00:32:20,338 --> 00:32:21,473 SYSTEMS THAT CAN DO THAT AND 748 00:32:21,473 --> 00:32:28,146 WE ARE STILL FAR FROM THAT. 749 00:32:28,146 --> 00:32:30,615 WHAT IS NEUROAI? NEUROAI IS THE 750 00:32:30,615 --> 00:32:36,988 MEETING OF THESE TWO STREAMS. 751 00:32:36,988 --> 00:32:38,890 NEUROSCIENCE HAS INSPIRED AI 752 00:32:38,890 --> 00:32:40,492 AND WILL LIKELY DRIVE PROGRESS. 753 00:32:40,492 --> 00:32:46,331 I'LL TOUCH UPON A FEW OF WHAT I 754 00:32:46,331 --> 00:32:48,400 SEE AS THE CHALLENGES AND 755 00:32:48,400 --> 00:32:51,203 OPPORTUNITIES DISCUSSED IN 756 00:32:51,203 --> 00:32:55,106 GREATER DETAIL IN THE VARIOUS 757 00:32:55,106 --> 00:32:56,208 PANELS. 758 00:32:56,208 --> 00:33:00,045 SO THE FIRST IS THE MORAVEC'S 759 00:33:00,045 --> 00:33:02,881 PARADOX LEARNING THE WAY ANIMALS 760 00:33:02,881 --> 00:33:05,917 DO -- I'LL TALK BRIEFLY ABOUT A 761 00:33:05,917 --> 00:33:08,119 SECOND IMPORTANT OPPORTUNITY AND 762 00:33:08,119 --> 00:33:18,630 CHALLENGE WHICH IS ENERGY USE. 763 00:33:20,165 --> 00:33:21,466 LOWEST HANGING FRUIT. A.I. SYSTEMS 764 00:33:21,466 --> 00:33:22,367 USE TREMENDOUS AMOUNTS OF ENERGY. 765 00:33:22,367 --> 00:33:25,370 THE AMOUNTS OF ENERGY REQUIRED 766 00:33:25,370 --> 00:33:28,306 FOR THESE SYSTEMS TO BE TRAINED 767 00:33:28,306 --> 00:33:32,010 UP NOW IS A SIGNIFICANT ITEM ON 768 00:33:32,010 --> 00:33:33,245 THE WORLD'S ENERGY BUDGET AND 769 00:33:33,245 --> 00:33:38,416 IT'S GOING UP RAPIDLY. 770 00:33:38,416 --> 00:33:40,552 SO THERE'S DIFFERENT WAYS OF 771 00:33:40,552 --> 00:33:42,854 ESTIMATING THIS. BY SOME MEASURES 772 00:33:42,854 --> 00:33:45,490 BRAINS ARE FOUR ORDERS OF 773 00:33:45,490 --> 00:33:52,130 MAGNITUDE MORE ENERGY EFFICIENT 774 00:33:52,130 --> 00:33:55,600 THAN A.I. SYSTEMS. 775 00:33:55,600 --> 00:33:58,904 THE AMOUNT OF ENERGY IT TAKES TO 776 00:33:58,904 --> 00:34:01,006 TRAIN A LARGE LANGUAGE MODEL IS 777 00:34:01,006 --> 00:34:02,007 NOT THERE BECAUSE IT'S 778 00:34:02,007 --> 00:34:06,978 PROPRIETARY BUT ON THE ORDER OF 779 00:34:06,978 --> 00:34:08,613 ENERGY USED TO RUN A MEDIUM 780 00:34:08,613 --> 00:34:11,716 SIZED CITY FOR A WEEK -- IT 781 00:34:11,716 --> 00:34:15,854 DOESN'T TAKE THAT MUCH ENERGY TO 782 00:34:15,854 --> 00:34:18,924 TRAIN A CHILD TO SPEAK. 783 00:34:18,924 --> 00:34:21,126 IT LEARNS WITH THE AMOUNT OF 784 00:34:21,126 --> 00:34:22,227 FOOD IT HAS IN THE FIRST TWO 785 00:34:22,227 --> 00:34:24,429 YEARS OF LIFE SO WHAT 786 00:34:24,429 --> 00:34:28,833 CONTRIBUTES TO THE ENERGY 787 00:34:28,833 --> 00:34:29,301 EFFICIENCY? 788 00:34:29,301 --> 00:34:30,702 PART IS THE HARDWARE AND OTHER 789 00:34:30,702 --> 00:34:30,902 STUFF. 790 00:34:30,902 --> 00:34:35,040 WE CAN TALK ABOUT THE HARDWARE. 791 00:34:35,040 --> 00:34:36,975 SO THE REASON A.I. SYSTEMS WE 792 00:34:36,975 --> 00:34:40,712 USE TODAY RUN ON THE KINDS OF 793 00:34:40,712 --> 00:34:41,646 HARDWARE THAT WE USE TODAY CAN 794 00:34:41,646 --> 00:34:43,882 BE ATTRIBUTED IN PART TO 795 00:34:43,882 --> 00:34:47,252 SOMETHING THAT SARAH HOOKER IN A 796 00:34:47,252 --> 00:34:48,954 BEAUTIFUL ARTICLE REFERS TO AS 797 00:34:48,954 --> 00:34:50,355 THE HARDWARE LOTTERY. 798 00:34:50,355 --> 00:34:52,023 THE ALGORITHMS WE DEVELOP, THE 799 00:34:52,023 --> 00:34:54,926 WAY WE DEVELOP AND THE ONES THAT 800 00:34:54,926 --> 00:34:57,495 WORK AND DON'T ARE NOT 801 00:34:57,495 --> 00:34:58,396 INDEPENDENT OF THE HARDWARE THEY 802 00:34:58,396 --> 00:34:59,931 RUN ON. 803 00:34:59,931 --> 00:35:04,302 IT TURNS OUT MOST MODERN A.I. 804 00:35:04,302 --> 00:35:07,172 RUNS ON GPUs ORIGINALLY 805 00:35:07,172 --> 00:35:10,442 DEVELOPED TO PROCESS IMAGES. 806 00:35:10,442 --> 00:35:11,509 ONE CAN IMAGINE A DIFFERENT 807 00:35:11,509 --> 00:35:17,182 WORLD IN WHICH DIFFERENT KINDS 808 00:35:17,182 --> 00:35:19,017 OF HARDWARE WOULD BE USEFUL OR 809 00:35:19,017 --> 00:35:21,519 DEVELOP THAT WOULD PROMOTE THE 810 00:35:21,519 --> 00:35:24,923 USE OF MORE BIO INSPIRED TYPE 811 00:35:24,923 --> 00:35:26,524 ALGORITHMS. 812 00:35:26,524 --> 00:35:28,059 WE KIND OF KNOW WHAT THE 813 00:35:28,059 --> 00:35:28,493 DIFFERENCES ARE. 814 00:35:28,493 --> 00:35:31,229 WE DON'T YET KNOW HOW TO BUILD 815 00:35:31,229 --> 00:35:34,933 THESE HARDWARE AND DON'T KNOW 816 00:35:34,933 --> 00:35:36,201 HOW TO DEVELOP ALGORITHMS 817 00:35:36,201 --> 00:35:38,803 BECAUSE WE DON'T YET HAVE THE 818 00:35:38,803 --> 00:35:39,471 HARDWARE ON WHICH TO TEST THESE 819 00:35:39,471 --> 00:35:49,581 IDEAS. I'LL MENTION A THIRD 820 00:35:49,781 --> 00:35:50,815 OPPORTUNITY AND CHALLENGE 821 00:35:50,815 --> 00:35:53,184 REFERRED TO AS THE ALIGNMENT 822 00:35:53,184 --> 00:35:53,618 PROBLEM. 823 00:35:53,618 --> 00:35:55,220 HOW DO WE GET A.I. SYSTEMS TO DO 824 00:35:55,220 --> 00:35:59,024 WHAT WE WANT THEM TO? THE EXTREME 825 00:35:59,024 --> 00:36:04,763 VERSION OF THIS IS CALLED THE 826 00:36:04,763 --> 00:36:06,431 PAPER CLIP PROBLEM. IT'S A 827 00:36:06,431 --> 00:36:11,469 THHOUGHT EXPERIMENT PROPOSED 828 00:36:11,469 --> 00:36:13,138 BY NICK BOSTROM ABOUT 20 829 00:36:13,138 --> 00:36:18,209 YEARS AGO IN WHICH SOMEBODY 830 00:36:18,209 --> 00:36:23,181 TRAINS AN A.I. SYSTEM TO MAXIMIZE 831 00:36:23,181 --> 00:36:26,284 THE NUMBERS OF PAPER CLIPS AND THE 832 00:36:26,284 --> 00:36:31,356 A.I. GETS OUT OF CONTROL AND IT 833 00:36:31,356 --> 00:36:33,358 REALIZES THE WAY TO MAXIMIZE IS 834 00:36:33,358 --> 00:36:35,760 TURN ALL MATTER ON EARTH TO 835 00:36:35,760 --> 00:36:37,696 PAPER CLIPS. 836 00:36:37,696 --> 00:36:39,431 THAT'S AN EXAMPLE OF AN A.I. 837 00:36:39,431 --> 00:36:41,466 SYSTEM NOT ALIGNED -- DOING WHAT 838 00:36:41,466 --> 00:36:42,667 WE ASKED IT TO DO BUT NOT WHAT 839 00:36:42,667 --> 00:36:48,239 WE WANTED IT TO DO. 840 00:36:48,239 --> 00:36:49,474 THERE ARE MANY POSSIBLE 841 00:36:49,474 --> 00:36:51,376 EXPLANATIONS, WAYS AROUND THIS. 842 00:36:51,376 --> 00:36:53,478 ONE INSPIRATION ONE WAY AROUND 843 00:36:53,478 --> 00:36:55,013 THIS FROM NEUROSCIENCE COMES 844 00:36:55,013 --> 00:36:56,414 FROM THE OBSERVATION THAT 845 00:36:56,414 --> 00:37:01,019 ANIMALS DON'T PURSUE A SINGLE ROLE 846 00:37:01,019 --> 00:37:02,821 THEY PURSUE MULTIPLE ROLES. 847 00:37:02,821 --> 00:37:06,224 REFERRED TO AS THE FOUR Fs: 848 00:37:06,224 --> 00:37:08,560 FEEDING, FLEEING, FIGHTING AND 849 00:37:08,560 --> 00:37:13,698 FAMILY. WE DON'T KNOW HOW TO BUILD 850 00:37:13,698 --> 00:37:21,139 SYSTEMS THAT PURSUE MULTIPLE GOALS. 851 00:37:21,139 --> 00:37:23,908 I WILL CLOSE BY SUMMARIZING WHAT WE 852 00:37:23,908 --> 00:37:29,047 MIGHT WANT OUR FUTURE A.I. SYSTEMS 853 00:37:29,047 --> 00:37:36,454 TO DO. TO RESOLVE MORAVEC'S PARADOX 854 00:37:36,454 --> 00:37:37,989 TO BE MORE ENERGY EFFICIENT AND 855 00:37:37,989 --> 00:37:39,257 BETTER ALIGNED TO OUR GOALS. 856 00:37:39,257 --> 00:37:41,759 I'LL CLOSE WITH A SHAMELESS PLUG 857 00:37:41,759 --> 00:37:45,463 NOW THAT I HAVE A PLATFORM. AT 858 00:37:45,463 --> 00:37:49,000 COLD SPRING HARBOR WE HAVE A 859 00:37:49,000 --> 00:37:50,869 NEUROAI PROGRAM FOR GRADUATE 860 00:37:50,869 --> 00:37:52,670 STUDENTS ALREADY INVOLVED IN A.I. 861 00:37:52,670 --> 00:37:54,572 TO SPEND THE SUMMER THERE. 862 00:37:54,572 --> 00:37:57,842 THIS IS THE NEUROAI INTERN 863 00:37:57,842 --> 00:37:59,611 PROGRAM AND AN INDEPENDENT 864 00:37:59,611 --> 00:38:01,713 SCHOLARS PROGRAM FOR THOSE WHO 865 00:38:01,713 --> 00:38:05,383 FINISHED THEIR TRAINING IN A.I. 866 00:38:05,383 --> 00:38:07,051 AND WOULD LIKE TO LEARN MORE 867 00:38:07,051 --> 00:38:07,385 ABOUT NEUROSCIENCE. 868 00:38:07,385 --> 00:38:08,853 WITH THAT I'D LIKE TO TURN THE 869 00:38:08,853 --> 00:38:13,024 FLOOR BACK OVER TO 870 00:38:13,024 --> 00:38:21,499 JOE. 871 00:38:21,499 --> 00:38:28,139 >> I'D LIKE TO TURN THE FLOOR 872 00:38:28,139 --> 00:38:38,683 OVER TO DIMITRI YATSENKO CEO 873 00:38:40,318 --> 00:38:47,292 AND CO-FOUNDER OF DATA JOINT 874 00:38:47,292 --> 00:38:49,460 >> GOOD MORNING IT'S GREAT TO BE 875 00:38:49,460 --> 00:38:53,998 HERE TO TALK ABOUT NEUROSCIENCE 876 00:38:53,998 --> 00:38:58,403 AND A.I. AND WHAT IT MEANS TO DO 877 00:38:58,403 --> 00:39:00,271 NEUROAI TOGETHER AND TALK ABOUT 878 00:39:00,271 --> 00:39:00,872 INFRASTRUCTURE OPERATIONS AND 879 00:39:00,872 --> 00:39:06,778 COLLABORATION. WE'VE COLLABORATED 880 00:39:06,778 --> 00:39:10,748 WITH VARIOUS PROJECTS TO DEFINE A 881 00:39:10,748 --> 00:39:13,484 UNIFIED FRAMEWORK FOR MANAGING 882 00:39:13,484 --> 00:39:18,990 DATA AND COMPUTATIONS AND CODE 883 00:39:18,990 --> 00:39:21,859 AND INTEGRATING THEM INTO 884 00:39:21,859 --> 00:39:26,197 STANDARDIZED PIPELINES FOR 885 00:39:26,197 --> 00:39:27,232 EXPERIMENTS. 886 00:39:27,232 --> 00:39:27,899 WHAT MAKES NEUROSCIENCE SO 887 00:39:27,899 --> 00:39:33,304 COMPLEX MORE THAN ANY OTHER 888 00:39:33,304 --> 00:39:35,873 FIELD IS IT'S THE MOST MODALITY 889 00:39:35,873 --> 00:39:37,008 RICH FIELD. WE COMBINE GENETIC AND 890 00:39:37,008 --> 00:39:43,281 NEUROPHYSIOLOGY, ANATOMICAL DATA, 891 00:39:43,281 --> 00:39:45,783 AND MOLECULAR METHODS. 892 00:39:45,783 --> 00:39:48,987 AND A.I. METHODS RIGHT 893 00:39:48,987 --> 00:39:50,588 NOW ARE BECOMING NOWHERE CLOSE 894 00:39:50,588 --> 00:39:51,522 TO THE RICHNESS OF THE DATA WE 895 00:39:51,522 --> 00:39:52,957 HAVE IN NEUROSCIENCE. 896 00:39:52,957 --> 00:39:55,059 SO HOW DO WE HANDLE THIS? 897 00:39:55,059 --> 00:40:00,265 RIGHT NOW THE APPROACH HAS BEEN 898 00:40:00,265 --> 00:40:02,433 AS A COMMUNITY DEVELOPING A LOT 899 00:40:02,433 --> 00:40:05,770 OF NEUROSCIENCE TOOLS. YOU WOULD 900 00:40:05,770 --> 00:40:11,276 SEE THESE MULTIMODEL EXPERIMENTS. 901 00:40:11,276 --> 00:40:15,413 THERE'S A SLIDE HERE FROM ODIN 902 00:40:15,413 --> 00:40:16,981 OPEN DATA NEUROSCIENCE INITIATIVE 903 00:40:16,981 --> 00:40:25,023 FROM MY TEAM AT McGOVERN INSTITUTE 904 00:40:25,023 --> 00:40:27,125 AND HAS A LOT OF OPEN SOURCE 905 00:40:27,125 --> 00:40:29,761 TOOLS AROUND THE DANDI ARCHIVE 906 00:40:29,761 --> 00:40:35,633 BUT DATA KEEPS GROWING AT A VERY 907 00:40:35,633 --> 00:40:38,436 FAST RATE. HARD TO ESTIMATE. 908 00:40:38,436 --> 00:40:42,206 WE FOLLOW MOORE'S LAW IN DATA 909 00:40:42,206 --> 00:40:43,174 IN NEUROSCIENCE DOUBLING EVERY 910 00:40:43,174 --> 00:40:45,777 18 MONTHS OR SO. BUT THE NUMBER 911 00:40:45,777 --> 00:40:47,645 OF FINDINGS DOESN'T GROW AS 912 00:40:47,645 --> 00:40:48,413 FAST. 913 00:40:48,413 --> 00:40:53,051 DOUBLING EVERY 10, 15 YEARS. 914 00:40:53,051 --> 00:40:55,853 THAT MEANS EVERY STUDY REQUIRES 915 00:40:55,853 --> 00:40:59,524 MORE DATA AND OPERATIONS. 916 00:40:59,524 --> 00:41:07,865 WE NEED TO SHIFT GEARS. SOMETIMES 917 00:41:07,865 --> 00:41:09,033 I THINK WE'RE LIVING IN SOMEBODY 918 00:41:09,033 --> 00:41:10,468 ELSE'S DREAM. WE WOULD THINK SO 919 00:41:10,468 --> 00:41:12,170 WE'RE LIVING IN A FUTURE WHERE THE 920 00:41:12,170 --> 00:41:18,009 TYPES OF EXPERIMENTS FRANCIS CRICK 921 00:41:18,009 --> 00:41:19,377 THOUGHT WERE IMPOSSIBLE ARE 922 00:41:19,377 --> 00:41:20,812 NOW BECOMING THE VERY SUBJECT 923 00:41:20,812 --> 00:41:29,253 OF OUR STUDIES. 924 00:41:29,253 --> 00:41:32,290 SO, WITH THIS COMBINATION OF 925 00:41:32,290 --> 00:41:38,329 NEUROSCIENCE AND A.I., NEUROAI 926 00:41:38,329 --> 00:41:41,899 ALLOWS TRANSFER BETWEEN DOMAINS 927 00:41:41,899 --> 00:41:45,503 WE CAN LEARN THE LESSONS FROM A.I. 928 00:41:45,503 --> 00:41:48,506 AND APPLY THEM TO NEUROSCIENCE 929 00:41:48,506 --> 00:41:51,009 EVEN TRANSFER THE APPROACH OF THE 930 00:41:51,009 --> 00:41:57,448 SCIENTIFIC METHOD. SO CLOSING THE 931 00:41:57,448 --> 00:42:00,351 DISCOVERY LOOP IS BECOMING 932 00:42:00,351 --> 00:42:01,919 POSSIBLE WITH MACHINE LEARNNG IN 933 00:42:01,919 --> 00:42:04,188 THE EXPERIMENT. NOT JUST LEARNING 934 00:42:04,188 --> 00:42:08,192 SEPARATELY BUT COMBINING STUDIES 935 00:42:08,192 --> 00:42:09,227 WHERE NEUROSCIENCE AND A.I. ARE 936 00:42:09,227 --> 00:42:10,428 WORKING TOGETHER. 937 00:42:10,428 --> 00:42:14,766 IT BECOMES POSSIBLE TO CLOSE THE 938 00:42:14,766 --> 00:42:15,433 DISCOVERY LOOP COVERING SCIENCE 939 00:42:15,433 --> 00:42:21,472 OR PORTIONS OF SCIENCE THAT WERE 940 00:42:21,472 --> 00:42:23,207 EXCLUSIVELY HUMAN. THINGS LIKE 941 00:42:23,207 --> 00:42:25,376 OBSERVING PHENOMENA, FORMULATING 942 00:42:25,376 --> 00:42:27,111 HYPOTHESIS ARE BECOMING PART OF THE 943 00:42:27,111 --> 00:42:30,014 AUTOMATED AND FORMALIZED APPROACH. 944 00:42:30,014 --> 00:42:30,281 ONE EXAMPLE IS FROM ANDREAS TOLIAS' 945 00:42:30,281 --> 00:42:37,989 LAB. USING MACHINE LEARNING IN THE LOOP WHERE A NEURAL NETWORK 946 00:42:37,989 --> 00:42:48,933 APPROXIMATES ACTIVITY OF A NEURAL CIRCUIT TO DEFINE THE NEXT EXPERIMENT 947 00:42:53,237 --> 00:43:03,548 SO THIS REQUIRES WE FIGURE OUT 948 00:43:03,548 --> 00:43:08,219 WHAT MADE THE A.I. REVOLUTION 949 00:43:08,219 --> 00:43:10,421 POSSIBLE. NOT JUST ALGORITHMS BUT 950 00:43:10,421 --> 00:43:11,622 INCREASED OPERATIONAL EFFICIENCY. 951 00:43:11,622 --> 00:43:16,994 WHAT WILL IT TAKE FOR US TO DO SAME 952 00:43:16,994 --> 00:43:18,830 IN NEUROSCIENCE? ARE WE GOING TO DO SAME TYPE OF OPERATIONS AS THE TECH GIANTS? 953 00:43:18,830 --> 00:43:29,273 WHEN THEY WORK WITH A.I.? 954 00:43:29,474 --> 00:43:32,977 AND DEV OPS FOR SOFTWARE AND 955 00:43:32,977 --> 00:43:35,947 DATA OPS AND ML OPS. THESE ARE 956 00:43:35,947 --> 00:43:39,517 TECHNICAL FIELDS THAT REQUIRE 957 00:43:39,517 --> 00:43:40,418 MANAGING INFRASTRUCTURE THAT RELY 958 00:43:40,418 --> 00:43:42,286 ON TECHNOLOGY, CONTAINERIZATION, 959 00:43:42,286 --> 00:43:46,757 VERSION CONTROL. FAST EVOLVING FOR 960 00:43:46,757 --> 00:43:49,927 SCALE UP. THE FACT THAT WE'RE NOW 961 00:43:49,927 --> 00:43:51,362 LIMITED BY ENERGY IN A.I. IS AN 962 00:43:51,362 --> 00:43:52,463 ACCOMPLISHMENT. WE'RE NOT LIMITED 963 00:43:52,463 --> 00:44:00,605 BY OUR ABILITY TO ORGANIZE, BUT WE 964 00:44:00,605 --> 00:44:02,707 ARE HITTING PHYSICAL LIMIT. IN 965 00:44:02,707 --> 00:44:03,541 NEUROSCIENCE WE'RE NOT THERE YET. 966 00:44:03,541 --> 00:44:09,380 WE'RE NOT LIMITED BY OUR 967 00:44:09,380 --> 00:44:14,418 ELECTRICITY BILL AS MUCH AS 968 00:44:14,418 --> 00:44:14,819 A.I. COMPANIES. 969 00:44:14,819 --> 00:44:16,053 SO CAN WE JUST TAKE WHAT A.I. 970 00:44:16,053 --> 00:44:21,092 COMPANIES HAVE DONE AND APPLY IT 971 00:44:21,092 --> 00:44:21,993 AND START USING IT IN OUR RESEARCH? 972 00:44:21,993 --> 00:44:24,562 IT MAY OR MAY NOT BE POSSIBLE. 973 00:44:24,562 --> 00:44:28,132 MAYBE THROUGH SOME COMBINATION 974 00:44:28,132 --> 00:44:28,666 OF COLLABORATING BETWEEN 975 00:44:28,666 --> 00:44:29,734 INDUSTRY AND SCIENCE. 976 00:44:29,734 --> 00:44:31,702 SO WE ASKED THESE QUESTIONS 977 00:44:31,702 --> 00:44:34,405 EXACTLY ABOUT TWO YEARS AGO. 978 00:44:34,405 --> 00:44:40,545 WE HAD A GRANT TO DEVELOP 979 00:44:40,545 --> 00:44:43,681 PLATFORMS FOR NEUROSCIENCE 980 00:44:43,681 --> 00:44:44,916 OPERATIONS AND THE PAPER GOT 981 00:44:44,916 --> 00:44:48,452 ACCEPTED FOR PEER REVIEW 982 00:44:48,452 --> 00:44:54,859 WE STARTED THINKING ABOUT HOW TO 983 00:44:54,859 --> 00:45:01,465 DEFINE SCIOPS LIKE ML OPS AND 984 00:45:01,465 --> 00:45:06,904 DATA OPS HAVE BEEN FORMALIZED IN 985 00:45:06,904 --> 00:45:09,473 THE A.I. INDUSTRY. WE SURMISE IT 986 00:45:09,473 --> 00:45:20,017 WILL PROGRESS TOWARD OPERATIONAL 987 00:45:22,153 --> 00:45:24,522 MATURITY AND ADOPT STANDARDS, OPEN 988 00:45:24,522 --> 00:45:28,659 SOURCE, AND FOLLOW FAIR WORKFLOWS 989 00:45:28,659 --> 00:45:29,093 THEY BECOME ABLE TO HAVE BIGGER 990 00:45:29,093 --> 00:45:31,862 COLLABORATIONS. 991 00:45:31,862 --> 00:45:33,764 AND THEN TO DO THE TYPES OF 992 00:45:33,764 --> 00:45:35,766 SCIENCE WE'RE DESCRIBING HERE 993 00:45:35,766 --> 00:45:40,504 REQUIRES ENTERING INTO THE NEW 994 00:45:40,504 --> 00:45:45,009 PHASES OF SCI OPS AT ADOPTING 995 00:45:45,009 --> 00:45:48,245 TECHNOLOGIES FOR MANAGING DATA 996 00:45:48,245 --> 00:45:53,451 PIPELINES AND AUTOMATING TASKS 997 00:45:53,451 --> 00:45:54,685 AND SCALING UP. AND THEN IMPLEMENT 998 00:45:54,685 --> 00:46:03,027 A.I. IN THE DISCOVERY LOOP. 999 00:46:03,027 --> 00:46:04,762 WHAT DOES IT MEAN IN PRACTICE? 1000 00:46:04,762 --> 00:46:07,431 ONE OF THE CONCLUSIONS WE MADE IS 1001 00:46:07,431 --> 00:46:10,334 THAT DATA, CODE, AND PROCESSES 1002 00:46:10,334 --> 00:46:14,105 HAVE TO BE MANAGED IN A UNIFIED 1003 00:46:14,105 --> 00:46:14,338 UNIFIED FRAME. IT'S REALLY HARD 1004 00:46:14,338 --> 00:46:19,877 TO DO THIS SEPARATELY. AND NEED 1005 00:46:19,877 --> 00:46:23,681 INFRASTRUCTURE THAT SEPARATES 1006 00:46:23,681 --> 00:46:29,253 THE SCIENTIFIC PROCESS FROM THE 1007 00:46:29,253 --> 00:46:29,854 INFRASTRUCTURE IMPLEMENTATION 1008 00:46:29,854 --> 00:46:36,093 SO THAT IT SCALES, CAN TRANSFER. 1009 00:46:36,093 --> 00:46:39,997 DATAJOINT FRAMEWORK BEGAN IN 1010 00:46:39,997 --> 00:46:44,635 ANDREAS LAB. I WORKED THERE ON THIS 1011 00:46:44,635 --> 00:46:45,836 PROBLEM HOW TO EXPRESS 1012 00:46:45,836 --> 00:46:49,473 SCIENTIFIC STUDIES AS A 1013 00:46:49,473 --> 00:46:52,877 COMPUTATIONAL SCHEMA AND DEPLOY 1014 00:46:52,877 --> 00:46:53,678 DIVERSE INFRASTRUCTURE. WITH 1015 00:46:53,678 --> 00:46:55,479 FUNDING FROM THE BRAIN 1016 00:46:55,479 --> 00:46:56,781 INITIATIVE AND COLLABORATING 1017 00:46:56,781 --> 00:46:58,149 WITH DOZENS OF NEUROSCIENCE 1018 00:46:58,149 --> 00:46:59,684 PROJECTS, WE EXPANDED TO A 1019 00:46:59,684 --> 00:47:01,852 PLATFORM THAT PROVIDES A COMMON 1020 00:47:01,852 --> 00:47:04,588 LANGUAGE FOR LABS TO EXPRESS 1021 00:47:04,588 --> 00:47:09,460 THEIR DATA ANALYSIS PIPELINES 1022 00:47:09,460 --> 00:47:13,931 FROM THE INSTRUMENT TO THE PAPER 1023 00:47:13,931 --> 00:47:17,535 AND SUBMITTED TO ARCHIVE AND 1024 00:47:17,535 --> 00:47:18,102 RELYING ON COMMON LANGUAGE 1025 00:47:18,102 --> 00:47:27,678 DESIGNS. OTHER PLATFORMS DEVELOPED 1026 00:47:27,678 --> 00:47:28,612 SIMILAR APPROACHES DATA AND 1027 00:47:28,612 --> 00:47:33,684 COMPUTATION MANAGED ON TOP OF THIS. 1028 00:47:33,684 --> 00:47:37,054 THIS IS AN EXAMPLE FROM VIRTUAL 1029 00:47:37,054 --> 00:47:39,690 BRAIN RUNNING ON EBRAINS PLATFORM 1030 00:47:39,690 --> 00:47:49,467 BY PETRA RITTER THAT COMBINE THESE 1031 00:47:49,467 --> 00:47:55,740 PRINCIPLES AND ALLOWS MANY COLLABORATORS TO MODEL BRAIN INPUTS.BRAINLIFE ALLOWS SIMILAR 1032 00:47:55,740 --> 00:47:59,510 APPROACH IN NEUROIMAGING. OTHER GATEWAYS AT UCSD ALLOWS ACCESS 1033 00:47:59,510 --> 00:48:04,648 TO HIGH PERFORMANCE COMPUTING TO 1034 00:48:04,648 --> 00:48:11,789 PERFORM ANALYSIS AT SCALE. 1035 00:48:11,789 --> 00:48:13,424 MOST ARE DONE IN THE LAB AND 1036 00:48:13,424 --> 00:48:21,065 WILL BE DONE IN THE LAB. 1037 00:48:21,065 --> 00:48:27,338 WE'RE WORKING MORE CLOSELY WITH 1038 00:48:27,338 --> 00:48:33,477 INSTITUTIONS TO PROVIDE A NEW 1039 00:48:33,477 --> 00:48:34,411 SERVICE. UNIVERSITIES PROVIDE I.T. 1040 00:48:34,411 --> 00:48:36,147 INFRASTRUCTURE BUT NOT 1041 00:48:36,147 --> 00:48:37,114 NECESSARILY THE DATA OPERATION 1042 00:48:37,114 --> 00:48:47,558 AND REQUIRES NEW LANGUAGE. A LOT 1043 00:48:47,558 --> 00:48:51,195 OF OPERATIONS ARE TIED TO THE GRANT 1044 00:48:51,195 --> 00:48:52,396 CYCLE AND IT REQUIRES A LITTLE 1045 00:48:52,396 --> 00:48:53,798 BIT OF LEADERSHIP WORKING TO 1046 00:48:53,798 --> 00:48:58,502 PROVIDE THIS AS A GENERAL THING. 1047 00:48:58,502 --> 00:49:02,573 WHAT DO WE SEE IN THE FUTURE? 1048 00:49:02,573 --> 00:49:05,476 IT'S HARD TO BUILD A.I. 1049 00:49:05,476 --> 00:49:07,578 COMBINATION ON TOP OF STICKY 1050 00:49:07,578 --> 00:49:11,482 NOTES AND SPREADSHEETS THAT 1051 00:49:11,482 --> 00:49:11,782 FLUCTUATES. 1052 00:49:11,782 --> 00:49:13,083 WE NEED TO IMPLEMENT THEM AT THE 1053 00:49:13,083 --> 00:49:15,119 INSTITUTIONAL LEVEL THAT ALLOWS 1054 00:49:15,119 --> 00:49:16,921 RESEARCHERS TO DEFINE THE LOGIC 1055 00:49:16,921 --> 00:49:21,926 AND STRUCTURE OF THEIR EXPERIMENTS 1056 00:49:21,926 --> 00:49:24,962 IN SUCH A WAY TO BE INTEGRATED 1057 00:49:24,962 --> 00:49:26,330 WITH COLLABORATIVE GROUPS WORKING 1058 00:49:26,330 --> 00:49:28,566 ON A.I.. NEEDS TO BE REPRODUCIBLE, 1059 00:49:28,566 --> 00:49:34,338 RELIABLE AND MANAGED WELL IN A 1060 00:49:34,338 --> 00:49:37,775 SUSTAINABLE WAY. THIS WILL LEAD TO 1061 00:49:37,775 --> 00:49:42,713 THE RECOGNITION OF TECHNOLOGY AND 1062 00:49:42,713 --> 00:49:53,224 ATTRACTING FUNDING AND TALENT. 1063 00:49:57,828 --> 00:50:08,272 >> IT'S A DANGEROUS OPERATION. 1064 00:50:09,540 --> 00:50:10,007 >> THANK YOU DIMITRI. 1065 00:50:10,007 --> 00:50:11,876 WITH THE FEW MINUTES REMAINING I 1066 00:50:11,876 --> 00:50:13,744 WANT TO GIVE A HIGH-LEVEL 1067 00:50:13,744 --> 00:50:16,747 OVERVIEW OF THE WORKSHOP AND 1068 00:50:16,747 --> 00:50:18,382 PARTICIPANT GUIDANCE WHAT THE 1069 00:50:18,382 --> 00:50:23,654 PANEL SESSION DISCUSSIONS LOOK 1070 00:50:23,654 --> 00:50:23,854 LIKE. 1071 00:50:23,854 --> 00:50:34,398 AND SHOW WHERE THE EXITS ARE SO 1072 00:50:41,772 --> 00:50:42,606 HOPEFULLY MANY HAVE SEEN THE 1073 00:50:42,606 --> 00:50:42,973 AGENDA. 1074 00:50:42,973 --> 00:50:44,275 IF YOU HAVEN'T SEEN THE AGENDA 1075 00:50:44,275 --> 00:50:51,448 IT'S ON THE WEB AT THIS QR CODE. 1076 00:50:51,448 --> 00:50:54,551 TODAY WE ARE THE OPENING REMARKS 1077 00:50:54,551 --> 00:50:55,753 AND ABOUT TO HIT INTO COFFEE 1078 00:50:55,753 --> 00:50:55,953 BREAK. 1079 00:50:55,953 --> 00:50:57,855 WE HAVE TWO SESSIONS. 1080 00:50:57,855 --> 00:51:00,557 MORNING SESSION, SESSION 1 IS 1081 00:51:00,557 --> 00:51:04,194 CONTINUING THIS DISCUSSION ABOUT 1082 00:51:04,194 --> 00:51:05,729 DEFINING NEUROAI FOR BRAIN AND 1083 00:51:05,729 --> 00:51:07,831 WHAT IS THE SCOPE THAT MATTERS 1084 00:51:07,831 --> 00:51:13,470 FOR THE FUTURE VISION AND NEARTERM 1085 00:51:13,470 --> 00:51:20,511 GAPS. WHAT ARE THE OPPORTUNITIES? 1086 00:51:20,511 --> 00:51:22,079 HELP US IDENTIFY PRIORITIES. 1087 00:51:22,079 --> 00:51:26,317 WE'LL HAVE LUNCH AND THEN END OF 1088 00:51:26,317 --> 00:51:32,456 LUNCH WE'LL HAVE A FUNDERS PANEL 1089 00:51:32,456 --> 00:51:33,324 REPRESENTATIVES FROM FUNDING 1090 00:51:33,324 --> 00:51:40,464 AGENCIES AND INVESTORS AND THEY 1091 00:51:40,464 --> 00:51:43,267 WILL HAVE A DISCUSSION MODERATED. 1092 00:51:43,267 --> 00:51:45,235 BY TERRY SEJNOWSKI THAT WILL HELP 1093 00:51:45,235 --> 00:51:47,037 LAY OUT THE FUNDING LANDSCAPE AND 1094 00:51:47,037 --> 00:51:51,141 FOLLOWING THAT WE'LL HAVE SESSION 1095 00:51:51,141 --> 00:51:54,311 TWO DIGGING INTO NEUROAI METHODS 1096 00:51:54,311 --> 00:51:57,881 METHODS OF MAPPING ARTIFICIAL 1097 00:51:57,881 --> 00:51:59,683 NEURAL NETWORK MODELS ONTO BRAIN 1098 00:51:59,683 --> 00:52:02,653 STRUCTURE, FUNCTION, AND ANATOMY. 1099 00:52:02,653 --> 00:52:05,422 WHAT CAN WE LEARN FROM METRICS AND 1100 00:52:05,422 --> 00:52:09,960 THEN WE'LL CLOSE OUT THE DAY 1101 00:52:09,960 --> 00:52:11,595 WITH A MODERATED DISCUSSION 1102 00:52:11,595 --> 00:52:12,429 BRINGING ALL THESE THINGS TOGETHER 1103 00:52:12,429 --> 00:52:13,197 FOR THE FUTURE. 1104 00:52:13,197 --> 00:52:18,702 SO DAY TWO FOLLOWS A SIMILAR 1105 00:52:18,702 --> 00:52:18,969 STRUCTURE. 1106 00:52:18,969 --> 00:52:20,938 WE START WITH EARLY CAREER 1107 00:52:20,938 --> 00:52:28,779 SESSION POSTER BLITZ STARTING AT 1108 00:52:28,779 --> 00:52:30,848 8:45 TOMORROW MORNING. SESSION 3 1109 00:52:30,848 --> 00:52:34,885 WILL INCLUDE ASPECTS OF EMBODIMENT, 1110 00:52:34,885 --> 00:52:36,387 PHYSICAL INTELLIGENCE, BIOMECHANICAL 1111 00:52:36,387 --> 00:52:40,744 FEEDBACK, THINKING ABOUT ALL THE PROBLEMS BIOLOGY 1112 00:52:40,744 --> 00:52:45,362 HAVE SOLVED THAT WE DON'T SEE THAT IN A.I. SYSTEMS c 1113 00:52:45,362 --> 00:52:50,100 AND HOW TO INCORPORATE THOSE INTO FUTURE VISIONS OF NEUROAI 1114 00:52:50,100 --> 00:52:53,570 SESSION 4 IS ABOUT TRANSITIONS IDEAS 1115 00:52:53,570 --> 00:52:56,703 INTO THEORIES AND TECHNOLOGIES THAT CAN MAKE 1116 00:52:56,703 --> 00:52:59,679 REAL IMPACT INTO HOW WEDO SCIENCE AND HOW 1117 00:52:59,679 --> 00:53:00,477 WE IMPROVE 1118 00:53:00,477 --> 00:53:02,012 BRAIN HEALTH. 1119 00:53:25,135 --> 00:53:26,370 WE ARE EXCITED TO HAVE EVERYONE HERE 1120 00:53:26,370 --> 00:53:26,937 FOR THIS AGENDA. 1121 00:53:26,937 --> 00:53:31,075 AS WE GO THROUGH THE SESSIONS, 1122 00:53:31,075 --> 00:53:34,411 WE'LL PUT UP A SLIDE THAT LOOKS 1123 00:53:34,411 --> 00:53:37,915 LIKE THIS. 1124 00:53:37,915 --> 00:53:39,783 THIS TELLS US ROUGHLY WHAT THE 1125 00:53:39,783 --> 00:53:41,685 GOALS ARE AND THERE WILL BE 1126 00:53:41,685 --> 00:53:43,020 QUESTIONS DURING THE DISCUSSION 1127 00:53:43,020 --> 00:53:44,521 PERIOD AND THINK ABOUT THE 1128 00:53:44,521 --> 00:53:46,423 QUESTIONS. THERE WILL BE A QR CODE 1129 00:53:46,423 --> 00:53:49,026 ON THE LEFT. IF YOU SCAN THAT DURING 1130 00:53:49,026 --> 00:53:51,428 THE SESSION IT WILL SEND AN 1131 00:53:51,428 --> 00:53:53,063 E-MAIL AND IF YOU'RE A VIRTUAL 1132 00:53:53,063 --> 00:53:55,566 AUDIENCE WE'LL READ YOUR QUESTIONS. 1133 00:53:55,566 --> 00:53:58,335 THERE ARE THREE MAIN PARTS TO EACH 1134 00:53:58,335 --> 00:54:01,205 SESSION. SHORT TALKS BY SPEAKERS, 1135 00:54:01,205 --> 00:54:07,478 QUESTIONS FROM INVITED DISCUSSANTS, 1136 00:54:07,478 --> 00:54:13,317 AND WE WILL OPEN IT UP TO AUDIENCE 1137 00:54:13,317 --> 00:54:19,957 Q&A AND IF YOU HAVE YOU 1138 00:54:19,957 --> 00:54:21,959 A CLEAR CONCRETE QUESTION READY 1139 00:54:21,959 --> 00:54:23,627 TO GO FOR A SPECIFIC SPEAKER 1140 00:54:23,627 --> 00:54:25,796 DURING THE TRANSITION, STEP UP 1141 00:54:25,796 --> 00:54:28,031 TO THE MIC OR SEND IN A 1142 00:54:28,031 --> 00:54:30,167 QUESTION AND WE'LL ASK THAT FOR 1143 00:54:30,167 --> 00:54:33,170 THE SPEAKERS OTHERWISE PLEASE 1144 00:54:33,170 --> 00:54:38,609 HOLD YOUR OPEN ENDED QUESTIONS 1145 00:54:38,609 --> 00:54:45,516 FOR AUDIENCE Q&A. 1146 00:54:45,516 --> 00:54:47,384 IF YOU'RE A PANELIST ON ZOOM, REMAIN 1147 00:54:47,384 --> 00:54:48,118 MUTED UNLESS YOU'RE PROMOTED TO 1148 00:54:48,118 --> 00:54:54,091 SPEAK. IF YOU ARE ON VIDEOCAST 1149 00:54:54,091 --> 00:54:59,830 AND HAVE A QUESTION HIT THE 1150 00:54:59,830 --> 00:55:02,499 Q&A BUTTON ON THE BOTTOM OF YOUR 1151 00:55:02,499 --> 00:55:09,706 SCREEN OR SEND USE THE SEND LIVE 1152 00:55:09,706 --> 00:55:19,082 FEEDBACK FORM. SO YOU KNOW WHERE 1153 00:55:19,082 --> 00:55:20,617 ARE. THERE ARE STAIRS GOING UP TO 1154 00:55:20,617 --> 00:55:26,056 THE ATRIUM WHERE THE POSTER SESSION 1155 00:55:26,056 --> 00:55:28,725 WILL BE TOMORROW. WITH THAT I WANT 1156 00:55:28,725 --> 00:55:32,829 TO THANK A SPONSORS SUPPORTING THE 1157 00:55:32,829 --> 00:55:38,769 NEUROAI WORKSHOP: THE KAVLI 1158 00:55:38,769 --> 00:55:48,212 FOUNDATION, JOHNS HOPKINS' KAVLI 1159 00:55:48,212 --> 00:55:52,783 NEUROSCIENCE DISCOVERY INSTITUTE, 1160 00:55:52,783 --> 00:55:55,652 WU TSAI INSTITUTE AT YALE AND DARPA. 1161 00:55:57,921 --> 00:56:03,994 WE'RE BEGINNING SESSION 1. 1162 00:56:03,994 --> 00:56:14,871 OUR FIRST SPEAKER IS ALI MINAI. 1163 00:56:21,078 --> 00:56:23,213 WE WILL HAVE A SEQUENCE OF SHORT 1164 00:56:23,213 --> 00:56:33,624 TALKS FOLLOWED BY A DISCUSSION 1165 00:56:58,248 --> 00:56:58,915 ALL RIGHT. 1166 00:56:58,915 --> 00:57:01,184 SO THANK YOU ALL FOR BEING HERE. 1167 00:57:01,184 --> 00:57:02,419 IT'S A GREAT PLEASURE TO BE AT 1168 00:57:02,419 --> 00:57:07,924 THIS WONDERFUL WORKSHOP. MY TOPIC 1169 00:57:07,924 --> 00:57:11,795 IS DEEP INTELLIGENCE.WHY A.I. MUST 1170 00:57:11,795 --> 00:57:13,230 LEARN FROM NATURE'S IMAGINATION. 1171 00:57:13,230 --> 00:57:17,334 THAT TERM COMES FROM THIS 1172 00:57:17,334 --> 00:57:19,836 WONDERFUL QUOTE FROM FEYNMAN THE 1173 00:57:19,836 --> 00:57:21,338 IMAGINATION OF NATURE IS FAR FAR 1174 00:57:21,338 --> 00:57:21,972 GREATER THAN THE IMAGINATION OF 1175 00:57:21,972 --> 00:57:23,407 MAN. 1176 00:57:23,407 --> 00:57:26,310 SO THE THING IS WHY SHOULD WE 1177 00:57:26,310 --> 00:57:34,117 NOT LEARN FROM THAT IMAGINATION? 1178 00:57:34,117 --> 00:57:35,619 HERE IS THE NEUROAI LOOP TONY 1179 00:57:35,619 --> 00:57:37,421 SHOWED AND OTHERS TOO. 1180 00:57:37,421 --> 00:57:47,964 I AM VERY MUCH GOING TO FOCUS ON 1181 00:57:48,865 --> 00:57:51,335 THE UPPER BRANCH ON THIS LOOP AND 1182 00:57:51,335 --> 00:57:52,736 ASKING MY A.I. COLLEAGUES TO PAY 1183 00:57:52,736 --> 00:57:53,937 MORE ATTENTION TO NEUROSCIENCE. 1184 00:57:53,937 --> 00:57:55,072 THERE'LL BE A LOT OF TALKS WHICH 1185 00:57:55,072 --> 00:57:55,972 WILL DEAL WITH THE OTHER 1186 00:57:55,972 --> 00:58:01,845 DIRECTION I'M SURE. 1187 00:58:01,845 --> 00:58:07,918 THE GOAL IS TO BUILD A MORE 1188 00:58:07,918 --> 00:58:09,586 NATURAL A.I. BY TAKING ADVANTAGE 1189 00:58:09,586 --> 00:58:11,655 OF WHAT WE HAVE LEARNED FROM 1190 00:58:11,655 --> 00:58:14,324 NEUROSCIENCE. 1191 00:58:14,324 --> 00:58:16,893 THE QUESTION IS DO WE WANT TO 1192 00:58:16,893 --> 00:58:23,500 BUILD A MORE NATURAL A.I.? 1193 00:58:23,500 --> 00:58:27,137 THAT IS NOT A VERY HOT IDEA IN ML 1194 00:58:27,137 --> 00:58:33,310 WHERE THE FOCUS IS ON BUILDING 1195 00:58:33,310 --> 00:58:35,512 HUMAN-LIKE SYSTEMS TO SOLVE 1196 00:58:35,512 --> 00:58:36,747 HUMAN-LIKE PROBLEMS. 1197 00:58:36,747 --> 00:58:44,321 A.I. AS TOOLS FOR HUMAN USE. 1198 00:58:44,321 --> 00:58:46,590 SOME OF US ARE HERE TO BUILD 1199 00:58:46,590 --> 00:58:51,695 REAL INTELLIGENCE WHICH IS MORE 1200 00:58:51,695 --> 00:58:53,597 LIKE THE INTELLIGENCE IN THE 1201 00:58:53,597 --> 00:58:55,432 ANIMAL WORLD. AND THE QUESTION IS 1202 00:58:55,432 --> 00:58:56,800 WHETHER WE'LL HAVE A STAR TREK 1203 00:58:56,800 --> 00:58:58,135 OR STAR WARS FUTURE. 1204 00:58:58,135 --> 00:58:59,936 I'M VERY MUCH TALKING ABOUT A 1205 00:58:59,936 --> 00:59:07,978 STAR WARS FUTURE WITH LOTS OF 1206 00:59:07,978 --> 00:59:10,080 INTELLIGENT THINGS GOING AROUND 1207 00:59:10,080 --> 00:59:12,949 AND MANY INTELLIGENT BEINGS. 1208 00:59:12,949 --> 00:59:15,886 HOW WOULD WE DEFINE NATURAL 1209 00:59:15,886 --> 00:59:16,219 INTELLIGENCE? 1210 00:59:16,219 --> 00:59:17,921 IT'S INTELLIGENCE ANIMALS HAVE. 1211 00:59:17,921 --> 00:59:20,957 WHAT ARE CHARACTERISTICS? 1212 00:59:20,957 --> 00:59:23,226 HERE ARE SOME. 1213 00:59:23,226 --> 00:59:27,097 IT'S BIOLOGICAL AND EMBODIED. 1214 00:59:27,097 --> 00:59:29,366 IT'S EMERGENT DEPENDS ON EMERGENCE. 1215 00:59:29,366 --> 00:59:32,302 IT'S HETEROGENOUS AND REQUIRES 1216 00:59:32,302 --> 00:59:34,137 HETEROGENOUS STRUCTURES AND 1217 00:59:34,137 --> 00:59:37,140 PROCESSES. NOT JUST A BLOB. 1218 00:59:37,140 --> 00:59:42,245 IT HAS MULTISCALE ORGANIZATION 1219 00:59:42,245 --> 00:59:45,282 IT HAS MULTISCALE AND ADAPTATION 1220 00:59:45,282 --> 00:59:47,651 AND IT'S AUTONOMOUS AND ALWAYS 1221 00:59:47,651 --> 00:59:49,719 REAL TIME AND IT'S ENVIRONMENT 1222 00:59:49,719 --> 00:59:51,221 SPECIFIC SO IT IS INTELLIGENT IN 1223 00:59:51,221 --> 00:59:57,994 ITS ENVIRONMENT. 1224 00:59:57,994 --> 01:00:00,797 IT'S INHERENTLY INTEGRATED UNLIKE 1225 01:00:00,797 --> 01:00:02,299 THE MACHINE LEARNING A.I. WE BUILD 1226 01:00:02,299 --> 01:00:03,834 NOW FOCUSSING ON INDIVIDUAL 1227 01:00:03,834 --> 01:00:09,206 TASKS AND HOPEFULLY WE'LL BE 1228 01:00:09,206 --> 01:00:11,641 ABLE OR HOPE TO PUT THEM ALL 1229 01:00:11,641 --> 01:00:14,444 TOGETHER. REAL INTELLIGENCE IS 1230 01:00:14,444 --> 01:00:15,679 ALWAYS INTEGRATED. THE INTELLIGENCE 1231 01:00:15,679 --> 01:00:17,614 OF A MOUSE OR SHRIMP IS JUST AS 1232 01:00:17,614 --> 01:00:20,750 INTEGRATED AS THE INTELLIGENCE OF 1233 01:00:20,750 --> 01:00:23,520 A HUMAN AND FINALLY IT'S EVOLVABLE. 1234 01:00:23,520 --> 01:00:28,625 THE COMPUTATIONAL A.I. IS VERY 1235 01:00:28,625 --> 01:00:31,428 DIFFERENT AND SELDOM EMBODIES 1236 01:00:31,428 --> 01:00:37,267 ANY OF THESE CHARACTERISTICS. 1237 01:00:37,267 --> 01:00:42,272 IF WE COMPARE THE TWO OF THESE 1238 01:00:42,272 --> 01:00:44,541 ARTIFICIAL AND NATURAL 1239 01:00:44,541 --> 01:00:45,208 INTELLIGENCE, OFTEN OUR MACHINE 1240 01:00:45,208 --> 01:00:48,144 LEARNING MODELS BEGIN WITH A 1241 01:00:48,144 --> 01:00:53,717 GENERAL UNDIFFERENTIATED REGULAR 1242 01:00:53,717 --> 01:01:04,261 STRUCTURE. THEN WE HOPE AND USES 1243 01:01:05,695 --> 01:01:11,134 LOTS OF DATA AND COMPUTATION AND 1244 01:01:11,134 --> 01:01:12,903 SOMEHOW THE ORGANIZATION NEEDED 1245 01:01:12,903 --> 01:01:13,803 TO ACCOMPLISH THE TASK WE'RE 1246 01:01:13,803 --> 01:01:20,777 TRAINING IT ON WILL APPEAR. 1247 01:01:20,777 --> 01:01:23,113 WE HOPE THE RESULTING SYSTEM 1248 01:01:23,113 --> 01:01:25,382 WILL BE USEFUL NOT JUST FOR THE 1249 01:01:25,382 --> 01:01:27,551 TASK YOU TRAIN IT ON BUT USEFUL 1250 01:01:27,551 --> 01:01:31,388 FOR LOTS OF OTHER THINGS LIKE 1251 01:01:31,388 --> 01:01:32,389 GENERATING TEXT AND REASONING 1252 01:01:32,389 --> 01:01:33,356 AND PLANNING AND ALL THE THINGS 1253 01:01:33,356 --> 01:01:39,663 WE TRY TO USE THE SYSTEMS FOR. 1254 01:01:39,663 --> 01:01:43,500 THE PROBLEM IS ONE, THE MODELS 1255 01:01:43,500 --> 01:01:46,036 ARE TASK SPECIFIC. 1256 01:01:46,036 --> 01:01:49,639 TWO, WE'RE GOING FROM A SYSTEM 1257 01:01:49,639 --> 01:01:52,108 THAT'S NAIVE TO A SYSTEM WITH 1258 01:01:52,108 --> 01:01:53,777 FUNCTIONALITY, DATA INTENSIVE AND 1259 01:01:53,777 --> 01:01:54,277 DIFFICULT IN GENERAL. 1260 01:01:54,277 --> 01:01:56,313 WE DON'T HAVE ANY CONTROL IN 1261 01:01:56,313 --> 01:01:57,747 THAT PROCESS OVER WHAT KIND OF 1262 01:01:57,747 --> 01:01:59,449 SYSTEM WILL EMERGE. 1263 01:01:59,449 --> 01:02:01,418 SO IN PRINCIPLE A WEIRD MODEL 1264 01:02:01,418 --> 01:02:02,085 COULD EMERGE. 1265 01:02:02,085 --> 01:02:08,625 IN SOME CASES, FOR EXAMPLE, 1266 01:02:08,625 --> 01:02:10,026 IN VISION THERE'S BEEN WORK WHICH 1267 01:02:10,026 --> 01:02:17,033 SHOWS THERE IS A CORRESPONDENCE 1268 01:02:17,033 --> 01:02:23,239 BETWEEN THE MODELS THAT EMERGE AND 1269 01:02:23,239 --> 01:02:25,642 THE MODELS BRAINS MAY BE USING. 1270 01:02:25,642 --> 01:02:26,610 DOES THIS HAPPEN IN LARGE LANGUAGE 1271 01:02:26,610 --> 01:02:28,511 MODELS? BECAUSE WE DON'T HAVE 1272 01:02:28,511 --> 01:02:31,481 CONTROL OVER THE MODEL, EVEN IF IT 1273 01:02:31,481 --> 01:02:32,749 WORKS WELL ON EVERYTHING WE TEST, 1274 01:02:32,749 --> 01:02:34,884 WE DON'T KNOW WHAT KIND OF HIDDEN 1275 01:02:34,884 --> 01:02:36,019 FAILURE MODES IT HAS AND WHEN 1276 01:02:36,019 --> 01:02:41,992 THEY MIGHT SHOW UP. 1277 01:02:41,992 --> 01:02:45,528 SO THIS SORT OF THING -- THE 1278 01:02:45,528 --> 01:02:47,297 QUESTION EXISTS WITHIN THE 1279 01:02:47,297 --> 01:02:48,031 MACHINE LEARNING COMMUNITY AND 1280 01:02:48,031 --> 01:02:50,433 THERE'S VARIOUS WAYS TO TRY TO 1281 01:02:50,433 --> 01:02:55,905 ADDRESS THAT WHICH I WILL PUT UP. 1282 01:02:55,905 --> 01:03:06,449 ML TODAY USES SHALLOW ADAPTATION. 1283 01:03:10,220 --> 01:03:14,557 USE DATA AND TRAIN THE SYSTEM ON 1284 01:03:14,557 --> 01:03:16,493 LOTS OF DATA AND HAVE A SYSTEM 1285 01:03:16,493 --> 01:03:17,894 THAT PERFORMS WELL ON ALL THE 1286 01:03:17,894 --> 01:03:22,098 TASKS WE PUT IT THROUGH. 1287 01:03:22,098 --> 01:03:26,803 IT'S UNLIKE HOW ANIMALS LEARN 1288 01:03:26,803 --> 01:03:30,040 WHICH IS THIS PROCESS WHICH I'M 1289 01:03:30,040 --> 01:03:33,243 CALLING DEEP ADAPTATION. 1290 01:03:33,243 --> 01:03:36,212 DEEP IN EVOLUTIONARY HISTORY AND 1291 01:03:36,212 --> 01:03:39,516 SOME SIMPLE ANIMALS EMERGED AND 1292 01:03:39,516 --> 01:03:42,686 THEY DEVELOPED OR EVOLVED INTO 1293 01:03:42,686 --> 01:03:44,421 MORE COMPLEX ANIMALS AND WHEN WE 1294 01:03:44,421 --> 01:03:50,193 GET TO A PARTICULAR ANIMAL WHETHER 1295 01:03:50,193 --> 01:04:00,737 A RAT CAT OR HUMAN, WE HAVE 1296 01:04:26,262 --> 01:04:28,064 WE HAVE 500 MILLION YEAR OF 1297 01:04:28,064 --> 01:04:30,800 EVOLUTIONARY ENGINEERING THAT 1298 01:04:30,800 --> 01:04:31,034 EXISTS. 1299 01:04:31,034 --> 01:04:32,035 THAT'S WHY SO LITTLE LEARNING 1300 01:04:32,035 --> 01:04:34,137 IS OFTEN NEEDED FOR ANIMALS TO 1301 01:04:34,137 --> 01:04:35,905 LEARN TO ACCOMPLISH COMPLICATED 1302 01:04:35,905 --> 01:04:37,440 TASKS AND WHY SO LITTLE ENERGY 1303 01:04:37,440 --> 01:04:39,075 IS NEEDED AND ALL SORTS OF OTHER 1304 01:04:39,075 --> 01:04:39,309 THINGS. AND EVEN AFTER THAT, THE 1305 01:04:39,309 --> 01:04:44,547 LEARNING IS NOT JUST DOING 1306 01:04:44,547 --> 01:04:55,291 GRADIENT DESCENT PROCESS OR 1307 01:04:59,596 --> 01:05:00,597 SUPERVISED LEARNING. IT IS GOING 1308 01:05:00,597 --> 01:05:01,998 THRU A DEVELOPMENT PROCESS. THE BRAIN AND BODY BECOMES COMPLEX TOGETHER AND LEARN IN STAGES 1309 01:05:01,998 --> 01:05:06,402 BUILDING UPON PREVIOUS STAGES. AND THEN OF COURSE DEVELOPMENT 1310 01:05:06,402 --> 01:05:08,772 AND NEURAL LEARNING GO HAND IN 1311 01:05:08,772 --> 01:05:21,050 HAND BASE ON EXPERIENCE OF WORLD. 1312 01:05:21,050 --> 01:05:31,027 WE SHOULD NOT JUST LEARN FROM 1313 01:05:31,027 --> 01:05:31,728 NEUROSCIENCE BUT DEVELOPMENTAL 1314 01:05:31,728 --> 01:05:31,928 AND EVOLUTIONARY BIOLOGY AS WELL. 1315 01:05:31,928 --> 01:05:34,931 WHAT SHOULD WE LEARN FROM NATURE'S 1316 01:05:34,931 --> 01:05:35,632 IMAGINATION? FROM NEUROSCIENCE? 1317 01:05:35,632 --> 01:05:41,704 ARCHITECTURES, MODULES, LEARNING 1318 01:05:41,704 --> 01:05:46,976 RULES. THINGS WE ALREADY LEARNED 1319 01:05:46,976 --> 01:05:50,213 FROM NEUROSCIENCE BUT ALSO FOCUS 1320 01:05:50,213 --> 01:05:51,915 ON HOW EVOLUTION BUILDS COMPLEXITY 1321 01:05:51,915 --> 01:05:53,850 IN ANIMALS AND BRAINS AND BODIES 1322 01:05:53,850 --> 01:05:57,754 HOW ONTOGENESIS DEVELOPS FROM A 1323 01:05:57,754 --> 01:05:59,756 SIMPLE CELL TO A COMPLEX ORGANISM. 1324 01:05:59,756 --> 01:06:01,257 DEVELOPMENTAL LEARNING, 1325 01:06:01,257 --> 01:06:02,926 UNSUPERVISED AND REINFORCEMENT 1326 01:06:02,926 --> 01:06:07,530 LEARNING THEN AND AT END OF THAT 1327 01:06:07,530 --> 01:06:08,832 SUPERVISED LEARNING ON A SYSTEM 1328 01:06:08,832 --> 01:06:10,300 READY FOR SUPERVISED LEARNING. 1329 01:06:10,300 --> 01:06:15,405 RIGHT NOW WE'RE TRYING TO GET 1330 01:06:15,405 --> 01:06:16,206 ALMOST EVERYTHING FROM THE LAST 1331 01:06:16,206 --> 01:06:17,006 KIND OF LEARNING. 1332 01:06:17,006 --> 01:06:19,542 AND WE SKIP ALL THE OTHER 1333 01:06:19,542 --> 01:06:27,183 THINGS. 1334 01:06:27,183 --> 01:06:37,694 SO INSTEAD OF A NEUROAI TO HAVE 1335 01:06:40,797 --> 01:06:44,000 A EVO DEVO NEUROAI. NOT SAYING WE 1336 01:06:44,000 --> 01:06:45,268 SHOULD DO COMPUTATIONAL EVOLUTION 1337 01:06:45,268 --> 01:06:47,103 BUT I'M SUGGESTING WE SHOULD 1338 01:06:47,103 --> 01:06:49,372 LEARN THE LESSONS FROM EVOLUTION 1339 01:06:49,372 --> 01:06:56,112 AND MINE WHAT WE KNOW FROM 1340 01:06:56,112 --> 01:06:56,379 EVOLUTION. SO WE HAVE 1341 01:06:56,379 --> 01:06:57,881 SIMPLE AND MORE COMPLEX SPECIES 1342 01:06:57,881 --> 01:06:59,282 AND WE SHOULD LOOK AT ALL OF 1343 01:06:59,282 --> 01:07:04,554 THOSE AND THEIR BODIES AND HOW 1344 01:07:04,554 --> 01:07:06,122 THEY WORK AND INCORPORATE INTO 1345 01:07:06,122 --> 01:07:12,729 OUR A.I. MODELS. TO DEVELOP A CANONICAL SET OF STRUCTURAL MODULES, FUNCTIONAL PRIMITIVES AND 1346 01:07:12,729 --> 01:07:14,030 COMPLEXIFICATION RULES THAT WOULD 1347 01:07:14,030 --> 01:07:15,565 ENABLE US TO BUILD MORE COMPLEX 1348 01:07:15,565 --> 01:07:18,835 A.I. SYSTEMS IN A MORE BIOLOGICAL 1349 01:07:18,835 --> 01:07:27,443 AND SYSTEMATIC WAY. 1350 01:07:32,548 --> 01:07:37,687 HERE'S COMPLEXIFICATION GOING ON 1351 01:07:37,687 --> 01:07:39,188 ONE EXAMPLE DEVELOPMENTAL LEARNING. 1352 01:07:39,188 --> 01:07:46,296 THE POINT IS WHEN WE DO ALIGNMENT 1353 01:07:46,296 --> 01:07:48,798 AND IF WE DID THAT IN A MORE 1354 01:07:48,798 --> 01:07:51,067 DEVELOPMENT WAY WE'D BE ABLE TO 1355 01:07:51,067 --> 01:07:53,836 DO IT BETTER AND TO HAVE IT 1356 01:07:53,836 --> 01:07:54,203 BE MORE NATURAL. 1357 01:07:54,203 --> 01:07:59,642 ALL RIGHT. 1358 01:07:59,642 --> 01:08:09,986 THANK YOU VERY MUCH. 1359 01:08:24,801 --> 01:08:28,671 >> I'M SAD TO HEAR I ONLY HAVE 1360 01:08:28,671 --> 01:08:30,006 59 MINUTES AND 12 SECONDS TO 1361 01:08:30,006 --> 01:08:30,206 SPEAK. 1362 01:08:30,206 --> 01:08:33,776 I LOOK FORWARD TO DISCUSSING THE 1363 01:08:33,776 --> 01:08:39,382 POTENTIAL ROLE OF ASTROCYTES IN 1364 01:08:39,382 --> 01:08:41,985 BIOCOMPUTING. I'VE BEEN PART OF 1365 01:08:41,985 --> 01:08:45,088 THE BRAIN INITIATIVE CENTER TO 1366 01:08:45,088 --> 01:08:51,227 STUDY CROWD CODING IN THE BRAIN 1367 01:08:51,227 --> 01:09:01,671 IN THE DATA SCIENCE CORE. 1368 01:09:02,638 --> 01:09:04,440 AND LOOKING AT NATURE BRAINS CAN 1369 01:09:04,440 --> 01:09:07,377 WE LOOK AT ASTROCYTES AND EXPLORE 1370 01:09:07,377 --> 01:09:09,178 THEIR ROLES IN MAKING SENSE OF 1371 01:09:09,178 --> 01:09:10,613 THEIR SURROUNDING WORLD AND 1372 01:09:10,613 --> 01:09:16,586 FUNDAMENTALLY IN BIOCOMPUTING. 1373 01:09:16,586 --> 01:09:18,154 THE REASON I MIGHT BE WORTH 1374 01:09:18,154 --> 01:09:20,656 LOOKING AT OTHER MODES OF 1375 01:09:20,656 --> 01:09:22,025 COMPUTING IS THOUGH A.I. HAS 1376 01:09:22,025 --> 01:09:25,028 GOTTEN INSPIRATION FROM 1377 01:09:25,028 --> 01:09:27,497 NEUROSCIENCE, HUMANS STILL 1378 01:09:27,497 --> 01:09:29,966 OUTPERFORM MACHINES IN MINIMAL 1379 01:09:29,966 --> 01:09:31,801 SUPERVISION AND TRAINING, 1380 01:09:31,801 --> 01:09:32,602 LIMITED DATA. 1381 01:09:32,602 --> 01:09:35,371 IT WOULD BE VERY SAD IF HUMANS 1382 01:09:35,371 --> 01:09:36,205 COULD LEARN ALL THE KNOWLEDGE IN 1383 01:09:36,205 --> 01:09:39,642 THE WORLD AND THEN NOT DECIDE 1384 01:09:39,642 --> 01:09:45,481 WHAT TO DO WITH IT. A HAWK WITH 1385 01:09:45,481 --> 01:09:48,785 1 WATT OF POWER AND 15 GRAMS OF 1386 01:09:48,785 --> 01:09:50,453 BRAIN MASS CAN RECEIVE DATA FROM 1387 01:09:50,453 --> 01:09:55,525 THOUSAND OF SENSORS AND LEARN 1388 01:09:55,525 --> 01:10:00,096 RAPIDLY. THE QUESTION IS DO WE 1389 01:10:00,096 --> 01:10:08,638 NEED NEW ARCHITECTURES TO HARNESS 1390 01:10:08,638 --> 01:10:11,607 THE COMPLEXITY OF THE HUMAN BRAIN? 1391 01:10:11,607 --> 01:10:13,342 I WANT TO TAKE SIMPLER APPROACH 1392 01:10:13,342 --> 01:10:15,111 AND ASK THE QUESTION 1393 01:10:15,111 --> 01:10:15,945 WHAT ABOUT ASTROCYTES? 1394 01:10:15,945 --> 01:10:19,449 50% OF OUR BRAIN MASS OR SO ARE 1395 01:10:19,449 --> 01:10:19,916 ASTROCYTES. 1396 01:10:19,916 --> 01:10:21,117 WHAT ROLE COULD THEY PLAY? 1397 01:10:21,117 --> 01:10:22,485 THE OTHER QUESTION I'M GOING TO 1398 01:10:22,485 --> 01:10:25,822 ASK IS WHAT ABOUT NON-ELECTRICAL 1399 01:10:25,822 --> 01:10:28,791 SIGNALS IN THE BRAIN? 1400 01:10:28,791 --> 01:10:30,259 HOW DO THEY CONTRIBUTE TO NEUROAI? 1401 01:10:30,259 --> 01:10:30,493 I WANT TO THINK ABOUT NEUROAI AS A 1402 01:10:30,493 --> 01:10:41,037 BIOCOMPUTING SYSTEM. WE HAVE 1403 01:10:44,073 --> 01:10:46,008 NEURONS AND ASTROCYTES. THEY'RE 1404 01:10:46,008 --> 01:10:47,477 DIFFERENT FROM HOW A QUANTUM 1405 01:10:47,477 --> 01:10:52,782 COMPUTER IS EMERGING AS A NEW 1406 01:10:52,782 --> 01:10:55,218 WAY OF PROCESSING INFORMATION 1407 01:10:55,218 --> 01:10:58,054 WITH ATOMS, PHOTONS, ELECTRONS. 1408 01:10:58,054 --> 01:11:01,190 WHAT'S AMAZING ABOUT 1409 01:11:01,190 --> 01:11:04,093 BIOCOMPUTING WITH ONLY 139,000 1410 01:11:04,093 --> 01:11:09,432 NEURONS A FLY CAN ALREADY CARRY 1411 01:11:09,432 --> 01:11:14,003 OUT VERY COMPLEX TASKS, DO 1412 01:11:14,003 --> 01:11:16,873 SOMETHING INTELLIGENT. THAT'S 1413 01:11:16,873 --> 01:11:17,640 SIMILAR TO A QUANTUM COMPUTER 1414 01:11:17,640 --> 01:11:18,241 REACHING THE THRESHOLD FOR 1415 01:11:18,241 --> 01:11:19,942 INTELLIGENCE AND VERY DIFFERENT 1416 01:11:19,942 --> 01:11:25,515 FROM THE BILLIONS OF NODES THAT 1417 01:11:25,515 --> 01:11:31,654 HAVE BEEN NEEDED FOR CHATGPT 4. 1418 01:11:31,654 --> 01:11:36,759 THE ADVANTAGES OF THE DIFFERENT 1419 01:11:36,759 --> 01:11:41,330 MODALITIES MAY BE DISTINCT. 1420 01:11:41,330 --> 01:11:46,736 CHATGPT 4 IS AMAZING TO HAVE 1421 01:11:46,736 --> 01:11:48,771 DATA STORED BUT BIOCOMPUTERS ARE 1422 01:11:48,771 --> 01:11:50,573 POWERFUL MAKING DECISIONS FROM 1423 01:11:50,573 --> 01:11:54,076 LIMITED DYNAMIC DATA, LEARNING, 1424 01:11:54,076 --> 01:12:04,720 EXTRAPOLATING TO NEW SITUATIONS. 1425 01:12:04,954 --> 01:12:05,721 ARE THERE OTHER SIGNALS? 1426 01:12:05,721 --> 01:12:09,692 AND WE'VE BEEN LOOKING AT 1427 01:12:09,692 --> 01:12:13,529 CHEMICAL SIGNALS OF ASTROCYTE. 1428 01:12:13,529 --> 01:12:15,998 THEY'RE NOT JUST PASSIVE HOMEO- 1429 01:12:15,998 --> 01:12:19,368 STATIC PROVIDERS OF NUTRITION TO 1430 01:12:19,368 --> 01:12:25,541 THE BRAIN BUT ALSO YIELDING THEIR 1431 01:12:25,541 --> 01:12:29,512 OWN INTRINSIC RHYTHM AND IF YOU 1432 01:12:29,512 --> 01:12:32,782 LOOK AT THE CYTOSKELETON OF 1433 01:12:32,782 --> 01:12:37,920 INDIVIDUAL CELLS, ACTIN POLYMERIZATION, IT'S NOT JUST FORMING SYNAPSES ON DEMAND 1434 01:12:37,920 --> 01:12:41,390 BUT HAS IT'S OWN INTRINSIC 1435 01:12:41,390 --> 01:12:41,624 RHYTHM ON THE MINUTE SCALE. 1436 01:12:41,624 --> 01:12:45,461 WHAT EMERGES IS A PICTURE OF THE 1437 01:12:45,461 --> 01:12:46,762 BRAIN WHERE YOU HAVE DIFFERENT 1438 01:12:46,762 --> 01:12:49,699 SIGNALS, ELECTRICAL ON THE MS 1439 01:12:49,699 --> 01:12:52,268 SCALE, CHEMICAL SIGNALS FROM 1440 01:12:52,268 --> 01:12:58,541 ASTROCYTES ON 10-20 SECOND SCALE, 1441 01:12:58,541 --> 01:13:06,015 SLOWER MINUTE TIME SCALE RHYTHMS 1442 01:13:06,015 --> 01:13:06,449 OF THE CYTOSKELETON. DO SLOW 1443 01:13:06,449 --> 01:13:11,287 RHYTHMS CONTRIBUTE OR A SUBSTRATE 1444 01:13:11,287 --> 01:13:15,658 THAT ELECTRICAL SIGNALS OPERATE? 1445 01:13:15,658 --> 01:13:20,429 AND THERE'S BEEN MECHANICAL 1446 01:13:20,429 --> 01:13:23,633 SIGNALS OF SYNAPSE AND EVIDENCE OF 1447 01:13:23,633 --> 01:13:25,534 DIRECT COUPLING. AND WE ASKED THE 1448 01:13:25,534 --> 01:13:27,870 SIMPLE QUESTION, IF YOU HAVE 1449 01:13:27,870 --> 01:13:29,372 MECHANICAL TRANSMISSION THROUGH 1450 01:13:29,372 --> 01:13:31,173 SYNAPSES WHAT IF THE MECHANICAL 1451 01:13:31,173 --> 01:13:33,809 SYSTEM IS NOT A STEADY SYSTEM. 1452 01:13:33,809 --> 01:13:37,079 IT'S AN EXCITABLE SYSTEM IN ITS 1453 01:13:37,079 --> 01:13:39,081 OWN RIGHT. 1454 01:13:39,081 --> 01:13:42,118 RUNNING ON THE MINUTE TIME SCALE 1455 01:13:42,118 --> 01:13:43,519 WHAT WOULD IT DO? 1456 01:13:43,519 --> 01:13:49,325 WE LOOKED AT A SIMPLE ALGORITHM. 1457 01:13:49,325 --> 01:13:57,533 WHAT IF WE HAD MULTI MODAL 1458 01:13:57,533 --> 01:14:00,436 SIGNAL AND WHAT IF YOU VARIED IT 1459 01:14:00,436 --> 01:14:03,973 AND THE QUESTION IS IT SUCKS. 1460 01:14:03,973 --> 01:14:05,608 IT GETS WORSE BECAUSE YOU'RE 1461 01:14:05,608 --> 01:14:09,278 VARING IT AND YOUR FORCED TO 1462 01:14:09,278 --> 01:14:14,750 STORE IN THE SYSTEM AND MORE 1463 01:14:14,750 --> 01:14:17,119 ROBUST AS TIME GOES ON THEN ASK 1464 01:14:17,119 --> 01:14:19,822 YOURSELF HOW DO YOU LEARN IN A 1465 01:14:19,822 --> 01:14:21,824 SEESAWING RHYTHMIC SYSTEM. 1466 01:14:21,824 --> 01:14:27,630 WE DISCOVERED YOU CAN COME UP 1467 01:14:27,630 --> 01:14:32,868 WITH A NEW LEARNING ALGORITHM 1468 01:14:32,868 --> 01:14:37,139 BASED ON SYNCHRONIZING AND THE 1469 01:14:37,139 --> 01:14:38,641 SIMPLE ALGORITHM WE'RE ABLE TO 1470 01:14:38,641 --> 01:14:44,246 GENERATE DYNAMIC TWINS THAT 1471 01:14:44,246 --> 01:14:46,449 RAPIDLY EXTRAPOLATE DYNAMICS AND 1472 01:14:46,449 --> 01:14:48,217 PROMISING TO GO IN THE RIGHT 1473 01:14:48,217 --> 01:14:50,119 DIRECTION OF EXTRAPOLATION AND 1474 01:14:50,119 --> 01:14:50,553 RAPID LEARNING. THIS IS A SMALL 1475 01:14:50,553 --> 01:14:56,592 EXAMPLE OF LOOKING BEYOND THE 1476 01:14:56,592 --> 01:14:56,859 ELECTRICAL SIGNAL OF NEURONS 1477 01:14:56,859 --> 01:15:07,303 AND EXPLORING THE BIOLOGICAL 1478 01:15:10,005 --> 01:15:11,974 RHYTHMS AT MULTIPLE LENGTHS AND 1479 01:15:11,974 --> 01:15:12,308 TIME SCALE. NOW WITH THAT, WHAT 1480 01:15:12,308 --> 01:15:15,945 WE WOULD NEED IS TECHNOLOGY TO 1481 01:15:15,945 --> 01:15:21,751 SYSTEMICALLY LOOK AT THESE RHYTHMS 1482 01:15:21,751 --> 01:15:24,553 ON THE MILLISECOND TIME SCALE 1483 01:15:24,553 --> 01:15:26,455 AND IMAGING ON SECOND TO MINUTE 1484 01:15:26,455 --> 01:15:28,557 TIME SCALE AND SPATIAL SCALE. 1485 01:15:28,557 --> 01:15:31,026 AND TO ACCOMPLISH THIS I WANTED 1486 01:15:31,026 --> 01:15:32,128 TO HIGHLIGHT THE BRAIN 1487 01:15:32,128 --> 01:15:34,563 INITIATIVE HAS BEEN GREAT IN 1488 01:15:34,563 --> 01:15:36,499 SUPPORTING INTERDISCIPLINARY 1489 01:15:36,499 --> 01:15:38,300 WORK AT THE BOUNDARIES BETWEEN 1490 01:15:38,300 --> 01:15:40,069 BIOENGINEERING AND BIOLOGY AND 1491 01:15:40,069 --> 01:15:47,643 NEUROSCIENCE AND PHYSICS AND I 1492 01:15:47,643 --> 01:15:49,545 WANT TO ACKNOWLEDGE SUPPORT FROM 1493 01:15:49,545 --> 01:15:59,855 AFOSR AND ARL AND THANK YOU. 1494 01:16:30,152 --> 01:16:31,587 >> I'LL SET MY TIMER. 1495 01:16:31,587 --> 01:16:33,689 THANKS FOR HAVING ME HERE. 1496 01:16:33,689 --> 01:16:40,529 SO MY NAME IS PATRICK MINEAULT 1497 01:16:40,529 --> 01:16:41,764 AND WORKING ON ISSUES WITH A.I., 1498 01:16:41,764 --> 01:16:43,632 NEUROSCIENCE AND A.I. SAFETY. 1499 01:16:43,632 --> 01:16:47,470 NOW, SUCH AN AMAZING OCCASION TO 1500 01:16:47,470 --> 01:16:48,938 MEET UP WITH YOU INVOLVED IN THE 1501 01:16:48,938 --> 01:16:49,738 BRAIN INITIATIVE. 1502 01:16:49,738 --> 01:16:51,540 WE HAVE LEARNED SO MUCH ABOUT 1503 01:16:51,540 --> 01:16:55,444 THE BRAIN IN THE LAST 10 YEARS 1504 01:16:55,444 --> 01:16:56,779 FROM THE CONNECTOME WITH FLIES 1505 01:16:56,779 --> 01:17:01,517 AND HAVING THE ABILITY TO HAVE A 1506 01:17:01,517 --> 01:17:02,251 LOT OF INFORMATION ABOUT CELL 1507 01:17:02,251 --> 01:17:11,660 TYPES AND FUNCTIONAL ACTIVITY AND 1508 01:17:11,660 --> 01:17:13,529 APPLICATIONS IN HUMAN NEUROSCIENCE. 1509 01:17:13,529 --> 01:17:15,130 MEANWHILE A.I. HAS ADVANCED 1510 01:17:15,130 --> 01:17:16,232 TREMENDOUSLY DURING IN THE SAME 1511 01:17:16,232 --> 01:17:16,632 TIME PERIOD. 1512 01:17:16,632 --> 01:17:18,100 OF COURSE, I DON'T NEED TO TELL 1513 01:17:18,100 --> 01:17:24,507 YOU THAT WHEN YOU LOOK AT 1514 01:17:24,507 --> 01:17:25,541 BENCHMARKS IN A.I., THERE'S 1515 01:17:25,541 --> 01:17:26,542 TREMENDOUS IMPROVEMENTS IN THE 1516 01:17:26,542 --> 01:17:27,776 LAST COUPLE OF YEARS. 1517 01:17:27,776 --> 01:17:29,378 IT SEEMS EVERY TIME THERE'S A 1518 01:17:29,378 --> 01:17:32,648 NEW BENCHMARK THAT COMES OUT IT 1519 01:17:32,648 --> 01:17:35,918 ONLY TAKES A FEW YEARS AND 1520 01:17:35,918 --> 01:17:37,286 SOMETIMES MONTHS TO ACTUALLY 1521 01:17:37,286 --> 01:17:38,854 SATURATE IT. 1522 01:17:38,854 --> 01:17:42,224 WE'VE NOW INTEGRATED LARGE 1523 01:17:42,224 --> 01:17:43,893 LANGUAGE MODELS TO EVERYDAY 1524 01:17:43,893 --> 01:17:44,193 WORK. 1525 01:17:44,193 --> 01:17:46,996 WE'VE SEEN THE NOBEL PRIZE TO 1526 01:17:46,996 --> 01:17:53,903 ALPHAFOLD FOR APPLICATIONS IN 1527 01:17:53,903 --> 01:17:57,973 PROTEIN FOLDING. GENERATE PHOTOS 1528 01:17:57,973 --> 01:17:59,775 THE ASPIRATION OF NEUROAI IS TO 1529 01:17:59,775 --> 01:18:01,510 BRING TWO FIELDS TOGETHER AND I'D 1530 01:18:01,510 --> 01:18:04,780 BE REMISSED NOT TO HAVE A FIGURE 1531 01:18:04,780 --> 01:18:10,052 LIKE THIS. WE'RE LEARNING FROM 1532 01:18:10,052 --> 01:18:20,563 NEUROSCIENCE TO BRING INTO A.I. 1533 01:18:21,931 --> 01:18:26,035 BUT WE ARE GETTING A LOT MORE FROM 1534 01:18:26,035 --> 01:18:27,236 A.I. THAN A.I. IS GIVING US. 1535 01:18:27,236 --> 01:18:33,242 I WANT TO TRY AND DIAGNOSE WHY 1536 01:18:33,242 --> 01:18:37,046 THIS IS THE CASE. THE FACT THAT 1537 01:18:37,046 --> 01:18:39,648 WE HAVE TO DO EXPERIMENTS IN THE 1538 01:18:39,648 --> 01:18:44,887 REAL WORLD ON BIOLOGICAL SYSTEMS 1539 01:18:44,887 --> 01:18:46,922 MEANS NEUROSCIENCE IS ALWAYS 1540 01:18:46,922 --> 01:18:47,923 GOING TO BE SLOWER THAN 1541 01:18:47,923 --> 01:18:51,293 ARTIFICIAL INTELLIGENCE. 1542 01:18:51,293 --> 01:18:55,030 AS A POINT HERE, IF YOU LOOK AT 1543 01:18:55,030 --> 01:18:58,500 THE PAPER SUBMITTED IN JULY 2023 1544 01:18:58,500 --> 01:19:00,869 IT NOW HAS 9,500 CITATIONS. THIS 1545 01:19:00,869 --> 01:19:02,304 IS AN A.I. PAPER BUT I DON'T KNOW 1546 01:19:02,304 --> 01:19:04,106 OF A SINGLE PAPER IN 1547 01:19:04,106 --> 01:19:04,940 NEUROSCIENCE THAT WOULD HAVE 1548 01:19:04,940 --> 01:19:08,510 THAT NUMBER OF CITATIONS IN A 1549 01:19:08,510 --> 01:19:09,078 YEAR AND A HALF SINCE THE 1550 01:19:09,078 --> 01:19:13,382 PUBLICATION. THINGS GO A LOT 1551 01:19:13,382 --> 01:19:14,817 FASTER IN A.I. BECAUSE 1552 01:19:14,817 --> 01:19:19,288 WE HAVE THE ABILITY TO DUPLICATE 1553 01:19:19,288 --> 01:19:21,790 MODELS AND DON'T HAVE TO GO 1554 01:19:21,790 --> 01:19:23,192 THROUGH THE BIOLOGICAL LOOP. 1555 01:19:23,192 --> 01:19:26,495 THE RATE OF PUBLICATION IS 1556 01:19:26,495 --> 01:19:30,833 HIGHER AND HOW CAN WE CONTINUE 1557 01:19:30,833 --> 01:19:33,135 TO INSPIRE A.I.? AND THERE'S GOOD 1558 01:19:33,135 --> 01:19:35,170 REASONS WHY WE WANT TO CONTINUE 1559 01:19:35,170 --> 01:19:36,839 TO IN SPIRE A.I. 1560 01:19:36,839 --> 01:19:39,008 ONE IS HUMAN BRAIN IS THE ONLY 1561 01:19:39,008 --> 01:19:42,544 EXAMPLE OF TRUE GENERAL 1562 01:19:42,544 --> 01:19:44,346 INTELLIGENCE THAT'S FLEXIBLE AND 1563 01:19:44,346 --> 01:19:46,515 COMPOSABLE AND EMBODIED. 1564 01:19:46,515 --> 01:19:48,684 THE HUMAN BRAIN IS THE ONLY 1565 01:19:48,684 --> 01:19:50,486 TRUE SAFE GENERAL INTELLIGENCE 1566 01:19:50,486 --> 01:19:53,022 SO WE'RE ABLE TO NEGOTIATE, 1567 01:19:53,022 --> 01:19:58,160 COOPERATE AND UNDERSTAND EACH 1568 01:19:58,160 --> 01:20:04,466 OTHER'S ACTIONS. OUR BRAINS ARE 1569 01:20:04,466 --> 01:20:05,100 ENERGY EFFICIENT. TO ACCELERATE 1570 01:20:05,100 --> 01:20:06,502 NEUROSCIENCE TO CONTINUE BUILDING 1571 01:20:06,502 --> 01:20:06,969 THE BRIDGE TOWARDS A.I. 1572 01:20:06,969 --> 01:20:11,106 WHAT I ADVOCATE FOR IS TO 1573 01:20:11,106 --> 01:20:11,607 LEVERAGE VIRTUAL NEUROSCIENCE. 1574 01:20:11,607 --> 01:20:16,345 SO IF WE'RE ABLE TO STUDY 1575 01:20:16,345 --> 01:20:17,646 VIRTUAL BRAINS RATHER THAN REAL 1576 01:20:17,646 --> 01:20:21,550 BRAINS WE'D HAVE ACCESS TO A 1577 01:20:21,550 --> 01:20:24,720 FULLY OBSERVABLE STATE AND 1578 01:20:24,720 --> 01:20:26,555 REPRODUCIBLE AND HAVE CAUSAL 1579 01:20:26,555 --> 01:20:28,357 ABLATIONS, INTERVENTIONS, EDITS AND 1580 01:20:28,357 --> 01:20:30,559 BE FASTER THAN THE CURRENT MO. 1581 01:20:30,559 --> 01:20:31,927 AND REALLY THE TEMPLATE I'M 1582 01:20:31,927 --> 01:20:37,466 THINKING ABOUT FOR THIS IS WHAT 1583 01:20:37,466 --> 01:20:42,838 WE'VE SEEN IN FLY NEUROSCIENCE. 1584 01:20:42,838 --> 01:20:44,073 IT IS FAR FROM SYSTEMS NEUROSCIENCE. 1585 01:20:44,073 --> 01:20:46,375 BUT I'M ASTOUNDED BY THE 1586 01:20:46,375 --> 01:20:48,911 ADVANCES THAT WE'VE BEEN ABLE TO 1587 01:20:48,911 --> 01:20:51,580 DO FROM CREATING THE CONNECTOME 1588 01:20:51,580 --> 01:20:56,685 OF A FLY AND ABLE TO RECONSTRUCT 1589 01:20:56,685 --> 01:20:59,655 TRACKS AND VIRTUALIZE THE BRAINS 1590 01:20:59,655 --> 01:21:02,224 AND BODIES AND FIGURE OUT 1591 01:21:02,224 --> 01:21:04,827 ESSENTIALLY THE ENTIRE TRACK OF 1592 01:21:04,827 --> 01:21:07,529 TRANSFORMATIONS FROM THE FIRST 1593 01:21:07,529 --> 01:21:08,764 SYNAPSE TO THE MORE CENTRAL 1594 01:21:08,764 --> 01:21:11,300 LOCATION INSIDE THE FLY'S BRAIN. 1595 01:21:11,300 --> 01:21:12,901 IF WE HAD SOMETHING LIKE THAT 1596 01:21:12,901 --> 01:21:14,403 FOR MAMMAL NEUROSCIENCE I THINK IT 1597 01:21:14,403 --> 01:21:19,141 WOULD BE TRULY TRANSFORMATIVE. SO 1598 01:21:19,141 --> 01:21:22,044 WHAT DOES THAT MEAN TO HAVE ACCESS 1599 01:21:22,044 --> 01:21:23,245 TO VIRTUAL NEUROSCIENCE? 1600 01:21:23,245 --> 01:21:25,347 THERE'S MANY THINGS TO TAKE INTO 1601 01:21:25,347 --> 01:21:25,581 ACCOUNT INCLUDING A WORLD MODEL. 1602 01:21:25,581 --> 01:21:28,517 WE WOULD NEED TO TAKE INTO ACCOUNT 1603 01:21:28,517 --> 01:21:29,785 THE VIRTUALIZATION OF THE WORLD 1604 01:21:29,785 --> 01:21:32,254 AND INPUTS OF THE 1605 01:21:32,254 --> 01:21:32,488 SYSTEM. 1606 01:21:32,488 --> 01:21:34,957 WE WOULD NEED TO TAKE INTO 1607 01:21:34,957 --> 01:21:35,924 ACCOUNT THE BRAIN AND 1608 01:21:35,924 --> 01:21:39,728 MEASUREMENTS WE TAKE OUT OF IT 1609 01:21:39,728 --> 01:21:41,497 AND THE POSITION OF THE BODY 1610 01:21:41,497 --> 01:21:42,631 WITHIN ITS ENVIRONMENT. 1611 01:21:42,631 --> 01:21:46,635 BUT WE'VE SEEN A LOT OF THE 1612 01:21:46,635 --> 01:21:48,771 BASIC BUILDING BLOCKS BEING 1613 01:21:48,771 --> 01:21:48,971 BUILT. 1614 01:21:48,971 --> 01:21:51,874 THAT INCLUDES THIS MAPPING FROM 1615 01:21:51,874 --> 01:21:54,009 INPUTS TO MEASUREMENT. WE NOW HAVE 1616 01:21:54,009 --> 01:21:55,577 ACCESS TO NEURAL NETWORKS THAT 1617 01:21:55,577 --> 01:21:58,180 CAN ACCOUNT FOR PROCESSING 1618 01:21:58,180 --> 01:22:03,419 THAT WE SEE IN THE VENTRAL VISUAL 1619 01:22:03,419 --> 01:22:06,355 STREAM. AND THERE'S DATA DRIVEN 1620 01:22:06,355 --> 01:22:09,725 NEURAL NETWORKS AND WE WILL HEAR 1621 01:22:09,725 --> 01:22:10,859 HOW TO BUILD DIGITAL TWINS OF THE 1622 01:22:10,859 --> 01:22:17,666 SYSTEMS. WE HAVE ACCESS TO 1623 01:22:17,666 --> 01:22:18,901 AMAZING TECHNOLOGY TO TRACK THE 1624 01:22:18,901 --> 01:22:20,936 POSITION OF THE BODIES OF 1625 01:22:20,936 --> 01:22:22,471 ANIMALS AND VIRTUALIZATION 1626 01:22:22,471 --> 01:22:23,872 TECHNOLOGIES. 1627 01:22:23,872 --> 01:22:26,175 AND FINALLY WE'RE SEEING A 1628 01:22:26,175 --> 01:22:32,214 REVOLUTION IN OUR ABILITY TO 1629 01:22:32,214 --> 01:22:34,416 AGGREGATE INFORMATION THRU 1630 01:22:34,416 --> 01:22:35,851 FOUNDATION MODELS. THERE'S OVER 1631 01:22:35,851 --> 01:22:36,485 100,000 HOURS IN DATA ARCHIVES 1632 01:22:36,485 --> 01:22:39,054 MANY SUPPORTED BY THE BRAIN 1633 01:22:39,054 --> 01:22:40,389 INITIATIVE AND THE NIH. 1634 01:22:40,389 --> 01:22:45,561 SO IF WE COULD BRING ALL THESE 1635 01:22:45,561 --> 01:22:48,931 DATA TOGETHER I THINK WE'D HAVE 1636 01:22:48,931 --> 01:22:51,033 A FAST PATH TOWARDS VIRTUAL 1637 01:22:51,033 --> 01:22:51,366 NEUROSCIENCE. 1638 01:22:51,366 --> 01:22:53,168 WHAT'S THE RECIPE? 1639 01:22:53,168 --> 01:22:54,136 WE'RE TALKING ABOUT DIGITAL 1640 01:22:54,136 --> 01:22:58,941 TWINS, AUTO REGRESSIVE FOUNDATION 1641 01:22:58,941 --> 01:23:00,876 MODELS, VIRTUAL BODIES, AND 1642 01:23:00,876 --> 01:23:01,910 VIRTUAL ENVIRONMENTS. 1643 01:23:01,910 --> 01:23:03,045 TO BUILD THE VIRTUAL BRAINS WE 1644 01:23:03,045 --> 01:23:06,682 WANT TO LEVERAGE THE EXISTING 1645 01:23:06,682 --> 01:23:07,049 DATA ARCHIVES. 1646 01:23:07,049 --> 01:23:10,118 BRING CLOSED DATA IN ADDITION TO 1647 01:23:10,118 --> 01:23:11,353 THAT. ACCELERATE COST-CUTTING DATA 1648 01:23:11,353 --> 01:23:13,288 ACQUISITION TECHNOLOGIES AND 1649 01:23:13,288 --> 01:23:16,191 CONNECTOMICS, WIRELESS 1650 01:23:16,191 --> 01:23:18,093 ELECTROPHYSIOLOGY, AND POOL DATA 1651 01:23:18,093 --> 01:23:21,864 ACQUISITION ACROSS MULTIPLE LABS 1652 01:23:21,864 --> 01:23:23,665 UNDER HIGH-ENTROPY CONDITIONS. 1653 01:23:23,665 --> 01:23:25,467 AND POOLED RESOURCES TO BUILD 1654 01:23:25,467 --> 01:23:26,802 TWINS, FOUNDATION MODELS AND 1655 01:23:26,802 --> 01:23:31,139 VIRTUAL BODIES AND DISSEMINATE 1656 01:23:31,139 --> 01:23:34,743 TO THE RESEARCH COMMUNITY. 1657 01:23:34,743 --> 01:23:36,178 THE REALITY OF NEUROAI MAY SEEM 1658 01:23:36,178 --> 01:23:37,913 LIKE THERE'S MORE WE'RE GETTING 1659 01:23:37,913 --> 01:23:42,150 OUT OF A.I. BUT THROUGH THE PATH 1660 01:23:42,150 --> 01:23:45,420 I THINK WE COULD POTENTIALLY 1661 01:23:45,420 --> 01:23:46,455 MAKE THE FUTURE OF NEUROAI WHERE 1662 01:23:46,455 --> 01:23:49,558 THE ARROW AT THE TOP IS LARGER. 1663 01:23:49,558 --> 01:23:59,902 THANK YOU VERY MUCH. 1664 01:24:18,520 --> 01:24:19,521 >> THANK YOU. 1665 01:24:19,521 --> 01:24:21,323 THANK YOU FOR THE ORGANIZERS FOR 1666 01:24:21,323 --> 01:24:27,129 THIS GREAT WORKSHOP. 1667 01:24:27,129 --> 01:24:32,601 SO, ARE THESE DOGS OR MUFFINS? 1668 01:24:32,601 --> 01:24:36,638 WHEN YOU ASSEMBLE THEM AND SHOW 1669 01:24:36,638 --> 01:24:39,441 THEM TO THE MODELS IN A.I., THEY 1670 01:24:39,441 --> 01:24:45,580 FAIL AND EVEN SOME CANNOT EVEN 1671 01:24:45,580 --> 01:24:48,350 PARSE THE IMAGES TO GET THE 1672 01:24:48,350 --> 01:24:49,351 RIGHT NUMBER OF ROWS AND COLUMNS. 1673 01:24:49,351 --> 01:24:51,486 THE REASON THEY FAIL IS BECAUSE 1674 01:24:51,486 --> 01:24:52,120 OF GENERALIZATION. 1675 01:24:52,120 --> 01:24:55,057 IT'S THE KEY INGREDIENT OF 1676 01:24:55,057 --> 01:24:55,958 INTELLIGENCE OF ARTIFICIAL AND 1677 01:24:55,958 --> 01:24:59,928 NATURAL AND WHAT I MEAN BY 1678 01:24:59,928 --> 01:25:01,496 GENERALIZATIONS IS THE ABILITY OF 1679 01:25:01,496 --> 01:25:02,831 AGENTS, HUMANS, A.I. SYSTEMS TO 1680 01:25:02,831 --> 01:25:06,168 DRAW INFERENCES AND MAKE 1681 01:25:06,168 --> 01:25:08,103 PREDICTIONS OF NEW DATA YOU'VE 1682 01:25:08,103 --> 01:25:09,504 NEVER SEEN BEFORE OR ON DATA 1683 01:25:09,504 --> 01:25:09,972 OUTSIDE THE TRAINING 1684 01:25:09,972 --> 01:25:14,009 DISTRIBUTION. AND BIG TECH'S 1685 01:25:14,009 --> 01:25:16,578 ANSWER TO THIS GENERALIZATION IS 1686 01:25:16,578 --> 01:25:20,215 ILLUSTRATED HERE WHERE FREDDY ASKED 1687 01:25:20,215 --> 01:25:22,050 JOHNNY DEPP MY MODEL CANNOT DO IT. 1688 01:25:22,050 --> 01:25:29,224 ANSWER IS TRAIN IT UNTIL THERE IS 1689 01:25:29,224 --> 01:25:30,926 NOTHING LEFT OUT OF DISTRIBUTION. 1690 01:25:30,926 --> 01:25:33,095 THIS IS WHY WE NEED 1691 01:25:33,095 --> 01:25:34,796 LARGE COMPUTE AND MORE LARGE 1692 01:25:34,796 --> 01:25:36,898 DATA TO BUILD THESE MODELS. 1693 01:25:36,898 --> 01:25:39,234 AND WHILE THIS IS GOING TO 1694 01:25:39,234 --> 01:25:42,137 CONTINUE HAPPENING AND WE'LL SEE 1695 01:25:42,137 --> 01:25:43,538 VERY IMPORTANT IMPROVEMENTS WE 1696 01:25:43,538 --> 01:25:46,475 MAY BE RUNNING OUT OF DATA. 1697 01:25:46,475 --> 01:25:47,209 FOR EXAMPLE, THERE'S EVIDENCE 1698 01:25:47,209 --> 01:25:48,910 WE'RE RUNNING OUT OF HIGH 1699 01:25:48,910 --> 01:25:50,512 QUALITY TEXT DATA AND IN MANY 1700 01:25:50,512 --> 01:25:54,916 CASES WE JUST DON'T HAVE ENOUGH 1701 01:25:54,916 --> 01:25:56,585 DATA TO BUILD POWERFUL A.I. 1702 01:25:56,585 --> 01:25:56,818 MODELS. 1703 01:25:56,818 --> 01:25:58,720 SO WE NEED NEW ALGORITHMS. 1704 01:25:58,720 --> 01:25:59,688 AS WE HEARD IN THE PREVIOUS 1705 01:25:59,688 --> 01:26:00,756 TALKS WE NEED TO DRAW 1706 01:26:00,756 --> 01:26:03,992 INSPIRATION FROM THE BRAIN. 1707 01:26:03,992 --> 01:26:06,328 NEUROSCIENCE IS ALL ABOUT FINDING 1708 01:26:06,328 --> 01:26:09,431 CORRESPONDENCES, IN SOME CASES, 1709 01:26:09,431 --> 01:26:11,333 ISOMORPHISMS BETWEEN THE EXTERNAL 1710 01:26:11,333 --> 01:26:15,637 WORLD OUR PERCEPTION OUR ACTIONS 1711 01:26:15,637 --> 01:26:17,139 OUR NEUROCOGNITIVE STATES AT 1712 01:26:17,139 --> 01:26:18,807 DIFFERENT LEVELS OF ANALYSIS. 1713 01:26:18,807 --> 01:26:24,046 A PARTICULAR LEVEL THAT'S BEEN 1714 01:26:24,046 --> 01:26:30,819 FRUITFUL, AT THE REPRESENTATIONAL 1715 01:26:30,819 --> 01:26:31,386 LEVEL, WE FOUND CORRESPONDENCES 1716 01:26:31,386 --> 01:26:33,622 BETWEEN OUR PERCEPTIONS OUR ACTIONS 1717 01:26:33,622 --> 01:26:35,023 OUR LOCATIONS IN SPACE AND 1718 01:26:35,023 --> 01:26:45,400 NEURAL ACTIVITY IN THE BRAIN. 1719 01:26:46,401 --> 01:26:49,404 AT THIS LEVEL THERE'S AN 1720 01:26:49,404 --> 01:26:51,406 INTERESTING FACT OF NEURAL 1721 01:26:51,406 --> 01:26:53,375 CODING THERE ARE PRINCIPLES THAT 1722 01:26:53,375 --> 01:26:54,743 TRANSCEND THE SUBSTRATE. 1723 01:26:54,743 --> 01:26:59,181 WHAT DO I MEAN BY THIS IS IF YOU 1724 01:26:59,181 --> 01:27:00,816 TAKE ARTIFICIAL NEURAL NETWORK AND 1725 01:27:00,816 --> 01:27:03,718 TRAIN IT IN OBJECT RECOGNITION AND 1726 01:27:03,718 --> 01:27:08,223 LOOK INSIDE THE BLACK BOX YOU'LL 1727 01:27:08,223 --> 01:27:09,891 FIND REPRESENTATIONS THAT LOOK 1728 01:27:09,891 --> 01:27:12,394 SIMILAR TO WHAT WAS STUDIED IN 1729 01:27:12,394 --> 01:27:14,262 THE BRAIN. 1730 01:27:14,262 --> 01:27:14,863 HERE'S DIFFERENT ARTIFICIAL 1731 01:27:14,863 --> 01:27:17,933 NEURAL NETWORKS AND YOU SEE IF 1732 01:27:17,933 --> 01:27:20,635 YOU LOOK AT THE CERTAIN LAYER 1733 01:27:20,635 --> 01:27:23,839 WITHIN THE ARTIFICIAL NEURAL 1734 01:27:23,839 --> 01:27:26,007 NETWORK YOU HAVE DETECTED 1735 01:27:26,007 --> 01:27:26,274 CURVATURE OR DIFFERENT PATTERNS 1736 01:27:26,274 --> 01:27:29,945 AND IF YOU DO THE SAME FOR THE 1737 01:27:29,945 --> 01:27:33,448 AREA BEFORE A MID LEVEL AREA 1738 01:27:33,448 --> 01:27:35,083 IN THE MACAQUE BRAIN YOU START 1739 01:27:35,083 --> 01:27:40,655 FINDING SIMILAR FEATURES. 1740 01:27:40,655 --> 01:27:43,325 HERE'S ANOTHER EXAMPLE OF THE GRID 1741 01:27:43,325 --> 01:27:53,435 CELLS IN THE ENTORHINAL CORTEX 1742 01:27:53,435 --> 01:27:55,403 THAT ENCODE SPATIAL INFORMATION 1743 01:27:55,403 --> 01:27:55,670 IN VISION YOU FIND SIMILAR 1744 01:27:55,670 --> 01:28:00,575 NOW WE HAVE A WAY TO COMPARE 1745 01:28:00,575 --> 01:28:01,743 ARTIFICIAL NEURAL NETWORKS IN 1746 01:28:01,743 --> 01:28:03,645 BRAINS IN THE SAME WAY AND THEN 1747 01:28:03,645 --> 01:28:09,651 LOOK FOR SIMILARITIES AND 1748 01:28:09,651 --> 01:28:09,985 DIFFERENCES. 1749 01:28:09,985 --> 01:28:12,954 THIS IS A FRUITFUL APPROACH TO 1750 01:28:12,954 --> 01:28:14,189 TRAIN ADVANCED A.I. AND 1751 01:28:14,189 --> 01:28:18,760 DIFFERENCES IN THE PRESENTATION 1752 01:28:18,760 --> 01:28:24,166 IS WHAT ENABLES A.I. SYSTEMS 1753 01:28:24,166 --> 01:28:27,369 IF YOU LOOK AT THE HISTORY OF 1754 01:28:27,369 --> 01:28:28,036 STUDYING REPRESENTATION AND 1755 01:28:28,036 --> 01:28:30,472 INFORMATION PROCESSING OF NEURAL 1756 01:28:30,472 --> 01:28:35,610 CODE, IT'S MOSTLY SERENDIPITOUS. 1757 01:28:35,610 --> 01:28:44,085 HERE'S AN EXAMPLE OF HUBEL AND 1758 01:28:44,085 --> 01:28:49,691 WEISEL WHERE THEY DISCOVERED 1759 01:28:49,691 --> 01:28:54,763 ORIENTATION SELECTIVITY AND WON 1760 01:28:54,763 --> 01:28:56,898 THE NOBEL PRIZE FOR IT. THIS 60-MINUTE CLIP DESCRIBED THE SERENDIPITOUS DISCOVERY THAT ONE DAY THE SLIDE 1761 01:28:56,898 --> 01:28:59,301 THEY WERE USING PRODUCED THIS THIN LINE. AND THEY FOUND THAT THIS WAS DRIVING THE NEURONS. 1762 01:28:59,301 --> 01:29:07,943 NOW BELIEVE IT OR NOT, THIS INSPIRED FUKUSHIMA AND GAVE RISE 1763 01:29:07,943 --> 01:29:09,010 TO CONVOLUTIONAL NEURAL NETWORKS. 1764 01:29:09,010 --> 01:29:10,579 A HALLMARK OF THE A.I. REVOLUTION. 1765 01:29:10,579 --> 01:29:15,784 IF YOU LOOK AT THE BRAIN OF THE 1766 01:29:15,784 --> 01:29:17,519 MACAQUE THERE'S THOUSANDS OF VISUAL 1767 01:29:17,519 --> 01:29:26,728 AREAS, MANY CORTICAL LAYERS, 1768 01:29:26,728 --> 01:29:31,166 HUNDREDS OF CELL TYPES. SO HOW TO 1769 01:29:31,166 --> 01:29:32,133 DECIPHER REPRESENTATIONS? FOLLOW 1770 01:29:32,133 --> 01:29:35,637 FOLLOW THIS SERENDIPITOUS PROCESS. 1771 01:29:35,637 --> 01:29:37,439 IF YOU HAVE TO MONITOR, 256 X 256 1772 01:29:37,439 --> 01:29:39,207 PIXELS AND JUST WANT TO STUDY THE 1773 01:29:39,207 --> 01:29:44,012 SYSTEM YOU CAN GENERATE TWO TO 1774 01:29:44,012 --> 01:29:45,880 65,000 POSSIBLE PATTERNS AND IF YOU 1775 01:29:45,880 --> 01:29:47,482 TAKE INTO ACCOUNT BEHAVIORAL 1776 01:29:47,482 --> 01:29:49,017 STATES AND DIFFERENT INTERNAL 1777 01:29:49,017 --> 01:29:50,485 STATES AND MODALITIES IT BECOMES 1778 01:29:50,485 --> 01:30:00,896 A VERY COMPLEX PROBLEM. WE ARE IN 1779 01:30:04,466 --> 01:30:05,901 A GOLDERN ERA BECAUSE OF COLLECTING 1780 01:30:05,901 --> 01:30:06,901 DATA AT SCALE. THANKS TO THE BRAIN 1781 01:30:06,901 --> 01:30:08,570 INITIATIVE IN THE LAST DECADE WE 1782 01:30:08,570 --> 01:30:08,903 HAVE THESE AMAZING TECHNOLOGIES. 1783 01:30:08,903 --> 01:30:15,477 AND THE OTHER IS THAT WE HAVE 1784 01:30:15,477 --> 01:30:16,077 THE A.I REVOLUTION WITH THE 1785 01:30:16,077 --> 01:30:18,546 TECHNOLOGY TO ENABLE TO INGEST 1786 01:30:18,546 --> 01:30:19,547 LARGE SCALE DATA AND FIND 1787 01:30:19,547 --> 01:30:29,758 RELATIONSHIPS AND MAKE PREDICTIONS. 1788 01:30:31,159 --> 01:30:33,995 WE CREATE PLATFORMS TO COLLECT 1789 01:30:33,995 --> 01:30:36,998 LARGE SCALE DATA, USE MACHINE 1790 01:30:36,998 --> 01:30:39,501 LEARINGTO BUILD PREDICTIVE MODELS, 1791 01:30:39,501 --> 01:30:40,735 USE LATENT AND WORLD VARIABLES, 1792 01:30:40,735 --> 01:30:42,671 BRAIN ACIVITY AND BEHAVIOR 1793 01:30:42,671 --> 01:30:45,340 THEN ONCE YOU HAVE THE DIGITAL 1794 01:30:45,340 --> 01:30:47,108 TWINS, NOW YOU CAN IMAGINE YOU 1795 01:30:47,108 --> 01:30:48,977 CAN SCALE THEM AND RUN WHAT IT 1796 01:30:48,977 --> 01:30:51,246 WOULD TAKE HUNDREDS OF YEARS AND 1797 01:30:51,246 --> 01:30:57,285 SOMETIMES THOUSAND OF YEARS. IF 1798 01:30:57,285 --> 01:31:00,522 YOU HAVE A LARGE ENOUGH 1799 01:31:00,522 --> 01:31:02,057 INFRASTRUCTURE, YOU CAN RUN IT IN 1800 01:31:02,057 --> 01:31:07,495 2 DAYS. MAKE DISCOVERIES, GENERATE 1801 01:31:07,495 --> 01:31:08,263 HYPOTHESIS THAT YOU CAN TEST. 1802 01:31:08,263 --> 01:31:10,732 HERE'S AN EXAMPLE OF A STUDY WE 1803 01:31:10,732 --> 01:31:12,534 DID A COUPLE YEARS AGO WHERE WE 1804 01:31:12,534 --> 01:31:14,736 WANTED TO STUDY HOW THE 1805 01:31:14,736 --> 01:31:19,808 SENSITIVITY OF THE MOUSE VISUAL 1806 01:31:19,808 --> 01:31:23,945 SYSTEM TUNING CHANGES DURING THE 1807 01:31:23,945 --> 01:31:25,347 CHANGE OF COGNITIVE STATES. 1808 01:31:25,347 --> 01:31:29,050 WE SHOWED NATURAL IMAGES WITH 1809 01:31:29,050 --> 01:31:31,686 COLOR AND BUILD A DIGITAL TWIN OF 1810 01:31:31,686 --> 01:31:34,856 THIS MODEL. YOU CAN SEE HERE WE 1811 01:31:34,856 --> 01:31:36,791 DISCOVERED WHEN THE PUPIL 1812 01:31:36,791 --> 01:31:40,362 DILATES THE COLOR SELECTIVITY 1813 01:31:40,362 --> 01:31:48,937 CHANGES TO UV BUT SPATIAL STRUCTURE 1814 01:31:48,937 --> 01:31:51,639 DIDN'T. THIS WASN'T A HYPOTHESIS 1815 01:31:51,639 --> 01:31:53,808 DRIVEN DISCOVERY. IT WAS GENERATED 1816 01:31:53,808 --> 01:32:03,885 BY THE MODEL AND COULD BE VERIFIED 1817 01:32:03,885 --> 01:32:07,355 INSILICO. IN THE NEXT DECADE WE'RE 1818 01:32:07,355 --> 01:32:08,223 IN THIS TIME IN NEUROSCIENCE AND 1819 01:32:08,223 --> 01:32:10,625 A.I. WHERE IF YOU THINK ABOUT IT, 1820 01:32:10,625 --> 01:32:13,561 A.I. IS ALL ABOUT GETTING HUMAN 1821 01:32:13,561 --> 01:32:17,365 BEHAVIORAL DATA AND TRAINING MODELS 1822 01:32:17,365 --> 01:32:19,234 TO MEET HUMANS INTELLIGENCE SYSTEMS. 1823 01:32:19,234 --> 01:32:21,336 NOW WE HAVE OPPORTUNITIES TO GET 1824 01:32:21,336 --> 01:32:26,541 NEURAL DATA AT LARGE SCALE AND 1825 01:32:26,541 --> 01:32:28,209 PRODUCE A.I. SYSTEMS THAT WILL 1826 01:32:28,209 --> 01:32:33,181 NOT JUST HAVE BEHAVIORAL 1827 01:32:33,181 --> 01:32:34,816 ALIGNMENT WITH OUR BRAINS BUT ALSO 1828 01:32:34,816 --> 01:32:35,650 NEURAL ALIGNMENT. 1829 01:32:35,650 --> 01:32:40,255 I'D LIKE TO CLOSE WITH THIS QUOTE 1830 01:32:40,255 --> 01:32:42,557 FROM DANIEL DENNETT: WE HOPE TO 1831 01:32:42,557 --> 01:32:46,094 BUILD A.I. SYSTEMS AND NOT JUST BE 1832 01:32:46,094 --> 01:32:51,299 COMPETENT BUT WILL COMPREHEND THE 1833 01:32:51,299 --> 01:32:55,503 WORLD. HE'S USING AN EXAMPLE OF 1834 01:32:55,503 --> 01:32:58,673 TERMITE MOUNDS. THEY ARE BEAUTIFUL 1835 01:32:58,673 --> 01:33:02,610 STRUCTURE AND ONE BUILT FROM SIMPLE 1836 01:33:02,610 --> 01:33:08,049 RULES NOT COMPREHENDING BUILDING 1837 01:33:08,049 --> 01:33:11,453 THE MASTER PIECE. THANK YOU TO THE 1838 01:33:11,453 --> 01:33:12,086 BRAIN INITIATIVE FOR PROVIDING 1839 01:33:12,086 --> 01:33:22,230 FUNDING. 1840 01:33:38,112 --> 01:33:40,014 >> ALL RIGHT. 1841 01:33:40,014 --> 01:33:40,615 HELLO, EVERYONE. EXCITING 1842 01:33:40,615 --> 01:33:42,016 WORKSHOP. GLAD TO BE HERE. 1843 01:33:42,016 --> 01:33:42,617 THANKS TO THE ORGANIZERS FOR 1844 01:33:42,617 --> 01:33:44,686 INVITING ME. 1845 01:33:44,686 --> 01:33:49,224 SO, WHAT I WOULD LIKE TO DO IS 1846 01:33:49,224 --> 01:33:52,694 TO MAKE A PITCH FOR AN IDEA OF 1847 01:33:52,694 --> 01:33:54,762 WHAT WE AS A COMMUNITY COULD 1848 01:33:54,762 --> 01:33:55,597 STUDY AMONG OTHER THINGS OF 1849 01:33:55,597 --> 01:33:59,133 COURSE. 1850 01:33:59,133 --> 01:34:02,637 IN THE NEXT DECADE OR SO AND 1851 01:34:02,637 --> 01:34:05,473 NAMELY WHAT I WANT TO ARGUE FOR 1852 01:34:05,473 --> 01:34:08,977 IS HIGHLY DETAILED BIOLOGICALLY 1853 01:34:08,977 --> 01:34:13,915 REALISTIC SIMULATIONS OF BRAINS AT 1854 01:34:13,915 --> 01:34:22,423 THE WHOLE BRAIN LEVEL OF MAMMALS 1855 01:34:22,423 --> 01:34:22,590 IN THE NEXT 10 YEARS AND WHY? 1856 01:34:22,590 --> 01:34:25,693 WE CARE ABOUT UNDERSTANDING THE 1857 01:34:25,693 --> 01:34:29,297 BRAINS AND UNDERSTANDING HOW 1858 01:34:29,297 --> 01:34:31,499 THEY WORK AND FIXING THAT. 1859 01:34:31,499 --> 01:34:37,372 AND I THINK SUCH SIMULATIONS 1860 01:34:37,372 --> 01:34:39,374 COULD CONSTITUTE EXCELLENT 1861 01:34:39,374 --> 01:34:44,178 PLATFORMS FOR DISCOVERY AND 1862 01:34:44,178 --> 01:34:46,581 IN SILICO EXPERIMENTATION. 1863 01:34:46,581 --> 01:34:48,216 I'LL SAY MORE WORDS ABOUT THAT 1864 01:34:48,216 --> 01:34:49,317 AT THE END BUT FIRST LET'S THINK 1865 01:34:49,317 --> 01:34:50,084 ABOUT WHAT'S POSSIBLE NOW. 1866 01:34:50,084 --> 01:34:52,487 I WANTED TO SHOW THE MOVIE WHICH 1867 01:34:52,487 --> 01:34:54,622 ILLUSTRATES THE MODEL OF THE 1868 01:34:54,622 --> 01:34:56,991 MOUSE PRIMARY VISUAL CORTEX WE 1869 01:34:56,991 --> 01:34:57,859 DEVELOPED AT THE ALLEN INSTITUTE 1870 01:34:57,859 --> 01:35:03,698 AND PUBLISHED A FEW YEARS AGO. 1871 01:35:03,698 --> 01:35:05,700 THIS INTEGRATED WORK FROM OTHER 1872 01:35:05,700 --> 01:35:10,805 SOURCES FROM THE CELL TYPES AND 1873 01:35:10,805 --> 01:35:16,044 DATA PROPERTIES, CONNECTIVITY, 1874 01:35:16,044 --> 01:35:21,282 IN VIVO ACTIVITY. THIS MODEL COMES 1875 01:35:21,282 --> 01:35:24,252 WITH VISUAL INPUT SO WE CAN STUDY 1876 01:35:24,252 --> 01:35:25,587 STUDY NEUROPHYSIOLOGICAL ACTIVITY 1877 01:35:25,587 --> 01:35:31,526 IN THIS SIMULATION. 1878 01:35:31,526 --> 01:35:33,294 WE MADE THE MODELS PUBLICLY 1879 01:35:33,294 --> 01:35:33,728 AVAILABLE. WE HAVE TWO VERSIONS. 1880 01:35:33,728 --> 01:35:39,233 WE HAVE A DETAILED VERSION AND A 1881 01:35:39,233 --> 01:35:46,074 POINT-NEURON VERSION. FOR THE 1882 01:35:46,074 --> 01:35:53,615 METRICS WE CAN COMPUTE SPIKING 1883 01:35:53,615 --> 01:35:55,383 ACTIVITIES IT REPRODUCES WELL. 1884 01:35:55,383 --> 01:35:56,918 BOTH MODELS HAVE A QUARTER 1885 01:35:56,918 --> 01:35:57,485 MILLON NEURONS AND A LOT OF 1886 01:35:57,485 --> 01:36:00,421 COMPLEXITY. 1887 01:36:00,421 --> 01:36:05,660 OTHER GROUPS CAME OUT WITH 1888 01:36:05,660 --> 01:36:10,164 SIMILAR MODELS AND I FEEL LIKE 1889 01:36:10,164 --> 01:36:15,436 WE'RE AT THE POINT WHERE THIS IS 1890 01:36:15,436 --> 01:36:16,871 POSSIBLE AND CAN MAKE LARGER 1891 01:36:16,871 --> 01:36:22,110 SIMULATIONS AND MODELS. 1892 01:36:22,110 --> 01:36:23,911 MULTIPLE GROUPS ARE USING OTHER 1893 01:36:23,911 --> 01:36:24,479 OUR AND OTHER MODELS AVAILABLE. 1894 01:36:24,479 --> 01:36:31,085 I WANT TO HIGHLIGHT THE WORK OF 1895 01:36:31,085 --> 01:36:34,122 A COLLABORATOR WHO PORTED OUR 1896 01:36:34,122 --> 01:36:42,096 MODEL TO TENSOR FLOW AND WERE ABLE 1897 01:36:42,096 --> 01:36:45,500 TO TRAIN IT. ONE EXAMPLE HERE 1898 01:36:45,500 --> 01:36:48,870 THE CURVE IN BLUE SHOWS THE BIO- 1899 01:36:48,870 --> 01:36:54,842 REALISTIC MODEL IS ORDERS OF 1900 01:36:54,842 --> 01:37:05,386 MAGNITUDE MORE ROBUST TO NOISE 1901 01:37:07,388 --> 01:37:08,589 IN THE HANDWRITTEN DIGIT 1902 01:37:08,589 --> 01:37:10,625 RECOGNITION TASK. A LOT CAN BE DONE 1903 01:37:10,625 --> 01:37:11,659 ON THIS MODEL STILL KEEPING THE 1904 01:37:11,659 --> 01:37:17,265 BIOLOGICAL CONSTRAINTS. 1905 01:37:17,265 --> 01:37:18,299 WHAT I BELIEVE IS POSSIBLE IN 1906 01:37:18,299 --> 01:37:19,767 THE NEXT DECADE IS GOING BIGGER 1907 01:37:19,767 --> 01:37:22,270 AND GOING TO THE WHOLE BRAIN. 1908 01:37:22,270 --> 01:37:24,939 WHAT'S ALREADY POSSIBLE NOW ARE 1909 01:37:24,939 --> 01:37:27,408 SIMULATIONS OF THE BRAINS OF A 1910 01:37:27,408 --> 01:37:30,078 A WORM, A FLY AS WE'VE SEEN IN 1911 01:37:30,078 --> 01:37:30,712 BEAUTIFUL PUBLICATIONS IN RECENT 1912 01:37:30,712 --> 01:37:35,650 MONTHS. 1913 01:37:35,650 --> 01:37:38,019 I BELIEVE ZEBRAFISH IS GETTING 1914 01:37:38,019 --> 01:37:40,388 THERE. WHAT CAN BE POSSIBLE SOON 1915 01:37:40,388 --> 01:37:45,526 IS A SIMULATION OF THE WHOLE MOUSE 1916 01:37:45,526 --> 01:37:49,330 BRAIN OR SUBSTANTIAL PIECE, 1 CC, 1917 01:37:49,330 --> 01:37:52,033 OF A HUMAN OR MONKEY BRAIN 1918 01:37:52,033 --> 01:37:53,167 AT CELLULAR RESOLUTION. 1919 01:37:53,167 --> 01:38:00,108 WE HAVE THE TOOLS TO DO THAT AND 1920 01:38:00,108 --> 01:38:01,242 THERE HAVE BEEN PREPRINTS WHERE 1921 01:38:01,242 --> 01:38:06,748 SIMULATIONS ON THE SCALE OF A 1922 01:38:06,748 --> 01:38:13,421 HUMAN BRAIN WAS DEMONSTRATED. 1923 01:38:13,421 --> 01:38:18,392 TOOLS SHOWN HERE AND FOR FAST 1924 01:38:18,392 --> 01:38:20,895 SIMULATIONS, DENDRITIC HIERARCHICAL 1925 01:38:20,895 --> 01:38:23,097 SCHEDULING, OR JAXLEY TOOL THAT 1926 01:38:23,097 --> 01:38:26,234 SHOW EXCITING RESULTS OF BIO- 1927 01:38:26,234 --> 01:38:27,635 PHYSICALLY DETAILED SIMULATIONS 1928 01:38:27,635 --> 01:38:30,438 OF GPUs AND OF COURSE WE NEED 1929 01:38:30,438 --> 01:38:33,508 THE DATA TO MAKE IT REALITY AND 1930 01:38:33,508 --> 01:38:35,977 A LOT OF THE DATA AS A BEEN 1931 01:38:35,977 --> 01:38:37,645 COLLECTED ALREADY AS PART OF THE 1932 01:38:37,645 --> 01:38:41,816 BRAIN INITIATIVE. 1933 01:38:41,816 --> 01:38:46,521 ON ONE HAND ONE KEY COMPONENT IS 1934 01:38:46,521 --> 01:38:48,589 GOING TO BE THE CONNECTOME OF 1935 01:38:48,589 --> 01:38:49,457 THE WHOLE MOUSE BRAIN. 1936 01:38:49,457 --> 01:38:54,028 THERE'S GROUPS WORKING ON THIS 1937 01:38:54,028 --> 01:38:57,832 ALREADY EITHER USING ELECTRON 1938 01:38:57,832 --> 01:38:58,032 OR EXPANSION LIGHT MICROSCOPY. 1939 01:38:58,032 --> 01:39:00,034 AND THAT'S NOT ENOUGH. 1940 01:39:00,034 --> 01:39:04,205 WE NEED TO IMBUE THE STRUCTURE 1941 01:39:04,205 --> 01:39:08,242 ONE LIFE BY COLLECTING DATA ON 1942 01:39:08,242 --> 01:39:10,511 ELECTROPHYSIOLOGY AND BIOCHEMICAL 1943 01:39:10,511 --> 01:39:13,047 PROPERTIES OF CELLS AND CONNECTION 1944 01:39:13,047 --> 01:39:13,981 AND DATA CONNECTIVITY IN VIVO. 1945 01:39:13,981 --> 01:39:15,183 THAT'S IMPRACTICAL TO ACHIEVE AT 1946 01:39:15,183 --> 01:39:16,517 THE LEVEL OF THE WHOLE BRAIN FOR 1947 01:39:16,517 --> 01:39:18,519 EVERY SINGLE CELL IN THE NEXT 10 1948 01:39:18,519 --> 01:39:18,786 YEARS. 1949 01:39:18,786 --> 01:39:25,259 BUT WHAT CAN BE DONE, I BELIEVE, 1950 01:39:25,259 --> 01:39:34,936 IS FOCUSING EXPERIMENTS ON 1951 01:39:34,936 --> 01:39:36,470 TRAINING A.I. TO PREDICT FROM 1952 01:39:36,470 --> 01:39:39,207 DATA WHAT THE PROPERTIES OF 1953 01:39:39,207 --> 01:39:40,741 THOSE BRAIN CELLS AND 1954 01:39:40,741 --> 01:39:44,645 CONNECTIONS ARE GOING TO BE 1955 01:39:44,645 --> 01:39:46,080 BASED ON A SERIES OF EXPERIMENTS 1956 01:39:46,080 --> 01:39:48,049 IN VARIOUS PARTS OF THE BRAIN. 1957 01:39:48,049 --> 01:39:53,354 ULTIMATELY, I BELIEVE THERE'S A 1958 01:39:53,354 --> 01:39:59,227 SPACE HERE FOR A PIPELINE THAT 1959 01:39:59,227 --> 01:40:04,665 CREATES BLUEPRINTS FOR 1960 01:40:04,665 --> 01:40:08,202 SIMULATIONS REGARDING WHAT 1961 01:40:08,202 --> 01:40:09,437 PROPERTIES THOSE COMPONENTS 1962 01:40:09,437 --> 01:40:10,204 SHOULD HAVE. 1963 01:40:10,204 --> 01:40:11,539 THIS IS FEASIBLE AND WE MAY BE 1964 01:40:11,539 --> 01:40:13,274 ABLE TO ACHIEVE WHOLE BRAIN AT 1965 01:40:13,274 --> 01:40:15,309 LEAST FOR THE MOUSE SIMULATIONS 1966 01:40:15,309 --> 01:40:21,449 IN THE NEXT TEN YEARS OR SO. TO 1967 01:40:21,449 --> 01:40:24,118 CONCLUDE I WANT TO ECHO WHY IT'S 1968 01:40:24,118 --> 01:40:24,385 IMPORTANT. 1969 01:40:24,385 --> 01:40:28,022 TO GIVE A COUPLE EXAMPLES, WE 1970 01:40:28,022 --> 01:40:30,358 CAN PRODUCE HUGE AMOUNTS OF 1971 01:40:30,358 --> 01:40:34,996 ENERGY BUT STILL CANNOT REALLY 1972 01:40:34,996 --> 01:40:39,200 PUT TOGETHER MITOCHONDRIA ON 1973 01:40:39,200 --> 01:40:41,702 DESIGN. WE CAN MOVE WITH CARS BUT 1974 01:40:41,702 --> 01:40:47,208 IF SOMETHING GOES WRONG IT'S NOT 1975 01:40:47,208 --> 01:40:49,343 STRAIGHTFORWARD TO FIX. A.I. IS 1976 01:40:49,343 --> 01:40:52,580 AMAZING GROWING FAST AND PROBABLY 1977 01:40:52,580 --> 01:40:55,049 WILL DEVELOP SUBSTANTIALLY WITHOUT 1978 01:40:55,049 --> 01:40:55,983 TOO MUCH INPUT FROM THE 1979 01:40:55,983 --> 01:41:01,155 NEUROSCIENCE BUT WE CAN 1980 01:41:01,155 --> 01:41:03,658 LEVERAGE IT ALSO AND LEARN A 1981 01:41:03,658 --> 01:41:04,692 LOT OF FROM THE BRAIN. BUT 1982 01:41:04,692 --> 01:41:07,662 WE CAN LEVERAGE THE DEVELOPMENT 1983 01:41:07,662 --> 01:41:10,197 OF A.I. TO CREATE DETAILED 1984 01:41:10,197 --> 01:41:11,465 SIMULATIONS OF THE BRAINS TO 1985 01:41:11,465 --> 01:41:14,502 LEARN ABOUT THE MECHANISMS OF THE 1986 01:41:14,502 --> 01:41:19,640 BRAIN THAT GIVE RISE TO THE 1987 01:41:19,640 --> 01:41:22,009 REPRESENTATIONS THAT ANDREAS 1988 01:41:22,009 --> 01:41:24,946 MENTIONED AND STUDY THAT 1989 01:41:24,946 --> 01:41:29,684 COMPREHENSIVELY IN HEALTHY 1990 01:41:29,684 --> 01:41:32,787 CONDITIONS AS WELL AS IN DISEASE 1991 01:41:32,787 --> 01:41:33,087 AND TRY TO FIX WHAT'S GOING ON 1992 01:41:33,087 --> 01:41:33,721 BASED ON SIMULATIONS. 1993 01:41:33,721 --> 01:41:35,456 SO WITH THAT, THANK YOU FOR 1994 01:41:35,456 --> 01:41:45,733 YOUR ATTENTION. 1995 01:42:04,118 --> 01:42:09,724 >> AS A BIOLOGIST AND 1996 01:42:09,724 --> 01:42:13,661 COMPUTATIONAL SCIENTIST -- 1997 01:42:13,661 --> 01:42:23,971 [AUDIO DIGITIZING] 1998 01:42:28,175 --> 01:42:30,044 >> WE HAVE A VIRTUAL 1999 01:42:30,044 --> 01:42:31,212 PRESENTATION AND THAT IS WHAT IS 2000 01:42:31,212 --> 01:42:33,080 BEING SET UP NOW. 2001 01:42:33,080 --> 01:42:34,081 WELCOME BING, VIRTUALLY. 2002 01:42:34,081 --> 01:42:42,590 THANK YOU. 2003 01:42:42,590 --> 01:42:48,295 AS A BIOLOGIST AND COMPUTATIONAL 2004 01:42:48,295 --> 01:42:51,632 SCIENTIST THE NERVOUS SYSTEMS 2005 01:42:51,632 --> 01:42:53,934 GIVES US THE AUTONOMY TO DECIDE 2006 01:42:53,934 --> 01:42:55,736 WHERE TO GO AND THE AGENCY. 2007 01:42:55,736 --> 01:42:56,270 YET NEUROSCIENCE HAS BEEN 2008 01:42:56,270 --> 01:42:57,838 CONSIDERING THE BRAIN IN 2009 01:42:57,838 --> 01:42:59,707 ISOLATION FROM THE BIOMECHANICS 2010 01:42:59,707 --> 01:43:01,409 OF THE BODY THOUGH WE KNOW THE 2011 01:43:01,409 --> 01:43:02,743 BRAIN IS NOT DIRECTLY CONNECTED 2012 01:43:02,743 --> 01:43:03,644 TO THE WORLD. 2013 01:43:03,644 --> 01:43:05,346 IT RECEIVES NO SENSORY INPUT AND 2014 01:43:05,346 --> 01:43:08,082 HAS NO ACTION OUTPUTS EXCEPT 2015 01:43:08,082 --> 01:43:08,282 THROUGH THE BODY. 2016 01:43:08,282 --> 01:43:09,183 SO MY RESEARCH PROGRAM IS 2017 01:43:09,183 --> 01:43:10,918 FOCUSSED ON EMBODIED 2018 01:43:10,918 --> 01:43:12,520 INTELLIGENCE INTEGRATING MODELS 2019 01:43:12,520 --> 01:43:13,988 OF THE BRAIN AND NERVOUS SYSTEM 2020 01:43:13,988 --> 01:43:15,956 WITH THE BODY TO UNDERSTAND THE 2021 01:43:15,956 --> 01:43:17,892 NEURAL BASIS OF NATURAL 2022 01:43:17,892 --> 01:43:18,159 BEHAVIOR. 2023 01:43:18,159 --> 01:43:22,630 IT MAY SEEM THE MOST MYSTERIOUS 2024 01:43:22,630 --> 01:43:23,831 THING ABOUT THE BRAIN IS 2025 01:43:23,831 --> 01:43:24,799 COGNITION AND INTELLIGENCE. 2026 01:43:24,799 --> 01:43:27,034 WHERE MOST TAKE FOR GRANTED IT'S 2027 01:43:27,034 --> 01:43:28,469 RELATIVELY EASY TO WALK UP AND 2028 01:43:28,469 --> 01:43:31,105 DOWN THE STAIRS AND DOING THAT 2029 01:43:31,105 --> 01:43:33,307 CARRYING A LAPTOP AND A CUP OF 2030 01:43:33,307 --> 01:43:35,409 COFFEE, SOMETHING I DO ALL THE 2031 01:43:35,409 --> 01:43:35,609 TIME. 2032 01:43:35,609 --> 01:43:38,179 BUT I CONTEND THE COMPLEXITY OF 2033 01:43:38,179 --> 01:43:42,983 CONTROLLING MY BODY EXCEEDS THE 2034 01:43:42,983 --> 01:43:47,655 COMPLEXITY TO SOLVING CALCULUS 2035 01:43:47,655 --> 01:43:49,957 AND WE NEUROSCIENTISTS STUDY A 2036 01:43:49,957 --> 01:43:57,231 SYSTEM THAT EVOLVED OVER 500 MILLION 2037 01:43:57,231 --> 01:44:01,235 YEARS AGO. SO ANIMALS CAN SENSE 2038 01:44:01,235 --> 01:44:04,538 AND DECIDE WHERE TO GO. TO UNDER- 2039 01:44:04,538 --> 01:44:06,407 STAND ARCHITECTURE, FUNCTION, AND 2040 01:44:06,407 --> 01:44:08,242 DYSFUNCTION OF THE NERVOUS SYSTEM 2041 01:44:08,242 --> 01:44:10,111 THU NEURAL COMPUTATION. ONE MUST 2042 01:44:10,111 --> 01:44:12,613 EMBRACE THAT THE BRAIN AND BODY 2043 01:44:12,613 --> 01:44:15,216 HAVE BEEN INEXTRICABLE IN SOLVING 2044 01:44:15,216 --> 01:44:16,550 THE CHALLENGES FACED BY THE ANIMAL 2045 01:44:16,550 --> 01:44:18,519 FROM MOMENT TO MOMENT AND 2046 01:44:18,519 --> 01:44:19,987 PLANNING FOR THE FUTURE. 2047 01:44:19,987 --> 01:44:23,657 TO UNDERLINE THE POINT WE SEE A 2048 01:44:23,657 --> 01:44:26,994 HUNGRY EAGLE HUNTING INTEGRATING 2049 01:44:26,994 --> 01:44:30,297 VISUAL AND SOMATOSENSORY INPUTS 2050 01:44:30,297 --> 01:44:31,832 PREDICTING THE FUTURE POSITION OF 2051 01:44:31,832 --> 01:44:33,200 THE FISH, COORDINATED MOVEMENT TO 2052 01:44:33,200 --> 01:44:34,935 EXECUTE THIS BANKING MANEUVER AND 2053 01:44:34,935 --> 01:44:37,138 COLLIDE HIS CLAWS WITH THE FISH. 2054 01:44:37,138 --> 01:44:39,473 ALL THIS IS HAPPENING WITH 2055 01:44:39,473 --> 01:44:41,609 SENSORS AND ACTUATORS EMBEDDED 2056 01:44:41,609 --> 01:44:42,209 IN THIS BODY PLAN. 2057 01:44:42,209 --> 01:44:46,947 TO TACKLE THIS INTEGRATION WE'VE 2058 01:44:46,947 --> 01:44:48,916 BUILT BODY MODELS OF THE BRAIN 2059 01:44:48,916 --> 01:44:50,518 WHERE A NERVOUS SYSTEM INTERFACES 2060 01:44:50,518 --> 01:44:55,656 WITH MODELS OF THE BODY AND 2061 01:44:55,656 --> 01:44:59,960 KINEMATICS, DYNAMICS, AND BIO- 2062 01:44:59,960 --> 01:45:07,001 MECHANICS. OUR MODELS ARE AGENT- 2063 01:45:07,001 --> 01:45:11,138 BASED AND CAN PREDICT RESPONSES TO 2064 01:45:11,138 --> 01:45:11,605 MANIPULATIONS NEVER BEFORE SEEN. 2065 01:45:11,605 --> 01:45:13,407 THE LAYERS OF THE BRAIN, BODY 2066 01:45:13,407 --> 01:45:15,075 AND WORLD AND MANY INTERACTIONS 2067 01:45:15,075 --> 01:45:17,144 ARE THE MOST ESSENTIAL LAYERS TO 2068 01:45:17,144 --> 01:45:18,445 MAKE SURE WE'RE NOT COMPROMISING 2069 01:45:18,445 --> 01:45:20,614 ON OUR CORE PRINCIPLES. 2070 01:45:20,614 --> 01:45:23,117 FIRST, FEEDBACK IS ESSENTIAL AND 2071 01:45:23,117 --> 01:45:23,651 UBIQUITOUS. 2072 01:45:23,651 --> 01:45:25,119 WE'RE TALKING ABOUT MORE THAN ME 2073 01:45:25,119 --> 01:45:27,054 PUSHING ON THE TABLE AND TABLE 2074 01:45:27,054 --> 01:45:28,489 PUSHES BACK. 2075 01:45:28,489 --> 01:45:32,193 THE FEEDBACK LOOP GOES BACK TO 2076 01:45:32,193 --> 01:45:34,862 CONNECTION BETWEEN BRAIN REGIONS 2077 01:45:34,862 --> 01:45:41,135 AND SYNAPTIC PLASTICITY. SECOND, 2078 01:45:41,135 --> 01:45:44,071 BODY MATTERS. WE SEE A TROUT THAT'S 2079 01:45:44,071 --> 01:45:46,106 SWIMMING UP STREAM AND WILL GO 2080 01:45:46,106 --> 01:45:48,008 UP TOWARDS THE ROCK. 2081 01:45:48,008 --> 01:45:49,610 IT LOOKS LIKE BEAUTIFUL BEHAVIOR 2082 01:45:49,610 --> 01:45:51,011 BUT THAT TROUT IS ALREADY DEAD. 2083 01:45:51,011 --> 01:45:52,580 THERE'S NO NEURONS AND NO 2084 01:45:52,580 --> 01:45:52,947 CONTROL. 2085 01:45:52,947 --> 01:45:54,748 JUST THE BODY INTERACTING WITH 2086 01:45:54,748 --> 01:45:55,583 THE WATER. 2087 01:45:55,583 --> 01:46:01,188 WE AS NEUROSCIENTISTS CANNOT 2088 01:46:01,188 --> 01:46:04,191 IGNORE THIS. THIRD, LOTS OF PARTS 2089 01:46:04,191 --> 01:46:11,665 CAN BE BROKEN DOWN INTO MODULES 2090 01:46:11,665 --> 01:46:12,499 THAT INTERFACE THRU INFORMATIONAL 2091 01:46:12,499 --> 01:46:14,435 BOTTLENECKS. EXAMPLE INCLUDE 2092 01:46:14,435 --> 01:46:15,736 RETINAL GANGLION CELLS ARE ONLY 2093 01:46:15,736 --> 01:46:19,106 CELL TYPE THAT CONNECT TO VISUAL 2094 01:46:19,106 --> 01:46:19,940 SYSTEM. ENTIRE MOTOR SYSTEM CONVERGES ON A FEW MOTOR NEURONS. 2095 01:46:19,940 --> 01:46:27,581 I EMPHASIZE MODULARITY BECAUSE IT 2096 01:46:27,581 --> 01:46:29,283 POINTS TO A TRACTABLE SOLUTION OF 2097 01:46:29,283 --> 01:46:30,918 MODELLING EMBODIED INTELLIGENCE. 2098 01:46:30,918 --> 01:46:33,220 AND RIGHT KNOW IN 2024 IS A 2099 01:46:33,220 --> 01:46:35,723 SPECIAL TIME IN COMPUTATIONAL 2100 01:46:35,723 --> 01:46:38,525 NEUROSCIENCE BECAUSE I SEE A 2101 01:46:38,525 --> 01:46:40,794 UNIQUE CONVERGENCE OF IDEAS, 2102 01:46:40,794 --> 01:46:42,029 DATA AND TECHNOLOGY AND CULTURE 2103 01:46:42,029 --> 01:46:43,564 ALL COMING TOGETHER TO MOTIVATE 2104 01:46:43,564 --> 01:46:45,866 THE TIMELINESS OF MY APPROACH 2105 01:46:45,866 --> 01:46:47,501 WHILE UNDERSTANDING NEURAL 2106 01:46:47,501 --> 01:46:49,069 COMPUTATION THROUGH BUILDING 2107 01:46:49,069 --> 01:46:50,204 EMBODIED SYSTEM OF SYSTEMS MODEL 2108 01:46:50,204 --> 01:46:52,072 CAPABLE OF MIMICKING THE NATURAL 2109 01:46:52,072 --> 01:47:02,616 BEHAVIOR OF ANIMALS IN A VIRTUAL 2110 01:47:03,717 --> 01:47:06,620 FACSIMILE OF REALITY. AND LOOKING AT 2111 01:47:06,620 --> 01:47:07,755 EXTENSIONS ALL THESE ARE READY 2112 01:47:07,755 --> 01:47:09,957 TO BE INTEGRATED IN A SIMULATION 2113 01:47:09,957 --> 01:47:11,659 FRAMEWORK GROUNDED IN REAL DATA 2114 01:47:11,659 --> 01:47:14,995 WITH MODULES THAT WILL LEARN FROM 2115 01:47:14,995 --> 01:47:16,330 DATA SUCH AS THE MANY SENSORY 2116 01:47:16,330 --> 01:47:21,335 CAPABILITY OF ANIMALS, MEMORY, 2117 01:47:21,335 --> 01:47:23,404 AROUSAL STATE, MOTOR PLANNING, 2118 01:47:23,404 --> 01:47:25,706 EXECUTION AND MUSCLES AND 2119 01:47:25,706 --> 01:47:26,140 SKELETONS. TO MAKE AN ANALOGY, 2120 01:47:26,140 --> 01:47:29,977 THE MODELS ARE FOR ANIMALS 2121 01:47:29,977 --> 01:47:32,513 CALLED DIGITAL TWIN MODELS IN 2122 01:47:32,513 --> 01:47:32,813 ENGINEERING. 2123 01:47:32,813 --> 01:47:39,019 IT'S A SYSTEM OF SYSTEMS 2124 01:47:39,019 --> 01:47:42,156 STIMULATION WITH MANY PARTS. 2125 01:47:42,156 --> 01:47:43,257 NOW VIRTUAL COMPUTATIONAL 2126 01:47:43,257 --> 01:47:44,058 SIMULATION IS CRITICAL. 2127 01:47:44,058 --> 01:47:47,661 BY LOOKING AT A WIRING DIAGRAM 2128 01:47:47,661 --> 01:47:48,963 OF AN AIRPLANE ONLY MAKES US 2129 01:47:48,963 --> 01:47:51,332 MORE CONFUSED HOW A PLANE WORKS. 2130 01:47:51,332 --> 01:47:52,633 TO PREDICT ANYTHING WE MUST KNOW 2131 01:47:52,633 --> 01:47:54,568 ABOUT THE BODY TOO. 2132 01:47:54,568 --> 01:47:56,370 WHICH WIRES GOES TO THE FUEL 2133 01:47:56,370 --> 01:47:59,373 PUMPS AND WHICH GOES TO THE TAIL 2134 01:47:59,373 --> 01:47:59,673 FLAPS? 2135 01:47:59,673 --> 01:48:01,375 >> IT'S THE FEEDBACK 2136 01:48:01,375 --> 01:48:02,343 INTERACTIONS THAT ACCOMPLISH FIGHT. 2137 01:48:02,343 --> 01:48:04,745 IN NEUROSCIENCE, WE ARE WEALTHY 2138 01:48:04,745 --> 01:48:05,813 IN DATA. 2139 01:48:05,813 --> 01:48:07,581 YET THIS WEALTH THREATENS TO 2140 01:48:07,581 --> 01:48:09,550 DRIVE US FURTHER APART BECAUSE 2141 01:48:09,550 --> 01:48:11,418 OUR FAVORITE TOOLS AND PREFERRED 2142 01:48:11,418 --> 01:48:12,586 ANALYSES ARE NOT READILY 2143 01:48:12,586 --> 01:48:14,755 TRANSFERABLE BETWEEN THE DATA 2144 01:48:14,755 --> 01:48:15,656 SETS. 2145 01:48:15,656 --> 01:48:17,691 FOR THIS REASON I SEE A PRESSING 2146 01:48:17,691 --> 01:48:19,493 NEED FOR EMBODIED MODELS TO BE 2147 01:48:19,493 --> 01:48:20,527 DEVELOPED AS COMPUTATIONAL FRAME- 2148 01:48:20,527 --> 01:48:23,497 WORKS AND INTEGRATING ALL THIS 2149 01:48:23,497 --> 01:48:25,065 DATA TO MAKE QUANTITATIVE 2150 01:48:25,065 --> 01:48:27,634 PREDICTIONS AND GUIDE THE DESIGN 2151 01:48:27,634 --> 01:48:38,145 OF INCREASINGLY EXPENSIVE AND 2152 01:48:40,481 --> 01:48:40,914 COMBINATORIALLY COMPLEX 2153 01:48:40,914 --> 01:48:43,417 EXPERIMENTS. WHEN PREDICTIONS 2154 01:48:43,417 --> 01:48:44,618 DEVIATE FROM EXPERIMENTS, WE WILL 2155 01:48:44,618 --> 01:48:46,086 REFINE OUR MODELS. WE NEVER 2156 01:48:46,086 --> 01:48:46,387 COMPROMISE OUR CORE PRINCIPLES. 2157 01:48:46,387 --> 01:48:49,656 THE BRAIN AND BODY CANNOT BE 2158 01:48:49,656 --> 01:48:51,392 UNDERSTAND IN ISOLATION FROM 2159 01:48:51,392 --> 01:48:53,560 EACH OTHER AND THE GOAL OF THE 2160 01:48:53,560 --> 01:48:54,995 OVERALL AGENT IS ALWAYS THE 2161 01:48:54,995 --> 01:48:56,964 COORDINATED MOVEMENT OF ITS BODY 2162 01:48:56,964 --> 01:48:58,799 TO TACKLE THE CHALLENGES OF 2163 01:48:58,799 --> 01:48:59,433 SURVIVAL. 2164 01:48:59,433 --> 01:49:01,235 FINALLY, THIS IS THE ONLY WAY 2165 01:49:01,235 --> 01:49:04,772 WE'LL BE ABLE TO GAIN HOLISTIC 2166 01:49:04,772 --> 01:49:08,609 UNDERSTANDING OF THE EMBODIED 2167 01:49:08,609 --> 01:49:09,576 INTELLIGENCE OF BRAIN AND BEHAVIOR 2168 01:49:09,576 --> 01:49:10,244 MODELS ARE GENERALIZABLE BECAUSE 2169 01:49:10,244 --> 01:49:11,945 THEY'RE MODULAR AND THE MODULE 2170 01:49:11,945 --> 01:49:14,114 REFLECTS THE PHYSICS OF THE BODY. 2171 01:49:14,114 --> 01:49:15,516 I'M EXCITED ABOUT WHERE WE'RE 2172 01:49:15,516 --> 01:49:16,884 GOING WITH THIS BECAUSE THIS IS 2173 01:49:16,884 --> 01:49:20,521 MY PROPOSED MODELLING APPROACH 2174 01:49:20,521 --> 01:49:22,956 TO TACKLE MORE INTERESTING 2175 01:49:22,956 --> 01:49:23,524 NATURALISTIC BEHAVIORS AND 2176 01:49:23,524 --> 01:49:24,958 COMPLEXITY OF NATURAL BEHAVIORS 2177 01:49:24,958 --> 01:49:27,161 OF ANIMALS INCLUDING HUMANS AS 2178 01:49:27,161 --> 01:49:29,329 WE DO IN THE REAL WORLD. 2179 01:49:29,329 --> 01:49:31,098 WE'LL DO THIS BY EXTENDING THE 2180 01:49:31,098 --> 01:49:32,800 MODEL WITH MORE MODULES, 2181 01:49:32,800 --> 01:49:34,301 IMPROVING THE FIDELITY OF THE 2182 01:49:34,301 --> 01:49:35,235 MODULES AND WORK WITH 2183 01:49:35,235 --> 01:49:36,670 COLLABORATORS TO INTEGRATE ON 2184 01:49:36,670 --> 01:49:38,906 MAKING TESTABLE PREDICTIONS AND 2185 01:49:38,906 --> 01:49:40,941 REFINE THE UNDERSTANDING OF HOW 2186 01:49:40,941 --> 01:49:42,109 THE COMPLEX SYSTEMS MAY BE 2187 01:49:42,109 --> 01:49:43,877 WORKING TOGETHER. 2188 01:49:43,877 --> 01:49:46,447 NOW, FOR REAL WELL STUDIED MODEL 2189 01:49:46,447 --> 01:49:48,749 ORGANISMS LIKE FLIES AND 2190 01:49:48,749 --> 01:49:49,450 RODENTS, THE NEXT COUPLE YEARS 2191 01:49:49,450 --> 01:49:51,084 WILL BE EXCITING BECAUSE OF ALL 2192 01:49:51,084 --> 01:49:52,986 THE COMPREHENSIVE DATA SETS 2193 01:49:52,986 --> 01:49:54,855 PUBLISHED OR WILL BE PUBLISHED 2194 01:49:54,855 --> 01:49:55,088 SOON. 2195 01:49:55,088 --> 01:49:56,957 AND FOR EACH MODEL ORGANISM THE 2196 01:49:56,957 --> 01:49:58,859 WORK WE'RE DOING LEANS ON THE 2197 01:49:58,859 --> 01:50:00,594 STRENGTH OF THE AVAILABLE DATA 2198 01:50:00,594 --> 01:50:01,662 SETS. AND BIOPHYSICAL KNOWLEDGE 2199 01:50:01,662 --> 01:50:03,230 IS AVAILABLE: THE FLY CONNECTOME 2200 01:50:03,230 --> 01:50:04,731 OF THE BRAIN AND VENTRAL NERVOUS 2201 01:50:04,731 --> 01:50:05,732 SYSTEM AND WE'LL INCORPORATE 2202 01:50:05,732 --> 01:50:09,369 THAT INTO OUR DYNAMIC MODELS. 2203 01:50:09,369 --> 01:50:11,038 WHERE NEEDED, WE'LL APPROXIMATE 2204 01:50:11,038 --> 01:50:13,407 WITH MACHINE LEARNING TO MATCH THE 2205 01:50:13,407 --> 01:50:15,576 INPUTS AND OUTPUTS OF THOSE 2206 01:50:15,576 --> 01:50:15,809 MODULES. FOR THE RODENT AND HUMAN 2207 01:50:15,809 --> 01:50:20,180 WE'LL MODEL AT THE MESO SCALE BUT 2208 01:50:20,180 --> 01:50:26,019 GROUND MODELS IN THE BEHAVIOR OF 2209 01:50:26,019 --> 01:50:26,887 ANIMALS TRACKED BY COMPUTER VISION 2210 01:50:26,887 --> 01:50:28,555 FURTHERMORE I ENVISION THE 2211 01:50:28,555 --> 01:50:30,958 FRAMEWORK AS A WAY TO ASK 2212 01:50:30,958 --> 01:50:32,259 QUESTIONS TO CONNECT SHORT TERM 2213 01:50:32,259 --> 01:50:33,861 BIOPHYSICAL MECHANISMS WITH 2214 01:50:33,861 --> 01:50:36,897 LARGER AND LONGER TERM SCALE 2215 01:50:36,897 --> 01:50:37,498 IDEAS. 2216 01:50:37,498 --> 01:50:39,333 FOR EXAMPLE, THE NERVOUS SYSTEM 2217 01:50:39,333 --> 01:50:40,400 MODIFIES ITSELF THROUGH PLASTICITY 2218 01:50:40,400 --> 01:50:41,134 AND LIFELONG LEARNING. 2219 01:50:41,134 --> 01:50:43,337 THE BODY CHANGES DURING 2220 01:50:43,337 --> 01:50:45,339 DEVELOPMENT AND AGING AS WELL AS 2221 01:50:45,339 --> 01:50:49,743 ACTIVITY DEPENDENT CHANGES WITH 2222 01:50:49,743 --> 01:50:51,111 EXERCISE OR THE CHANGING 2223 01:50:51,111 --> 01:50:52,379 ENVIRONMENT, SOCIETY AND 2224 01:50:52,379 --> 01:50:52,613 CLIMATE. 2225 01:50:52,613 --> 01:50:55,015 TO SUMMARIZE, THIS ENTIRE 2226 01:50:55,015 --> 01:50:56,950 RESEARCH PLAN IS BUILT ON THE 2227 01:50:56,950 --> 01:51:00,120 IDEA BRAINS AND BODIES CANNOT BE 2228 01:51:00,120 --> 01:51:02,055 UNDERSTOOD IN ISOLATION AND 2229 01:51:02,055 --> 01:51:02,789 HOLISTIC MODELS MUST SERVE THE 2230 01:51:02,789 --> 01:51:05,425 GOALS OF THE ANIMAL. 2231 01:51:05,425 --> 01:51:06,426 I FEEL INCREDIBLY FORTUNATE TO 2232 01:51:06,426 --> 01:51:07,928 BE A MEMBER OF THIS COMMUNITY 2233 01:51:07,928 --> 01:51:10,097 RIGHT NOW AND OPTIMISTIC ABOUT THE 2234 01:51:10,097 --> 01:51:11,598 SIGNIFICANCE OF WHAT WE'LL BE 2235 01:51:11,598 --> 01:51:12,933 BUILDING TOGETHER. 2236 01:51:12,933 --> 01:51:15,269 I THINK THE POTENTIAL IMPACT OF 2237 01:51:15,269 --> 01:51:18,071 MY PROPOSED WORK AND AS I'VE 2238 01:51:18,071 --> 01:51:19,106 DESCRIBED TODAY IS NOTHING SHORT 2239 01:51:19,106 --> 01:51:20,741 OF CHANGING THE WAY 2240 01:51:20,741 --> 01:51:21,942 NEUROSCIENCE IS DONE IN THE NEXT 2241 01:51:21,942 --> 01:51:23,177 YEARS TO DECADES. 2242 01:51:23,177 --> 01:51:26,313 I WOULD LIKE TO GIVE MY SINCERE 2243 01:51:26,313 --> 01:51:28,815 THANKS TO EVERYONE IN MY LAB, 2244 01:51:28,815 --> 01:51:32,886 PAST AND PRESENT 2245 01:51:32,886 --> 01:51:33,687 AND ALL OF OUR 2246 01:51:33,687 --> 01:51:34,154 COLLABORATORS. 2247 01:51:34,154 --> 01:51:35,088 I'M GRATEFUL TO YOU AND THANK 2248 01:51:35,088 --> 01:51:35,923 YOU FOR TRUSTING ME ENOUGH TO 2249 01:51:35,923 --> 01:51:46,166 WORK WITH ME. 2250 01:52:15,729 --> 01:52:17,497 >> I'M DOMINIQUE DUNCAN AND 2251 01:52:17,497 --> 01:52:18,899 WANTED TO TALK ABOUT NEUROAI 2252 01:52:18,899 --> 01:52:20,534 FROM A DATA SHARING PERSPECTIVE 2253 01:52:20,534 --> 01:52:21,735 AND HOW IMPORTANT IT IS TO SHARE 2254 01:52:21,735 --> 01:52:24,571 DATA AND THE CONSIDERATIONS 2255 01:52:24,571 --> 01:52:25,205 AROUND THAT AND CHALLENGES AND 2256 01:52:25,205 --> 01:52:30,410 OPPORTUNITIES. 2257 01:52:30,410 --> 01:52:35,649 THE BRAIN INITIATIVE FUNDED DATA 2258 01:52:35,649 --> 01:52:43,223 ARCHIVES AND SOME ARE OPENNEURO, 2259 01:52:43,223 --> 01:52:46,026 NEMAR, NEMO, BRAIN IMAGE LIBRARY, 2260 01:52:46,026 --> 01:52:46,226 DANDI AND DABI. I'VE BEEN WORKING 2261 01:52:46,226 --> 01:52:47,628 ON DABI. I'LL GO OVER THAT 2262 01:52:47,628 --> 01:52:57,904 INFRASTRUCTURE IN A LITTLE BIT. 2263 01:53:01,041 --> 01:53:02,643 MORE INFORMATION ABOUT THE DATA 2264 01:53:02,643 --> 01:53:04,745 STORED IN THE DIFFERENT DATA 2265 01:53:04,745 --> 01:53:05,012 ARCHIVES. 2266 01:53:05,012 --> 01:53:07,247 THIS IS A LIST OF THOSE BUT SOME 2267 01:53:07,247 --> 01:53:11,184 DATA ARCHIVES HAVE ALSO EXPANDED 2268 01:53:11,184 --> 01:53:12,486 INTO OTHER DATA TYPES. 2269 01:53:12,486 --> 01:53:14,388 SO THERE'S MANY CHALLENGES IN 2270 01:53:14,388 --> 01:53:16,623 LARGE SCALE DATA REPOSITORIES. 2271 01:53:16,623 --> 01:53:19,660 THESE ARE SOME OF THEM. 2272 01:53:19,660 --> 01:53:22,295 SO DATA ARE HETEROGENEOUS AND 2273 01:53:22,295 --> 01:53:23,497 CAN BE DIFFICULT TO COMPARE 2274 01:53:23,497 --> 01:53:24,498 ACROSS SITES. 2275 01:53:24,498 --> 01:53:26,400 WE'RE DEALING WITH LARGE FILE 2276 01:53:26,400 --> 01:53:26,600 SIZES. 2277 01:53:26,600 --> 01:53:29,403 SOME OF THE DATA ARE HIGHLY 2278 01:53:29,403 --> 01:53:30,037 SPECIFIC. 2279 01:53:30,037 --> 01:53:35,075 PEOPLE USE DIFFERENT FORMATS. 2280 01:53:35,075 --> 01:53:37,010 THERE CAN BE PRIVACY CONSTRAINTS 2281 01:53:37,010 --> 01:53:39,579 AND NEED TO SECURE RESOURCE TO 2282 01:53:39,579 --> 01:53:41,048 SUSTAIN THE ARCHIVE EFFORT 2283 01:53:41,048 --> 01:53:42,716 BEYOND THE INITIAL FUNDING 2284 01:53:42,716 --> 01:53:43,950 PERIOD IN CASE FUNDING ENDS OR 2285 01:53:43,950 --> 01:53:45,886 THE PROJECT IS NOT RENEWED. 2286 01:53:45,886 --> 01:53:47,854 SO HERE ON THE LEFT I WANTED TO 2287 01:53:47,854 --> 01:53:50,457 BRING UP THE APPROACH WE TAKE 2288 01:53:50,457 --> 01:53:53,293 FOR DABI IN TERMS OF DATA 2289 01:53:53,293 --> 01:53:53,560 STORAGE MODELS. 2290 01:53:53,560 --> 01:53:56,063 WE'RE AGNOSTIC TO THE WAY THE 2291 01:53:56,063 --> 01:53:57,030 DATA ARE STORED. 2292 01:53:57,030 --> 01:54:00,067 SO WE GIVE FLEXIBILITY TO OUR 2293 01:54:00,067 --> 01:54:02,302 DATA PROVIDERS. 2294 01:54:02,302 --> 01:54:04,504 AND ONE OPTION IS CENTRALIZED 2295 01:54:04,504 --> 01:54:05,038 DATA STORAGE MODEL. 2296 01:54:05,038 --> 01:54:07,507 IN THAT CASE THE DATA ARE STORED 2297 01:54:07,507 --> 01:54:11,645 AT OUR SITE AT USC. THE OTHER DATA 2298 01:54:11,645 --> 01:54:13,080 STORAGE MODEL IS FEDERATED. 2299 01:54:13,080 --> 01:54:15,449 THAT MEANS WE HAVE THE METADATA 2300 01:54:15,449 --> 01:54:18,552 AND RAW DATA AT THE SITE WHERE 2301 01:54:18,552 --> 01:54:20,020 THEY'RE COLLECTED OR CLOUD 2302 01:54:20,020 --> 01:54:20,587 BASED. 2303 01:54:20,587 --> 01:54:22,656 CURRENTLY WE'RE MOVING OUR DATA 2304 01:54:22,656 --> 01:54:30,864 TO AWS. THE NEXT FEW SLIDES 2305 01:54:30,864 --> 01:54:32,733 I WANTED TO TALK ABOUT THE 2306 01:54:32,733 --> 01:54:38,972 INFRASTRUCTURE. DABI IS FOR 2307 01:54:38,972 --> 01:54:43,343 INVASIVE HUMAN NEUROPHYSIOLOGY. 2308 01:54:43,343 --> 01:54:45,512 WE GIVE FLEXIBILITY TO OUR DATA 2309 01:54:45,512 --> 01:54:47,948 PROVIDERS IN TERMS OF STANDARDS 2310 01:54:47,948 --> 01:54:48,148 USED. 2311 01:54:48,148 --> 01:54:51,651 WE ENCOURAGE DATA PROVIDERS TO 2312 01:54:51,651 --> 01:54:54,788 USE EITHER BIDS OR NEURODATA 2313 01:54:54,788 --> 01:54:57,791 WITHOUT BORDERS OR HAVE OUR OWN 2314 01:54:57,791 --> 01:54:59,493 DABI FRAMEWORK THE DATA 2315 01:54:59,493 --> 01:55:00,961 PROVIDERS CAN USE. WE WORK CLOSELY 2316 01:55:00,961 --> 01:55:04,131 WITH THE NEUROSURGEON DATA 2317 01:55:04,131 --> 01:55:05,866 PROVIDERS TO FIGURE OUT WHAT 2318 01:55:05,866 --> 01:55:07,400 THEIR NEEDS ARE, WHAT WOULD MAKE 2319 01:55:07,400 --> 01:55:13,039 THEIR LIVES EASIER IN TERMS OF 2320 01:55:13,039 --> 01:55:21,615 SHARING DATA. FOR ANALYSIS WE CAN 2321 01:55:21,615 --> 01:55:24,518 HELP THEM DEVELOP ANALYTIC TOOLS 2322 01:55:24,518 --> 01:55:25,085 INCLUDING VISUALIZATION. 2323 01:55:25,085 --> 01:55:27,654 WE HAVE MOSTLY EEG. WE HAVE ECOG 2324 01:55:27,654 --> 01:55:33,493 AND DBS AND OTHER DATA TYPES, 2325 01:55:33,493 --> 01:55:39,399 CLINICAL, IMAGING, PATHOLOGY, 2326 01:55:39,399 --> 01:55:43,637 DEMOGRAPHIC AND BEHAVIORIAL. 2327 01:55:43,637 --> 01:55:52,479 USE JSON TO HARMONIZE METADATA 2328 01:55:52,479 --> 01:55:59,986 JSON WAS CHOSEN BY ITS USE OF BIDS 2329 01:55:59,986 --> 01:56:02,155 WE'VE WORKED CLOSELY WITH THE DATA 2330 01:56:02,155 --> 01:56:07,294 PROVIDERS TO FIGURE OUT WHAT 2331 01:56:07,294 --> 01:56:17,838 WOULD MAKE THEIR LIVES EASY. IF 2332 01:56:22,209 --> 01:56:32,686 THERE'S REQUESTS FOR ACCESS AND 2333 01:56:34,921 --> 01:56:36,423 ADJUSTMENTS TO SHARE THE DATA. 2334 01:56:36,423 --> 01:56:39,793 IF YOU GET ON DABI ANYONE CAN 2335 01:56:39,793 --> 01:56:40,560 CREATE A FREE ACCOUNT. 2336 01:56:40,560 --> 01:56:42,262 IF YOU WANT TO BUILD A COHORT 2337 01:56:42,262 --> 01:56:45,131 YOU CAN QUERY THE DATA WE HAVE 2338 01:56:45,131 --> 01:56:48,869 ON DABI AND LOOK AT CRITERIA YOU 2339 01:56:48,869 --> 01:56:50,637 MIGHT BE INTERESTED IN. 2340 01:56:50,637 --> 01:56:52,205 SELECT THE CRITERIA IF THEY'RE 2341 01:56:52,205 --> 01:56:54,407 DATA MODALITIES, BRAIN REGIONS, 2342 01:56:54,407 --> 01:57:04,751 DIAGNOSIS, ETCETERA. 2343 01:57:10,156 --> 01:57:12,759 AND YOU CAN LOOK AT THE IMAGING 2344 01:57:12,759 --> 01:57:14,494 AND THE SIGNAL AND DECIDE IF YOU 2345 01:57:14,494 --> 01:57:17,197 WANT TO DISCARD ANY OF THE 2346 01:57:17,197 --> 01:57:19,266 SUBJECTS FROM YOUR COHORT. 2347 01:57:19,266 --> 01:57:27,340 SO ONCE YOU DO THAT, THEN YOU 2348 01:57:27,340 --> 01:57:28,909 CAN SAVE YOUR COHORT AND MOVE ON 2349 01:57:28,909 --> 01:57:31,645 TO THE NEXT STEP OF DOING SOME 2350 01:57:31,645 --> 01:57:41,288 ANALYSIS. 2351 01:57:41,288 --> 01:57:49,696 AND WE HAVE A LIBRARY OF TOOLS 2352 01:57:49,696 --> 01:57:50,797 AND PROCESSING TOOLS AND YOU 2353 01:57:50,797 --> 01:57:52,799 CAN SELECT THE ONES YOU WANT TO 2354 01:57:52,799 --> 01:57:59,639 USE FOR YOUR COHORT AND PULL OUT 2355 01:57:59,639 --> 01:58:01,608 THE MODULES AND CONNECT THEM AND 2356 01:58:01,608 --> 01:58:04,344 RUN YOUR COHORT THROUGH THAT AND 2357 01:58:04,344 --> 01:58:06,680 AT EACH STEP CAN DO A QUICK 2358 01:58:06,680 --> 01:58:08,481 VISUAL INSPECTION. 2359 01:58:08,481 --> 01:58:10,583 SAY YOU DO FILTERING AND DO A 2360 01:58:10,583 --> 01:58:12,252 QUICK INSPECTION TO SEE HOW THAT 2361 01:58:12,252 --> 01:58:22,595 CHANGED THE SIGNAL. 2362 01:58:26,433 --> 01:58:33,573 ONE OF THE TOOLS WAS IEEG SYNC. 2363 01:58:33,573 --> 01:58:33,773 IT'S LIKE GOOGLE DOCS FOR IEEG 2364 01:58:33,773 --> 01:58:37,544 WHERE YOU CAN COLLABORATIVELY 2365 01:58:37,544 --> 01:58:41,848 MARK UP DATA AND MARK UP ANY 2366 01:58:41,848 --> 01:58:43,950 EVENTS OR RANGES OF INTEREST AND 2367 01:58:43,950 --> 01:58:54,394 PERFORM FURTHER ANALYSIS. 2368 01:58:57,430 --> 01:58:59,599 WE CAN INTEGRATE THE VARIOUS 2369 01:58:59,599 --> 01:59:01,668 DATA ARCHIVES AND ALLOW USERS TO 2370 01:59:01,668 --> 01:59:04,604 QUERY ACROSS THE ARCHIVES AND 2371 01:59:04,604 --> 01:59:08,441 LEVERAGE THESE RICH DATA SETS 2372 01:59:08,441 --> 01:59:10,176 AND BE ABLE TO PERFORM A.I. 2373 01:59:10,176 --> 01:59:11,077 METHODS AND ANALYSIS ON THE 2374 01:59:11,077 --> 01:59:13,947 DIFFERENT DATA SETS. 2375 01:59:13,947 --> 01:59:16,850 SO BRAIN INITIATIVE AS YOU HEARD 2376 01:59:16,850 --> 01:59:20,420 FROM MANY PEOPLE ENCOURAGED 2377 01:59:20,420 --> 01:59:23,656 MULTIDISCIPLINARY TEAMS AND MY 2378 01:59:23,656 --> 01:59:25,458 BACKGROUND IS ELECTRICAL 2379 01:59:25,458 --> 01:59:26,860 ENGINEERING AND MATH. I WORK WITH 2380 01:59:26,860 --> 01:59:29,029 NEUROSURGEONS, CLINICIAN, PEOPLE 2381 01:59:29,029 --> 01:59:34,000 WITH VARIOUS TYPES OF EXPERTISE. 2382 01:59:34,000 --> 01:59:37,203 WE NEED STANDARDIZED CROSS MODAL DATA INTEGRATION, 2383 01:59:37,203 --> 01:59:39,672 COMMON THEORETICAL FRAMEWORKS, 2384 01:59:39,672 --> 01:59:42,542 METRICS TO COMPARE BIOLOGICAL AND 2385 01:59:42,542 --> 01:59:44,477 ARTIFICIAL NEURAL NETWORKS AND 2386 01:59:44,477 --> 01:59:46,413 IF NEW ARCHIVES ARE NEEDED TO 2387 01:59:46,413 --> 01:59:46,980 INCORPORATE ADDITIONAL DATA 2388 01:59:46,980 --> 01:59:57,090 TYPES. WE ALSO LINK TO BRAIN TOOLS 2389 01:59:58,224 --> 02:00:04,397 DABI HAS INTEGRATED RAVE AND A 2390 02:00:04,397 --> 02:00:08,134 TOOL FOR VISUALIZATION OF IEEG 2391 02:00:08,134 --> 02:00:09,936 I WANTED TO END WITH THE 2392 02:00:09,936 --> 02:00:16,242 NEWEST BRAIN DATA ARCHIVE FOR 2393 02:00:16,242 --> 02:00:18,478 THE BBQS PROGRAM. IT LEVERAGES 2394 02:00:18,478 --> 02:00:22,849 EXISTING DATA ARCHIVES SUCH AS 2395 02:00:22,849 --> 02:00:28,121 DABI, DANDI, AND STANDARDS LIKE 2396 02:00:28,121 --> 02:00:30,557 BIDS AND NWB. IT'LL USE A 2397 02:00:30,557 --> 02:00:34,727 CLOUD-BASED ENVIRONMENT SO 2398 02:00:34,727 --> 02:00:41,668 PEOPLE CAN PERFORM A.I. ANALYSIS. 2399 02:00:41,668 --> 02:00:42,936 SO THANK YOU. 2400 02:00:42,936 --> 02:00:44,337 AND THESE ARE SOME OF MY LAB 2401 02:00:44,337 --> 02:00:46,639 MEMBERS AND WE'RE GRATEFUL FOR 2402 02:00:46,639 --> 02:00:48,241 THE FUNDING FROM NIH AND NSF. 2403 02:00:48,241 --> 02:00:57,450 I'LL END THERE. 2404 02:00:57,450 --> 02:00:59,052 >> THANK YOU TO ALL OF OUR 2405 02:00:59,052 --> 02:00:59,319 SPEAKERS. 2406 02:00:59,319 --> 02:01:03,556 TONY'S UP. 2407 02:01:03,556 --> 02:01:05,391 AT THIS POINT I WANTED TO INVITE 2408 02:01:05,391 --> 02:01:07,127 ALL OUR INVITED DISCUSSANTS TO 2409 02:01:07,127 --> 02:01:10,497 THE STAGE. FRANCES CHANCE, SUEYEON 2410 02:01:10,497 --> 02:01:12,866 CHUNG AND PAUL MIDDLEBROOKS IF 2411 02:01:12,866 --> 02:01:23,009 YOU CAN JOIN US ON THE STAGE. 2412 02:01:23,009 --> 02:01:24,410 WE SHOULD HAVE THE GUIDING 2413 02:01:24,410 --> 02:01:34,854 QUESTION SLIDE DISPLAYED. 2414 02:01:48,034 --> 02:01:51,704 YOU CAN SEE THE GOAL FOR THE 2415 02:01:51,704 --> 02:01:54,240 SESSION WITH THE TALKS WE'VE 2416 02:01:54,240 --> 02:01:58,378 HEARD IS TO HELP ESTABLISH A 2417 02:01:58,378 --> 02:02:00,747 VISION FOR NEUROAI, A CONVERGENT 2418 02:02:00,747 --> 02:02:01,681 FRAMEWORK TO DRIVE RECIPROCAL 2419 02:02:01,681 --> 02:02:02,415 ADVANCES THRU SHARED PRINCIPLES 2420 02:02:02,415 --> 02:02:07,554 AND WE NEED TO IDENTIFY CONCRETE 2421 02:02:07,554 --> 02:02:09,722 OPPORTUNITIES FOR BRAIN TO BRIDGE 2422 02:02:09,722 --> 02:02:10,823 BIOLOGICAL AND ARTIFICIAL 2423 02:02:10,823 --> 02:02:14,928 SYSTEMS AND WE HEARD GREAT 2424 02:02:14,928 --> 02:02:16,129 PERSPECTIVES ABOUT THAT AND THE 2425 02:02:16,129 --> 02:02:18,898 AUDIENCE MEMBERS CAN SEE THE 2426 02:02:18,898 --> 02:02:20,200 QUESTIONS WE'VE PUT TOGETHER 2427 02:02:20,200 --> 02:02:23,036 BEFORE THE WORKSHOP TO GUIDE 2428 02:02:23,036 --> 02:02:25,438 YOUR THINKING BUT WE WANT TO 2429 02:02:25,438 --> 02:02:31,511 START OUR PANEL DISCUSSION WITH 2430 02:02:31,511 --> 02:02:41,654 FRANCES. 2431 02:02:49,796 --> 02:02:50,630 >> 2432 02:02:50,630 --> 02:02:52,131 >> ONE THING I APPRECIATED ABOUT 2433 02:02:52,131 --> 02:02:53,533 THE SESSION IS WE'RE SEEING AN 2434 02:02:53,533 --> 02:02:55,335 APPRECIATION OF THE BIODIVERSITY 2435 02:02:55,335 --> 02:02:56,903 THAT EXISTS IN NEURAL SYSTEMS 2436 02:02:56,903 --> 02:02:59,639 FROM THE PEOPLE COMING FROM MORE 2437 02:02:59,639 --> 02:03:03,676 THE A.I. CAMP. AND NEUROSCIENCE 2438 02:03:03,676 --> 02:03:13,086 CAMP GETTING DIVERSITY IN DATASETS 2439 02:03:13,086 --> 02:03:14,554 INTO A.I. MODELS AND FACILITATING 2440 02:03:14,554 --> 02:03:14,921 THAT. 2441 02:03:14,921 --> 02:03:18,024 THE QUESTION I'D LIKE TO ASK IS 2442 02:03:18,024 --> 02:03:20,860 HOW DO YOU CHOOSE YOUR 2443 02:03:20,860 --> 02:03:22,795 CONSTRAINTS? 2444 02:03:22,795 --> 02:03:25,198 ON ONE HAND IF YOU'RE JUST -- 2445 02:03:25,198 --> 02:03:26,766 THAT'S HOW I'LL LEAVE IT. 2446 02:03:26,766 --> 02:03:31,337 MAYBE IF SOMEONE FROM THE A.I. 2447 02:03:31,337 --> 02:03:33,039 SIDE, ALI, WOLFGANG FOR PATRICK, 2448 02:03:33,039 --> 02:03:34,374 HOW DO YOU CHOOSE WHAT YOU TRY 2449 02:03:34,374 --> 02:03:36,876 TO REPLICATE FROM A BRAIN? 2450 02:03:36,876 --> 02:03:41,314 OTHERWISE YOU'RE JUST TRYING TO 2451 02:03:41,314 --> 02:03:45,051 RECREATE A BRAIN IN A SIMULATION 2452 02:03:45,051 --> 02:03:49,455 AND IF THERE'S TIME WE CAN MOVE 2453 02:03:49,455 --> 02:03:55,361 TO ANDREAS, ANTONE, DOMINIQUE ON 2454 02:03:55,361 --> 02:03:56,329 HOW YOU'LL FACILITATE A.I. LOOKING 2455 02:03:56,329 --> 02:03:57,930 TO THE LARGE DATA SETS FOR 2456 02:03:57,930 --> 02:04:00,933 SPECIFIC ASPECTS OF HOW 2457 02:04:00,933 --> 02:04:01,534 BIOLOGICAL NEURONS ARE DOING 2458 02:04:01,534 --> 02:04:11,744 PROCESSING. 2459 02:04:13,513 --> 02:04:19,652 >> I RECENTLY UNEARTH THE 2460 02:04:19,652 --> 02:04:24,090 QUOTE IT'S A SHORT STORY ABOUT 2461 02:04:24,090 --> 02:04:25,391 A KINGDOM MAP THE SAME SIZE AND 2462 02:04:25,391 --> 02:04:27,293 RESOLUTION AS THE KINGDOM ITSELF 2463 02:04:27,293 --> 02:04:31,130 AND IT TALKS ABOUT HOW IN THE 2464 02:04:31,130 --> 02:04:33,766 DESERT THEY HAVE A MAP THAT'S 2465 02:04:33,766 --> 02:04:38,438 NOW CRUMBLING BECAUSE IT'S ISO 2466 02:04:38,438 --> 02:04:40,640 MORPHIC AND I THINK THAT'S MADE 2467 02:04:40,640 --> 02:04:43,743 SOME REALLY INTERESTING POINTS 2468 02:04:43,743 --> 02:04:46,312 ABOUT THE FACT THAT WE NOW HAVE 2469 02:04:46,312 --> 02:04:47,714 3457Z FROM THE REAL WORLD AT THE 2470 02:04:47,714 --> 02:04:48,681 SAME RESOLUTION AS THE REAL 2471 02:04:48,681 --> 02:04:51,651 THING AND IT'S HELPFUL. 2472 02:04:51,651 --> 02:04:53,152 SO IS THERE SCALE SEPARATION IN 2473 02:04:53,152 --> 02:04:55,021 THE BRAIN? 2474 02:04:55,021 --> 02:04:56,255 WE CAN SAY ANYTHING BELOW THE 2475 02:04:56,255 --> 02:04:57,957 SCALE DOESN'T MATTER OR THERE'S 2476 02:04:57,957 --> 02:05:00,693 A CLEAR DIFFERENCE OF THE KINDS 2477 02:05:00,693 --> 02:05:04,330 OF STUFF THAT YOU'RE FOCUSING 2478 02:05:04,330 --> 02:05:05,231 ON. 2479 02:05:05,231 --> 02:05:07,266 THERE'S A CLEAR SEPARATION OF 2480 02:05:07,266 --> 02:05:09,302 SCALE IN THE SENSE THAT NEURONS 2481 02:05:09,302 --> 02:05:11,637 HAVE CELL WALLS AND THERE'S A 2482 02:05:11,637 --> 02:05:14,307 CLEAR INSIDE AND OUTSIDE AND IT 2483 02:05:14,307 --> 02:05:16,008 SEEMS LIKE A NATURAL SCALE TO 2484 02:05:16,008 --> 02:05:17,610 FOCUS ON AT THE LEVEL OF NEURONS 2485 02:05:17,610 --> 02:05:28,121 THEIR CONNECTIONS AND ACTIVITY. 2486 02:05:31,724 --> 02:05:39,198 THE MESO AND MICROSCOPIC SCALES 2487 02:05:39,198 --> 02:05:41,401 ARE SUBSERVIENT TO THIS LEVEL. 2488 02:05:41,401 --> 02:05:43,636 NATURAL SCALE LEVEL IS NEURONS. 2489 02:05:43,636 --> 02:05:48,574 >> MOVING ON TO WOLFGANG AND 2490 02:05:48,574 --> 02:05:48,741 ALI. 2491 02:05:48,741 --> 02:05:50,843 YOU WERE ADVOCATING FOR BRINGING 2492 02:05:50,843 --> 02:05:52,412 EVOLUTION TO YOUR MODELS. 2493 02:05:52,412 --> 02:05:53,379 >> RIGHT. 2494 02:05:53,379 --> 02:05:55,348 NOT TO EXPLICITLY PUT EVOLUTION 2495 02:05:55,348 --> 02:05:58,418 IN THE MODEL BUT TO MINE WHAT 2496 02:05:58,418 --> 02:06:00,052 EVOLUTION HAS DONE TO INFORM OUR 2497 02:06:00,052 --> 02:06:00,286 MODELS. 2498 02:06:00,286 --> 02:06:01,587 NUMBER ONE I DON'T THINK THE 2499 02:06:01,587 --> 02:06:04,857 GOAL OF A.I. SHOULD BE TO 2500 02:06:04,857 --> 02:06:15,401 REPLICATE ANY BIOLOGICAL SYSTEM 2501 02:06:16,602 --> 02:06:21,407 AND I THINK A FLY IS JUST AS 2502 02:06:21,407 --> 02:06:22,241 GENERALLY INTELLIGENT IN IT'S 2503 02:06:22,241 --> 02:06:24,010 OWN ENVIRONMENT AND NICHE AS 2504 02:06:24,010 --> 02:06:25,044 A HUMAN AS IN THEIRS. 2505 02:06:25,044 --> 02:06:28,915 I THINK WHEN WE BUILD THE A.I. 2506 02:06:28,915 --> 02:06:30,716 SYSTEMS WE SHOULD BE THINKING 2507 02:06:30,716 --> 02:06:35,087 ABOUT BUILDING NEW SPECIES OF 2508 02:06:35,087 --> 02:06:36,722 WHICH HAVE INTELLIGENCES 2509 02:06:36,722 --> 02:06:39,225 EQUIVALENT TO OR THAT WOULD FIT 2510 02:06:39,225 --> 02:06:42,762 INTO THE SAME SCALE OF 2511 02:06:42,762 --> 02:06:43,429 BEHAVIORAL COMPLEXITY AND 2512 02:06:43,429 --> 02:06:45,097 COGNITIVE COMPLEXITY WE SEE IN 2513 02:06:45,097 --> 02:06:47,133 LIVING SYSTEMS. 2514 02:06:47,133 --> 02:06:51,637 AND AS WE SUCCEED IN DOING THAT 2515 02:06:51,637 --> 02:06:52,972 WE'RE CREATING ARTIFICIAL 2516 02:06:52,972 --> 02:06:53,306 INTELLIGENCES. 2517 02:06:53,306 --> 02:06:54,974 THAT'S THE WAY I WOULD LOOK AT 2518 02:06:54,974 --> 02:06:55,174 THAT. 2519 02:06:55,174 --> 02:06:56,876 >> AND I'M HAPPY TO AGREE THAT 2520 02:06:56,876 --> 02:06:58,411 FLIES HAVE BEHAVIORS WE WANT TO 2521 02:06:58,411 --> 02:07:04,817 BRING INTO OUR A.I. MODELS BUT 2522 02:07:04,817 --> 02:07:11,123 HOW ARE WE GOING TO LOOK 2523 02:07:11,123 --> 02:07:15,528 AT THE DATA FROM NEUROSCIENCE TO 2524 02:07:15,528 --> 02:07:16,229 RECREATE THESE INVALUABLE 2525 02:07:16,229 --> 02:07:22,368 BEHAVIORS WE GET? 2526 02:07:22,368 --> 02:07:23,336 >> IT'S A GREAT QUESTION AND 2527 02:07:23,336 --> 02:07:24,837 ALREADY A.I. HAS PROVEN WHAT IT 2528 02:07:24,837 --> 02:07:26,439 CAN DO WITH STRUCTURAL DATA 2529 02:07:26,439 --> 02:07:26,906 ABOUT THE WIRING OF THE BRAIN. 2530 02:07:26,906 --> 02:07:31,143 WHAT I WOULD ARGUE IS WE FOCUS 2531 02:07:31,143 --> 02:07:37,149 ON NEURONS AND STRUCTURAL 2532 02:07:37,149 --> 02:07:41,387 CONNECTIONS. EXPLORE NON-NEURONAL 2533 02:07:41,387 --> 02:07:48,327 CELLS, NON-ELECTRICAL SIGNALS AND 2534 02:07:48,327 --> 02:07:48,995 DYNAMICS AS INFORMATIONAL CARRIER. 2535 02:07:48,995 --> 02:07:50,930 THERE'S ROOM FOR THE EVO-DEVO SIDE 2536 02:07:50,930 --> 02:07:52,665 AND BACK TO SIGNALS IN THE BRAIN 2537 02:07:52,665 --> 02:07:55,768 AND QUESTIONS ABOUT HOW THEY CAN BE 2538 02:07:55,768 --> 02:08:01,374 INTEGRATED IN AN A.I. MODEL 2539 02:08:01,374 --> 02:08:04,644 >> IF I CAN PULL ON THAT MORE 2540 02:08:04,644 --> 02:08:06,045 AND ONE MORE QUESTION, HOW DO 2541 02:08:06,045 --> 02:08:07,480 YOU KNOW WHEN YOU'RE DONE? 2542 02:08:07,480 --> 02:08:10,883 WHEN YOU'VE RECREATED THE RIGHT 2543 02:08:10,883 --> 02:08:14,287 DYNAMICS OR AS MANY NEURONAL 2544 02:08:14,287 --> 02:08:16,122 SUBTYPES OBSERVES IN THE DATASETS 2545 02:08:16,122 --> 02:08:17,924 OR WHEN YOU GET THE RIGHT 2546 02:08:17,924 --> 02:08:26,799 DYNAMICS IN THE CONNECTIONS? 2547 02:08:26,799 --> 02:08:28,968 >> I'M ASKING YOU BECAUSE YOUR 2548 02:08:28,968 --> 02:08:30,570 MIC IS ON. 2549 02:08:30,570 --> 02:08:31,904 >> I ALWAYS ASK STUDENTS TO DRAW 2550 02:08:31,904 --> 02:08:35,374 A PROTEIN AND EACH ONE DRAWS A 2551 02:08:35,374 --> 02:08:41,080 DIFFERENT MODEL OF A PROTEIN. 2552 02:08:41,080 --> 02:08:42,882 COMPLEXITY OF THE MODEL HAS TO 2553 02:08:42,882 --> 02:08:47,653 MATCH WHAT WE'RE TRYING TO ACHIEVE 2554 02:08:47,653 --> 02:08:50,856 THAT'S AN OVERLY SIMPLISTIC 2555 02:08:50,856 --> 02:09:01,267 ANSWER TO YOUR QUESTION. 2556 02:09:02,301 --> 02:09:06,572 >> I WOULD LIKE TO ASK A 2557 02:09:06,572 --> 02:09:11,711 QUESTION TO THE SPEAKERS AND 2558 02:09:11,711 --> 02:09:22,254 ANYONE AT THE TABLE FROM THE 2559 02:09:24,023 --> 02:09:29,962 PERSPECTIVE OF A THEORIST. I LIKE 2560 02:09:29,962 --> 02:09:40,339 THE ARROW OF NEUROSCIENCE TO A.I. 2561 02:09:40,339 --> 02:09:45,378 AND ALSO WANT THE VIRTUAL LOOP TO 2562 02:09:45,378 --> 02:09:46,445 HAVE BETTER THEORIES ABOUT 2563 02:09:46,445 --> 02:09:54,487 THE BRAIN AND HOW IT COMPUTES. 2564 02:09:54,487 --> 02:09:56,455 HOW SHOULD WE EXPAND OUR 2565 02:09:56,455 --> 02:09:57,423 UNDERSTANDING OF NEURAL 2566 02:09:57,423 --> 02:09:58,824 COMPUTATION AND COMPUTING IN THE 2567 02:09:58,824 --> 02:10:03,462 BRAIN TO INCORPORATE MANY 2568 02:10:03,462 --> 02:10:04,964 DIFFERENT TYPES OF BIOLOGICAL 2569 02:10:04,964 --> 02:10:15,441 SYSTEMS LIKE GLIA, NON-NEURON 2570 02:10:16,142 --> 02:10:17,810 COMPUTATIONS OR NEUROMODULATION, 2571 02:10:17,810 --> 02:10:26,485 EVO-DEVO, TO INFORM THEORETICAL 2572 02:10:26,485 --> 02:10:34,360 ADVANCES AND PRACTICAL NEUROAI 2573 02:10:34,360 --> 02:10:34,894 IMPLICATIONS 2574 02:10:34,894 --> 02:10:36,562 WHAT QUESTIONS SHOULD WE HAVE IN 2575 02:10:36,562 --> 02:10:40,800 THINKING ABOUT THIS TYPE OF 2576 02:10:40,800 --> 02:10:42,702 NEUROAI AND IF YOU'RE A THEORIST 2577 02:10:42,702 --> 02:10:49,041 IF YOU HAVE THOUGHTS ON THAT. 2578 02:10:49,041 --> 02:10:55,648 TOWARDS THE END OF A TALK AND 2579 02:10:55,648 --> 02:11:00,720 ANDREAS WAS TALKING ABOUT 2580 02:11:00,720 --> 02:11:03,422 REPRESENTATIONS SO MAYBE I 2581 02:11:03,422 --> 02:11:09,528 COULD POINT THE QUESTION TO ANDREAS 2582 02:11:09,528 --> 02:11:11,764 IF YOU LOOK AT THE NEXT FIVE TO 2583 02:11:11,764 --> 02:11:14,166 TEN YEARS AT THE LEVEL OF NEURAL 2584 02:11:14,166 --> 02:11:19,371 REPRESENTATIONS AND NEURAL 2585 02:11:19,371 --> 02:11:21,574 DYNAMICS, WE HAVE A DIRECT WAY 2586 02:11:21,574 --> 02:11:25,344 TO MAKE LINKS TO A.I. IN THEORY 2587 02:11:25,344 --> 02:11:28,013 I THINK IT'S ONE OF THE MOST 2588 02:11:28,013 --> 02:11:30,416 FRUITFUL DIRECTION AND AT THAT 2589 02:11:30,416 --> 02:11:32,985 LEVEL WE CAN ANALYZE NEURAL 2590 02:11:32,985 --> 02:11:35,888 DYNAMICS, MANIFOLDS, BRINGING 2591 02:11:35,888 --> 02:11:37,723 THEORY AND WE CAN DO THE SAME 2592 02:11:37,723 --> 02:11:47,967 WITH ARTIFICIAL NEURAL NETWORKS. 2593 02:11:49,368 --> 02:11:52,204 WE CAN DO THIS LEVEL OF ANALYSIS 2594 02:11:52,204 --> 02:11:57,176 WOULD LIKE TO SEE THEORY NOT JUST 2595 02:11:57,176 --> 02:11:58,477 WITH STATISTICAL TOOLS IN ANALYZING THE GEOMETRY OF NEURAL MANIFOLDS 2596 02:11:58,477 --> 02:12:04,817 BUT TRY TO RELATE THEORETICAL WAYS 2597 02:12:04,817 --> 02:12:15,361 TO UNDERSTAND DIGITAL TWINS IN THE NEURAL MANIFOLDS IN RELATIONSHIP TO WORLD MODELS. 2598 02:12:15,361 --> 02:12:16,762 OUR BRAINS HAVE CAUSAL UNDERSTANDING 2599 02:12:16,762 --> 02:12:18,864 THE WORLD MODEL IS INCORPORATED IN 2600 02:12:18,864 --> 02:12:20,065 NEURAL DYNAMICS. 2601 02:12:20,065 --> 02:12:21,167 IF WE UNDERSTAND THESE PRINCIPLES 2602 02:12:21,167 --> 02:12:23,836 THEN WE'LL BE ABLE TO HAVE 2603 02:12:23,836 --> 02:12:25,171 IMPACT IN A.I. 2604 02:12:25,171 --> 02:12:27,039 AT LEAST THAT'S ONE AREA I 2605 02:12:27,039 --> 02:12:32,878 THINK WE SHOULD PUT EFFORT IN 2606 02:12:32,878 --> 02:12:41,654 BECAUSE IT'S AN AREA WE SEE EVIDENCE 2607 02:12:41,654 --> 02:12:46,625 THAT SOMETHING IS POSSIBLE. I WAS ALSO WONDERING PATRICK IF YOU HAVE 2608 02:12:46,625 --> 02:12:51,997 THOUGHTS ON ALIGNMENT OR INTERPRETABILITY? 2609 02:12:54,800 --> 02:12:57,736 THERE'S A REVOLUTION IN MECHANISTIC 2610 02:12:57,736 --> 02:12:58,571 INTERPRETABILITY. IF YOU'VE BEEN 2611 02:12:58,571 --> 02:13:00,005 FOLLOWING THE LITERATURE THE LAST 2612 02:13:00,005 --> 02:13:00,806 FIVE YEARS, THE LAST YEAR LOOKS VERY 2613 02:13:00,806 --> 02:13:10,716 DIFFERENT FROM THE FIRST FOUR. 2614 02:13:10,716 --> 02:13:11,383 MECHANISTIC INTERPRETABILITY IS 2615 02:13:11,383 --> 02:13:16,655 LIKE NEUROSCIENCE FOR A.I. 2616 02:13:16,655 --> 02:13:20,893 PEOPLE ARE DOING LOOKING AT 2617 02:13:20,893 --> 02:13:24,196 REPRESENTATION AND ABLATIONS OF 2618 02:13:24,196 --> 02:13:27,666 SINGLE NEURONS IN A.I. MODELS AND 2619 02:13:27,666 --> 02:13:32,872 MAPPING OUT RECEPTIVE FIELDS 2620 02:13:32,872 --> 02:13:33,973 MECHANISTIC INTERPRETABILITY HAS 2621 02:13:33,973 --> 02:13:35,574 MOVED FROM SINGLE NEURON VIEWS 2622 02:13:35,574 --> 02:13:38,177 TO POPULATION-BASED VIEWS 2623 02:13:38,177 --> 02:13:43,782 WHETHER IT'S WORK ON SPARSE AUTO- 2624 02:13:43,782 --> 02:13:47,653 ENCODERS TO EXPLAIN LLM TO DOING 2625 02:13:47,653 --> 02:13:51,056 POPULATION ABLATIONS AND 2626 02:13:51,056 --> 02:13:53,792 ACTIVATION VECTORS IN ANOTHER 2627 02:13:53,792 --> 02:13:54,460 SUBPART OF THAT FIELD. 2628 02:13:54,460 --> 02:13:56,161 IF THERE'S ANY ARROW OF 2629 02:13:56,161 --> 02:13:57,663 INFLUENCE IN THAT FIELD, TAKING 2630 02:13:57,663 --> 02:14:00,599 THOSE TECHNIQUES BACK INTO 2631 02:14:00,599 --> 02:14:01,867 NEUROSCIENCE WOULD SEEM VERY 2632 02:14:01,867 --> 02:14:02,134 FRUITFUL. 2633 02:14:02,134 --> 02:14:02,935 THE KIND OF THINGS THEY CAN 2634 02:14:02,935 --> 02:14:03,969 DO IS AMAZING. 2635 02:14:03,969 --> 02:14:05,471 YOU CAN TAKE THE REPRESENTATIONS 2636 02:14:05,471 --> 02:14:10,676 FROM ONE NETWORK LOOKING AT A 2637 02:14:10,676 --> 02:14:12,044 PARTICULAR STIMULATION AND THEN 2638 02:14:12,044 --> 02:14:13,145 GRAFT IT ONTO ANOTHER ONE AND 2639 02:14:13,145 --> 02:14:18,417 SWAP OUT LAYERS AND DO ABLATIONS 2640 02:14:18,417 --> 02:14:20,219 ZAP SOME NEURONS. IF WE CAN DO 2641 02:14:20,219 --> 02:14:30,396 IN NEUROSCIENCE, WE'D BE GOLDEN. 2642 02:14:30,396 --> 02:14:32,498 IS THERE A MINIMUM SET OF TOOLS 2643 02:14:32,498 --> 02:14:37,469 STRUCTURED OPTOGENETICS, STIMULA- 2644 02:14:37,469 --> 02:14:40,105 TION OR THE ABILITY TO GROW NET- 2645 02:14:40,105 --> 02:14:50,516 WORKS AND BRING INTO NEUROSCIENCE? 2646 02:14:52,418 --> 02:14:56,388 EVEN IN A.I., PHYSICAL INTELLIGENCE 2647 02:14:56,388 --> 02:15:02,461 AND EMBODIED A.I. ARE THE HARDER 2648 02:15:02,461 --> 02:15:05,130 QUESTIONS COMPARED TO VISUAL ANN. 2649 02:15:05,130 --> 02:15:08,801 WHAT ARE SOME TYPES OF DATASETS 2650 02:15:08,801 --> 02:15:12,905 THAT MAY BE NEEDED FOR OR TYPES 2651 02:15:12,905 --> 02:15:16,542 OF THEORIES OR CONCEPTUAL 2652 02:15:16,542 --> 02:15:19,511 FRAMEWORKS NEEDED IN THIS DOMAIN 2653 02:15:19,511 --> 02:15:30,055 IF ANYBODY HAS ANYTHING TO ADD. 2654 02:15:43,535 --> 02:15:45,404 EMBRACE HIGH-ENTROPY DATA. 2655 02:15:45,404 --> 02:15:50,876 NEED ANIMALS TO INTERACT IN 2656 02:15:50,876 --> 02:15:57,249 NATURALISTIC ENVIRONMENT. 2657 02:15:57,249 --> 02:15:59,651 NEUROSCIENTISTS ARE TRAINED TO 2658 02:15:59,651 --> 02:16:05,391 COLLECT HYPOTHESIS-DRIVEN DATA. 2659 02:16:05,391 --> 02:16:07,760 WE NEED TO GET AWAY FROM THAT AND 2660 02:16:07,760 --> 02:16:18,237 DO NATURALISTIC DATASET COLLECTION 2661 02:16:21,206 --> 02:16:21,807 >> GREAT TO SEE THOSE TYPES OF DATASETS OPEN TO PUBLIC TO 2662 02:16:21,807 --> 02:16:22,474 DEVELOP MODELS AND THEORIES TO TEST IDEAS, ANYONE? 2663 02:16:22,474 --> 02:16:25,978 >> MAY I SAY SOMETHING ABOUT 2664 02:16:25,978 --> 02:16:26,245 ALIGNMENT? 2665 02:16:26,245 --> 02:16:28,313 THE WAY WE ARE THINKING ABOUT 2666 02:16:28,313 --> 02:16:33,786 ALIGNMENT IS WE TRAIN THESE 2667 02:16:33,786 --> 02:16:35,821 LARGE MODELS WHICH ARE UNALIGNED. 2668 02:16:35,821 --> 02:16:36,588 AND THEN WE ALIGN THEM USING 2669 02:16:36,588 --> 02:16:39,591 RLHF OR 2670 02:16:39,591 --> 02:16:41,794 SOMETHING LIKE THAT WHICH IS 2671 02:16:41,794 --> 02:16:43,328 VERY UNNATURAL. 2672 02:16:43,328 --> 02:16:48,767 THE REASON YOU AND I ARE ALIGNED 2673 02:16:48,767 --> 02:16:50,936 IS BECAUSE WE WENT THROUGH A 2674 02:16:50,936 --> 02:16:52,271 LONG DEVELOPMENTAL PROCESS AND 2675 02:16:52,271 --> 02:16:53,172 LEARNED IN THE EARLY STAGE IN 2676 02:16:53,172 --> 02:16:56,975 OUR LIFE HOW TO BEHAVE AND WHAT 2677 02:16:56,975 --> 02:16:57,910 ARE GOOD THINGS TO THINK ABOUT 2678 02:16:57,910 --> 02:17:01,947 AND BAD WAYS. 2679 02:17:01,947 --> 02:17:03,649 THE ALIGNMENT PROFESSION HAS TO 2680 02:17:03,649 --> 02:17:06,051 BE INTEGRATED WITH THE LEARNING 2681 02:17:06,051 --> 02:17:06,351 PROCESS. 2682 02:17:06,351 --> 02:17:10,255 THE SYSTEMS AS THEY LEARN ARE 2683 02:17:10,255 --> 02:17:16,762 INHERENTLY ASSIGNED. MAYBE NOT 2684 02:17:16,762 --> 02:17:18,697 PERFECTLY. SOME OF US NEED TUNING 2685 02:17:18,697 --> 02:17:28,006 BUT IT'S NOT AN AFTERTHOUGHT. 2686 02:17:28,006 --> 02:17:31,176 WE DON'T FIRST TRAIN BY WOLVES AND 2687 02:17:31,176 --> 02:17:37,583 THEN TRY TO CIVILIZE IT. THIS IS 2688 02:17:37,583 --> 02:17:38,851 A.I. PEOPLE SHOULD THINK ABOUT. 2689 02:17:38,851 --> 02:17:40,586 >> WHAT WOULD BE GOOD METRICS TO 2690 02:17:40,586 --> 02:17:48,594 MEASURE SUCH ALIGNMENT? 2691 02:17:48,594 --> 02:17:50,162 >> THAT'S THE MAIN TOPIC OF 2692 02:17:50,162 --> 02:17:51,663 SESSION 2 AND NEED TO STAY ON 2693 02:17:51,663 --> 02:17:54,366 TIME WE'LL NEED TO MOVE TO PAUL 2694 02:17:54,366 --> 02:17:56,135 AS AN INVITED DISCUSSANT BUT 2695 02:17:56,135 --> 02:17:59,638 WE'LL HAVE MORE ON METRICS FOR 2696 02:17:59,638 --> 02:18:03,142 SURE. 2697 02:18:03,142 --> 02:18:05,144 >> A.I. IS SUPERFAST. 2698 02:18:05,144 --> 02:18:08,046 SCIENCE IS SUPER SLOW AND 2699 02:18:08,046 --> 02:18:14,720 DEPENDENT ON LUCK. GIVEN THE 2700 02:18:14,720 --> 02:18:16,722 DELUGE OF DATA AND PIPELINES, 2701 02:18:16,722 --> 02:18:17,589 HOW DO WE MAKE TRAINEES MORE 2702 02:18:17,589 --> 02:18:23,095 EFFICIENT? 2703 02:18:23,095 --> 02:18:25,397 IF YOU HAD A PERFECT AVATAR OF 2704 02:18:25,397 --> 02:18:26,165 THE SKILLS NECESSARY TO 2705 02:18:26,165 --> 02:18:31,036 IMPLEMENT WHAT YOU'RE TALKING 2706 02:18:31,036 --> 02:18:32,171 ABOUT, WHAT WOULD THAT LOOK LIKE? 2707 02:18:32,171 --> 02:18:37,743 SO TRAINEES ARE PART OF THE 2708 02:18:37,743 --> 02:18:43,048 INFRASTRUCTURE WE NEED. WHAT DO 2709 02:18:43,048 --> 02:18:43,815 YOU NEED IN TERMS OF SKILLS, 2710 02:18:43,815 --> 02:18:53,992 ET CETERA? 2711 02:19:01,533 --> 02:19:03,835 I THINK THE MENTAL MODELS FOR 2712 02:19:03,835 --> 02:19:06,071 HOW WE'LL DO SCIENCE ARE IN THE 2713 02:19:06,071 --> 02:19:07,172 NEAR FUTURE MIGHT BE A LITTLE 2714 02:19:07,172 --> 02:19:14,313 BIT OFF. 2715 02:19:14,313 --> 02:19:15,781 THE WAY I DID SCIENCE IN GRAD 2716 02:19:15,781 --> 02:19:19,885 SCHOOL AND THE WAY MY 2717 02:19:19,885 --> 02:19:20,953 FOREFATHERS DID SCIENCE IS 2718 02:19:20,953 --> 02:19:25,924 SIMILAR BUT MIGHT CHANGE A 2719 02:19:25,924 --> 02:19:26,558 LITTLE BIT. 2720 02:19:26,558 --> 02:19:28,460 WE'RE SEEING FUNDAMENTAL ADVANCES 2721 02:19:28,460 --> 02:19:33,432 IN A.I. AND WE'RE STARTING TO SEE 2722 02:19:33,432 --> 02:19:38,904 USING A.I. FOR SCIENCE IN RADICAL 2723 02:19:38,904 --> 02:19:49,381 WAYS. TRAINEES WILL HAVE ACCESS TO 2724 02:19:52,184 --> 02:19:56,955 MORE KNOWLEDGEABLE, POWERFUL TOOLS 2725 02:19:56,955 --> 02:19:59,791 THE WAY WE TRAIN TRAINEES IT'S 2726 02:19:59,791 --> 02:20:01,260 HARD TO BUILD A ROAD GIVEN WE 2727 02:20:01,260 --> 02:20:04,029 DON'T KNOW WHERE IT'S GOING TO 2728 02:20:04,029 --> 02:20:05,664 END UP THE ADVANCE OF A.I. BUT WE 2729 02:20:05,664 --> 02:20:11,236 SHOULD TRY TO INCORPORATE IT AND 2730 02:20:11,236 --> 02:20:21,513 THINK ABOUT IT AND ADAPT. 2731 02:20:23,048 --> 02:20:25,117 >> WE'RE STILL BETTER AT 2732 02:20:25,117 --> 02:20:26,618 EXPLORING THAN A.I. AND ONE OF 2733 02:20:26,618 --> 02:20:30,589 THE THINGS THERE'S ALSO THE NEED 2734 02:20:30,589 --> 02:20:34,793 TO COLLABORATE ACROSS DIFFERENT 2735 02:20:34,793 --> 02:20:43,035 LEVELS OF EXPERTISE. 2736 02:20:43,035 --> 02:20:45,971 AND HAVE TO FOLLOW GUIDANCE AND 2737 02:20:45,971 --> 02:20:48,073 EXPLORE THINGS IN A RESEARCH 2738 02:20:48,073 --> 02:20:48,240 LAB. 2739 02:20:48,240 --> 02:20:51,243 WE CAN'T GO TO LIKE A 2740 02:20:51,243 --> 02:20:53,245 MATHEMATICIANS MODEL TO EXPLORE 2741 02:20:53,245 --> 02:20:54,613 WHATEVER YOU WANT FOR YOUR Ph.D. 2742 02:20:54,613 --> 02:21:00,319 BUT HAS TO BE MORE ABILITY TO BE 2743 02:21:00,319 --> 02:21:06,458 CURIOUS AT THE TRAINING LEVEL. 2744 02:21:06,458 --> 02:21:09,428 >> CURIOSITY WOULD DRIVE YOU TO 2745 02:21:09,428 --> 02:21:11,964 A MILLION THINGS SO ONE QUESTION 2746 02:21:11,964 --> 02:21:14,833 IS HOW TO FOCUS THE EXPLORATION 2747 02:21:14,833 --> 02:21:21,306 I SUPPOSE. 2748 02:21:21,306 --> 02:21:22,741 >> HOW TO FOCUS THE EXPLORATION 2749 02:21:22,741 --> 02:21:28,680 IS THE ROLE OF THE MENTOR TO 2750 02:21:28,680 --> 02:21:33,518 HELP DIRECT TRAINEES TO WHAT IS 2751 02:21:33,518 --> 02:21:39,491 MOST FRUITFUL, INTERESTING, AND 2752 02:21:39,491 --> 02:21:48,100 DOABLE. MORE GENERALLY WE NEED 2753 02:21:48,100 --> 02:21:49,434 TO GIVE PEOPLE THE TOOLS EARLY 2754 02:21:49,434 --> 02:21:51,870 TO FLOURISH AND BECOME GOOD 2755 02:21:51,870 --> 02:21:52,504 SCIENTISTS. WE ARE TALKING ABOUT 2756 02:21:52,504 --> 02:21:54,439 SOPHISTICATED SYSTEMS 2757 02:21:54,439 --> 02:21:55,674 WHETHER THAT'S A.I. SYSTEMS. SO 2758 02:21:55,674 --> 02:22:00,846 WHAT WE'RE DOING IN MY GROUP IS 2759 02:22:00,846 --> 02:22:04,383 BIOREALISTIC SIMULATIONS OF BRAIN 2760 02:22:04,383 --> 02:22:04,649 CIRCUITS. 2761 02:22:04,649 --> 02:22:11,323 IT'S SUCH A COMPLICATED THING TO 2762 02:22:11,323 --> 02:22:12,724 BE PRODUCTIVE. YOU HAVE TO LEARN 2763 02:22:12,724 --> 02:22:13,759 A LOT OF AND A LOT OF THE STEPS 2764 02:22:13,759 --> 02:22:18,897 IN GETTING THERE ARE TAKEN CARE 2765 02:22:18,897 --> 02:22:22,868 OF BY THE DEVELOPING TOOLS. 2766 02:22:22,868 --> 02:22:28,607 NEEDS TO TRAIN PEOPLE IN PHYSICS 2767 02:22:28,607 --> 02:22:37,249 AND USING THE TOOLS EARLY ON. 2768 02:22:37,249 --> 02:22:47,793 >> I WANT TO MAKE A COMMENT AND 2769 02:22:48,794 --> 02:22:53,365 >> A.I. HAS RELIED ON HUMANS 2770 02:22:53,365 --> 02:22:57,903 THOUSANDS OF YEARS PRODUCING 2771 02:22:57,903 --> 02:23:00,338 DATA AND IS THE TECHNOLOGY THAT'S 2772 02:23:00,338 --> 02:23:01,373 ACCELERATED THE PROGRESS. 2773 02:23:01,373 --> 02:23:04,810 NOW WE'RE IN A SPECIAL ERA IN 2774 02:23:04,810 --> 02:23:10,315 NEUROSCIENCE WHERE IF WE CREATE 2775 02:23:10,315 --> 02:23:17,022 LARGE CENTERS THAT WILL OUTPACE 2776 02:23:17,022 --> 02:23:18,290 THE AMOUNT OF DATA EVEN THE LARGE 2777 02:23:18,290 --> 02:23:18,824 LANGUAGE MODELS GET TRAINED ON. 2778 02:23:18,824 --> 02:23:20,826 IN THAT CASE IT'S POSSIBLE THAT 2779 02:23:20,826 --> 02:23:27,065 WE MAY BE IN A REGIME WE MAY BE 2780 02:23:27,065 --> 02:23:32,737 ABLE TO MOVE MAYBE EVEN FASTER 2781 02:23:32,737 --> 02:23:36,408 POSSIBLY THAN CURRENT A.I. 2782 02:23:36,408 --> 02:23:41,646 RESEARCH AND THE DATA BANDWIDTH 2783 02:23:41,646 --> 02:23:43,482 WILL BE HIGH LIKE RECORDING NEURAL 2784 02:23:43,482 --> 02:23:47,219 DATA VS BEHAVIORAL WHICH IS LOW 2785 02:23:47,219 --> 02:23:54,059 BIT, WHETHER YOUR WRITE TEXT OR 2786 02:23:54,059 --> 02:24:02,667 OR LABEL YOUR DATA. POSSIBLE FUTURE. 2787 02:24:02,667 --> 02:24:13,211 >> FOR ANTOINE AND ANDREAS WHAT 2788 02:24:14,279 --> 02:24:17,015 IS THE BALANCE FOR MORE DATA AND 2789 02:24:17,015 --> 02:24:17,949 ANALYZING DATA WE ALREADY HAVE FOR 2790 02:24:17,949 --> 02:24:23,421 A.I. SO WE STAY ON TARGET? 2791 02:24:23,421 --> 02:24:26,258 >> 70% FOR THE NEW DATA, 30% FOR 2792 02:24:26,258 --> 02:24:28,360 ANALYZING THE EXISTING ONES. 2793 02:24:28,360 --> 02:24:29,027 THERE'S A LOT OF TO BE DONE WITH 2794 02:24:29,027 --> 02:24:31,897 EXISTING ONES. 2795 02:24:31,897 --> 02:24:33,798 I'M VERY EXCITED WITH WHAT'S 2796 02:24:33,798 --> 02:24:42,407 COMING AND I AGREE WITH ANDREAS 2797 02:24:42,407 --> 02:24:43,675 THE BANDWIDTH IS GOING TO BE 2798 02:24:43,675 --> 02:24:43,942 TREMENDOUS. 2799 02:24:43,942 --> 02:24:46,011 >> CAN WE ACCELERATE 2800 02:24:46,011 --> 02:24:48,213 NEUROSCIENCE CONTRIBUTIONS BY 2801 02:24:48,213 --> 02:24:49,347 LOOKING AT EXISTING DATA? 2802 02:24:49,347 --> 02:24:58,156 >> I'M MORE PESSIMISTIC THERE. 2803 02:24:58,156 --> 02:24:59,057 A LOT OF THE DATA WE'VE 2804 02:24:59,057 --> 02:25:00,492 COLLECTED SO FAR ARE NOT THE 2805 02:25:00,492 --> 02:25:03,094 RIGHT KIND OF DATA FOR A.I. 2806 02:25:03,094 --> 02:25:03,995 THEY'RE BASICALLY VERY 2807 02:25:03,995 --> 02:25:14,406 HYPOTHESIS DRIVEN DATA. 2808 02:25:15,707 --> 02:25:22,147 THEY COULDN'T GENERALIZE AND I 2809 02:25:22,147 --> 02:25:23,715 ACTUALLY BELIEVE THE LAST DECADE 2810 02:25:23,715 --> 02:25:24,716 IS THE NEURAL TECHNOLOGY THAT 2811 02:25:24,716 --> 02:25:26,184 WAS THE BIG THING AND NOT THE 2812 02:25:26,184 --> 02:25:30,422 DATA WE'VE COLLECTED SO FAR. 2813 02:25:30,422 --> 02:25:31,957 I THINK ALSO THE WAY WE COLLECT 2814 02:25:31,957 --> 02:25:35,427 THE DATA MAY NEED TO CHANGE. 2815 02:25:35,427 --> 02:25:37,762 I'M NOT SURE IF THE LAB BASED 2816 02:25:37,762 --> 02:25:38,363 ORGANIZATION IS THE RIGHT WAY 2817 02:25:38,363 --> 02:25:48,506 FORWARD. WE HAVE TO THINK LIKE 2818 02:25:48,773 --> 02:25:49,874 PHYSICISTS DO. I THINK WE NEED 2819 02:25:49,874 --> 02:25:52,077 NEW TYPES OF DATA ACTUALLY. 2820 02:25:52,077 --> 02:25:53,345 >> GREAT AND WITH THAT I 2821 02:25:53,345 --> 02:25:54,980 WOULD LIKE TO OPEN IT UP TO 2822 02:25:54,980 --> 02:26:05,457 AUDIENCE Q&A AT THIS POINT. 2823 02:26:08,627 --> 02:26:12,163 >> I SEE A FACE I RECOGNIZE. 2824 02:26:12,163 --> 02:26:13,231 >> KURT THOROUGHMAN FROM 2825 02:26:13,231 --> 02:26:15,467 WASHINGTON UNIVERSITY IN ST. LOUIS 2826 02:26:15,467 --> 02:26:20,572 I WANT TO ASK ABOUT EXPERIMENTAL 2827 02:26:20,572 --> 02:26:20,905 MANIPULATION. IN NEUROSCIENCE 2828 02:26:20,905 --> 02:26:22,574 THERE'S VOLTAGE CLAMPS AND 2829 02:26:22,574 --> 02:26:23,742 OPTOGENETIC MANIPULATION AND IT 2830 02:26:23,742 --> 02:26:24,643 SEEMS LIKE A.I. SHOULD BE A 2831 02:26:24,643 --> 02:26:30,482 SYSTEM WHERE YOU CAN USE THAT 2832 02:26:30,482 --> 02:26:33,785 TO CONTROL COMPLEX SYSTEMS IN 2833 02:26:33,785 --> 02:26:37,689 WAY THAT WILL PAY BENEFITS AS 2834 02:26:37,689 --> 02:26:39,691 NEUROSCIENCE. WHAT NEWFANGALED WAYS 2835 02:26:39,691 --> 02:26:50,368 DOES NEUROSCIENCE LEARN ABOUT 2836 02:26:50,935 --> 02:26:54,406 TO BE ABLE TO UNDERSTAND A.I. AND 2837 02:26:54,406 --> 02:26:57,842 HOW CAN AI THEN USE VOLTAGE CLAMPS 2838 02:26:57,842 --> 02:26:58,443 TO FIGURE OUT THEIR OWN SYSTEMS? 2839 02:26:58,443 --> 02:26:59,044 >> I CAN ANSWER THAT A LITTLE 2840 02:26:59,044 --> 02:27:08,420 BIT. 2841 02:27:08,420 --> 02:27:13,525 >> MOST OF A.I. MODELS WHICH ARE 2842 02:27:13,525 --> 02:27:14,059 CURRENT ARE NOT RECURRENT 2843 02:27:14,059 --> 02:27:21,833 MODELS. 2844 02:27:21,833 --> 02:27:23,635 I THINK THERE WAS A LOT OF 2845 02:27:23,635 --> 02:27:26,705 REALLY INTERESTING WORK FROM 2846 02:27:26,705 --> 02:27:28,540 PEOPLE ON TRYING TO UNDERSTAND 2847 02:27:28,540 --> 02:27:32,744 WHAT'S GOING ON INSIDE OF 2848 02:27:32,744 --> 02:27:34,646 RECURRENT NEURAL NETWORKS AND 2849 02:27:34,646 --> 02:27:37,649 GIVEN THE SHIFTED TOWARDS 2850 02:27:37,649 --> 02:27:39,617 TRANSFORMERS THEY DON'T HAVE 2851 02:27:39,617 --> 02:27:44,456 AUTO REGRESSION. IT'S DIFFICULT 2852 02:27:44,456 --> 02:27:46,157 TO APPLY THE SAME TOOLS. VOLTAGE 2853 02:27:46,157 --> 02:27:51,529 CLAMP WOULD BE EMBEDDED IN CONTROL 2854 02:27:51,529 --> 02:27:58,937 THEORY AND MAKE MORE SENSE FOR RNN 2855 02:27:58,937 --> 02:28:03,641 PEOPLE ARE LOOKING AT STATE-SPACE 2856 02:28:03,641 --> 02:28:04,909 MODELS, A NEWFANGLED VERSION OF 2857 02:28:07,712 --> 02:28:08,713 RNN, REACHING INTERESTING PRO- 2858 02:28:09,047 --> 02:28:10,982 PERTIES AND WORKING WELL AND MAY 2859 02:28:10,982 --> 02:28:16,688 BE TIME TO UNEARTH THOSE PRIOR 2860 02:28:16,688 --> 02:28:19,090 TOOLS AND APPLY THEM TO THE NEW 2861 02:28:19,090 --> 02:28:19,624 SYSTEMS. 2862 02:28:19,624 --> 02:28:21,092 >> I'LL PUSH BACK BECAUSE 2863 02:28:21,092 --> 02:28:22,127 TRANSFORMERS WE BUILT. 2864 02:28:22,127 --> 02:28:23,461 IT'S OUR FAULT THEY LOOK THIS 2865 02:28:23,461 --> 02:28:27,799 WAY. WHEN I TEACH HODGKIN AND 2866 02:28:27,799 --> 02:28:30,468 HUXLEY THERE ARE SPACE CLAMPS, 2867 02:28:30,468 --> 02:28:37,809 ADJUSTMENTS OF EXTRACELLULAR FLUID 2868 02:28:37,809 --> 02:28:40,245 TO BALANCE IONIC MIX. WE SHOULD BE 2869 02:28:40,245 --> 02:28:44,849 ABLE TO DO THIS FOR A.I. WE 2870 02:28:44,849 --> 02:28:49,788 BUILT THE DARN THING. HOW CAN WE 2871 02:28:49,788 --> 02:28:56,628 MOVE BACK TOWARDS THOSE SORTS 2872 02:28:56,628 --> 02:28:57,061 OF MANIPULATABILITIES? 2873 02:28:57,061 --> 02:29:03,835 >> LIKE HH MODEL FOR A.I.? 2874 02:29:03,835 --> 02:29:05,136 >> OR LOOK AT WHAT'S HAPPENING 2875 02:29:05,136 --> 02:29:07,438 UNDER THE HOOD AND PUT THE 2876 02:29:07,438 --> 02:29:09,174 CONNECTIONS IN THE A.I. TO MAKE 2877 02:29:09,174 --> 02:29:12,377 THAT MORE INTERPRETABLE THE WAY 2878 02:29:12,377 --> 02:29:13,711 THAT THEY BUILT VOLTAGE MAPS. 2879 02:29:13,711 --> 02:29:24,222 >> IT DEPENDS ON THE QUESTION. 2880 02:29:29,828 --> 02:29:32,697 AND PREDICTING NEURAL ACTIVITY 2881 02:29:32,697 --> 02:29:34,532 DYNAMICS BY ADDING MORE 2882 02:29:34,532 --> 02:29:35,266 PARAMETERS FROM A PRACTICAL 2883 02:29:35,266 --> 02:29:40,438 POINT OF VIEW IT'S NOT HELPING 2884 02:29:40,438 --> 02:29:42,640 BECAUSE THE NEURAL NETWORKS ARE 2885 02:29:42,640 --> 02:29:45,143 A FUNCTION OF APPROXIMATORS. 2886 02:29:45,143 --> 02:29:47,412 WE HAVE BETTER WAYS TO OPTIMIZE 2887 02:29:47,412 --> 02:29:49,647 THEM, AS PATRICK IS SAYING, BEING 2888 02:29:49,647 --> 02:29:49,981 GENERIC ARCHITECTURES. 2889 02:29:49,981 --> 02:29:53,184 THERE'S A MATCH BETWEEN THE 2890 02:29:53,184 --> 02:29:54,118 OPTIMIZATION METHOD IN MACHINE 2891 02:29:54,118 --> 02:29:58,923 LEARNING AND THE ARCHITECTURE. 2892 02:29:58,923 --> 02:30:01,259 BECAUSE OF HOW THINGS EVOLVED WE 2893 02:30:01,259 --> 02:30:03,294 HAVE METHODS WORKING BEST WITH 2894 02:30:03,294 --> 02:30:04,829 EXISTING ARCHITECTURE AND GIVES 2895 02:30:04,829 --> 02:30:06,798 THE BEST MODELS. 2896 02:30:06,798 --> 02:30:07,932 WHEN WE ADD HIGHER LEVEL 2897 02:30:07,932 --> 02:30:11,636 COMPLEXITIES THEY GET HARDER TO 2898 02:30:11,636 --> 02:30:14,339 FIT TO COMPLEX DATA AND MANY 2899 02:30:14,339 --> 02:30:14,572 NEURONS. 2900 02:30:14,572 --> 02:30:15,273 >> THANK YOU. 2901 02:30:15,273 --> 02:30:15,640 TERRY. 2902 02:30:15,640 --> 02:30:23,648 >> THIS HAS BEEN A WONDERFUL 2903 02:30:23,648 --> 02:30:26,084 DISCUSSION AND 2904 02:30:26,084 --> 02:30:27,185 LOTS OF ISSUES 2905 02:30:27,185 --> 02:30:28,586 BROUGHT UP. 2906 02:30:28,586 --> 02:30:30,488 WHEN THE WRIGHT BROTHERS WERE 2907 02:30:30,488 --> 02:30:34,025 DESIGNING THE PLANE POWERED WITH 2908 02:30:34,025 --> 02:30:35,560 MAN FLIGHT WHAT DID THEY LOOK AT 2909 02:30:35,560 --> 02:30:36,194 FROM BIRDS? 2910 02:30:36,194 --> 02:30:40,532 WHAT PRINCIPLES DID THEY 2911 02:30:40,532 --> 02:30:40,932 EXTRACT? 2912 02:30:40,932 --> 02:30:46,104 THEY EXTRACTED HOW BIRDS COULD 2913 02:30:46,104 --> 02:30:46,905 GLIDE LONG DISTANCES WITHOUT 2914 02:30:46,905 --> 02:30:49,107 MUCH ENERGY AND THEN LOOKED AT 2915 02:30:49,107 --> 02:30:55,647 TO THE WINGS AND THE 2916 02:30:55,647 --> 02:30:57,682 AERODYNAMICS OF WINGS. 2917 02:30:57,682 --> 02:31:00,752 THEY LOOKED AT THE FEATHER. 2918 02:31:00,752 --> 02:31:04,555 THE DIDN'T TAKE AWAY THE DETAILS 2919 02:31:04,555 --> 02:31:06,925 OF HOW NATURE BUILT THE 2920 02:31:06,925 --> 02:31:09,460 MATERIALS BUT THEY TOOK AWAY 2921 02:31:09,460 --> 02:31:10,895 SOMETHING STIFF THAT HAD A LOT 2922 02:31:10,895 --> 02:31:15,633 OF AREA AND SO THEY BUILT WINGS 2923 02:31:15,633 --> 02:31:18,703 THAT HAD WOOD SPARS AND CANVAS 2924 02:31:18,703 --> 02:31:19,237 LIGHTWEIGHT BETWEEN IT. 2925 02:31:19,237 --> 02:31:20,872 IT WAS CLEAR THE GOAL WAS TO 2926 02:31:20,872 --> 02:31:26,110 EXTRACT PRINCIPLES FROM ALL 2927 02:31:26,110 --> 02:31:29,948 THOSE DETAILS. HERE'S THE 2928 02:31:29,948 --> 02:31:30,214 QUESTION. 2929 02:31:30,214 --> 02:31:31,582 WHAT ADDITIONAL PRINCIPLES CAN 2930 02:31:31,582 --> 02:31:38,222 WE EXTRACT FROM ALL THE 2931 02:31:38,222 --> 02:31:40,391 FANTASTIC DISCOVERIES MADE IN 2932 02:31:40,391 --> 02:31:42,093 NEUROSCIENCES AND THE DATABASES 2933 02:31:42,093 --> 02:31:43,261 WE'RE BUILDING THAT COULD BE OF 2934 02:31:43,261 --> 02:31:48,533 USE IN A.I.? WE'VE HEARD A 2935 02:31:48,533 --> 02:31:57,842 FEW. I AGREE LARGE LANGUAGE MODELS 2936 02:31:57,842 --> 02:32:02,714 HAVE TO BE ALIGNED WHILE THEY'RE 2937 02:32:02,714 --> 02:32:03,214 BEING BUILT. 2938 02:32:03,214 --> 02:32:04,515 THERE'S ONLY TWO TIME SCALES IN 2939 02:32:04,515 --> 02:32:07,652 THEM. TRAINING AT THE BEGINNING 2940 02:32:07,652 --> 02:32:11,122 AND THEN THE INFERENCE. 2941 02:32:11,122 --> 02:32:13,157 THAT'S LEAVING OUT A TREMENDOUS 2942 02:32:13,157 --> 02:32:14,559 NUMBER OF TIME SCALES IN BETWEEN 2943 02:32:14,559 --> 02:32:15,927 WHAT ELSE CAN WE LEARN ABOUT 2944 02:32:15,927 --> 02:32:19,597 TIME SCALES FROM NEUROSCIENCE 2945 02:32:19,597 --> 02:32:30,108 GENERAL PRINCIPLES? 2946 02:32:35,346 --> 02:32:40,118 THEY EXTRACTED LIFT, NOT FLIGHT 2947 02:32:40,118 --> 02:32:41,652 TONY MAKES THIS EXPLICIT. PLANES 2948 02:32:41,652 --> 02:32:44,522 DON'T FLY THE WAY BIRDS FLY SO THE 2949 02:32:44,522 --> 02:32:47,158 TARGET IN THAT INSTANCE -- SO 2950 02:32:47,158 --> 02:32:48,826 THE PRINCIPLE IS LIFT NOT FLIGHT. 2951 02:32:48,826 --> 02:32:50,962 >> IT'S AIR FOILS. 2952 02:32:50,962 --> 02:32:52,296 THAT'S WHAT THEY EXTRACTED. 2953 02:32:52,296 --> 02:32:53,998 THEY LOOKED AT THE AIR FOIL OF 2954 02:32:53,998 --> 02:32:58,903 THE BIRD WING. IN FACT THE 2955 02:32:58,903 --> 02:33:03,274 REAL PROBLEM GETTING PLANES SAFE 2956 02:33:03,274 --> 02:33:06,844 WAS CONTROL AND IT WAS MENTIONED 2957 02:33:06,844 --> 02:33:08,579 IN WHAT NATURE GOT RIGHT. 2958 02:33:08,579 --> 02:33:10,948 HOW TO DO CONTROL SAFELY. WE 2959 02:33:10,948 --> 02:33:21,459 DON'T KNOW HOW TO DO THAT IN A.I. 2960 02:33:22,193 --> 02:33:24,829 WHAT OTHER PRINCIPLES CAN WE 2961 02:33:24,829 --> 02:33:28,299 EXTRACT FROM NATURE FROM ALL THE 2962 02:33:28,299 --> 02:33:29,901 FACTS AND WE HAVE HUGE 2963 02:33:29,901 --> 02:33:30,368 DATABASES. 2964 02:33:30,368 --> 02:33:34,205 WE WANT TO EXTRACT PRINCIPLES. 2965 02:33:34,205 --> 02:33:36,407 BY THE WAY NOT ONLY TO HELP A.I. 2966 02:33:36,407 --> 02:33:37,375 BUT HELP OURSELVES UNDERSTAND 2967 02:33:37,375 --> 02:33:42,547 HOW BRAINS ARE ORGANIZED. 2968 02:33:42,547 --> 02:33:47,885 >> A COUPLE THINGS THAT COME TO 2969 02:33:47,885 --> 02:33:55,827 MIND ARE LEARNING, LOCAL LEARNING 2970 02:33:55,827 --> 02:33:59,330 RULES VS. HOW A.I. SYSTEMS WORK. 2971 02:33:59,330 --> 02:33:59,797 NEUROMODULATOR CHANGING STATE 2972 02:33:59,797 --> 02:34:05,503 AND CONTRIBUTING TO LEARNING. 2973 02:34:05,503 --> 02:34:07,705 THERE'S PROBABLY MORE BUT ALSO I 2974 02:34:07,705 --> 02:34:10,274 WANT TO SAY FOR BROADER CONTEXT 2975 02:34:10,274 --> 02:34:13,511 MY PERSONAL OPINION IS PROBABLY 2976 02:34:13,511 --> 02:34:21,853 A.I. CAN DEVELOP MUCH FASTER NOT 2977 02:34:21,853 --> 02:34:24,689 NECESSARILY BORROWING FROM THE 2978 02:34:24,689 --> 02:34:31,662 BRAIN. BUT THE HOPE IS THAT WE 2979 02:34:31,662 --> 02:34:32,930 CAN LEARN SOMETHING THAT'LL BE 2980 02:34:32,930 --> 02:34:34,165 USEFUL FOR THE A.I. MY WORK IS 2981 02:34:34,165 --> 02:34:37,135 ABOUT USING A.I. TO LEARN ABOUT 2982 02:34:37,135 --> 02:34:38,369 THE BRAIN, HOW THE BRAIN WORKS, 2983 02:34:38,369 --> 02:34:39,971 WE WANT TO UNDERSTAND THE BRAIN. 2984 02:34:39,971 --> 02:34:41,773 >> IT GOES IN BOTH DIRECTION. 2985 02:34:41,773 --> 02:34:43,608 WE SHOULD BE ABLE TO EXTRACT 2986 02:34:43,608 --> 02:34:44,475 PRINCIPLES FROM THE LARGE 2987 02:34:44,475 --> 02:34:46,277 LANGUAGE MODELS TO HELP ASK NEW 2988 02:34:46,277 --> 02:34:46,978 QUESTIONS ABOUT THE BRAIN. 2989 02:34:46,978 --> 02:34:54,085 I AGREE WITH THAT. 2990 02:34:54,085 --> 02:34:55,987 WE HAVE LEARNED A LOT OF THINGS 2991 02:34:55,987 --> 02:34:57,488 FROM THE BRAIN THAT A.I. IS NOT 2992 02:34:57,488 --> 02:35:03,628 NOT USING ANYMORE. FOR EXAMPLE, 2993 02:35:03,628 --> 02:35:07,465 ATTRACTOR NETWORKS, DENDRITIC 2994 02:35:07,465 --> 02:35:08,733 COMPUTATION, SUPERVISED LEARNING 2995 02:35:08,733 --> 02:35:09,000 FROM MAP FORMATION. 2996 02:35:09,000 --> 02:35:10,368 THEY'RE OLD METHOD THAT WE LEARN 2997 02:35:10,368 --> 02:35:13,371 FROM THE BRAIN AND PUT INTO A.I. 2998 02:35:13,371 --> 02:35:16,174 AND THEN A.I. HAS ABANDONED FOR 2999 02:35:16,174 --> 02:35:18,242 NOW BUT THEY MIGHT NEED TO COME 3000 02:35:18,242 --> 02:35:18,476 BACK. 3001 02:35:18,476 --> 02:35:20,945 AND THERE'S A WHOLE LOT OF 3002 02:35:20,945 --> 02:35:22,213 THINGS A.I. HAS NOT 3003 02:35:22,213 --> 02:35:23,614 INCORPORATED. 3004 02:35:23,614 --> 02:35:25,550 I MENTIONED DEVELOPMENTAL LEARN 3005 02:35:25,550 --> 02:35:30,188 AND THINGS AT OTHER SCALES WE 3006 02:35:30,188 --> 02:35:32,857 COULD TAKE FROM BIOLOGY NOT JUST 3007 02:35:32,857 --> 02:35:33,457 NEUROSCIENCE BUT BIOLOGY IN 3008 02:35:33,457 --> 02:35:36,727 GENERAL AND INCORPORATE INTO 3009 02:35:36,727 --> 02:35:37,428 A.I. 3010 02:35:37,428 --> 02:35:38,529 >> I'D LIKE TO ADD ANOTHER 3011 02:35:38,529 --> 02:35:41,199 PRINCIPLE THAT WAS LEARNED FROM 3012 02:35:41,199 --> 02:35:46,237 THE BRAIN NOW BEING USED IN A.I. 3013 02:35:46,237 --> 02:35:48,472 WHICH IS SELF-SUPERVISION AND 3014 02:35:48,472 --> 02:35:51,642 REWARD PREDICTION ERROR THAT CAME 3015 02:35:51,642 --> 02:35:53,477 FIRST FROM TEMPORAL DIFFERENCE 3016 02:35:53,477 --> 02:35:56,314 LEARNING IN BASAL GANGLIA AND 3017 02:35:56,314 --> 02:36:00,351 DOPAMINE AND THE IDEA THOUGH YOU 3018 02:36:00,351 --> 02:36:02,019 COULD LEARN A LOT OF BY 3019 02:36:02,019 --> 02:36:03,120 PREDICTING THE NEXT SENSORY 3020 02:36:03,120 --> 02:36:06,924 INPUT OR PREDICTING IF YOU MAKE 3021 02:36:06,924 --> 02:36:08,226 A MOTOR ACTION WHAT'S THE 3022 02:36:08,226 --> 02:36:11,629 CONSEQUENCE IN THE CEREBELLUM? 3023 02:36:11,629 --> 02:36:13,064 >> THEY'RE IMPORTANT IDEAS AND 3024 02:36:13,064 --> 02:36:14,665 WE WANT TO GET TO A ZOOM 3025 02:36:14,665 --> 02:36:16,801 QUESTION AND WE MAY GO A FEW 3026 02:36:16,801 --> 02:36:18,069 MINUTES LATE BEFORE WE SEND YOU 3027 02:36:18,069 --> 02:36:19,036 TO LUNCH BUT A ZOOM QUESTION 3028 02:36:19,036 --> 02:36:19,403 NOW. 3029 02:36:19,403 --> 02:36:21,172 THANK YOU. 3030 02:36:21,172 --> 02:36:23,574 >> THIS QUESTION COMES FROM 3031 02:36:23,574 --> 02:36:23,808 ONLINE. 3032 02:36:23,808 --> 02:36:25,810 AND THE QUESTION IS I WONDER HOW 3033 02:36:25,810 --> 02:36:28,079 THE DATA COLLECTION COULD BE 3034 02:36:28,079 --> 02:36:30,181 ORGANIZED BEST IF NOT LAB-BASED 3035 02:36:30,181 --> 02:36:32,850 AND HOW TO INCREASE NEURO 3036 02:36:32,850 --> 02:36:34,719 DATA INTEROPERABILITY AND MAKE 3037 02:36:34,719 --> 02:36:38,689 IT DESCRIBE A NATURAL STATE. 3038 02:36:38,689 --> 02:36:40,791 >> WHAT WAS WITH THE SECOND 3039 02:36:40,791 --> 02:36:43,194 QUESTION? 3040 02:36:43,194 --> 02:36:47,064 >> HOW TO INCREASE NEUROSCIENTIFIC 3041 02:36:47,064 --> 02:36:47,698 DATA INTRAOPERABILITY AND INCREASE 3042 02:36:47,698 --> 02:36:48,633 THE NATURAL STATE. 3043 02:36:48,633 --> 02:36:52,870 >> THE FIRST QUESTION WAS HOW TO 3044 02:36:52,870 --> 02:36:55,640 COLLECT DATA NOT BASED ON 3045 02:36:55,640 --> 02:36:59,210 INDIVIDUAL LABS. 3046 02:36:59,210 --> 02:37:00,745 THIS WOULD BE THE CREATION OF 3047 02:37:00,745 --> 02:37:02,213 NATIONAL LABS WOULD BE ONE WAY 3048 02:37:02,213 --> 02:37:04,415 IN THE U.S. WE HAVE 3049 02:37:04,415 --> 02:37:05,216 EXPERIENCED THIS AND THAT WOULD 3050 02:37:05,216 --> 02:37:11,255 BE ONE APPROACH. CREATE A 3051 02:37:11,255 --> 02:37:12,690 NATIONAL LAB FOR THE LARGE-SCALE 3052 02:37:12,690 --> 02:37:13,925 DATA COLLECTION EXPERIMENTS AND 3053 02:37:13,925 --> 02:37:15,893 DATA AND EVERYTHING WOULD BE 3054 02:37:15,893 --> 02:37:19,630 MANAGED AT THAT LEVEL. 3055 02:37:19,630 --> 02:37:20,698 AND THEN ORGANIZATIONAL SYSTEM 3056 02:37:20,698 --> 02:37:22,066 WHERE PEOPLE HAVE ACCESS TO THE 3057 02:37:22,066 --> 02:37:26,337 DATA AND MAYBE THEY PROPOSE 3058 02:37:26,337 --> 02:37:27,171 EXPERIMENTS AFTER THEY ANALYZE 3059 02:37:27,171 --> 02:37:29,674 AND RUN INSILICO SIMULATIONS. 3060 02:37:29,674 --> 02:37:35,479 I THINK THIS IS A WAY FORWARD 3061 02:37:35,479 --> 02:37:36,714 JUST NEEDS TO RETHINK THE WHOLE 3062 02:37:36,714 --> 02:37:37,915 INFRASTRUCTURE OF HOW WE DO 3063 02:37:37,915 --> 02:37:48,259 THINGS RIGHT NOW. 3064 02:37:48,259 --> 02:37:49,994 >> DO WE HAVE A QUICK QUESTION 3065 02:37:49,994 --> 02:37:51,395 IN THE ROOM? 3066 02:37:51,395 --> 02:37:53,097 >> YEAH, FOLLOWING UP ON THAT MY 3067 02:37:53,097 --> 02:37:56,534 QUESTION ORIENTED TOWARDS 3068 02:37:56,534 --> 02:37:57,868 DOMINIQUE AND DIMITRI 3069 02:37:57,868 --> 02:37:58,169 ESPECIALLY. 3070 02:37:58,169 --> 02:38:01,839 I'M CURIOUS WHEN THINKING ABOUT 3071 02:38:01,839 --> 02:38:05,109 THE MASSIVE HETEROGENEITY OF 3072 02:38:05,109 --> 02:38:07,478 DATA SETS AND INCONSISTENT DATA 3073 02:38:07,478 --> 02:38:12,083 FILE FORMATS, METADATA, SCHEMAS, 3074 02:38:12,083 --> 02:38:15,653 WHAT OPPORTUNITIES ARE THERE TO 3075 02:38:15,653 --> 02:38:18,689 LEVERAGE LARGE SCALE LLMs TO 3076 02:38:18,689 --> 02:38:21,759 HELP FACILITATE THAT 3077 02:38:21,759 --> 02:38:22,960 HARMONIZATION AND CURATION 3078 02:38:22,960 --> 02:38:24,128 PROCESS IN ORDER TO PULL THE 3079 02:38:24,128 --> 02:38:25,229 DATA INTO THE TYPES OF LARGE 3080 02:38:25,229 --> 02:38:33,237 MODELS WE'RE TALKING ABOUT. 3081 02:38:33,237 --> 02:38:36,440 >> YES, OPPORTUNITIES ARE THERE 3082 02:38:36,440 --> 02:38:40,444 FOR REUSING THE DATA. 3083 02:38:40,444 --> 02:38:44,882 WE'RE FOCUSING MORE ON THE 3084 02:38:44,882 --> 02:38:47,418 QUESTIONS OF THE APPLICATION 3085 02:38:47,418 --> 02:38:48,452 THAT'S EXISTING. THE MAIN 3086 02:38:48,452 --> 02:38:53,624 PURPOSE OF THIS DISCUSSION IS 3087 02:38:53,624 --> 02:38:59,630 ABOUT DESIGNING STUDIES, COST 3088 02:38:59,630 --> 02:39:03,300 FUNCTIONS TO LEARN AND EXTRACT THE 3089 02:39:03,300 --> 02:39:05,136 MAXIMUM AMOUNT OF INFORMATION 3090 02:39:05,136 --> 02:39:08,773 AND DESIGN CLOSED-LOOP 3091 02:39:08,773 --> 02:39:10,374 EXPERIMENTS AND MAKING SENSE OF 3092 02:39:10,374 --> 02:39:11,642 DIVERSE EXPERIMENTS. 3093 02:39:11,642 --> 02:39:13,911 IT'S A SLIGHTLY DIFFERENT 3094 02:39:13,911 --> 02:39:15,813 PROBLEM. 3095 02:39:15,813 --> 02:39:17,782 THAT'S MORE OF THE TECHNOLOGICAL 3096 02:39:17,782 --> 02:39:25,322 SIDE OF THINGS AND WORKING 3097 02:39:25,322 --> 02:39:28,192 ACROSS LABS AND YES EVERY LAB 3098 02:39:28,192 --> 02:39:32,630 UNLIKE OTHER RELATES DISCIPLINES 3099 02:39:32,630 --> 02:39:34,632 LIKE BIOINFORMATICS, NEUROSCIENCE 3100 02:39:34,632 --> 02:39:39,637 WILL LIKELY NOT CONVERGE TO A 3101 02:39:39,637 --> 02:39:43,574 SET OF STANDARD EXPERIMENTS. 3102 02:39:43,574 --> 02:39:44,108 WE'LL ALWAYS BE EXPLORING DIVERSE 3103 02:39:44,108 --> 02:39:52,149 PARADIGMS FOR EXPERIMENTS. 3104 02:39:52,149 --> 02:39:52,650 THESE APPROACHES BECOME 3105 02:39:52,650 --> 02:40:00,891 NECESSARY. 3106 02:40:00,891 --> 02:40:02,159 >> THAT'S A REAL CHALLENGE 3107 02:40:02,159 --> 02:40:04,829 IN DABI PEOPLE HAD THEIR OWN 3108 02:40:04,829 --> 02:40:07,131 WAYS OF USING FORMATS AND 3109 02:40:07,131 --> 02:40:11,635 STANDARDS AND NOT EVERYONE WANTS 3110 02:40:11,635 --> 02:40:13,437 TO USE BIDS FOR EXAMPLE. 3111 02:40:13,437 --> 02:40:18,075 IT'S A HUGE EFFORT TO CONVERT 3112 02:40:18,075 --> 02:40:18,943 THEIR DATA INTO BIDS. 3113 02:40:18,943 --> 02:40:20,077 IT'S A TIME AND RESOURCES AND 3114 02:40:20,077 --> 02:40:25,483 THAT MAY NOT BE AVAILABLE. 3115 02:40:25,483 --> 02:40:31,155 FINDING WAYS TO HELP IN THAT 3116 02:40:31,155 --> 02:40:33,791 CONVERSION AND HAVE CONVERTERS 3117 02:40:33,791 --> 02:40:35,659 THAT MAKE IT EASIER TO CONVERT 3118 02:40:35,659 --> 02:40:41,265 TO A STANDARDIZED FORMAT AND 3119 02:40:41,265 --> 02:40:51,775 STRUCTURE WOULD BE BENEFICIAL. 3120 02:40:54,078 --> 02:40:56,847 >> AL XIN, I'M A TRAINEE AT THE 3121 02:40:56,847 --> 02:41:00,751 MACHINE LEARNING CORE AT NIMH 3122 02:41:00,751 --> 02:41:04,288 SEVERAL PRESENTERS MENTIONED THE 3123 02:41:04,288 --> 02:41:05,789 DESIDERATA OF NATURAL INTELLIGENCE 3124 02:41:05,789 --> 02:41:07,091 REQUIRE LESS TRAINING ON LESS 3125 02:41:07,091 --> 02:41:09,593 DATA BUT I WANTED TO INTERROGATE 3126 02:41:09,593 --> 02:41:12,863 THE ROBUSTNESS OF THIS CLAIM. 3127 02:41:12,863 --> 02:41:14,164 IN TERMS OF DATA NATURAL 3128 02:41:14,164 --> 02:41:14,798 INTELLIGENCE IS TRAINED ON. IT'S 3129 02:41:14,798 --> 02:41:18,502 VERY DIFFERENT FROM THE TYPE OF 3130 02:41:18,502 --> 02:41:20,571 DATA SETS WE CAN PRODUCE IN TERMS 3131 02:41:20,571 --> 02:41:22,439 OF TEXT DATA AND IMAGE DATA AND 3132 02:41:22,439 --> 02:41:24,074 EVEN VIDEO DATA. I WAS WONDERING 3133 02:41:24,074 --> 02:41:25,976 IF ANYONE CAN COMMENT ON EFFORTS 3134 02:41:25,976 --> 02:41:27,745 TO CREATE DATA SETS THAT CAPTURE 3135 02:41:27,745 --> 02:41:31,181 THE COMPLEXITY OF NATURAL 3136 02:41:31,181 --> 02:41:36,287 INTELLIGENT TRAINED DATA IN ORDER 3137 02:41:36,287 --> 02:41:42,459 TO CREATE MODELS THAT HAVE THESE 3138 02:41:42,459 --> 02:41:46,263 CHARACTERISTICS. HOW ROBUST IS THE 3139 02:41:46,263 --> 02:41:48,098 CLAIM THAT IT REQUIRES LESS 3140 02:41:48,098 --> 02:41:49,300 TRAINING TIME ON LESS DATA? 3141 02:41:49,300 --> 02:41:51,669 IS THERE EFFORTS TO CURATE 3142 02:41:51,669 --> 02:41:55,472 NATURALLY INTELLIGENT DATA SETS? 3143 02:41:55,472 --> 02:42:00,110 >> YEAH, THERE'S A DATA SET 3144 02:42:00,110 --> 02:42:03,347 COLLECTED BY PEOPLE AT 3145 02:42:03,347 --> 02:42:05,182 STANFORD WHERE THEY PUT 3146 02:42:05,182 --> 02:42:07,751 HEAD CAMERAS IN CHILDREN AND 3147 02:42:07,751 --> 02:42:11,255 COLLECTED THIS KIND OF DATA. 3148 02:42:11,255 --> 02:42:14,158 WHERE BASICALLY YOU COLLECT HOW 3149 02:42:14,158 --> 02:42:15,125 CHILDREN INTERACT WITH THE WORLD 3150 02:42:15,125 --> 02:42:19,430 AND USE THE DATA TO TRAIN A.I. 3151 02:42:19,430 --> 02:42:19,663 SYSTEMS. 3152 02:42:19,663 --> 02:42:21,231 YOU RAISE A VERY GOOD QUESTION. 3153 02:42:21,231 --> 02:42:24,468 I THINK IT'S MOSTLY LIKE WHEN 3154 02:42:24,468 --> 02:42:26,070 PEOPLE SAY THAT WE GET TRAINED 3155 02:42:26,070 --> 02:42:27,738 ON THE DATA, A LOT OF IT HAS TO 3156 02:42:27,738 --> 02:42:28,872 DO WITH LANGUAGE. 3157 02:42:28,872 --> 02:42:34,111 THAT'S A CLEAR CASE. 3158 02:42:34,111 --> 02:42:44,655 LIKE NOBODY AND THERE'S EVIDENCE 3159 02:42:48,258 --> 02:42:50,494 WE HUMANS ARE MUCH BETTER AT 3160 02:42:50,494 --> 02:42:52,463 GENERALIZING OUR TRAINING 3161 02:42:52,463 --> 02:42:53,797 DISTRIBUTION WHICH IS THE HALL- 3162 02:42:53,797 --> 02:42:58,936 MARK OF INTELLIGENCE IN MANY 3163 02:42:58,936 --> 02:42:59,770 WAYS. AND THERE'S PSYCHOPHYSICAL 3164 02:42:59,770 --> 02:43:01,238 AND BEHAVIORAL EVIDENCE FOR 3165 02:43:01,238 --> 02:43:01,438 THAT. 3166 02:43:01,438 --> 02:43:03,574 >> THAT'S GREAT. 3167 02:43:03,574 --> 02:43:04,508 I THINK WE'VE GOT A LOT OF 3168 02:43:04,508 --> 02:43:06,543 MATERIAL TO CONTINUE OUR 3169 02:43:06,543 --> 02:43:07,478 CONVERSATIONS OVER LUNCH AND 3170 02:43:07,478 --> 02:43:08,345 OVER ALL THE REMAINING SESSIONS 3171 02:43:08,345 --> 02:43:09,213 OF THE WORKSHOP. 3172 02:43:09,213 --> 02:43:11,348 SO ONCE AGAIN I WANT TO THANK 3173 02:43:11,348 --> 02:43:13,584 ALL OF OUR SPEAKERS AND INVITED 3174 02:43:13,584 --> 02:43:14,785 DISCUSSANTS AND ALL THE 3175 02:43:14,785 --> 02:43:16,420 PANELISTS AND AUDIENCE MEMBERS 3176 02:43:16,420 --> 02:43:18,689 IN THE ROOM AND WATCHING ON ZOOM 3177 02:43:18,689 --> 02:43:21,158 AND VIDEOCAST. 3178 02:43:21,158 --> 02:43:22,526 PANELISTS AND ORGANIZERS, PLEASE 3179 02:43:22,526 --> 02:43:25,963 STAND BY FOR A GROUP PHOTO BY 3180 02:43:25,963 --> 02:43:27,831 THE STAIRWELL AFTERWARDS BUT 3181 02:43:27,831 --> 02:43:28,899 EVERYONE ELSE IS FREE TO FIND LUNCH. 3182 02:43:28,899 --> 02:43:32,236 THANK YOU SO MUCH. 3183 02:43:32,236 --> 02:43:32,736 All right. 3184 02:43:32,736 --> 02:43:33,904 It's one 115. 3185 02:43:33,904 --> 02:43:38,575 So I'd like to welcome everyone back to the first program 3186 02:43:38,609 --> 02:43:41,612 in the afternoon of the NEURO AI workshop. 3187 02:43:44,081 --> 02:43:46,617 We're delighted to have gathered a bunch 3188 02:43:46,617 --> 02:43:49,920 of different types of funders, investors and sponsors. 3189 02:43:50,721 --> 02:43:53,057 And many of you may have noticed the carousel slides 3190 02:43:53,057 --> 02:43:56,060 earlier with lots of funding opportunities. 3191 02:43:56,126 --> 02:44:00,497 So, we will not repeat those slides 3192 02:44:00,631 --> 02:44:04,134 unless one of our panelists need to reference to them. 3193 02:44:04,134 --> 02:44:06,904 During the course of this discussion. 3194 02:44:06,904 --> 02:44:08,739 So let me introduce myself. 3195 02:44:08,739 --> 02:44:10,174 I am Grace Hwang. 3196 02:44:10,174 --> 02:44:11,975 I am a Program Director 3197 02:44:11,975 --> 02:44:13,243 at the National Institute 3198 02:44:13,243 --> 02:44:15,245 of Neurological Disorders and Stroke, 3199 02:44:15,245 --> 02:44:17,147 and I support the BRAIN Initiative, 3200 02:44:17,147 --> 02:44:20,417 where I manage a portfolio of projects on neuromodulation 3201 02:44:20,417 --> 02:44:23,420 recording and stimulation technologies. 3202 02:44:23,754 --> 02:44:26,824 And I'm honored to introduce our moderator 3203 02:44:26,824 --> 02:44:31,362 today, Doctor TERRENCE SEJNOWSKI 3204 02:44:31,362 --> 02:44:33,964 He is a professor and laboratory 3205 02:44:33,964 --> 02:44:38,235 head of the Computational Neurobiology Laboratory. 3206 02:44:38,235 --> 02:44:39,603 HE IS A PIONEER IN 3207 02:44:39,603 --> 02:44:42,673 COMPUTATIONAL NEUROSCIENCE, HE HAS MANY ACCOLADES 3208 02:44:42,673 --> 02:44:45,576 that you can read about IN THE PROGRAM BOOK. 3209 02:44:45,576 --> 02:44:48,946 HIS MOST RECENT RECOGNITION IS the 2024 3210 02:44:48,946 --> 02:44:51,815 LUNDBEck FOUNDATION’s BRAIN PRIZE FOR HIS 3211 02:44:51,815 --> 02:44:55,786 CONTRIBUTIONS TO THEORETICAL AND COMPUTATIONAL NEUROSCIENCE. 3212 02:44:55,786 --> 02:45:00,357 HIS COLLABORATIVE WORK WITH GEOFF HINTON ON LEARNING ALGORITHMS 3213 02:45:00,357 --> 02:45:04,261 FOR BOLTZMANN MACHINES FROM 1985 WHICH WAS FEATURED 3214 02:45:04,261 --> 02:45:07,865 IN TONY ZADOR’S TALK EARLIER TODAY, ALSO SUBSTANTIALLY 3215 02:45:07,865 --> 02:45:10,234 CONTRIBUTED TO THE RECENT NOBEL PRIZE IN PHYSICS 3216 02:45:10,234 --> 02:45:13,137 FOR FOUNDATIONAL DISCOVERIES AND INVENTIONS THAT 3217 02:45:13,137 --> 02:45:16,373 ENABLE MACHINE LEARNING WITH ARTIFICIAL NEURAL NETWORKS. 3218 02:45:19,076 --> 02:45:21,111 I'M GOING TO GO OVER MINOR 3219 02:45:21,111 --> 02:45:23,113 HOUSEKEEPING HERE. 3220 02:45:27,918 --> 02:45:29,987 AND LET THE AUDIENCE KNOW WE'LL 3221 02:45:29,987 --> 02:45:31,955 FIRST DO INTRODUCTIONS AND GO TO 3222 02:45:31,955 --> 02:45:33,724 THE FUNDER PANEL GUIDING 3223 02:45:33,724 --> 02:45:35,492 QUESTIONS AND DISCUSSION SLIDE. 3224 02:45:35,492 --> 02:45:36,794 ANYBODY WHO WANTS TO SEND THEIR 3225 02:45:36,794 --> 02:45:39,062 RESPONSE IS WELCOME TO SCAN THE 3226 02:45:39,062 --> 02:45:41,165 QR CODE, WHETHER YOU'RE WATCHING 3227 02:45:41,165 --> 02:45:42,699 THIS PROGRAM LIVE OR AS A 3228 02:45:42,699 --> 02:45:45,102 RECORDING BECAUSE WE WILL BE 3229 02:45:45,102 --> 02:45:46,436 MONITORING RESPONSES IN E-MAIL 3230 02:45:46,503 --> 02:45:49,039 FOR SOME WEEKS TO COME. 3231 02:45:49,039 --> 02:45:52,576 AND WITH THAT, TERRY, WHY DON'T 3232 02:45:52,576 --> 02:45:55,612 WE GO BACK TO YOU. 3233 02:45:55,612 --> 02:45:57,114 >> I'D LIKE TO BEGIN BY ASKING 3234 02:45:57,114 --> 02:45:58,515 ALL OF THE PANEL DISCUSSANTS TO 3235 02:45:58,515 --> 02:45:59,583 INTRODUCE THEMSELVES. 3236 02:45:59,583 --> 02:46:02,119 NAME, RANK, SERIAL NUMBER, SO WE 3237 02:46:02,119 --> 02:46:07,491 DON'T TAKE TOO MUCH TIME FROM 3238 02:46:07,491 --> 02:46:08,292 OUR DISCUSSION. 3239 02:46:08,292 --> 02:46:15,933 >> I CAN HEAR MY VOICE. 3240 02:46:15,933 --> 02:46:20,804 I'M STEVEN ZENDER, NATIONAL 3241 02:46:20,804 --> 02:46:22,973 SCIENCE FOUNDATION. 3242 02:46:22,973 --> 02:46:24,975 >> I'M ALYSSA SCHAFER, SENIOR 3243 02:46:24,975 --> 02:46:26,076 SCIENTIST AND VICE PRESIDENT, 3244 02:46:26,076 --> 02:46:29,546 SIMONS FOUNDATION, WHERE I 3245 02:46:29,546 --> 02:46:32,182 DIRECT THE NEUROSCIENCE 3246 02:46:32,182 --> 02:46:33,083 COLLABORATION. 3247 02:46:33,083 --> 02:46:35,152 >> GOOD AFTERNOON, STEPHANIE 3248 02:46:35,152 --> 02:46:35,686 GAGE, ASSOCIATE PROGRAM 3249 02:46:35,686 --> 02:46:36,253 DIRECTOR, NATIONAL SCIENCE 3250 02:46:36,253 --> 02:46:38,121 FOUNDATION, WORK IN THE 3251 02:46:38,121 --> 02:46:40,624 COMPUTER SCIENCE DIRECTORATE IN 3252 02:46:40,624 --> 02:46:42,726 THE COMPUTING AND COMMUNICATIONS 3253 02:46:42,726 --> 02:46:52,903 FOUNDATIONS DIVISION. 3254 02:46:52,903 --> 02:46:53,170 >> OKAY. 3255 02:46:53,170 --> 02:46:54,238 THANK YOU. 3256 02:46:54,238 --> 02:47:01,979 GOOD AFTERNOON, ROBINSON PINO, 3257 02:47:01,979 --> 02:47:05,048 PROGRAM DIRECTOR, DEPARTMENT OF 3258 02:47:05,048 --> 02:47:05,282 ENERGY. 3259 02:47:05,282 --> 02:47:07,918 THANK YOU. 3260 02:47:07,918 --> 02:47:08,485 >> HAL GREENWALD, PROGRAM 3261 02:47:08,485 --> 02:47:13,056 OFFICER AT THE AIR FORCE OFFICE 3262 02:47:13,056 --> 02:47:14,024 OF SCIENTIFIC RESEARCH, WHICH IS 3263 02:47:14,024 --> 02:47:15,125 AIR FORCE'S RESEARCH FUNDING ARM. 3264 02:47:15,125 --> 02:47:16,994 I FUND HIGH RISK HIGH IMPACT 3265 02:47:16,994 --> 02:47:19,596 RESEARCH IN THE COGNITIVE AND 3266 02:47:19,596 --> 02:47:20,497 COMPUTATIONAL NEUROSCIENCE 3267 02:47:20,497 --> 02:47:23,367 PROGRAM WHICH FUNDS WORK ON THE 3268 02:47:23,367 --> 02:47:25,636 MECHANISMS OF PERCEPTION, 3269 02:47:25,636 --> 02:47:27,738 COGNITION, BEHAVIOR, ALSO AT THE 3270 02:47:27,738 --> 02:47:30,340 INTERSECTION OF NEUROSCIENCE AND 3271 02:47:30,340 --> 02:47:32,209 A.I., I ALSO MANAGE THE 3272 02:47:32,209 --> 02:47:36,947 COMPUTATIONAL COGNITION PROGRAM, 3273 02:47:36,947 --> 02:47:37,381 AN A.I. PROGRAM. 3274 02:47:37,381 --> 02:47:38,882 >> GOOD AFTERNOON, THANK YOU, 3275 02:47:38,882 --> 02:47:39,783 GRACE, FOR THE INVITATION AND 3276 02:47:39,783 --> 02:47:42,853 OPPORTUNITY TO BE HERE. 3277 02:47:42,853 --> 02:47:43,954 I'M CHRISTINE EDWARDS, NATIONAL 3278 02:47:43,954 --> 02:47:46,390 SECURITY AGENCY, I'VE BEEN THERE 3279 02:47:46,390 --> 02:47:48,492 FOR ABOUT 30 YEARS. 3280 02:47:48,492 --> 02:47:52,429 MY MOST CURRENT ROLE IS DARPA, 3281 02:47:52,429 --> 02:47:53,397 I'M A NSA REPRESENTATIVE AT DARPA, 3282 02:47:53,397 --> 02:47:55,832 A NEW POSITION FOR THEM SO IT'S A 3283 02:47:55,832 --> 02:47:56,700 LIAISON TYPE ROLE. 3284 02:47:56,700 --> 02:48:00,070 SO I GET TO WORK WITH PEOPLE 3285 02:48:00,070 --> 02:48:01,338 ACROSS THE DIFFERENT DARPA 3286 02:48:01,338 --> 02:48:02,506 TECHNICAL OFFICES, THERE'S SIX 3287 02:48:02,506 --> 02:48:03,707 OF THEM TOTAL. 3288 02:48:03,707 --> 02:48:06,009 AND I SIT WITHIN THE INFORMATION 3289 02:48:06,009 --> 02:48:07,344 INNOVATION OFFICE. 3290 02:48:07,344 --> 02:48:09,746 I LOOK FORWARD TO TALKING WITH 3291 02:48:09,746 --> 02:48:11,381 PEOPLE DURING THIS SESSION BUT 3292 02:48:11,381 --> 02:48:15,886 ALSO AFTERWARDS. THERE'S A TABLE 3293 02:48:15,886 --> 02:48:21,925 WITH HANDOUTS IF YOU'RE 3294 02:48:21,925 --> 02:48:22,326 INTERESTED. 3295 02:48:22,326 --> 02:48:24,861 >> I'M CHOU HUNG, PROGRAM 3296 02:48:24,861 --> 02:48:25,862 MANAGER WITH ARMY RESEARCH OFFICE. 3297 02:48:25,862 --> 02:48:27,965 EXTRAMURAL ARM OF ARMY RESEARCH 3298 02:48:27,965 --> 02:48:30,267 LABORATORY. MY FUNDING TOPIC AREA 3299 02:48:30,267 --> 02:48:31,468 IS NEUROPHYSIOLOGY OF COGNITION, 3300 02:48:31,468 --> 02:48:35,739 AREAS THAT COVER INCLUDES 3301 02:48:35,739 --> 02:48:36,606 EVOLUTIONARY BIOLOGY, 3302 02:48:36,606 --> 02:48:40,777 INTERACTIONS WITH REAL AND 3303 02:48:40,777 --> 02:48:42,179 AUGMENTED WORLD, BIO-INSPIRED 3304 02:48:42,179 --> 02:48:43,046 AND NEUROMORPHIC COMPUTING. 3305 02:48:43,046 --> 02:48:43,947 THANK YOU. 3306 02:48:43,947 --> 02:48:46,683 >> DAVID WYATT, THIS IS MY 3307 02:48:46,683 --> 02:48:48,885 SECOND NIH CONFERENCE. 3308 02:48:48,885 --> 02:48:51,288 I'M A RECOVERING ENGINEER, SO I 3309 02:48:51,288 --> 02:48:55,225 USED TO WORK IN SILICON VALLEY, 3310 02:48:55,225 --> 02:48:55,792 AT NVIDIA. 3311 02:48:55,792 --> 02:48:58,295 AND THESE DAYS I'M RUNNING THE 3312 02:48:58,295 --> 02:49:02,366 OPENGPU FOUNDATION AS WELL AS 3313 02:49:02,366 --> 02:49:05,669 OTHER START-UPS IN AUSTIN WHICH 3314 02:49:05,669 --> 02:49:07,871 IS NOW SILICON HILLS. 3315 02:49:07,871 --> 02:49:10,140 >> WE'RE LIVING IN EXTRAORDINARY 3316 02:49:10,140 --> 02:49:10,707 TIMES. 3317 02:49:10,707 --> 02:49:16,079 EVOLUTION WAS BROUGHT UP IN THE 3318 02:49:16,079 --> 02:49:17,247 FIRST SESSION. 3319 02:49:17,247 --> 02:49:19,249 AND JUDGING BY THE TIME SCALE OF 3320 02:49:19,249 --> 02:49:20,017 EVOLUTION, WHAT WE'RE 3321 02:49:20,017 --> 02:49:21,351 EXPERIENCING RIGHT NOW OVER THE 3322 02:49:21,351 --> 02:49:24,421 LAST DECADE IS LIKE AN 3323 02:49:24,421 --> 02:49:24,855 EXPLOSION. 3324 02:49:24,855 --> 02:49:29,659 AND EVEN JUDGING BY THE TIME 3325 02:49:29,659 --> 02:49:32,095 SCALE OF CULTURAL CHANGE, THE 3326 02:49:32,095 --> 02:49:34,364 INDUSTRIAL REVOLUTION TOOK 3327 02:49:34,364 --> 02:49:40,404 HUNDREDS OF YEARS TO PERMEATE 3328 02:49:40,404 --> 02:49:42,272 THROUGH SOCIETY, AND NOW A.I. IS 3329 02:49:42,272 --> 02:49:45,642 SPREADING ON THE TIME SCALE OF 3330 02:49:45,642 --> 02:49:46,643 DECADES. 3331 02:49:46,643 --> 02:49:49,813 SO, THIS IS WITHIN, YOU KNOW, 3332 02:49:49,813 --> 02:49:52,449 LIFETIMES OF PEOPLE WHO ARE 3333 02:49:52,449 --> 02:49:58,021 GOING TO BE BOTH USERS AND 3334 02:49:58,021 --> 02:49:59,456 DEVELOPERS SO THIS IS AN 3335 02:49:59,456 --> 02:50:00,090 EXTRAORDINARY GROUP OF FUNDERS 3336 02:50:00,090 --> 02:50:00,524 WE HAVE. 3337 02:50:00,524 --> 02:50:03,293 WE HAVE THE CHANCE TO FIND OUT A 3338 02:50:03,293 --> 02:50:04,027 LITTLE BIT ABOUT WHAT 3339 02:50:04,027 --> 02:50:06,696 OPPORTUNITIES ARE GOING TO BE 3340 02:50:06,696 --> 02:50:08,098 LIKE IN THE FUTURE. 3341 02:50:08,098 --> 02:50:10,834 I'LL START BY A QUESTION, WHY IS 3342 02:50:10,834 --> 02:50:14,471 YOUR AGENCY OR ORGANIZATION 3343 02:50:14,471 --> 02:50:15,439 INTERESTED IN NEUROAI? 3344 02:50:15,439 --> 02:50:18,608 ANYONE WANT TO TAKE THAT 3345 02:50:18,608 --> 02:50:22,913 QUESTION? 3346 02:50:22,913 --> 02:50:24,114 ANYBODY CAN JUMP IN. 3347 02:50:24,114 --> 02:50:26,416 >> YOU KNOW, WHY NOT? 3348 02:50:26,416 --> 02:50:28,852 OH, GO AHEAD. 3349 02:50:28,852 --> 02:50:29,619 >> THANK YOU. 3350 02:50:29,619 --> 02:50:31,455 SO, I'VE BEEN AT DARPA FOR ABOUT 3351 02:50:31,455 --> 02:50:35,092 A YEAR SO FAR, TWO MORE TO GO AT 3352 02:50:35,092 --> 02:50:35,392 LEAST. 3353 02:50:35,392 --> 02:50:40,430 BUT DARPA AS MANY OF YOU KNOW IS 3354 02:50:40,430 --> 02:50:42,966 ABOUT HIGH-RISK HIGH-PAYOFF 3355 02:50:42,966 --> 02:50:43,834 RESEARCH. 3356 02:50:43,834 --> 02:50:44,935 I'VE BEEN IN NEUROSCIENCE AND 3357 02:50:44,935 --> 02:50:47,671 A.I. FOR DECADES NOW, THIS IS A 3358 02:50:47,671 --> 02:50:50,640 REALLY EXCITING TIME, AS YOU 3359 02:50:50,640 --> 02:50:50,974 STATED. 3360 02:50:50,974 --> 02:50:55,445 AND SO DARPA IS LOOKING AT THE 3361 02:50:55,445 --> 02:50:57,314 BIG IDEAL THAT ARE COMING. 3362 02:50:57,314 --> 02:50:59,049 MY PERSONAL PERSPECTIVE, I THINK 3363 02:50:59,049 --> 02:51:01,485 OTHERS AT DARPA SHARE THAT, 3364 02:51:01,485 --> 02:51:03,787 GROWING MOMENTUM IN THE FACE OF 3365 02:51:03,787 --> 02:51:09,259 NEUROSCIENCE AND A.I., 3366 02:51:09,259 --> 02:51:12,329 NEUROMORPHIC COMPUTING, 3367 02:51:12,329 --> 02:51:13,997 HUMAN-MACHINE COLLABORATION, 3368 02:51:13,997 --> 02:51:16,833 THAT TIGHTLY COUPLED LOOP. THE 3369 02:51:16,833 --> 02:51:20,637 OODA LOOP, AND THE A.I. POWERED 3370 02:51:20,637 --> 02:51:26,476 OODA LOOP. AND HOW DO YOU DO THAT 3371 02:51:26,476 --> 02:51:27,244 IN A TRUSTWORTHY WAY. 3372 02:51:27,244 --> 02:51:28,879 TO GET TO NEXTGEN A.I. WE CAN 3373 02:51:28,879 --> 02:51:32,749 BENEFIT BY UNDERSTANDING HOW 3374 02:51:32,749 --> 02:51:33,750 BRAINS WORK, NEUROAI IS PERFECT 3375 02:51:33,750 --> 02:51:36,453 FOR THAT. SCALING IS GREAT, BUT 3376 02:51:36,453 --> 02:51:38,755 CAN ONLY GO SO FAR. SO WHAT'S 3377 02:51:38,755 --> 02:51:39,623 BEYOND THAT? ALGORITHMS, ARCHI- 3378 02:51:39,623 --> 02:51:41,625 TECTURE, HARDWARE, THE FULLSTACK. 3379 02:51:41,625 --> 02:51:44,227 >> WHICH PARTS OF DARPA? 3380 02:51:44,227 --> 02:51:48,899 >> SO, I SIT WITHIN I2O, CARES 3381 02:51:48,899 --> 02:51:52,969 ABOUT NEXT GENERATION A.I., 3382 02:51:52,969 --> 02:51:54,304 CYBERSECURITY APPLICATIONS, 3383 02:51:54,304 --> 02:51:54,938 RESILIENCY. 3384 02:51:54,938 --> 02:51:56,373 THAT'S ANOTHER THING THAT WE'RE 3385 02:51:56,373 --> 02:51:58,775 THINKING ABOUT AS WELL AS THE 3386 02:51:58,775 --> 02:52:00,310 SECURITY OF LET'S SAY DEVICES 3387 02:52:00,310 --> 02:52:02,312 THAT ARE BASED ON NEUROMORPHIC 3388 02:52:02,312 --> 02:52:04,281 HARDWARE, WE SHOULD BE THINKING 3389 02:52:04,281 --> 02:52:06,883 ABOUT ARE THERE BENEFITS FROM A 3390 02:52:06,883 --> 02:52:08,318 SECURITY PERSPECTIVE TO MAKE 3391 02:52:08,318 --> 02:52:09,653 THINGS PRIVATE AND MORE SECURE 3392 02:52:09,653 --> 02:52:10,287 GOING FORWARD. 3393 02:52:10,287 --> 02:52:11,821 I THINK NOW IS THE TIME TO THINK 3394 02:52:11,821 --> 02:52:14,891 ABOUT THAT WHEN WE'RE THINKING 3395 02:52:14,891 --> 02:52:16,760 ABOUT NEUROMORPHIC COMPUTE 3396 02:52:16,760 --> 02:52:16,960 STACK. 3397 02:52:16,960 --> 02:52:19,596 AND THEN MTO IS THE MICROSYSTEMS 3398 02:52:19,596 --> 02:52:22,766 TECHNOLOGY OFFICE, THEY THINK 3399 02:52:22,766 --> 02:52:26,069 ABOUT DIFFERENT TYPES OF 3400 02:52:26,069 --> 02:52:30,807 HARDWARE, THAT INCLUDES QUANTUM, 3401 02:52:30,807 --> 02:52:33,610 CAN INCLUDE BIO, NEUROMORPHIC 3402 02:52:33,610 --> 02:52:35,045 COMPUTING, ANALOG, NEURAL 3403 02:52:35,045 --> 02:52:36,346 NETWORK HARDWARE. 3404 02:52:36,346 --> 02:52:37,847 >> IT'S ACROSS MANY -- 3405 02:52:37,847 --> 02:52:40,550 >> YEAH, AND THERE'S BTO, 3406 02:52:40,550 --> 02:52:41,418 BIOLOGICAL TECHNOLOGY OFFICE, A 3407 02:52:41,418 --> 02:52:42,619 NATURAL FIT. 3408 02:52:42,619 --> 02:52:43,386 >> GREAT. 3409 02:52:43,386 --> 02:52:48,024 HAL WAS GOING TO RESPOND TOO. 3410 02:52:48,024 --> 02:52:48,291 >> SURE. 3411 02:52:48,291 --> 02:52:50,327 SO, IN GENERAL, SO THE FIRST 3412 02:52:50,327 --> 02:52:53,163 QUESTION IS WHY IS THE AIR FORCE 3413 02:52:53,163 --> 02:52:53,830 INTERESTED IN NEUROSCIENCE. 3414 02:52:53,830 --> 02:52:55,131 AND THERE ARE A NUMBER OF 3415 02:52:55,131 --> 02:52:55,565 REASONS FOR THAT. 3416 02:52:55,565 --> 02:52:56,766 WHEN YOU THINK ABOUT THE AIR 3417 02:52:56,766 --> 02:52:59,836 FORCE, YOU MAY THINK FIRST ABOUT 3418 02:52:59,836 --> 02:53:01,271 AIRCRAFT BUT THE AIR FORCE 3419 02:53:01,271 --> 02:53:02,772 REALLY DEPENDS FIRST AND 3420 02:53:02,772 --> 02:53:04,341 FOREMOST ON ITS PEOPLE, NOT JUST 3421 02:53:04,341 --> 02:53:06,776 TO PILOT THE AIRCRAFT CURRENTLY 3422 02:53:06,776 --> 02:53:09,813 BUT FOR ALL THE OTHER 3423 02:53:09,813 --> 02:53:11,581 OPERATIONS, THE OTHER ASPECTS OF 3424 02:53:11,581 --> 02:53:13,016 OPERATIONS THE AIR FORCE AND NOW 3425 02:53:13,016 --> 02:53:15,285 THE SPACE FORCE ARE INVOLVED IN. 3426 02:53:15,285 --> 02:53:18,822 SO ONE OBJECTIVE IS TO 3427 02:53:18,822 --> 02:53:21,424 UNDERSTAND HOW HUMANS PROCESS 3428 02:53:21,424 --> 02:53:24,494 THE INFORMATION AROUND US, TO 3429 02:53:24,494 --> 02:53:25,662 MAKE HUMANSS SMARTER, MORE 3430 02:53:25,662 --> 02:53:27,330 CAPABLE, FITS WITH THE IDEAS TO 3431 02:53:27,330 --> 02:53:28,899 MAKE COMPUTERS SMARTER AND MORE 3432 02:53:28,899 --> 02:53:29,332 CAPABLE. 3433 02:53:29,332 --> 02:53:32,302 AND AS THE SPEAKERS THIS MORNING 3434 02:53:32,302 --> 02:53:35,805 DISCUSSED, IT MAKES SENSE TO 3435 02:53:35,805 --> 02:53:40,277 HAVE RESEARCH IN HUMAN -- 3436 02:53:40,277 --> 02:53:44,014 NEUROSCIENCE, WHETHER HUMANS OR 3437 02:53:44,014 --> 02:53:48,051 ANIMAL MODELS, COMPUTATIONAL 3438 02:53:48,051 --> 02:53:49,686 MODELS AS WELL, INVESTMENTS IN 3439 02:53:49,686 --> 02:53:52,088 A.I. LOOKING AT PRINCIPLES FROM 3440 02:53:52,088 --> 02:53:55,258 THE BRAIN, COMPUTATIONS THE 3441 02:53:55,258 --> 02:53:56,993 BRAIN PERFORMS, TO HAVE NICE 3442 02:53:56,993 --> 02:53:59,029 CYCLES BETWEEN THEM SO 3443 02:53:59,029 --> 02:54:01,398 PREDICTIONS FROM ONE PART CAN BE 3444 02:54:01,398 --> 02:54:04,367 TESTED IN THE OTHER. 3445 02:54:04,367 --> 02:54:05,669 >> GO AHEAD. 3446 02:54:05,669 --> 02:54:07,304 >> THANK YOU, TERRY. 3447 02:54:07,304 --> 02:54:08,738 THE SIMONS FOUNDATION'S MISSION 3448 02:54:08,738 --> 02:54:10,940 IS TO ADVANCE THE FRONTIERS OF 3449 02:54:10,940 --> 02:54:12,842 RESEARCH IN MATHEMATICS AND THE 3450 02:54:12,842 --> 02:54:13,777 BASIC SCIENCES. 3451 02:54:13,777 --> 02:54:16,046 I THINK OUR INTEREST IN A.I., 3452 02:54:16,046 --> 02:54:18,014 LIKE YOU GUYS AT DARPA, LIES 3453 02:54:18,014 --> 02:54:20,984 ACROSS A BUNCH OF OUR DIVISIONS. 3454 02:54:20,984 --> 02:54:23,853 SO WE HAVE AN INTRAMURAL 3455 02:54:23,853 --> 02:54:24,554 DIVISION, THE FLATIRON 3456 02:54:24,554 --> 02:54:27,057 INSTITUTE, WHICH IS AN IN-HOUSE 3457 02:54:27,057 --> 02:54:28,458 COMPUTATIONAL RESEARCH INSTITUTE 3458 02:54:28,458 --> 02:54:32,529 THAT HOUSES OUR FIVE CENTERS, 3459 02:54:32,529 --> 02:54:35,065 COMPUTATIONAL NEUROSCIENCE, 3460 02:54:35,065 --> 02:54:37,834 COMPUTATIONAL BIOLOGY, 3461 02:54:37,834 --> 02:54:38,702 COMPUTATIONAL MATHEMATICS, QUANTUM 3462 02:54:38,702 --> 02:54:39,235 CHEMISTRY, ASTROPHYSICS 3463 02:54:39,235 --> 02:54:45,041 THEN ON THE EXTRAMURAL SIDE, OUR 3464 02:54:45,041 --> 02:54:46,309 OUTWARD-FACING GRANTING 3465 02:54:46,309 --> 02:54:49,412 DIVISION, I SIT IN THE 3466 02:54:49,412 --> 02:54:50,180 NEUROSCIENCE DIVISION, OUR 3467 02:54:50,180 --> 02:54:51,648 NEUROSCIENCE COLLABORATIONS THAT 3468 02:54:51,648 --> 02:54:54,150 I DIRECT ARE FOCUSED ON 3469 02:54:54,150 --> 02:54:54,617 UNCOVERING FUNDAMENTAL 3470 02:54:54,617 --> 02:54:56,353 PRINCIPLES OF BRAIN FUNCTION. 3471 02:54:56,353 --> 02:54:58,321 MY COLLEAGUES IN MY SISTER 3472 02:54:58,321 --> 02:55:00,190 DIVISION IN AUTISM ARE FOCUSED 3473 02:55:00,190 --> 02:55:02,359 ON BASIC AUTISM RESEARCH. 3474 02:55:02,359 --> 02:55:04,494 AND SO OUR INTEREST IN NEUROAI 3475 02:55:04,494 --> 02:55:08,498 AND A.I. IS SPREAD ACROSS THESE 3476 02:55:08,498 --> 02:55:08,932 DIVISIONS. 3477 02:55:08,932 --> 02:55:11,568 SO BOTH NEUROSCIENCE 3478 02:55:11,568 --> 02:55:12,369 COLLABORATIONS AND FLATIRON 3479 02:55:12,369 --> 02:55:14,204 INSTITUTE ON FOCUSED ON USING 3480 02:55:14,204 --> 02:55:15,105 NEUROAI TO UNDERSTAND THE BRAIN 3481 02:55:15,105 --> 02:55:19,576 AND HOW THE BRAIN WORKS, WHETHER 3482 02:55:19,576 --> 02:55:20,310 STUDYING BIOLOGICAL CONSTRAINED 3483 02:55:20,310 --> 02:55:21,177 NEURAL NETS AS MODEL SYSTEMS 3484 02:55:21,177 --> 02:55:25,615 UNDERSTAND THE BRAIN OR HARNESSING 3485 02:55:25,615 --> 02:55:26,683 THE POWER OF NEUROINSPIRED A.I. 3486 02:55:26,683 --> 02:55:27,117 AND A.I. MODELS TO ANALYZE DATA. 3487 02:55:27,117 --> 02:55:29,519 IN THE AUTISM RESEARCH 3488 02:55:29,519 --> 02:55:33,790 INITIATIVE, OUR INTEREST IN A.I. 3489 02:55:33,790 --> 02:55:36,126 EXTENDS SPECIFICALLY TO 3490 02:55:36,126 --> 02:55:43,466 AUTISM-RELATED WORK, USING A.I. 3491 02:55:43,466 --> 02:55:49,606 TO SPOTLIGHT PATHWAYS FROM 3492 02:55:49,606 --> 02:55:59,149 GENOMICS TO BIOLOGY. 3493 02:55:59,149 --> 02:56:02,218 FLATIRON INSTITUTE, THE WORK ON 3494 02:56:02,218 --> 02:56:04,287 MODERN DEEP NETWORK THEORY, ALSO IN 3495 02:56:04,287 --> 02:56:11,761 OUR CENTER FOR COMPUTATIONAL 3496 02:56:11,761 --> 02:56:18,401 BIOLOGY WE HAVE RESEARCHERS WITH 3497 02:56:18,401 --> 02:56:27,310 PROTEIN PROTEIN INTERACTIONS. 3498 02:56:27,310 --> 02:56:29,045 WE HAVE A LARGE INTEREST IN 3499 02:56:29,045 --> 02:56:31,581 INVESTMENT IN A.I. AT THE SIMONS 3500 02:56:31,581 --> 02:56:33,349 FOUNDATION, FROM LOTS OF 3501 02:56:33,349 --> 02:56:33,783 DIFFERENT ANGLES. 3502 02:56:33,783 --> 02:56:35,084 >> THAT'S ONE OF THE THINGS 3503 02:56:35,084 --> 02:56:39,589 ABOUT A.I., IT SEEMS TO PERMEATE 3504 02:56:39,589 --> 02:56:41,124 EVERYTHING, RIGHT? 3505 02:56:41,124 --> 02:56:44,160 EVERYTHING FROM ENHANCING HUMAN 3506 02:56:44,160 --> 02:56:47,030 PERFORMANCE, COMPUTERS, AND 3507 02:56:47,030 --> 02:56:49,199 ESPECIALLY PEOPLE WHO SUFFER 3508 02:56:49,199 --> 02:56:50,033 FROM MENTAL DISORDERS, 3509 02:56:50,033 --> 02:56:50,834 NEUROLOGICAL DISEASES. 3510 02:56:50,834 --> 02:56:52,602 THIS IS REALLY EXCITING. 3511 02:56:52,602 --> 02:56:52,802 OKAY. 3512 02:56:52,802 --> 02:56:58,842 OTHERS THAT WANT TO JUMP IN? 3513 02:56:58,842 --> 02:56:59,609 STEVE? 3514 02:56:59,609 --> 02:57:01,578 >> OKAY, THANKS. 3515 02:57:01,578 --> 02:57:02,745 TERRY, WHAT YOU JUST SAID STOLE 3516 02:57:02,745 --> 02:57:03,880 MY ANSWER. 3517 02:57:03,880 --> 02:57:05,582 A.I. PERMEATES EVERYTHING. 3518 02:57:05,582 --> 02:57:07,817 AT THE NATIONAL SCIENCE 3519 02:57:07,817 --> 02:57:09,552 FOUNDATION, WE ARE, YOU KNOW, 3520 02:57:09,552 --> 02:57:11,187 OUR MISSION IS TO ADVANCE 3521 02:57:11,187 --> 02:57:13,823 SCIENCE FOR THE BENEFIT OF THE 3522 02:57:13,823 --> 02:57:14,891 NATION, FULL STOP. 3523 02:57:14,891 --> 02:57:16,025 WE'RE TRYING TO POSITION 3524 02:57:16,025 --> 02:57:19,529 OURSELVES TO BE A LEADER IN THE 3525 02:57:19,529 --> 02:57:21,297 FUTURE OF A.I., PREMIER FEDERAL 3526 02:57:21,297 --> 02:57:26,903 FUNDING, FUTURE OF A.I., VARIOUS 3527 02:57:26,903 --> 02:57:28,204 AREAS OF BIOTECHNOLOGIES ALIGNED 3528 02:57:28,204 --> 02:57:28,571 HERE. 3529 02:57:28,571 --> 02:57:30,106 THE PROBLEMS WE'VE BEEN TALKING 3530 02:57:30,106 --> 02:57:31,774 ABOUT TODAY, YOU KNOW, LIKE THE 3531 02:57:31,774 --> 02:57:34,244 POWER ISSUES WITH A.I., RIGHT? 3532 02:57:34,244 --> 02:57:37,413 YOU KNOW, MAYBE WE NEED HARDWARE 3533 02:57:37,413 --> 02:57:39,949 SYSTEMS, WE NEED FUNDAMENTAL 3534 02:57:39,949 --> 02:57:45,922 RESEARCH INTO HARDWARE SYSTEMS 3535 02:57:45,922 --> 02:57:46,623 FASTER, CHEAPER, UNDERSTANDING 3536 02:57:46,623 --> 02:57:48,858 INTERACTION AND CONTROL OF 3537 02:57:48,858 --> 02:57:49,893 PHYSICAL SYSTEMS THAT INTERFACE 3538 02:57:49,893 --> 02:57:53,897 WITH A.I., ALL OF THESE SORTS OF 3539 02:57:53,897 --> 02:57:55,298 ISSUES COULD GENERATE 3540 02:57:55,298 --> 02:57:56,199 FUNDAMENTAL SCIENCE PROPOSALS 3541 02:57:56,199 --> 02:57:59,469 THAT COULD BE SUBMITTED TO THE 3542 02:57:59,469 --> 02:57:59,903 NSF. 3543 02:57:59,903 --> 02:58:02,972 I DON'T KNOW, STEPHANIE, IF YOU 3544 02:58:02,972 --> 02:58:04,941 WANT TO ADD ANYTHING? 3545 02:58:04,941 --> 02:58:06,042 IT REALLY WAS, TERRY SUMMED IT 3546 02:58:06,042 --> 02:58:06,276 UP. 3547 02:58:06,276 --> 02:58:08,778 I DON'T THINK THERE'S AN AREA OF 3548 02:58:08,778 --> 02:58:11,614 NSF THAT ISN'T INTERESTED IN 3549 02:58:11,614 --> 02:58:12,949 NEUROAI. WE HAVE A NEW 3550 02:58:12,949 --> 02:58:16,319 UNDERSTANDING THE BRAIN WEBSITE 3551 02:58:16,319 --> 02:58:18,621 TO HELP THE PUBLIC THINK ABOUT 3552 02:58:18,621 --> 02:58:22,892 HOW NSF THINKS ABOUT NEUROSCIENCE 3553 02:58:22,892 --> 02:58:26,629 AND BRAIN RESEARCH. FROM NEURAL 3554 02:58:26,629 --> 02:58:28,164 MECHANISMS AND PROCESSING TO 3555 02:58:28,164 --> 02:58:28,932 UNDERSTANDING INTELLIGENCE, 3556 02:58:28,932 --> 02:58:31,768 BRAIN-INSPIRED CONCEPTS OF 3557 02:58:31,768 --> 02:58:32,569 DESIGN, THEORY, COMPUTATION, 3558 02:58:32,569 --> 02:58:34,304 ANALYSIS, MODELING OF THE BRAIN, 3559 02:58:34,304 --> 02:58:35,838 HUMAN COGNITION OF SOCIETY. 3560 02:58:35,838 --> 02:58:39,342 I'M SURE I'M MISSING OTHERS. 3561 02:58:39,342 --> 02:58:39,809 NEUROTECHNOLOGY AND 3562 02:58:39,809 --> 02:58:40,243 INFRASTRUCTURE. 3563 02:58:40,243 --> 02:58:42,545 EVERY TALK THAT I'VE HEARD SO 3564 02:58:42,545 --> 02:58:45,915 FAR HAS A PLACE AT NSF FOR 3565 02:58:45,915 --> 02:58:49,152 FUNDING THOSE BASIC QUESTIONS. 3566 02:58:49,152 --> 02:58:54,591 >> YES, GO AHEAD. 3567 02:58:54,591 --> 02:58:54,857 >> OKAY. 3568 02:58:54,857 --> 02:58:57,427 SO, IN ADDITION TO OVERALL 3569 02:58:57,427 --> 02:59:02,031 INTEREST IN A.I., THERE'S A LOT 3570 02:59:02,031 --> 02:59:04,801 OF INTEREST, ENHANCEMENTS, BUT 3571 02:59:04,801 --> 02:59:06,202 LOOKING INTO THE FUTURE, ONE OF 3572 02:59:06,202 --> 02:59:07,971 THE ISSUES, WE'RE NOT REALLY 3573 02:59:07,971 --> 02:59:09,606 EVOLVED FOR THE WORLD WE'RE 3574 02:59:09,606 --> 02:59:11,674 LIVING IN NOW. 3575 02:59:11,674 --> 02:59:14,677 ALL THE A.I., AUGMENTED REALITY, 3576 02:59:14,677 --> 02:59:16,179 YOU KNOW. 3577 02:59:16,179 --> 02:59:18,681 AND WHAT THEY FORESEE IN THE 3578 02:59:18,681 --> 02:59:21,217 FUTURE WARFARE IS MORE AUTONOMY, 3579 02:59:21,217 --> 02:59:22,619 MORE A.I.-ENABLED DECISION 3580 02:59:22,619 --> 02:59:22,986 SUPPORT. 3581 02:59:22,986 --> 02:59:25,355 FOR THAT YOU NEED A.I. THAT'S 3582 02:59:25,355 --> 02:59:28,992 SMART ENOUGH TO BE MORE 3583 02:59:28,992 --> 02:59:29,359 AUTONOMOUS. 3584 02:59:29,359 --> 02:59:31,828 YOU KNOW, WITH THE LIMITATIONS 3585 02:59:31,828 --> 02:59:33,162 WITH EXISTING DATA, EXISTING 3586 02:59:33,162 --> 02:59:35,465 A.I., WE NEED SOME NEW 3587 02:59:35,465 --> 02:59:42,572 PRINCIPLES, NEW THEORIES, NEW 3588 02:59:42,572 --> 02:59:42,839 APPROACHES 3589 02:59:42,839 --> 02:59:47,777 >> AUTONOMY CAME UP IN THE 3590 02:59:47,777 --> 02:59:48,845 PREVIOUS SESSION. 3591 02:59:48,845 --> 02:59:50,780 A.I. RIGHT NOW CAN TALK THE TALK 3592 02:59:50,780 --> 02:59:52,248 BUT CAN'T WALK THE WALK BECAUSE 3593 02:59:52,248 --> 02:59:55,518 IT DOESN'T HAVE A BODY. 3594 02:59:55,518 --> 02:59:58,254 BUT AUTONOMY MEANS MORE THAN 3595 02:59:58,254 --> 03:00:02,125 JUST A BODY. 3596 03:00:02,125 --> 03:00:07,030 IT MEANS IT HAS ITS OWN SELF, 3597 03:00:07,030 --> 03:00:08,765 IT'S SELF-GENERATED ACTIVITY. 3598 03:00:08,765 --> 03:00:10,867 WHEN YOU STOP TALKING, YOU KNOW, 3599 03:00:10,867 --> 03:00:14,771 TO ChatGPT, YOU PRESS THE 3600 03:00:14,771 --> 03:00:16,105 BUTTON, IT COMPLETELY 3601 03:00:16,105 --> 03:00:16,506 DISAPPEARS. 3602 03:00:16,506 --> 03:00:17,874 ChatGPT JUST GOES BLANK. 3603 03:00:17,874 --> 03:00:20,143 THERE'S NO INTERNAL THOUGHT 3604 03:00:20,143 --> 03:00:20,376 THERE. 3605 03:00:20,376 --> 03:00:20,943 RIGHT? 3606 03:00:20,943 --> 03:00:23,112 AND SO WE HAVE A LONG WAY TO GO 3607 03:00:23,112 --> 03:00:28,251 BEFORE WE GET ANY PLACE NEAR TO 3608 03:00:28,251 --> 03:00:28,518 AUTONOMY. 3609 03:00:28,518 --> 03:00:29,252 GO AHEAD. 3610 03:00:29,252 --> 03:00:30,253 >> I'D LIKE TO RECAST THE 3611 03:00:30,253 --> 03:00:33,189 QUESTION SO INSTEAD OF THE WORD 3612 03:00:33,189 --> 03:00:35,525 NEUROAI IN THERE, LET'S GO BACK 3613 03:00:35,525 --> 03:00:36,926 TO THE, SAY, EARLY INDUSTRIAL 3614 03:00:36,926 --> 03:00:41,097 REVOLUTION AND REPLACE THE WORD 3615 03:00:41,097 --> 03:00:42,498 WITH THE RAILWAY, TRAINS. 3616 03:00:42,498 --> 03:00:44,701 WHY ARE WE INTERESTED IN TRAINS? 3617 03:00:44,701 --> 03:00:46,969 WELL, PERHAPS WE'D LIKE TO CROSS 3618 03:00:46,969 --> 03:00:48,204 THE CONTINENT, BUILD A COUNTRY, 3619 03:00:48,204 --> 03:00:48,971 LET'S CALL IT AMERICA. 3620 03:00:48,971 --> 03:00:51,040 WE NEED TO GET FROM ONE SIDE TO 3621 03:00:51,040 --> 03:00:52,141 THE OTHER, EFFECTIVELY. 3622 03:00:52,141 --> 03:00:54,644 IT'S GOING TO AFFECT EVERY 3623 03:00:54,644 --> 03:00:59,816 VERTICAL, WHETHER THE ARMY, 3624 03:00:59,816 --> 03:01:02,118 WHETHER IT'S PASSENGER 3625 03:01:02,118 --> 03:01:02,885 MOTIVATION. 3626 03:01:02,885 --> 03:01:04,954 A.I. IS HORIZONTAL TECHNOLOGY 3627 03:01:04,954 --> 03:01:06,155 THAT AFFECTS EVERY VERTICAL. 3628 03:01:06,155 --> 03:01:10,326 WE HAVE TO FIGURE OUT HOW TO GET 3629 03:01:10,326 --> 03:01:14,263 INVOLVED EARLY TO MAKE IT MORE 3630 03:01:14,263 --> 03:01:15,465 EFFICIENT, MORE PROLIFIC. 3631 03:01:15,465 --> 03:01:16,899 >> YEAH, THIS IS OFTEN THE CASE 3632 03:01:16,899 --> 03:01:19,202 WITH I MEAN THE RAILROADS ARE A 3633 03:01:19,202 --> 03:01:20,069 GOOD EXAMPLE. 3634 03:01:20,069 --> 03:01:21,704 THE INTERNET IS AN EVEN BETTER 3635 03:01:21,704 --> 03:01:22,105 EXAMPLE. 3636 03:01:22,105 --> 03:01:24,674 YOU REMEMBER THERE WAS A TIME 3637 03:01:24,674 --> 03:01:26,943 WHEN THEY WERE PUTTING IN 3638 03:01:26,943 --> 03:01:28,611 FIBEROPTICS EVERYWHERE, WHEN OF 3639 03:01:28,611 --> 03:01:32,081 COURSE THERE WAS THE INTERNET 3640 03:01:32,081 --> 03:01:34,283 BUBBLE BURST. 3641 03:01:34,283 --> 03:01:36,285 THAT ALL DISAPPEARED. 3642 03:01:36,285 --> 03:01:37,920 IT WENT SILENT, DARK, THEY CALL 3643 03:01:37,920 --> 03:01:40,022 IT DARK. 3644 03:01:40,022 --> 03:01:41,524 NOTHING GOING THROUGH THE FIBER. 3645 03:01:41,524 --> 03:01:43,159 THIS MIGHT HAPPEN RIGHT NOW, 3646 03:01:43,159 --> 03:01:45,261 IT'S THE CURRENT A.I., IT'S JUST 3647 03:01:45,261 --> 03:01:49,198 ACCELERATING SO MUCH. 3648 03:01:49,198 --> 03:01:51,567 $100 BILLION, LARGE COMPANIES 3649 03:01:51,567 --> 03:01:52,502 ARE PUTTING IN $100 BILLION 3650 03:01:52,502 --> 03:01:54,470 INVESTMENTS IN DATA CENTERS. 3651 03:01:54,470 --> 03:01:56,639 SO MUCH POWER, GIGAWATTS OF 3652 03:01:56,639 --> 03:01:58,741 POWER THEY NEED TO BUILD NUCLEAR 3653 03:01:58,741 --> 03:01:59,709 POWER PLANTS NEXT TO THEM. 3654 03:01:59,709 --> 03:02:01,344 SO I DON'T THINK THAT'S 3655 03:02:01,344 --> 03:02:03,212 SUSTAINABLE BUT MAYBE I'M WRONG. 3656 03:02:03,212 --> 03:02:05,948 >> I'D LIKE TO BRING THIS 3657 03:02:05,948 --> 03:02:07,283 CONVERSATION BACK TO NEUROAI, 3658 03:02:07,283 --> 03:02:10,319 WHICH IS NOT A.I., RIGHT? 3659 03:02:10,319 --> 03:02:13,089 >> NO, NO, RIGHT NOW A.I. IS 3660 03:02:13,089 --> 03:02:14,724 PULLING THE LOAD, RIGHT? 3661 03:02:14,724 --> 03:02:18,227 AND WHERE WE GO AS NEUROAI IS 3662 03:02:18,227 --> 03:02:19,428 COUPLED, YOU KNOW. WE'RE ACTUALLY 3663 03:02:19,428 --> 03:02:21,030 IN TERMS OF AMOUNT OF FUNDING 3664 03:02:21,030 --> 03:02:21,931 INVOLVED, I DON'T 3665 03:02:21,931 --> 03:02:24,667 KNOW WHAT THAT NUMBER IS, BUT 3666 03:02:24,667 --> 03:02:26,669 WE'RE THE CABOOSE. 3667 03:02:26,669 --> 03:02:27,403 OKAY? 3668 03:02:27,403 --> 03:02:27,870 >> YEAH, OKAY. 3669 03:02:27,870 --> 03:02:32,975 SORRY, I BROUGHT UP THE RAILWAY 3670 03:02:32,975 --> 03:02:33,209 ANALOGY. 3671 03:02:33,209 --> 03:02:38,548 JUMP TO CARS AND MOTORCYCLES. 3672 03:02:38,548 --> 03:02:39,782 INTERNET ANALOGY, MORE RECENT, 3673 03:02:39,782 --> 03:02:42,518 SURE, WE ALL STARTED WITH 3674 03:02:42,518 --> 03:02:42,819 MODEMS. 3675 03:02:42,819 --> 03:02:46,656 I THINK GRACE'S POINT IS WE'RE 3676 03:02:46,656 --> 03:02:48,758 LOOKING FORWARD, PERHAPS TO A 3677 03:02:48,758 --> 03:02:51,127 DAY WE CAN HAVE VIDEO 3678 03:02:51,127 --> 03:02:54,764 TELECONFERENCE AS PORTRAYED IN 3679 03:02:54,764 --> 03:02:56,866 THE MOVIE 2001 A SPACE ODYSSEY, 3680 03:02:56,866 --> 03:02:57,066 RIGHT? 3681 03:02:57,066 --> 03:02:59,268 IT'S GOING TO TAKE FIBER. 3682 03:02:59,268 --> 03:03:01,971 PEOPLE KNEW, GRACE IS POINTING 3683 03:03:01,971 --> 03:03:03,239 OUT NEUROAI THERE'S AN IMPLICIT 3684 03:03:03,239 --> 03:03:05,041 UNDERSTANDING WE'VE GOT TO DO IT 3685 03:03:05,041 --> 03:03:06,142 BETTER, EARLIER, AS FAST AS 3686 03:03:06,142 --> 03:03:08,311 POSSIBLE, TO YOUR POINT WITH 3687 03:03:08,311 --> 03:03:08,678 EFFICIENCY. 3688 03:03:08,678 --> 03:03:09,745 >> IT'S TRUE. 3689 03:03:09,745 --> 03:03:12,215 IT'S HEADING IN THAT DIRECTION. 3690 03:03:12,215 --> 03:03:13,583 I THINK THAT IN THE LONG TERM 3691 03:03:13,583 --> 03:03:16,319 WE'LL GET THERE BUT IT'S GOING 3692 03:03:16,319 --> 03:03:17,220 TO BE INCREMENTAL. 3693 03:03:17,220 --> 03:03:18,855 MAYBE THERE WILL BE SOME HUGE 3694 03:03:18,855 --> 03:03:20,056 BREAKTHROUGH, I DON'T THINK YOU 3695 03:03:20,056 --> 03:03:21,023 CAN PREDICT THAT. 3696 03:03:21,023 --> 03:03:22,925 >> EVEN BETTER ANSWER TO THE 3697 03:03:22,925 --> 03:03:24,327 FIRST QUESTION, BECAUSE I WANT 3698 03:03:24,327 --> 03:03:25,761 TO MAKE IT HAPPEN FASTER. 3699 03:03:25,761 --> 03:03:26,395 >> OKAY, GOOD. 3700 03:03:26,395 --> 03:03:30,800 NO, THAT'S A GREAT ANSWER. 3701 03:03:30,800 --> 03:03:34,704 BECAUSE WHAT WE'RE TRYING TO DO. 3702 03:03:34,704 --> 03:03:39,108 >> NIH IS INTERESTED IN NEUROAI 3703 03:03:39,108 --> 03:03:41,844 OR BRAIN WOULDN'T BE HOSTING 3704 03:03:41,844 --> 03:03:44,380 THIS WORKSHOP. 3705 03:03:44,380 --> 03:03:45,248 NEUROAI REQUIRES TRUE 3706 03:03:45,248 --> 03:03:46,015 INTERDISCIPLINARY COLLABORATION. 3707 03:03:46,015 --> 03:03:48,517 BEYOND WHAT BRAIN HAS ACHIEVED. 3708 03:03:48,517 --> 03:03:52,021 NIH BRAIN IS A GREAT EXAMPLE OF 3709 03:03:52,021 --> 03:03:57,293 CONVERGENT SCIENCE, WE BROUGHT 3710 03:03:57,293 --> 03:03:58,494 TOGETHER ENGINEERS, CLINICIANS, 3711 03:03:58,494 --> 03:03:59,595 NEUROSCIENTISTS TO BUILD AMAZING 3712 03:03:59,595 --> 03:04:01,898 TECHNOLOGY, ALL IN THE SERVICE 3713 03:04:01,898 --> 03:04:02,965 OF NEUROSCIENCE. 3714 03:04:02,965 --> 03:04:04,166 NEUROAI IS DIFFERENT. 3715 03:04:04,166 --> 03:04:06,836 IT'S ABOUT TWO TO THREE FIELDS 3716 03:04:06,836 --> 03:04:12,308 COMING TOGETHER TO TEACH EACH 3717 03:04:12,308 --> 03:04:15,278 OTHER, THAT RECIPROCAL VIRTUOUS 3718 03:04:15,278 --> 03:04:16,812 LOOP HAL MENTIONED. 3719 03:04:16,812 --> 03:04:22,518 NEUROAI WILL GIVE US NEW KNOWLEDGE 3720 03:04:22,518 --> 03:04:23,152 ABOUT ROBOTICS, NEUROSCIENCE, 3721 03:04:23,152 --> 03:04:25,121 THAT'S THE PROMISE. 3722 03:04:25,121 --> 03:04:27,757 >> WE HAD SLIDES SHOWING THAT 3723 03:04:27,757 --> 03:04:29,725 VIRTUOUS CIRCLE, BUT ALSO 3724 03:04:29,725 --> 03:04:32,361 POINTED OUT THAT IN FACT MODERN 3725 03:04:32,361 --> 03:04:35,531 A.I. WAS BUILT FROM 20th 3726 03:04:35,531 --> 03:04:38,034 CENTURY NEUROSCIENCE, AND NOW 3727 03:04:38,034 --> 03:04:39,468 THE QUESTION IS -- THAT THIN 3728 03:04:39,468 --> 03:04:42,305 STRAND ON THE TOP, HOW IS THAT 3729 03:04:42,305 --> 03:04:43,873 GOING TO BE FATTENED UP? 3730 03:04:43,873 --> 03:04:46,375 BUT, YOU KNOW, I TAKE YOUR 3731 03:04:46,375 --> 03:04:49,045 POINT, IT'S A TWO-WAY STREET. 3732 03:04:49,045 --> 03:04:53,282 BUT REALLY THINK IT'S NOT JUST 3733 03:04:53,282 --> 03:04:55,685 NEUROSCIENCE AND A.I. 3734 03:04:55,685 --> 03:04:57,887 WE'RE TALKING ABOUT, YOU KNOW, 3735 03:04:57,887 --> 03:04:59,755 THE FUTURE, AND IT'S GOING HAVE 3736 03:04:59,755 --> 03:05:01,457 AN INFLUENCE ON EVERY PERSON ON 3737 03:05:01,457 --> 03:05:01,958 THE PLANET. 3738 03:05:01,958 --> 03:05:05,695 THIS IS GOING TO HAVE HUGE 3739 03:05:05,695 --> 03:05:05,995 IMPACT. 3740 03:05:05,995 --> 03:05:07,229 SO REGULATORY ISSUES, WE JUST 3741 03:05:07,229 --> 03:05:08,097 TOUCHED ON THOSE. 3742 03:05:08,097 --> 03:05:11,167 BUT, YOU KNOW, WE'RE REALLY JUST 3743 03:05:11,167 --> 03:05:12,368 BEGINNING THAT PROCESS AND OF 3744 03:05:12,368 --> 03:05:13,903 COURSE THIS HAS HAPPENED MANY 3745 03:05:13,903 --> 03:05:15,104 TIMES WITH DIFFERENT 3746 03:05:15,104 --> 03:05:15,438 TECHNOLOGIES. 3747 03:05:15,438 --> 03:05:16,839 DOES ANYONE ELSE WANT TO FINISH 3748 03:05:16,839 --> 03:05:18,407 THIS BEFORE WE GO TO THE NEXT 3749 03:05:18,407 --> 03:05:19,275 TOPIC? 3750 03:05:19,275 --> 03:05:19,475 OKAY. 3751 03:05:19,475 --> 03:05:22,011 NEXT TOPIC IS WHAT MULTI-AGENCY 3752 03:05:22,011 --> 03:05:23,079 EFFORTS ALREADY EXISTS THAT COULD 3753 03:05:23,079 --> 03:05:29,285 BE A MODEL FOR THE COLLABORATION 3754 03:05:29,285 --> 03:05:30,519 IN NEUROAI? 3755 03:05:30,519 --> 03:05:33,456 >> I'M NOT AWARE OF ANY. 3756 03:05:33,456 --> 03:05:34,657 >> WHY IS THAT? 3757 03:05:34,657 --> 03:05:36,659 WHY IS THAT? 3758 03:05:36,659 --> 03:05:38,594 >> I THINK IT'S, AGAIN, A CHANCE 3759 03:05:38,594 --> 03:05:39,729 TO REPLACE THE WORD WITH 3760 03:05:39,729 --> 03:05:40,896 SOMETHING ELSE. 3761 03:05:40,896 --> 03:05:43,132 AND THEN REFLECT ON THE 3762 03:05:43,132 --> 03:05:43,833 SITUATION. 3763 03:05:43,833 --> 03:05:44,033 OKAY? 3764 03:05:44,033 --> 03:05:45,968 SO, I'LL TAKE LIBERTIES. 3765 03:05:45,968 --> 03:05:49,138 REPLACE THE WORD NEUROAI WITH 3766 03:05:49,138 --> 03:05:50,139 USB. 3767 03:05:50,139 --> 03:05:51,440 WE ALL USE IT. 3768 03:05:51,440 --> 03:05:55,277 I'VE GOT IT HERE. 3769 03:05:55,277 --> 03:05:56,345 WHAT MULTIAGENCY EFFORTS EXIST 3770 03:05:56,345 --> 03:05:59,315 THAT COULD BE A MODEL FOR 3771 03:05:59,315 --> 03:06:00,950 COLLABORATION IN USB? 3772 03:06:00,950 --> 03:06:03,586 >> WELL, WE ACTUALLY HAVE 3773 03:06:03,586 --> 03:06:04,153 EXISTING MULTIAGENCY 3774 03:06:04,153 --> 03:06:08,624 COLLABORATION IN PLACE FOR 3775 03:06:08,624 --> 03:06:11,927 NEUROAI, SCH, FOLKS IN THIS 3776 03:06:11,927 --> 03:06:14,630 ROOM PROBABLY FAMILIAR WITH 3777 03:06:14,630 --> 03:06:16,065 SMART-CONNECTED HEALTH PROGRAM, 3778 03:06:16,065 --> 03:06:17,299 COLLABORATIVE RESEARCH AND 3779 03:06:17,299 --> 03:06:18,567 COMPUTATION NEUROSCIENCE 3780 03:06:18,567 --> 03:06:18,834 PROGRAM. 3781 03:06:18,834 --> 03:06:22,638 I REALIZE YOU'RE NOT AN NSF NIH 3782 03:06:22,638 --> 03:06:24,373 GRANTEE, BUT THERE'S OBVIOUS 3783 03:06:24,373 --> 03:06:26,909 ANSWERS TO NUMBER 2, AND MORE 3784 03:06:26,909 --> 03:06:28,878 RECENTLY NSF AND OUR COLLEAGUES 3785 03:06:28,878 --> 03:06:32,381 AT NIBIB CREATED A THIRD MODEL 3786 03:06:32,381 --> 03:06:33,949 THAT'S REALLY INTERESTING. 3787 03:06:33,949 --> 03:06:34,250 >> OKAY. 3788 03:06:34,250 --> 03:06:37,086 I THINK WHEN I READ THE WORD AGENCY 3789 03:06:37,086 --> 03:06:39,255 I INTERPRETED IT-PRIVATE SECTOR 3790 03:06:39,255 --> 03:06:41,390 >> BY THE WAY, IEEE HAD A LOT TO 3791 03:06:41,390 --> 03:06:42,625 DO WITH THE STANDARDS. 3792 03:06:42,625 --> 03:06:43,492 >> ACTUALLY THEY HAD NOTHING TO 3793 03:06:43,492 --> 03:06:44,360 DO WITH IT. 3794 03:06:44,360 --> 03:06:44,827 >> REALLY? 3795 03:06:44,827 --> 03:06:49,765 >> I TAKE THAT BACK. 3796 03:06:49,765 --> 03:06:56,672 IEEE 3094 WAS COMPETING STANDARD TO 3797 03:06:56,672 --> 03:06:58,841 USB, WANTED TO CHARGE A ROYALTY, 3798 03:06:58,841 --> 03:07:00,376 SEVERAL SAID NO. 3799 03:07:00,376 --> 03:07:03,546 MY FRIEND IN THE FOLSOM FACTORY AT 3800 03:07:03,546 --> 03:07:04,980 INTEL SAID LET'S DO IT FOR FREE. 3801 03:07:04,980 --> 03:07:06,515 >> THIS IS REALITY. 3802 03:07:06,515 --> 03:07:07,817 WE'RE TALKING ABOUT IDEAL -- 3803 03:07:07,817 --> 03:07:08,150 OKAY. 3804 03:07:08,150 --> 03:07:09,452 >> YEAH. 3805 03:07:09,452 --> 03:07:10,553 THE PROBLEM WITH MULTIAGENCY 3806 03:07:10,553 --> 03:07:11,987 WAS I READ PRIVATE SECTOR AS 3807 03:07:11,987 --> 03:07:12,988 WELL INTO THAT. 3808 03:07:12,988 --> 03:07:15,591 BUT YOU'RE RIGHT, IT COULD BE AN 3809 03:07:15,591 --> 03:07:16,725 ENTIRE GROUP OF AGENCY 3810 03:07:16,725 --> 03:07:17,893 COLLABORATIONS THAT WE KNOW 3811 03:07:17,893 --> 03:07:19,662 NOTHING ABOUT. 3812 03:07:19,662 --> 03:07:20,763 PERHAPS TURN THAT AROUND, HOW 3813 03:07:20,763 --> 03:07:23,165 COULD WE FIX THAT SO THE PRIVATE 3814 03:07:23,165 --> 03:07:28,270 SECTOR KNOWS MORE ABOUT IT? 3815 03:07:28,270 --> 03:07:29,939 >> I KNOW NSF AND NIH ALWAYS 3816 03:07:29,939 --> 03:07:31,140 COLLABORATE WITH EACH OTHER, 3817 03:07:31,140 --> 03:07:32,141 RIGHT? 3818 03:07:32,141 --> 03:07:33,342 >> THAT'S TWO AGENCIES, NOT WITH 3819 03:07:33,342 --> 03:07:33,576 PRIVATE 3820 03:07:33,576 --> 03:07:34,743 >> I WAS HOPING THIS QUESTION 3821 03:07:34,743 --> 03:07:38,280 WOULD GO THAT WE LOOK AT WHAT WE 3822 03:07:38,280 --> 03:07:40,349 KNOW ABOUT SCH AND CRCNS TODAY, 3823 03:07:40,349 --> 03:07:42,751 DO WE THINK IT WOULD BE A GOOD 3824 03:07:42,751 --> 03:07:45,755 MODEL FOR SUPPORTING FUTURE 3825 03:07:45,755 --> 03:07:47,490 NEUROAI RESEARCH THAT'S 3826 03:07:47,490 --> 03:07:50,459 INCLUSIVE OF EXTERNAL 3827 03:07:50,459 --> 03:07:52,728 PARTNERSHIPS WITH INDUSTRY. 3828 03:07:52,728 --> 03:07:54,930 >> BETWEEN GOVERNMENT AND 3829 03:07:54,930 --> 03:07:55,998 INDUSTRY. 3830 03:07:55,998 --> 03:08:02,705 >> RIGHT. 3831 03:08:02,705 --> 03:08:05,541 BECAUSE CRNCS AND SCH TODAY IS 3832 03:08:05,541 --> 03:08:06,742 PURELY GOVERNMENT FUNDING AGENCIES, 3833 03:08:06,742 --> 03:08:09,044 I DON'T BELIEVE THERE'S A COMMERCIAL 3834 03:08:09,044 --> 03:08:10,045 PARTNER ON THOSE OPPORTUNITIES. 3835 03:08:10,045 --> 03:08:11,213 >> THAT MIGHT BE AN OPPORTUNITY 3836 03:08:11,213 --> 03:08:13,315 TO GO BACK TO FIRST PRINCIPLES. 3837 03:08:13,315 --> 03:08:16,152 WHY DO WE WANT TO INCREASE 3838 03:08:16,152 --> 03:08:17,253 COLLABORATION AND PERHAPS THE 3839 03:08:17,253 --> 03:08:22,825 REASON IS BECAUSE WE WANT TO 3840 03:08:22,825 --> 03:08:27,997 PROLIFERATE IT MORE EFFICIENTS, 3841 03:08:27,997 --> 03:08:28,364 COST EFFICIENT. 3842 03:08:28,364 --> 03:08:29,198 >> CYBERSECURITY, IF YOU THINK 3843 03:08:29,198 --> 03:08:30,633 ABOUT OUR MISSION SPACE WE MIGHT 3844 03:08:30,633 --> 03:08:33,569 HAVE DIFFERENT MISSION SPACE BUT 3845 03:08:33,569 --> 03:08:34,670 THERE'S CORE COMMONALITIES, 3846 03:08:34,670 --> 03:08:37,206 SECURITY IS ONE. 3847 03:08:37,206 --> 03:08:39,708 POWER EFFICIENCY, DATA 3848 03:08:39,708 --> 03:08:40,009 EFFICIENCY. 3849 03:08:40,009 --> 03:08:41,343 ASPECTS OF ETHICS. 3850 03:08:41,343 --> 03:08:43,179 I'M MISSING ONE MORE 3851 03:08:43,179 --> 03:08:43,946 COMMONALITY. 3852 03:08:43,946 --> 03:08:48,984 >> PROBABLY COST. 3853 03:08:48,984 --> 03:08:49,418 >> COST, YES. 3854 03:08:49,418 --> 03:08:50,219 LEVERAGED RESOURCES. 3855 03:08:50,219 --> 03:08:52,288 >> I THOUGHT I WAS IN 3856 03:08:52,288 --> 03:08:52,855 WASHINGTON. 3857 03:08:52,855 --> 03:08:55,891 >> THERE'S AN EXAMPLE, META 3858 03:08:55,891 --> 03:08:59,828 RELEASED LLAMA 3, OPEN SOURCE, 3859 03:08:59,828 --> 03:09:00,829 ONE OF THEIR BIGGEST LARGE 3860 03:09:00,829 --> 03:09:01,697 LANGUAGE MODELS. 3861 03:09:01,697 --> 03:09:03,899 IT'S SPURRING A LOT OF START-UPS 3862 03:09:03,899 --> 03:09:05,067 AND ACADEMICS WHO COULDN'T 3863 03:09:05,067 --> 03:09:07,937 AFFORD TO BUILD THEIR OWN, 3864 03:09:07,937 --> 03:09:09,371 RIGHT? 3865 03:09:09,371 --> 03:09:10,472 THAT'S AN EXAMPLE WHERE, YOU 3866 03:09:10,472 --> 03:09:15,010 KNOW, BIG -- YOU KNOW, HIGH TECH 3867 03:09:15,010 --> 03:09:18,447 INDUSTRY IS HELPING ALL OF US TO 3868 03:09:18,447 --> 03:09:19,348 EXPLORE THIS NEW DIRECTION. 3869 03:09:19,348 --> 03:09:20,649 >> SORRY TO PUT A DOSE OF 3870 03:09:20,649 --> 03:09:24,253 REALITY BACK IN THAT ONE AGAIN. 3871 03:09:24,253 --> 03:09:25,921 AS AN IEEE VERSUS USB, REALITY IS 3872 03:09:25,921 --> 03:09:28,557 THEY ARE NOT SHARING SOME OF THE 3873 03:09:28,557 --> 03:09:29,558 CORE TRAINING -- NOT SHARING 3874 03:09:29,558 --> 03:09:30,726 NEAR ENOUGH OF THE CORE TRAINING 3875 03:09:30,726 --> 03:09:34,163 DATA TO MAKE IT ACTUALLY USEFUL. 3876 03:09:34,163 --> 03:09:35,030 >> OKAY. 3877 03:09:35,030 --> 03:09:36,765 YEAH, IT DEPENDS ON WHAT THEY 3878 03:09:36,765 --> 03:09:38,300 ARE RELEASING, YOU'RE RIGHT. 3879 03:09:38,300 --> 03:09:40,402 BUT THEY HAVE RELEASED ACTUAL 3880 03:09:40,402 --> 03:09:40,936 NETWORK. 3881 03:09:40,936 --> 03:09:41,203 >> YEAH. 3882 03:09:41,203 --> 03:09:42,638 >> PEOPLE CAN FINE TUNE IT FOR 3883 03:09:42,638 --> 03:09:44,106 THEIR PURPOSES AND A LOT OF 3884 03:09:44,106 --> 03:09:45,874 PEOPLE ARE DOING THAT, FOR 3885 03:09:45,874 --> 03:09:46,542 SPECIFIC APPLICATIONS 3886 03:09:46,542 --> 03:09:48,143 >> WELL, THE TRAINING. 3887 03:09:48,143 --> 03:09:50,546 AGAIN, WITH RESPECT TO TRAINING 3888 03:09:50,546 --> 03:09:52,014 TO CREATE WHAT THE WEIGHTS ARE 3889 03:09:52,014 --> 03:09:57,152 IN THAT MODEL IS WHAT REQUIRES 3890 03:09:57,152 --> 03:09:59,088 DATA CENTERS GPUs, BILLIONS OF 3891 03:09:59,088 --> 03:09:59,455 DOLLARS. 3892 03:09:59,455 --> 03:10:00,990 >> YEAH, BUT -- OKAY. 3893 03:10:00,990 --> 03:10:03,726 ONCE YOU HAVE THAT INVESTMENT IN 3894 03:10:03,726 --> 03:10:05,127 THIS LARGE LANGUAGE MODEL, FINE 3895 03:10:05,127 --> 03:10:06,328 TUNING IS ONLY A SMALL FRACTION 3896 03:10:06,328 --> 03:10:09,298 OF THAT BECAUSE YOU'RE NOW FINE 3897 03:10:09,298 --> 03:10:12,468 TUNING WITH A MUCH SMALLER 3898 03:10:12,468 --> 03:10:14,436 CORPUS OF DATA. 3899 03:10:14,436 --> 03:10:15,738 WE'RE DIVERTING OURSELVES FROM 3900 03:10:15,738 --> 03:10:18,173 THE LARGER QUESTION THAT GRACE 3901 03:10:18,173 --> 03:10:18,374 ASKED. 3902 03:10:18,374 --> 03:10:20,109 >> MAY I ADD A COMMENT? 3903 03:10:20,109 --> 03:10:20,676 >> GO AHEAD. 3904 03:10:20,676 --> 03:10:22,444 >> I CAN THINK OF SEVERAL 3905 03:10:22,444 --> 03:10:24,313 DIFFERENT THINGS THAT WOULD BE 3906 03:10:24,313 --> 03:10:25,748 GOOD EXAMPLES OF COLLABORATION 3907 03:10:25,748 --> 03:10:27,716 ACROSS DIFFERENT DISCIPLINES 3908 03:10:27,716 --> 03:10:28,917 THAT WOULD BENEFIT NEUROAI. 3909 03:10:28,917 --> 03:10:35,257 ONE OBVIOUS TO ME IS NSF A.I. 3910 03:10:35,257 --> 03:10:37,026 INSTITUTE, ESPECIALLY THE ARNI 3911 03:10:37,026 --> 03:10:39,194 ARTIFICIAL AND NATURAL 3912 03:10:39,194 --> 03:10:41,196 INTELLIGENCE -- DID I GET THAT 3913 03:10:41,196 --> 03:10:41,497 RIGHT? 3914 03:10:41,497 --> 03:10:43,499 THERE'S ALSO AT MY HOME AGENCY, 3915 03:10:43,499 --> 03:10:45,901 FOR EXAMPLE, WE HAVE SCIENCE OF 3916 03:10:45,901 --> 03:10:47,036 SECURITY. 3917 03:10:47,036 --> 03:10:48,470 IT'S A COLLABORATION ACROSS 3918 03:10:48,470 --> 03:10:52,174 DIFFERENT LABS THAT INCLUDE 3919 03:10:52,174 --> 03:10:53,609 INDUSTRY, ACADEMIA, ALSO UNDER 3920 03:10:53,609 --> 03:10:56,879 THE WHITE HOUSE OFFICE OF 3921 03:10:56,879 --> 03:11:00,849 SCIENCE TECHNOLOGY POLICY 3922 03:11:00,849 --> 03:11:04,787 THERE'S THE NITRD PROGRAM, STANDS 3923 03:11:04,787 --> 03:11:05,788 FOR NETWORKING INFORMATION 3924 03:11:05,788 --> 03:11:08,824 TECHNOLOGY RESEARCH AND 3925 03:11:08,824 --> 03:11:09,491 DEVELOPMENT. 3926 03:11:09,491 --> 03:11:12,461 SO THAT HELPS WITH 3927 03:11:12,461 --> 03:11:14,296 COLLABORATIONS ACROSS ALL OF 3928 03:11:14,296 --> 03:11:14,897 FEDERAL GOVERNMENT, NOT JUST 3929 03:11:14,897 --> 03:11:16,332 DEPARTMENT OF DEFENSE OR NIH, 3930 03:11:16,332 --> 03:11:17,599 IT'S ACROSS THE FEDERAL 3931 03:11:17,599 --> 03:11:18,133 GOVERNMENT. 3932 03:11:18,133 --> 03:11:25,174 THAT'S ANOTHER EXAMPLE. 3933 03:11:25,174 --> 03:11:28,377 THERE'S LOTS OF OPTIONS. 3934 03:11:28,377 --> 03:11:30,079 >> IS THIS ADVERTISED? 3935 03:11:30,079 --> 03:11:34,249 >> SO IT IS ADVERTISED. 3936 03:11:34,249 --> 03:11:39,588 NITRD, PEOPLE PRONOUNCE THE 3937 03:11:39,588 --> 03:11:40,255 ACRONYM DIFFERENT WAYS, 3938 03:11:40,255 --> 03:11:50,666 NITRD. IT'S NITRD.GOV 3939 03:11:50,933 --> 03:11:51,633 OR JUST TO THE WHITE HOUSE 3940 03:11:51,633 --> 03:11:55,237 >> I WANTED TO ADD GRACE HWANG 3941 03:11:55,237 --> 03:11:58,640 HAD THE BRAID EFRI WHEN SHE WAS AT 3942 03:11:58,640 --> 03:12:04,346 THE NSF A FEW YEAR AO, THE OF TWO US 3943 03:12:04,346 --> 03:12:07,182 PARTNERED AND SO WE WORKED TOGETHER 3944 03:12:07,182 --> 03:12:08,617 TO SHAPE SOLICITATION AND THEN GO 3945 03:12:08,617 --> 03:12:12,121 THROUGH THE SELECTION PROCESS AS 3946 03:12:12,121 --> 03:12:12,788 WELL. 3947 03:12:12,788 --> 03:12:14,823 I ALSO WORKED WITH CHOU. 3948 03:12:14,823 --> 03:12:16,592 WE'RE ALWAYS LOOKING FOR 3949 03:12:16,592 --> 03:12:17,826 OPPORTUNITIES TO WORK ACROSS THE 3950 03:12:17,826 --> 03:12:18,127 GOVERNMENT. 3951 03:12:18,127 --> 03:12:20,229 IT HAS A LOT OF BENEFITS IN 3952 03:12:20,229 --> 03:12:21,864 TERMS OF JUST YOU HAVE MORE 3953 03:12:21,864 --> 03:12:23,732 PEOPLE THINKING ABOUT A PROBLEM. 3954 03:12:23,732 --> 03:12:25,567 BUT ALSO YOU'RE AWARE OF WHAT'S 3955 03:12:25,567 --> 03:12:27,102 GOING ON ACROSS THESE AGENCIES. 3956 03:12:27,102 --> 03:12:29,204 WE TRY TO KEEP UP. 3957 03:12:29,204 --> 03:12:30,172 BUT IT'S NOT EASY. 3958 03:12:30,172 --> 03:12:32,708 AND SO WE WANT TO MAKE SURE 3959 03:12:32,708 --> 03:12:34,676 WE'RE NOT DUPLICATING EFFORTS 3960 03:12:34,676 --> 03:12:37,479 AND THINGS WE DO CAN BENEFIT 3961 03:12:37,479 --> 03:12:38,313 EACH OTHER. 3962 03:12:38,313 --> 03:12:39,381 >> AS A FOLLOW-UP QUESTION, HOW 3963 03:12:39,381 --> 03:12:45,187 DIFFICULT IS IT TO DO THAT KIND 3964 03:12:45,187 --> 03:12:46,321 OF INTERAGENCY COOPERATION? 3965 03:12:46,321 --> 03:12:47,756 >> SO TALKING IS EASY. 3966 03:12:47,756 --> 03:12:48,557 EXCHANGING MONEY IS HARDER 3967 03:12:48,557 --> 03:12:49,158 BECAUSE THERE ARE AGREEMENTS 3968 03:12:49,158 --> 03:12:51,060 THAT HAVE TO BE PUT IN PLACE. 3969 03:12:51,060 --> 03:12:52,795 IT'S NOT EVEN -- JUST BECAUSE 3970 03:12:52,795 --> 03:12:54,096 YOU'RE BOTH GOVERNMENT AGENCIES 3971 03:12:54,096 --> 03:12:55,731 YOU CAN EXCHANGE THINGS, BUT YOU 3972 03:12:55,731 --> 03:12:57,399 HAVE TO MEMORANDA AND THOSE HAVE 3973 03:12:57,399 --> 03:12:58,367 TO BE APPROVED. 3974 03:12:58,367 --> 03:12:59,802 AND IT CAN BE TRICKY. 3975 03:12:59,802 --> 03:13:02,738 BUT IT'S POSSIBLE TO DO. 3976 03:13:02,738 --> 03:13:04,940 BUT, YEAH, SO MONEY IS ONE WAY, 3977 03:13:04,940 --> 03:13:06,041 BUT EXCHANGING IDEAS, THERE'S 3978 03:13:06,041 --> 03:13:06,608 VARIOUS WAYS. 3979 03:13:06,608 --> 03:13:09,211 >> I ALSO HAVE A PERSONAL 3980 03:13:09,211 --> 03:13:13,048 EXAMPLE, WHICH IS IT USED TO BE 3981 03:13:13,048 --> 03:13:16,018 THE CASE THAT NIH STUDY SECTIONS 3982 03:13:16,018 --> 03:13:17,352 WOULD THROW OUT ANY 3983 03:13:17,352 --> 03:13:18,654 COMPUTATIONAL GRANTS BECAUSE 3984 03:13:18,654 --> 03:13:22,491 THEY COULDN'T JUDGE IT. 3985 03:13:22,491 --> 03:13:22,825 BIOLOGISTS. 3986 03:13:22,825 --> 03:13:25,227 BUT THE NSF WAS ASKED TO DO THE 3987 03:13:25,227 --> 03:13:25,794 REVIEWS. 3988 03:13:25,794 --> 03:13:28,297 AND IT WAS A CRCNS, MUST HAVE 3989 03:13:28,297 --> 03:13:31,800 BEEN SOME AGREEMENT, BUT THE 3990 03:13:31,800 --> 03:13:33,469 BEAUTY IS THEY HAVE 3991 03:13:33,469 --> 03:13:35,104 MATHEMATICIANS TO GO THROUGH THE 3992 03:13:35,104 --> 03:13:36,872 PROPOSAL, PICK OUT ONES THE 3993 03:13:36,872 --> 03:13:38,173 HIGHEST RANK, SEND THEM BACK TO 3994 03:13:38,173 --> 03:13:39,374 NIH AND THOSE WERE FUNDED. 3995 03:13:39,374 --> 03:13:41,343 IT WAS REALLY A GREAT PROGRAM. 3996 03:13:41,343 --> 03:13:44,213 JUMP STARTED OUR WHOLE FIELD, 3997 03:13:44,213 --> 03:13:44,746 COMPUTATIONAL SCIENCE. 3998 03:13:44,746 --> 03:13:46,748 STEVEN HAS BEEN STANDING THERE. 3999 03:13:46,748 --> 03:13:48,584 >> YES, GRACE WAS TRYING TO GO 4000 03:13:48,584 --> 03:13:52,087 TO ME FIVE MINUTES AGO. 4001 03:13:52,087 --> 03:13:53,455 CRCNS IS STILL ONGOING. 4002 03:13:53,455 --> 03:13:55,591 YOU'RE SPEAKING LIGHT OF LIKE IT 4003 03:13:55,591 --> 03:13:56,458 DOESN'T EXIST. 4004 03:13:56,458 --> 03:13:58,560 NSF IS STILL RUNNING THE REVIEW, 4005 03:13:58,560 --> 03:13:59,761 WORKING CLOSELY WITH NIH, IT'S A 4006 03:13:59,761 --> 03:14:02,598 FANTASTIC PROGRAM. 4007 03:14:02,598 --> 03:14:06,869 IF YOU CAN FIND KEN WHANG IN THE 4008 03:14:06,869 --> 03:14:09,204 AUDIENCE, YEAH, HE JUST WAVED. 4009 03:14:09,204 --> 03:14:10,339 >> HE STOOD UP. 4010 03:14:10,339 --> 03:14:11,807 >> SO I WANTED TO TALK ABOUT 4011 03:14:11,807 --> 03:14:12,774 THAT WE'RE ALWAYS LOOKING FOR 4012 03:14:12,774 --> 03:14:15,010 NEW WAYS TO WORK WITH OUR OTHER 4013 03:14:15,010 --> 03:14:16,111 FEDERAL PARTNERS. 4014 03:14:16,111 --> 03:14:18,080 HAL BROUGHT UP REALLY WELL HOW 4015 03:14:18,080 --> 03:14:20,249 DIFFICULT THAT CAN BE, TO HAVE 4016 03:14:20,249 --> 03:14:21,483 MONEY CHANGING HANDS, OR TO WORK 4017 03:14:21,483 --> 03:14:22,217 TOGETHER. 4018 03:14:22,217 --> 03:14:25,621 WE HAVE MODELS LIKE CRCNS. 4019 03:14:25,621 --> 03:14:28,023 WE HAVE NEW COLLABORATIONS WITH 4020 03:14:28,023 --> 03:14:29,124 NIH WHERE WE'RE THINKING ABOUT 4021 03:14:29,124 --> 03:14:31,660 WELL, HOW CAN IT BE THAT NSF 4022 03:14:31,660 --> 03:14:34,029 COULD MAYBE DO ONE REVIEW SO 4023 03:14:34,029 --> 03:14:36,231 THERE'S NOT DOUBLE JEOPARDY AND 4024 03:14:36,231 --> 03:14:39,601 WE COULD HAVE A PIPELINE FROM 4025 03:14:39,601 --> 03:14:41,169 INITIAL NSF FUNDING FOR A COUPLE 4026 03:14:41,169 --> 03:14:43,138 YEARS WITH A SORT OF INTERIM 4027 03:14:43,138 --> 03:14:47,943 REVIEW BY NIH, NOT BY A PANEL, 4028 03:14:47,943 --> 03:14:49,711 ADDITIONAL TWO YEARS OF FUNDING 4029 03:14:49,711 --> 03:14:50,913 BY NIH, LEVERAGING UNIQUE 4030 03:14:50,913 --> 03:14:52,347 MISSIONS, WHAT WE'RE LOOKING TO 4031 03:14:52,347 --> 03:14:53,448 SUPPORT TO SUPPORT SOMETHING 4032 03:14:53,448 --> 03:15:00,656 BIGGER THAN WHAT WE COULD DO 4033 03:15:00,656 --> 03:15:02,224 ALONE. 4034 03:15:02,224 --> 03:15:04,393 SO YEAH, NOT BE A DOUBLE TAX, 4035 03:15:04,393 --> 03:15:05,260 DOUBLE REVIEWED. 4036 03:15:05,260 --> 03:15:06,528 YEAH, NO, YOU KNOW, SO THE THING 4037 03:15:06,528 --> 03:15:08,063 I WANT TO SAY THAT WE WANT TO 4038 03:15:08,063 --> 03:15:09,464 HEAR FROM YOU. 4039 03:15:09,464 --> 03:15:11,433 THAT CAME FROM LISTENING TO THE 4040 03:15:11,433 --> 03:15:12,100 SYNTHETIC BIOLOGY COMMUNITY, 4041 03:15:12,100 --> 03:15:13,835 COMING TO US AND SAYING, HEY, 4042 03:15:13,835 --> 03:15:16,738 WHY ISN'T THERE A MECHANISM LIKE 4043 03:15:16,738 --> 03:15:17,072 THIS? 4044 03:15:17,072 --> 03:15:18,507 SO TELL US. 4045 03:15:18,507 --> 03:15:19,942 PLEASE, EVERYBODY IN THE 4046 03:15:19,942 --> 03:15:21,043 AUDIENCE, WHAT ARE YOUR ACTUAL 4047 03:15:21,043 --> 03:15:23,011 NEEDS IN TERMS OF FUNDING? 4048 03:15:23,011 --> 03:15:30,452 WE'LL TRY TO FIGURE OUT A WAY. 4049 03:15:30,452 --> 03:15:31,753 >> STEPHANIE HAS BEEN WAITING. 4050 03:15:31,753 --> 03:15:34,723 >> I WANTED TO ADD ONE POINT 4051 03:15:34,723 --> 03:15:37,659 ABOUT CRCNS, THERE'S ALSO 4052 03:15:37,659 --> 03:15:38,894 INTERNATIONAL AGENCIES ALSO 4053 03:15:38,894 --> 03:15:40,629 PARTNERS SO IN ADDITION TO OUR 4054 03:15:40,629 --> 03:15:42,631 FEDERAL AGENCIES WE DO HAVE 4055 03:15:42,631 --> 03:15:46,134 COLLABORATION WITH THOSE 4056 03:15:46,134 --> 03:15:46,768 INTERNATIONAL PARTNERS WHICH 4057 03:15:46,768 --> 03:15:51,039 GIVEN THE IMPORTANCE OF THIS 4058 03:15:51,039 --> 03:15:53,775 TOPIC IT'S HELPFUL TO GET THOSE 4059 03:15:53,775 --> 03:15:54,209 OPINIONS. 4060 03:15:54,209 --> 03:15:56,078 >> FROM THE PRIVATE FOUNDATION 4061 03:15:56,078 --> 03:15:58,247 PERSPECTIVE, THE SIMONS 4062 03:15:58,247 --> 03:15:59,481 FOUNDATION, IT'S MOSTLY 4063 03:15:59,481 --> 03:16:01,416 ASTROPHYSICS BUT WE'VE PUT OUT 4064 03:16:01,416 --> 03:16:05,287 JOINT CALLS WITH NSF, 4065 03:16:05,287 --> 03:16:06,588 SPECIFICALLY NSF AND SIMONS 4066 03:16:06,588 --> 03:16:09,324 FOUNDATION LAUNCHED TWO A.I. 4067 03:16:09,324 --> 03:16:16,665 INSTITUTE TO HELP ASTRONOMERS 4068 03:16:16,665 --> 03:16:19,267 UNDERSTAND DATA, WE HAVE BLUE- 4069 03:16:19,267 --> 03:16:24,640 PRINTS BETWEEN THE TWO AGENCIES. 4070 03:16:24,640 --> 03:16:28,143 AND WE COULD COLLABORATE WITH OTHER 4071 03:16:28,143 --> 03:16:29,878 FOUNDATIONS ON SIMILAR CAUSE AND 4072 03:16:29,878 --> 03:16:37,119 USE THOSE MECHANISMS AS WELL. 4073 03:16:37,119 --> 03:16:37,452 >> GREAT. 4074 03:16:37,452 --> 03:16:39,521 >> SO, WE HAVE STANDARDIZATION 4075 03:16:39,521 --> 03:16:42,157 FOR DATA EFFICIENCY FOR TESTBEDS, 4076 03:16:42,157 --> 03:16:44,226 THAT'S GREAT. WE HAVE THOSE 4077 03:16:44,226 --> 03:16:46,628 SHARED HPC RESOURCES 4078 03:16:46,628 --> 03:16:47,829 ACROSS THE DOD. 4079 03:16:47,829 --> 03:16:50,132 BUT IF YOU HAVE IDEAS WHAT OTHER 4080 03:16:50,132 --> 03:16:53,969 KINDS OF TESTBEDS WE NEED, HOW TO 4081 03:16:53,969 --> 03:17:04,379 MAKE THOSE AVAILABLE FOR 4082 03:17:04,880 --> 03:17:07,015 RESEARCH. 4083 03:17:07,015 --> 03:17:07,382 >> CHRISTINE. 4084 03:17:07,382 --> 03:17:08,350 >> TWO THINGS. 4085 03:17:08,350 --> 03:17:10,085 I WANTED TO MENTION DARPA SINCE 4086 03:17:10,085 --> 03:17:13,889 I'M REPRESENTING DARPA TODAY. 4087 03:17:13,889 --> 03:17:14,656 DARPA, MY EXPERIENCE, PROBABLY 4088 03:17:14,656 --> 03:17:17,859 THE EXPERIENCE OF PEOPLE IN THE 4089 03:17:17,859 --> 03:17:18,760 ROOM, IT'S COLLABORATIVE, 4090 03:17:18,760 --> 03:17:19,494 MULTIDISCIPLINARY PROGRAMS THEY 4091 03:17:19,494 --> 03:17:20,495 PULL TOGETHER. 4092 03:17:20,495 --> 03:17:23,765 BUT ONE I WANTED TO MENTION WAS 4093 03:17:23,765 --> 03:17:27,369 THEIR A.I. CYBER CHALLENGE THEY 4094 03:17:27,369 --> 03:17:29,171 HAD RECENTLY. 4095 03:17:29,171 --> 03:17:33,075 SO THEY BROUGHT TOGETHER THE BIG 4096 03:17:33,075 --> 03:17:38,613 FOUNDATION COMPANIES LIKE 4097 03:17:38,613 --> 03:17:40,182 GOOGLE, ANTHROPIC, OPENAI, 4098 03:17:40,182 --> 03:17:44,886 THEY SUPPORTED THIS CHALLENGE. 4099 03:17:44,886 --> 03:17:46,988 SEMIFINAL COMPETITION AT DEVCON 4100 03:17:46,988 --> 03:17:47,923 THIS YEAR, FINAL COMPETITION 4101 03:17:47,923 --> 03:17:50,292 NEXT YEAR, BUT IT'S A REALLY 4102 03:17:50,292 --> 03:17:53,028 GOOD EXAMPLE OF DARPA WORKING 4103 03:17:53,028 --> 03:17:55,530 TOGETHER WITH ARPA-H AND ALSO 4104 03:17:55,530 --> 03:17:56,298 INDUSTRY PARTNERS. 4105 03:17:56,298 --> 03:17:58,166 SO THAT'S ONE REALLY GOOD 4106 03:17:58,166 --> 03:17:59,935 EXAMPLE OF A DARPA CHALLENGE 4107 03:17:59,935 --> 03:18:06,641 PROBLEM THAT ENDED UP BRINGING 4108 03:18:06,641 --> 03:18:10,445 IN ARPA-H TO USE A.I. TO DETECT 4109 03:18:10,445 --> 03:18:11,880 AND FIX SOFTWARE VULNERABILITIES 4110 03:18:11,880 --> 03:18:12,981 IN CRITICAL SOFTWARE. 4111 03:18:12,981 --> 03:18:17,052 >> COMPETITION AND STANDARDS. 4112 03:18:17,052 --> 03:18:17,352 >> RIGHT. 4113 03:18:17,352 --> 03:18:18,487 >> AND TESTBEDS. 4114 03:18:18,487 --> 03:18:20,188 SO, I THINK THIS ILLUSTRATES 4115 03:18:20,188 --> 03:18:21,256 THERE'S A LOT OF OPPORTUNITIES, 4116 03:18:21,256 --> 03:18:23,158 A LOT OF THINGS WE PROBABLY 4117 03:18:23,158 --> 03:18:25,560 HAVEN'T EVEN COVERED YET, BUT I 4118 03:18:25,560 --> 03:18:27,095 THINK WE COULD PROBABLY BENEFIT 4119 03:18:27,095 --> 03:18:28,964 IF THERE WERE MORE COOPERATION. 4120 03:18:28,964 --> 03:18:33,902 NOW WE'RE GOING TO GO OVER TO 4121 03:18:33,902 --> 03:18:36,071 OUR ZOOM WEBINAR QUESTIONS. 4122 03:18:36,071 --> 03:18:42,844 I THINK THAT JOE IS GOING TO -- 4123 03:18:42,844 --> 03:18:44,479 IS GOING TO SHOOT THE BALL UP. 4124 03:18:44,479 --> 03:18:47,816 >> HINTING WE GO SHOULD OPEN 4125 03:18:47,816 --> 03:18:49,584 UP TO AUDIENCE Q&A, EITHER IN 4126 03:18:49,584 --> 03:18:51,119 THE ROOM OR ZOOM. 4127 03:18:51,119 --> 03:18:51,486 UNLESS -- 4128 03:18:51,486 --> 03:18:53,388 >> I DON'T SEE ANYBODY OUT THERE 4129 03:18:53,388 --> 03:18:54,489 ASKING QUESTIONS. 4130 03:18:54,489 --> 03:18:57,125 WE HAVE ONE A VOLUNTEER. 4131 03:18:57,125 --> 03:18:57,893 >> ALL RIGHT. 4132 03:18:57,893 --> 03:18:58,894 I'LL START OFF. 4133 03:18:58,894 --> 03:19:02,264 SO, ONE OF THE THINGS ABOUT THIS 4134 03:19:02,264 --> 03:19:03,265 MEETING THAT WILL CONTINUES TO 4135 03:19:03,265 --> 03:19:06,201 BE MORE CLEAR AS WE GO FORWARD 4136 03:19:06,201 --> 03:19:09,738 IS THAT THIS IS A VERY -- GOING 4137 03:19:09,738 --> 03:19:11,740 TO BE VERY INTERDISCIPLINARY, 4138 03:19:11,740 --> 03:19:15,444 WE'RE TALKING ABOUT BIOLOGISTS, 4139 03:19:15,444 --> 03:19:15,977 NEUROSCIENTISTS, COMPUTER 4140 03:19:15,977 --> 03:19:16,745 SCIENTISTS, MATHEMATICIANS, YOU 4141 03:19:16,745 --> 03:19:17,646 NAME IT. 4142 03:19:17,646 --> 03:19:21,449 ONE OF THE CHALLENGES OFTEN WITH 4143 03:19:21,449 --> 03:19:25,287 FUNDING IS THAT IT'S RELATIVELY 4144 03:19:25,287 --> 03:19:27,489 EASY, IN THE FUNDING SENSE, TO 4145 03:19:27,489 --> 03:19:29,224 GET MONEY IN A SWIM LANE. 4146 03:19:29,224 --> 03:19:30,859 BUT WHEN YOU CROSS SWIM LANES, 4147 03:19:30,859 --> 03:19:32,661 WHERE YOU NEED TO BRIDGE MONEY 4148 03:19:32,661 --> 03:19:35,030 FOR COMPUTER SCIENTISTS AND 4149 03:19:35,030 --> 03:19:36,364 NEUROSCIENTISTS, SAY, IT BECOMES 4150 03:19:36,364 --> 03:19:36,698 HARD. 4151 03:19:36,698 --> 03:19:38,333 UNLESS THERE'S A SPECIAL CALL. 4152 03:19:38,333 --> 03:19:39,968 BUT THOSE SPECIAL CALLS USUALLY 4153 03:19:39,968 --> 03:19:42,070 HAVE LIKE A TIME FRAME LIKE 4154 03:19:42,070 --> 03:19:43,038 THREE YEARS OR SOMETHING. 4155 03:19:43,038 --> 03:19:47,742 YET WE'RE TRYING TO BUILD THESE 4156 03:19:47,742 --> 03:19:48,844 COLLABORATIONS, IF THEY WORK, 4157 03:19:48,844 --> 03:19:53,348 LAST DECADES, HOW DO WE GET PAST 4158 03:19:53,348 --> 03:19:55,550 THIS SILOED SWIM LANE SHORT FUSED 4159 03:19:55,550 --> 03:20:04,326 FUNDING CYCLES OPPOSITE WHAT IS 4160 03:20:04,326 --> 03:20:05,527 NEEDED FOR BIG INTERDISCIPLINARY 4161 03:20:05,527 --> 03:20:05,760 EFFORTS. 4162 03:20:05,760 --> 03:20:11,132 >> GOOD QUESTION, ALWAYS A 4163 03:20:11,132 --> 03:20:15,871 PROBLEM. 4164 03:20:15,871 --> 03:20:18,139 >> PRIVATE SECTOR. 4165 03:20:18,139 --> 03:20:18,907 TURN IT INTO A PRODUCT 4166 03:20:18,907 --> 03:20:21,443 >> FOR PROFIT. 4167 03:20:21,443 --> 03:20:26,481 >> WITHIN THE DOD FIRST THING 4168 03:20:26,481 --> 03:20:28,250 OBVIOUS IS MULTIDISCIPLINARY 4169 03:20:28,250 --> 03:20:29,217 UNIVERSITY RESEARCH INITIATIVE. 4170 03:20:29,217 --> 03:20:30,886 SO THAT HAS TOPICS THAT CHANGE 4171 03:20:30,886 --> 03:20:31,720 EVERY YEAR. 4172 03:20:31,720 --> 03:20:32,954 THE PROGRAM OFFICERS ACROSS THE 4173 03:20:32,954 --> 03:20:34,155 THREE SERVICES COMPETE FOR 4174 03:20:34,155 --> 03:20:34,723 TOPICS. 4175 03:20:34,723 --> 03:20:37,659 BUT ONE OF THE KEY ASPECTS IS 4176 03:20:37,659 --> 03:20:39,494 THOSE PROJECTS WHICH ARE UP TO 4177 03:20:39,494 --> 03:20:40,829 $1.5 MILLION A YEAR FOR UP TO 4178 03:20:40,829 --> 03:20:42,530 FIVE YEARS, THEY ARE FAIRLY BIG, 4179 03:20:42,530 --> 03:20:44,299 BUT THEY HAVE TO BE FOR BASIC 4180 03:20:44,299 --> 03:20:47,135 RESEARCH, IT HAS TO BE 4181 03:20:47,135 --> 03:20:47,569 MULTIDISCIPLINARY 4182 03:20:47,569 --> 03:20:50,238 IN TERMS OF THE TEAM MAKEUP, 4183 03:20:50,238 --> 03:20:51,139 APPROACH, SO WE SPECIFICALLY 4184 03:20:51,139 --> 03:20:51,773 LOOK FOR THAT. 4185 03:20:51,773 --> 03:20:55,243 OUTSIDE OF THAT WE'RE ALWAYS 4186 03:20:55,243 --> 03:20:56,478 LOOKING FOR MULTIDISCIPLINARY 4187 03:20:56,478 --> 03:20:57,679 ACTIVITIES, AND THERE'S -- WE 4188 03:20:57,679 --> 03:21:00,849 DON'T HAVE THOSE SORTS OF SILOS 4189 03:21:00,849 --> 03:21:02,918 WITHIN OUR PROGRAMS. 4190 03:21:02,918 --> 03:21:03,451 NEUROSCIENCE IS A 4191 03:21:03,451 --> 03:21:04,219 MULTIDISCIPLINARY FIELD. 4192 03:21:04,219 --> 03:21:06,855 I WOULD SAY WITHIN MY PROGRAM I 4193 03:21:06,855 --> 03:21:10,825 HAVE PEOPLE REPRESENTING VARIOUS 4194 03:21:10,825 --> 03:21:11,159 DISCIPLINES. 4195 03:21:11,159 --> 03:21:12,994 BUT SOMETIMES YOU'LL FIND THERE 4196 03:21:12,994 --> 03:21:15,530 ARE EFFORTS THAT -- WHERE THE 4197 03:21:15,530 --> 03:21:17,732 IDEAS WILL SPAN MULTIPLE 4198 03:21:17,732 --> 03:21:20,468 PROGRAMS WITHIN AN AGENCY, AND 4199 03:21:20,468 --> 03:21:23,305 SO I'M ABLE TO COLLABORATE 4200 03:21:23,305 --> 03:21:24,839 FAIRLY EASILY WITH MY COLLEAGUES 4201 03:21:24,839 --> 03:21:26,574 BECAUSE WE DON'T HAVE MEMORANDA 4202 03:21:26,574 --> 03:21:31,179 IN PLACE, IF THINGS FIT MULTIPLE 4203 03:21:31,179 --> 03:21:32,314 PORTFOLIOS WE CAN CO-FUND ACROSS 4204 03:21:32,314 --> 03:21:32,580 PROGRAMS. 4205 03:21:32,580 --> 03:21:35,050 I DON'T THINK THAT'S AN ISSUE. 4206 03:21:35,050 --> 03:21:38,353 WE'RE TRYING TO ENCOURAGE 4207 03:21:38,353 --> 03:21:38,954 MULTIDISCIPLINARY ACTIVITIES. 4208 03:21:38,954 --> 03:21:42,490 I DON'T THINK WE HAVE BARRIERS 4209 03:21:42,490 --> 03:21:43,024 TO THAT. 4210 03:21:43,024 --> 03:21:44,926 >> NSF HAS THE A.I. CENTERS 4211 03:21:44,926 --> 03:21:47,195 PROGRAM, MEANT TO BE LONGER 4212 03:21:47,195 --> 03:21:51,366 LASTING LIKE FIVE YEARS 4213 03:21:51,366 --> 03:21:52,567 RENEWABLE, IS THAT RIGHT? 4214 03:21:52,567 --> 03:21:57,372 YOU'RE RIGHT. 4215 03:21:57,372 --> 03:21:58,373 IT IS POSSIBLE. 4216 03:21:58,373 --> 03:22:00,542 WITH CENTERS PROGRAMS TO HAVE 4217 03:22:00,542 --> 03:22:04,379 SOMETHING MORE EXTENDED BUT 4218 03:22:04,379 --> 03:22:08,183 THERE'S STILL AN ISSUE, I THINK, 4219 03:22:08,183 --> 03:22:10,819 VERY FEW PEOPLE CAN TAKE 4220 03:22:10,819 --> 03:22:12,587 ADVANTAGE OF THAT, RIGHT? 4221 03:22:12,587 --> 03:22:17,158 IT'S NOT OPEN TO EVERYBODY. 4222 03:22:17,158 --> 03:22:20,128 >> THE MURIES ARE LIMITED, FOR 4223 03:22:20,128 --> 03:22:21,796 U.S. INSTITUTIONS. 4224 03:22:21,796 --> 03:22:25,400 TRADITIONAL GRANTS ARE OPEN, WE 4225 03:22:25,400 --> 03:22:27,802 FUND RESEARCH THAT'S OPEN TO 4226 03:22:27,802 --> 03:22:28,036 ANYONE. 4227 03:22:28,036 --> 03:22:30,338 >> NO DEADLINES, NO DATA 4228 03:22:30,338 --> 03:22:34,042 REQUIREMENTS, TRYING TO BE 4229 03:22:34,042 --> 03:22:34,609 INTERDISCIPLINARY. 4230 03:22:34,609 --> 03:22:36,578 >> THESE PROGRAMS ARE ONGOING. 4231 03:22:36,578 --> 03:22:40,615 SO IT'S NOT LIKE THEY ARE -- 4232 03:22:40,615 --> 03:22:42,817 VERY GOOD, OKAY. 4233 03:22:42,817 --> 03:22:43,918 NEXT QUESTION. 4234 03:22:43,918 --> 03:22:45,887 >> SO PIGGYBACKING ON THAT 4235 03:22:45,887 --> 03:22:47,422 THERE'S TWO FACETS OF 4236 03:22:47,422 --> 03:22:48,323 INTERDISCIPLINARY, RIGHT? 4237 03:22:48,323 --> 03:22:49,958 THERE'S INTERDISCIPLINARY WORK 4238 03:22:49,958 --> 03:22:50,859 AND THERE'S ALSO 4239 03:22:50,859 --> 03:22:53,361 INTERDISCIPLINARY OF TRAINING, 4240 03:22:53,361 --> 03:22:53,561 RIGHT? 4241 03:22:53,561 --> 03:22:55,864 SO ONE WAY TO DEAL WITH THIS IS 4242 03:22:55,864 --> 03:22:57,799 TRAINING NEW GENERATIONS OF 4243 03:22:57,799 --> 03:22:59,934 ENGINEERS AND SCIENTISTS THAT 4244 03:22:59,934 --> 03:23:01,703 ARE COMFORTABLE IN BOTH WORLDS. 4245 03:23:01,703 --> 03:23:04,105 DO ANY FUNDERS HAVE INSIGHTS OR 4246 03:23:04,105 --> 03:23:06,207 PERSPECTIVES ON HOW TO 4247 03:23:06,207 --> 03:23:08,243 INCENTIVIZE FUNDING AND TRAINING 4248 03:23:08,243 --> 03:23:14,582 OF INTERDISCIPLINARY RESEARCHERS 4249 03:23:14,582 --> 03:23:14,983 BETWEEN THESE FIELDS? 4250 03:23:14,983 --> 03:23:16,484 >> I THINK THE BRAIN INITIATIVE 4251 03:23:16,484 --> 03:23:18,053 IS DOING A PRETTY GOOD JOB. 4252 03:23:18,053 --> 03:23:21,623 >> IT'S A REALLY IMPORTANT 4253 03:23:21,623 --> 03:23:21,990 POINT. 4254 03:23:21,990 --> 03:23:24,159 WE'RE AT A PLACE READY TO TRAIN 4255 03:23:24,159 --> 03:23:26,661 THE NEXT GENERATION OF 4256 03:23:26,661 --> 03:23:29,064 NEUROSCIENTISTS THROUGH 4257 03:23:29,064 --> 03:23:29,330 NEUROAI. 4258 03:23:29,330 --> 03:23:30,665 YOU KNOW, EARLY CAREER SCHOLARS 4259 03:23:30,665 --> 03:23:32,367 ARE NOT AFRAID OF COMPUTERS. 4260 03:23:32,367 --> 03:23:39,240 THEY ARE NOT AFRAID OF DATA, TO 4261 03:23:39,240 --> 03:23:42,977 WORK WITH ANIMALS. 4262 03:23:42,977 --> 03:23:44,946 IT'S LIKE THREE IN ONE GOING 4263 03:23:44,946 --> 03:23:45,280 FORWARD. 4264 03:23:45,280 --> 03:23:46,781 >> I WAS TALKING TO PEOPLE AT 4265 03:23:46,781 --> 03:23:48,116 LUNCH, THEY NORMALLY TEACH A 4266 03:23:48,116 --> 03:23:49,884 COURSE, THEY GET A COUPLE DOZEN 4267 03:23:49,884 --> 03:23:51,519 STUDENTS . 4268 03:23:51,519 --> 03:23:55,356 AS SOON AS A.I. IS THROWN INTO 4269 03:23:55,356 --> 03:23:57,425 THE POT THEY HAD 50, 100. 4270 03:23:57,425 --> 03:23:59,394 THE YOUNG GENERATION IS REALLY 4271 03:23:59,394 --> 03:24:01,830 EXCITED ABOUT THIS NEUROAI. 4272 03:24:01,830 --> 03:24:04,766 THEY ARE READY. THEY WANT TO BE 4273 03:24:04,766 --> 03:24:05,767 INTERDISCIPLINARY. SO I THINK 4274 03:24:05,767 --> 03:24:09,370 THAT'S WHERE THE FUTURE IS. 4275 03:24:09,370 --> 03:24:10,438 >> THE OVERWHELMING RESPONSE FOR 4276 03:24:10,438 --> 03:24:14,309 EARLY CAREER SCHOLAR PROGRAM THAT 4277 03:24:14,309 --> 03:24:15,410 JESSICA MOLLICK AND COURTNEY PINARD 4278 03:24:15,410 --> 03:24:18,446 PUT TOGETHER SPEAKS TO THAT. 4279 03:24:18,446 --> 03:24:21,850 YOU'LL HEAR MORE TOMORROW. 4280 03:24:21,850 --> 03:24:22,250 >> OKAY. 4281 03:24:22,250 --> 03:24:22,851 ANOTHER QUESTION? 4282 03:24:22,851 --> 03:24:24,252 WAIT, DO YOU WANT TO SAY 4283 03:24:24,252 --> 03:24:24,752 SOMETHING? 4284 03:24:24,752 --> 03:24:25,120 >> REAL QUICK. 4285 03:24:25,120 --> 03:24:31,826 ONE OF THE THINGS I MENTIONED 4286 03:24:31,826 --> 03:24:35,864 FOR THE DSO, A PROGRAM CALLED 4287 03:24:35,864 --> 03:24:38,800 INNOVATION FELLOWSSHIP PROGRAM, IT 4288 03:24:38,800 --> 03:24:40,235 IS A 2-YEAR ASSIGNMENT FOR PEOPLE 4289 03:24:40,235 --> 03:24:41,870 WITHIN -- WHO HAVE THEIR 4290 03:24:41,870 --> 03:24:43,104 Ph.D., LIKE A POSTDOC 4291 03:24:43,104 --> 03:24:45,306 POSITION, BUT IT'S A GREAT 4292 03:24:45,306 --> 03:24:49,911 OPPORTUNITY FOR PEOPLE TO GET 4293 03:24:49,911 --> 03:24:50,545 EXPERIENCE OVERSEEING PORTFOLIO 4294 03:24:50,545 --> 03:24:52,647 OF HIGH RISK, HIGH REWARD 4295 03:24:52,647 --> 03:24:55,416 RESEARCH, SO I WANT TO PUT A 4296 03:24:55,416 --> 03:24:59,320 PLUG, POSTCARD ON THE TABLE DESCRIBE 4297 03:24:59,320 --> 03:25:02,724 A CALL FOR THEIR - IT'S CALLED 4298 03:25:02,724 --> 03:25:04,926 INSPIRE, ACRONYM, BUT ONE OF THE 4299 03:25:04,926 --> 03:25:06,461 INNOVATION FELLOWS IS HERE 4300 03:25:06,461 --> 03:25:06,928 TODAY. 4301 03:25:06,928 --> 03:25:09,063 HER NAME IS ALISSA LING. 4302 03:25:09,063 --> 03:25:11,599 I DON'T SEE HER BECAUSE THE 4303 03:25:11,599 --> 03:25:13,801 LIGHTS ARE BLINDING ME. 4304 03:25:13,801 --> 03:25:14,469 THERE SHE IS. 4305 03:25:14,469 --> 03:25:16,004 SHE WOULD BE HAPPY TO TELL YOU 4306 03:25:16,004 --> 03:25:17,405 MORE ABOUT THAT PROGRAM AND 4307 03:25:17,405 --> 03:25:17,639 TOPIC. 4308 03:25:17,639 --> 03:25:20,809 THAT'S A REALLY GOOD EXAMPLE OF 4309 03:25:20,809 --> 03:25:22,777 CROSS-TRAINING. 4310 03:25:22,777 --> 03:25:26,281 AND MY HOME AGENCY WE REALLY 4311 03:25:26,281 --> 03:25:28,783 VALUE GETTING THE INTELLIGENCE 4312 03:25:28,783 --> 03:25:30,885 ANALYST PERSPECTIVE, WOULD HAVE 4313 03:25:30,885 --> 03:25:31,853 CROSS-TRAINING WITHIN OUR OWN 4314 03:25:31,853 --> 03:25:33,955 AGENCY BUT I THINK THAT CONCEPT 4315 03:25:33,955 --> 03:25:35,857 OF THE RESEARCHERS GETTING 4316 03:25:35,857 --> 03:25:40,128 INVOLVED IN END USER SPACE, VICE 4317 03:25:40,128 --> 03:25:41,129 VERSA, IS REALLY VALUABLE. 4318 03:25:41,129 --> 03:25:44,165 >> I WANT TO MENTION THE ARMY 4319 03:25:44,165 --> 03:25:50,972 HAS A PROGRAM, A FUNDING MECHANISM, 4320 03:25:50,972 --> 03:25:54,475 $60,000 FOR NINE MONTHS. POSTDOC 4321 03:25:54,475 --> 03:25:58,079 GETS A CRAZY IDEA, NO PILOT DATA, 4322 03:25:58,079 --> 03:25:59,514 HIGH RISK, IF IT HAS A CHANCE TO 4323 03:25:59,514 --> 03:26:02,016 OPEN A NEW FIELD, GO FOR IT. 4324 03:26:02,016 --> 03:26:06,421 >> IT'S A POSTDOC. 4325 03:26:06,421 --> 03:26:08,156 >> YES, A POSTDOC OR FACULTY 4326 03:26:08,156 --> 03:26:09,023 COULD APPLY FOR IT HIMSELF. 4327 03:26:09,023 --> 03:26:12,393 >> I CAN BRING IN A QUESTION 4328 03:26:12,393 --> 03:26:14,295 FROM ZOOM. 4329 03:26:14,295 --> 03:26:18,733 WELL, NOTING THAT THE R&D BUDGET 4330 03:26:18,733 --> 03:26:23,037 OF BIG TECH COMPANIES ARE SUPER 4331 03:26:23,037 --> 03:26:25,773 HIGH, WHAT DO WE THINK ABOUT 4332 03:26:25,773 --> 03:26:27,642 CREATING SMALLER SCALE COMPANIES 4333 03:26:27,642 --> 03:26:30,078 FOR INDIVIDUAL DOMAINS AND 4334 03:26:30,078 --> 03:26:32,347 APPLICATIONS LIKE BCIs, BASED 4335 03:26:32,347 --> 03:26:37,051 ON NEUROAI MODELS OR THEORIES, 4336 03:26:37,051 --> 03:26:40,121 AND HOW DO YOU SUPPORT DIVERSE 4337 03:26:40,121 --> 03:26:42,290 ECOSYSTEM OF SMALLER COMPANIES? DO 4338 03:26:42,290 --> 03:26:45,260 THEY GO TO VC OR IS THERE PUBLIC 4339 03:26:45,260 --> 03:26:47,562 SECTOR SUPPORT OR OTHER PRIVATE 4340 03:26:47,562 --> 03:26:48,429 SECTOR MECHANISMS? 4341 03:26:48,429 --> 03:26:49,964 HOW DOES THIS FALL OUT AGAINST 4342 03:26:49,964 --> 03:26:56,771 THE DIFFERENT LINES AND SECTORS 4343 03:26:56,771 --> 03:26:58,840 OF SUPPORT ACROSS AGENCIES? 4344 03:26:58,840 --> 03:27:01,376 >> OKAY, I COULD TALK TO THE 4345 03:27:01,376 --> 03:27:02,777 PRIVATE SECTOR ASPECT. 4346 03:27:02,777 --> 03:27:06,748 SO, PART OF OPENGPU FOUNDATION, 4347 03:27:06,748 --> 03:27:10,018 IS A SPINOFF OPENGPU VENTURES, 4348 03:27:10,018 --> 03:27:11,219 WHICH IS SPECIFICALLY FOCUSED AT 4349 03:27:11,219 --> 03:27:13,221 FUNDING STARTUPS DOING AND 4350 03:27:13,221 --> 03:27:14,088 SUPPORTING OPEN STANDARDS WORK. 4351 03:27:14,088 --> 03:27:18,359 ONE OF THE CHALLENGES AS A SMALL 4352 03:27:18,359 --> 03:27:20,128 STARTUP YOU CAN FUND MANY POINT 4353 03:27:20,128 --> 03:27:22,730 EFFORTS BUT THEY TEND NOT TO BE 4354 03:27:22,730 --> 03:27:24,932 EXTENSIBLE, IT'S DIFFICULT TO 4355 03:27:24,932 --> 03:27:25,900 STAND ON SHOULDERS WHEN EVERY 4356 03:27:25,900 --> 03:27:27,468 ONE IS AT KNEE LEVEL. 4357 03:27:27,468 --> 03:27:29,837 PART OF THE EFFORT I THINK TO 4358 03:27:29,837 --> 03:27:34,442 GROW THIS HAS TO INVOLVE SOME 4359 03:27:34,442 --> 03:27:36,210 LEVEL OF, YEAH, STARTUPS CAN 4360 03:27:36,210 --> 03:27:38,179 PROLIFERATE WITH SOME POINT 4361 03:27:38,179 --> 03:27:41,015 SOLUTION BUT IT'S GOT TO BE 4362 03:27:41,015 --> 03:27:46,054 SOMETHING OPEN, REUSABLE, 4363 03:27:46,054 --> 03:27:48,156 LEVERAGABLE, EXTENSIBLE INTO 4364 03:27:48,156 --> 03:27:48,389 FUTURE EFFORTS. 4365 03:27:48,389 --> 03:27:52,193 >> I WOULD POINT TO THE SBIR AND 4366 03:27:52,193 --> 03:27:56,364 STTR PROGRAMS, SMALL BUSINESS 4367 03:27:56,364 --> 03:27:58,566 INNOVATIVE RESEARCH AND THESE 4368 03:27:58,566 --> 03:28:01,069 ARE AWARDS, CONTRACTS, GIVEN TO 4369 03:28:01,069 --> 03:28:04,138 SMALL BUSINESSES UNDER CERTAIN 4370 03:28:04,138 --> 03:28:05,139 CONDITIONS TO INVESTIGATE 4371 03:28:05,139 --> 03:28:07,875 TOPICS OF INTEREST. 4372 03:28:07,875 --> 03:28:09,310 STTRs ARE LIKE SBIRs, BUT IT'S A 4373 03:28:09,310 --> 03:28:10,478 PARTNERSHIP BETWEEN A SMALL 4374 03:28:10,478 --> 03:28:12,447 BUSINESS AND A UNIVERSITY. 4375 03:28:12,447 --> 03:28:14,315 AND SO THEY -- THESE TYPES OF 4376 03:28:14,315 --> 03:28:24,859 OPPORTUNITIES MAY BE OF INTEREST 4377 03:28:26,427 --> 03:28:27,095 TO THIS COMMUNITY. 4378 03:28:27,095 --> 03:28:31,265 >> MY QUESTION WAS ON QUESTION 4379 03:28:31,265 --> 03:28:32,133 2, WHAT MULTI-AGENCIES EFFORTS 4380 03:28:32,133 --> 03:28:32,834 ALREADY EXIST. AND 4381 03:28:32,834 --> 03:28:34,902 IT HAD MORE TO DO WITH 4382 03:28:34,902 --> 03:28:35,670 ESTABLISHING STANDARDS BECAUSE 4383 03:28:35,670 --> 03:28:38,306 WE'RE ALL HAVING TO LIKE PROCESS 4384 03:28:38,306 --> 03:28:40,174 REFORMAT ALL THIS DIFFERENT 4385 03:28:40,174 --> 03:28:40,608 DATA. 4386 03:28:40,608 --> 03:28:44,011 AND I WAS WONDERING IF AN 4387 03:28:44,011 --> 03:28:45,947 EXAMPLE LIKE THE APACHE 4388 03:28:45,947 --> 03:28:48,049 FOUNDATION, WHERE MULTIPLE 4389 03:28:48,049 --> 03:28:51,119 COMPANIES ARE COMING TOGETHER TO 4390 03:28:51,119 --> 03:28:52,153 STANDARDIZE DATABASE FORMATS, 4391 03:28:52,153 --> 03:28:53,321 FILE TYPES, WHERE THEY VOTE, 4392 03:28:53,321 --> 03:28:55,957 THEY ALL CONTRIBUTE TO A LARGER 4393 03:28:55,957 --> 03:28:59,227 BODY OF WORK WOULD BE A GOOD 4394 03:28:59,227 --> 03:29:00,228 EXAMPLE. 4395 03:29:00,228 --> 03:29:01,729 >> I DIDN'T QUITE CATCH THE 4396 03:29:01,729 --> 03:29:02,730 QUESTION IN THERE. 4397 03:29:02,730 --> 03:29:05,900 >> LIKE IS THE WAY THAT OPERATES 4398 03:29:05,900 --> 03:29:08,770 AN EXAMPLE OF A MODEL FOR 4399 03:29:08,770 --> 03:29:10,405 COLLABORATION IN NEUROAI? 4400 03:29:10,405 --> 03:29:13,474 >> I DELIBERATELY PICKED OUT THE 4401 03:29:13,474 --> 03:29:15,343 USB EXAMPLE BEFORE BECAUSE JUST 4402 03:29:15,343 --> 03:29:16,611 LIKE BLUETOOTH THERE ARE ZERO 4403 03:29:16,611 --> 03:29:19,046 GOVERNMENT AGENCIES INVOLVED IN 4404 03:29:19,046 --> 03:29:20,181 THE USB STANDARD. 4405 03:29:20,181 --> 03:29:21,916 AND IN BLUETOOTH. 4406 03:29:21,916 --> 03:29:24,752 BUT IT'S EVERYWHERE, RIGHT? 4407 03:29:24,752 --> 03:29:28,990 SO I THINK THE CORE POINT, YOU 4408 03:29:28,990 --> 03:29:30,224 KNOW, ESPECIALLY WITH THE GROWTH 4409 03:29:30,224 --> 03:29:33,027 IN INDUSTRY, WITH ABILITY TO 4410 03:29:33,027 --> 03:29:35,029 INCREASE, YOU KNOW, TO BRAD'S 4411 03:29:35,029 --> 03:29:36,230 QUESTION, MORE PEOPLE ONLINE, 4412 03:29:36,230 --> 03:29:37,899 AND TO TRAIN THE NEXT GENERATION 4413 03:29:37,899 --> 03:29:41,269 PART HAS TO BE, YOU KNOW, WHAT 4414 03:29:41,269 --> 03:29:44,672 ARE THE STANDARDS, WHAT COULD 4415 03:29:44,672 --> 03:29:46,441 YOU GIVE A TEENAGER AS A BOOK TO 4416 03:29:46,441 --> 03:29:48,042 GET HIM EXCITED, TO GO TO 4417 03:29:48,042 --> 03:29:49,944 UNIVERSITY, TO WANT TO DO AND 4418 03:29:49,944 --> 03:29:50,945 APPLY FOR THE GRANT. 4419 03:29:50,945 --> 03:29:52,580 THAT'S GREAT BUT WE HAVE TO GET 4420 03:29:52,580 --> 03:29:53,781 IT MUCH EARLIER. 4421 03:29:53,781 --> 03:29:56,317 THE WAY THAT'S HAPPENED IN THE 4422 03:29:56,317 --> 03:29:57,618 INDUSTRY IS THROUGH CREATING 4423 03:29:57,618 --> 03:29:59,287 OPEN STANDARDS. 4424 03:29:59,287 --> 03:30:06,060 YEAH, IT'S ABSOLUTELY ESSENTIAL. 4425 03:30:06,060 --> 03:30:06,828 >> THANK YOU. 4426 03:30:06,828 --> 03:30:08,029 >> OH, SORRY. 4427 03:30:08,029 --> 03:30:09,464 I FORGOT TO DO THE PLUG. 4428 03:30:09,464 --> 03:30:11,599 BY THE WAY, IT'S CALLED OPENGPU 4429 03:30:11,599 --> 03:30:13,501 BECAUSE IT IS AN OPEN STANDARD. 4430 03:30:13,501 --> 03:30:18,439 THAT'S WHAT I REPRESENT HERE 4431 03:30:18,439 --> 03:30:18,673 TODAY. 4432 03:30:18,673 --> 03:30:19,440 >> SINCE WE'RE PLUGGING THINGS. 4433 03:30:19,440 --> 03:30:20,741 I HAVE A NEW BOOK. 4434 03:30:20,741 --> 03:30:23,478 THIS WAS LAUNCHED LAST MONTH. 4435 03:30:23,478 --> 03:30:25,913 ChatGPT AND THE FUTURE OF 4436 03:30:25,913 --> 03:30:29,984 A.I., IT'S WRITTEN FOR YOU. 4437 03:30:29,984 --> 03:30:32,687 ALL THESE ISSUES THAT HAVE COME 4438 03:30:32,687 --> 03:30:35,423 UP ARE ONES THAT I HAVE TALKED 4439 03:30:35,423 --> 03:30:37,959 ABOUT. 4440 03:30:37,959 --> 03:30:38,192 OKAY. 4441 03:30:38,192 --> 03:30:39,393 ON THIS SIDE? 4442 03:30:39,393 --> 03:30:45,299 >> JUST SORT OF AN OBSERVATION 4443 03:30:45,299 --> 03:30:47,735 AND PERHAPS SORT OF REQUEST 4444 03:30:47,735 --> 03:30:48,836 REACTION FROM PANELISTS. 4445 03:30:48,836 --> 03:30:52,807 THE FIELD OF A.I. GROWS 4446 03:30:52,807 --> 03:30:53,241 EXTREMELY FAST. 4447 03:30:53,241 --> 03:30:56,978 YOU KNOW, THERE ARE SORT OF NeRF 4448 03:30:56,978 --> 03:31:00,348 WAS THE BIG THING IN GRAPHICS. IT'S 4449 03:31:00,348 --> 03:31:04,185 NO LONGER THE BIG THING, IT'S 4450 03:31:04,185 --> 03:31:04,952 CALLED SPLATTING. 4451 03:31:04,952 --> 03:31:06,821 PAPERS PUBLISHED A YEAR AGO HAVE 4452 03:31:06,821 --> 03:31:08,456 10,000 CITATIONS AND SO ON. 4453 03:31:08,456 --> 03:31:13,928 THE REASON FOR THAT IS 4454 03:31:13,928 --> 03:31:15,696 CODE AND DATA ARE EASILY 4455 03:31:15,696 --> 03:31:17,899 AVAILABLE AND PEOPLE CAN ITERATE 4456 03:31:17,899 --> 03:31:18,099 ON IT. 4457 03:31:18,099 --> 03:31:20,535 ON THE OTHER HAND NEUROSCIENCE 4458 03:31:20,535 --> 03:31:22,503 HAS A DIFFERENT CULTURE, PEOPLE 4459 03:31:22,503 --> 03:31:24,572 SPEND YEARS AND CAREFUL 4460 03:31:24,572 --> 03:31:26,207 EXPERIMENTATION TO GET DATA, AND 4461 03:31:26,207 --> 03:31:27,408 THEY WANT TO EXTRACT VALUE OUT 4462 03:31:27,408 --> 03:31:30,044 OF THE DATA AS THEY SHOULD FOR 4463 03:31:30,044 --> 03:31:31,245 THE AMOUNT OF EFFORT THEY HAVE 4464 03:31:31,245 --> 03:31:32,446 PUT IN. 4465 03:31:32,446 --> 03:31:40,254 HOW DO WE BRIDGE THIS CULTURAL 4466 03:31:40,254 --> 03:31:42,590 DIVIDE, IF YOU WOULD? 4467 03:31:42,590 --> 03:31:46,193 >> ANYONE WANT TO BRIDGE THE 4468 03:31:46,193 --> 03:31:47,962 DIVIDE? 4469 03:31:47,962 --> 03:31:50,565 >> BETWEEN WHAT AND WHAT? 4470 03:31:50,565 --> 03:31:52,733 >> BETWEEN THE DESIRE FOR 4471 03:31:52,733 --> 03:31:54,468 RESEARCHER TO MAXIMIZE THEIR 4472 03:31:54,468 --> 03:31:55,570 BENEFIT FROM THE PAINSTAKINGLY 4473 03:31:55,570 --> 03:32:00,408 COLLECTED DATA ON THE ONE SIDE, 4474 03:32:00,408 --> 03:32:03,911 AND THE DESIRE OF FAST 4475 03:32:03,911 --> 03:32:05,646 DEVELOPING FIELD TO QUICKLY HAVE 4476 03:32:05,646 --> 03:32:10,351 MANY, MANY PEOPLE EXTRACT THE 4477 03:32:10,351 --> 03:32:13,454 JUICE FROM THE NEW CODE AND DATA. 4478 03:32:13,454 --> 03:32:15,623 >> I THINK IN MANY WAYS THE 4479 03:32:15,623 --> 03:32:17,825 NEUROSCIENCE FIELD HAS WEIGHED 4480 03:32:17,825 --> 03:32:20,261 IN ON THAT WITH THE PUSH TO OPEN 4481 03:32:20,261 --> 03:32:22,096 DATA AND OPEN SCIENCE, RIGHT? 4482 03:32:22,096 --> 03:32:24,098 I THINK MOST OF NEUROSCIENCE IS 4483 03:32:24,098 --> 03:32:28,035 MOVING IN THAT DIRECTION WHICH I 4484 03:32:28,035 --> 03:32:30,671 THINK LEANS TOWARDS LESS SORT OF 4485 03:32:30,671 --> 03:32:31,439 HOARDING DATA, EXTRACTING ALL 4486 03:32:31,439 --> 03:32:33,541 THE JUICE AS YOU SAID, VERSUS 4487 03:32:33,541 --> 03:32:35,910 GETTING IT OUT THERE FOR THE 4488 03:32:35,910 --> 03:32:37,979 WORLD TO ALLOW OTHERS TO 4489 03:32:37,979 --> 03:32:38,579 INTERACT WITH THE DATA. 4490 03:32:38,579 --> 03:32:39,947 SO I DON'T KNOW THAT THERE'S 4491 03:32:39,947 --> 03:32:41,816 MUCH MORE THAT WE CAN DO AT 4492 03:32:41,816 --> 03:32:44,352 LEAST FROM THE FOUNDATION 4493 03:32:44,352 --> 03:32:44,685 PERSPECTIVE. 4494 03:32:44,685 --> 03:32:52,026 WE ALL HAVE DIFFERENT IDEAS. 4495 03:32:52,026 --> 03:32:53,227 I DON'T KNOW. 4496 03:32:53,227 --> 03:32:54,095 NEUROSCIENCE HAS ALREADY MOVED 4497 03:32:54,095 --> 03:32:57,398 IN THAT DIRECTION. 4498 03:32:57,398 --> 03:32:59,166 >> THAT WAS NOT THE TRADITION, 4499 03:32:59,166 --> 03:33:00,334 BEFORE THE BRAIN INITIATIVE. 4500 03:33:00,334 --> 03:33:04,305 AND I KNOW BECAUSE I WROTE THAT 4501 03:33:04,305 --> 03:33:04,672 SECTION. 4502 03:33:04,672 --> 03:33:07,608 AND THE REASON I WROTE THAT 4503 03:33:07,608 --> 03:33:10,711 SECTION, BRAIN 2025, WAS THAT 4504 03:33:10,711 --> 03:33:11,779 TRADITIONALLY, YOU KNOW, YOU DO 4505 03:33:11,779 --> 03:33:14,649 AN EXPERIMENT IN YOUR LAB, 4506 03:33:14,649 --> 03:33:15,616 COLLECT THE DATA, IT NEVER 4507 03:33:15,616 --> 03:33:16,717 LEAVES THE LAB. 4508 03:33:16,717 --> 03:33:18,252 NOW THERE'S SO MUCH DATA, OTHER 4509 03:33:18,252 --> 03:33:19,887 PEOPLE SHOULD HAVE AN 4510 03:33:19,887 --> 03:33:21,756 OPPORTUNITY TO ANALYZE, OTHER 4511 03:33:21,756 --> 03:33:22,723 TOOLS AND TECHNIQUES. 4512 03:33:22,723 --> 03:33:25,693 NOW THAT'S BECOME THE STANDARD. 4513 03:33:25,693 --> 03:33:26,894 IT'S REALLY CHANGED COMPLETELY; 4514 03:33:26,894 --> 03:33:27,662 YOU'RE ABSOLUTELY RIGHT ABOUT 4515 03:33:27,662 --> 03:33:28,195 THAT. 4516 03:33:28,195 --> 03:33:31,365 >> A QUESTION FROM THE VIDEOCAST 4517 03:33:31,365 --> 03:33:32,800 MONITOR, DANA. 4518 03:33:32,800 --> 03:33:33,567 >> YES. 4519 03:33:33,567 --> 03:33:37,405 THERE'S A QUESTION THAT IS 4520 03:33:37,405 --> 03:33:40,474 ASKING IF THERE ARE ANY GRANTS 4521 03:33:40,474 --> 03:33:41,242 THAT STUDENTS, PhDs CAN 4522 03:33:41,242 --> 03:33:44,745 APPLY FOR IN ORDER TO GAIN 4523 03:33:44,745 --> 03:33:46,313 ACCESS TO NEUROMORPHIC HARDWARE 4524 03:33:46,313 --> 03:33:56,824 IN ORDER TO TRAIN THESE NEUROAI 4525 03:33:59,393 --> 03:33:59,560 MODEL. 4526 03:33:59,560 --> 03:34:01,862 >> MORE GENERALLY, WHAT IS THE 4527 03:34:01,862 --> 03:34:03,497 LANDSCAPE SUPPORT LOOKING LIKE 4528 03:34:03,497 --> 03:34:07,435 FOR BREAKING BARRIERS TO ACCESS 4529 03:34:07,435 --> 03:34:08,102 TECHNOLOGY? 4530 03:34:08,102 --> 03:34:09,403 CERTAINLY IN NEUROMORPHIC 4531 03:34:09,403 --> 03:34:11,372 COMPUTING WHICH WE'LL HEAR MORE 4532 03:34:11,372 --> 03:34:12,907 ABOUT TOMORROW, PEOPLE HAVE 4533 03:34:12,907 --> 03:34:14,642 TROUBLE ACCESSING TESTBEDS AND 4534 03:34:14,642 --> 03:34:18,245 DEVICES IN ORDER TO TRY OUT NEW 4535 03:34:18,245 --> 03:34:19,914 DESIGNS SO THAT'S GOING TO BE AN 4536 03:34:19,914 --> 03:34:23,284 IMPORTANT PART OF TRAINING AS 4537 03:34:23,284 --> 03:34:24,919 WELL AS EARLY STAGE DEVELOPMENT 4538 03:34:24,919 --> 03:34:27,888 OF THE SCIENTIFIC MODELS AND AS 4539 03:34:27,888 --> 03:34:28,422 TECHNOLOGICAL PRODUCTS. 4540 03:34:28,422 --> 03:34:31,826 SO IT WOULD BE GREAT TO HEAR 4541 03:34:31,826 --> 03:34:33,027 FROM MANY OF YOU WITH NEUROMORPHIC 4542 03:34:33,027 --> 03:34:35,629 PROGRAMS WHAT THAT LOOKS LIKE 4543 03:34:35,629 --> 03:34:38,399 GETTING THINGS OUT AT THESE 4544 03:34:38,399 --> 03:34:41,469 EARLY STAGES FOR STUDENTS AND 4545 03:34:41,469 --> 03:34:42,336 OTHERS. 4546 03:34:42,336 --> 03:34:44,405 >> THERE'S A PRECEDENT IN 4547 03:34:44,405 --> 03:34:44,872 SEMICONDUCTORS. SO MOSIS WAS 4548 03:34:44,872 --> 03:34:52,546 SET UP, THESE BIG COMPANIES, VERY 4549 03:34:52,546 --> 03:34:53,581 EXPENSIVE, BUT CARVER MEAD WAS 4550 03:34:53,581 --> 03:34:56,250 INVOLVED IN GETTING THEM TO TAKE 4551 03:34:56,250 --> 03:34:59,186 ACADEMIC CHIPS, A WAY FOR 4552 03:34:59,186 --> 03:35:01,288 TEACHING STUDENTS HOW TO DESIGN 4553 03:35:01,288 --> 03:35:02,990 CHIPS, THE COMPANIES BENEFIT 4554 03:35:02,990 --> 03:35:04,592 BECAUSE NOW THEY HAVE PEOPLE 4555 03:35:04,592 --> 03:35:06,761 THAT WERE COMING IN, ALREADY 4556 03:35:06,761 --> 03:35:07,995 TRAINED, HOW TO DO THAT. 4557 03:35:07,995 --> 03:35:09,730 MAYBE WE COULD DO SOMETHING 4558 03:35:09,730 --> 03:35:11,398 ALONG THOSE LINES. 4559 03:35:11,398 --> 03:35:14,034 >> SO, I'M GOING TO REFLECT ON 4560 03:35:14,034 --> 03:35:16,303 MY EXPERIENCE AT NVIDIA. THERE 4561 03:35:16,303 --> 03:35:24,411 WERE MORE THAN 10 YEARS OF PATH- 4562 03:35:24,411 --> 03:35:26,547 FINDING AND SEEDING RESEARCHERS 4563 03:35:26,547 --> 03:35:27,615 IN INDUSTRY, UNIVERSITIES WITH GPUs 4564 03:35:27,615 --> 03:35:28,816 AND CUDA. JUST GO FOR IT. GIVE 4565 03:35:28,816 --> 03:35:30,584 US BACK THE IP LICENSE, BUT GO FOR 4566 03:35:30,584 --> 03:35:32,219 IT, DO WHATEVER YOU WANT. 4567 03:35:32,219 --> 03:35:34,755 THE FACT AS A RESEARCHER YOU 4568 03:35:34,755 --> 03:35:38,459 COULD BUY A GPU FROM BEST BUY 4569 03:35:38,459 --> 03:35:39,727 AND DOWNLOAD THE CUDA LIBRARIES, 4570 03:35:39,727 --> 03:35:40,895 THE LATEST ONES AND JUST GO. 4571 03:35:40,895 --> 03:35:44,064 I THINK THE QUESTION WAS HOW TO 4572 03:35:44,064 --> 03:35:45,499 GET ACCESS TO NEUROMORPHIC HARDWARE 4573 03:35:45,499 --> 03:35:47,568 NOT HOW TO DESIGN A CHIP. 4574 03:35:47,568 --> 03:35:49,403 SURE, WE WOULD LOVE HAVE MORE 4575 03:35:49,403 --> 03:35:53,574 PEOPLE AWARE OF SEMICONDUCTORS 4576 03:35:53,574 --> 03:35:57,478 BUT THE CORE PROBLEM IS GETTING 4577 03:35:57,478 --> 03:36:01,816 PROLIFIC ACCESS TO NEXTGEN HARDWARE 4578 03:36:01,816 --> 03:36:02,516 IN A.I. TO GET OFF THE GROUND. 4579 03:36:02,516 --> 03:36:05,786 I DON'T HAVE AN ANSWER TO HOW TO 4580 03:36:05,786 --> 03:36:06,854 GET TO NEUROMORPHIC HARDWARE 4581 03:36:06,854 --> 03:36:07,188 TODAY. 4582 03:36:07,188 --> 03:36:10,157 THAT'S KIND OF WHY WE DO OPENGPU 4583 03:36:10,157 --> 03:36:14,228 VENTURES, BECAUSE WE WANT TO 4584 03:36:14,228 --> 03:36:16,730 CREATE PROLIFIC HARDWARE, AS 4585 03:36:16,730 --> 03:36:17,832 CHEAP AS GPU. 4586 03:36:17,832 --> 03:36:18,032 RIGHT? 4587 03:36:18,032 --> 03:36:22,069 SO MY ANSWER TO THE QUESTION 4588 03:36:22,069 --> 03:36:24,605 WOULD BE I CAN'T HELP YOU GET 4589 03:36:24,605 --> 03:36:25,272 ACCESS TO SPINNAKER, I 4590 03:36:25,272 --> 03:36:27,074 UNDERSTAND IT'S IN EUROPE AND 4591 03:36:27,074 --> 03:36:29,810 VERY EXPENSIVE, CAN'T HELP YOU 4592 03:36:29,810 --> 03:36:33,848 GET ACCESS TO SANDIA LABS' LOIHI 4593 03:36:33,848 --> 03:36:35,182 INSTALLATION, BUT MAYBE WE COULD 4594 03:36:35,182 --> 03:36:36,917 TALK ABOUT HOW TO MAKE PROLIFIC 4595 03:36:36,917 --> 03:36:38,552 HARDWARE THAT'S EVERYWHERE AND 4596 03:36:38,552 --> 03:36:41,121 EASILY ACCESSIBLE AND CHEAP. 4597 03:36:41,121 --> 03:36:46,560 >> YEAH, I THINK ABSOLUTELY. 4598 03:36:46,560 --> 03:36:48,963 OKAY. 4599 03:36:48,963 --> 03:36:51,932 NEXT QUESTION HERE? 4600 03:36:51,932 --> 03:36:53,400 >> JUST A PRACTICAL QUESTION 4601 03:36:53,400 --> 03:36:54,969 SINCE WE HAVE INDIVIDUALS FROM 4602 03:36:54,969 --> 03:36:57,938 FUNDING AGENCIES, WHEN WE DO 4603 03:36:57,938 --> 03:36:59,240 MULTIDISCIPLINARY RESEARCH IT 4604 03:36:59,240 --> 03:37:00,241 REQUIRES A MULTITUDE OF SKILLS, 4605 03:37:00,241 --> 03:37:03,410 AND MANY THINGS WE DO IN THE 4606 03:37:03,410 --> 03:37:04,378 FIELD HAVE BIOLOGISTS, PEOPLE 4607 03:37:04,378 --> 03:37:06,280 THAT ARE EITHER EXPERTS IN 4608 03:37:06,280 --> 03:37:08,883 COMPUTER SCIENCE OR MATHEMATICS 4609 03:37:08,883 --> 03:37:09,450 OR PHYSICS AND CLINICIANS. 4610 03:37:09,450 --> 03:37:12,586 BUT THE WAY EACH OF US ARE 4611 03:37:12,586 --> 03:37:14,688 REINFORCED WHEN WE NEED TO GO UP 4612 03:37:14,688 --> 03:37:16,657 FOR TENURE REVIEW, THESE KINDS 4613 03:37:16,657 --> 03:37:18,225 OF THINGS, DEPENDS ON HAVING 4614 03:37:18,225 --> 03:37:20,094 SPECIFIC LABELS FROM THE FUNDING 4615 03:37:20,094 --> 03:37:25,232 AGENCIES, NAMELY ARE YOU A P.I., 4616 03:37:25,232 --> 03:37:25,466 CO-P.I. 4617 03:37:25,466 --> 03:37:28,636 WOULD THERE BE A ROLE FOR GRANT 4618 03:37:28,636 --> 03:37:30,938 CALLS FOR INDIVIDUALS WHERE THERE'S 4619 03:37:30,938 --> 03:37:36,410 MULTI-P.I. STRUCTURE BUILT IN? 4620 03:37:36,410 --> 03:37:37,611 PI CLINICIAN, ENGINEER, BIOLOGIST, 4621 03:37:37,611 --> 03:37:39,680 THAT KIND OF THING IS JUST GOING 4622 03:37:39,680 --> 03:37:44,385 TO BE HARDBAKED INTO ANY CALLS? 4623 03:37:44,385 --> 03:37:50,224 DOES THAT MAKE SENSE? 4624 03:37:50,224 --> 03:37:55,763 >> NSF, NIH? 4625 03:37:55,763 --> 03:37:59,266 >> I THINK NIH HAS SOMETHING 4626 03:37:59,266 --> 03:38:04,238 LIKE THAT ON RM1, IS KAREN DAVID 4627 03:38:04,238 --> 03:38:05,673 IN THE AUDIENCE? MAYBE YOU 4628 03:38:05,673 --> 03:38:08,742 COULD SPEAK TO THAT. SHE PIONEERED 4629 03:38:08,742 --> 03:38:13,013 THE COLLABORATIVE OPPORTUNITIES 4630 03:38:13,013 --> 03:38:13,981 FOR MULTI-DISCIPLINARY, BOLD, 4631 03:38:13,981 --> 03:38:15,316 AND INNOVATIVE NEUROSCIENCE. 4632 03:38:15,316 --> 03:38:17,284 >> THE INSPIRATION WAS OUR BRAIN 4633 03:38:17,284 --> 03:38:18,919 INITIATIVE TEAM SCIENCE PROGRAM, 4634 03:38:18,919 --> 03:38:21,188 THE BRAIN CIRCUITS PORTFOLIO. 4635 03:38:21,188 --> 03:38:23,724 WE HAVE DIFFERENT SCALES AND 4636 03:38:23,724 --> 03:38:25,392 SIZES, ONE IS FOR INDIVIDUAL LABS, 4637 03:38:25,392 --> 03:38:27,795 THE OTHER SIDE IS PURELY FOR 4638 03:38:27,795 --> 03:38:30,731 TEAM SCIENCE WHERE WE REQUIRE AT 4639 03:38:30,731 --> 03:38:33,567 LEAST THREE UP TO SIX 4640 03:38:33,567 --> 03:38:35,436 MULTI-P.I.s, AND BY DESIGN THE 4641 03:38:35,436 --> 03:38:39,740 SCIENTIST WOULD NEED AN 4642 03:38:39,740 --> 03:38:40,941 EXPERIMENTALIST, A COMPUTATIONAL 4643 03:38:40,941 --> 03:38:41,909 NEUROBIOLOGIST, AND OTHER 4644 03:38:41,909 --> 03:38:44,979 EXPERTISE THAT'S NEEDED TO 4645 03:38:44,979 --> 03:38:46,080 ANSWER THE PROPOSED QUESTIONS OR 4646 03:38:46,080 --> 03:38:48,415 PROJECT IN MIND. 4647 03:38:48,415 --> 03:38:50,584 SO THE SCIENCE NECESSITATES THE 4648 03:38:50,584 --> 03:38:51,919 CONVERGENCE OF THE EXPERTISE 4649 03:38:51,919 --> 03:38:55,622 THAT'S NEEDED TO SOLVE THAT. 4650 03:38:55,622 --> 03:38:56,623 AND THAT SERVES AS INSPIRATION 4651 03:38:56,623 --> 03:39:01,261 FOR NINDS TEAM SCIENCE PROGRAM, 4652 03:39:01,261 --> 03:39:02,997 WHICH IS THE RM1 COMBINE 4653 03:39:02,997 --> 03:39:04,198 PROGRAM, MORE GENERAL, NOT 4654 03:39:04,198 --> 03:39:06,066 PRESCRIBED WHAT TYPE OF 4655 03:39:06,066 --> 03:39:07,701 COLLABORATION OR TYPE OF 4656 03:39:07,701 --> 03:39:09,770 SCIENCE, IT'S MORE WHATEVER IS 4657 03:39:09,770 --> 03:39:12,072 IN THE NINDS SPACE. 4658 03:39:12,072 --> 03:39:13,374 THAT ONE WE REQUIRE SIMILAR, 4659 03:39:13,374 --> 03:39:21,181 THREE TO SIX P.I.s, HAS TO BE 4660 03:39:21,181 --> 03:39:21,749 MULTIDISCIPLINARY OR 4661 03:39:21,749 --> 03:39:22,383 TRANSDISCIPLINARY, CONVERGENC OF 4662 03:39:22,383 --> 03:39:26,553 EXPERTISE IS NECESSARY TO SOLVE 4663 03:39:26,553 --> 03:39:28,622 THE PROPOSED QUESTION. 4664 03:39:28,622 --> 03:39:30,157 TEAM SCIENCE BY DESIGN, WE 4665 03:39:30,157 --> 03:39:32,793 REQUIRE THEM TO THINK ABOUT 4666 03:39:32,793 --> 03:39:33,994 GLUES THAT SERVE TO BRIDGE 4667 03:39:33,994 --> 03:39:35,195 EXPERTISE AT LEVEL OF SCIENCE, 4668 03:39:35,195 --> 03:39:37,164 THEN AT THE LEVEL OF THE PEOPLE, 4669 03:39:37,164 --> 03:39:38,799 LEVEL OF THE APPROACHES. 4670 03:39:38,799 --> 03:39:40,334 I THINK SOCIAL ASPECTS OF THE 4671 03:39:40,334 --> 03:39:42,636 TEAM MANAGEMENT PLAN IS AS 4672 03:39:42,636 --> 03:39:48,776 IMPORTANT AS THE SCIENCE BEING 4673 03:39:48,776 --> 03:39:49,343 PROPOSED. 4674 03:39:49,343 --> 03:39:51,745 >> LET'S GO TO THE RIGHT SIDE 4675 03:39:51,745 --> 03:39:52,046 HERE. 4676 03:39:52,046 --> 03:39:55,115 WE HAVE TO FINISH IN FIVE 4677 03:39:55,115 --> 03:39:55,349 MINUTES. 4678 03:39:55,349 --> 03:39:57,985 GO AHEAD. 4679 03:39:57,985 --> 03:39:58,952 >> THANK YOU. 4680 03:39:58,952 --> 03:40:07,094 I FULLY UNDERSTAND FOR THE 4681 03:40:07,094 --> 03:40:07,961 NEUROMORPHIC COMPUTING AND 4682 03:40:07,961 --> 03:40:12,666 NEUROAI COMMUNITY, MANY PEOPLE 4683 03:40:12,666 --> 03:40:17,071 FROM HARDWARE, SOFTWARE AND BIOLOGY 4684 03:40:17,071 --> 03:40:20,007 I'M WONDERING FOR THE AGENCY 4685 03:40:20,007 --> 03:40:23,977 THAT, FOR EXAMPLE, FOR THE NSF 4686 03:40:23,977 --> 03:40:28,682 TO BUILD A NEW PROGRAM SPECIFIC 4687 03:40:28,682 --> 03:40:33,287 TO NEUROAI, AND MAYBE FOR 4688 03:40:33,287 --> 03:40:34,288 EXAMPLE LIKE DIFFERENT THINGS 4689 03:40:34,288 --> 03:40:36,023 LIKE THAT HAVE THEIR OWN 4690 03:40:36,023 --> 03:40:38,192 FOUNDATION, SO IF WE ARE ALL 4691 03:40:38,192 --> 03:40:43,797 INTERESTED IN THESE KIND OF 4692 03:40:43,797 --> 03:40:47,835 NEUROAI OR NEUROMORPHIC APPROACH, 4693 03:40:47,835 --> 03:40:49,503 POSSIBLE ALL AGENCIES TO WORK 4694 03:40:49,503 --> 03:40:50,604 WITH THE COMMUNITY TO BUILD KIND 4695 03:40:50,604 --> 03:40:54,141 OF THE FOUNDATION FOR THIS. 4696 03:40:54,141 --> 03:40:57,211 THANK YOU. 4697 03:40:57,211 --> 03:40:57,344 4698 03:40:57,344 --> 03:41:02,716 >> THAT WOULD BE WONDERFUL. 4699 03:41:02,716 --> 03:41:05,419 BUT UNFORTUNATELY IT TAKES A LOT 4700 03:41:05,419 --> 03:41:06,887 OF RESOURCES, AND A LOT OF 4701 03:41:06,887 --> 03:41:08,288 POLITICS IN ORDER TO GET 4702 03:41:08,288 --> 03:41:09,723 SOMETHING LIKE THAT STARTED. 4703 03:41:09,723 --> 03:41:11,892 THE BIGGER THE THING YOU TRY TO 4704 03:41:11,892 --> 03:41:12,893 START, THE LONGER IT'S GOING 4705 03:41:12,893 --> 03:41:13,527 TO TAKE. 4706 03:41:13,527 --> 03:41:14,862 WE'RE TRYING TO START HERE FROM 4707 03:41:14,862 --> 03:41:15,395 GRASSROOTS. 4708 03:41:15,395 --> 03:41:19,032 IN OTHER WORDS, WE HAVE -- IT IS 4709 03:41:19,032 --> 03:41:19,333 INTERESTING. 4710 03:41:19,333 --> 03:41:21,435 THERE SHOULD BE PROBABLY 4711 03:41:21,435 --> 03:41:26,773 SOMEWHERE A SPECIFIC PROGRAM FOR 4712 03:41:26,773 --> 03:41:27,107 NEUROAI. 4713 03:41:27,107 --> 03:41:30,177 THAT WOULD BE SOME AGENCY, MAYBE 4714 03:41:30,177 --> 03:41:33,013 FOUNDATION, SHOULD ACTUALLY TAKE 4715 03:41:33,013 --> 03:41:36,984 THAT ON SO THE KERNEL IS THERE 4716 03:41:36,984 --> 03:41:42,055 FOR -- YOU KNOW, MAYBE IT'S 4717 03:41:42,055 --> 03:41:44,224 ALREADY THERE, WE HAVEN'T 4718 03:41:44,224 --> 03:41:45,993 FERTILIZED IT YET. 4719 03:41:45,993 --> 03:41:47,461 >> COULD I JUMP IN? 4720 03:41:47,461 --> 03:41:49,696 YOU KNOW, GOING BACK TO 4721 03:41:49,696 --> 03:41:50,998 SOMETHING WHERE I PLEADED WITH 4722 03:41:50,998 --> 03:41:53,534 THE AUDIENCE WE NEED TO HEAR 4723 03:41:53,534 --> 03:41:54,835 WHAT YOU WANT, RIGHT? 4724 03:41:54,835 --> 03:41:57,070 EARLIER THERE WAS A TALK ABOUT 4725 03:41:57,070 --> 03:41:58,605 THESE SMALLER TARGETED PROGRAMS 4726 03:41:58,605 --> 03:41:59,606 THAT DON'T LAST FOREVER. 4727 03:41:59,606 --> 03:42:04,411 A LOT OF THOSE SMALLER TARGETED 4728 03:42:04,411 --> 03:42:08,215 PROGRAMS CAN LEAD TO BIGGER 4729 03:42:08,215 --> 03:42:08,749 THINGS. 4730 03:42:08,749 --> 03:42:09,783 THERE'S A PIPELINE, YOU AS A 4731 03:42:09,783 --> 03:42:11,718 COMMUNITY THINK THIS IS AN AREA 4732 03:42:11,718 --> 03:42:14,588 WORTH INVESTING IN, REACHING 4733 03:42:14,588 --> 03:42:16,156 OUT, TALKING TO PEOPLE, IN 4734 03:42:16,156 --> 03:42:17,624 FUNDING PROGRAMS, WHETHER IN THE 4735 03:42:17,624 --> 03:42:21,028 GOVERNMENT OR IN PRIVATE 4736 03:42:21,028 --> 03:42:22,863 FUNDERS, SAY THIS IS IMPORTANT. 4737 03:42:22,863 --> 03:42:26,833 THERE'S A PATH TO EVENTUALLY 4738 03:42:26,833 --> 03:42:29,903 GENERATING SOMETHING. 4739 03:42:29,903 --> 03:42:32,506 SO REACH OUT TO ALL OF US, 4740 03:42:32,506 --> 03:42:33,941 SAYING THIS IS IMPORTANT, 4741 03:42:33,941 --> 03:42:35,242 SOMETHING WE WANT, THINGS MOVE 4742 03:42:35,242 --> 03:42:36,910 SLOWLY BUT THEY DO MOVE. 4743 03:42:36,910 --> 03:42:39,513 >> SCAN THE QR CODE AND SEND US 4744 03:42:39,513 --> 03:42:40,013 YOUR THOUGHTS. 4745 03:42:40,013 --> 03:42:40,814 >> YES. 4746 03:42:40,814 --> 03:42:42,349 >> WE HAVE ONE MORE QUESTION IN 4747 03:42:42,349 --> 03:42:49,489 THE ROOM BEFORE WE HAVE TO WRAP 4748 03:42:49,489 --> 03:42:49,856 UP. 4749 03:42:49,856 --> 03:42:53,760 >> JULIE, COMING TO LONDON. 4750 03:42:53,760 --> 03:43:01,902 YOU'VE MENTIONED INTERDISCIPLINARY 4751 03:43:01,902 --> 03:43:02,803 AND INTERNATIONAL COLLABORATION. 4752 03:43:02,803 --> 03:43:05,606 THE BRAIN INITIATIVE PIVOTAL IN 4753 03:43:05,606 --> 03:43:09,109 DEVELOPING MY RESEARCH QUESTION 4754 03:43:09,109 --> 03:43:11,211 AND DOING WELL, MY QUESTION IS WHAT 4755 03:43:11,211 --> 03:43:16,116 SORT OF FUNDING YOU HAVE FOR AN 4756 03:43:16,116 --> 03:43:17,951 INTERNATIONAL SCHOLAR, AND WHAT 4757 03:43:17,951 --> 03:43:22,389 CAN YOU DO TO ACTUALLY HELP 4758 03:43:22,389 --> 03:43:24,124 PEOPLE ABROAD TO EXPAND RESEARCH 4759 03:43:24,124 --> 03:43:25,425 IF THEY ARE ACTUALLY DRAWING ON 4760 03:43:25,425 --> 03:43:30,130 SOME OF THE DATA THAT THE BRAIN 4761 03:43:30,130 --> 03:43:33,200 INITIATIVE IS ACTUALLY 4762 03:43:33,200 --> 03:43:33,533 DISSEMINATING. 4763 03:43:33,533 --> 03:43:34,167 THANK YOU. 4764 03:43:34,167 --> 03:43:35,836 >> WITHIN THE DEPARTMENT OF 4765 03:43:35,836 --> 03:43:37,571 DEFENSE, SO ALL THREE SERVICES, 4766 03:43:37,571 --> 03:43:40,207 ARMY, NAVY, AIR FORCE, HAVE 4767 03:43:40,207 --> 03:43:40,907 INTERNATIONAL OFFICES. 4768 03:43:40,907 --> 03:43:42,843 AND SO I THINK ALL THREE, YOU 4769 03:43:42,843 --> 03:43:45,445 TALKED ABOUT THINGS IN LONDON, 4770 03:43:45,445 --> 03:43:47,114 ALL THREE HAVE OFFICES IN 4771 03:43:47,114 --> 03:43:51,051 LONDON, BUT I KNOW FOR THE AIR 4772 03:43:51,051 --> 03:43:56,123 FORCE WE'VE GOT OFFICES IN TOKYO 4773 03:43:56,123 --> 03:44:01,261 SANTIAGO, CHILE, SAO PAULO, BRAZIL 4774 03:44:01,261 --> 03:44:02,763 MELBOURNE, AUSTRALIA. ARMY HAS 4775 03:44:02,763 --> 03:44:04,631 OWN LOCATIONS SOME OVERLAP. 4776 03:44:04,631 --> 03:44:07,367 SAME WITH ONR GLOBAL FOR THE NAVY. 4777 03:44:07,367 --> 03:44:09,803 SO, WE FUND SOME INTERNATIONAL 4778 03:44:09,803 --> 03:44:11,004 WORK FROM WHERE WE'RE BASED IN 4779 03:44:11,004 --> 03:44:14,174 THE U.S. 4780 03:44:14,174 --> 03:44:16,576 BUT ALSO THE INTERNATIONAL 4781 03:44:16,576 --> 03:44:18,979 PROGRAM OFFICERS ARE OUR EYES 4782 03:44:18,979 --> 03:44:20,614 AND EARS ABROAD, FOLLOWING 4783 03:44:20,614 --> 03:44:22,482 TRENDS, SCIENCE IN THEIR REGION, 4784 03:44:22,482 --> 03:44:23,784 HELPING TO INFORM US BUT THEY 4785 03:44:23,784 --> 03:44:25,085 ALSO HAVE THEIR OWN BUDGET SO 4786 03:44:25,085 --> 03:44:27,187 THEY ARE ABLE TO FUND WORK FROM 4787 03:44:27,187 --> 03:44:28,388 THERE BUT THEY CAN HELP DIRECT 4788 03:44:28,388 --> 03:44:31,758 YOU TO THE RIGHT PLACES AND MAKE 4789 03:44:31,758 --> 03:44:35,595 SURE THAT YOU'RE ABLE TO 4790 03:44:35,595 --> 03:44:37,864 NAVIGATE OUR PROCESSES. 4791 03:44:37,864 --> 03:44:39,766 >> THANK YOU. 4792 03:44:39,766 --> 03:44:40,867 >> SIMONS FOUNDATION FUNDS 4793 03:44:40,867 --> 03:44:42,035 INTERNATIONALLY BY DESIGN AND 4794 03:44:42,035 --> 03:44:45,772 MANY PRIVATE FOUNDATIONS DO AS 4795 03:44:45,772 --> 03:44:47,274 WELL. 4796 03:44:47,274 --> 03:44:48,275 >> SO TO FINISH, THE DEPARTMENT 4797 03:44:48,275 --> 03:44:53,313 OF ENERGY DOES IT HAVE 4798 03:44:53,313 --> 03:44:58,018 INTERNATIONAL PROGRAMS? 4799 03:44:58,018 --> 03:45:02,522 4800 03:45:02,522 --> 03:45:08,996 >> RELATED TO COMPUTATIONAL 4801 03:45:08,996 --> 03:45:14,401 NEUROSCIENCE, WE ARE PART OF CRCNS 4802 03:45:14,401 --> 03:45:15,602 WITH NSF, NIH AND DOE, SPAIN, GERMANY, ISRAEL, JAPAN, AND FRANCE. 4803 03:45:15,602 --> 03:45:20,407 >> ONE LAST THING IF I COULD. 4804 03:45:20,407 --> 03:45:22,943 WE'RE BY NATURE INTERNATIONAL, 4805 03:45:22,943 --> 03:45:27,748 OPENGPU IS TRADEMARKED 4806 03:45:27,748 --> 03:45:28,181 INTERNATIONAL. 4807 03:45:28,181 --> 03:45:29,516 WE BELIEVE INSIGHTS AND 4808 03:45:29,516 --> 03:45:30,050 DEVELOPMENT COMES FROM 4809 03:45:30,050 --> 03:45:30,350 EVERYWHERE. 4810 03:45:30,350 --> 03:45:33,553 >> THIS IS A GOOD WAY TO END. 4811 03:45:33,553 --> 03:45:34,755 THANKS TO THE PARTICIPANTS HERE 4812 03:45:34,755 --> 03:45:36,623 AND THE AUDIENCE FOR THE 4813 03:45:36,623 --> 03:45:39,126 QUESTIONS. 4814 03:45:39,126 --> 03:45:39,393 THANK YOU. 4815 03:45:39,393 --> 03:45:43,096 [APPLAUSE] 4816 03:45:43,096 --> 03:45:50,670 >> LET'S TRANSITION TO SESSION 4817 03:45:50,670 --> 03:45:53,407 2. 4818 03:45:54,741 --> 03:45:58,178 IT IS MY GREAT PLEASURE TO 4819 03:45:58,178 --> 03:45:59,246 INTRODUCE DR. KAREN 4820 03:45:59,246 --> 03:46:00,113 ROMMELFANGER, FOUNDER AND 4821 03:46:00,113 --> 03:46:03,116 DIRECTOR OF THE INSTITUTE OF 4822 03:46:03,116 --> 03:46:04,851 NEUROETHICS THINK AND DO TANK, 4823 03:46:04,851 --> 03:46:06,820 SHE'S ALSO A FOUNDING MEMBER OF 4824 03:46:06,820 --> 03:46:10,557 THE NIH BRAIN NEUROETHICS 4825 03:46:10,557 --> 03:46:11,958 WORKING GROUP, NEWG, SINCE 2015. 4826 03:46:11,958 --> 03:46:13,927 SHE WILL TEACH US ABOUT 4827 03:46:13,927 --> 03:46:15,128 NEUROETHICS TODAY AND WHAT SHE 4828 03:46:15,128 --> 03:46:20,233 HAS TO SAY ABOUT APPLY TO ALL 4829 03:46:20,233 --> 03:46:22,702 FOUR SESSIONS. 4830 03:46:22,702 --> 03:46:23,670 >> THANK YOU, GRACE. 4831 03:46:23,670 --> 03:46:25,739 THANK YOU FOR HAVING ME HERE 4832 03:46:25,739 --> 03:46:25,972 TODAY. 4833 03:46:25,972 --> 03:46:28,074 THANK YOU, GRACE AND JOE FOR 4834 03:46:28,074 --> 03:46:29,176 YOUR EXCELLENT ORGANIZATION AND 4835 03:46:29,176 --> 03:46:30,710 TO THE EXTERNAL ADVISORS AS 4836 03:46:30,710 --> 03:46:32,679 WELL, WITH THE BIG NIH TEAM THAT 4837 03:46:32,679 --> 03:46:33,547 MADE THIS HAPPEN. 4838 03:46:33,547 --> 03:46:34,648 I'LL THRILLED TO HAVE YOUR 4839 03:46:34,648 --> 03:46:34,915 ATTENTION. 4840 03:46:34,915 --> 03:46:35,916 WE'RE IN THE PART OF THE DAY 4841 03:46:35,916 --> 03:46:38,585 WHERE YOU NEED A LITTLE I.V. 4842 03:46:38,585 --> 03:46:40,220 DOSE OF CAFFEINE BUT I'LL TRY MY 4843 03:46:40,220 --> 03:46:43,290 BEST TO GET YOU THROUGH HERE. 4844 03:46:43,290 --> 03:46:43,523 OKAY. 4845 03:46:43,523 --> 03:46:45,459 SO, WHAT I'M HOPING YOU GET BY 4846 03:46:45,459 --> 03:46:48,762 THE END OF THIS IS THAT YOU 4847 03:46:48,762 --> 03:46:49,963 UNDERSTAND SOMETHING WE'VE HAD 4848 03:46:49,963 --> 03:46:51,932 AS A CORE VALUE IN THE 4849 03:46:51,932 --> 03:46:53,133 NEUROETHICS WORKING GROUP FROM 4850 03:46:53,133 --> 03:46:54,100 ITS INCEPTION. 4851 03:46:54,100 --> 03:46:56,403 THAT IS GOOD SCIENCE IS 4852 03:46:56,403 --> 03:46:58,505 ETHICAL NEUROSCIENCE. 4853 03:46:58,505 --> 03:47:00,474 THAT MEANS THAT ETHICS IS 4854 03:47:00,474 --> 03:47:01,908 INTEGRATED THROUGHOUT THE LIFE 4855 03:47:01,908 --> 03:47:03,977 CYCLE OF DESIGNING AN IDEA 4856 03:47:03,977 --> 03:47:06,179 IDEA, CARRYING OUT THE RESEARCH, 4857 03:47:06,179 --> 03:47:10,016 AND ALSO THROUGH DISSEMINATION 4858 03:47:10,016 --> 03:47:11,651 AND TRANSLATION OF FINDINGS. 4859 03:47:11,651 --> 03:47:14,921 SO I WON'T SHOW THE SAME 4860 03:47:14,921 --> 03:47:15,689 VIRTUOUS CYCLE THING EVERYONE 4861 03:47:15,689 --> 03:47:18,024 ELSE HAS SHOWN BUT IN A 4862 03:47:18,024 --> 03:47:19,993 DIFFERENT WAY, WHICH IS THE 4863 03:47:19,993 --> 03:47:22,195 REASON WHY THIS SPACE AND WHY 4864 03:47:22,195 --> 03:47:24,498 NEUROETHICS AS A FIELD EXISTS IS 4865 03:47:24,498 --> 03:47:26,800 BECAUSE THE BRAIN IS NOT JUST 4866 03:47:26,800 --> 03:47:28,902 BIOLOGICALLY SPECIAL. 4867 03:47:28,902 --> 03:47:30,337 IT'S CULTURALLY SPECIAL. 4868 03:47:30,337 --> 03:47:33,707 SO, AT LEAST FOR OVER 2500 YEARS 4869 03:47:33,707 --> 03:47:36,376 WE'VE THOUGHT OF THE BRAIN AS 4870 03:47:36,376 --> 03:47:38,778 THE PLACE FROM WHICH 4871 03:47:38,778 --> 03:47:40,547 INTELLIGENCE DERIVES, A PLACE 4872 03:47:40,547 --> 03:47:43,149 TODAY THAT'S THE HOME OF 4873 03:47:43,149 --> 03:47:45,051 PERSONAL IDENTITY, OF THOUGHTS, 4874 03:47:45,051 --> 03:47:47,187 EMOTIONS, HUMAN EXPERIENCE, OF 4875 03:47:47,187 --> 03:47:50,290 THE THINGS THAT WE PRIZE MOST IN 4876 03:47:50,290 --> 03:47:51,458 LIFE. 4877 03:47:51,458 --> 03:47:53,560 AND THE MOST ADVANCED TECHNOLOGY 4878 03:47:53,560 --> 03:47:55,195 THROUGHOUT HISTORY OF TIME HAS 4879 03:47:55,195 --> 03:47:58,665 ALWAYS BEEN COMPARED TO THE 4880 03:47:58,665 --> 03:47:59,466 BRAIN. 4881 03:47:59,466 --> 03:48:01,434 SO, ONE EXAMPLE, THE BRAIN WAS A 4882 03:48:01,434 --> 03:48:01,701 TELEGRAPH. 4883 03:48:01,701 --> 03:48:04,371 NOW IT'S A COMPUTER. 4884 03:48:04,371 --> 03:48:06,706 NOW THE TWIST IS THAT THE MOST 4885 03:48:06,706 --> 03:48:07,340 SOPHISTICATED TECHNOLOGY MIGHT 4886 03:48:07,340 --> 03:48:09,543 BE THE BRAIN AND COMPUTER NEEDS 4887 03:48:09,543 --> 03:48:12,712 TO BE MORE LIKE THE BRAIN. 4888 03:48:12,712 --> 03:48:14,781 THAT MEANS ANYTHING INTERVENING 4889 03:48:14,781 --> 03:48:17,417 WITH THIS BIOLOGICALLY AND 4890 03:48:17,417 --> 03:48:18,618 CULTURALLY SPECIAL ORGAN MEANS 4891 03:48:18,618 --> 03:48:21,254 THAT IT BECOMES EXTRA ETHICALLY 4892 03:48:21,254 --> 03:48:21,922 COMPLEX. 4893 03:48:21,922 --> 03:48:25,525 AND HERE IN NEUROAI WE BRING 4894 03:48:25,525 --> 03:48:27,060 TOGETHER TWO FIELDS, ACTUALLY, 4895 03:48:27,060 --> 03:48:27,961 SURPRISINGLY THAT DON'T REALLY 4896 03:48:27,961 --> 03:48:29,596 WORK TOGETHER THAT MUCH, MAYBE 4897 03:48:29,596 --> 03:48:33,166 IN WAYS THAT OFTEN WE DON'T SEE 4898 03:48:33,166 --> 03:48:35,602 NEUROSCIENCE AND A.I. SCIENCES 4899 03:48:35,602 --> 03:48:37,270 WORK TOGETHER. 4900 03:48:37,270 --> 03:48:39,205 NEUROETHICS BROADLY EXPLORES 4901 03:48:39,205 --> 03:48:41,174 ETHICAL, LEGAL, SOCIAL 4902 03:48:41,174 --> 03:48:42,609 IMPLICATIONS OF NEUROSCIENCE IN 4903 03:48:42,609 --> 03:48:43,677 APPLICATION. 4904 03:48:43,677 --> 03:48:44,911 A.I. ETHICS DOES SIMILAR THINGS 4905 03:48:44,911 --> 03:48:47,213 AROUND DEVELOPMENT OF A.I. AND 4906 03:48:47,213 --> 03:48:52,352 APPLICATIONS AND IT'S A FIELD 30 4907 03:48:52,352 --> 03:48:55,221 YEARS OLDER THAN NEUROETHICS. 4908 03:48:55,221 --> 03:48:57,624 IT HAS ROUTES IN COMPUTER ETHICS 4909 03:48:57,624 --> 03:48:59,492 ROBOT ETHICS. MOST OF YOU FAMILIAR 4910 03:48:59,492 --> 03:49:00,927 WITH TOPICS IN THE BLUE THAT 4911 03:49:00,927 --> 03:49:02,662 ADDRESS A.I. ETHICS. 4912 03:49:02,662 --> 03:49:08,168 THIS IS NOT EXHAUSTIVE BUT 4913 03:49:08,168 --> 03:49:09,569 THINGS LIKE FAIRNESS AND 4914 03:49:09,569 --> 03:49:11,438 EXPLAINABILITY. FOR NEUROETHICS 4915 03:49:11,438 --> 03:49:16,142 RELATED TO INDIVIDUAL EVENT, 4916 03:49:16,142 --> 03:49:17,377 FREE WILL, MENTAL PRIVACY, 4917 03:49:17,377 --> 03:49:18,445 PERSONAL IDENTITY. 4918 03:49:18,445 --> 03:49:19,980 THE INTERSECTION WE SEE THINGS 4919 03:49:19,980 --> 03:49:21,748 RELATED TO SAFETY, HUMAN RIGHTS, 4920 03:49:21,748 --> 03:49:23,717 CHANGES TO THE STATUS QUO, 4921 03:49:23,717 --> 03:49:26,252 CHALLENGES TO HUMAN NATURE, AND 4922 03:49:26,252 --> 03:49:28,088 AGENCY AND AUTONOMY WHICH WE'VE 4923 03:49:28,088 --> 03:49:31,925 HEARD MANY WORDS THROWN AROUND 4924 03:49:31,925 --> 03:49:32,258 TODAY. 4925 03:49:32,258 --> 03:49:34,227 AS WE THINK ABOUT THESE ISSUES 4926 03:49:34,227 --> 03:49:38,264 THERE'S MANY FLAVORS OF ETHICAL 4927 03:49:38,264 --> 03:49:39,899 ISSUES LUMPED IN THE VENN 4928 03:49:39,899 --> 03:49:40,800 DIAGRAM, CATEGORIZED IN A NUMBER 4929 03:49:40,800 --> 03:49:42,102 OF WAYS. 4930 03:49:42,102 --> 03:49:44,437 THE WAY WE CATEGORIZE, AND THEY 4931 03:49:44,437 --> 03:49:45,839 OVERLAP BY THE WAY, WHAT BUCKET 4932 03:49:45,839 --> 03:49:47,374 WE PUT THEM IN MIGHT DETERMINE 4933 03:49:47,374 --> 03:49:48,675 THE KIND OF INTERVENTION WE 4934 03:49:48,675 --> 03:49:50,877 WOULD HAVE TO ADDRESS THAT 4935 03:49:50,877 --> 03:49:51,978 ETHICAL ISSUE. 4936 03:49:51,978 --> 03:49:54,581 SO THERE'S ETHICAL ISSUES 4937 03:49:54,581 --> 03:49:56,149 RELATED TO HOW TECHNOLOGY WORKS. 4938 03:49:56,149 --> 03:49:59,853 WE MIGHT BE TALKING ABOUT 4939 03:49:59,853 --> 03:50:00,286 ENVIRONMENT IMPACT, 4940 03:50:00,286 --> 03:50:03,089 TRANSPARENCY, AND THERE ARE 4941 03:50:03,089 --> 03:50:03,823 THOSE THAT ARE ETHICAL-LEGAL, 4942 03:50:03,823 --> 03:50:06,226 HOW NEUROAI IS GOVERNED, 4943 03:50:06,226 --> 03:50:10,964 THERE ARE THOSE THAT ARE HOW 4944 03:50:10,964 --> 03:50:12,365 NEUROAI IS APPLIED, INDIVIDUAL 4945 03:50:12,365 --> 03:50:13,900 MEDICAL CONTEXT OR MAKING 4946 03:50:13,900 --> 03:50:15,101 ASSUMPTIONS ABOUT BROAD GROUPS 4947 03:50:15,101 --> 03:50:17,737 OF PEOPLE AND DEFINING DISEASE. 4948 03:50:17,737 --> 03:50:18,471 THEN THERE'S RELATIONAL 4949 03:50:18,471 --> 03:50:20,774 METAPHYSICAL, THESE ARE MORE 4950 03:50:20,774 --> 03:50:23,743 AROUND HOW WE CONCEPTUALIZE 4951 03:50:23,743 --> 03:50:24,944 OURSELVES, CHANGING WAYS WE 4952 03:50:24,944 --> 03:50:26,146 INTERACT WITH ONE ANOTHER, HOW 4953 03:50:26,146 --> 03:50:29,649 WE MIGHT EVEN THINK ABOUT 4954 03:50:29,649 --> 03:50:32,185 TECHNOLOGY ITSELF. 4955 03:50:32,185 --> 03:50:35,021 SO AS AN ETHICIST, WHO HAS 4956 03:50:35,021 --> 03:50:36,790 WORKED AS A FORMER SCIENTIST 4957 03:50:36,790 --> 03:50:40,393 TURNED ETHICIST AND HAVING 4958 03:50:40,393 --> 03:50:44,364 WORKED WITH MANY SCIENTISTS, 4959 03:50:44,364 --> 03:50:45,732 IT'S EASY, IT'S TEMPTING TO ASK 4960 03:50:45,732 --> 03:50:47,300 CAN YOU GIVE ME A LIST OF THING 4961 03:50:47,300 --> 03:50:49,135 I CAN DO, A CHECKLIST, CAN'T 4962 03:50:49,135 --> 03:50:50,503 TELL YOU HOW MANY TIMES I'VE 4963 03:50:50,503 --> 03:50:51,137 HEARD THAT. 4964 03:50:51,137 --> 03:50:53,540 I WISH IT WERE THAT SIMPLE, I 4965 03:50:53,540 --> 03:50:55,175 WOULDN'T EXIST IF IT WERE BLACK 4966 03:50:55,175 --> 03:50:57,844 AND WHITE LIKE THAT. 4967 03:50:57,844 --> 03:51:01,681 SO MY INSTITUTE, ROOTS ARE IN 4968 03:51:01,681 --> 03:51:03,316 INTERNATIONAL BRAIN INITIATIVE, 4969 03:51:03,316 --> 03:51:04,718 A LARGE SCALE COLLABORATION 4970 03:51:04,718 --> 03:51:06,386 BETWEEN SEVEN EXISTING AND 4971 03:51:06,386 --> 03:51:07,454 EMERGING NATIONAL LEVEL BRAIN 4972 03:51:07,454 --> 03:51:08,755 RESEARCH PROJECTS AROUND THE 4973 03:51:08,755 --> 03:51:08,988 GLOBE. 4974 03:51:08,988 --> 03:51:11,191 AND AT THAT TIME THAT GROUP 4975 03:51:11,191 --> 03:51:12,726 WANTED TO LEVERAGE ITS RESOURCES 4976 03:51:12,726 --> 03:51:15,362 AND ITS DATA TO MAKE SURE THAT 4977 03:51:15,362 --> 03:51:18,098 THEY MOVE THE FIELD FORWARD, 4978 03:51:18,098 --> 03:51:19,099 PERHAPS ACCELERATE WITH 4979 03:51:19,099 --> 03:51:19,799 COLLABORATION. 4980 03:51:19,799 --> 03:51:23,236 AND A KEY FOUNDATIONAL PIECE 4981 03:51:23,236 --> 03:51:25,271 THAT WAS NECESSARY IN BRIDGING 4982 03:51:25,271 --> 03:51:28,174 CULTURES WAS GOING TO BE ETHICS, 4983 03:51:28,174 --> 03:51:30,043 THE VALUES UNDERLYING IT. 4984 03:51:30,043 --> 03:51:33,113 WE WORKED WITH THE SCIENTISTS, 4985 03:51:33,113 --> 03:51:33,980 ETHICISTS, CULTURAL SCHOLARS IN 4986 03:51:33,980 --> 03:51:36,282 THAT SPACE AND DEVELOPED A 4987 03:51:36,282 --> 03:51:37,684 SERIES OF QUESTIONS THAT COULD 4988 03:51:37,684 --> 03:51:41,421 BE USED TO GUIDE EACH OF THE 4989 03:51:41,421 --> 03:51:45,692 RESEARCH PROJECTS, SCIENCE AND 4990 03:51:45,692 --> 03:51:49,162 ETHICAL QUESTIONS. WE HAVE A 4991 03:51:49,162 --> 03:51:50,830 SPECIAL SERIES IN NEURON, BRAIN 4992 03:51:50,830 --> 03:51:52,265 PROJECTS TALK ABOUT HOW THEY 4993 03:51:52,265 --> 03:51:54,267 IMPLEMENT THESE. FEATURED IN THE 4994 03:51:54,267 --> 03:51:56,202 BRAIN 2.0 NEUROETHICS ROADMAP. 4995 03:51:56,202 --> 03:51:58,838 I'LL GO THROUGH WHAT I CAN, 4996 03:51:58,838 --> 03:52:00,273 EXAMPLES, AND I'LL GIVE ANYONE 4997 03:52:00,273 --> 03:52:01,574 WHO WANTS ACCESS TO THESE SLIDES 4998 03:52:01,574 --> 03:52:03,443 BECAUSE WE WON'T HAVE TIME TO GO 4999 03:52:03,443 --> 03:52:04,544 THROUGH IT ALL TODAY. 5000 03:52:04,544 --> 03:52:09,048 ONE AREA THAT WE'VE TALKED ABOUT 5001 03:52:09,048 --> 03:52:11,918 THIS MORNING WAS DIGITAL TWINS. 5002 03:52:11,918 --> 03:52:22,462 THERE'S PROJECTS RELATED TO HOW 5003 03:52:23,363 --> 03:52:24,097 BRAIN WORKS AND THOSE APPLIED TO 5004 03:52:24,097 --> 03:52:27,167 PRECISION MEDICINE IN BODY SYSTEMS. 5005 03:52:27,167 --> 03:52:28,568 QUESTIONS WE FACE WITH DIGITAL 5006 03:52:28,568 --> 03:52:30,036 TWINS ARE RELATED TO CERTAINLY 5007 03:52:30,036 --> 03:52:34,808 HOW IT WORKS, SO BIAS, WHAT 5008 03:52:34,808 --> 03:52:36,442 CONSTITUTES ACCURATE 5009 03:52:36,442 --> 03:52:38,311 REPRESENTATION, IF DATA ISOLATED 5010 03:52:38,311 --> 03:52:41,247 IN THESE VARIOUS EXPERIMENTS 5011 03:52:41,247 --> 03:52:42,615 ACTUALLY ARE SITUATED IN THE 5012 03:52:42,615 --> 03:52:43,917 CONTEXT OF PUBLIC HEALTH 5013 03:52:43,917 --> 03:52:45,919 FEATURES AND OTHER DATA THAT 5014 03:52:45,919 --> 03:52:47,220 MAKE THEN MEANINGFUL IN REAL 5015 03:52:47,220 --> 03:52:47,453 LIFE. 5016 03:52:47,453 --> 03:52:49,088 THERE ARE OTHER NEW CONCERNS 5017 03:52:49,088 --> 03:52:50,423 THAT ARISE AROUND WHAT 5018 03:52:50,423 --> 03:52:52,058 SAFEGUARDS, FOR EXAMPLE, MIGHT 5019 03:52:52,058 --> 03:52:56,429 BE IN PLACE WITH EVEN WHAT 5020 03:52:56,429 --> 03:52:57,564 PATIENTS AND PARTICIPANTS MIGHT 5021 03:52:57,564 --> 03:52:59,766 CONSENT TO IN OWNING OR GIVING 5022 03:52:59,766 --> 03:53:03,169 AWAY RIGHTS TO A DIGITAL TWIN. 5023 03:53:03,169 --> 03:53:11,845 THEN WE SEE EXAMPLES IN 5024 03:53:11,845 --> 03:53:13,046 NEUROMORPHIC INSPIRED BCIs. 5025 03:53:13,046 --> 03:53:22,922 SOMEBODY IS DRIVING MY SCREEN. 5026 03:53:22,922 --> 03:53:24,157 NEUROMORPHIC-BCI RELATING TO 5027 03:53:24,157 --> 03:53:25,758 CHALLENGES TO AUTONOMY AND 5028 03:53:25,758 --> 03:53:26,025 AGENCIES. 5029 03:53:26,025 --> 03:53:28,628 A QUESTION WE MIGHT ASK IS THERE 5030 03:53:28,628 --> 03:53:30,930 ARE MEANINGFUL DIFFERENCES IN 5031 03:53:30,930 --> 03:53:33,666 RISKS AND BENEFITS BETWEEN 5032 03:53:33,666 --> 03:53:36,636 NEUROMORPHIC BCI AND TRADITIONAL 5033 03:53:36,636 --> 03:53:38,171 BCI, IS THERE SOMETHING 5034 03:53:38,171 --> 03:53:38,671 DIFFERENT ABOUT NATIVE 5035 03:53:38,671 --> 03:53:42,442 COMMUNICATION TO THE BRAIN THAT 5036 03:53:42,442 --> 03:53:45,178 WARRANTS A DIFFERENT PROCESS HOW 5037 03:53:45,178 --> 03:53:48,882 WE MIGHT GOVERN AND HAVE FUTURE 5038 03:53:48,882 --> 03:53:49,782 CIRCUMSTANCES FOR INTEGRATING 5039 03:53:49,782 --> 03:53:50,316 TECHNOLOGIES. 5040 03:53:50,316 --> 03:53:53,519 AND THEN OF COURSE ONGOING 5041 03:53:53,519 --> 03:53:55,154 CONVERSATIONS ABOUT LONG-TERM 5042 03:53:55,154 --> 03:53:56,189 RESPONSIBILITIES TO DEVICES 5043 03:53:56,189 --> 03:53:57,490 PARTICULARLY IMPLANTED WHICH THE 5044 03:53:57,490 --> 03:53:58,458 BRAIN NEUROETHICS WORKING GROUP 5045 03:53:58,458 --> 03:54:00,426 HAS WORKED ON QUITE A BIT. 5046 03:54:00,426 --> 03:54:02,195 THE LAST QUESTION, THIS IS 5047 03:54:02,195 --> 03:54:04,163 PROBABLY ONE OF THE TOUGHEST TO 5048 03:54:04,163 --> 03:54:07,333 CRACK, THIS ONE AROUND EMBODIED 5049 03:54:07,333 --> 03:54:07,600 NEUROAI. 5050 03:54:07,600 --> 03:54:09,535 WE'VE TALKED ABOUT THE VALUE OF 5051 03:54:09,535 --> 03:54:13,139 THAT, VALUE OF THAT CREATING 5052 03:54:13,139 --> 03:54:14,340 MORE LEGITIMATE SYSTEMS TO 5053 03:54:14,340 --> 03:54:15,775 BETTER UNDERSTAND THE BRAIN, 5054 03:54:15,775 --> 03:54:17,977 CREATE A BETTER MODEL, SOMETHING 5055 03:54:17,977 --> 03:54:20,480 THAT HAS MOTIVATION, MAYBE 5056 03:54:20,480 --> 03:54:21,281 SOMETHING THAT'S BETTER 5057 03:54:21,281 --> 03:54:22,482 REPRESENTED ACROSS SCALES. 5058 03:54:22,482 --> 03:54:25,184 A FUNDAMENTAL THING HERE IS WE 5059 03:54:25,184 --> 03:54:27,086 HAVE VERY LOOSELY IN THE FIELD 5060 03:54:27,086 --> 03:54:31,491 AND OTHERS BEEN USING WORDS LIKE 5061 03:54:31,491 --> 03:54:31,991 INTELLIGENCE, BEHAVIOR, 5062 03:54:31,991 --> 03:54:32,625 MOTIVATION, CONSCIOUSNESS, AND 5063 03:54:32,625 --> 03:54:34,193 MIND, AS IF WE MEAN THE SAME 5064 03:54:34,193 --> 03:54:36,496 THING WHEN WE SAY THAT. 5065 03:54:36,496 --> 03:54:38,131 I DON'T MEAN ACROSS SCIENTISTS, 5066 03:54:38,131 --> 03:54:39,332 I MEAN ACROSS SOCIETY. 5067 03:54:39,332 --> 03:54:42,435 AND WE HAVE ALREADY SEEN THIS 5068 03:54:42,435 --> 03:54:43,736 BORNE OUT IN A SCIENTIFIC 5069 03:54:43,736 --> 03:54:46,572 COMMUNITY IN MANY WAYS. RECENT 5070 03:54:46,572 --> 03:54:47,840 EXAMPLE NOT A NEUROAI-RELATED 5071 03:54:47,840 --> 03:54:51,044 ONE, FOR THOSE THAT MAY BE AWARE 5072 03:54:51,044 --> 03:54:52,512 WITH THE BRAIN, DISH BRAIN, 5073 03:54:52,512 --> 03:54:54,247 CELLS IN A DISH THAT WERE ABLE 5074 03:54:54,247 --> 03:54:55,782 TO PLAY PONG, SAID TO BE 5075 03:54:55,782 --> 03:55:06,259 EMBODIED IN A VIRTUAL WORLD, 5076 03:55:06,592 --> 03:55:08,261 DEMONSTRATE SENTIENCE, AND WE 5077 03:55:08,261 --> 03:55:18,771 SAW CALLS FROM THE ETHICISTS. 5078 03:55:33,519 --> 03:55:39,359 A PATHWAY TO EXPLORE NEUROAI ETHICS 5079 03:55:39,359 --> 03:55:40,593 ITERATE ISSUES, PILOT SOLUTIONS 5080 03:55:40,593 --> 03:55:42,128 THRU THE RESEARCH LIFE CYCLE, NOT 5081 03:55:42,128 --> 03:55:46,165 JUST ONCE AT THE STAGE OF A 5082 03:55:46,165 --> 03:55:47,700 PROTOTYPE, HAVE EMBEDDED ETHICS 5083 03:55:47,700 --> 03:55:50,336 RESEARCH THAT CAN FACILITATE 5084 03:55:50,336 --> 03:55:54,407 THAT, TOOLS AND FRAMEWORKS. WE 5085 03:55:54,407 --> 03:55:55,274 HAVE THE NEUROETHICS PRINCIPLES, 5086 03:55:55,274 --> 03:55:56,909 THE QUESTIONS I SHOWED. I SUSPECT 5087 03:55:56,909 --> 03:56:00,780 MORE WOULD BE NEEDED IN THE NIH 5088 03:56:00,780 --> 03:56:05,151 FOR NEUROAI AND NEUROAI ENGAGEMENT 5089 03:56:05,151 --> 03:56:07,820 EXPERIENCES INVOLVING BROAD 5090 03:56:07,820 --> 03:56:08,921 GROUPS, NOT JUST SCIENTISTS, 5091 03:56:08,921 --> 03:56:10,890 ACROSS DISPRESENCE TO HAVE 5092 03:56:10,890 --> 03:56:13,760 BETTER UNDERSTANDING HOW TO 5093 03:56:13,760 --> 03:56:14,761 REPRESENT THE SCIENCE AND 5094 03:56:14,761 --> 03:56:15,962 DIRECTION THE SCIENCE SHOULD GO. 5095 03:56:15,962 --> 03:56:18,264 WITH THAT, THANK YOU FOR YOUR 5096 03:56:18,264 --> 03:56:26,005 TIME, I WILL PASS TO THE NEXT 5097 03:56:26,005 --> 03:56:26,606 PANEL. 5098 03:56:26,606 --> 03:56:29,575 >> I'D LIKE TO INTRODUCE BLAKE 5099 03:56:29,575 --> 03:56:31,310 RICHARDS FROM MILA, KICKING OFF 5100 03:56:31,310 --> 03:56:33,613 THE FIRST SESSION ON EXPLORING 5101 03:56:33,613 --> 03:56:37,316 THE STRUCTURAL AND FUNCTIONAL 5102 03:56:37,316 --> 03:56:45,324 CONVERGENCE OF DEEP NEURAL 5103 03:56:45,324 --> 03:56:46,192 NETWORKS AND BRAINS. 5104 03:56:46,192 --> 03:56:49,595 >> I WANTED TO DO SOMETHING LIKE 5105 03:56:49,595 --> 03:56:52,331 WHAT TONY DID, AND TALK ABOUT 5106 03:56:52,331 --> 03:56:53,966 WHAT THIS FIELD THAT WE HAVE 5107 03:56:53,966 --> 03:56:55,802 COME TOGETHER TO DISCUSS TODAY 5108 03:56:55,802 --> 03:56:57,603 ACTUALLY IS BECAUSE I THINK IT'S 5109 03:56:57,603 --> 03:57:01,574 NOT ALWAYS CLEAR TO EVERYONE. 5110 03:57:01,574 --> 03:57:02,942 SO, I'M GOING TO DO LIKE 5111 03:57:02,942 --> 03:57:06,245 EVERYONE ELSE DID AND PUT THIS 5112 03:57:06,245 --> 03:57:07,046 CYCLE UP. 5113 03:57:07,046 --> 03:57:10,616 BUT I'M GOING TO DO SOMETHING 5114 03:57:10,616 --> 03:57:13,686 WHICH IS MAYBE MEAN AND SAY THAT 5115 03:57:13,686 --> 03:57:14,587 IT'S VACUOUS. 5116 03:57:14,587 --> 03:57:15,888 I DON'T THINK THIS ANSWER 5117 03:57:15,888 --> 03:57:17,857 PROVIDES A GREAT DEAL OF MEANING 5118 03:57:17,857 --> 03:57:18,858 TO US. 5119 03:57:18,858 --> 03:57:21,160 SURE, THIS SOUNDS NICE, BUT IN 5120 03:57:21,160 --> 03:57:22,462 PRACTICE WHAT DOES THIS MEAN? 5121 03:57:22,462 --> 03:57:25,731 I USUALLY GIVE A STORY ABOUT HOW 5122 03:57:25,731 --> 03:57:29,969 A.I. HELPED US INTERPRET AND 5123 03:57:29,969 --> 03:57:30,837 ANALYZE DATA, NEUROSCIENCE 5124 03:57:30,837 --> 03:57:32,004 INSPIRES SOLUTIONS BUT THAT'S 5125 03:57:32,004 --> 03:57:32,772 HAND WAVY. 5126 03:57:32,772 --> 03:57:37,276 PRACTICAL ANSWER IS THAT IT'S 5127 03:57:37,276 --> 03:57:38,911 GENERALLY RESEARCH THAT USES ANN 5128 03:57:38,911 --> 03:57:43,182 TO MODEL THE BRAIN OR WHICH USES 5129 03:57:43,182 --> 03:57:45,151 NEURO IDEAS TO IMPROVE ANNs AND 5130 03:57:45,151 --> 03:57:53,292 A.I. 5131 03:57:53,292 --> 03:57:57,563 THAT MIGHT SEEM A BORING ANSWER. 5132 03:57:57,563 --> 03:58:01,300 BUT I THINK IT HIDES MUCH MORE 5133 03:58:01,300 --> 03:58:02,468 INTERESTING AND DEEPER 5134 03:58:02,468 --> 03:58:03,369 DISCIPLINARY TERM, WHAT I WANT 5135 03:58:03,369 --> 03:58:09,175 TO DESCRIBE FOR YOU TODAY. 5136 03:58:09,175 --> 03:58:12,778 SO IN THE POST WAR ERA, THERE 5137 03:58:12,778 --> 03:58:14,981 WAS A FUNDAMENTAL QUESTION. 5138 03:58:14,981 --> 03:58:17,583 SHOULD WE ENGINEER SYSTEMS BASED 5139 03:58:17,583 --> 03:58:22,855 ON INSPIRATION FROM THE BRAIN 5140 03:58:22,855 --> 03:58:24,524 ITSELF OR FROM LOGIC. 5141 03:58:24,524 --> 03:58:27,693 AT THE START IDEAS WERE UNIFIED 5142 03:58:27,693 --> 03:58:30,096 WITH THE MCCULLEN PITTS MODEL, 5143 03:58:30,096 --> 03:58:31,430 PEOPLE THOUGHT YOU COULD DO BOTH 5144 03:58:31,430 --> 03:58:33,599 AT THE SAME TIME. 5145 03:58:33,599 --> 03:58:37,770 BUT AS THE FIELD PROGRESSED, A 5146 03:58:37,770 --> 03:58:39,805 BATTLE BETWEEN TWO IDEOLOGICAL 5147 03:58:39,805 --> 03:58:41,607 CAMPS EMERGED. IN MY EDUCATION 5148 03:58:41,607 --> 03:58:43,809 IN COGNITIVE SCIENCE, I LEARNED 5149 03:58:43,809 --> 03:58:45,344 THESE CAMPS WERE REFERRED TO AS 5150 03:58:45,344 --> 03:58:48,181 GOOD OLD-FASHIONED A.I. OR 5151 03:58:48,181 --> 03:58:52,151 GOFAI, AND CYBERNETICS OR 5152 03:58:52,151 --> 03:58:54,387 CONNECTIONISM. 5153 03:58:54,387 --> 03:58:56,956 THESE TWO CAMPS HAD VERY 5154 03:58:56,956 --> 03:58:58,157 DIFFERENT ANSWERS TO HOW EXACTLY 5155 03:58:58,157 --> 03:59:03,196 YOU SHOULD TRY TO UNDERSTAND OR 5156 03:59:03,196 --> 03:59:04,397 MODEL INTELLIGENCE. GOFAI 5157 03:59:04,397 --> 03:59:07,366 FOLKS SAID LOOK INTELLIGENCE IS 5158 03:59:07,366 --> 03:59:10,636 ABOUT LOGICAL THINKING, SO CLEAR 5159 03:59:10,636 --> 03:59:11,737 SYMBOLIC STRUCTURES, LOGICAL 5160 03:59:11,737 --> 03:59:13,272 REASONING PROGRAMS, THESE ARE 5161 03:59:13,272 --> 03:59:17,743 THE WAYS TO GO. 5162 03:59:17,743 --> 03:59:22,782 EMPHASIS WAS ON PRE-PROGRAMMED, 5163 03:59:22,782 --> 03:59:24,083 HIGHLY INTERPRETABLE SYSTEMS. 5164 03:59:24,083 --> 03:59:25,084 IN CONTRAST THE IDEA IT HAD TO 5165 03:59:25,084 --> 03:59:27,820 BE DONE IN THE SAME WAY AS THE 5166 03:59:27,820 --> 03:59:30,890 BRAIN, THAT IS USING PARALLEL 5167 03:59:30,890 --> 03:59:33,092 DISTRIBUTED PROCESSING SYSTEMS. 5168 03:59:33,092 --> 03:59:34,594 AND THESE RESEARCHERS GENERALLY 5169 03:59:34,594 --> 03:59:36,362 TENDED TO EMPHASIZE LEARNING 5170 03:59:36,362 --> 03:59:38,130 FROM FEEDBACK AND CONTROL LOOPS, 5171 03:59:38,130 --> 03:59:39,966 AND SORT OF LOW LEVEL THINGS 5172 03:59:39,966 --> 03:59:41,167 LIKE MOTOR CONTROL THAT A LOT OF 5173 03:59:41,167 --> 03:59:44,170 US HAVE TALKED ABOUT TODAY. 5174 03:59:44,170 --> 03:59:45,905 SO REALLY IT COULD BE BOILED 5175 03:59:45,905 --> 03:59:48,374 DOWN TO ONE QUESTION. 5176 03:59:48,374 --> 03:59:52,378 DO WE NEED THE BRAIN TO 5177 03:59:52,378 --> 03:59:56,315 UNDERSTAND OR ENGINEER 5178 03:59:56,315 --> 03:59:56,882 INTELLIGENCE? 5179 03:59:56,882 --> 04:00:00,052 AND THE ORIGINAL KIND OF LIKE 5180 04:00:00,052 --> 04:00:03,089 GODFATHERS OF A.I., NOT MODERN 5181 04:00:03,089 --> 04:00:04,090 ONES, OLD SCHOOL ONES, LET'S 5182 04:00:04,090 --> 04:00:05,391 CALL THEM THE FOUNDERS OF A.I., 5183 04:00:05,391 --> 04:00:06,492 HAD A CLEAR ANSWER. 5184 04:00:06,492 --> 04:00:08,261 THE ANSWER WAS NO. 5185 04:00:08,261 --> 04:00:12,765 YOU DID NOT NEED THE BRAIN TO 5186 04:00:12,765 --> 04:00:13,833 ENGINEER OR UNDERSTAND 5187 04:00:13,833 --> 04:00:14,300 INTELLIGENCE. 5188 04:00:14,300 --> 04:00:17,136 THIS IS A PICTURE OF A GROUP OF 5189 04:00:17,136 --> 04:00:20,306 RESEARCHERS AT THE DARTMOUTH 5190 04:00:20,306 --> 04:00:21,841 WORKSHOP, THE FIRST WORKSHOP I 5191 04:00:21,841 --> 04:00:23,676 THINK 54 OR 55, WHERE THEY 5192 04:00:23,676 --> 04:00:25,678 COINED THE TERM ARTIFICIAL 5193 04:00:25,678 --> 04:00:27,446 INTELLIGENCE, IN FACT. 5194 04:00:27,446 --> 04:00:30,082 AND NOTABLY, THEY MADE A POINT 5195 04:00:30,082 --> 04:00:34,253 OF NOT INVITING ANY OF THE 5196 04:00:34,253 --> 04:00:35,454 CYBERNETICISTS OR CONNECTIONISTS 5197 04:00:35,454 --> 04:00:37,490 TO THEIR WORKSHOP. 5198 04:00:37,490 --> 04:00:43,229 I WAS AT A CFAR MEETING WITH 5199 04:00:43,229 --> 04:00:45,998 TERRY THE OTHER DAY, DOING A 5200 04:00:45,998 --> 04:00:46,599 CELEBRATION FOR GEOFF HINTON'S 5201 04:00:46,599 --> 04:00:47,733 NOBEL PRIZE. 5202 04:00:47,733 --> 04:00:48,934 GEOFF HAD AN INTERESTING 5203 04:00:48,934 --> 04:00:49,201 ANECDOTE. IN HIS OPINION 5204 04:00:49,201 --> 04:00:54,840 HE FELT THAT HAD TURING AND 5205 04:00:54,840 --> 04:00:56,942 VON NEUMANN SURVIVED TO 5206 04:00:56,942 --> 04:00:58,577 OLDER AGE THIS WOULD NOT HAVE 5207 04:00:58,577 --> 04:01:00,212 HAPPENED BECAUSE THEY WOULD HAVE 5208 04:01:00,212 --> 04:01:01,414 BEEN ABLE TO ADVOCATE BRAIN 5209 04:01:01,414 --> 04:01:02,715 SHOULD BE INCLUDED IN 5210 04:01:02,715 --> 04:01:03,582 UNDERSTANDING INTELLIGENCE, AND 5211 04:01:03,582 --> 04:01:05,484 THEY BOTH THOUGHT THAT AND 5212 04:01:05,484 --> 04:01:09,088 UNDERSTOOD IT AND HAD GOOD 5213 04:01:09,088 --> 04:01:09,955 LOGICAL REASONS BUT DID DIE 5214 04:01:09,955 --> 04:01:10,156 YOUNG. 5215 04:01:10,156 --> 04:01:12,258 AS A RESULT THERE WAS NO ONE TO 5216 04:01:12,258 --> 04:01:13,993 CORRECT THESE FOLKS AND SAY, NO, 5217 04:01:13,993 --> 04:01:17,396 NO, YOU HAVE TO BRING BRAINS IN. 5218 04:01:17,396 --> 04:01:18,497 SO, THEY BASICALLY IGNORED THE 5219 04:01:18,497 --> 04:01:21,634 BRAIN FOR A COUPLE DECADES, AND, 5220 04:01:21,634 --> 04:01:24,937 WELL, WE ALL KNOW WHERE THAT GOT 5221 04:01:24,937 --> 04:01:25,371 THEM. NOW, PART OF THE REASON WHY 5222 04:01:25,371 --> 04:01:30,976 IS THAT THE CYBERNETICS PEOPLE 5223 04:01:30,976 --> 04:01:32,278 AND CONNECTIONISTS PEOPLE HAVE 5224 04:01:32,278 --> 04:01:38,017 A TENDENCY TO OVERPROMISE AND 5225 04:01:38,017 --> 04:01:43,255 UNDERDELIVER. WEINER'S THE 5226 04:01:43,255 --> 04:01:44,357 HUMAN USE OF HUMAN BEINGS 5227 04:01:44,357 --> 04:01:46,759 IS A BRILLIANT BOOK, BUT 5228 04:01:46,759 --> 04:01:47,860 IT'S FAR OUT, EVERYTHING IS 5229 04:01:47,860 --> 04:01:51,564 GOING TO BE SOLVED FROM THE 5230 04:01:51,564 --> 04:01:52,598 ECONOMY TO CHEMISTRY, ELECTRICAL 5231 04:01:52,598 --> 04:01:57,703 POWER PLANTS, THE BRAIN, A.I., 5232 04:01:57,703 --> 04:01:58,904 EVERYTHING, CYBERNETIC WILL 5233 04:01:58,904 --> 04:02:01,207 SOLVE IT ALL, CONNECTIONS WILL 5234 04:02:01,207 --> 04:02:01,874 SOLVE IT ALL. 5235 04:02:01,874 --> 04:02:04,176 THAT WAS HARD TO SUPPORT IN THE 5236 04:02:04,176 --> 04:02:10,750 FACE OF SMART AND CLEAR THINKERS 5237 04:02:10,750 --> 04:02:16,122 LIKE MINKSY, PAPERT, CHOMSKY, 5238 04:02:16,122 --> 04:02:17,923 ET CETERA. 5239 04:02:17,923 --> 04:02:24,563 BUT THE STORY HAS CHANGED OVER 5240 04:02:24,563 --> 04:02:25,865 THE LAST COUPLE DECADES. 5241 04:02:25,865 --> 04:02:28,067 NOW A.I. IS ALMOST ENTIRELY 5242 04:02:28,067 --> 04:02:33,105 ASSOCIATED WITH PARALLEL 5243 04:02:33,105 --> 04:02:33,973 DISTRIBUTED PROCESSING MODELS. 5244 04:02:33,973 --> 04:02:37,810 THE GOFAI APPROACH HAS BEEN LEFT 5245 04:02:37,810 --> 04:02:38,177 BY THE WAYSIDE BOTH IN 5246 04:02:38,177 --> 04:02:41,247 ARTIFICIAL INTELLIGENCE 5247 04:02:41,247 --> 04:02:43,349 ITSELF AND ARGUABLY IN PSYCHOLOGY 5248 04:02:43,349 --> 04:02:44,417 AND NEUROSCIENCE. SO THIS BRINGS 5249 04:02:44,417 --> 04:02:54,760 ME TO WHAT I SAY NEUROAI IS, 5250 04:02:54,760 --> 04:02:56,595 THE REALIZATION OF THE ORIGINAL 5251 04:02:56,595 --> 04:02:59,832 PROMISE OF CYBERNETICS AND 5252 04:02:59,832 --> 04:03:02,902 CONNECTIONISM, IT IS THE FIELD 5253 04:03:02,902 --> 04:03:04,336 THAT ATTEMPTS TO UNDERSTAND 5254 04:03:04,336 --> 04:03:06,071 INTELLIGENCE USING PARALLEL 5255 04:03:06,071 --> 04:03:08,274 DISTRIBUTED PROCESSING SYSTEMS 5256 04:03:08,274 --> 04:03:11,577 WITH A FOCUS ON CONTROL LOOPS 5257 04:03:11,577 --> 04:03:12,845 AND LEARNING FROM FEEDBACK. 5258 04:03:12,845 --> 04:03:14,847 RATHER THAN CONTINUING TO BATTLE 5259 04:03:14,847 --> 04:03:17,349 WITH FOLKS IN A.I. AND COGNITIVE 5260 04:03:17,349 --> 04:03:18,350 SCIENCE AND OTHER PLACES, WHAT 5261 04:03:18,350 --> 04:03:21,086 MANY OF US IN THE NEUROAI FIELD 5262 04:03:21,086 --> 04:03:23,389 HAVE DONE IS SIMPLY BYPASS IT. 5263 04:03:23,389 --> 04:03:24,790 WE SAID, OKAY, WE'RE DONE WITH 5264 04:03:24,790 --> 04:03:26,325 YOU GUYS. 5265 04:03:26,325 --> 04:03:27,460 WE DON'T WANT THIS DISCUSSION 5266 04:03:27,460 --> 04:03:27,793 ANYMORE. 5267 04:03:27,793 --> 04:03:29,862 WE'RE GOING TO DO OUR OWN THING 5268 04:03:29,862 --> 04:03:31,297 AND WE'VE CREATED A NEW FIELD. 5269 04:03:31,297 --> 04:03:33,866 REALLY, THIS WAS A LESSON WE 5270 04:03:33,866 --> 04:03:35,768 LEARNED FROM THE QUOTE/UNQUOTE 5271 04:03:35,768 --> 04:03:39,705 NOW GODFATHERS OF A.I., LIKE 5272 04:03:39,705 --> 04:03:41,040 GEOFF HINTON WHO SAID, WELL, 5273 04:03:41,040 --> 04:03:43,008 OKAY, I'M GOING TO STOP HAVING 5274 04:03:43,008 --> 04:03:44,310 DEBATES WITH THESE FOLKS AND 5275 04:03:44,310 --> 04:03:46,412 JUST GOING TO ENGINEER STUFF 5276 04:03:46,412 --> 04:03:47,379 THAT WORKS. 5277 04:03:47,379 --> 04:03:49,381 AND THAT WILL WIN THE DAY. 5278 04:03:49,381 --> 04:03:52,751 >> ONE MINUTE WARNING. 5279 04:03:52,751 --> 04:03:53,586 >> YEP. 5280 04:03:53,586 --> 04:03:54,186 THANKS. 5281 04:03:54,186 --> 04:03:57,223 SO, WITHIN THIS NEUROAI PROGRAM 5282 04:03:57,223 --> 04:04:01,427 WE HAVE I WOULD ARGUE A LAKATOSIAN 5283 04:04:01,427 --> 04:04:05,564 RESEARCH PROGRAM, AN IDEA MY 5284 04:04:05,564 --> 04:04:08,534 COLLEAGUES ADVOCATED FOR IN THIS 5285 04:04:08,534 --> 04:04:12,004 PAPER ON NEUROCONNECTIONISM, 5286 04:04:12,004 --> 04:04:14,540 IT'S A BEAUTIFUL CONCEPT. 5287 04:04:14,540 --> 04:04:16,208 THE IDEA WITH THE LAKATOSIAN 5288 04:04:16,208 --> 04:04:19,044 RESEARCH PROGRAM IS THAT UNLIKE 5289 04:04:19,044 --> 04:04:21,914 THE EXTREME FALSIFIABILITY OF 5290 04:04:21,914 --> 04:04:28,354 KARL POPPER PERSPECTIVE OR IT'S 5291 04:04:28,354 --> 04:04:30,890 ABOUT PARADIGMS LIKE KUNIAN 5292 04:04:30,890 --> 04:04:32,191 SCIENCE EVOLVES WITH CORE 5293 04:04:32,191 --> 04:04:33,759 ASSUMPTIONS CRITICAL FOR MAKING 5294 04:04:33,759 --> 04:04:35,494 PROCESS, A SERIES OF AUXILIARY 5295 04:04:35,494 --> 04:04:39,732 HYPOTHESES THAT ARE OPEN FOR 5296 04:04:39,732 --> 04:04:46,639 TESTING AND FALSIFICATION, YES, 5297 04:04:46,639 --> 04:04:48,607 INTELLIGENCE IS BEST MODELED WITH 5298 04:04:48,607 --> 04:04:49,375 PARALLEL DISTRIBUTED PROCESSING 5299 04:04:49,375 --> 04:04:51,577 SYSTEMS AND FOCUS ON CONTROL AND 5300 04:04:51,577 --> 04:04:51,844 LEARNING. 5301 04:04:51,844 --> 04:04:53,646 AND THEN THERE'S A SERIES OF 5302 04:04:53,646 --> 04:04:56,382 HYPOTHESES AROUND THAT WE CAN 5303 04:04:56,382 --> 04:04:58,450 TEST AND FALSIFY. 5304 04:04:58,450 --> 04:04:59,885 SO RAPIDLY WE CAN START WITH 5305 04:04:59,885 --> 04:05:06,792 THAT CORE AND BUILD UP THINGS 5306 04:05:06,792 --> 04:05:09,395 LIKE NEOCOGNITRON OR CNNs TO MODEL 5307 04:05:09,395 --> 04:05:10,996 THE VENTRAL VISUAL STREAM BUT 5308 04:05:10,996 --> 04:05:13,232 WE'LL DISCOVER PROBLEMS AS WE 5309 04:05:13,232 --> 04:05:15,000 HAVE WITH CNNs. WE CAN FALSIFY 5310 04:05:15,000 --> 04:05:17,303 AND MODIFY AND MOVE THE FIELD 5311 04:05:17,303 --> 04:05:20,706 FORWARD WITHOUT DITCHING THAT 5312 04:05:20,706 --> 04:05:21,373 CORE ESSENCE OF THE PARADIGM. 5313 04:05:21,373 --> 04:05:23,342 SO I SHOULD END NOW BECAUSE I 5314 04:05:23,342 --> 04:05:24,209 HAVE FOUR SECONDS. 5315 04:05:24,209 --> 04:05:32,418 THANK YOU VERY MUCH. 5316 04:05:32,418 --> 04:05:36,488 >> I'D LIKE TO INTRODUCE THE 5317 04:05:36,488 --> 04:05:40,793 NEXT SPEAKER, DORIS TSAO FROM 5318 04:05:40,793 --> 04:05:51,036 UC-BERKELEY. 5319 04:05:51,036 --> 04:05:54,106 >> GREAT TO BE HERE. 5320 04:05:54,106 --> 04:05:55,541 SO, TODAY I WANT TO TALK ABOUT 5321 04:05:55,541 --> 04:06:03,248 THE CURRENT STATE OF NEUROAI. 5322 04:06:03,248 --> 04:06:06,051 SORRY ABOUT THAT. 5323 04:06:06,051 --> 04:06:07,353 IS THAT BETTER? 5324 04:06:07,353 --> 04:06:08,120 ALL RIGHT. 5325 04:06:08,120 --> 04:06:09,288 YES I FIRST WANT TO TALK ABOUT 5326 04:06:09,288 --> 04:06:11,490 CURRENT STATE OF NEUROAI BY SHARING 5327 04:06:11,490 --> 04:06:12,925 HIGHLIGHTS FROM OUR WORK IN THE 5328 04:06:12,925 --> 04:06:15,427 SPACE, AND THEN I WANT TO TELL 5329 04:06:15,427 --> 04:06:17,496 YOU ABOUT MY VISION FOR THE 5330 04:06:17,496 --> 04:06:18,564 FUTURE OF NEUROAI. 5331 04:06:18,564 --> 04:06:19,598 DEEP NETWORKS ARE POWERFUL 5332 04:06:19,598 --> 04:06:23,002 BECAUSE THEY LET US DESCRIBE 5333 04:06:23,002 --> 04:06:27,473 THINGS HARD TO DESCRIBE. 5334 04:06:27,473 --> 04:06:28,907 ORIENTATION OF AN EDGE, 5335 04:06:28,907 --> 04:06:34,213 SIMULATE DIFFUSION OF SMOKE BUT 5336 04:06:34,213 --> 04:06:37,383 CAN'T DERIVE A FORMULA FROM FIRST 5337 04:06:37,383 --> 04:06:41,754 PRINCIPLES FOR A TREE. HARD FOR 5338 04:06:41,754 --> 04:06:42,121 CHARLIE GROSS TO GET TRACTION. 5339 04:06:42,121 --> 04:06:45,591 WE CAN DESCRIBE COMPLEX OBJECTS 5340 04:06:45,591 --> 04:06:47,126 PREVIOUSLY INNEFABLE WITH THESE 5341 04:06:47,126 --> 04:06:47,459 DEEP NETWORKS. 5342 04:06:47,459 --> 04:06:48,661 THIS TRANSFORMED OUR 5343 04:06:48,661 --> 04:06:49,928 UNDERSTANDING OF MECHANISMS FOR 5344 04:06:49,928 --> 04:06:56,135 OBJECT RECOGNITION IN I.T. 5345 04:06:56,135 --> 04:06:56,468 CORTEX. 5346 04:06:56,468 --> 04:06:59,204 NEURONS REPRESENT FACIAL 5347 04:06:59,204 --> 04:07:00,039 IDENTITY USING SO-CALLED 5348 04:07:00,039 --> 04:07:01,740 SHAPE-APPEARANCE CODE THAT COMES 5349 04:07:01,740 --> 04:07:02,608 FROM CLASSIC COMPUTER VISION. 5350 04:07:02,608 --> 04:07:08,080 AND THIS IS A WAY OF 5351 04:07:08,080 --> 04:07:12,918 PARAMETERIZING FACES THAT SEPARATE 5352 04:07:12,918 --> 04:07:15,554 FROM TEXTURE. SINGLE FACE CELLS 5353 04:07:15,554 --> 04:07:18,957 PROJECTING TO SPECIFIC PREFERRED 5354 04:07:18,957 --> 04:07:20,392 AXIS, ENCODING A GENERATIVE 5355 04:07:20,392 --> 04:07:23,662 MODEL THAT LETS YOU RECONSTRUCT 5356 04:07:23,662 --> 04:07:25,364 FACES FROM NEURAL ACTIVITY. 5357 04:07:25,364 --> 04:07:26,665 A BEAUTIFUL INSIGHT THAT DEEP 5358 04:07:26,665 --> 04:07:28,634 NETWORKS GAVE US IS ALL I.T. 5359 04:07:28,634 --> 04:07:33,072 CORTEX WORKS IN THE SAME WAY. 5360 04:07:33,072 --> 04:07:40,379 CELLS IN IT CORTEX ARE USING AXIS 5361 04:07:40,379 --> 04:07:43,215 TO ENCODE, PASSING THROUGH A DEEP 5362 04:07:43,215 --> 04:07:43,882 NETWORK TO PRODUCE COMPRESSED 5363 04:07:43,882 --> 04:07:49,688 REPRESENTATION IN THE DEEP 5364 04:07:49,688 --> 04:07:50,556 LAYER, RECONSTRUCTING ARBITRARY 5365 04:07:50,556 --> 04:07:54,059 OBJECTS FROM NEURAL ACTIVITY. 5366 04:07:54,059 --> 04:07:55,794 FIRST TWO DIMENSIONS DERIVE FROM 5367 04:07:55,794 --> 04:07:58,197 DEEP NETWORKS SPAN A 2D SPACE. 5368 04:07:58,197 --> 04:08:00,199 AMAZINGLY WE FOUND PRECISE 5369 04:08:00,199 --> 04:08:01,834 CORRESPONDENCE BETWEEN LAYOUT OF 5370 04:08:01,834 --> 04:08:12,244 THE 2D OBJECT SPACE AND TOPO- 5371 04:08:18,550 --> 04:08:18,917 GRAPHIC LAYOUT. EACH QUADRANT OF 5372 04:08:18,917 --> 04:08:21,220 SPACE MAP TO OWN NETWORK OF I.T. 5373 04:08:21,220 --> 04:08:21,520 ORGANIZATION OF THE FACE PATCH SYSTEM GENERALIZES TO THE REST OF I.T. CORTEX. 5374 04:08:21,520 --> 04:08:24,389 AND WE INITIALLY MADE THIS 5375 04:08:24,389 --> 04:08:25,357 DISCOVER USING ALEXNET, DOESN'T 5376 04:08:25,357 --> 04:08:26,558 MATTER WHICH NETWORK YOU USE, 5377 04:08:26,558 --> 04:08:28,560 THEY ALL REPRESENT THE SAME 5378 04:08:28,560 --> 04:08:34,266 OBJECT SPACE, AND MULTIPLE OTHER 5379 04:08:34,266 --> 04:08:38,203 GROUPS MADE RELATED FINDINGS 5380 04:08:38,203 --> 04:08:39,838 ABOUT CONVERGENCE ACROSS 5381 04:08:39,838 --> 04:08:40,339 DIFFERENT NETWORKS. 5382 04:08:40,339 --> 04:08:41,373 AFTER I.T. CORTEX, VISUAL 5383 04:08:41,373 --> 04:08:44,009 INFORMATION ABOUT FACES GOES TO 5384 04:08:44,009 --> 04:08:46,178 MULTIPLE BRAIN REGIONS FOR PROCES- 5385 04:08:46,178 --> 04:08:47,412 SING HIGHER LEVEL SOCIAL STIMULI. 5386 04:08:47,412 --> 04:08:49,248 IT TURNS OUT OF ACTIVITY IN THESE 5387 04:08:49,248 --> 04:08:51,617 REGIONS CAN'T BE EXPLAINED BY 5388 04:08:51,617 --> 04:08:55,387 ALEXNET BUT INSTEAD CAN BE WELL 5389 04:08:55,387 --> 04:08:58,457 EXPLAINED BY GENERATIVE VIDEO 5390 04:08:58,457 --> 04:08:58,957 NETWORK. INDEED USING LATENCE 5391 04:08:58,957 --> 04:09:02,027 OF THE VIDEO NETWORK WE COULD 5392 04:09:02,027 --> 04:09:08,700 RECONSTRUCT SOCIAL STIMULI FROM 5393 04:09:08,700 --> 04:09:14,072 NEURAL ACTIVITY IN FRONTAL 5394 04:09:14,072 --> 04:09:17,242 CORTEX AND INSULA, THE LEVEL OF 5395 04:09:17,242 --> 04:09:18,777 DETAIL WAS ASTONISHING TO US AND 5396 04:09:18,777 --> 04:09:20,846 WE COULD NEVER HAVE DISCOVERED 5397 04:09:20,846 --> 04:09:25,651 THIS WITHOUT HELP OF DEEP 5398 04:09:25,651 --> 04:09:25,984 NETWORK. 5399 04:09:25,984 --> 04:09:27,486 WE'RE MAKING EXCITING PROCESS, 5400 04:09:27,486 --> 04:09:28,687 THE WHOLE FIELD OF NEUROAI 5401 04:09:28,687 --> 04:09:30,923 BUT I BELIEVE THIS IS THE 5402 04:09:30,923 --> 04:09:37,029 BEGINNING OF WHAT'S POSSIBLE. 5403 04:09:37,029 --> 04:09:40,232 CURRENT A.I. NETWORK IS BEHIND. 5404 04:09:40,232 --> 04:09:47,506 COMPARING START OF THE ART 5405 04:09:47,506 --> 04:09:49,241 MULTIMODAL LLMs TO HUMANS. 5406 04:09:49,241 --> 04:09:50,242 FOR EXAMPLE, ANSWERING A 5407 04:09:50,242 --> 04:09:51,543 MULTIPLE CHOICE QUESTION WHY DID 5408 04:09:51,543 --> 04:09:53,846 THE CHILD MOVE THE STEP TOOL 5409 04:09:53,846 --> 04:10:00,619 NEAR THE KITCHEN COUNTERTOP? 5410 04:10:00,619 --> 04:10:04,022 HUMANS ARE WAY BETTER. 5411 04:10:04,022 --> 04:10:05,324 85% COMPARED TO 37%. 5412 04:10:05,324 --> 04:10:08,293 AND TO TAKE EXAMPLE FROM MY OWN 5413 04:10:08,293 --> 04:10:12,331 WORK, MY PANDEMIC PROJECT, WAS TO 5414 04:10:12,331 --> 04:10:13,765 COMPUTATIONAL IMPLEMENT THEORY OF 5415 04:10:13,765 --> 04:10:15,601 VISUAL SEGMENTATION TRACKING 5416 04:10:15,601 --> 04:10:17,169 DERIVED FROM FIRST PRINCIPLES ABOUT 5417 04:10:17,169 --> 04:10:18,670 TOPOLOGY OF VISUAL SURFACES. 5418 04:10:18,670 --> 04:10:20,772 MY PROGRAM WAS PAINFULLY SLOW, 5419 04:10:20,772 --> 04:10:23,208 RUNNING A THOUSAND COPIES IN 5420 04:10:23,208 --> 04:10:25,811 PARALLEL, BUT WAS ABLE TO SHOW 5421 04:10:25,811 --> 04:10:27,479 IT COULD SUCCEED ON THIS 5422 04:10:27,479 --> 04:10:29,648 SYNTHETIC VIDEO ON THE LEFT, AS 5423 04:10:29,648 --> 04:10:30,649 PROOF OF CONCEPT. 5424 04:10:30,649 --> 04:10:35,220 WHEN I FED THIS EXACT SAME VIDEO 5425 04:10:35,220 --> 04:10:36,855 TO SAM, A DEEPNET FOR SEGMENTATION 5426 04:10:36,855 --> 04:10:39,191 TRACKING DEVELOPED BY META, IT 5427 04:10:39,191 --> 04:10:39,658 COMPLELTELY FAILED. 5428 04:10:39,658 --> 04:10:41,693 LIKELY BECAUSE THIS VIDEO WAS 5429 04:10:41,693 --> 04:10:43,695 OUTSIDE THE TRAINING DISTRIBUTION 5430 04:10:43,695 --> 04:10:46,531 SO WHATEVER HACKS THE NETWORK 5431 04:10:46,531 --> 04:10:47,399 WAS USING DIDN'T WORK. 5432 04:10:47,399 --> 04:10:48,800 IT'S NOT POPULAR TO SAY THIS IN 5433 04:10:48,800 --> 04:10:52,137 OUR AGE OF BIG DATA BUT I TRULY 5434 04:10:52,137 --> 04:10:54,873 BELIEVE WE WANT NETWORKS THAT 5435 04:10:54,873 --> 04:10:55,540 INCORPORATE PRINCIPLED 5436 04:10:55,540 --> 04:10:55,874 UNDERSTANDING. 5437 04:10:55,874 --> 04:10:58,277 I BELIEVE THIS IS THE ONLY PATH 5438 04:10:58,277 --> 04:10:59,378 TO REPLICATE NATURAL 5439 04:10:59,378 --> 04:11:01,346 INTELLIGENCE AND TO THEREBY 5440 04:11:01,346 --> 04:11:01,880 UNDERSTAND THE BRAIN. 5441 04:11:01,880 --> 04:11:04,049 WE HAVE TO REMEMBER THAT 5442 04:11:04,049 --> 04:11:05,918 NEUROSCIENCE AND A.I. HAVE 5443 04:11:05,918 --> 04:11:07,686 FUNDAMENTALLY DIFFERENT GOALS. 5444 04:11:07,686 --> 04:11:08,787 NEUROSCIENCE IS FOCUSED ON 5445 04:11:08,787 --> 04:11:09,855 UNDERSTANDING. 5446 04:11:09,855 --> 04:11:12,291 A.I. IS FOCUSED ON PERFORMANCE. 5447 04:11:12,291 --> 04:11:15,761 THE TITLE OF OUR SESSION HERE IS 5448 04:11:15,761 --> 04:11:21,033 EXPLORING THE STRUCTURE AND 5449 04:11:21,033 --> 04:11:21,800 STRUCTURAL CONVERGENCE CAPTURING 5450 04:11:21,800 --> 04:11:29,875 WHAT'S HAPPENING NOW, OFF THE 5451 04:11:29,875 --> 04:11:31,176 SHELF NETWORKS ARE COMPARED TO 5452 04:11:31,176 --> 04:11:33,278 THE BRAIN, WE MUST ACTIVELY 5453 04:11:33,278 --> 04:11:34,146 CREATE THIS CONVERGENCE. 5454 04:11:34,146 --> 04:11:37,316 I BELIEVE THE PATH FORWARD IS TO 5455 04:11:37,316 --> 04:11:39,184 INTENTIONALLY ALIGN INSIGHTS 5456 04:11:39,184 --> 04:11:41,153 FROM A.I., NEUROSCIENCE, AND 5457 04:11:41,153 --> 04:11:42,054 COGNITIVE SCIENCE. 5458 04:11:42,054 --> 04:11:44,990 SO AS A VISION SCIENTIST, SOME 5459 04:11:44,990 --> 04:11:46,091 IDEAS FROM COGNITIVE SCIENCE ARE 5460 04:11:46,091 --> 04:11:47,626 DEEP, NEED TO BE INCORPORATED 5461 04:11:47,626 --> 04:11:51,129 INTO A.I. VISION SYSTEMS, THE 5462 04:11:51,129 --> 04:11:52,030 CONCEPT OF SURFACES AS 5463 04:11:52,030 --> 04:11:55,634 FUNDAMENTAL UNITS OF VISUAL 5464 04:11:55,634 --> 04:11:57,402 REPRESENTATION, IDEA OF EXPLICIT 5465 04:11:57,402 --> 04:11:59,137 OBJECT FILES AS CRITICAL 5466 04:11:59,137 --> 04:12:03,842 STRUCTURE FOR ORGANIZING SENSORY 5467 04:12:03,842 --> 04:12:06,144 INFORMATION. INVARIANCE IS NOT 5468 04:12:06,144 --> 04:12:08,347 PERCEIVED BY POOLING BUT BY 5469 04:12:08,347 --> 04:12:09,314 COMPUTING OBJECT TRANSFORMATION. 5470 04:12:09,314 --> 04:12:11,750 THESE CONCEPTS ARE SUPPORTED BY 5471 04:12:11,750 --> 04:12:13,385 WEALTH OF BEHAVIORAL DATA, ALL 5472 04:12:13,385 --> 04:12:14,252 MISSING FROM CURRENT A.I. 5473 04:12:14,252 --> 04:12:15,921 AND THE FACT THEY ARE MISSING 5474 04:12:15,921 --> 04:12:26,498 HIGHLIGHTS A WEAKNESS OF CURRENT 5475 04:12:31,336 --> 04:12:31,603 A.I. 5476 04:12:31,603 --> 04:12:34,106 I THINK WE SHOULDN'T JUST WAIT 5477 04:12:34,106 --> 04:12:36,308 FOR CONCEPTS TO MAGICALLY APPEAR 5478 04:12:36,308 --> 04:12:40,345 IN THE DEEP NETWORKS ENGINEERED 5479 04:12:40,345 --> 04:12:42,581 IN SILICON VALLEY BUT SHOULD 5480 04:12:42,581 --> 04:12:45,083 INCORPORATE STRUCTORAL PRIORS INTO 5481 04:12:45,083 --> 04:12:45,350 OUR NEURAL NETWORKS. 5482 04:12:45,350 --> 04:12:49,554 IT'S TRICKY TO FIGURE OUT HOW TO 5483 04:12:49,554 --> 04:12:51,423 INCORPORATE THESE PRIORS BUT 5484 04:12:51,423 --> 04:12:54,926 NEUROSCIENCE COMES IN HERE. 5485 04:12:54,926 --> 04:12:57,996 INSIGHTS FROM NEUROSCIENCE GUIDES 5486 04:12:57,996 --> 04:13:00,298 US TO IMPLEMENT THESE PRIORS IN 5487 04:13:00,298 --> 04:13:00,732 THESE DISTRIBUTED SYSTEMS. 5488 04:13:00,732 --> 04:13:02,768 FOR EXAMPLE, WORK ON SURFACE 5489 04:13:02,768 --> 04:13:06,738 REPRESENTATION LED TO THE 5490 04:13:06,738 --> 04:13:09,574 DISCOVERY OF BORDER-OWNERSHIP 5491 04:13:09,574 --> 04:13:09,775 CELLS THAT CODE WHICH SIDE OF A 5492 04:13:09,775 --> 04:13:12,411 FIGURE OWNS A BORDER. WE COULD 5493 04:13:12,411 --> 04:13:13,412 INCORPORATE BORDER-OWNERSHIP 5494 04:13:13,412 --> 04:13:16,481 CELLS INTO A VISUAL SURFACE 5495 04:13:16,481 --> 04:13:17,315 REPRESENTATION NETWORK. 5496 04:13:17,315 --> 04:13:18,717 SO THERE'S A SAYING, RENDER TO 5497 04:13:18,717 --> 04:13:20,419 CAESAR THE THINGS THAT ARE 5498 04:13:20,419 --> 04:13:22,387 CAESAR'S, TO GOD THE THINGS THAT 5499 04:13:22,387 --> 04:13:23,588 ARE GOD'S. 5500 04:13:23,588 --> 04:13:25,991 AND HERE IS MY PARAPHRASE. 5501 04:13:25,991 --> 04:13:27,926 EVERYTHING RIGHT NOW IN A.I. IS 5502 04:13:27,926 --> 04:13:29,194 ABOUT DATA. 5503 04:13:29,194 --> 04:13:30,829 I THINK NEUROSCIENCE AND 5504 04:13:30,829 --> 04:13:32,297 COGNITIVE SCIENCE CAN HELP US 5505 04:13:32,297 --> 04:13:34,533 DISCOVER INNATE STRUCTURE OF THE 5506 04:13:34,533 --> 04:13:37,602 BRAIN AND RESTORE THE BALANCE. 5507 04:13:37,602 --> 04:13:40,772 AND I WANT TO NOTE THESE THREE 5508 04:13:40,772 --> 04:13:43,208 STRUCTURES IN PARTICULAR REALLY 5509 04:13:43,208 --> 04:13:45,610 BECOME CRITICAL FOR HANDLING 5510 04:13:45,610 --> 04:13:47,379 SENSORY INPUT IN DYNAMIC 3D 5511 04:13:47,379 --> 04:13:49,114 WORLD, LOTS OF OBJECT GENERATING 5512 04:13:49,114 --> 04:13:53,285 OCCLUSIONS AND PERSPECTIVE 5513 04:13:53,285 --> 04:13:56,254 CHANGES IN ACTIVE EXPLORATION. 5514 04:13:56,254 --> 04:13:58,090 THIS SCENARIO YOU NEED REAL 5515 04:13:58,090 --> 04:14:00,292 UNDERSTANDING OF 3D WORLD WHERE 5516 04:14:00,292 --> 04:14:02,494 CURRENT DEEP NETWORKS FAIL 5517 04:14:02,494 --> 04:14:04,663 MISERABLY AS WE ALREADY 5518 04:14:04,663 --> 04:14:04,996 DISCUSSED. 5519 04:14:04,996 --> 04:14:06,765 >> TWO-MINUTE WARNING. 5520 04:14:06,765 --> 04:14:08,934 >> I'M LUCKY TO TACKLE THIS 5521 04:14:08,934 --> 04:14:12,404 CHALLENGE OF REAL 3D WORLD 5522 04:14:12,404 --> 04:14:15,607 UNDERSTANDING AS PART OF THE 5523 04:14:15,607 --> 04:14:24,683 BRAIN INITIATIVE GRANT ON INTERNAL 5524 04:14:24,683 --> 04:14:26,118 MODELS LED BY ADRIENNE FAIRHALL. 5525 04:14:26,118 --> 04:14:29,387 THE MONKEYS LOVE PLAYING THIS 5526 04:14:29,387 --> 04:14:29,588 GAME. WITH NEUROPIXELS PROBES 5527 04:14:29,588 --> 04:14:30,989 WE ARE RECORDING ACROSS THE 5528 04:14:30,989 --> 04:14:34,926 MONKEY BRAIN TO UNDERSTAND HOW 5529 04:14:34,926 --> 04:14:38,196 A DYNAMIC SPATIAL ENVIRONMENT OF 5530 04:14:38,196 --> 04:14:40,165 OBJECTS IS REPRESENTED. 5531 04:14:40,165 --> 04:14:44,202 I HOPE THE DATA CAN CONTRIBUTE 5532 04:14:44,202 --> 04:14:45,770 TO FOUNDATION MODELS AND AID IN 5533 04:14:45,770 --> 04:14:48,507 THIS QUEST TO ALIGN 5534 04:14:48,507 --> 04:14:50,275 NEUROSCIENCE, COGNITIVE SCIENCE, 5535 04:14:50,275 --> 04:14:53,645 AND A.I. 5536 04:14:53,645 --> 04:14:55,714 WE NEED A NEW CROSS-DISCIPLINARY 5537 04:14:55,714 --> 04:14:58,683 TEAM GRANTS TO SUPPORT THIS 5538 04:14:58,683 --> 04:15:00,118 EFFORT, ALSO COMPUTING RESOURCES 5539 04:15:00,118 --> 04:15:04,923 SO TEAMS CAN IMPLEMENT IDEAS AT 5540 04:15:04,923 --> 04:15:07,259 SCALE TO COMPETE WITH MODELS IN 5541 04:15:07,259 --> 04:15:09,528 SILICON VALLEY. 5542 04:15:09,528 --> 04:15:13,231 THANK YOU. 5543 04:15:13,231 --> 04:15:23,742 >> I'D LIKE TO INTRODUCE THE 5544 04:15:26,811 --> 04:15:29,247 NEXT SPEAKER YIOTA POIRAZI FROM 5545 04:15:29,247 --> 04:15:30,015 IMBB-FORTH. 5546 04:15:30,015 --> 04:15:31,516 >> THANK YOU VERY MUCH FOR THIS 5547 04:15:31,516 --> 04:15:32,317 INVITATION. 5548 04:15:32,317 --> 04:15:35,253 GREAT TO BE HERE. 5549 04:15:35,253 --> 04:15:36,454 THANKS FOR THE WONDERFUL 5550 04:15:36,454 --> 04:15:39,524 INTRODUCTION, THANKS TO DORIS. 5551 04:15:39,524 --> 04:15:43,228 I'LL TALK ABOUT HOW TO INCORPORATE 5552 04:15:43,228 --> 04:15:44,996 NEW PRINCIPLES INTO NEUROAI. 5553 04:15:44,996 --> 04:15:48,400 I WILL TALK ABOUT DENDRITES, MY 5554 04:15:48,400 --> 04:15:50,802 FAVORITE SUBJECT. 5555 04:15:50,802 --> 04:15:52,437 WHY DENDRITES? FIRST, DENDRITIC 5556 04:15:52,437 --> 04:15:55,273 AND MORPHOLOGICAL COMPLEXITY 5557 04:15:55,273 --> 04:15:57,242 CORRELATES WITH OUR COGNITIVE 5558 04:15:57,242 --> 04:15:57,809 ABILITIES. 5559 04:15:57,809 --> 04:16:01,546 AS YOU CAN SEE IN THE SLIDE WHEN 5560 04:16:01,546 --> 04:16:03,215 WE'RE YOUNG CHILDREN, DENDRITIC 5561 04:16:03,215 --> 04:16:04,249 MORPHOLOGY IS SIMPLE, PEAKS 5562 04:16:04,249 --> 04:16:07,319 AROUND THE SAME TIME THE 5563 04:16:07,319 --> 04:16:09,421 COGNITIVE ABILITIES PEAK, STARTS 5564 04:16:09,421 --> 04:16:12,490 TO DECLINE WITH AGING AND 5565 04:16:12,490 --> 04:16:13,225 NEURODEGENERATION. 5566 04:16:13,225 --> 04:16:15,193 IT'S NOT ONLY MORPHOLOGY THAT 5567 04:16:15,193 --> 04:16:17,095 CORRELATES WITH COGNITIVE 5568 04:16:17,095 --> 04:16:17,362 ABILITIES. 5569 04:16:17,362 --> 04:16:18,830 THERE'S A NUMBER OF COMPUTATIONS 5570 04:16:18,830 --> 04:16:20,932 KNOWN TO TAKE PLACE WITHIN 5571 04:16:20,932 --> 04:16:23,201 DENDRITES THAT ARE IMPORTANT. 5572 04:16:23,201 --> 04:16:23,868 FOR EXAMPLE, PLASTICITY 5573 04:16:23,868 --> 04:16:26,104 CONSIDERED TO BE THE MAIN 5574 04:16:26,104 --> 04:16:27,405 SUBSTRATE OF LEARNING TAKES 5575 04:16:27,405 --> 04:16:30,909 PLACE AT SYNAPSES HOSTED WITHIN 5576 04:16:30,909 --> 04:16:32,978 DENDRITES, IT'S FACILITATED BY A 5577 04:16:32,978 --> 04:16:35,513 NUMBER OF NONLINEAR EVENTS, 5578 04:16:35,513 --> 04:16:37,249 DENDRITIC SPIKES THAT OPERATE AT 5579 04:16:37,249 --> 04:16:37,582 VERY DIFFERENT TIMESCALES. 5580 04:16:37,582 --> 04:16:40,218 WE AND OTHERS HAVE SHOWN THAT 5581 04:16:40,218 --> 04:16:41,987 THERE ARE STRUCTURAL PLASTICITY 5582 04:16:41,987 --> 04:16:43,822 LEARNING RULES THAT CAN HELP 5583 04:16:43,822 --> 04:16:45,590 NEURONS ORGANIZE THEIR INPUTS IN A 5584 04:16:45,590 --> 04:16:51,429 WAY THAT STRONGLY GENERATE SPIKES 5585 04:16:51,429 --> 04:16:53,398 GROUPING THEIR SYNAPTIC 5586 04:16:53,398 --> 04:16:54,733 INPUTS INTO NEARBY LOCATIONS IN 5587 04:16:54,733 --> 04:16:59,537 SPACE AND TIME. AND BY DRIVING 5588 04:16:59,537 --> 04:17:00,839 SPIKES UNDER SPECIFIC CONDITIONS 5589 04:17:00,839 --> 04:17:03,241 DENDRITES CAN ACT AS SMALL 5590 04:17:03,241 --> 04:17:06,878 COMPUTING UNITS HIGHLY 5591 04:17:06,878 --> 04:17:10,515 NONLINEAR AND PRODUCING LOCAL 5592 04:17:10,515 --> 04:17:11,716 SPIKES, SPATIAL ARRANGEMENT 5593 04:17:11,716 --> 04:17:13,818 BECOMES IMPORTANT READOUT IN 5594 04:17:13,818 --> 04:17:15,120 THESE NONLINEAR MODELS. 5595 04:17:15,120 --> 04:17:17,422 IF DENDRITES ACT AS SMALL 5596 04:17:17,422 --> 04:17:20,925 COMPUTING UNITS THEN NEURONS 5597 04:17:20,925 --> 04:17:23,795 ACT AS NEURAL NETWORKS, EVERY 5598 04:17:23,795 --> 04:17:26,298 DENDRITIC UNIT IS OFFERING ANOTHER 5599 04:17:26,298 --> 04:17:29,134 LEVEL OF NONLINEAR COMPUTATION. 5600 04:17:29,134 --> 04:17:32,237 THIS MEANS DENDRITES ALLOW 5601 04:17:32,237 --> 04:17:34,105 BIOLOGICAL BRAINS TO BECOME MUCH 5602 04:17:34,105 --> 04:17:34,973 MORE COMPUTATIONALLY POWERFUL 5603 04:17:34,973 --> 04:17:39,678 AND ALSO EFFICIENT AND LESS 5604 04:17:39,678 --> 04:17:41,346 ENERGY COSTLY BY PACKING 5605 04:17:41,346 --> 04:17:42,881 EVERYTHING INTO A SIMPLE NEURON. 5606 04:17:42,881 --> 04:17:44,416 AND THERE ARE MANY OTHER REASONS 5607 04:17:44,416 --> 04:17:46,818 WHY WE'RE INTERESTED IN DENDRITES. 5608 04:17:46,818 --> 04:17:47,085 THEY UNDERLIE A LOT OF COMPLEX 5609 04:17:47,085 --> 04:17:49,020 FUNCTIONS LIKE SENSORY PERCEPTION, 5610 04:17:49,020 --> 04:17:56,127 MEMORY, COINCIDENCE DETECTION, 5611 04:17:56,127 --> 04:17:57,796 TEMPORAL SEQUENCE, AND SO FORTH 5612 04:17:57,796 --> 04:17:59,397 WE LOVE DENDRITES BUT THIS IS 5613 04:17:59,397 --> 04:18:00,165 ABOUT BIOLOGICAL INTELLIGENCE. 5614 04:18:00,165 --> 04:18:03,468 WHAT IS THE ROLE OF DENDRITES IN 5615 04:18:03,468 --> 04:18:04,135 ARTIFICIAL INTELLIGENCE? 5616 04:18:04,135 --> 04:18:05,537 WELL, THEY HAVE SEEMED TO BE 5617 04:18:05,537 --> 04:18:06,504 QUITE PROMISING THERE AS WELL, 5618 04:18:06,504 --> 04:18:17,015 A NUMBER OF STUDIES ADOPT FOR A.I. 5619 04:18:17,849 --> 04:18:23,922 ABILITY TO SEGREGATE INFORMATION 5620 04:18:23,922 --> 04:18:25,457 INPUT INTO DIFFERENT DENDRITIC 5621 04:18:25,457 --> 04:18:26,624 COMPARTMENTS, ONLY INTEGRATING 5622 04:18:26,624 --> 04:18:29,594 TWO INFORMATION SOURCES UNDER 5623 04:18:29,594 --> 04:18:30,695 SPECIFIC CONDITIONS THROUGH THE 5624 04:18:30,695 --> 04:18:33,998 DENDRITIC SPIKES HAS BEEN USED 5625 04:18:33,998 --> 04:18:37,268 TO IMPLEMENT MULTIPLEXING OR 5626 04:18:37,268 --> 04:18:38,069 CREDIT ASSIGNMENT INTO SYNAPSES 5627 04:18:38,069 --> 04:18:39,003 RESPONSIBLE FOR LEARNING. 5628 04:18:39,003 --> 04:18:40,138 UNFORTUNATELY, I DON'T HAVE TIME 5629 04:18:40,138 --> 04:18:44,743 TO GO INTO THE FINDINGS BUT WE 5630 04:18:44,743 --> 04:18:46,044 HAVE RECENTLY WRITTEN A REVIEW YOU 5631 04:18:46,044 --> 04:18:49,013 CAN READ ABOUT IT IF YOU WANT. 5632 04:18:49,013 --> 04:18:52,417 IN ESSENCE, HAVING DENDRITES 5633 04:18:52,417 --> 04:18:55,587 MEANS THAT INDIVIDUAL NEURONS 5634 04:18:55,587 --> 04:18:58,790 ARE EMPOWERED WITH POWERFUL 5635 04:18:58,790 --> 04:18:59,057 MULTI-TOOL. 5636 04:18:59,057 --> 04:19:02,594 THEY CAN IMPLEMENT DIFFERENT 5637 04:19:02,594 --> 04:19:04,396 FUNCTIONS LIKE CONTEXT-GATING, 5638 04:19:04,396 --> 04:19:05,730 LOCAL LEARNING, THRESHOLDING, 5639 04:19:05,730 --> 04:19:10,168 MULTIPLEXING, AND SO ON, NOT 5640 04:19:10,168 --> 04:19:13,138 ALLOWED BY USING POINT NEURONS 5641 04:19:13,138 --> 04:19:14,105 WITHOUT DENDRITIC STRUCTURES 5642 04:19:14,105 --> 04:19:18,510 WHILE MAINTAINING A VERY LOW 5643 04:19:18,510 --> 04:19:20,812 ACTIVITY LEVEL, HIGH SPARSITY IN 5644 04:19:20,812 --> 04:19:23,014 THE BRAIN THUS MAINTAINING A 5645 04:19:23,014 --> 04:19:23,415 VERY LOWER POWER CONSUMPTION. 5646 04:19:23,415 --> 04:19:28,253 LET ME GIVE AN EXAMPLE OF HOW 5647 04:19:28,253 --> 04:19:33,725 INCORPORATING DENDRITES INTO 5648 04:19:33,725 --> 04:19:34,826 CLASSICAL ARTIFICIAL NEURAL 5649 04:19:34,826 --> 04:19:37,328 NETWORKS CAN IMPROVE EFFICIENCY. 5650 04:19:37,328 --> 04:19:39,731 TAKE A CLASSICAL FULLY CONNECTED 5651 04:19:39,731 --> 04:19:41,833 NEURAL NETWORK, INCORPORATE TWO 5652 04:19:41,833 --> 04:19:43,468 DENDRITIC PROPERTIES, ONE IS 5653 04:19:43,468 --> 04:19:44,536 STRUCTURE OF THE CONNECTIVITY, 5654 04:19:44,536 --> 04:19:46,438 SO SOME OF THE NODES OF THE 5655 04:19:46,438 --> 04:19:52,343 NETWORK ONLY TALK TO THEIR CELL 5656 04:19:52,343 --> 04:19:54,212 BODIES, SO IT'S SPARSELY CONNECTED. 5657 04:19:54,212 --> 04:19:56,214 LET DENDRITES SAMPLE INFORMATION 5658 04:19:56,214 --> 04:20:00,585 FROM THE INPUT IN A RESTRICTED 5659 04:20:00,585 --> 04:20:01,886 MANNER USING RECEPTIVE FIELDS. 5660 04:20:01,886 --> 04:20:04,522 THEN WE COMPARE HOW DO THESE 5661 04:20:04,522 --> 04:20:06,357 NETWORKS BEHAVE IF WE PUT THEM 5662 04:20:06,357 --> 04:20:07,892 NEXT TO A FULLY CONNECTED NEURAL 5663 04:20:07,892 --> 04:20:08,126 NETWORK. 5664 04:20:08,126 --> 04:20:11,062 WHAT YOU SEE HERE IS 5665 04:20:11,062 --> 04:20:12,764 PERFORMANCE, GRAY IS FULLY 5666 04:20:12,764 --> 04:20:13,731 CONNECTED, COLORED ARE THE 5667 04:20:13,731 --> 04:20:14,499 DENDRITIC ONES. 5668 04:20:14,499 --> 04:20:16,134 YOU NEED ABOUT A MILLION 5669 04:20:16,134 --> 04:20:17,769 PARAMETERS TO ACHIEVE THE BEST 5670 04:20:17,769 --> 04:20:22,040 PERFORMANCE OF A FULLY CONNECTED 5671 04:20:22,040 --> 04:20:23,641 NETWORK, WHEREAS YOU NEED 100 5672 04:20:23,641 --> 04:20:25,677 TIMES FEWER PARAMETERS TO 5673 04:20:25,677 --> 04:20:28,112 ACHIEVE THE SAME PERFORMANCE IF 5674 04:20:28,112 --> 04:20:29,180 YOU INCORPORATE DENDRITES INTO 5675 04:20:29,180 --> 04:20:29,848 THESE NETWORKS. 5676 04:20:29,848 --> 04:20:34,686 THIS IS ONE KEY EXAMPLE OF HOW 5677 04:20:34,686 --> 04:20:36,554 DENDRITIC COMPUTATION CAN ALLOW 5678 04:20:36,554 --> 04:20:38,323 PERFORMANCE IMPROVEMENT EVEN IN 5679 04:20:38,323 --> 04:20:40,091 CLASSICAL NEURAL NETWORKS THAT 5680 04:20:40,091 --> 04:20:41,960 RAN ON GPUs. NOT NEUROMORPHIC 5681 04:20:41,960 --> 04:20:43,895 HARDWARE OR MORE BIOLOGICAL 5682 04:20:43,895 --> 04:20:45,330 HARDWARE. THERE'S ALSO AN 5683 04:20:45,330 --> 04:20:47,398 IMPORTANT BENEFIT WITH RESPECT 5684 04:20:47,398 --> 04:20:49,734 TO OVERFITTING. CLASSICAL FULLY 5685 04:20:49,734 --> 04:20:51,836 CONNECTED NETWORKS TEND TO INCREASE 5686 04:20:51,836 --> 04:20:56,140 LOSS ONCE THEY GET LARGER, 5687 04:20:56,140 --> 04:20:57,509 DENDRITIC DO NOT SEEM TO SUFFER 5688 04:20:57,509 --> 04:20:59,310 SO MUCH. BY HAVING DENDRITES 5689 04:20:59,310 --> 04:21:01,279 INTO FULLY CONNECTED NEURAL 5690 04:21:01,279 --> 04:21:02,280 NETWORKS YOU CAN MAKE THEM MORE 5691 04:21:02,280 --> 04:21:11,256 EFFICIENT, MORE RESILIENT TO OVER- 5692 04:21:11,256 --> 04:21:12,590 FITTING AND EVEN MORE ACCURATE. 5693 04:21:12,590 --> 04:21:14,325 WE CAN STUDY HOW THESE NETWORKS 5694 04:21:14,325 --> 04:21:18,396 LEARN AND SEE THAT THEY USE A 5695 04:21:18,396 --> 04:21:19,464 VERY DIFFERENT LEARNING 5696 04:21:19,464 --> 04:21:20,565 STRATEGY. 5697 04:21:20,565 --> 04:21:21,966 GRAY IS FULLY CONNECTED. 5698 04:21:21,966 --> 04:21:24,302 WHAT NETWORKS TRY TO DO IS EACH 5699 04:21:24,302 --> 04:21:27,539 ONE OF THEIR NODES TRY TO LEARN 5700 04:21:27,539 --> 04:21:29,807 ONE CLASS, TO BE CLASS SPECIFIC. 5701 04:21:29,807 --> 04:21:34,345 DENDRITIC ONES, COLORED ONES, 5702 04:21:34,345 --> 04:21:35,980 REVERSE DISTRIBUTION, STRIVE TO 5703 04:21:35,980 --> 04:21:38,616 ENCODE INFORMATION ABOUT ALL 5704 04:21:38,616 --> 04:21:39,918 CLASSES. BY INCORPORATING THESE 5705 04:21:39,918 --> 04:21:41,019 PRINCIPLES IN THESE NETWORKS NOT 5706 04:21:41,019 --> 04:21:42,453 ONLY WE CAN MAKE THEM MORE 5707 04:21:42,453 --> 04:21:44,856 EFFICIENT BUT PUSH THEM IN A 5708 04:21:44,856 --> 04:21:45,857 COMPLETELY DIFFERENT LEARNING 5709 04:21:45,857 --> 04:21:47,358 REGIME THAT MAYBE TELLS US THAT 5710 04:21:47,358 --> 04:21:49,761 THE BRAIN IS USING A COMPLETELY 5711 04:21:49,761 --> 04:21:50,795 DIFFERENT LEARNING STRATEGY TO 5712 04:21:50,795 --> 04:21:51,996 SOLVE THE SAME PROBLEM. 5713 04:21:51,996 --> 04:21:57,702 AND WE CAN DO THIS ALSO IN 5714 04:21:57,702 --> 04:22:04,542 SPIKING NEURAL NETWORKS, ADDING 5715 04:22:04,542 --> 04:22:05,743 BIOLOGICAL RULES, SYNAPTIC 5716 04:22:05,743 --> 04:22:07,178 TURNOVER, AND LOOKING HOW 5717 04:22:07,178 --> 04:22:08,479 THESE PROPERTIES CONTRIBUTE TO 5718 04:22:08,479 --> 04:22:09,480 LEARNING IN THESE NETWORKS. 5719 04:22:09,480 --> 04:22:11,449 I DON'T HAVE TIME TO GO INTO 5720 04:22:11,449 --> 04:22:15,119 DETAILS BUT YOU CAN READ THE 5721 04:22:15,119 --> 04:22:16,754 PAPER. IF WE ADD SYNAPTIC 5722 04:22:16,754 --> 04:22:20,491 TURNOVER INTO SNN WE CAN GET 5723 04:22:20,491 --> 04:22:21,459 HIGHER PERFORMANCE, USING FEWER 5724 04:22:21,459 --> 04:22:22,560 EPOCHS RATHER THAN IF WE DON'T 5725 04:22:22,560 --> 04:22:27,932 HAVE IT THERE. 5726 04:22:27,932 --> 04:22:29,801 SO AGAIN ENSURING MORE ACCURATE 5727 04:22:29,801 --> 04:22:31,102 AND FASTER LEARNING HAVING 5728 04:22:31,102 --> 04:22:32,103 BIOLOGICAL PROPERTIES. 5729 04:22:32,103 --> 04:22:38,109 WE STILL DON'T UNDERSTAND EXACTLY 5730 04:22:38,109 --> 04:22:38,810 HOW DENDRITES CONTRIBUTE TO 5731 04:22:38,810 --> 04:22:40,311 CIRCUITS FOR LEARNING, THIS IS 5732 04:22:40,311 --> 04:22:43,281 REALLY IMPORTANT TO BUILD A.I. 5733 04:22:43,281 --> 04:22:45,149 SYSTEMS. TO DO THAT 5734 04:22:45,149 --> 04:22:45,450 WE NEED TO BRIDGE GAPS. 5735 04:22:45,450 --> 04:22:48,286 WE NEED TO HAVE MORE INFORMATION 5736 04:22:48,286 --> 04:22:49,921 ABOUT HOW DENDRITIC PROPERTIES 5737 04:22:49,921 --> 04:22:51,689 LOOK LIKE, AND UNFORTUNATELY 5738 04:22:51,689 --> 04:22:56,427 PEOPLE DON'T RECORD FROM 5739 04:22:56,427 --> 04:22:57,395 DENDRITES ANYMORE SO WE DON'T 5740 04:22:57,395 --> 04:22:59,597 KNOW FOR DIFFERENT CELL TYPES. 5741 04:22:59,597 --> 04:23:02,800 WE NEED TO DESCRIBE INPUT-OUTPUT 5742 04:23:02,800 --> 04:23:07,805 FUNCTION, IMPACT AT NEURONAL AND 5743 04:23:07,805 --> 04:23:08,706 BEHAVIORAL LEVELS. EXPERIMENTAL 5744 04:23:08,706 --> 04:23:10,541 TECHNIQUES TO HELP US WITH 5745 04:23:10,541 --> 04:23:13,911 DETAILED BIOPHYSICAL MODELS CAN 5746 04:23:13,911 --> 04:23:15,546 SHED NEW LIGHT. 5747 04:23:15,546 --> 04:23:17,849 AND WITH RESPECT TO A.I. FIELD 5748 04:23:17,849 --> 04:23:21,019 WE NEED TO IDENTIFY WHICH OF 5749 04:23:21,019 --> 04:23:22,553 THESE DENDRITIC PROPERTIES ARE 5750 04:23:22,553 --> 04:23:25,523 LIKELY TO BRING THE MOST IMPORTANT 5751 04:23:25,523 --> 04:23:27,058 BENEFITS, DESCRIBE THEM 5752 04:23:27,058 --> 04:23:28,593 MATHEMATICALLY, PUT THEM INTO 5753 04:23:28,593 --> 04:23:32,764 NEW A.I. SYSTEMS AND COME UP 5754 04:23:32,764 --> 04:23:35,366 WITH METRICS TO MEASURE ADVANTAGE 5755 04:23:35,366 --> 04:23:37,368 WE NEED COLLABORATION BETWEEN 5756 04:23:37,368 --> 04:23:38,970 DISCIPLINES TO ACHIEVE THIS GOAL 5757 04:23:38,970 --> 04:23:44,475 FROM . 5758 04:23:44,475 --> 04:23:46,878 >> TWO-MINUTE WARNING. 5759 04:23:46,878 --> 04:23:51,015 >> WE'VE BUILT TOOLS, FIRST 5760 04:23:51,015 --> 04:23:54,085 CALLED DENDROTWEAKS TO UPLOAD 5761 04:23:54,085 --> 04:23:54,919 EXISTING PREVIOUSLY DEVELOPED 5762 04:23:54,919 --> 04:23:57,088 BIOPHYSICAL MODEL AND PLAY WITH 5763 04:23:57,088 --> 04:23:59,390 ITS PARAMETERS IN A VISUALLY 5764 04:23:59,390 --> 04:24:00,892 EASY TO MANIPULATE SETTING. 5765 04:24:00,892 --> 04:24:03,428 YOU CAN CHANGE PARAMETERS AND 5766 04:24:03,428 --> 04:24:05,830 LOOK AT RESPONSES AT DIFFERENT 5767 04:24:05,830 --> 04:24:07,365 LEVELS OF THE MODEL, STANDARDIZE 5768 04:24:07,365 --> 04:24:09,801 ION CHANNELS THAT ARE THERE, YOU 5769 04:24:09,801 --> 04:24:12,870 CAN REDUCE MORPHOLOGY AT 5770 04:24:12,870 --> 04:24:13,938 DIFFERENT STAGES, KEEPING THE 5771 04:24:13,938 --> 04:24:15,707 RESPONSE SIMILAR TO ORIGINAL ONE 5772 04:24:15,707 --> 04:24:18,443 AND YOU CAN USE SCRIPTS TO 5773 04:24:18,443 --> 04:24:19,744 VALIDATE THE MODEL THAT YOU'RE 5774 04:24:19,744 --> 04:24:20,511 INTERESTED IN. 5775 04:24:20,511 --> 04:24:22,680 AND THIS IS REALLY IMPORTANT FOR 5776 04:24:22,680 --> 04:24:23,581 ADVANCING OUR UNDERSTANDING 5777 04:24:23,581 --> 04:24:25,783 BECAUSE YOU CAN CHANGE 5778 04:24:25,783 --> 04:24:27,185 PARAMETERS AND LOOK AT THE 5779 04:24:27,185 --> 04:24:28,619 IMPACT AT THE SINGLE NEURON 5780 04:24:28,619 --> 04:24:33,791 LEVEL AT LEAST. YOU CAN GO TO 5781 04:24:33,791 --> 04:24:36,928 USE DENDRIFY, THE OTHER TOOL TO 5782 04:24:36,928 --> 04:24:40,531 ADD ONLY THE MOST IMPORTANT 5783 04:24:40,531 --> 04:24:42,633 PROPERTIES OF DENDRITES TO POINT 5784 04:24:42,633 --> 04:24:44,869 NEURONS. LIF NEURONS ADD COMPARTMENT 5785 04:24:44,869 --> 04:24:47,438 OF DENDRITIC PROCESSES IN A 5786 04:24:47,438 --> 04:24:51,142 SIMPLE MANNER, REALLY FAST, 5787 04:24:51,142 --> 04:24:57,081 RUNNING ON BRIAN, CAN CAPTURE 5788 04:24:57,081 --> 04:24:58,082 KEY PROPERTIES, DENDRITIC NON- 5789 04:24:58,082 --> 04:24:58,683 LINEARITIES, BACKPROPAGATING 5790 04:24:58,683 --> 04:25:00,518 ACTION POTENTIAL, NONLINEAR 5791 04:25:00,518 --> 04:25:02,687 INTEGRATION EFFICIENTLY SO 5792 04:25:02,687 --> 04:25:05,423 YOU CAN RUN IT ON LARGE NETWORKS 5793 04:25:05,423 --> 04:25:05,690 AND ASSESS THE BENEFITS OF 5794 04:25:05,690 --> 04:25:06,724 DENDRITES. WITH THAT I'D LIKE 5795 04:25:06,724 --> 04:25:10,394 TO THANK YOU VERY MUCH FOR YOUR 5796 04:25:10,394 --> 04:25:10,661 ATTENTION. 5797 04:25:10,661 --> 04:25:20,104 [APPLAUSE] 5798 04:25:20,104 --> 04:25:29,180 >> I'D LIKE TO INTRODUCE CARINA 5799 04:25:29,180 --> 04:25:38,489 CURTO FROM BROWN UNIVERSITY. 5800 04:25:38,489 --> 04:25:39,690 >> IS THE SOUND OKAY? 5801 04:25:39,690 --> 04:25:45,596 NOW IT'S ON? 5802 04:25:45,596 --> 04:25:45,797 OKAY. 5803 04:25:45,797 --> 04:25:55,940 GREAT. 5804 04:26:00,144 --> 04:26:01,646 ALL RIGHT, GREAT. 5805 04:26:01,646 --> 04:26:05,249 LET'S GET STARTED. 5806 04:26:05,249 --> 04:26:08,019 SO, I'D LIKE TO FOCUS ON 5807 04:26:08,019 --> 04:26:11,856 SOMETHING SLIGHTLY DIFFERENT 5808 04:26:11,856 --> 04:26:14,158 THAN WHAT THE PREVIOUS TALKS 5809 04:26:14,158 --> 04:26:16,227 TALKED ABOUT, THE ROLE OF MATH 5810 04:26:16,227 --> 04:26:17,862 IN SMALL NETWORKS. 5811 04:26:17,862 --> 04:26:20,665 I'M A MATHEMATICIAN BY TRAINING, 5812 04:26:20,665 --> 04:26:21,432 TRAINED IN MATHEMATICAL PHYSICS. 5813 04:26:21,432 --> 04:26:24,068 I THROUGH MY Ph.D., SWITCHED 5814 04:26:24,068 --> 04:26:26,537 TO NEUROSCIENCE, A VERY EXCITING 5815 04:26:26,537 --> 04:26:28,506 FIELD, I DID NOT REGRET IT. 5816 04:26:28,506 --> 04:26:31,475 MORE EXCITING THAN EVER RIGHT 5817 04:26:31,475 --> 04:26:31,676 NOW. 5818 04:26:31,676 --> 04:26:34,645 AND I WANT TO THINK ABOUT THE 5819 04:26:34,645 --> 04:26:36,514 ROLE THAT DEVELOPING MORE 5820 04:26:36,514 --> 04:26:38,616 ADVANCED MATHEMATICS CAN PLAY IN 5821 04:26:38,616 --> 04:26:40,952 HELPING WITH NEUROAI. 5822 04:26:40,952 --> 04:26:42,453 SO, HERE I, YOU KNOW, ASKED 5823 04:26:42,453 --> 04:26:44,989 GOOGLE THE A.I. OVERVIEW, HOW 5824 04:26:44,989 --> 04:26:46,190 BIG THE NETWORKS ARE. 5825 04:26:46,190 --> 04:26:51,796 I GOT SOME NUMBERS. 5826 04:26:51,796 --> 04:26:52,864 NOTHING NEW. 5827 04:26:52,864 --> 04:26:54,232 TRILLIONS OF WEIGHTS HAVE TO BE 5828 04:26:54,232 --> 04:26:57,068 STORED FOR SOME OF THESE MODELS, 5829 04:26:57,068 --> 04:26:58,836 VERY LARGE DATASETS. 5830 04:26:58,836 --> 04:27:00,338 HIGH ENERGY USAGE. 5831 04:27:00,338 --> 04:27:01,672 WE'RE TALKING ABOUT NUCLEAR 5832 04:27:01,672 --> 04:27:04,308 REACTORS NOW FOR SOME STUFF. 5833 04:27:04,308 --> 04:27:06,377 AND THERE'S A HOPE 5834 04:27:06,377 --> 04:27:09,580 THAT WE CAN DO BETTER BY GETTING 5835 04:27:09,580 --> 04:27:10,581 INSPIRATION FROM NEUROSCIENCE 5836 04:27:10,581 --> 04:27:12,750 AND I TOTALLY AGREE WITH THAT. 5837 04:27:12,750 --> 04:27:14,518 SOME INGREDIENTS THAT A LOT OF 5838 04:27:14,518 --> 04:27:16,821 CURRENT A.I. SYSTEMS SEEM TO 5839 04:27:16,821 --> 04:27:20,124 IGNORE WE JUST HEARD A LOVELY 5840 04:27:20,124 --> 04:27:23,294 TALK ABOUT DENDRITES AND OTHER 5841 04:27:23,294 --> 04:27:24,495 ASPECTS, NEUROMODULATION IS 5842 04:27:24,495 --> 04:27:25,696 REALLY IMPORTANT BECAUSE IT 5843 04:27:25,696 --> 04:27:26,797 ALLOWS YOU TO BASICALLY ENCODE 5844 04:27:26,797 --> 04:27:28,866 MULTIPLE NETWORKS 5845 04:27:28,866 --> 04:27:30,935 IN ONE. 5846 04:27:30,935 --> 04:27:34,138 BECAUSE VIA NEUROMODULATION THE 5847 04:27:34,138 --> 04:27:36,974 SAME NETWORK CAN HAVE MULTIPLE 5848 04:27:36,974 --> 04:27:37,375 DYNAMIC REGIMES. 5849 04:27:37,375 --> 04:27:38,609 RHYTHMS AND OSCILLATIONS ARE 5850 04:27:38,609 --> 04:27:40,378 ALSO A REALLY IMPORTANT ASPECT 5851 04:27:40,378 --> 04:27:42,980 OF WHAT NETWORKS IN THE BRAIN 5852 04:27:42,980 --> 04:27:43,748 ACTUALLY DO. 5853 04:27:43,748 --> 04:27:45,983 WHEN YOU THINK ABOUT NEURAL 5854 04:27:45,983 --> 04:27:50,521 CIRCUITS THAT MAINTAIN CENTRAL 5855 04:27:50,521 --> 04:27:51,355 PATTERN GENERATORS, THESE CIRCUITS, 5856 04:27:51,355 --> 04:27:52,590 MAINTAIN DIFFERENT PATTERNS OF 5857 04:27:52,590 --> 04:27:53,858 ACTIVITY FOR LOCOMOTION AND SO 5858 04:27:53,858 --> 04:27:55,927 ON WHEN TALKING ABOUT EMBODIED 5859 04:27:55,927 --> 04:28:00,865 A.I., THESE ARE THINGS WE NEED 5860 04:28:00,865 --> 04:28:02,199 TO THINK ABOUT. 5861 04:28:02,199 --> 04:28:05,269 IT IS VERY INSPIRING TO THINK 5862 04:28:05,269 --> 04:28:06,370 ABOUT THE CONNECTIVITY 5863 04:28:06,370 --> 04:28:09,540 STRUCTURES WE'VE ALREADY WORKED 5864 04:28:09,540 --> 04:28:11,742 OUT, THE FLY COMPASS SYSTEM A 5865 04:28:11,742 --> 04:28:13,811 LOVELY PART OF THE FLY BRAIN, 5866 04:28:13,811 --> 04:28:16,013 WHICH HAS BEEN MENTIONED ALREADY 5867 04:28:16,013 --> 04:28:16,881 SEVERAL TIMES TODAY. 5868 04:28:16,881 --> 04:28:18,749 I THINK OTHER SMALL CREATURES, 5869 04:28:18,749 --> 04:28:23,988 ZEBRAFISH IS ONE I WORK ON WITH 5870 04:28:23,988 --> 04:28:25,089 THE CRCNS GRANT THROUGH THE BRAIN 5871 04:28:25,089 --> 04:28:26,824 INITIATIVE, THAT'S BEEN GREAT. 5872 04:28:26,824 --> 04:28:27,925 THERE ARE LOTS OF DIFFERENT 5873 04:28:27,925 --> 04:28:30,328 CIRCUITS THAT I THINK ARE VERY 5874 04:28:30,328 --> 04:28:32,830 INTERESTING TO STUDY, HOW THESE 5875 04:28:32,830 --> 04:28:34,932 SMALL CREATURES DO INCREDIBLY 5876 04:28:34,932 --> 04:28:35,566 COMPLEX BEHAVIORS, MAKE 5877 04:28:35,566 --> 04:28:39,303 DECISIONS AND SO ON AND DON'T 5878 04:28:39,303 --> 04:28:43,674 NEED TRILLIONS OF PARAMETERS TO 5879 04:28:43,674 --> 04:28:44,508 DO IT. 5880 04:28:44,508 --> 04:28:45,309 MATHEMATICAL ISSUES, IT'S WHERE 5881 04:28:45,309 --> 04:28:51,115 I WANT TO GET TO IN THIS TALK. 5882 04:28:51,115 --> 04:28:53,351 I THINK A LOT OF THE POTENTIAL 5883 04:28:53,351 --> 04:28:56,153 THAT WE'VE SEEN IN THE PAST AS 5884 04:28:56,153 --> 04:28:57,221 WELL OF GETTING NEUROSCIENCE 5885 04:28:57,221 --> 04:29:00,057 IDEAS INTO THE A.I. SPACE GOES 5886 04:29:00,057 --> 04:29:02,893 THROUGH KIND OF THE STEP OF 5887 04:29:02,893 --> 04:29:03,928 MATHEMATICAL ABSTRACTION. IF YOU 5888 04:29:03,928 --> 04:29:06,330 THINK ABOUT MCCULLOCH-PITTS AND 5889 04:29:06,330 --> 04:29:10,267 THESE EXAMPLES WE'VE TALKED ABOUT 5890 04:29:10,267 --> 04:29:12,570 THE HOPFIELD MODEL, BOLTZMANN 5891 04:29:12,570 --> 04:29:16,507 MACHINES, THERE ARE INSPIRATIONS 5892 04:29:16,507 --> 04:29:17,041 FROM NEUROSCIENCE, THERE'S THIS 5893 04:29:17,041 --> 04:29:17,708 MATHEMATICAL ABSTRACTION, THEN THE 5894 04:29:17,708 --> 04:29:19,610 MOST USEFUL MODELS TEND TO BE THE 5895 04:29:19,610 --> 04:29:21,245 ONES THAT WE CAN UNDERSTAND. 5896 04:29:21,245 --> 04:29:24,648 I WANT TO TAKE ONE MOMENT TO 5897 04:29:24,648 --> 04:29:26,517 IMAGINE THE FOLLOWING SCENARIO. 5898 04:29:26,517 --> 04:29:28,919 IMAGINE A WORLD WHERE WE DIDN'T 5899 04:29:28,919 --> 04:29:30,454 HAVE LINEAR ALGEBRA. 5900 04:29:30,454 --> 04:29:30,788 OKAY? 5901 04:29:30,788 --> 04:29:32,857 SO IMAGINE YOU HAVE A LINEAR 5902 04:29:32,857 --> 04:29:34,058 SYSTEM OF DIFFERENTIAL EQUATIONS 5903 04:29:34,058 --> 04:29:37,028 CAN WRITE IT DOWN, WE HAVE 5904 04:29:37,028 --> 04:29:38,863 CALCULUS, WE KNOW ABOUT RATES OF 5905 04:29:38,863 --> 04:29:42,199 CHANGE, AND WE HAVE SOME MATRIX 5906 04:29:42,199 --> 04:29:47,004 A OF INTERACTION, AND LOTS AND 5907 04:29:47,004 --> 04:29:51,375 LOTS OF -- LOTS AN LOTS OF 5908 04:29:51,375 --> 04:29:55,780 VARIABLES INTERACTING IN COMPLEX 5909 04:29:55,780 --> 04:29:57,848 WAYS, ARBITRARILY COMPLEX AND 5910 04:29:57,848 --> 04:29:59,050 DIFFICULT TO UNDERSTAND. 5911 04:29:59,050 --> 04:30:01,485 YET IF THE SYSTEM IS LINEAR, WE 5912 04:30:01,485 --> 04:30:03,888 KNOW WHAT TO DO, RIGHT? 5913 04:30:03,888 --> 04:30:07,158 EVEN OUR SECOND YEAR 5914 04:30:07,158 --> 04:30:09,994 UNDERGRADS CAN STUDY THAT 5915 04:30:09,994 --> 04:30:18,436 MATRIX, WE KNOW HOW TO COMPUTE 5916 04:30:18,436 --> 04:30:19,003 VALUES, EIGENVECTORS, THIS 5917 04:30:19,003 --> 04:30:23,307 MATHEMATICS HAS BEEN USED TO 5918 04:30:23,307 --> 04:30:24,175 UNDERSTAND MARKOV CHAINS, 5919 04:30:24,175 --> 04:30:26,043 ENGINEERING APPLICATIONS, SO ON. 5920 04:30:26,043 --> 04:30:29,413 WE UNDERSTAND LINEAR SYSTEMS. 5921 04:30:29,413 --> 04:30:32,483 VERY, VERY WELL. 5922 04:30:32,483 --> 04:30:35,019 THAT HAS LED TO REMARKABLE 5923 04:30:35,019 --> 04:30:37,988 PROGRESS IN ENGINEERING, 5924 04:30:37,988 --> 04:30:39,824 PHYSICS, EVEN IN BIOLOGY. 5925 04:30:39,824 --> 04:30:43,928 NOW THINKING ABOUT 5926 04:30:43,928 --> 04:30:44,562 NEUROSCIENCE, FUNDAMENTALLY THE 5927 04:30:44,562 --> 04:30:47,765 NEURONS TELL US WE NEED TO LOOK 5928 04:30:47,765 --> 04:30:48,632 AT NONLINEARITIES. 5929 04:30:48,632 --> 04:30:52,803 WE HEARD ABOUT IT IN THE 5930 04:30:52,803 --> 04:30:55,873 DENDRITE TALK AS WELL. 5931 04:30:55,873 --> 04:30:57,074 WHEN YOU ADD NONLINEARITY EVEN 5932 04:30:57,074 --> 04:30:58,175 IN THE MOST SIMPLIFIED VERSION 5933 04:30:58,175 --> 04:31:00,244 OF THE PROBLEM IT CHANGES 5934 04:31:00,244 --> 04:31:01,445 EVERYTHING. 5935 04:31:01,445 --> 04:31:01,679 OKAY? 5936 04:31:01,679 --> 04:31:03,647 NOW SUDDENLY WE DO NOT HAVE THE 5937 04:31:03,647 --> 04:31:04,849 HANDLE THAT WE HAD ON THE 5938 04:31:04,849 --> 04:31:07,017 LEFT-HAND SIDE IN THE LINEAR 5939 04:31:07,017 --> 04:31:07,218 CASE. 5940 04:31:07,218 --> 04:31:14,024 NOW SUDDENLY WE DON'T KNOW WHAT 5941 04:31:14,024 --> 04:31:19,497 THE ASYMPTOTIC DYNAMICS WILL DO. 5942 04:31:19,497 --> 04:31:22,032 WE DON'T KNOW HOW MANY EQUILIBRIUM 5943 04:31:22,032 --> 04:31:23,234 POINTS. IN NONLINEAR DYNAMICS, 5944 04:31:23,234 --> 04:31:24,335 THERE CAN BE MULTIPLE, LIMIT CYCLES 5945 04:31:24,335 --> 04:31:26,103 PERIODIC PATTERNS OF ACTIVITY, 5946 04:31:26,103 --> 04:31:28,305 LOTS OF STUFF CAN GO ON, AND 5947 04:31:28,305 --> 04:31:30,141 WITHOUT A BETTER THEORY IT'S VERY 5948 04:31:30,141 --> 04:31:31,675 DIFFICULT TO UNDERSTAND WHAT'S 5949 04:31:31,675 --> 04:31:32,910 POSSIBLE IN THIS PARAMETER SPACE 5950 04:31:32,910 --> 04:31:36,046 EVEN WITH THE SAME NUMBER OF 5951 04:31:36,046 --> 04:31:36,347 PARAMETERS. 5952 04:31:36,347 --> 04:31:38,349 THAT W MATRIX AND B VECTOR CAN 5953 04:31:38,349 --> 04:31:40,084 HAVE THE SAME SIZE AS THE A AND 5954 04:31:40,084 --> 04:31:44,488 B ON THE LEFT BUT BECAUSE OF THE 5955 04:31:44,488 --> 04:31:46,223 NONLINEARITY SUDDENLY IT'S 5956 04:31:46,223 --> 04:31:46,857 INTRACTABLE. 5957 04:31:46,857 --> 04:31:47,224 OKAY. 5958 04:31:47,224 --> 04:31:50,494 SO, ONE OF THE THINGS THAT I 5959 04:31:50,494 --> 04:31:54,098 GUESS I WANT TO ADVOCATE IS THAT 5960 04:31:54,098 --> 04:31:55,032 NEUROSCIENCE AND NEURAL CIRCUITS 5961 04:31:55,032 --> 04:31:58,102 THAT WE SEE IN THE BEHAVIOR AND 5962 04:31:58,102 --> 04:32:00,704 INTERACTIONS WE SEE CAN LEAD US 5963 04:32:00,704 --> 04:32:03,340 TO VERSIONS OF THIS WHERE EVEN 5964 04:32:03,340 --> 04:32:04,542 THOUGH WE HAVE NONLINEARITIES 5965 04:32:04,542 --> 04:32:06,944 THEY ARE SURPRISINGLY 5966 04:32:06,944 --> 04:32:09,246 TRACTABLE. 5967 04:32:09,246 --> 04:32:11,215 ONE EXAMPLE FROM MY WORK, ONE 5968 04:32:11,215 --> 04:32:15,252 VERSION OF THE FIRING RATE MODEL 5969 04:32:15,252 --> 04:32:16,353 WITH A SIMPLE RELU NONLINEARITY, 5970 04:32:16,353 --> 04:32:16,987 MACHINE LEARNING PEOPLE LIKE 5971 04:32:16,987 --> 04:32:19,156 THAT ONE TOO. 5972 04:32:19,156 --> 04:32:20,758 IN THIS SITUATION THERE ARE -- 5973 04:32:20,758 --> 04:32:23,360 THERE'S A FULL RICHNESS OF 5974 04:32:23,360 --> 04:32:24,228 NONLINEAR DYNAMICS, NETWORKS 5975 04:32:24,228 --> 04:32:26,197 USED SUCCESSFULLY FOR LOTS OF 5976 04:32:26,197 --> 04:32:27,097 MODELING APPLICATIONS IN 5977 04:32:27,097 --> 04:32:27,765 NEUROSCIENCE. 5978 04:32:27,765 --> 04:32:28,599 WE DIDN'T INVENT THEM. 5979 04:32:28,599 --> 04:32:30,267 THEY HAVE BEEN AROUND AT LEAST 5980 04:32:30,267 --> 04:32:34,138 40 YEARS. FIRST MATHEMATICAL 5981 04:32:34,138 --> 04:32:34,838 RESULTS IN EARNEST MORE THAN 5982 04:32:34,838 --> 04:32:41,679 20 YEARS FROM SEUNG, HAHNLOSER, 5983 04:32:41,679 --> 04:32:47,284 SLOTINE. YOU CAN THINK OF THEM 5984 04:32:47,284 --> 04:32:49,119 AS SORT OF A 21ST CENTURY HOPFIELD 5985 04:32:49,119 --> 04:32:52,623 MODEL THAT HAS ALL OF THE SORT OF 5986 04:32:52,623 --> 04:32:53,490 STATIC, MULTI-STABILITY AND 5987 04:32:53,490 --> 04:32:57,194 MULTIPLE-STATIC ATTRACTORS LIKE 5988 04:32:57,194 --> 04:32:59,863 HOPFIELD MODEL DOES, CAN ENCODE 5989 04:32:59,863 --> 04:33:02,666 DYNAMIC ATTRACTORS, RICH 5990 04:33:02,666 --> 04:33:04,201 PATTERNS OF ACTIVITY AND DO LOTS 5991 04:33:04,201 --> 04:33:05,035 OF THINGS. 5992 04:33:05,035 --> 04:33:07,271 SO THE PROBLEM OF COURSE IS 5993 04:33:07,271 --> 04:33:08,606 BECAUSE OF THE NONLINEARITY 5994 04:33:08,606 --> 04:33:10,207 IT'S HARD TO UNDERSTAND. 5995 04:33:10,207 --> 04:33:13,077 SO IN MY GROUP WE HAVE BEEN 5996 04:33:13,077 --> 04:33:16,180 WORKING ON THESE AS LIKE KIND OF 5997 04:33:16,180 --> 04:33:18,282 THE BABY STEPS TOWARDS 5998 04:33:18,282 --> 04:33:22,886 UNDERSTANDING HIGH DIMENSIONAL 5999 04:33:22,886 --> 04:33:24,622 NONLINEAR DYNAMICAL SYSTEMS, 6000 04:33:24,622 --> 04:33:25,222 FOR NEUROSCIENCE INSPIRED 6001 04:33:25,222 --> 04:33:26,824 NETWORKS AND WE'VE BEEN ABLE TO 6002 04:33:26,824 --> 04:33:31,629 PROVE A BUNCH OF THEOREMS TO 6003 04:33:31,629 --> 04:33:32,963 CONNECT GRAPH STRUCTURE AND 6004 04:33:32,963 --> 04:33:34,598 ARCHITECTURE TO THE DYNAMICS. 6005 04:33:34,598 --> 04:33:36,033 THIS GIVES AN EXAMPLE OF WHERE A 6006 04:33:36,033 --> 04:33:37,434 LITTLE BIT OF MATHEMATICS CAN GO 6007 04:33:37,434 --> 04:33:39,637 A LONG WAY. FOR EXAMPLE, MY FORMER 6008 04:33:39,637 --> 04:33:43,607 STUDENT, JULIANA LONDONO ALVAREZ 6009 04:33:43,607 --> 04:33:46,010 WHO'S HERE GIVING A POSTER 6010 04:33:46,010 --> 04:33:48,612 TOMORROW MORNING I BELIEVE -- 6011 04:33:48,612 --> 04:33:50,614 >> TWO-MINUTE WARNING. 6012 04:33:50,614 --> 04:33:52,016 >> OKAY. 6013 04:33:52,016 --> 04:33:53,517 SHE HAD THE VERY INTERESTING 6014 04:33:53,517 --> 04:33:55,519 PROJECT AS A GRADUATE STUDENT OF 6015 04:33:55,519 --> 04:34:02,426 TRYING TO DESIGN A NETWORK THAT 6016 04:34:02,426 --> 04:34:03,560 ENCODED MULTIPLE QUADRUPED GAITS. 6017 04:34:03,560 --> 04:34:09,867 SO LIMIT CYCLES, DYNAMIC ATTRACTORS 6018 04:34:09,867 --> 04:34:11,835 INVOLVED THE SAME CONTROL 6019 04:34:11,835 --> 04:34:14,705 NEURONS FOR THE LIMB AND COULD 6020 04:34:14,705 --> 04:34:17,207 ENCODE MULTIPLE GAITS AT ONCE, A 6021 04:34:17,207 --> 04:34:22,046 PROBLEM MANY GROUPS TACKLED, AND 6022 04:34:22,046 --> 04:34:23,013 TYPICALLY THE SOLUTIONS INVOLVE 6023 04:34:23,013 --> 04:34:24,648 A FAMILY OF MODELS WHERE YOU 6024 04:34:24,648 --> 04:34:27,418 HAVE TO CHANGE SOME PARAMETER IN 6025 04:34:27,418 --> 04:34:35,392 ORDER TO GO FROM ONE PATTERN TO 6026 04:34:35,392 --> 04:34:36,293 ANOTHER, SYNAPTIC PLASTICITY OR 6027 04:34:36,293 --> 04:34:38,829 EXTERNAL DRIVE, SOMETHING HAS TO 6028 04:34:38,829 --> 04:34:39,196 CHANGE. 6029 04:34:39,196 --> 04:34:40,798 AND BECAUSE OF THE 6030 04:34:40,798 --> 04:34:42,333 SMALL AMOUNT OF MATHEMATICAL 6031 04:34:42,333 --> 04:34:44,735 UNDERSTANDING THAT WE HAVE AND 6032 04:34:44,735 --> 04:34:46,236 SPECIAL FAMILY OF THRESHOLD 6033 04:34:46,236 --> 04:34:48,038 LINEAR NETWORKS SHE IS ABLE TO 6034 04:34:48,038 --> 04:34:50,007 QUICKLY DESIGN A NETWORK 6035 04:34:50,007 --> 04:34:52,843 ENCODING FIVE DIFFERENT KINDS OF 6036 04:34:52,843 --> 04:34:54,611 GAITS, ALL SIMULTANEOUSLY, NO 6037 04:34:54,611 --> 04:34:56,480 CHANGE OF PARAMETERS, REALLY IS 6038 04:34:56,480 --> 04:34:58,649 FIVE ATTRACTORS IN THE NETWORK 6039 04:34:58,649 --> 04:34:59,750 AND DIFFERENT INITIAL CONDITIONS 6040 04:34:59,750 --> 04:35:02,152 WILL DROP YOU IN EACH ONE. 6041 04:35:02,152 --> 04:35:03,921 THIS IS JUST AN ADVERTISEMENT 6042 04:35:03,921 --> 04:35:13,530 FOR HER WORK, BUT POINT IS 6043 04:35:13,530 --> 04:35:15,165 SOMETHING COMPLICATED, WE DID NOT NEED A NETWORK OF 20,000 TO 100,000 NEURONS WITH LOTS OF TRAINING. 6044 04:35:15,165 --> 04:35:16,800 WE NEEDED A LITTLE BIT OF MATHEMATICAL 6045 04:35:16,800 --> 04:35:17,868 INSIGHT, AND I THINK THAT, YOU 6046 04:35:17,868 --> 04:35:22,439 KNOW, IN THE LONG RUN THERE WILL 6047 04:35:22,439 --> 04:35:23,640 BE MANY OPPORTUNITIES TO REALLY 6048 04:35:23,640 --> 04:35:26,143 TAKE ADVANTAGE OF A LITTLE BIT 6049 04:35:26,143 --> 04:35:29,680 OF EXTRA MATHEMATICAL INSIGHT IN 6050 04:35:29,680 --> 04:35:31,181 ORDER TO SEARCH BETTER THE 6051 04:35:31,181 --> 04:35:32,416 PARAMETER SPACE. 6052 04:35:32,416 --> 04:35:33,617 SAME MODELS, RIGHT? 6053 04:35:33,617 --> 04:35:36,453 SAME MODELS THAT WE'RE USING 6054 04:35:36,453 --> 04:35:37,554 THAT FIND BETTER 6055 04:35:37,554 --> 04:35:39,056 PLACES IN THE PARAMETER SPACE 6056 04:35:39,056 --> 04:35:42,559 WHERE WE CAN START BETTER 6057 04:35:42,559 --> 04:35:43,794 ARCHITECTURES THAT LEAD US TO 6058 04:35:43,794 --> 04:35:45,429 THE KIND OF DYNAMICS WE WANT. 6059 04:35:45,429 --> 04:35:48,065 I THINK WITH THAT I WILL END. 6060 04:35:48,065 --> 04:35:58,442 THANK YOU VERY MUCH. 6061 04:36:01,845 --> 04:36:04,581 >> NEXT SPEAKER IS EV 6062 04:36:04,581 --> 04:36:08,986 FEDORENKO FROM M.I.T., AND USE 6063 04:36:08,986 --> 04:36:17,394 PODIUM MICS, BETTER SOUND. 6064 04:36:17,394 --> 04:36:20,464 >> THANKS FOR HAVING ME HERE. 6065 04:36:20,464 --> 04:36:22,433 ANIMAL MODELS HAVE BEEN 6066 04:36:22,433 --> 04:36:24,401 INVALUABLE IN BOTH BASIC 6067 04:36:24,401 --> 04:36:25,068 NEUROSCIENCE AND MEDICAL 6068 04:36:25,068 --> 04:36:25,436 RESEARCH. 6069 04:36:25,436 --> 04:36:28,005 OF COURSE IN MANY CASES THE HOPE 6070 04:36:28,005 --> 04:36:29,773 IS TO UNDERSTAND SOMETHING ABOUT 6071 04:36:29,773 --> 04:36:32,509 HOW THE HUMAN BRAIN WORKS. 6072 04:36:32,509 --> 04:36:32,709 THE REASON WE USE ANIMAL MODELS 6073 04:36:32,709 --> 04:36:34,812 IS BECAUSE THE 6074 04:36:34,812 --> 04:36:36,447 HARDWARE IS SIMILAR ENOUGH, WE 6075 04:36:36,447 --> 04:36:38,982 CAN USE APPROACHES NOT AVAILABLE 6076 04:36:38,982 --> 04:36:40,717 IN HUMANS. WE RECORD FROM 6077 04:36:40,717 --> 04:36:44,555 HUNDREDS OF CELLS AT ONCE, USE 6078 04:36:44,555 --> 04:36:45,189 OPTOGENETICS TO CONTROL CELLS, 6079 04:36:45,189 --> 04:36:46,089 SO ON. 6080 04:36:46,089 --> 04:36:51,995 WHAT IF THE BEHAVIOR OF INTEREST 6081 04:36:51,995 --> 04:36:54,531 IS LANGUAGE, SOCIAL REASONING, 6082 04:36:54,531 --> 04:36:54,798 OR MATHEMATICAL COGNITION? 6083 04:36:54,798 --> 04:36:57,267 WE DON'T DO WELL WITH A MOUSE OR 6084 04:36:57,267 --> 04:36:59,169 MACAQUE AND WE'RE STUCK WITH 6085 04:36:59,169 --> 04:36:59,503 HUMANS. 6086 04:36:59,503 --> 04:37:01,872 THE REASON IT'S SO CRITICAL TO 6087 04:37:01,872 --> 04:37:03,207 UNDERSTAND HOW LANGUAGE WORKS IS 6088 04:37:03,207 --> 04:37:04,508 THAT IT'S SUCH A FUNDAMENTAL 6089 04:37:04,508 --> 04:37:08,011 PART OF THE HUMAN EXPERIENCE. 6090 04:37:08,011 --> 04:37:12,349 USING LANGUAGE WE BUILD DEEP 6091 04:37:12,349 --> 04:37:14,117 INTERPERSONAL RELATIONSHIPS AND 6092 04:37:14,117 --> 04:37:15,085 ESTABLISH INTERNATIONAL 6093 04:37:15,085 --> 04:37:16,854 POLITICAL ORDER, SEND ROCKETS TO 6094 04:37:16,854 --> 04:37:18,922 SPACE, SUBMARINES TO THE OCEAN, 6095 04:37:18,922 --> 04:37:23,260 CURE DISEASES, ENACT SOCIAL AND 6096 04:37:23,260 --> 04:37:24,094 POLITICAL CHANGE, AND MAKE A 6097 04:37:24,094 --> 04:37:26,830 RECORD OF THE HISTORY OF OUR 6098 04:37:26,830 --> 04:37:28,465 PLANET AND OUR SPECIES. 6099 04:37:28,465 --> 04:37:30,868 LOSING LANGUAGE EARLY ON IN LIFE 6100 04:37:30,868 --> 04:37:32,302 DUE TO DEVELOPMENTAL 6101 04:37:32,302 --> 04:37:33,370 COMMUNICATIVE DISORDERS IS 6102 04:37:33,370 --> 04:37:34,671 DEVASTATING FOR OUR ABILITIES TO 6103 04:37:34,671 --> 04:37:36,440 LEARN ABOUT THE WORLD, AND 6104 04:37:36,440 --> 04:37:40,577 LOSING LANGUAGE DUE TO STROKE OR 6105 04:37:40,577 --> 04:37:41,678 DEGENERATION IS INCREDIBLY 6106 04:37:41,678 --> 04:37:46,517 PAINFUL FOR BOTH PATIENTS AND 6107 04:37:46,517 --> 04:37:47,484 THEIR LOVED ONES. 6108 04:37:47,484 --> 04:37:50,320 SO OF COURSE IF WE'RE STUCK WITH 6109 04:37:50,320 --> 04:37:51,889 HUMANS WE'RE LIMITED TO THE 6110 04:37:51,889 --> 04:37:53,624 KINDS OF APPROACHES THAT WE CAN 6111 04:37:53,624 --> 04:37:53,891 USE. 6112 04:37:53,891 --> 04:37:56,360 BUT STILL WE'VE BEEN ABLE TO 6113 04:37:56,360 --> 04:37:57,561 MAKE SOME PROGRESS, I'LL TELL 6114 04:37:57,561 --> 04:37:59,096 YOU IN ONE MINUTE, SO WE'VE 6115 04:37:59,096 --> 04:38:01,698 LEARNED THAT THERE'S A NETWORK 6116 04:38:01,698 --> 04:38:06,303 IN THE LEFT HEMISPHERE THAT 6117 04:38:06,303 --> 04:38:07,404 SUPPORTS LANGUAGE PROCESSING. 6118 04:38:07,404 --> 04:38:09,273 THESE AREAS ARE ENGAGED WHEN WE 6119 04:38:09,273 --> 04:38:10,274 UNDERSTAND LANGUAGE, REGARDLESS 6120 04:38:10,274 --> 04:38:12,009 OF WHETHER IT'S SPOKEN, WRITTEN, 6121 04:38:12,009 --> 04:38:12,342 SIGNED. 6122 04:38:12,342 --> 04:38:14,745 AND WHEN WE PRODUCE LANGUAGE 6123 04:38:14,745 --> 04:38:16,947 ALSO ACROSS MODALITIES, WHICH 6124 04:38:16,947 --> 04:38:18,382 SUGGESTS THAT THESE AREAS LIKELY 6125 04:38:18,382 --> 04:38:19,950 STORE OUR KNOWLEDGE OF LANGUAGE, 6126 04:38:19,950 --> 04:38:21,852 WHAT WORDS MEAN, HOW TO PUT THEM 6127 04:38:21,852 --> 04:38:22,853 TOGETHER TO MAKE COMPLEX 6128 04:38:22,853 --> 04:38:25,155 MEANINGS, AND OF COURSE WE NEED 6129 04:38:25,155 --> 04:38:27,658 THIS KNOWLEDGE TO CONVERT OUR 6130 04:38:27,658 --> 04:38:30,827 THOUGHTS INTO SEQUENCES OF WORDS 6131 04:38:30,827 --> 04:38:31,595 AND EXTRACT MEANING. 6132 04:38:31,595 --> 04:38:33,797 WHAT I TOLD YOU SHOULD MAKE IT 6133 04:38:33,797 --> 04:38:34,865 CLEAR THAT THESE AREAS ARE 6134 04:38:34,865 --> 04:38:36,533 DISTINCT FROM THE AREAS THAT 6135 04:38:36,533 --> 04:38:43,006 HAVE BEEN HISTORICALLY DISCUSSED 6136 04:38:43,006 --> 04:38:47,578 BROCA'S AREA IN RED AND WERNICKE'S 6137 04:38:47,578 --> 04:38:50,414 AREA IN BLUE AND ARE INSENSITIVE 6138 04:38:50,414 --> 04:39:00,857 TO THE MEANING OF LANGUAGE. 6139 04:39:08,599 --> 04:39:10,500 WORK IN MY GROUP ESTABLISHED 6140 04:39:10,500 --> 04:39:12,369 LANGUAGE SYSTEM IS DISTINCT FROM 6141 04:39:12,369 --> 04:39:13,804 SYSTEMS OF KNOWLEDGE AND 6142 04:39:13,804 --> 04:39:15,105 REASONING, I'M SHOWING THREE 6143 04:39:15,105 --> 04:39:18,442 SUCH SYSTEMS BUT THERE ARE MORE. 6144 04:39:18,442 --> 04:39:19,810 OF COURSE LANGUAGE HAS TO WORK 6145 04:39:19,810 --> 04:39:20,811 CLOSELY WITH THESE SYSTEMS BUT 6146 04:39:20,811 --> 04:39:22,546 IT'S NOT THE CASE THAT YOUR 6147 04:39:22,546 --> 04:39:23,981 LANGUAGE SYSTEM IS DOING THE 6148 04:39:23,981 --> 04:39:25,949 THINKING. YOU CAN STILL THINK IF 6149 04:39:25,949 --> 04:39:33,090 YOU LOSE YOUR LANGUAGE SYSTEM. 6150 04:39:33,090 --> 04:39:35,459 IT'S THERE BY AGE 3 BUT WE CAN'T 6151 04:39:35,459 --> 04:39:37,561 MEASURE IN 2-YEAR-OLDS, MUCH 6152 04:39:37,561 --> 04:39:37,894 HARDER. IT'S SIMILAR ACROSS 6153 04:39:37,894 --> 04:39:39,463 TYPOLOGICALLY DIVERSE 6154 04:39:39,463 --> 04:39:41,264 LANGUAGES, MANY OF US IN THE 6155 04:39:41,264 --> 04:39:42,699 AUDIENCE HAVE TWO LANGUAGES 6156 04:39:42,699 --> 04:39:44,134 SYSTEMS, TWO OR MORE LIVING IN 6157 04:39:44,134 --> 04:39:46,203 THEIR BRAIN, THEY ALL LIVE IN 6158 04:39:46,203 --> 04:39:48,839 THE SAME PLACE. 6159 04:39:48,839 --> 04:39:50,641 DAMAGE TO THESE REGIONS IN 6160 04:39:50,641 --> 04:39:51,475 ADULTHOOD LEADS TO LOSS OF 6161 04:39:51,475 --> 04:39:53,010 FUNCTION SO THEY ARE CRITICALLY 6162 04:39:53,010 --> 04:39:54,745 IMPORTANT FOR LANGUAGE. 6163 04:39:54,745 --> 04:39:56,046 FINALLY BASED ON DOZENS OF 6164 04:39:56,046 --> 04:39:57,481 STUDIES WE HAVE A BASIC 6165 04:39:57,481 --> 04:39:58,548 UNDERSTANDING OF WHAT THE 6166 04:39:58,548 --> 04:40:00,217 REGIONS DO WHICH IS RELATED TO 6167 04:40:00,217 --> 04:40:01,852 RETRIEVING WORDS FROM MEMORY AND 6168 04:40:01,852 --> 04:40:05,489 PUTTING THEM TOGETHER TO MAKE 6169 04:40:05,489 --> 04:40:07,357 NEW COMPLEX MEANINGS. BUT WHAT 6170 04:40:07,357 --> 04:40:11,395 DATA STRUCTURES SUPPORT THESE 6171 04:40:11,395 --> 04:40:12,596 COMPUTATIONS IN COMPREHENSION AND 6172 04:40:12,596 --> 04:40:16,633 PRODUCTION? AND HOW ARE THEY 6173 04:40:16,633 --> 04:40:18,001 INSTANTIATED IN NEURONS? 6174 04:40:18,001 --> 04:40:19,603 WELL, UNTIL RECENTLY IT WAS NOT 6175 04:40:19,603 --> 04:40:23,507 CLEAR HOW TO MAKE PROGRESS BUT 6176 04:40:23,507 --> 04:40:25,242 THEN EVERYTHING 6177 04:40:25,242 --> 04:40:25,942 CHANGED. 6178 04:40:25,942 --> 04:40:27,878 NEURAL NETWORK LANGUAGE MODELS 6179 04:40:27,878 --> 04:40:31,682 EMERGED PRODUCING BEHAVIOR THAT 6180 04:40:31,682 --> 04:40:32,516 WAS HUMAN-LIKE, LINGUISTIC 6181 04:40:32,516 --> 04:40:33,784 BEHAVIOR THAT WAS HUMAN-LIKE. 6182 04:40:33,784 --> 04:40:37,054 JUST LIKE THAT WE HAD A MODEL 6183 04:40:37,054 --> 04:40:37,721 ORGANISM ADMITTEDLY 6184 04:40:37,721 --> 04:40:39,056 IN SILICO, NOT BIOLOGICAL, BUT 6185 04:40:39,056 --> 04:40:40,791 FOR THE FIRST TIME IN THE 6186 04:40:40,791 --> 04:40:42,526 HISTORY OF HUMANKIND WE HAD A 6187 04:40:42,526 --> 04:40:44,428 SYSTEM OTHER THAN THE HUMAN 6188 04:40:44,428 --> 04:40:45,595 BRAIN THAT WOULD PRODUCE 6189 04:40:45,595 --> 04:40:47,931 BEHAVIOR OF INTEREST, MEANING 6190 04:40:47,931 --> 04:40:51,068 OUR LANGUAGE SCIENTIST BEHAVIOR 6191 04:40:51,068 --> 04:40:54,538 OF INTEREST. INSPIRED BY THESE 6192 04:40:54,538 --> 04:40:56,740 MODELS, MARTIN SCHRIMPF WHO 6193 04:40:56,740 --> 04:40:59,176 WILL WHO TALK AFTER ME WANTED TO 6194 04:40:59,176 --> 04:41:02,212 SEE IF SIMILARITY IN BEHAVIORS 6195 04:41:02,212 --> 04:41:03,847 EXPANDS TO REPRESENTATIONS. 6196 04:41:03,847 --> 04:41:05,849 HE PRESENTED SENTENCES TO BOTH 6197 04:41:05,849 --> 04:41:07,684 MODELS IN HUMANS FROM THE BRAIN, 6198 04:41:07,684 --> 04:41:08,485 EXTRACTED RESPONSES FROM 6199 04:41:08,485 --> 04:41:11,955 LANGUAGE AREAS, FROM THE MODELS 6200 04:41:11,955 --> 04:41:14,257 INTERNAL REPRESENTATIONS AS THEY 6201 04:41:14,257 --> 04:41:15,092 PROCESSED SIMILARLY, TESTING 6202 04:41:15,092 --> 04:41:16,026 SIMILARITY BETWEEN THEM. 6203 04:41:16,026 --> 04:41:18,195 EVALUATED A LARGE NUMBER OF 6204 04:41:18,195 --> 04:41:19,730 LANGUAGE MODELS, FOUND QUITE A 6205 04:41:19,730 --> 04:41:22,265 FEW OF THEM DO PRETTY WELL IN 6206 04:41:22,265 --> 04:41:26,636 ALIGNING WITH HUMAN NEURAL 6207 04:41:26,636 --> 04:41:28,939 REPRESENTATION, EACH BAR IS A 6208 04:41:28,939 --> 04:41:31,241 MODEL, THE GRAY LINE IS THE CEILING BASED ON THE INTERNAL RELIABILITY OF THE DATASET. 6209 04:41:31,241 --> 04:41:34,211 NOW WE AS LANGUAGE SCIENTISTS 6210 04:41:34,211 --> 04:41:37,147 HAVE A MODEL SYSTEM, IN SILICO 6211 04:41:37,147 --> 04:41:39,750 SPECIES, THAT PRODUCES THE SAME 6212 04:41:39,750 --> 04:41:40,851 BEHAVIORS AS HUMANS, 6213 04:41:40,851 --> 04:41:42,519 QUOTE/UNQUOTE BRAIN CAN BE 6214 04:41:42,519 --> 04:41:44,921 ANALYZED AT THE CELLULAR AND 6215 04:41:44,921 --> 04:41:46,590 CIRCUIT LEVELS USING THE 6216 04:41:46,590 --> 04:41:54,498 EVER-GROWING SUITE OF MODEL 6217 04:41:54,498 --> 04:41:55,232 INTERPRETABILITY TOOLS WHICH 6218 04:41:55,232 --> 04:41:57,134 CAN BE TAKEN APART LIKE WE CAN'T 6219 04:41:57,134 --> 04:41:58,235 WITH BIOLOGICAL BRAINS. 6220 04:41:58,235 --> 04:41:59,870 WE CAN NOW UNDERSTAND HOW 6221 04:41:59,870 --> 04:42:01,338 LANGUAGE WORKS WITH THE 6222 04:42:01,338 --> 04:42:02,606 PRECISION AND SCALE COULD NOT 6223 04:42:02,606 --> 04:42:09,146 IMAGINE POSSIBLE IN MY 6224 04:42:09,146 --> 04:42:09,546 LIFETIME. 6225 04:42:09,546 --> 04:42:11,882 WE CAN UNDERSTAND HOW IT WORKS. 6226 04:42:11,882 --> 04:42:13,650 WE RECENTLY REVIEWED EFFORTS IN 6227 04:42:13,650 --> 04:42:20,190 THE SPACE IN THE RECENT PAPER IN 6228 04:42:20,190 --> 04:42:27,898 HER WORK GRETA TUCKUTE SHOWED YOU 6229 04:42:27,898 --> 04:42:30,367 CAN MODULATE RESPONSE IN CLOSED 6230 04:42:30,367 --> 04:42:33,270 LOOP MANNER. EGHBAL HOSSEINI IS 6231 04:42:33,270 --> 04:42:36,640 DISCOVERING THE CORE CODING ACCESS 6232 04:42:36,640 --> 04:42:40,911 USING CONTROVERSIAL STIMULI. 6233 04:42:40,911 --> 04:42:41,978 WE USE INTRACRANIAL RECORDINGS TO 6234 04:42:41,978 --> 04:42:43,547 EVALUATE ALGORITHMS FOR HOW 6235 04:42:43,547 --> 04:42:48,985 REPRESENTATIONS EVOLVE AS NEW WORDS 6236 04:42:48,985 --> 04:42:51,188 COME IN. WE ALSO ASK QUESTIONS ABOUT 6237 04:42:51,188 --> 04:42:52,389 HOW LANGUAGE SYSTEM DEVELOPS AND 6238 04:42:52,389 --> 04:42:54,357 CAN BREAK. ON DEVELOPMENT, WE USE 6239 04:42:54,357 --> 04:42:55,559 CONTROLLED REARING APPROACH TO 6240 04:42:55,559 --> 04:42:57,594 ISOLATE DIFFERENT VARIABLES THAT 6241 04:42:57,594 --> 04:42:59,396 AFFECT TRAJECTORY AND ENDPOINT 6242 04:42:59,396 --> 04:43:01,832 OF LANGUAGE LEARNING, INCLUDING 6243 04:43:01,832 --> 04:43:05,569 USING MODELS THAT LEARN IN MORE 6244 04:43:05,569 --> 04:43:08,839 HUMAN-LIKE WAYS. 6245 04:43:08,839 --> 04:43:11,875 WITH RESPECT TO BEHAVIOR, 6246 04:43:11,875 --> 04:43:13,210 WE CAN RELAY REPRESENTATIONS 6247 04:43:13,210 --> 04:43:15,946 FROM BABY LANGUAGE MODELS TO 6248 04:43:15,946 --> 04:43:17,147 NEURAL DATA ACROSS DEVELOPMENTAL 6249 04:43:17,147 --> 04:43:17,781 TRAJECTORY. 6250 04:43:17,781 --> 04:43:19,416 WITH RESPECT TO LINGUISTIC 6251 04:43:19,416 --> 04:43:22,385 DISORDERS, WE CAN BREAK MODELS IN 6252 04:43:22,385 --> 04:43:24,788 VARIOUS WAYS AND TEST THEIR 6253 04:43:24,788 --> 04:43:29,226 BEHAVIOR TO SEE IF WE MAY 6254 04:43:29,226 --> 04:43:30,060 SIMULATE PATTERNS OF IMPAIRMENT 6255 04:43:30,060 --> 04:43:31,328 AND TRY TO UNDERSTAND WAYS THEY 6256 04:43:31,328 --> 04:43:33,697 BREAK SUCH THAT THEY LEAD TO 6257 04:43:33,697 --> 04:43:35,699 TO PARTICULAR KINDS OF 6258 04:43:35,699 --> 04:43:35,999 LINGUISTIC DEFICITS. 6259 04:43:35,999 --> 04:43:37,968 FINALLY WE CAN START ASKING HOW 6260 04:43:37,968 --> 04:43:39,703 DIFFERENT PIECES OF LANGUAGE 6261 04:43:39,703 --> 04:43:40,704 PROCESSING PIPELINE WORK 6262 04:43:40,704 --> 04:43:40,971 TOGETHER. 6263 04:43:40,971 --> 04:43:42,639 FOR EXAMPLE HOW DOES LANGUAGE 6264 04:43:42,639 --> 04:43:47,777 SYSTEM WORK WITH REASONING 6265 04:43:47,777 --> 04:43:47,978 SYSTEMS. 6266 04:43:47,978 --> 04:43:50,247 HOW DOES INFORMATION GO FROM THE 6267 04:43:50,247 --> 04:43:51,081 LANGUAGE SYSTEM TO DOWNSTREAM 6268 04:43:51,081 --> 04:43:52,515 SYSTEM WHEN WE HAVE TO REASON ON 6269 04:43:52,515 --> 04:43:54,351 THE INFORMATION THAT WE GET 6270 04:43:54,351 --> 04:43:59,422 THROUGH LANGUAGE FOR EXAMPLE TO 6271 04:43:59,422 --> 04:44:02,092 SOLVE A VERBAL MATH PROBLEM OR 6272 04:44:02,092 --> 04:44:02,626 UPDATE OUR KNOWLEDGE OF SOME 6273 04:44:02,626 --> 04:44:03,927 SOCIAL AGENT? ALMOST DONE. 6274 04:44:03,927 --> 04:44:06,096 SO WHAT DO WE NEED TO MAKE ALL 6275 04:44:06,096 --> 04:44:07,364 THESE COOL THINGS? 6276 04:44:07,364 --> 04:44:09,165 THERE'S FOUR PIECES HERE. 6277 04:44:09,165 --> 04:44:10,367 I THINK WE NEED MUCH BETTER DATA 6278 04:44:10,367 --> 04:44:11,268 THAN WHAT WE HAVE. 6279 04:44:11,268 --> 04:44:13,069 I WOULD ARGUE THE DATA THAT WE 6280 04:44:13,069 --> 04:44:14,304 NEED TO UNDERSTAND HUMAN 6281 04:44:14,304 --> 04:44:16,306 LANGUAGE AND OTHER HIGH LEVEL 6282 04:44:16,306 --> 04:44:18,375 COGNITION DOESN'T EXIST HARDLY 6283 04:44:18,375 --> 04:44:20,744 AT ALL YET. 6284 04:44:20,744 --> 04:44:25,749 WE NEED MORE BIOLOGICALLY AND 6285 04:44:25,749 --> 04:44:26,483 COGNITIVELY PLAUSIBLE MODELS, 6286 04:44:26,483 --> 04:44:28,451 THERE'S QUITE A FEW MORE THAN 6287 04:44:28,451 --> 04:44:30,654 WHAT I LISTED. 6288 04:44:30,654 --> 04:44:33,290 WE NEED SOPHISTICATED MODELS TO 6289 04:44:33,290 --> 04:44:35,125 METRICS, CURRENT METRICS ATTEMPT 6290 04:44:35,125 --> 04:44:37,327 TO MAKE CONNECTIONS WHICH LINK 6291 04:44:37,327 --> 04:44:40,030 BETTER TO IMPLEMENTATION LEVEL 6292 04:44:40,030 --> 04:44:41,498 COMPARED TO REPRESENTATIONAL LEVEL. 6293 04:44:41,498 --> 04:44:43,333 AND THESE MODELS ARE POWERFUL, 6294 04:44:43,333 --> 04:44:44,868 AND THAT'S FINE BUT THAT MEANS 6295 04:44:44,868 --> 04:44:49,406 THE BURDEN IS ON US TO BE GOOD 6296 04:44:49,406 --> 04:44:51,474 EXPERIMENTALISTS AND DRAW THE 6297 04:44:51,474 --> 04:44:52,576 RIGHT CONCLUSIONS AND NOT 6298 04:44:52,576 --> 04:44:55,312 OVERINTERPRET WHAT WE CAN FIND 6299 04:44:55,312 --> 04:44:57,681 FROM MODELS WITH A LOT OF 6300 04:44:57,681 --> 04:44:59,416 EXCITING DISCOVERIES THAT LAY 6301 04:44:59,416 --> 04:45:08,325 AHEAD. 6302 04:45:08,325 --> 04:45:09,893 THANK YOU. 6303 04:45:09,893 --> 04:45:14,064 NEXT SPEAKER IS MARTIN SC 6304 04:45:14,064 --> 04:45:15,131 HRIMPF FROM EPFL. 6305 04:45:15,131 --> 04:45:17,867 >> THANKS TO THE ORGANIZERS, 6306 04:45:17,867 --> 04:45:20,537 USUALLY, TO GIVE BACKGROUND WHY 6307 04:45:20,537 --> 04:45:25,809 WE WANT MODELS, I THOUGHT WE 6308 04:45:25,809 --> 04:45:27,077 WOULD ALL AGREE MODELS ARE 6309 04:45:27,077 --> 04:45:27,544 GREAT. 6310 04:45:27,544 --> 04:45:31,247 I DON'T THINK THE GOAL IS TO BE 6311 04:45:31,247 --> 04:45:33,350 BUILD AGI, JUST BUILDING A MODEL 6312 04:45:33,350 --> 04:45:35,986 OF THE BRAIN IS USEFUL. 6313 04:45:35,986 --> 04:45:38,355 IT'S USEFUL FROM ONE 6314 04:45:38,355 --> 04:45:40,090 PERSPECTIVE, BUT ALSO FROM THE 6315 04:45:40,090 --> 04:45:43,727 PERSPECTIVE THIS COULD LEAD TO 6316 04:45:43,727 --> 04:45:44,361 FUTURE CLINICAL APPLICATIONS, I 6317 04:45:44,361 --> 04:45:45,996 WANT TO TALK ABOUT WHAT WE CAN 6318 04:45:45,996 --> 04:45:47,831 DO ON THE MORE BRAIN SIDE OF 6319 04:45:47,831 --> 04:45:50,533 THINGS IF WE HAD A REALLY GOOD 6320 04:45:50,533 --> 04:45:52,002 MODEL OF THE BRAIN. 6321 04:45:52,002 --> 04:45:53,837 FIRST BACKGROUND, ONE OF THE 6322 04:45:53,837 --> 04:46:03,313 CORE FINDINGS THAT HAS COME OUT, 6323 04:46:03,313 --> 04:46:04,748 WHEN WE SHOW THE SAME STIMULI TO MODELS THAT WERE SHOWN TO HUMANS OR ANIMALS AND MEASURE 6324 04:46:04,748 --> 04:46:07,117 INTERNAL ACTIVITY, THAT JUST TENDS TO ALIGN, TRUE 6325 04:46:07,117 --> 04:46:11,121 EVEN THOUGH A.I. MODELS WERE NOT 6326 04:46:11,121 --> 04:46:13,490 BUILT FOR NEUROSCIENCE THE LAST 6327 04:46:13,490 --> 04:46:14,758 10-20 YEARS OR SO. A.I. DOES NOT 6328 04:46:14,758 --> 04:46:16,026 BUILD THINGS IN A BRAIN-LIKE 6329 04:46:16,026 --> 04:46:21,231 WAY BUT THE WAY THEY OPTIMIZE 6330 04:46:21,231 --> 04:46:22,499 MODELS LEADS TO REPRESENTATIONS THAT ARE SURPRISINGLY BRAIN LIKE. 6331 04:46:22,499 --> 04:46:24,300 THIS IS AN EXAMPLE FROM VISION. 6332 04:46:24,300 --> 04:46:27,670 Y AXIS IS ALIGNMENT TO A SET OF 6333 04:46:27,670 --> 04:46:31,141 DIFFERENT NEURAL RECORDINGS FROM 6334 04:46:31,141 --> 04:46:33,510 PRIMATES AS WELL AS BEHAVIOR. 6335 04:46:33,510 --> 04:46:38,515 AND ONE DOT IS ONE MODEL. 6336 04:46:38,515 --> 04:46:41,951 TURNS OUT IF YOU MEASURE THEIR 6337 04:46:41,951 --> 04:46:45,321 ALIGNMENT, THAT'S PREDICTED BY 6338 04:46:45,321 --> 04:46:46,222 THEIR TASK PERFORMANCE. 6339 04:46:46,222 --> 04:46:49,626 THAT IS WHAT MACHINE LEARNING 6340 04:46:49,626 --> 04:46:52,695 CARES ABOUT, AND THIS ALIGNS 6341 04:46:52,695 --> 04:46:54,431 WITH THE BRAIN LIKENESS OF THE 6342 04:46:54,431 --> 04:46:54,631 MODEL. THIS IS NOT JUST IN 6343 04:46:54,631 --> 04:47:01,805 VISION. RECENTLY THERE HAS BEEN 6344 04:47:01,805 --> 04:47:03,440 A PAPER FOR MOTOR. 6345 04:47:03,440 --> 04:47:06,076 THIS SEEMS TO BE A THEME WHERE 6346 04:47:06,076 --> 04:47:07,911 MODELS TEND TO BE GOOD AT 6347 04:47:07,911 --> 04:47:11,648 PERFORMING A TASK ARE THE ONES 6348 04:47:11,648 --> 04:47:13,283 THAT ARE BRAIN-LIKE IN THEIR 6349 04:47:13,283 --> 04:47:13,917 INTERNAL REPRESENTATIONS. 6350 04:47:13,917 --> 04:47:15,819 THIS IS A MORE RECENT FINDING. 6351 04:47:15,819 --> 04:47:20,190 IF YOU LOOK AT THEM, MOST RECENT 6352 04:47:20,190 --> 04:47:21,724 MODELS, ONES GETTING BETTER AT 6353 04:47:21,724 --> 04:47:23,093 MACHINE LEARNING THE LAST COUPLE 6354 04:47:23,093 --> 04:47:24,360 YEARS, THEY ARE NOT TURNING OUT 6355 04:47:24,360 --> 04:47:26,930 TO BE MORE BRAIN-LIKE. 6356 04:47:26,930 --> 04:47:34,471 IF ANYTHING WE'RE SEEING 6357 04:47:34,471 --> 04:47:37,474 INVERSE, THE MODELS STARTING TO 6358 04:47:37,474 --> 04:47:41,010 APPROACH 100% ARE LESS AND LESS 6359 04:47:41,010 --> 04:47:41,311 BRAIN-LIKE. 6360 04:47:41,311 --> 04:47:42,312 RECENTLY WE FIND THAT IF YOU 6361 04:47:42,312 --> 04:47:43,947 LOOK AT THE SCALE OF MODELS, 6362 04:47:43,947 --> 04:47:48,017 THIS IS WHAT MACHINE LEARNING 6363 04:47:48,017 --> 04:47:49,486 OPTIMIZES FOR VERSUS AVERAGE 6364 04:47:49,486 --> 04:47:50,754 ALIGNMENT OF MODELS THEN THERE'S 6365 04:47:50,754 --> 04:47:54,591 A BIT OF DISTINCTION. THE GRAY 6366 04:47:54,591 --> 04:47:59,295 LINE IS A RANDOM HIGH DIMENSIONAL 6367 04:47:59,295 --> 04:48:03,600 EMBEDDING. MANY FEATURES ALONE DO NOT HELP. THEY DO NOT GIVE YOU HIGH ALIGNMENT. 6368 04:48:03,600 --> 04:48:05,869 IF YOU LOOK AT THE ALIGNMENT OF MODELS WITH RESPECT TO BEHAVIOR AND TO NEURONS SEPARATELY, 6369 04:48:05,869 --> 04:48:07,737 IT DOES SEEM TO CONTINUE TO SCALE 6370 04:48:07,737 --> 04:48:10,874 AT LEAST IN VISION AGAIN. 6371 04:48:10,874 --> 04:48:16,513 FOR INTERNAL ALIGNMENT OF MODELS 6372 04:48:16,513 --> 04:48:18,314 IT DOES SEEM WE'RE NOT GOING TO 6373 04:48:18,314 --> 04:48:20,049 GET TO A MODEL THAT'S REALLY 6374 04:48:20,049 --> 04:48:20,750 LIKE THE BRAIN. 6375 04:48:20,750 --> 04:48:22,619 I THINK THERE'S SOMETHING THAT 6376 04:48:22,619 --> 04:48:24,154 NEEDS TO BE DONE AND TO ME THAT 6377 04:48:24,154 --> 04:48:28,158 MEANS THIS IS A FRONTIER FOR OUR 6378 04:48:28,158 --> 04:48:30,693 FIELD OF ACTUALLY BUILDING 6379 04:48:30,693 --> 04:48:31,528 BRAINLIKE MODELS TRULY 6380 04:48:31,528 --> 04:48:33,429 RESPECTING BREADTH AND DEPTH OF 6381 04:48:33,429 --> 04:48:34,998 THE BRAIN MECHANISMS, I'LL SPEAK 6382 04:48:34,998 --> 04:48:38,268 TO THAT IN MORE DETAIL BUT I 6383 04:48:38,268 --> 04:48:40,637 WANT TO SAY OUR FIELD NEEDS TO 6384 04:48:40,637 --> 04:48:43,339 BUILD MODELS THAT DO NOT ABANDON 6385 04:48:43,339 --> 04:48:49,179 WHAT WE'VE LEARNED FROM A.I. BUT 6386 04:48:49,179 --> 04:48:49,712 ENHANCE IT WITH BRAIN INTERNALS. 6387 04:48:49,712 --> 04:48:51,014 I'LL WALK THROUGH A FEW THINGS 6388 04:48:51,014 --> 04:48:54,417 WE NEED FOR THAT. 6389 04:48:54,417 --> 04:48:56,352 ONE IS ACCESSIBLE HIGH-FIDELITY 6390 04:48:56,352 --> 04:48:56,786 DATA. 6391 04:48:56,786 --> 04:48:58,188 BRAIN AND OTHER INITIATIVES HAVE 6392 04:48:58,188 --> 04:49:00,557 BEEN GREAT IN GIVING LARGE SCALE 6393 04:49:00,557 --> 04:49:00,824 DATA. 6394 04:49:00,824 --> 04:49:01,925 THERE'S STILL CHALLENGES IN EVEN 6395 04:49:01,925 --> 04:49:03,293 THOUGH DATA IS SUPPOSED TO BE 6396 04:49:03,293 --> 04:49:05,929 OPEN, WE DON'T HAVE STANDARDS. 6397 04:49:05,929 --> 04:49:08,731 ONES WE DO DON'T REINFORCE 6398 04:49:08,731 --> 04:49:10,833 METADATA, OFTEN FROM PERSONAL 6399 04:49:10,833 --> 04:49:11,467 PAINSTAKING EXPERIENCE, IT'S 6400 04:49:11,467 --> 04:49:12,835 EXTREMELY HARD TO REALLY GET 6401 04:49:12,835 --> 04:49:17,040 DATA USEFUL, TO BE USEFUL, OFTEN 6402 04:49:17,040 --> 04:49:18,675 NOT COLLECTED WITH MODELS IN 6403 04:49:18,675 --> 04:49:18,842 MIND. 6404 04:49:18,842 --> 04:49:21,311 IT'S SO LIMITED IN ITS SCALE OR 6405 04:49:21,311 --> 04:49:23,313 IN STORABILITY YOU CAN'T RELEASE 6406 04:49:23,313 --> 04:49:25,582 IT OR TEST MODELS AT SCALE. 6407 04:49:25,582 --> 04:49:27,750 THERE ARE EFFORTS TO TRY AND FIX 6408 04:49:27,750 --> 04:49:28,084 THAT. 6409 04:49:28,084 --> 04:49:29,586 ALL THE POWER TO THEM. 6410 04:49:29,586 --> 04:49:32,455 ALSO IF YOU HAVE DATA THAT'S 6411 04:49:32,455 --> 04:49:35,592 SMALL SCALE, OFTEN THAT DOES NOT 6412 04:49:35,592 --> 04:49:37,227 HELP TO DIFFERENTIATE BETWEEN 6413 04:49:37,227 --> 04:49:39,829 MODELS, THIS IS ABOUT RAISING 6414 04:49:39,829 --> 04:49:41,097 THE CEILING. 6415 04:49:41,097 --> 04:49:43,766 CURRENTLY IN LANGUAGE I THINK 6416 04:49:43,766 --> 04:49:45,501 ALSO OTHER FIELDS, SUBFIELDS, 6417 04:49:45,501 --> 04:49:49,072 WE'RE GETTING TO A POINT FOR A 6418 04:49:49,072 --> 04:49:51,874 LOT OF DATASETS, WHICH MODELS 6419 04:49:51,874 --> 04:49:52,508 ARE BETTER, WHICH WORSE, 6420 04:49:52,508 --> 04:49:53,743 EVERYTHING LOOKS TO BE THE SAME. 6421 04:49:53,743 --> 04:49:55,245 WE NEED TO DEVELOP MORE DATA 6422 04:49:55,245 --> 04:50:00,850 WITH MODELS IN MIND AND SEPARATE 6423 04:50:00,850 --> 04:50:01,985 BETWEEN MODEL TYPES. 6424 04:50:01,985 --> 04:50:06,089 THE OPPORTUNITY TO ME IN THIS, 6425 04:50:06,089 --> 04:50:08,625 WE SHOULD COLLECT NEW DATA, THAT 6426 04:50:08,625 --> 04:50:11,628 IS MAYBE NOT JUST FOCUS ON A A 6427 04:50:11,628 --> 04:50:13,997 TINY CIRCUIT OR ASPECT OF BRAIN 6428 04:50:13,997 --> 04:50:15,098 PROCESSING BUT IS MULTI-SYSTEM 6429 04:50:15,098 --> 04:50:16,866 FROM VISION TO LANGUAGE TO MAYBE 6430 04:50:16,866 --> 04:50:20,737 ACTUAL REASONING, THIS IS ALSO 6431 04:50:20,737 --> 04:50:21,904 MULTI-DOMAIN, IT SPANS DIFFERENT 6432 04:50:21,904 --> 04:50:22,272 BRAIN AREAS. ALSO DATA THAT WE 6433 04:50:22,272 --> 04:50:26,776 SHOULD COLLECT INVOLVE 6434 04:50:26,776 --> 04:50:29,746 CAUSAL MANIPULATION TO GET 6435 04:50:29,746 --> 04:50:31,281 US AT SOMETHING TO TURN TO 6436 04:50:31,281 --> 04:50:34,183 TREATMENT FOR PEOPLE WITH 6437 04:50:34,183 --> 04:50:34,751 DISORDERS. 6438 04:50:34,751 --> 04:50:36,386 THE SECOND POINT THAT I THINK WE 6439 04:50:36,386 --> 04:50:39,656 NEED IS WE NEED TO HAVE SOME 6440 04:50:39,656 --> 04:50:40,456 REFERENCE POINT FOR WHETHER OR 6441 04:50:40,456 --> 04:50:43,192 NOT CHANGES TO THE MODELS WILL 6442 04:50:43,192 --> 04:50:45,028 MAKE THEM BETTER. 6443 04:50:45,028 --> 04:50:47,430 WE'RE PROBABLY ALL DISAGREE ON 6444 04:50:47,430 --> 04:50:48,765 METRICS BUT TESTING MODELS IN 6445 04:50:48,765 --> 04:50:52,902 COMPARABLE WAY AND SEEING IF 6446 04:50:52,902 --> 04:50:55,805 MODEL UPDATE X IS BETTER IS 6447 04:50:55,805 --> 04:50:56,039 USEFUL. 6448 04:50:56,039 --> 04:51:01,044 FROM EXPERIENCE, VERY EFFICIENT 6449 04:51:01,044 --> 04:51:03,146 WAY TO MAKE PROGRESS. 6450 04:51:03,146 --> 04:51:04,781 BRAIN SCORE PLATFORM HAS A LOT 6451 04:51:04,781 --> 04:51:06,683 OF DATA AND ONE THEME THAT I 6452 04:51:06,683 --> 04:51:10,687 WANT TO HIGHLIGHT RECENTLY HAD A 6453 04:51:10,687 --> 04:51:12,455 COMPETITION WHERE WE TURNED 6454 04:51:12,455 --> 04:51:13,489 AROUND AND ASKED 6455 04:51:13,489 --> 04:51:14,324 EXPERIMENTALISTS TO SUBMIT 6456 04:51:14,324 --> 04:51:17,226 BENCHMARKS THAT FALSIFY THE 6457 04:51:17,226 --> 04:51:17,493 MODELS. 6458 04:51:17,493 --> 04:51:19,862 THAT IS INCENTIVE WE NEED MORE, 6459 04:51:19,862 --> 04:51:23,166 NOT JUST ABOUT MODELERS, BUT 6460 04:51:23,166 --> 04:51:25,568 PRODUCING DATA THAT CAN FALSIFY 6461 04:51:25,568 --> 04:51:26,602 MODELS AND SHOW MODELERS AGAIN 6462 04:51:26,602 --> 04:51:30,606 WHERE WE NEED TO PUT MORE 6463 04:51:30,606 --> 04:51:30,973 RESOURCES. 6464 04:51:30,973 --> 04:51:35,812 AND THEN FINAL POINT WE NEED IS 6465 04:51:35,812 --> 04:51:36,813 MODELS THAT RESPECT THE 6466 04:51:36,813 --> 04:51:39,182 INTRICACIES OF THE BRAIN MORE, I 6467 04:51:39,182 --> 04:51:40,183 WANT TO GIVE EXAMPLES. 6468 04:51:40,183 --> 04:51:45,355 AGAIN THE THEME OF BREADTH AND 6469 04:51:45,355 --> 04:51:49,359 DEPTH. 6470 04:51:49,359 --> 04:51:51,994 BREADTH BUILDING ON FEDORENKO'S 6471 04:51:51,994 --> 04:51:53,529 REVIEWS, MULTIPLE SYSTEMS, WILL 6472 04:51:53,529 --> 04:51:57,900 HELP US TO CONNECT FROM SENSORY 6473 04:51:57,900 --> 04:51:59,102 PROCESSING THROUGH REPRESENTATIONS 6474 04:51:59,102 --> 04:52:00,103 TO REASONING BEHAVIORS. 6475 04:52:00,103 --> 04:52:02,839 I THINK WE SHOULD ALSO RESPECT 6476 04:52:02,839 --> 04:52:05,375 MORE THE BIOLOGIC DETAIL THAT 6477 04:52:05,375 --> 04:52:07,110 COMES FROM THE BRAIN AND REFERS 6478 04:52:07,110 --> 04:52:11,013 TO DIFFERENT THEMES THAT COULD 6479 04:52:11,013 --> 04:52:11,481 BE. 6480 04:52:11,481 --> 04:52:13,549 ONE THEME WOULD BE TOPOGRAPHY, 6481 04:52:13,549 --> 04:52:14,917 THE BRAIN IS SPATIALLY 6482 04:52:14,917 --> 04:52:17,220 ORGANIZED, RATHER THAN A CLUMP 6483 04:52:17,220 --> 04:52:18,921 OF FACTORS. 6484 04:52:18,921 --> 04:52:21,391 BUT I THINK FOR WHAT DEPTH 6485 04:52:21,391 --> 04:52:23,659 REALLY MEANS, HOW TO BUILD 6486 04:52:23,659 --> 04:52:26,429 MODELS THAT ARE FUNCTION ALIGNED 6487 04:52:26,429 --> 04:52:28,297 SHOULD BE A THEME BUT WE SHOULD 6488 04:52:28,297 --> 04:52:30,266 AGREE ON TESTING, AT LEAST TELL 6489 04:52:30,266 --> 04:52:31,501 WHICH OF THE DIFFERENT CHANGES 6490 04:52:31,501 --> 04:52:34,137 LEADS TO IMPROVEMENT IN THE 6491 04:52:34,137 --> 04:52:34,670 MODEL SPACE. 6492 04:52:34,670 --> 04:52:37,940 AND SO THEN IF WE FOLLOW THROUGH 6493 04:52:37,940 --> 04:52:40,009 ON THE THREE POINTS I THINK 6494 04:52:40,009 --> 04:52:41,544 THERE'S AN EXCITING TIME FOR THE 6495 04:52:41,544 --> 04:52:44,914 FIELD. WE CAN TURN MODELS WE'RE 6496 04:52:44,914 --> 04:52:46,249 BUILDING INTO CLINICAL APPLICATIONS 6497 04:52:46,249 --> 04:52:48,251 TO ME, THERE'S BEEN TALK OF 6498 04:52:48,251 --> 04:52:50,086 FOUNDATION MODELS BUT IF WE HAD 6499 04:52:50,086 --> 04:52:54,157 A TRUE FOUNDATION MODEL IN OUR 6500 04:52:54,157 --> 04:52:55,925 FIELD IT WOULD ACCEPT DIFFERENT 6501 04:52:55,925 --> 04:53:03,533 SENSORY INPUT. VISUAL BEHAVIOR AND 6502 04:53:03,533 --> 04:53:05,668 CORRESPONDING ACTIVITY, SINCE WE 6503 04:53:05,668 --> 04:53:07,270 WANT TO GO BROADER, WE SHOULD 6504 04:53:07,270 --> 04:53:12,909 ALSO INCLUDE A LOT OF DIFFERENT 6505 04:53:12,909 --> 04:53:14,911 BEHAVIORS, LANGUAGE ABILITIES THAT 6506 04:53:14,911 --> 04:53:16,279 ACCEPTS DIFFERENT SENSORY INPUTS. 6507 04:53:16,279 --> 04:53:18,915 DEPTH WOULD ENABLE NEW LEVERS, 6508 04:53:18,915 --> 04:53:20,616 FOR INSTANCE NEURAL INTERVENTIONS 6509 04:53:20,616 --> 04:53:24,987 THAT PREDICT BEHAVIORAL EFFECT 6510 04:53:24,987 --> 04:53:27,190 OF MUSCIMOL OR MICROSTIMULATION 6511 04:53:27,190 --> 04:53:28,224 AS YOU STIMULATE SOMEONE'S BRAIN. 6512 04:53:28,224 --> 04:53:31,727 >> TWO-MINUTE WARNING. 6513 04:53:31,727 --> 04:53:34,597 >> TO GIVE TWO EXAMPLES, LET'S 6514 04:53:34,597 --> 04:53:36,098 SAY WE HAVE A PERFECT MODEL, 6515 04:53:36,098 --> 04:53:38,201 TURN IT AROUND AND ASK WHAT KIND 6516 04:53:38,201 --> 04:53:39,936 OF STIMULI CAN WE PRESENT TO DRIVE 6517 04:53:39,936 --> 04:53:42,138 OR SUPPRESS NEURAL ACTIVITY, AND 6518 04:53:42,138 --> 04:53:45,575 WE'VE DONE A PRELIMINARY VERSION 6519 04:53:45,575 --> 04:53:46,976 IN LANGUAGE, WHERE IF YOU TAKE 6520 04:53:46,976 --> 04:53:48,611 THE BEST MODEL THAT'S MOST 6521 04:53:48,611 --> 04:53:50,246 PREDICTIVE OF BRAIN DATA IN THE 6522 04:53:50,246 --> 04:53:54,317 LANGUAGE SYSTEM AND ASK IT TO 6523 04:53:54,317 --> 04:53:57,854 PRODUCE STIMULI THAT DRIVE OR 6524 04:53:57,854 --> 04:53:59,489 SUPPRESS ACTIVITY, RUN IN HUMANS, 6525 04:53:59,489 --> 04:54:03,326 THAT WORKS WEKK. THIS IS A 6526 04:54:03,326 --> 04:54:04,393 FORM OF NON-INVASIVE NEURAL 6527 04:54:04,393 --> 04:54:05,962 CONTROL THAT I THINK WITHOUT 6528 04:54:05,962 --> 04:54:08,231 THESE MODELS WOULD NOT 6529 04:54:08,231 --> 04:54:09,132 BE POSSIBLE. FINALLY THINK OF 6530 04:54:09,132 --> 04:54:11,334 MODELS THAT PREDICT NOT JUST NEURAL 6531 04:54:11,334 --> 04:54:18,174 ACTIVITY BUT EFFECT OF CAUSAL 6532 04:54:18,174 --> 04:54:21,677 INTERVENTION. WHAT IF YOU MICRO- 6533 04:54:21,677 --> 04:54:23,513 STIMULATE, APPLY OPTOGENETICS? 6534 04:54:23,513 --> 04:54:28,384 WE'VE DONE THIS FOR TOPOGRAPHIC 6535 04:54:28,384 --> 04:54:29,151 MODELS IN VISION, APPLY MICRO- 6536 04:54:29,151 --> 04:54:31,354 STIMULATION AND IT PREDICTS 6537 04:54:31,354 --> 04:54:31,554 BIOLOGICAL DATA QUITE NICELY. 6538 04:54:31,554 --> 04:54:36,526 SO TO ME I THINK THE 10-YEAR 6539 04:54:36,526 --> 04:54:37,693 PROCESS OF OUR FIELD WILL 6540 04:54:37,693 --> 04:54:41,998 CONTINUE TO NEED LARGE SCALE 6541 04:54:41,998 --> 04:54:45,034 DATA, WE'LL NEED DETAILED DATA, 6542 04:54:45,034 --> 04:54:47,169 STILL DISCUSSING METRICS, SHARED 6543 04:54:47,169 --> 04:54:48,371 SET OF REFERENCE POINTS. WE 6544 04:54:48,371 --> 04:54:53,376 CAN BUILD ON A.I. BY BRINGING OUR 6545 04:54:53,376 --> 04:54:57,813 OWN EXPERTISE IN BUILDING MODELS 6546 04:54:57,813 --> 04:54:59,649 TOWARD CLINICAL APPLICATIONS TO 6547 04:54:59,649 --> 04:55:01,717 WHERE WE CAN DRIVE A LOT OF 6548 04:55:01,717 --> 04:55:06,155 VALUE FOR US AND HUMANS IN 6549 04:55:06,155 --> 04:55:12,695 GENERAL. 6550 04:55:12,695 --> 04:55:15,198 >> I'D LIKE TO INVITE SESSION 2 6551 04:55:15,198 --> 04:55:19,402 DISCUSSANTS TO THE STAGE. 6552 04:55:19,402 --> 04:55:21,671 >> THIS IS A GOOD TIME TO SAY IF 6553 04:55:21,671 --> 04:55:26,142 YOU'RE NOT SPEAKING YOUR 6554 04:55:26,142 --> 04:55:27,810 MICROPHONE SHOULD BE MUTED. 6555 04:55:27,810 --> 04:55:30,980 ONLY SO MANY CAN BE ACTIVATED AT 6556 04:55:30,980 --> 04:55:41,257 THE SAME TIME. 6557 04:55:50,032 --> 04:55:54,337 WE'RE GOING TO HAVE EACH 6558 04:55:54,337 --> 04:55:59,108 DISCUSSANT ASK THEIR 6559 04:55:59,108 --> 04:55:59,375 QUESTION. 6560 04:55:59,375 --> 04:56:05,481 I'LL START WITH MARK HISTED FROM 6561 04:56:05,481 --> 04:56:07,650 THE NIMH INTRAMURAL PROGRAM. 6562 04:56:07,650 --> 04:56:08,217 >> THANK YOU. 6563 04:56:08,217 --> 04:56:10,186 THANKS TO ALL THE SPEAKERS. 6564 04:56:10,186 --> 04:56:12,688 THAT WAS VERY I THINK A LOT OF 6565 04:56:12,688 --> 04:56:13,689 REALLY INTERESTING IDEAS 6566 04:56:13,689 --> 04:56:14,023 INTRODUCED. 6567 04:56:14,023 --> 04:56:16,659 I HAVE A QUESTION THAT'S REALLY 6568 04:56:16,659 --> 04:56:20,262 FOR THE WHOLE PANEL, INSPIRED BY 6569 04:56:20,262 --> 04:56:22,565 A COUPLE DIFFERENT COMMENTS, A 6570 04:56:22,565 --> 04:56:23,766 COUPLE DIFFERENT TALKS THAT HAD 6571 04:56:23,766 --> 04:56:25,401 SORT OF A COMMON THREAD RUNNING 6572 04:56:25,401 --> 04:56:30,106 THROUGH ABOUT THIS KIND OF LIKE 6573 04:56:30,106 --> 04:56:31,874 WORSE IS BETTER APPROACH BLAKE 6574 04:56:31,874 --> 04:56:33,943 REFERRED TO, TRYING TO BUILD 6575 04:56:33,943 --> 04:56:35,277 MODELS THAT WORK AND SEE WHERE 6576 04:56:35,277 --> 04:56:36,712 WE GET TO. 6577 04:56:36,712 --> 04:56:38,347 I THINK I'M SYMPATHETIC TO THE 6578 04:56:38,347 --> 04:56:43,386 IDEA OF CONTROL LOOPS, LEARNING 6579 04:56:43,386 --> 04:56:44,487 FROM FEEDBACK, RECURRENT 6580 04:56:44,487 --> 04:56:44,754 NETWORKS. 6581 04:56:44,754 --> 04:56:48,524 AND WHEN IT COMES TO USING THESE 6582 04:56:48,524 --> 04:56:51,394 APPROACHES, WE THINK ABOUT USING 6583 04:56:51,394 --> 04:56:53,462 NEUROAI TO UNDERSTAND THE 6584 04:56:53,462 --> 04:56:55,297 BRAIN, FROM MY PLACE 6585 04:56:55,297 --> 04:56:57,933 IN THE MENTAL HEALTH INSTITUTE, 6586 04:56:57,933 --> 04:57:00,336 I'LL SAY WE'VE HAD A LONG EFFORT 6587 04:57:00,336 --> 04:57:05,941 OVER THE PAST TWO DECADES TO USE 6588 04:57:05,941 --> 04:57:07,476 GENETIC DATA TO UNDERSTAND 6589 04:57:07,476 --> 04:57:08,811 MENTAL DISEASE AND BEHAVIOR. 6590 04:57:08,811 --> 04:57:15,484 WE HAVE APPLIED A SORT OF DATA- 6591 04:57:15,484 --> 04:57:17,987 BASED APPROACH TO COLLECT 6592 04:57:17,987 --> 04:57:20,956 GENETIC DATA AND TRY TOPREDICT 6593 04:57:20,956 --> 04:57:22,491 DISEASE PATHOLOGY, LET'S SAY, 6594 04:57:22,491 --> 04:57:23,592 PREDICT WHEN DISEASE WILL 6595 04:57:23,592 --> 04:57:25,261 HAPPEN, ALSO TRY TO DEVELOP WAYS 6596 04:57:25,261 --> 04:57:26,562 TO TREAT THEM. 6597 04:57:26,562 --> 04:57:29,632 I THINK THERE'S WIDESPREAD 6598 04:57:29,632 --> 04:57:30,933 FEELING THAT THAT HASN'T 6599 04:57:30,933 --> 04:57:36,305 PRODUCED EFFECTS THAT PEOPLE 6600 04:57:36,305 --> 04:57:36,806 WANTED. 6601 04:57:36,806 --> 04:57:39,175 I THINK -- SO WHEN WE THINK -- 6602 04:57:39,175 --> 04:57:42,712 ONE CAN THINK ABOUT THAT FROM 6603 04:57:42,712 --> 04:57:44,213 LIKE DATA-CENTERED APPROACH 6604 04:57:44,213 --> 04:57:46,348 WHERE THEY JUST TRY TO BRUTE 6605 04:57:46,348 --> 04:57:47,316 FORCE THROUGH GENETIC DATA AT 6606 04:57:47,316 --> 04:57:48,718 THINGS BUT YOU CAN ALSO THINK 6607 04:57:48,718 --> 04:57:51,787 ABOUT THERE'S A LOT OF THEORIES 6608 04:57:51,787 --> 04:57:53,422 THAT WERE DEVELOPED AROUND THAT 6609 04:57:53,422 --> 04:57:57,226 FROM THE DSM ABOUT DIFFERENT 6610 04:57:57,226 --> 04:57:57,493 PATHOLOGY. 6611 04:57:57,493 --> 04:58:00,896 MY QUESTION IS GIVEN WHAT WE 6612 04:58:00,896 --> 04:58:05,134 HAVE HERE FROM NEUROAI 6613 04:58:05,134 --> 04:58:06,736 APPROACHES, HOW CAN WE APPLY 6614 04:58:06,736 --> 04:58:08,437 UNDERSTANDING OF NEURAL CIRCUITS 6615 04:58:08,437 --> 04:58:10,172 AND NEURAL COMPUTATION AND BRAIN 6616 04:58:10,172 --> 04:58:14,810 FUNCTION AND NETWORKS TO MAKE 6617 04:58:14,810 --> 04:58:22,785 PROGRESS ON UNDERSTANDING OF 6618 04:58:22,785 --> 04:58:24,887 MENTAL DISEASE? 6619 04:58:24,887 --> 04:58:26,622 >> FROM MY PERSPECTIVE I THINK 6620 04:58:26,622 --> 04:58:28,057 WE'LL DEFINITELY NEED TO GO 6621 04:58:28,057 --> 04:58:29,125 BROADER THAN TODAY. 6622 04:58:29,125 --> 04:58:33,496 RIGHT NOW THE WAY WE'RE TACKLING 6623 04:58:33,496 --> 04:58:36,799 MODELING THE BRAIN, DIVIDE-AND- 6624 04:58:36,799 --> 04:58:38,000 CONQUER APPROACH, ONE GROUP ON 6625 04:58:38,000 --> 04:58:39,835 VISION, ONE GROUP ON LANGUAGE. 6626 04:58:39,835 --> 04:58:42,805 FOR SOMETHING AS COMPLEX AS 6627 04:58:42,805 --> 04:58:45,307 MENTAL DISORDERS, MOOD IN 6628 04:58:45,307 --> 04:58:47,042 GENERAL, YOU NEED MORE 6629 04:58:47,042 --> 04:58:47,309 INTERPLAY. 6630 04:58:47,309 --> 04:58:48,611 UNFORTUNATELY THAT WOULD REQUIRE 6631 04:58:48,611 --> 04:58:52,314 US TO GO FROM THESE MORE MODULAR 6632 04:58:52,314 --> 04:58:53,649 APPROACHES TO A MORE INTEGRATED 6633 04:58:53,649 --> 04:58:57,520 LIKE MAYBE TOWARDS THE WHOLE 6634 04:58:57,520 --> 04:59:00,589 BRAIN MODEL THAT HAS MORE 6635 04:59:00,589 --> 04:59:01,791 COMPLICATED BEHAVIORS, WE'LL HAVE 6636 04:59:01,791 --> 04:59:03,559 TO BUILD THOSE KINDS OF MODELS 6637 04:59:03,559 --> 04:59:05,594 OURSELVES, I DON'T THINK WE CAN 6638 04:59:05,594 --> 04:59:07,363 RELY ON MACHINE LEARNING FOR 6639 04:59:07,363 --> 04:59:17,306 THAT BUT BASICALLY GOING MORE 6640 04:59:17,306 --> 04:59:17,773 BROAD. 6641 04:59:17,773 --> 04:59:19,642 >> I DON'T KNOW IN I'M ALLOWED 6642 04:59:19,642 --> 04:59:20,743 TO ANSWER THIS TOO. 6643 04:59:20,743 --> 04:59:24,246 ONE THING YOU TALKED ABOUT 6644 04:59:24,246 --> 04:59:26,215 TRYING TO PREDICT MENTAL HEALTH 6645 04:59:26,215 --> 04:59:27,249 TRAJECTORY WITH GENETIC DATA BUT 6646 04:59:27,249 --> 04:59:28,951 I THINK A LOT OF WHAT'S GOING TO 6647 04:59:28,951 --> 04:59:31,153 HAPPEN HERE, WE DIDN'T TALK 6648 04:59:31,153 --> 04:59:32,121 ABOUT COMPUTATIONAL PSYCHIATRY, 6649 04:59:32,121 --> 04:59:35,191 OR EVEN THE BRAIN BEHAVIORAL 6650 04:59:35,191 --> 04:59:37,493 QUANTIFICATION SYNCHRONIZATION BIT 6651 04:59:37,493 --> 04:59:39,695 WHERE WE'RE NOT JUST USING DATA 6652 04:59:39,695 --> 04:59:41,197 TO DERIVE SOME DATA COLLECTING 6653 04:59:41,197 --> 04:59:46,469 IN THE RESEARCH LAB TO BUILD 6654 04:59:46,469 --> 04:59:47,903 MODELS BUT ACTUALLY USING DATA 6655 04:59:47,903 --> 04:59:49,638 COLLECTED IN THE WILD, ALSO 6656 04:59:49,638 --> 04:59:50,973 USING DATA COLLECTED FROM 6657 04:59:50,973 --> 04:59:52,842 MEDICAL HEALTH RECORDS, TRYING 6658 04:59:52,842 --> 04:59:55,678 TO PUT THEM IN. DATA THAT COMES 6659 04:59:55,678 --> 04:59:57,780 FROM THOSE POPULATIONS TENDS TO 6660 04:59:57,780 --> 04:59:59,682 BE NOT REPRESENTATIVE, SO IF 6661 04:59:59,682 --> 05:00:00,850 WE'RE BUILDING MODELS BASED ON 6662 05:00:00,850 --> 05:00:02,918 BRAIN DATA THAT ISN'T 6663 05:00:02,918 --> 05:00:04,987 REPRESENTATIVE WE'LL HAVE EQUAL 6664 05:00:04,987 --> 05:00:07,289 HOLES, NOT JUST THAT DATA ISN'T 6665 05:00:07,289 --> 05:00:17,466 REPRESENTATIVE, NOT INCLUDE SOCIAL 6666 05:00:17,466 --> 05:00:18,234 FACTORS THAT INFLUENCE DISORDERS. 6667 05:00:18,234 --> 05:00:21,303 >> THERE'S A GAP BETWEEN GENE 6668 05:00:21,303 --> 05:00:26,242 AND BEHAVIOR, SEVERAL 6669 05:00:26,242 --> 05:00:26,809 INTERMEDIATE LEVELS, I THINK 6670 05:00:26,809 --> 05:00:34,817 THAT'S ONE OF THE BIGGEST 6671 05:00:34,817 --> 05:00:35,784 REASONS FOR FAILURE. 6672 05:00:35,784 --> 05:00:41,257 AT LEAST WE HAVE GOOD 6673 05:00:41,257 --> 05:00:43,125 BIOPHYSICAL MODEL OF NEURONAL 6674 05:00:43,125 --> 05:00:45,895 AND CIRCUIT LEVEL. LOOK AT IMPACT 6675 05:00:45,895 --> 05:00:53,569 OF THE GENE ON THE COMPUTATION. 6676 05:00:53,569 --> 05:00:54,970 HAVE A MECHECHANISTIC UNDER- 6677 05:00:54,970 --> 05:00:56,438 STANDING OR HYPOTHESIS AND THEN 6678 05:00:56,438 --> 05:01:06,882 GO TO BIGGER A.I. MASSIVE 6679 05:01:07,783 --> 05:01:08,984 APPROACH OF MULTIPLE BRAIN AREAS AND INTEGRATE THESE TWO APPROACHES 6680 05:01:08,984 --> 05:01:09,919 TO COME UP WITH BETTER INSIGHTS. 6681 05:01:09,919 --> 05:01:11,120 >> TO BRIDGE THAT GAP 6682 05:01:11,120 --> 05:01:12,988 SPECIFICALLY WHAT DO WE SEE 6683 05:01:12,988 --> 05:01:14,189 THESE KIND OF NEUROAI 6684 05:01:14,189 --> 05:01:15,391 APPROACHES THAT WE'VE DISCUSSED 6685 05:01:15,391 --> 05:01:18,227 TODAY DOING? LIKE WHERE ARE THEY 6686 05:01:18,227 --> 05:01:18,727 POINTING? 6687 05:01:18,727 --> 05:01:19,962 TO ME IT FEELS OPTIMISTIC. 6688 05:01:19,962 --> 05:01:21,630 I FEEL LIKE THIS IS THE 6689 05:01:21,630 --> 05:01:23,365 DIRECTION THAT'S GOING TO GIVE 6690 05:01:23,365 --> 05:01:25,134 US UNDERSTANDING OF NEURAL 6691 05:01:25,134 --> 05:01:26,869 COMPUTATIONS, GOING TO LEAD TO 6692 05:01:26,869 --> 05:01:28,170 DEEPER UNDERSTANDING OF WHAT'S 6693 05:01:28,170 --> 05:01:30,372 HAPPENING, I WONDER WHAT THE 6694 05:01:30,372 --> 05:01:33,542 PANEL THINKS, THE REST OF 6695 05:01:33,542 --> 05:01:35,210 EVERYONE. 6696 05:01:35,210 --> 05:01:37,713 >> WHY IS ONE PERSON DEPRESSING, 6697 05:01:37,713 --> 05:01:40,015 ANOTHER HAS FULL HOPE, WHY DO 6698 05:01:40,015 --> 05:01:41,884 YOU SEE AN IMAGE AS A FACE AT 6699 05:01:41,884 --> 05:01:43,252 ONE MOMENT AND A VASE AT 6700 05:01:43,252 --> 05:01:46,789 ANOTHER, COULD BE THE SAME BASIC 6701 05:01:46,789 --> 05:01:48,557 MECHANISM TOP-DOWN PRIORS, ONE OF 6702 05:01:48,557 --> 05:01:49,758 THE MOST AMAZING INSIGHTS FROM A.I. 6703 05:01:49,758 --> 05:01:51,961 HAS BEEN HOW THE SAME 6704 05:01:51,961 --> 05:01:53,162 ARCHITECTURE LIKE TRANSFORMERS 6705 05:01:53,162 --> 05:01:54,563 CAN SOLVE SUCH DIVERSE PROBLEMS, 6706 05:01:54,563 --> 05:01:57,433 I SUSPECT WHEN WE UNDERSTAND THE 6707 05:01:57,433 --> 05:01:58,734 BASIC MECHANISMS FOR EVEN 6708 05:01:58,734 --> 05:02:00,803 VISUAL PERCEPTION 6709 05:02:00,803 --> 05:02:03,572 IT'S GOING TO GIVE 6710 05:02:03,572 --> 05:02:07,076 BIG CLUES INTO MENTAL ILLNESS. 6711 05:02:07,076 --> 05:02:10,312 >> I'D LIKE TO TRANSITION TO THE 6712 05:02:10,312 --> 05:02:19,154 NEXT DISCUSSANT, FROM YALE 6713 05:02:19,154 --> 05:02:20,255 UNIVERSITY, STEVE ZUCKER. 6714 05:02:20,255 --> 05:02:25,060 >> I LOVED THE TALKS THIS 6715 05:02:25,060 --> 05:02:25,828 AFTERNOON. 6716 05:02:25,828 --> 05:02:27,363 DORIS PROVIDED TWO EXAMPLES OF 6717 05:02:27,363 --> 05:02:29,999 DEEP NETWORKS THAT SHE USED IN 6718 05:02:29,999 --> 05:02:32,167 HER WORK TO GAIN INSPIRATION FOR 6719 05:02:32,167 --> 05:02:34,269 WHAT WAS GOING IN NEUROSCIENCE, 6720 05:02:34,269 --> 05:02:38,507 AND IN HER WORDS SHE SAID DEEP 6721 05:02:38,507 --> 05:02:41,043 NETWORKS LET US DESCRIBE COMPLEX 6722 05:02:41,043 --> 05:02:41,276 OBJECTS. 6723 05:02:41,276 --> 05:02:45,314 AT THE END OF THE SESSION, 6724 05:02:45,314 --> 05:02:47,049 MARTIN TALKED ABOUT BRAIN SCORES 6725 05:02:47,049 --> 05:02:49,952 AND THE NEED FOR METRICS AND IN 6726 05:02:49,952 --> 05:02:52,421 FACT SEVERAL PEOPLE MENTIONED 6727 05:02:52,421 --> 05:02:54,723 THE NEED FOR METRICS TO COMPARE 6728 05:02:54,723 --> 05:02:56,058 SYSTEMS. 6729 05:02:56,058 --> 05:02:58,994 SO MY QUESTION IS, AND LET ME 6730 05:02:58,994 --> 05:03:04,299 PHRASE IT TWO WAYS. 6731 05:03:04,299 --> 05:03:08,637 INFORMALLY, DO SUCH METRICS 6732 05:03:08,637 --> 05:03:09,104 EXIST? 6733 05:03:09,104 --> 05:03:12,174 TECHNICALLY OVER WHAT SPACE OF 6734 05:03:12,174 --> 05:03:15,711 MODELS MIGHT ONE FIND THE 6735 05:03:15,711 --> 05:03:15,944 METRIC? 6736 05:03:15,944 --> 05:03:19,281 TO DO WHAT DORIS DID WHICH IS 6737 05:03:19,281 --> 05:03:22,017 TO GAIN INSPIRATION FROM AN 6738 05:03:22,017 --> 05:03:24,086 ARTIFACT, DO WE ACTUALLY NEED 6739 05:03:24,086 --> 05:03:24,686 METRICS? 6740 05:03:24,686 --> 05:03:27,156 THE FOURTH VERSION OF THE SAME 6741 05:03:27,156 --> 05:03:32,061 QUESTION IS DOES IT MAKE SENSE 6742 05:03:32,061 --> 05:03:34,797 TO RANK NETWORKS LIKE WE RANK 6743 05:03:34,797 --> 05:03:35,064 SWIMMERS? 6744 05:03:35,064 --> 05:03:39,735 SO WHO DID 100 METERS OF 6745 05:03:39,735 --> 05:03:40,536 FREESTYLE THE FASTEST? 6746 05:03:40,536 --> 05:03:42,905 THEY ARE ALL THE SAME QUESTION 6747 05:03:42,905 --> 05:03:44,673 BUT WITH DIFFERENT CLOTHING. 6748 05:03:44,673 --> 05:03:50,045 I DON'T KNOW, YEAH. 6749 05:03:50,045 --> 05:03:54,850 >> MAYBE I CAN START. 6750 05:03:54,850 --> 05:03:59,021 I MEAN SHORT QUESTION, WE NEED 6751 05:03:59,021 --> 05:04:01,223 METRICS, DEFINITELY YES TO 6752 05:04:01,223 --> 05:04:02,991 ASSESS PRODUCTS, WHETHER WE'RE 6753 05:04:02,991 --> 05:04:04,626 IMPROVING, WE NEED TO HAVE A 6754 05:04:04,626 --> 05:04:05,260 WAY TO MEASURE IMPROVEMENT. 6755 05:04:05,260 --> 05:04:06,295 I DON'T THINK IT'S AN EASY 6756 05:04:06,295 --> 05:04:06,829 ANSWER. 6757 05:04:06,829 --> 05:04:08,964 I DON'T KNOW WHAT THE RIGHT 6758 05:04:08,964 --> 05:04:10,632 METRIC IS. 6759 05:04:10,632 --> 05:04:13,068 I THINK IT DEPENDS A LOT ON KIND 6760 05:04:13,068 --> 05:04:14,703 OF NETWORKS, EASIER TO COMPARE 6761 05:04:14,703 --> 05:04:18,841 LET'S SAY WITHIN THE SAME SYSTEM 6762 05:04:18,841 --> 05:04:20,375 ARCHITECTURE, LEARNING, WHAT IS 6763 05:04:20,375 --> 05:04:22,644 THE ADVANTAGE OF ADDING MORE 6764 05:04:22,644 --> 05:04:24,646 BIOLOGY TO THESE NETWORKS, IT'S 6765 05:04:24,646 --> 05:04:26,615 AN EASIER COMPARISON, LET'S SAY 6766 05:04:26,615 --> 05:04:29,351 THAN GOING TO MULTIPLE DIFFERENT 6767 05:04:29,351 --> 05:04:31,587 MODELS SOLVING THE SAME TASK 6768 05:04:31,587 --> 05:04:34,890 BECAUSE THEN IT DEPENDS ON 6769 05:04:34,890 --> 05:04:36,525 ARCHITECTURURAL DIFFERENCES, SO 6770 05:04:36,525 --> 05:04:37,593 THAT WOULD BE MORE DIFFICULT TASK 6771 05:04:37,593 --> 05:04:38,794 TO SOLVE. WHAT'S IMPORTANT IS TO 6772 05:04:38,794 --> 05:04:42,064 CONSIDER THE BRAIN IS OPERATING 6773 05:04:42,064 --> 05:04:43,932 UNDER A SET OF CONSTRAINTS, 6774 05:04:43,932 --> 05:04:45,567 ENERGY, SPACE, SPEED 6775 05:04:45,567 --> 05:04:45,767 LET'SSAY. 6776 05:04:45,767 --> 05:04:48,370 SO THE METRICS THAT WE'RE USING 6777 05:04:48,370 --> 05:04:52,040 TO COMPARE A.I. AND BIOLOGICAL 6778 05:04:52,040 --> 05:04:55,110 SYSTEMS SHOULD OPERATE UNDER 6779 05:04:55,110 --> 05:04:56,011 VERY SIMILAR CONSTRAINTS FOR 6780 05:04:56,011 --> 05:04:56,979 THIS TO MAKE SENSE. 6781 05:04:56,979 --> 05:04:59,047 OTHERWISE IF WE DON'T HAVE THE 6782 05:04:59,047 --> 05:05:00,682 CONSTRAINTS WHAT IS THE 6783 05:05:00,682 --> 05:05:02,017 COMPARISON, IT'S NOT REALLY 6784 05:05:02,017 --> 05:05:03,752 FAIR, SO I DON'T HAVE AN ANSWER 6785 05:05:03,752 --> 05:05:06,622 BUT THESE ARE ALL THINGS THAT WE 6786 05:05:06,622 --> 05:05:07,556 NEED TO CONSIDER CAREFULLY TO 6787 05:05:07,556 --> 05:05:11,760 COME UP WITH THE RIGHT METRICS. 6788 05:05:11,760 --> 05:05:14,530 >> I CAN ADD TO THIS QUICKLY? 6789 05:05:14,530 --> 05:05:17,566 I HAVE A COUPLE POINTS IN 6790 05:05:17,566 --> 05:05:17,833 RESPONSE. 6791 05:05:17,833 --> 05:05:19,434 ONE IS THAT I THINK IT'S 6792 05:05:19,434 --> 05:05:21,737 BECOMING LESS USEFUL PERHAPS 6793 05:05:21,737 --> 05:05:25,007 TO JUST TAKE OFF-THE-SHELF 6794 05:05:25,007 --> 05:05:26,108 MODELS AGAINST NEURAL DATA. 6795 05:05:26,108 --> 05:05:26,542 RELATED TO WHAT DORIS WAS 6796 05:05:26,542 --> 05:05:27,876 POINTING TO. IF WE TREAT MODEL 6797 05:05:27,876 --> 05:05:31,980 ARCHITECTURES AS A WAY TO TEST 6798 05:05:31,980 --> 05:05:34,883 HYPOTHESES, WE CAN CREATE MINIMAL 6799 05:05:34,883 --> 05:05:36,218 PAIR VARIANTS. SO MANY THINGS 6800 05:05:36,218 --> 05:05:38,420 VARY ACROSS MODELS BUT WE CAN 6801 05:05:38,420 --> 05:05:41,590 MAKE MINIMAL PAIRS WHICH VARY IN 6802 05:05:41,590 --> 05:05:43,358 THE PRESENCE OR ABSENCE OF 6803 05:05:43,358 --> 05:05:46,762 INDUCTIVE BIAS OR IN THE PRESENCE 6804 05:05:46,762 --> 05:05:47,563 OF CERTAIN ARCHITECTURAL MOTIF AND 6805 05:05:47,563 --> 05:05:51,900 CONNECTIVITY PATTERN OR PRESENCE 6806 05:05:51,900 --> 05:05:56,838 OF BIOLOGICAL FEATURES, AND SEE 6807 05:05:56,838 --> 05:05:58,040 NECESSARY AND SUFFICIENT CONSTRAINTS 6808 05:05:58,040 --> 05:06:02,945 TO YIELD A REPRESENTATIONAL SPACE 6809 05:06:02,945 --> 05:06:04,112 OR HOW ALGORITHMS ARE INSTANTIATED. 6810 05:06:04,112 --> 05:06:05,948 ANOTHER SMALL POINT, MAYBE NOT A 6811 05:06:05,948 --> 05:06:07,282 SMALL POINT -- 6812 05:06:07,282 --> 05:06:10,419 >> I HAVE TO INTERRUPT. 6813 05:06:10,419 --> 05:06:11,820 REPRESENTATIONAL SPACE, WE'RE 6814 05:06:11,820 --> 05:06:12,487 COMPARING REPRESENTATIONS, ISN'T 6815 05:06:12,487 --> 05:06:15,224 THAT A VERSION OF THE SAME 6816 05:06:15,224 --> 05:06:15,791 PROBLEM? 6817 05:06:15,791 --> 05:06:17,993 >> WHAT DO YOU MEAN BY THAT. 6818 05:06:17,993 --> 05:06:20,829 >> TO COMPARE MODELS OR 6819 05:06:20,829 --> 05:06:22,130 REPRESENTATIONS IN TWO MODELS, 6820 05:06:22,130 --> 05:06:24,633 TO CLAIM THERE IS SOME METRIC 6821 05:06:24,633 --> 05:06:27,402 BETWEEN THEM, AREN'T THEY 6822 05:06:27,402 --> 05:06:29,171 EQUIVALENT PROBLEMS? 6823 05:06:29,171 --> 05:06:31,006 >> PEOPLE HAVE SOLVED SOME OF 6824 05:06:31,006 --> 05:06:31,640 THESE COMPARISON ISSUES, LIKE 6825 05:06:31,640 --> 05:06:35,143 THERE ARE WAYS TO COMPARE 6826 05:06:35,143 --> 05:06:35,777 REPRESENTATIONS WITH MULTIPLE 6827 05:06:35,777 --> 05:06:38,280 DIMENSIONS LIKE IT'S NOT A 6828 05:06:38,280 --> 05:06:38,780 TECHNICAL ISSUE. 6829 05:06:38,780 --> 05:06:41,283 ARE YOU ASKING ABOUT A 6830 05:06:41,283 --> 05:06:42,050 CONCEPTUAL PROBLEM? 6831 05:06:42,050 --> 05:06:42,684 >> YES, MATHEMATICAL QUESTION, 6832 05:06:42,684 --> 05:06:48,290 OVER WHAT SPACE ARE METRICS 6833 05:06:48,290 --> 05:06:49,658 DEFINED? 6834 05:06:49,658 --> 05:06:50,392 >> WELL, BLAKE DO YOU WANT TO GO 6835 05:06:50,392 --> 05:06:52,027 >> I'D LOVE TO JUMP IN. 6836 05:06:52,027 --> 05:06:53,462 THERE'S TWO THINGS. 6837 05:06:53,462 --> 05:06:55,731 FIRST TO ANSWER MOST DIRECTLY 6838 05:06:55,731 --> 05:06:59,067 YOUR QUESTION, YOU JUST ASKED 6839 05:06:59,067 --> 05:07:00,902 NOW, THEY ARE COMPARED IN THE 6840 05:07:00,902 --> 05:07:02,971 SPACE OF THE VECTORS THAT WE 6841 05:07:02,971 --> 05:07:05,307 TAKE TO BE THE REPRESENTATIONS 6842 05:07:05,307 --> 05:07:07,175 OF EITHER THE MODEL OR THE 6843 05:07:07,175 --> 05:07:07,376 BRAIN. 6844 05:07:07,376 --> 05:07:09,478 IN BOTH CASES WE'RE MAKING SOME 6845 05:07:09,478 --> 05:07:11,280 ASSUMPTIONS ABOUT THE NATURE OF 6846 05:07:11,280 --> 05:07:12,748 INFORMATION ENCODING, IN THE 6847 05:07:12,748 --> 05:07:14,283 MODEL OR SUBJECT WE'RE STUDYING. 6848 05:07:14,283 --> 05:07:17,119 THIS HAS BEEN A FRUITFUL WAY OF 6849 05:07:17,119 --> 05:07:18,453 STUDYING THE BRAIN, FRUITFUL WAY 6850 05:07:18,453 --> 05:07:20,656 OF BUILDING A.I. MODELS, 6851 05:07:20,656 --> 05:07:21,723 SUFFICIENTLY GROUNDED AT THIS 6852 05:07:21,723 --> 05:07:25,894 POINT IN TIME DUE TO SCIENTIFIC 6853 05:07:25,894 --> 05:07:27,195 SUCCESSES, NOT AN INAPPROPRIATE 6854 05:07:27,195 --> 05:07:28,964 ASSUMPTION NAMELY WE CAN 6855 05:07:28,964 --> 05:07:29,598 UNDERSTAND INFORMATION ENCODING 6856 05:07:29,598 --> 05:07:32,668 IN THE BRAIN AS RELATING TO A 6857 05:07:32,668 --> 05:07:34,536 VECTOR OF NEURAL ACTIVITY, LIKE- 6858 05:07:34,536 --> 05:07:38,607 WISE WE CAN UNDERSTAND INFORMATION 6859 05:07:38,607 --> 05:07:42,878 ENCODING IN ANN AS VECTOR 6860 05:07:42,878 --> 05:07:44,413 ACTIVATIONS OF EACH UNIT IN THE 6861 05:07:44,413 --> 05:07:46,815 NETWORK AND COMPARE THESE SPACES. 6862 05:07:46,815 --> 05:07:50,953 TO GET TO THE BROADER QUESTION, 6863 05:07:50,953 --> 05:07:53,155 THE IMPORTANT THING TO REMEMBER 6864 05:07:53,155 --> 05:07:54,690 METRICS SHOULD NOT BE VIEWED AS 6865 05:07:54,690 --> 05:07:57,326 A WAY OF MEASURING WHO IS THE 6866 05:07:57,326 --> 05:07:57,893 BEST SWIMMER. 6867 05:07:57,893 --> 05:07:59,494 IN FACT I THINK THAT'S THE EXACT 6868 05:07:59,494 --> 05:08:01,596 WRONG WAY TO VIEW THESE SCORES. 6869 05:08:01,596 --> 05:08:02,898 THE WAY TO UNDERSTAND THESE 6870 05:08:02,898 --> 05:08:04,866 SCORES IS THAT EACH DIFFERENT 6871 05:08:04,866 --> 05:08:06,802 METRIC PROVIDES A DIFFERENT 6872 05:08:06,802 --> 05:08:09,338 INSIGHT INTO HOW WELL THE MODEL 6873 05:08:09,338 --> 05:08:11,506 IS CAPTURING SOME ASPECT OF THE 6874 05:08:11,506 --> 05:08:11,940 BRAIN. 6875 05:08:11,940 --> 05:08:14,676 SO, WHEN YOU DO REPRESENTATIONAL 6876 05:08:14,676 --> 05:08:15,477 SIMILARITY ANALYSIS, YOU ARE 6877 05:08:15,477 --> 05:08:17,145 GETTING AT THE QUESTION OF 6878 05:08:17,145 --> 05:08:18,547 WHETHER OR NOT THE 6879 05:08:18,547 --> 05:08:20,082 REPRESENTATIONS IN THOSE SPACES 6880 05:08:20,082 --> 05:08:21,917 THAT I JUST MENTIONED HAVE 6881 05:08:21,917 --> 05:08:22,684 SIMILAR GEOMETRY. 6882 05:08:22,684 --> 05:08:23,785 IT'S A VERY SPECIFIC QUESTION 6883 05:08:23,785 --> 05:08:24,252 YOU'RE ASKING. 6884 05:08:24,252 --> 05:08:26,421 IT'S NOT IS THIS THE BEST MODEL. 6885 05:08:26,421 --> 05:08:29,257 IT IS, DOES THIS MODEL HAVE THE 6886 05:08:29,257 --> 05:08:30,559 BEST GEOMETRICAL ALIGNMENT TO 6887 05:08:30,559 --> 05:08:31,893 THAT OF THE BRAIN? 6888 05:08:31,893 --> 05:08:33,662 LIKEWISE, IF YOU HAVE BEHAVIORAL 6889 05:08:33,662 --> 05:08:35,197 SCORES, YOU ARE ASKING A 6890 05:08:35,197 --> 05:08:38,133 QUESTION, TO WHAT EXTENT CAN THE 6891 05:08:38,133 --> 05:08:40,102 MODEL REPLICATE HUMAN BEHAVIOR, 6892 05:08:40,102 --> 05:08:40,869 ANDHUMAN COMPUTATION? 6893 05:08:40,869 --> 05:08:42,838 SO, I THINK THAT THE WAY WE NEED 6894 05:08:42,838 --> 05:08:46,541 TO MOVE FORWARD ON THIS IS TO 6895 05:08:46,541 --> 05:08:49,077 HAVE AN ENTIRE LIKE BASKET OF 6896 05:08:49,077 --> 05:08:50,412 METRICS, EACH ONE PROVIDING 6897 05:08:50,412 --> 05:08:52,147 DIFFERENT INSIGHT INTO THE MODEL 6898 05:08:52,147 --> 05:08:55,650 IN ITS RELATIONSHIP TO THE 6899 05:08:55,650 --> 05:08:59,187 BRAIN, EACH HELPING US GUIDE 6900 05:08:59,187 --> 05:08:59,955 SCIENTIFIC ADVANCES. 6901 05:08:59,955 --> 05:09:03,225 YOU SHOULD NEVER VIEW AS WHO IS 6902 05:09:03,225 --> 05:09:05,193 THE FASTERS SWIMMER, BECAUSE 6903 05:09:05,193 --> 05:09:07,062 THAT WOULD KILL SCIENCE. 6904 05:09:07,062 --> 05:09:12,634 >> IN DEFENSE OF THE SWIMMER, I 6905 05:09:12,634 --> 05:09:12,901 AGREE SHOUD NOT RELY ON ONE METRIC 6906 05:09:12,901 --> 05:09:15,470 >> WE WILL GO TO THE NEXT DISCUSSANT 6907 05:09:15,470 --> 05:09:19,541 >> SHOULD I SHUT UP? 6908 05:09:19,541 --> 05:09:20,842 >> GO AHEAD. 6909 05:09:20,842 --> 05:09:24,079 >> SO THE NEXT DISCUSSANT IS 6910 05:09:24,079 --> 05:09:29,484 JOSH VOGELSTEIN FROM JOHNS 6911 05:09:29,484 --> 05:09:33,321 HOPKINS UNIVERSITY. 6912 05:09:33,321 --> 05:09:33,855 >> OKAY. 6913 05:09:33,855 --> 05:09:36,058 THANK YOU FOR INVITING ME HERE. 6914 05:09:36,058 --> 05:09:37,692 IT'S BEEN FUN LISTENING TO YOU 6915 05:09:37,692 --> 05:09:38,026 ARGUE. 6916 05:09:38,026 --> 05:09:39,795 I'M EXCITED. 6917 05:09:39,795 --> 05:09:43,732 I WAS TOLD TO ASK SOMETHING 6918 05:09:43,732 --> 05:09:44,833 ARGUMENTATIVE EXACTLY, I DON'T 6919 05:09:44,833 --> 05:09:51,106 REMEMBER EXACT THE WORDS, 6920 05:09:51,106 --> 05:09:53,508 PARTICULARLY WITH MARTIN AND 6921 05:09:53,508 --> 05:09:54,276 BLAKE. 6922 05:09:54,276 --> 05:09:56,912 I LIKE WHAT WAS SAID AND 6923 05:09:56,912 --> 05:09:58,413 FIGHTING WITH BLAKE, ANECDOTE. 6924 05:09:58,413 --> 05:10:05,654 I TAKE ISSUE. 6925 05:10:05,654 --> 05:10:06,621 HISTORICALLY CYBERNETICS, AND 6926 05:10:06,621 --> 05:10:10,192 GOFAI PEOPLE ENGINEERED STUFF THAT 6927 05:10:10,192 --> 05:10:15,096 WORKS, THOSE PEOPLE LEFT THE 6928 05:10:15,096 --> 05:10:17,866 CYBERNETICS DOGMA OF FEEDBACK 6929 05:10:17,866 --> 05:10:23,104 LOOPS AND CONTROL AND JUMPED INTO 6930 05:10:23,104 --> 05:10:23,772 COMPUTATIONAL STATISTICS, AND 6931 05:10:23,772 --> 05:10:30,345 PEOPLE NOT AT THE FIRST A.I. 6932 05:10:30,345 --> 05:10:30,879 MEETING DEVELOPED NON-PARAMETRIC 6933 05:10:30,879 --> 05:10:32,547 STATISTICS, THOSE THINGS WORKS, 6934 05:10:32,547 --> 05:10:34,049 WORK REALLY WELL, BETTER THAN 6935 05:10:34,049 --> 05:10:38,353 EITHER OF THE CONNECTIONIST 6936 05:10:38,353 --> 05:10:40,989 STUFF OR GOFAI STUFF IN PRACTICE. 6937 05:10:40,989 --> 05:10:45,594 TO DORIS POINT ABOUT TRANSFORMERS, 6938 05:10:45,594 --> 05:10:46,561 THERE'S A NADARIA-WATSON KERNEL 6939 05:10:46,561 --> 05:10:49,865 CAME OUT OF THE FIELD FROM THAT 6940 05:10:49,865 --> 05:10:50,832 IN THE '60s. 6941 05:10:50,832 --> 05:10:52,701 SO MY ARGUMENT IS ACTUALLY MAYBE 6942 05:10:52,701 --> 05:10:53,835 WHAT'S HAPPENING NOW IN NEUROAI 6943 05:10:53,835 --> 05:10:57,873 IS A BUNCH OF PEOPLE IN 6944 05:10:57,873 --> 05:10:58,406 NEUROSCIENCE REALIZING THAT 6945 05:10:58,406 --> 05:10:59,608 COMPUTATIONAL STATISTICS ARE 6946 05:10:59,608 --> 05:11:00,976 VALUABLE FOR CHARACTERIZING, 6947 05:11:00,976 --> 05:11:01,810 EXPLAINING NEUROSCIENCE LIKE 6948 05:11:01,810 --> 05:11:04,012 MANY SPEAKERS HAVE DONE, AND A 6949 05:11:04,012 --> 05:11:06,314 BUNCH OF PEOPLE IN A.I. MAKING 6950 05:11:06,314 --> 05:11:07,616 NEW MODELS AND VAGUELY HAND- 6951 05:11:07,616 --> 05:11:12,020 WAVINGLY SAYING THESE THINGS ARE 6952 05:11:12,020 --> 05:11:12,420 RELATED TO BRAINS. 6953 05:11:12,420 --> 05:11:19,227 >> I'D LIKE TO COMMENT ON THIS. 6954 05:11:19,227 --> 05:11:20,328 CERTAINLY I'M WELL WITH YOU 6955 05:11:20,328 --> 05:11:21,763 TO THE POINT, YES, LIKE 6956 05:11:21,763 --> 05:11:23,965 THERE ARE OTHER WAYS OF 6957 05:11:23,965 --> 05:11:25,200 UNDERSTANDING KIND OF THE 6958 05:11:25,200 --> 05:11:27,035 ADVANCES THAT HAVE BEEN MADE IN 6959 05:11:27,035 --> 05:11:28,203 MACHINE LEARNING THAT DON'T 6960 05:11:28,203 --> 05:11:33,074 RELATE DIRECTLY TO THE HERITAGE 6961 05:11:33,074 --> 05:11:34,376 OF CYBERNETIC AND CONNECTIONISM. 6962 05:11:34,376 --> 05:11:36,011 MAYBE THIS ALSO ADDRESSES A 6963 05:11:36,011 --> 05:11:37,546 LITTLE BIT SOMETHING MARTIN SAID 6964 05:11:37,546 --> 05:11:40,048 THAT I WANT TO ATTEND TO, WHICH 6965 05:11:40,048 --> 05:11:41,683 IS THAT IT'S WORTH RECOGNIZING 6966 05:11:41,683 --> 05:11:42,817 THE HISTORY OF A LOT OF THE 6967 05:11:42,817 --> 05:11:45,320 IDEAS IN A.I. AND HOW THEY TRACE 6968 05:11:45,320 --> 05:11:46,855 THEIR ROOTS BACK TO 6969 05:11:46,855 --> 05:11:48,056 COMPUTATIONAL NEUROSCIENCE. 6970 05:11:48,056 --> 05:11:50,258 LET'S TALK ABOUT THE 6971 05:11:50,258 --> 05:11:50,559 TRANSFORMER. 6972 05:11:50,559 --> 05:11:51,927 EVEN THE TRANSFORMER, MANY SAY 6973 05:11:51,927 --> 05:11:53,495 IT DOESN'T WORK LIKE THE BRAIN, 6974 05:11:53,495 --> 05:11:54,996 IF YOU FOLLOW CITATION TRAIL 6975 05:11:54,996 --> 05:11:56,665 THROUGH THE TRANSFORMER PAPER IT 6976 05:11:56,665 --> 05:11:58,633 COMES FROM IDEAS THAT YOSHUA 6977 05:11:58,633 --> 05:12:00,502 BENGIO HAD ABOUT SOFT ATTENTION, 6978 05:12:00,502 --> 05:12:03,071 WHICH CAME FROM IDEAS CHRISTOF 6979 05:12:03,071 --> 05:12:04,339 KOCH HAD ABOUT ATTENTION IN THE 6980 05:12:04,339 --> 05:12:04,539 BRAIN. 6981 05:12:04,539 --> 05:12:06,875 THERE'S A CLEAR HERITAGE FROM 6982 05:12:06,875 --> 05:12:07,943 NEUROSCIENCE TO TRANSFORMERS 6983 05:12:07,943 --> 05:12:09,477 THAT IS JUST THERE IN THE RECORD 6984 05:12:09,477 --> 05:12:11,446 FOR EVERYONE TO SEE. 6985 05:12:11,446 --> 05:12:12,547 MOREOVER, YOU MENTIONED THAT 6986 05:12:12,547 --> 05:12:19,821 TRANSFORMERS CAN BE ANALYZED IN 6987 05:12:19,821 --> 05:12:22,324 KERNEL METHOD, BUT THEY CAN ALSO 6988 05:12:22,324 --> 05:12:23,658 BE VIEWED AS HOPFIELD NETWORKS. 6989 05:12:23,658 --> 05:12:27,596 SOFT ATTENTION MECHANISMS OF 6990 05:12:27,596 --> 05:12:27,929 TRANSFORMERS ARE EQUIVALENT TO 6991 05:12:27,929 --> 05:12:32,601 HOPFILED NETWORKS WHICH ARE 6992 05:12:32,601 --> 05:12:33,168 DIRECTLY INSPIRED BY THE BRAIN. 6993 05:12:33,168 --> 05:12:35,804 EVEN IN A.I. BROADLY THOUGH 6994 05:12:35,804 --> 05:12:38,440 WE'VE SEEN THE FIELD PERHAPS PAY 6995 05:12:38,440 --> 05:12:39,975 LESS ATTENTION TO NEUROSCIENCE 6996 05:12:39,975 --> 05:12:42,043 MORE GENERALLY AS MORE MONEY AND 6997 05:12:42,043 --> 05:12:44,980 ENGINEERS PILE INTO THAT FIELD, 6998 05:12:44,980 --> 05:12:54,055 A LOT OF LUMINARIES IN A.I. ARE 6999 05:12:54,055 --> 05:12:56,257 COMMITTED THAT NEUROSCIENCE IS 7000 05:12:56,257 --> 05:12:56,925 A KEY SOURCE OF INSPIRATION. 7001 05:12:56,925 --> 05:13:00,395 THEY KNOW THE HISTORY OF THESE 7002 05:13:00,395 --> 05:13:01,396 IDEAS DIDN'T JUST LEARN THEM ON 7003 05:13:01,396 --> 05:13:04,265 A TUTORIAL ON GOOGLE. THEY 7004 05:13:04,265 --> 05:13:04,933 FOLLOWED THE ACADEMIC RECORD. 7005 05:13:04,933 --> 05:13:05,800 DEMIS, GEORF, RICH SUTTON, THEY 7006 05:13:05,800 --> 05:13:11,172 WILL TELL YOU, YOU GOT TO PAY 7007 05:13:11,172 --> 05:13:14,442 ATTENTION TO THE BRAIN, THERE'S 7008 05:13:14,442 --> 05:13:16,511 FUNDAMENTAL DEEP RELATIONSHIPS 7009 05:13:16,511 --> 05:13:17,746 BETWEEN AND THE COMPUTATION, 7010 05:13:17,746 --> 05:13:20,448 LIKE A TRANSFORMER. 7011 05:13:20,448 --> 05:13:30,725 I GOT CUT OFF. 7012 05:13:30,992 --> 05:13:34,362 >> DO YOU WANT TO BORROW THIS 7013 05:13:34,362 --> 05:13:36,264 ONE? 7014 05:13:36,264 --> 05:13:46,541 THIS ONE WORKS. 7015 05:13:51,713 --> 05:13:53,114 7016 05:13:53,114 --> 05:13:54,949 >> SO, FANTASTIC CONVERSATION. 7017 05:13:54,949 --> 05:13:56,484 ENGINEERING BACKGROUND TO 7018 05:13:56,484 --> 05:13:58,553 MEASURE THE BRAIN, ALSO BUILD 7019 05:13:58,553 --> 05:14:02,290 HARDWARE TO SIMULATE THE BRAIN 7020 05:14:02,290 --> 05:14:03,358 DYNAMICS. 7021 05:14:03,358 --> 05:14:09,030 MY QUESTION HAVE TWO-FOLD. 7022 05:14:09,030 --> 05:14:12,500 WE HAVE THIS BRAIN INITIATIVE, 7023 05:14:12,500 --> 05:14:14,803 COLLECT TONS OF DATA, AND REALLY 7024 05:14:14,803 --> 05:14:16,104 THE QUESTION, BECAUSE THOSE DATA 7025 05:14:16,104 --> 05:14:17,739 WHEN THEY ARE COLLECTED, THEY 7026 05:14:17,739 --> 05:14:22,544 ARE NOT LIKE -- I MEAN FOR 7027 05:14:22,544 --> 05:14:25,280 NEUROAI, THOSE INTERESTED TO 7028 05:14:25,280 --> 05:14:28,149 UNDERSTAND BRAIN FUNCTION, BRAIN 7029 05:14:28,149 --> 05:14:29,350 PHYSICAL STRUCTURE, THOSE DATA 7030 05:14:29,350 --> 05:14:32,554 ARE GOOD ENOUGH, YOU KNOW, FOR 7031 05:14:32,554 --> 05:14:35,290 US TO DO THOSE LIKE NEUROAI 7032 05:14:35,290 --> 05:14:38,159 STUDY OR SINCE WE HAVE THE TOOLS 7033 05:14:38,159 --> 05:14:40,762 WE SHOULD LIKE VERY SPECIFICALLY 7034 05:14:40,762 --> 05:14:45,266 COLLECT ANOTHER SET OF THE DATA, 7035 05:14:45,266 --> 05:14:46,267 SPECIFICALLY LIKE NEUROAI 7036 05:14:46,267 --> 05:14:49,137 AND IF WE NEED WHAT KIND OF DATA 7037 05:14:49,137 --> 05:14:53,541 YOU THINK IS MOST IMPORTANT 7038 05:14:53,541 --> 05:14:54,376 FOR YOU RIGHT NOW. 7039 05:14:54,376 --> 05:14:58,546 SECOND YES, QUESTION, ON THE OR 7040 05:14:58,546 --> 05:15:01,850 HAND, TRYING BUILD MODEL BASED 7041 05:15:01,850 --> 05:15:04,252 ON THAT DATA DO YOU THINK THE 7042 05:15:04,252 --> 05:15:06,321 INFRASTRUCTURE WE HAVE NOW DEEP 7043 05:15:06,321 --> 05:15:09,624 DURAL NETWORK IS GOOD ENOUGH, 7044 05:15:09,624 --> 05:15:12,961 NOT GOOD ENOUGH BUT A DIRECTION 7045 05:15:12,961 --> 05:15:15,630 WE SHOULD GO, DO YOU THINK THAT 7046 05:15:15,630 --> 05:15:19,701 MAKES SENSE RIGHT NOW TO 7047 05:15:19,701 --> 05:15:21,002 ACTUALLY BECAUSE NEUROSCIENTISTS 7048 05:15:21,002 --> 05:15:23,004 RIGHT NOW REALIZE HOW IMPORTANT 7049 05:15:23,004 --> 05:15:26,174 A.I. IS IMPORTANT BUT PROBABLY 7050 05:15:26,174 --> 05:15:30,011 NOT ENOUGH FOR ACTUALLY HOW 7051 05:15:30,011 --> 05:15:30,979 IMPORTANT TO DESIGN FROM 7052 05:15:30,979 --> 05:15:34,716 SCRATCH, DO YOU THINK WE SHOULD 7053 05:15:34,716 --> 05:15:38,887 LIKE BRING MORE NEUROSCIENCE TO 7054 05:15:38,887 --> 05:15:40,155 SEMICONDUCTOR MANUFACTURING OR 7055 05:15:40,155 --> 05:15:43,725 DESIGN, TO CHIP DESIGN INDUSTRY? 7056 05:15:43,725 --> 05:15:45,126 THANK YOU. 7057 05:15:45,126 --> 05:15:47,195 >> I CAN TRY TO ANSWER THE FIRST 7058 05:15:47,195 --> 05:15:48,430 PART ABOUT DATA. 7059 05:15:48,430 --> 05:15:50,498 SO, I THINK THE DATA THAT WE'VE 7060 05:15:50,498 --> 05:15:53,134 ALREADY COLLECTED IN BRAIN 7061 05:15:53,134 --> 05:15:53,835 INITIATIVE IS VALUABLE, 7062 05:15:53,835 --> 05:15:57,205 DEFINITELY MORE DATA WOULD BE 7063 05:15:57,205 --> 05:15:58,106 BETTER. 7064 05:15:58,106 --> 05:16:02,477 I LIKE THE POINT FROM THE 7065 05:16:02,477 --> 05:16:04,546 MORNING COLLECTING UNDER 7066 05:16:04,546 --> 05:16:05,780 NATURALISTIC, HIGH ENTROPY, BUT 7067 05:16:05,780 --> 05:16:08,183 I FEEL LIKE THIS NOTION IS 7068 05:16:08,183 --> 05:16:09,384 CREEPING UP, NEUROSCIENTISTS ARE 7069 05:16:09,384 --> 05:16:11,019 JUST THERE TO INSTRUMENT ANIMAL 7070 05:16:11,019 --> 05:16:12,987 BRAINS AND COLLECT DATA AND 7071 05:16:12,987 --> 05:16:17,258 WE'LL PASS TO SOME A.I. PERSON 7072 05:16:17,258 --> 05:16:18,793 WHO CAN USE THAT DATA. 7073 05:16:18,793 --> 05:16:21,763 I REBEL AGAINST THAT NOTION. 7074 05:16:21,763 --> 05:16:25,733 I WANT TO PUT IN A PITCH FOR 7075 05:16:25,733 --> 05:16:27,435 OLD-FASHIONED HYPOTHESIS 7076 05:16:27,435 --> 05:16:28,536 TESTING, STRUCTURAL INNATE 7077 05:16:28,536 --> 05:16:34,142 PRIORS, AND YOU CAN'T PUT A 7078 05:16:34,142 --> 05:16:35,677 PROBE IN EVERY MILLIMETER OF A 7079 05:16:35,677 --> 05:16:39,180 MONKEY'S BRAIN, HAVE YOU TO BE 7080 05:16:39,180 --> 05:16:40,381 GUIDED BY HYPOTHESES, WE'RE 7081 05:16:40,381 --> 05:16:42,917 TESTING WITH THE IDEAS, WHICH IS 7082 05:16:42,917 --> 05:16:47,188 SHOULD BE INCORPORATED INTO 7083 05:16:47,188 --> 05:16:48,823 THE NEXT NEURAL NETWORK, NOT 7084 05:16:48,823 --> 05:16:50,358 JUST COLLECTING DATA. 7085 05:16:50,358 --> 05:16:51,993 >> I WAS GOING TO SAY SOMETHING 7086 05:16:51,993 --> 05:16:52,227 TOO. 7087 05:16:52,227 --> 05:16:54,429 I THINK THE QUESTION OF THE 7088 05:16:54,429 --> 05:16:55,864 ARCHITECTURE IS A REALLY 7089 05:16:55,864 --> 05:16:56,397 IMPORTANT SOMETHING WHERE 7090 05:16:56,397 --> 05:16:57,832 THERE'S A LOT WE CAN LEARN FROM 7091 05:16:57,832 --> 05:17:01,636 THE DATA THAT'S ALREADY BEEN 7092 05:17:01,636 --> 05:17:01,903 COLLECTED. 7093 05:17:01,903 --> 05:17:04,272 I THINK THAT PERSONALLY I FEEL 7094 05:17:04,272 --> 05:17:06,774 LIKE ENORMOUS PROGRESS CAN BE 7095 05:17:06,774 --> 05:17:07,876 MADE BY STUDYING DIFFERENT 7096 05:17:07,876 --> 05:17:09,143 ARCHITECTURES, WHAT THEY CAN DO, 7097 05:17:09,143 --> 05:17:12,714 GETTING AWAY FROM, YOU KNOW, THE 7098 05:17:12,714 --> 05:17:16,851 STANDARD CONVOLUTIONAL NEURAL 7099 05:17:16,851 --> 05:17:21,256 NETWORK AND EXPLORING GETTING 7100 05:17:21,256 --> 05:17:24,559 INSPIRATION LIKE THE FLY 7101 05:17:24,559 --> 05:17:29,163 CONNECTOME AND MATHEMATICAL 7102 05:17:29,163 --> 05:17:31,132 EXPLORATION. 7103 05:17:31,132 --> 05:17:34,536 >> IT DEPENDS ON THE LEVEL 7104 05:17:34,536 --> 05:17:42,944 OF ABSTRACTION. WITH THE IMAGING 7105 05:17:42,944 --> 05:17:44,479 TECHNIQUES WE STILL LACK A LOT 7106 05:17:44,479 --> 05:17:46,547 OF DATA ON SUBCELLULAR LEVEL 7107 05:17:46,547 --> 05:17:48,082 WHICH PEOPLE USED TO PROBE WITH 7108 05:17:48,082 --> 05:17:49,050 PATCH ELECTRODES TEN YEARS AGO. 7109 05:17:49,050 --> 05:17:51,386 THEY DON'T DO THAT ANYMORE. 7110 05:17:51,386 --> 05:17:54,455 SO WE DON'T HAVE A LOT OF 7111 05:17:54,455 --> 05:17:55,523 SUBCELLULAR DATA. COUNTING ON 7112 05:17:55,523 --> 05:17:56,724 VOLTAGE IMAGING. MAYBE WHEN 7113 05:17:56,724 --> 05:17:58,927 THAT COMES OF AGE WE'LL HAVE A 7114 05:17:58,927 --> 05:18:01,229 LOT OF DATA TO UNDERSTAND WHAT 7115 05:18:01,229 --> 05:18:02,330 HAPPENS AT THE SUBCELLULAR LEVEL 7116 05:18:02,330 --> 05:18:05,833 WHICH IS CRITICAL BUT I ALSO 7117 05:18:05,833 --> 05:18:07,802 LIKE ANTON'S IDEA IN PREVIOUS 7118 05:18:07,802 --> 05:18:11,105 SESSION, HE SAID WE CAN USE A.I. 7119 05:18:11,105 --> 05:18:12,707 TO PREDICT WHAT THE DATA WOULD 7120 05:18:12,707 --> 05:18:14,242 LOOK LIKE, LET'S SAY FOR 7121 05:18:14,242 --> 05:18:17,178 DIFFERENT CELL TYPES, IF WE HAVE 7122 05:18:17,178 --> 05:18:18,680 DATA FOR SOME OF THE NEURONS, 7123 05:18:18,680 --> 05:18:20,148 LET'S SAY, AND GO BACK AND CLOSE 7124 05:18:20,148 --> 05:18:22,116 THE LOOP AND THAT WOULD BE 7125 05:18:22,116 --> 05:18:23,451 REALLY USEFUL TO DO, TO BRING 7126 05:18:23,451 --> 05:18:25,320 A.I. TO THE SERVICE OF 7127 05:18:25,320 --> 05:18:26,955 NEUROSCIENCE LET'S SAY TO 7128 05:18:26,955 --> 05:18:27,622 COMPLEMENT EXPERIMENTS. 7129 05:18:27,622 --> 05:18:29,357 >> SO THIS IS A GOOD TIME TO 7130 05:18:29,357 --> 05:18:32,427 INTRODUCE A QUESTION FROM ZOOM. 7131 05:18:32,427 --> 05:18:35,596 IT FOLLOWS REALLY NICELY FROM 7132 05:18:35,596 --> 05:18:37,465 WHAT WAS JUST SAID. 7133 05:18:37,465 --> 05:18:43,104 TO WHAT EXTENT DO MORE ADVANCED 7134 05:18:43,104 --> 05:18:44,072 ARCHITECTURES NECESSITATE 7135 05:18:44,072 --> 05:18:45,473 DIFFERENT APPROACHES TO MODEL 7136 05:18:45,473 --> 05:18:45,773 PERFORMANCE VALIDATION? 7137 05:18:45,773 --> 05:18:47,575 THAT IS TO WHAT EXTENTS ARE 7138 05:18:47,575 --> 05:18:49,243 APPROPRIATE A.I. MODEL 7139 05:18:49,243 --> 05:18:54,415 VALIDATION TECHNIQUES DRIVEN BY 7140 05:18:54,415 --> 05:18:58,119 THE MODEL'S ARCHITECTURE VERSUS 7141 05:18:58,119 --> 05:19:03,257 DATA STRUCTURES OR APPLICATIONS? 7142 05:19:03,257 --> 05:19:04,258 THAT'S FOR ANYONE. 7143 05:19:04,258 --> 05:19:06,394 >> I'LL MAKE ONE COMMENT, WHICH 7144 05:19:06,394 --> 05:19:08,596 IS, THIS GETS BACK TO DIFFERENT 7145 05:19:08,596 --> 05:19:08,830 METRICS. 7146 05:19:08,830 --> 05:19:12,533 INDEED IF YOUR GOAL IS TO 7147 05:19:12,533 --> 05:19:13,334 CAPTURE SOMETHING MORE 7148 05:19:13,334 --> 05:19:16,771 SPECIFICALLY RELATED TO THE 7149 05:19:16,771 --> 05:19:19,474 BIOLOGICAL IMPLEMENTATION OF 7150 05:19:19,474 --> 05:19:21,442 COMPUTATION AND INTELLIGENCE IN 7151 05:19:21,442 --> 05:19:22,310 THE BRAIN YOU NEED METRICS THAT 7152 05:19:22,310 --> 05:19:23,745 SPEAK TO THAT GOAL. 7153 05:19:23,745 --> 05:19:26,714 IF YOU'RE TRYING TO GET MORE 7154 05:19:26,714 --> 05:19:27,682 BRAIN-LIKE MODELS, BY BUILDING 7155 05:19:27,682 --> 05:19:32,520 IN SOME OF THESE BIOLOGICAL 7156 05:19:32,520 --> 05:19:33,488 DETAILS LIKE DENDRITES, GLIA, 7157 05:19:33,488 --> 05:19:34,389 WHATEVER, PART OF WHAT YOU 7158 05:19:34,389 --> 05:19:36,891 SHOULD DO IS NOT JUST LOOK AT 7159 05:19:36,891 --> 05:19:39,327 BEHAVIORAL PERFORMANCE OR 7160 05:19:39,327 --> 05:19:40,461 ALIGNMENT AND REPRESENTATIONAL 7161 05:19:40,461 --> 05:19:45,133 SPACE BUT THINGS LIKE LOW LEVEL 7162 05:19:45,133 --> 05:19:47,335 DYNAMICS OVER TIME. THE 7163 05:19:47,335 --> 05:19:48,536 TRAJECTORY OF SYNAPSES THROUGH 7164 05:19:48,536 --> 05:19:50,772 LEARNING, MAYBE EVEN THINGS I'M 7165 05:19:50,772 --> 05:19:52,473 NOT SUCCESSFULLY GENERATING ON 7166 05:19:52,473 --> 05:19:54,008 THE SPOT HERE. 7167 05:19:54,008 --> 05:19:58,513 BUT I THINK THAT IT IS KEY THAT 7168 05:19:58,513 --> 05:20:00,048 YOU ASK THE QUESTION - FOR A 7169 05:20:00,048 --> 05:20:03,651 QUESTION THAT YOU WANT TO ASK 7170 05:20:03,651 --> 05:20:06,154 RELATED TO ARE WE MODELING THE 7171 05:20:06,154 --> 05:20:07,355 BRAIN SUCCESSFULLY WITH THIS MORE 7172 05:20:07,355 --> 05:20:10,124 DETAILED MODEL YOU NEED THE RIGHT 7173 05:20:10,124 --> 05:20:10,358 METRICS FOR THAT. 7174 05:20:10,358 --> 05:20:12,860 >> I HAVE DISCUSSION OF DETAILED 7175 05:20:12,860 --> 05:20:13,094 METRICS. 7176 05:20:13,094 --> 05:20:14,062 THIS RELATES TO THE QUESTION 7177 05:20:14,062 --> 05:20:15,129 WE DISCUSSED IN THE MORNING YOU 7178 05:20:15,129 --> 05:20:17,999 WHO DO WE KNOW WHEN WE'RE DONE? 7179 05:20:17,999 --> 05:20:20,268 I DON'T THINK IT'S WHEN A METRIC 7180 05:20:20,268 --> 05:20:21,703 PASSES A THRESHOLD. 7181 05:20:21,703 --> 05:20:23,237 I THINK IT'S WHEN WE CAN BUILD 7182 05:20:23,237 --> 05:20:25,239 BASED ON OUR KNOWLEDGE LIKE SOME 7183 05:20:25,239 --> 05:20:26,207 AMAZING ARTIFICIAL INTELLIGENCE 7184 05:20:26,207 --> 05:20:28,409 THAT CAN DO THINGS THAT WE JUST 7185 05:20:28,409 --> 05:20:30,812 WOULD NOT HAVE FATHOMED BEFORE, 7186 05:20:30,812 --> 05:20:31,012 RIGHT? 7187 05:20:31,012 --> 05:20:34,482 LIKE THAT'S GOING TO BE THE 7188 05:20:34,482 --> 05:20:34,849 METRIC. 7189 05:20:34,849 --> 05:20:37,151 >> HOW DO YOU GET THERE? 7190 05:20:37,151 --> 05:20:38,352 >> THIS ALSO GETS TO THE 7191 05:20:38,352 --> 05:20:40,455 QUESTION I WAS THINKING ABOUT, 7192 05:20:40,455 --> 05:20:43,624 HOW CAN THESE METRICS TAKE INTO 7193 05:20:43,624 --> 05:20:44,492 ACCOUNT MULTIPLE LAYERS, AND 7194 05:20:44,492 --> 05:20:45,793 CONNECTIONS BETWEEN THOSE 7195 05:20:45,793 --> 05:20:50,131 MULTIPLE LAYERS WHEN LOOKING AT 7196 05:20:50,131 --> 05:20:56,504 MODEL-TO-BRAIN SIMILARITIES. 7197 05:20:56,504 --> 05:20:57,905 >> TECHNICALLY YOU LOOK ACROSS 7198 05:20:57,905 --> 05:20:59,674 LAYERS, FIND THOSE THAT DO BEST, 7199 05:20:59,674 --> 05:21:02,276 GET INFERENCES WHY THE LATEST 7200 05:21:02,276 --> 05:21:04,479 MIGHT DO BEST AND VALIDATE 7201 05:21:04,479 --> 05:21:05,580 INDEPENDENT DATA. 7202 05:21:05,580 --> 05:21:06,447 BUT THERE'S INTERESTING 7203 05:21:06,447 --> 05:21:08,850 QUESTIONS ABOUT WHAT KIND OF 7204 05:21:08,850 --> 05:21:10,618 INFORMATION CONTAINED IN CROSS 7205 05:21:10,618 --> 05:21:12,587 LAYERS, CAN I COME BACK TO A 7206 05:21:12,587 --> 05:21:14,122 POINT ABOUT MODEL COMPARISONS? 7207 05:21:14,122 --> 05:21:17,525 THE THING I WANTED TO ADD, 7208 05:21:17,525 --> 05:21:19,160 ANDREAS MENTIONED EARLY IN THE 7209 05:21:19,160 --> 05:21:20,895 TALK THIS MORNING THAT THERE'S 7210 05:21:20,895 --> 05:21:23,197 ALSO THIS PHENOMENON THAT'S 7211 05:21:23,197 --> 05:21:25,066 HAPPENING ACROSS SEVERAL DOMAINS 7212 05:21:25,066 --> 05:21:26,367 WHERE REPRESENTATIONS ACROSS 7213 05:21:26,367 --> 05:21:27,969 HIGH PERFORMING MODELS SEEM TO 7214 05:21:27,969 --> 05:21:32,607 BE CONVERGING ON THE SAME AXIS. 7215 05:21:32,607 --> 05:21:35,576 THIS SEEMS REALLY IMPORTANT. 7216 05:21:35,576 --> 05:21:39,180 HOSSEINI HAS A POSTER ON THIS. 7217 05:21:39,180 --> 05:21:43,784 IF YOU WANT TO HEAR MORE IN THE 7218 05:21:43,784 --> 05:21:45,119 CONTEXT OF LANGUAGE AND VISION, 7219 05:21:45,119 --> 05:21:46,554 NO MATTER WHAT GIVEN THIS PROBLEM 7220 05:21:46,554 --> 05:21:47,955 THAT YOU'RE SOLVING YOU'RE 7221 05:21:47,955 --> 05:21:48,623 ALIGNING ONTO THE SAME SPACE. 7222 05:21:48,623 --> 05:21:49,624 THIS TELLS YOU SOMETHING REALLY 7223 05:21:49,624 --> 05:21:51,292 DEEP ABOUT WHAT IS NEEDED TO 7224 05:21:51,292 --> 05:21:53,694 SOLVE A PARTICULAR PROBLEM 7225 05:21:53,694 --> 05:21:58,199 SUCCESSFULLY. 7226 05:21:58,199 --> 05:22:01,035 >> THIS IS SOMETHING -- 7227 05:22:01,035 --> 05:22:03,638 THERE'S A THEORY RESULT THAT 7228 05:22:03,638 --> 05:22:04,839 TRANSFORMERS AND LSTMS ARE 7229 05:22:04,839 --> 05:22:06,140 EQUIVALENT IN REPRESENTATION IN 7230 05:22:06,140 --> 05:22:06,941 COMPARABLE TRAINING CONDITIONS. 7231 05:22:06,941 --> 05:22:15,917 COMES BACK TO HOW WE SHOULD 7232 05:22:15,917 --> 05:22:21,856 COLLECT, DO HYPOTHESIS TESTING. 7233 05:22:21,856 --> 05:22:22,757 BUT HYPOTHESES NEED TO BE ON THE LEVEL OF HOW THE WHOLE REPRESENTATIONAL SPACE HAS BEEN OPTIMIZED 7234 05:22:22,757 --> 05:22:22,957 IN THE FIRST PLACE, WHICH NECESSITATES MODELS WHICH 7235 05:22:22,957 --> 05:22:26,260 NECESSITATES US TO USE THE TOP 7236 05:22:26,260 --> 05:22:30,298 SWIMMERS, PIT THEM AGAINST ONE 7237 05:22:30,298 --> 05:22:34,135 ANOTHER, AND FOR THAT WE NEED TO 7238 05:22:34,135 --> 05:22:36,470 MAKE PROGRESS ON METRICS AND 7239 05:22:36,470 --> 05:22:39,207 PROBABLY WE NEED MANY OF THOSE, 7240 05:22:39,207 --> 05:22:41,275 PROBABLY NEED A MODEL, MOST 7241 05:22:41,275 --> 05:22:43,578 MODELS WILL NOT DO WELL IN MANY 7242 05:22:43,578 --> 05:22:46,881 OF THOSE, WE NEED TO PUT THOSE 7243 05:22:46,881 --> 05:22:49,383 REFERENCES UP BECAUSE OTHERWISE 7244 05:22:49,383 --> 05:22:51,052 I DON'T SEE US MAKING PROGRESS. 7245 05:22:51,052 --> 05:22:52,553 WE'LL SEE WHEN WE GET THERE BUT 7246 05:22:52,553 --> 05:22:55,423 THE GETTING THERE PART IS GOING 7247 05:22:55,423 --> 05:22:58,492 TO BE DIFFICULT-- IF WE DON'T HAVE 7248 05:22:58,492 --> 05:23:00,494 NUMERICAL GOALPOSTS TO TRACE 7249 05:23:00,494 --> 05:23:01,262 DOWN WITH. 7250 05:23:01,262 --> 05:23:03,464 >> WE HAVE FIVE MINUTES LEFT. 7251 05:23:03,464 --> 05:23:04,999 I'D LIKE TO OPEN UP TO Q&A BOTH 7252 05:23:04,999 --> 05:23:09,370 IN THE ROOM AND QUESTIONS FROM 7253 05:23:09,370 --> 05:23:13,241 THE VIRTUAL AUDIENCE. 7254 05:23:13,241 --> 05:23:13,574 >> HELLO. 7255 05:23:13,574 --> 05:23:16,310 I THANK YOU FOR YOUR WONDERFUL 7256 05:23:16,310 --> 05:23:17,945 PRESENTATIONS. MY NAME IS XINHE 7257 05:23:17,945 --> 05:23:21,782 ZHANG. I'M A GRADUATE STUDENT 7258 05:23:21,782 --> 05:23:22,984 FROM HARVARD 7259 05:23:22,984 --> 05:23:23,851 UNIVERSITY. WE'VE TALKED ABOUT 7260 05:23:23,851 --> 05:23:27,388 STRUCTURALLY HOW CAN WE IMPLEMENT 7261 05:23:27,388 --> 05:23:31,959 BRAIN-LIKE MODULES OR LAYERS INTO 7262 05:23:31,959 --> 05:23:33,494 THE NETWORK BUT BIOLOGICAL 7263 05:23:33,494 --> 05:23:34,662 LEARNING RULES ARE LACKING IN 7264 05:23:34,662 --> 05:23:38,099 THAT SENSE. FOR EXAMPLE, HEBBIAN 7265 05:23:38,099 --> 05:23:41,502 LEARNING RULES AREN'T PERFORMING AS 7266 05:23:41,502 --> 05:23:44,705 WELL AS HARDENDED-AND-TRUE 7267 05:23:44,705 --> 05:23:45,139 BACKPROPAGATION. 7268 05:23:45,139 --> 05:23:47,441 WHAT TECHNOLOGICAL ADVANCEMENTS 7269 05:23:47,441 --> 05:23:49,176 HARDWARE AND SOFTWARE-WISE CAN 7270 05:23:49,176 --> 05:23:52,446 HELP WITH THE DISCOVERIES OF 7271 05:23:52,446 --> 05:23:54,882 MORE BIOLOGICALLY EFFICIENT AND 7272 05:23:54,882 --> 05:23:55,917 PLAUSIBLE LEARNING RULES, AND 7273 05:23:55,917 --> 05:23:59,720 SECOND, DO YOU THINK THERE'S A 7274 05:23:59,720 --> 05:24:01,155 GLOBAL LEARNING RULE GOVERNING ALL 7275 05:24:01,155 --> 05:24:03,691 REGIONS OR DO YOU THINK THERE'S 7276 05:24:03,691 --> 05:24:07,094 SOME INSTRUMENTAL EFFECT OF 7277 05:24:07,094 --> 05:24:07,628 INDIVIDUAL BRAIN REGIONS, AND 7278 05:24:07,628 --> 05:24:08,996 COLLECTIVELY THEY LEARN IN THE 7279 05:24:08,996 --> 05:24:14,101 SHORT TERM AND LONG TERM, SOME 7280 05:24:14,101 --> 05:24:14,735 BEHAVIORS OR TASKS? 7281 05:24:14,735 --> 05:24:15,636 THANK YOU. 7282 05:24:15,636 --> 05:24:18,339 >> MAYBE I CAN START WITH THAT. 7283 05:24:18,339 --> 05:24:20,007 EXCELLENT QUESTION. 7284 05:24:20,007 --> 05:24:20,975 INDEED BIOLOGICAL LEARNING RULES 7285 05:24:20,975 --> 05:24:25,146 RIGHT NOW ARE NOT DOING AS WELL 7286 05:24:25,146 --> 05:24:27,648 AS GRADIENT DESCENT BUT MAYBE THAT'S 7287 05:24:27,648 --> 05:24:29,417 THE WRONG QUESTION TO ASK. 7288 05:24:29,417 --> 05:24:31,485 AS WE NOTICE FROM THE 7289 05:24:31,485 --> 05:24:33,454 PRESENTATION EARLIER TODAY BY 7290 05:24:33,454 --> 05:24:35,990 TONY, FOR EXAMPLE, HUMANS AND 7291 05:24:35,990 --> 05:24:37,491 ANIMALS LEARN AT DIFFERENT 7292 05:24:37,491 --> 05:24:38,259 SCALES. WE HAVE 7293 05:24:38,259 --> 05:24:40,127 EVOLUTIONARY SCALE, DEVELOPMENTAL 7294 05:24:40,127 --> 05:24:41,529 SCALE, AND THEN LEARNING WHICH 7295 05:24:41,529 --> 05:24:46,367 IS THE CLOSEST TO WHAT 7296 05:24:46,367 --> 05:24:49,670 ARTIFICIAL NEURONS ARE USING. 7297 05:24:49,670 --> 05:24:51,005 GRADIENT DESCENT 7298 05:24:51,005 --> 05:24:54,742 IS IDEAL FOR THE LAST STAGE 7299 05:24:54,742 --> 05:24:55,910 BUT WHAT BIOLOGICAL LEARNING 7300 05:24:55,910 --> 05:25:00,114 RULES ARE DOING IS OPERATING AT 7301 05:25:00,114 --> 05:25:00,981 DIFFERENT OR MULTIPLE SCALES AT 7302 05:25:00,981 --> 05:25:02,283 THE SAME TIME AND WE DON'T HAVE 7303 05:25:02,283 --> 05:25:03,250 A CLEAR UNDERSTAND HOW THAT WORKS 7304 05:25:03,250 --> 05:25:04,919 IN THE BRAIN RIGHT NOW AND 7305 05:25:04,919 --> 05:25:07,088 I THINK THAT IS WHAT WE'RE 7306 05:25:07,088 --> 05:25:11,058 MISSING IN ORDER TO TRANSFER 7307 05:25:11,058 --> 05:25:12,593 THIS KNOWLEDGE TO CLASSICAL A.I. 7308 05:25:12,593 --> 05:25:14,028 SO THAT'S ONE ANSWER. 7309 05:25:14,028 --> 05:25:17,498 I DON'T THINK IT'S A 7310 05:25:17,498 --> 05:25:18,299 TECHNOLOGICAL PROBLEM. 7311 05:25:18,299 --> 05:25:20,468 MAYBE IF WE HAVE BETTER BETTER 7312 05:25:20,468 --> 05:25:22,336 HARDWARE TO MIMIC HOW THE BRAIN 7313 05:25:22,336 --> 05:25:23,971 IS USING THE LEARNING RULES, 7314 05:25:23,971 --> 05:25:26,374 THAT WOULD PROVIDE NEW INSIGHTS, 7315 05:25:26,374 --> 05:25:28,142 BUT I THINK IT'S MOSTLY A 7316 05:25:28,142 --> 05:25:29,243 CONCEPTUAL PROBLEM AT THIS 7317 05:25:29,243 --> 05:25:29,677 MOMENT. 7318 05:25:29,677 --> 05:25:31,946 WE NEED TO LOOK AT RULES, 7319 05:25:31,946 --> 05:25:34,048 BIOLOGICAL ONES, FROM A 7320 05:25:34,048 --> 05:25:34,815 DIFFERENT PERSPECTIVE. 7321 05:25:34,815 --> 05:25:37,418 AND THERE ARE STUDIES RIGHT NOW 7322 05:25:37,418 --> 05:25:39,053 LOOKING AT SELF-SUPERVISED 7323 05:25:39,053 --> 05:25:42,223 LEARNING IF YOU LIKE, OR 7324 05:25:42,223 --> 05:25:44,225 STATISTICAL LEARNING FOR WHICH 7325 05:25:44,225 --> 05:25:45,326 THESE HEBBIAN-LIKE RULES ARE BEST. 7326 05:25:45,326 --> 05:25:48,162 I THINK WE STILL HAVE A LONG WAY 7327 05:25:48,162 --> 05:25:53,067 TO GO UNTIL WE SEE THEM 7328 05:25:53,067 --> 05:25:54,268 IMPLEMENTED SUCCESSFULLY IN 7329 05:25:54,268 --> 05:25:56,237 NEUROAI SYSTEMS BUT WE 7330 05:25:56,237 --> 05:25:57,405 UNDERSTAND WHY THEY DON'T WORK 7331 05:25:57,405 --> 05:26:02,243 WHEN USING THEM IN THE CONCEPT 7332 05:26:02,243 --> 05:26:04,211 OF PERFORMANCE ASSESSMENT. 7333 05:26:04,211 --> 05:26:07,615 >> TO JUMP IN, TO ACTUALLY ECHO 7334 05:26:07,615 --> 05:26:09,550 THE COMMENT I HEARD BLAKE MADE, 7335 05:26:09,550 --> 05:26:13,521 WHICH IS THAT IF YOU LOOK AT THE 7336 05:26:13,521 --> 05:26:14,655 THE RESEARCH IN NEUROSCIENCE 7337 05:26:14,655 --> 05:26:18,726 OVER THE LAST 20 YEARS, 10 YEARS 7338 05:26:18,726 --> 05:26:20,661 ESPECIALLY, ALMOST ALL OF IT IS 7339 05:26:20,661 --> 05:26:25,733 BASED ON GRADIENT DESCENT. 7340 05:26:25,733 --> 05:26:33,040 NOW IT'S RECURRENT NETWORKS OF 7341 05:26:33,040 --> 05:26:33,574 GRADIENT DESCENT. 7342 05:26:33,574 --> 05:26:35,543 BUT THERE MAY BE A REASON FOR 7343 05:26:35,543 --> 05:26:35,876 THAT. 7344 05:26:35,876 --> 05:26:38,512 I THINK THE BRAIN IS USING 7345 05:26:38,512 --> 05:26:40,681 GRADIENT DESCENT, MAYBE NOT BACK 7346 05:26:40,681 --> 05:26:42,883 PROP BUT IT IS TAKING 7347 05:26:42,883 --> 05:26:45,820 ADVANTAGE OF ERROR SIGNALS, 7348 05:26:45,820 --> 05:26:46,520 RIGHT? 7349 05:26:46,520 --> 05:26:52,193 AND THIS IS RELATED I THINK TO 7350 05:26:52,193 --> 05:26:54,161 THE FACT THAT IT'S REALLY 7351 05:26:54,161 --> 05:26:57,131 DYNAMICS THAT WE'RE TRYING TO 7352 05:26:57,131 --> 05:27:00,167 UNDERSTAND, THE BRAIN DYNAMICS, 7353 05:27:00,167 --> 05:27:03,137 AND THAT INVOLVES FEEDBACK 7354 05:27:03,137 --> 05:27:10,144 LOOPS, RECURRENCE, LAYER 2-3 IN 7355 05:27:10,144 --> 05:27:10,911 THE CORTEX, CA3 IN HIPPOCAMPUS. 7356 05:27:10,911 --> 05:27:13,047 CARINA MADE A BEAUTIFUL MODEL, 7357 05:27:13,047 --> 05:27:15,416 QUITE REMARKABLE, IF YOU DIDN'T 7358 05:27:15,416 --> 05:27:17,718 APPRECIATE IT, YOU MAY NOT HAVE 7359 05:27:17,718 --> 05:27:19,787 APPRECIATED IT, HOW REMARKABLE 7360 05:27:19,787 --> 05:27:23,057 IT IS, RELATIVELY SMALL MODEL, 7361 05:27:23,057 --> 05:27:25,693 YOU CAN EMBED FOUR DIFFERENT 7362 05:27:25,693 --> 05:27:27,128 DYNAMICAL PATTERNS OF ACTIVITY 7363 05:27:27,128 --> 05:27:30,831 AND I THINK THIS IS GOING TO BE 7364 05:27:30,831 --> 05:27:35,202 A GENERAL INSIGHT INTO HOW 7365 05:27:35,202 --> 05:27:37,505 BRAINS ARE ORGANIZED, THAT IS TO 7366 05:27:37,505 --> 05:27:39,273 SAY THEY ARE LOW DIMENSIONAL. 7367 05:27:39,273 --> 05:27:45,813 IN FACT, ONE 7368 05:27:45,813 --> 05:27:47,915 DIMENSIONAL. 7369 05:27:47,915 --> 05:27:50,084 FOUR DIFFERENT ONE-DIMENSIONAL 7370 05:27:50,084 --> 05:27:50,851 ATTRACTORS IN THAT NETWORK. 7371 05:27:50,851 --> 05:27:54,788 THERE'S A LOT OF EVIDENCE NOW 7372 05:27:54,788 --> 05:27:56,557 THAT FROM COMPLEX COGNITIVE 7373 05:27:56,557 --> 05:28:00,060 TASKS DONE IN RODENTS THAT THE 7374 05:28:00,060 --> 05:28:01,061 DIMENSIONALITY OF THE MANIFOLDS 7375 05:28:01,061 --> 05:28:04,398 THAT IS USED TO SOLVE THE TASK 7376 05:28:04,398 --> 05:28:06,066 IS LOW DIMENSIONAL, ON THE ORDER 7377 05:28:06,066 --> 05:28:07,835 OF FOUR TO SIX. 7378 05:28:07,835 --> 05:28:10,137 NOW, IF THAT'S TRUE WE'RE NOT 7379 05:28:10,137 --> 05:28:11,872 DEALING WITH NUMBERS 7380 05:28:11,872 --> 05:28:14,174 LIKE TEN TO THE TENTH ANYMORE. 7381 05:28:14,174 --> 05:28:15,943 WE'RE DEALING WITH LOW 7382 05:28:15,943 --> 05:28:16,710 DIMENSIONAL MANIFOLDS EMBEDDED 7383 05:28:16,710 --> 05:28:18,045 IN A LARGE NUMBER OF NEURONS, 7384 05:28:18,045 --> 05:28:19,747 BUT I THINK THAT MAKES IT MORE 7385 05:28:19,747 --> 05:28:22,049 TRACTABLE AND I THINK THAT MAYBE 7386 05:28:22,049 --> 05:28:25,986 THIS IS THE WAY TO GO, RATHER 7387 05:28:25,986 --> 05:28:28,155 THAN TRYING TO COMPARE THINGS TO 7388 05:28:28,155 --> 05:28:29,857 FEEDFORWARD NETWORKS. 7389 05:28:29,857 --> 05:28:34,328 VISION IS NOT IN OUR BRAINS 7390 05:28:34,328 --> 05:28:35,763 IT'S NOT FEEDFORWARD, IT'S AN 7391 05:28:35,763 --> 05:28:36,463 INCREDIBLE AMOUNT OF FEEDBACK. 7392 05:28:36,463 --> 05:28:38,365 >> I WANT TO FOLLOW THAT UP WITH 7393 05:28:38,365 --> 05:28:39,500 THEY'VE ARE ALREADY 7394 05:28:39,500 --> 05:28:42,236 SAID WHAT I WOULD HAVE SAID, BUT 7395 05:28:42,236 --> 05:28:43,771 THERE IS ONE TECHNOLOGICAL 7396 05:28:43,771 --> 05:28:44,905 INNOVATION I WANT TO SAY HERE AT 7397 05:28:44,905 --> 05:28:47,041 THE NIH WE NEED TO REALLY NAIL 7398 05:28:47,041 --> 05:28:47,508 THIS QUESTION. 7399 05:28:47,508 --> 05:28:50,311 WE NEED TO BE ABLE TO TRACK 7400 05:28:50,311 --> 05:28:52,813 SYNAPTIC WEIGHTS IN LEARNING 7401 05:28:52,813 --> 05:28:53,247 ANIMALS. 7402 05:28:53,247 --> 05:28:55,316 HONESTLY, IF I COULD GET ANY ONE 7403 05:28:55,316 --> 05:28:57,318 TOOL IN THE NEXT TEN YEARS TO 7404 05:28:57,318 --> 05:28:59,954 SOLVE THIS PROBLEM, IT WOULD BE 7405 05:28:59,954 --> 05:29:02,590 THAT. 7406 05:29:02,590 --> 05:29:04,858 >> SO WE'RE PRETTY MUCH OUT OF 7407 05:29:04,858 --> 05:29:05,059 TIME. 7408 05:29:05,059 --> 05:29:06,393 I THOUGHT IT WOULD BE A NICE 7409 05:29:06,393 --> 05:29:07,728 IDEA TO TAKE ONE OF THE 7410 05:29:07,728 --> 05:29:10,564 QUESTIONS WE DIDN'T GET TO AND 7411 05:29:10,564 --> 05:29:12,967 ACTUALLY TAKE IT FOR THE ROAD. 7412 05:29:12,967 --> 05:29:15,736 SO Y'ALL CAN DISCUSS DURING THE 7413 05:29:15,736 --> 05:29:17,238 COFFEE BREAK, BATHROOM BREAK, SO 7414 05:29:17,238 --> 05:29:19,440 THE QUESTION IS MULTIPLE 7415 05:29:19,440 --> 05:29:21,842 TEMPORAL SCALES FOR LEARNING AND 7416 05:29:21,842 --> 05:29:24,044 INFERENCE SEEM TO BE MISSING IN 7417 05:29:24,044 --> 05:29:26,981 MOST CURRENT MODELS AND MANY 7418 05:29:26,981 --> 05:29:27,748 NEUROSCIENCE APPROACHES. 7419 05:29:27,748 --> 05:29:29,416 DO YOU SEE THAT AS SOMETHING 7420 05:29:29,416 --> 05:29:30,484 THAT IS CRITICAL TO 7421 05:29:30,484 --> 05:29:32,786 UNDERSTANDING THE BRAIN AND BY 7422 05:29:32,786 --> 05:29:34,088 CORRESPONDENCE IN KEEPING WITH 7423 05:29:34,088 --> 05:29:34,989 THE THOUGHT PROCESSES OF THE 7424 05:29:34,989 --> 05:29:37,691 PANEL HERE WHAT IS NEEDED TO 7425 05:29:37,691 --> 05:29:40,127 MOVE MODELS IN NEUROSCIENCE 7426 05:29:40,127 --> 05:29:41,328 DATASETS TOWARDS ADDRESSING 7427 05:29:41,328 --> 05:29:43,964 MULTIPLE TEMPORAL SCALES OF 7428 05:29:43,964 --> 05:29:45,833 ECOLOGICAL BEHAVIOR? 7429 05:29:45,833 --> 05:29:49,770 >> THANK YOU SO MUCH, ALL 7430 05:29:49,770 --> 05:29:51,405 SPEAKERS AND DISCUSSANTS AND THE 7431 05:29:51,405 --> 05:29:53,307 AUDIENCE FOR YOUR PARTICIPATION AS WELL. 7432 05:30:02,316 --> 05:30:04,885 >> THE FIRST HALF AN HOUR WILL 7433 05:30:04,885 --> 05:30:06,520 GO TOWARDS ANSWERING UNRESOLVED 7434 05:30:06,520 --> 05:30:10,624 QUESTIONS FROM SESSIONS 1 AND 2. 7435 05:30:10,624 --> 05:30:14,428 STARTING AT 5:05 WE'LL LOOK 7436 05:30:14,428 --> 05:30:22,703 FORWARD INTO THE FUTURE, AND 7437 05:30:22,703 --> 05:30:33,113 ANDRES WILL HAVE A TALK. 7438 05:30:33,414 --> 05:30:35,182 I'M PLEASED TO INTRODUCE STEVEN 7439 05:30:35,182 --> 05:30:38,786 ZUCKER, YOU'VE ALREADY HEARD 7440 05:30:38,786 --> 05:30:45,592 FROM HIM TODAY, HE'S A PROFESSOR 7441 05:30:45,592 --> 05:30:51,165 OF COMPUTER SCIENCE, BIOMEDICAL 7442 05:30:51,165 --> 05:30:57,237 ENGINEERING YALE UNIVERSITY. 7443 05:30:57,237 --> 05:30:59,473 HE CO-FOUNDED WITH TERRENCE 7444 05:30:59,473 --> 05:31:02,743 SEJNOWSKI THE WOODS HOLE WORKSHOP 7445 05:31:02,743 --> 05:31:03,711 ON COMPUTATIONAL NEUROSCIENCE IN 7446 05:31:03,711 --> 05:31:08,315 1980S BLOOMING INTO A NOW. HE 7447 05:31:08,315 --> 05:31:15,122 STUDIED ROLE OF VISUAL SYSTEM, 7448 05:31:15,122 --> 05:31:17,024 ITS CONNECTION TO GEOMETRY. 7449 05:31:17,024 --> 05:31:25,999 WE'RE ALSO DELIGHTED TO HAVE AS 7450 05:31:25,999 --> 05:31:26,767 CO-MODERATOR KANAKA RAJAN HARVARD 7451 05:31:26,767 --> 05:31:27,735 UNIVERSITY, FOUNDING MEMBER OF 7452 05:31:27,735 --> 05:31:35,309 KEMPNER INSTITUTE FOR NATURAL 7453 05:31:35,309 --> 05:31:45,152 AND A.I. AT HARVARD. 7454 05:31:45,152 --> 05:31:48,655 >> WELCOME BACK, EVERYONE, TO 7455 05:31:48,655 --> 05:31:51,959 GET THE AFTERNOON SESSION GOING 7456 05:31:51,959 --> 05:31:54,928 I THOUGHT WE SHOULD ASK DORIS 7457 05:31:54,928 --> 05:31:59,600 WHAT SHE THOUGHT OF ANDREAS'S 7458 05:31:59,600 --> 05:32:01,135 TALK THIS MORNING. 7459 05:32:01,135 --> 05:32:03,637 >> IT WAS AMAZING. 7460 05:32:03,637 --> 05:32:09,042 POINTING US TO THE NEW 7461 05:32:09,042 --> 05:32:10,043 FUTURE, BUILDING DIGITAL TWINS, 7462 05:32:10,043 --> 05:32:12,012 THE DREAM THAT WE DON'T HAVE 7463 05:32:12,012 --> 05:32:13,447 TO -- AS A MONKEY PHYSIOLOGIST, 7464 05:32:13,447 --> 05:32:16,016 I KNOW HOW HARD IT IS TO DO 7465 05:32:16,016 --> 05:32:17,818 THESE EXPERIMENTS, IT WOULD BE 7466 05:32:17,818 --> 05:32:20,320 INCREDIBLE IF I COULD DOWNLOAD 7467 05:32:20,320 --> 05:32:23,257 THE MODEL AND JUST, YOU KNOW, 7468 05:32:23,257 --> 05:32:24,258 THINK. 7469 05:32:24,258 --> 05:32:25,893 BUT AS WE DISCUSSED TOGETHER IN 7470 05:32:25,893 --> 05:32:27,461 THE BREAK, THAT'S NOT THE 7471 05:32:27,461 --> 05:32:28,095 ENDPOINT, RIGHT? 7472 05:32:28,095 --> 05:32:29,630 I HOPE THAT'S CLEAR. 7473 05:32:29,630 --> 05:32:31,598 ONCE YOU HAVE THE DIGITAL TWIN, 7474 05:32:31,598 --> 05:32:33,033 THAT'S JUST THE BEGINNING. 7475 05:32:33,033 --> 05:32:35,235 ULTIMATELY WE WANT TO UNDERSTAND 7476 05:32:35,235 --> 05:32:36,537 THE PRINCIPLE, THAT'S WHAT IT 7477 05:32:36,537 --> 05:32:38,739 MEANS TO UNDERSTAND. 7478 05:32:38,739 --> 05:32:41,375 SOME PEOPLE THINK THAT 7479 05:32:41,375 --> 05:32:41,975 UNDERSTANDING IS NOT POSSIBLE, 7480 05:32:41,975 --> 05:32:44,411 IF YOU JUST HAVE A NETWORK AND 7481 05:32:44,411 --> 05:32:46,613 THAT'S IT, YOU CAN'T EXPLAIN IT 7482 05:32:46,613 --> 05:32:50,651 ANY FURTHER, EXCEPT IN TERMS OF 7483 05:32:50,651 --> 05:32:51,785 ARCHITECTURE AND LEARNING RULE. 7484 05:32:51,785 --> 05:32:53,187 I BELIEVE WE CAN UNDERSTAND THE 7485 05:32:53,187 --> 05:32:56,590 BRAIN THE WAY WE UNDERSTAND 7486 05:32:56,590 --> 05:32:59,193 EVERYTHING ELSE, AND, YEAH, I 7487 05:32:59,193 --> 05:33:01,161 THINK THAT ANDREAS IS GOING TO 7488 05:33:01,161 --> 05:33:07,334 MAKE A GOOD CONTRIBUTION TO THAT. 7489 05:33:07,334 --> 05:33:09,403 >> YEAH, NO, I FULLY AGREE WITH 7490 05:33:09,403 --> 05:33:10,904 DORIS. 7491 05:33:10,904 --> 05:33:13,240 I THINK THERE'S AN 7492 05:33:13,240 --> 05:33:14,141 INTERESTING -- INTERESTING TO 7493 05:33:14,141 --> 05:33:17,444 HEAR BLAKE ON THIS, WHEN YOU 7494 05:33:17,444 --> 05:33:27,721 LOOK AT THE PARALLEL DISTRIBUTED 7495 05:33:27,721 --> 05:33:28,922 PROCESSING BOOKS, IT'S 7496 05:33:28,922 --> 05:33:29,690 PHILOSOPHY. CAN WE HUMANS 7497 05:33:29,690 --> 05:33:30,457 UNDERSTAND AND INTERPRET THE 7498 05:33:30,457 --> 05:33:33,961 BRAIN IN A WAY WE CAN DESCRIBE 7499 05:33:33,961 --> 05:33:35,596 IT WITH BEAUTIFUL WORDS AND SAY, 7500 05:33:35,596 --> 05:33:37,231 YOU KNOW, WHEN I DO THIS OR I 7501 05:33:37,231 --> 05:33:39,433 SEE THIS, THIS IS HOW IT WORKS. 7502 05:33:39,433 --> 05:33:41,835 AND I AGREE WITH DORIS HERE. 7503 05:33:41,835 --> 05:33:44,671 I DON'T THINK I WOULD HAVE BEEN 7504 05:33:44,671 --> 05:33:46,306 A NEUROSCIENTIST IF I DIDN'T 7505 05:33:46,306 --> 05:33:47,641 BELIEVE THERE WERE PRINCIPLES TO 7506 05:33:47,641 --> 05:33:49,610 BE UNDERSTOOD, THAT CAN BE 7507 05:33:49,610 --> 05:33:53,547 DESCRIBED WITH, IN 7508 05:33:53,547 --> 05:33:55,849 CLASSICAL SCIENCE, DOWN TO 7509 05:33:55,849 --> 05:33:57,084 BEAUTY, SYMMETRY, MATH. 7510 05:33:57,084 --> 05:34:00,687 I DO THINK THAT FOR ME BUILDING 7511 05:34:00,687 --> 05:34:03,757 THESE MODELS AND MORE THINKING 7512 05:34:03,757 --> 05:34:04,658 ABOUT DATA DRIVEN MODELS, NOT 7513 05:34:04,658 --> 05:34:07,594 MORE LIKE THE OTHER KINDS OF 7514 05:34:07,594 --> 05:34:10,230 MODELS THAT YOU PRE-TRAIN IT ON 7515 05:34:10,230 --> 05:34:12,766 A TASK, SEE HOW WELL IT PREDICTS. 7516 05:34:12,766 --> 05:34:14,268 THOSE ARE GREAT BUT I'M THINKING 7517 05:34:14,268 --> 05:34:15,802 MORE ABOUT DATA DRIVEN MODELS 7518 05:34:15,802 --> 05:34:17,170 BECAUSE WE HAVE THE CAPABILITY 7519 05:34:17,170 --> 05:34:20,741 TO GET SO MUCH DATA, IF YOU LOOK 7520 05:34:20,741 --> 05:34:28,282 AT THE SCALING LAWS AS SHOWN, 7521 05:34:28,282 --> 05:34:29,816 THESE MODELS ARE SATURATING, BUT THE DATA DRIVEN-MODELS ARE NOT SATURATED. IT CAN BECAUSE 7522 05:34:29,816 --> 05:34:31,351 CAN GET MORE DATA, I THINK THE 7523 05:34:31,351 --> 05:34:33,487 MODEL IS JUST THE BEGINNING TO 7524 05:34:33,487 --> 05:34:35,956 BASICALLY BE ABLE TO DO 7525 05:34:35,956 --> 05:34:36,623 EXPERIMENTS WITH LESS 7526 05:34:36,623 --> 05:34:42,963 RESTRICTIONS TO GET TO 7527 05:34:42,963 --> 05:34:43,830 PRINCIPLES. 7528 05:34:43,830 --> 05:34:47,668 >> BLAKE, IS IT PRINCIPLES OR 7529 05:34:47,668 --> 05:34:48,568 LOSS FUNCTIONS? 7530 05:34:48,568 --> 05:34:50,737 >> OH, NO, I THINK IT'S 7531 05:34:50,737 --> 05:34:53,807 DEFINITELY . 7532 05:34:53,807 --> 05:34:54,575 PRINCIPLES. 7533 05:34:54,575 --> 05:35:00,180 BUT I THINK THAT THE KEY POINT 7534 05:35:00,180 --> 05:35:04,351 IS, WHAT ANDREAS WAS HINTING AT, 7535 05:35:04,351 --> 05:35:08,288 PRINCIPLES EXPRESSED IN 7536 05:35:08,288 --> 05:35:08,689 MATHEMATICS. 7537 05:35:08,689 --> 05:35:10,123 HUMAN LANGUAGE IS GOING TO BE 7538 05:35:10,123 --> 05:35:11,358 INSUFFICIENT TO EXPLAIN WHAT'S 7539 05:35:11,358 --> 05:35:12,759 GOING IN THE BRAIN, MUCH AS IT 7540 05:35:12,759 --> 05:35:14,928 IS WITH PHYSICS. 7541 05:35:14,928 --> 05:35:18,799 CAN'T DO ADVANCED PHYSICS PURELY 7542 05:35:18,799 --> 05:35:20,634 WITH LANGUAGE BUT IT'S WHAT WE 7543 05:35:20,634 --> 05:35:21,868 TRIED DO IN NEUROSCIENCE FOR 7544 05:35:21,868 --> 05:35:23,070 MANY YEARS, USE OUR NATURAL 7545 05:35:23,070 --> 05:35:25,038 LANGUAGE TO DESCRIBE WHAT THE 7546 05:35:25,038 --> 05:35:27,641 SYSTEM IS DOING. 7547 05:35:27,641 --> 05:35:29,843 I THINK NECESSARILY WHAT 7548 05:35:29,843 --> 05:35:31,311 NEUROAI IS PUSHING US TOWARDS, 7549 05:35:31,311 --> 05:35:34,014 REALIZATION THAT WE DESCRIBE 7550 05:35:34,014 --> 05:35:36,616 SYSTEMS USING MATHEMATICS OF 7551 05:35:36,616 --> 05:35:38,585 HIGH DIMENSIONAL VECTOR SPACES, 7552 05:35:38,585 --> 05:35:41,088 DYNAMICS IN THOSE HIGH DIMENSIONAL 7553 05:35:41,088 --> 05:35:41,989 VECTOR SPACES, ET CETERA. 7554 05:35:41,989 --> 05:35:44,825 THINGS LIKE LOST FUNCTIONS, 7555 05:35:44,825 --> 05:35:47,894 ARCHITECTURES, WHATEVER, ARE 7556 05:35:47,894 --> 05:35:49,663 ULTIMATELY WAYS OF GETTING A 7557 05:35:49,663 --> 05:35:53,033 HOLD MENTALLY ON THAT 7558 05:35:53,033 --> 05:35:54,001 MATHEMATICAL PRINCIPLES. 7559 05:35:54,001 --> 05:35:55,936 IT'S A MECHANISM FOR US TO 7560 05:35:55,936 --> 05:35:57,671 TRANSLATE SOME OF THOSE 7561 05:35:57,671 --> 05:35:58,572 MATHEMATICAL PRINCIPLES INTO 7562 05:35:58,572 --> 05:36:00,574 THINGS THAT WE CAN VERBALIZE IN 7563 05:36:00,574 --> 05:36:01,875 OUR NATURAL LANGUAGE FOR THE 7564 05:36:01,875 --> 05:36:02,876 SAKE OF COMMUNICATING THEM TO 7565 05:36:02,876 --> 05:36:03,210 EACH OTHER. 7566 05:36:03,210 --> 05:36:07,147 BUT AT THE END OF THE DAY, I DO 7567 05:36:07,147 --> 05:36:12,819 THINK WE CAN EXTRACT PRINCIPLES, 7568 05:36:12,819 --> 05:36:13,387 VERY MUCH SO. 7569 05:36:13,387 --> 05:36:15,255 >> I HAVE A QUESTION FOR THE 7570 05:36:15,255 --> 05:36:16,656 ENTIRE PANEL BECAUSE THERE'S 7571 05:36:16,656 --> 05:36:18,525 SOMETHING THAT WAS CIRCLED A 7572 05:36:18,525 --> 05:36:21,862 COUPLE TIMES TODAY AND THAT WAS 7573 05:36:21,862 --> 05:36:23,430 THE NOTION OF ARTIFICIAL 7574 05:36:23,430 --> 05:36:24,631 INTELLIGENCE AND NATURAL 7575 05:36:24,631 --> 05:36:27,167 INTELLIGENCE HAVING SORT OF 7576 05:36:27,167 --> 05:36:28,368 ORTHOGONAL INCENTIVE. 7577 05:36:28,368 --> 05:36:32,539 LIKE A.I. ENGINEER, YOU WANT TO 7578 05:36:32,539 --> 05:36:34,608 BUILD A PERFECT PERFORMING 7579 05:36:34,608 --> 05:36:37,911 TASK-ALIGNED MODEL BUT ANIMALS 7580 05:36:37,911 --> 05:36:39,980 AND HUMANS, IMPERFECTIONS AND 7581 05:36:39,980 --> 05:36:42,616 INDIVIDUAL DIFFERENCES LEAD US 7582 05:36:42,616 --> 05:36:44,718 AWAY FROM THE FIRST GOAL. 7583 05:36:44,718 --> 05:36:47,087 SO, IN THE RESULTS THAT TALK 7584 05:36:47,087 --> 05:36:49,623 ABOUT MULTIPLE MODELS HAVING 7585 05:36:49,623 --> 05:36:50,490 CONVERGED TO SIMILAR 7586 05:36:50,490 --> 05:36:52,659 REPRESENTATIONS WITH EACH OTHER 7587 05:36:52,659 --> 05:36:54,227 THAT ARE SOMEWHAT SUBSTRATE 7588 05:36:54,227 --> 05:36:57,264 AGNOSTIC AS WAS POINTED OUT THIS 7589 05:36:57,264 --> 05:36:58,131 MORNING BY ALI MINAI. 7590 05:36:58,131 --> 05:37:02,836 WHAT I BUY LESS PERHAPS IS BOTH 7591 05:37:02,836 --> 05:37:06,039 MODELS, SELF-ALIGNED TO COMMON 7592 05:37:06,039 --> 05:37:07,574 REPRESENTATIONS HAVE ANYTHING TO 7593 05:37:07,574 --> 05:37:10,210 DO WITH THE SPACES THAT NATURAL 7594 05:37:10,210 --> 05:37:12,079 INTELLIGENCE HAS FOUND WITH ITS 7595 05:37:12,079 --> 05:37:12,646 BUMPS AND WARTS. 7596 05:37:12,646 --> 05:37:15,615 AND SO I THINK I WOULD WANT TO 7597 05:37:15,615 --> 05:37:19,286 HEAR FIRST FROM DR. CHANCE OR DR. 7598 05:37:19,286 --> 05:37:24,891 TSAO, BOTH OF WHOM MENTIONED 7599 05:37:24,891 --> 05:37:26,326 THIS IN THEIR TALKS. 7600 05:37:26,326 --> 05:37:27,127 >> WELL, I ACTUALLY AM GOING TO 7601 05:37:27,127 --> 05:37:29,062 TOUCH ON THIS A LITTLE BIT 7602 05:37:29,062 --> 05:37:30,263 TOMORROW SO MAYBE I'LL TRY TO 7603 05:37:30,263 --> 05:37:34,167 ANSWER THE QUESTION IN THAT 7604 05:37:34,167 --> 05:37:35,268 CONTEXT. 7605 05:37:35,268 --> 05:37:37,904 I MIGHT DISAGREE NATURAL 7606 05:37:37,904 --> 05:37:39,473 INTELLIGENCE IS AN OPTIMAL 7607 05:37:39,473 --> 05:37:41,975 SOLUTION SO MUCH AS THAT IS 7608 05:37:41,975 --> 05:37:43,310 OPTIMIZED FOR DIFFERENT CONSTRAINTS 7609 05:37:43,310 --> 05:37:44,077 AND I THINK WE AGREE HERE. 7610 05:37:44,077 --> 05:37:45,212 I THINK THERE'S A LOT TO BE 7611 05:37:45,212 --> 05:37:46,813 LEARNED FROM UNDERSTANDING WHAT 7612 05:37:46,813 --> 05:37:49,182 THE CONSTRAINTS FACING THE 7613 05:37:49,182 --> 05:37:51,518 ANIMAL ARE AND HOW THEY HAVE 7614 05:37:51,518 --> 05:37:52,285 SHAPED THE SOLUTION THAT BIOLOGY 7615 05:37:52,285 --> 05:37:53,720 HAS FOUND. 7616 05:37:53,720 --> 05:37:56,356 AND THAT IS SOMETHING I THINK 7617 05:37:56,356 --> 05:37:57,023 A.I. CAN TAKE. 7618 05:37:57,023 --> 05:37:59,192 AND THE OTHER THING I WOULD ADD 7619 05:37:59,192 --> 05:38:03,363 IS I THINK SOMETIMES WHEN WE -- 7620 05:38:03,363 --> 05:38:06,500 I AGREE I WANTED TO RESPOND TO 7621 05:38:06,500 --> 05:38:10,270 BLAKE, I AGREE THAT MATH IS 7622 05:38:10,270 --> 05:38:14,274 NECESSARY TO MOVE FROM BIOLOGICAL 7623 05:38:14,274 --> 05:38:16,009 DETAIL INTO A.I. MODELS BUT I 7624 05:38:16,009 --> 05:38:17,310 THINK ONE THING THAT GETS LOST 7625 05:38:17,310 --> 05:38:19,179 WHICH I THINK IS WHAT MAYBE 7626 05:38:19,179 --> 05:38:22,015 DORIS WAS TOUCHING ON WITH THE 7627 05:38:22,015 --> 05:38:23,416 COMMENT EARLIER, IS THE INTUITION 7628 05:38:23,416 --> 05:38:24,784 BIOLOGISTS GET WHEN THEY STUDY 7629 05:38:24,784 --> 05:38:25,819 THEIR SYSTEMS. 7630 05:38:25,819 --> 05:38:29,022 I DON'T WANT TO DERAIL DORIS' 7631 05:38:29,022 --> 05:38:29,956 QUESTION, BUT DORIS' ANSWER TO 7632 05:38:29,956 --> 05:38:30,657 THE QUESTION, BUT MAYBE 7633 05:38:30,657 --> 05:38:32,292 SOMETHING WE CAN TALK ABOUT HOW 7634 05:38:32,292 --> 05:38:36,029 DO YOU TRANSLATE WITHOUT LOSING 7635 05:38:36,029 --> 05:38:43,470 THAT INTUITION THAT'S SITTING IN 7636 05:38:43,470 --> 05:38:44,771 THE EXPERIMENTALISTS' HEAD. 7637 05:38:44,771 --> 05:38:47,674 >> I AGREE WITH WHAT FRANCES 7638 05:38:47,674 --> 05:38:48,175 SAID. THE IDEA BY PHILLIP ISOLA 7639 05:38:48,175 --> 05:38:49,776 CALLED THE PLATONIC 7640 05:38:49,776 --> 05:38:51,811 REPRESENTATION HYPOTHESIS THAT 7641 05:38:51,811 --> 05:38:53,980 THERE'S LIKE ONE REALITY, IF YOU 7642 05:38:53,980 --> 05:38:55,715 TRAIN A NETWORK THAT'S LARGE 7643 05:38:55,715 --> 05:38:56,650 ENOUGH WITH ENOUGH TASKS THEN 7644 05:38:56,650 --> 05:38:58,919 IT'S GOING TO FIND THAT MODEL 7645 05:38:58,919 --> 05:38:59,986 THAT REALITY AND THAT'S GOING TO 7646 05:38:59,986 --> 05:39:04,424 MATCH WHAT'S IN THE BRAIN. AND 7647 05:39:04,424 --> 05:39:08,061 I DON'T THINK WE'RE NEAR THAT 7648 05:39:08,061 --> 05:39:09,796 FANTASY LAND. THESE NETWORKS THAT 7649 05:39:09,796 --> 05:39:10,931 ARE TRAINED ARE NOWHERE NEAR WHAT 7650 05:39:10,931 --> 05:39:12,766 OUR BRAIN IS DOING. 7651 05:39:12,766 --> 05:39:16,002 THEY MAKE SO MANY MISTAKES, SO I 7652 05:39:16,002 --> 05:39:19,139 REALLY THINK WE NEED TO MAKE USE 7653 05:39:19,139 --> 05:39:21,208 OF EVERYTHING THAT WE KNOW ABOUT 7654 05:39:21,208 --> 05:39:22,976 THE BRAIN, RIGHT? 7655 05:39:22,976 --> 05:39:24,978 LIKE WHY SHOULD WE THROW ALL THE 7656 05:39:24,978 --> 05:39:26,813 DECADES OF PSYCHOLOGY 7657 05:39:26,813 --> 05:39:29,983 EXPERIMENTS OUT THE WINDOW? 7658 05:39:29,983 --> 05:39:31,718 AND ALL OUR NEUROSCIENCE DATA. 7659 05:39:31,718 --> 05:39:32,719 WE SHOULD USE ALL THE NEUROSCIENCE 7660 05:39:32,719 --> 05:39:34,387 DATA INFORMATION WE HAVE TO FIND 7661 05:39:34,387 --> 05:39:36,323 THE BEST MODEL OF THE BRAIN. 7662 05:39:36,323 --> 05:39:38,391 >> CAN I MAKE ONE BRIEF COMMENT? 7663 05:39:38,391 --> 05:39:41,795 SO I REALLY LIKE THIS POINT 7664 05:39:41,795 --> 05:39:43,063 ABOUT BIOLOGISTS INTUITIONS. 7665 05:39:43,063 --> 05:39:44,664 I THINK I WANT TO COMMENT ON 7666 05:39:44,664 --> 05:39:47,634 THAT, THAT RELATES TO EARLIER 7667 05:39:47,634 --> 05:39:51,471 DISCUSSION ON TRAINING, THE BEST 7668 05:39:51,471 --> 05:39:57,911 WAY TO ENSURE BIOLOGIST'S 7669 05:39:57,911 --> 05:40:00,146 INTUITION ARE PRESERVED IS TO HAVE 7670 05:40:00,146 --> 05:40:01,214 BIOLOGISTS WHO SPEAK MATHEMATICS. 7671 05:40:01,214 --> 05:40:03,516 SO, I THINK ONE OF THE MOST 7672 05:40:03,516 --> 05:40:04,451 IMPORTANT THINGS MOVING FORWARD 7673 05:40:04,451 --> 05:40:07,320 IN THE FIELD OF NEUROSCIENCE IN 7674 05:40:07,320 --> 05:40:10,390 GENERAL IS THAT WE CANNOT VIEW 7675 05:40:10,390 --> 05:40:12,259 FORMAL TRAINING IN MATHEMATICS 7676 05:40:12,259 --> 05:40:14,327 AND COMPUTER SCIENCE ANYMORE AS 7677 05:40:14,327 --> 05:40:15,128 AN OPTION. 7678 05:40:15,128 --> 05:40:16,029 IT SHOULDN'T BE SOMETHING 7679 05:40:16,029 --> 05:40:17,063 STUDENTS CAN CHOOSE TO TAKE OR 7680 05:40:17,063 --> 05:40:17,497 NOT. 7681 05:40:17,497 --> 05:40:20,700 IT SHOULD BE AS FUNDAMENTAL AS 7682 05:40:20,700 --> 05:40:24,271 LEARNING HODGKIN-HUXLEY, YOU 7683 05:40:24,271 --> 05:40:25,272 KNOW, PRINCIPLES OF HUBEL AND 7684 05:40:25,272 --> 05:40:26,039 WIESEL, THESE SORTS OF THINGS. 7685 05:40:26,039 --> 05:40:28,108 IT SHOULD NOT BE SOMETHING WE 7686 05:40:28,108 --> 05:40:28,541 CONSIDER SECONDARY. YOU SHOULD 7687 05:40:28,541 --> 05:40:30,911 START FROM THE VERY GET-GO 7688 05:40:30,911 --> 05:40:39,786 LEARNING MULTIVARIATE CALCULUS, 7689 05:40:39,786 --> 05:40:40,654 LINEAR ALGEBRA, STATISTICS, TO 7690 05:40:40,654 --> 05:40:42,522 TURN BIOLOGY INTO MATH, THAT WILL 7691 05:40:42,522 --> 05:40:46,993 BE THE WAY TO ACHIEVE WHAT WE'RE 7692 05:40:46,993 --> 05:40:47,994 ULTIMATELY STRIVING FOR. 7693 05:40:47,994 --> 05:40:49,229 >> THE COUNTER IS ALSO TRUE. 7694 05:40:49,229 --> 05:40:53,767 I WISH I HADN'T WAITED UNTIL 7695 05:40:53,767 --> 05:40:56,636 POSTDOC TO KNOW THAT COGNITIVE 7696 05:40:56,636 --> 05:41:01,675 SCIENCE HAS SO MUCH TO OFFER, 7697 05:41:01,675 --> 05:41:04,511 GRADIENTS SINCE 1920, LEARNING 7698 05:41:04,511 --> 05:41:05,211 RATES AND STUFF. 7699 05:41:05,211 --> 05:41:08,248 FEWER SILOS WOULD BE BETTER. 7700 05:41:08,248 --> 05:41:11,418 >> A HAND UP ON ZOOM, I INVITE 7701 05:41:11,418 --> 05:41:12,419 HER TO UNMUTE. 7702 05:41:12,419 --> 05:41:13,520 >> THANK YOU, JOE. 7703 05:41:13,520 --> 05:41:18,491 I HAVE NO IDEA WHAT I SOUND LIKE 7704 05:41:18,491 --> 05:41:18,925 IN THE AUDITORIUM. 7705 05:41:18,925 --> 05:41:22,462 CAN YOU HEAR ME OKAY? 7706 05:41:22,462 --> 05:41:24,631 >> WE CAN'T HEAR YOU. 7707 05:41:24,631 --> 05:41:28,234 >> I AM TALKING AND I DON'T KNOW 7708 05:41:28,234 --> 05:41:29,836 WHAT THE SITUATION IS LIKE. 7709 05:41:29,836 --> 05:41:31,504 CAN SOMEONE GIVE ME A THUMBS UP 7710 05:41:31,504 --> 05:41:32,839 IF YOU CAN HEAR ME? 7711 05:41:32,839 --> 05:41:35,775 >> WE CAN HEAR YOU NOW. 7712 05:41:35,775 --> 05:41:36,776 >> THANK YOU, FRANCIS. 7713 05:41:36,776 --> 05:41:38,478 SORRY FOR BEING A DISEMBODIED 7714 05:41:38,478 --> 05:41:40,680 VOICE IN THE LARGE ROOM. 7715 05:41:40,680 --> 05:41:42,115 I AGREE WITH WHAT DORIS SAID 7716 05:41:42,115 --> 05:41:44,184 ABOUT NOT THROWING AWAY THE 7717 05:41:44,184 --> 05:41:45,652 DECADES IF NOT CENTURIES OF 7718 05:41:45,652 --> 05:41:48,154 KNOWLEDGE WE HAVE. 7719 05:41:48,154 --> 05:41:49,055 AS A COMPUTATIONALLIST, I THINK 7720 05:41:49,055 --> 05:41:52,625 ONE OF THE KEY THINGS I GET FROM 7721 05:41:52,625 --> 05:41:54,160 TALKING TO EXPERIMENTALISTS IS 7722 05:41:54,160 --> 05:41:57,530 KNOWING WHAT FACTS TO IGNORE. 7723 05:41:57,530 --> 05:41:58,898 SO FOR ME IT'S ESSENTIALLY LESS 7724 05:41:58,898 --> 05:42:01,835 ABOUT WHAT ARE ALL THE DATASETS 7725 05:42:01,835 --> 05:42:03,837 WE'RE GOING TO THROW INTO TRAINING 7726 05:42:03,837 --> 05:42:05,672 DATASET BUT USING KNOWLEDGE FROM 7727 05:42:05,672 --> 05:42:07,640 BIOLOGY TO FIGURE OUT THESE ARE 7728 05:42:07,640 --> 05:42:08,541 ACTUALLY FACTS WE KNOW TO BE 7729 05:42:08,541 --> 05:42:10,810 TRUE ABOUT BIOLOGY BUT MAYBE WE 7730 05:42:10,810 --> 05:42:11,811 CAN SAFELY IGNORE THEM IN 7731 05:42:11,811 --> 05:42:13,146 MODELING FOR NOW BECAUSE THEY 7732 05:42:13,146 --> 05:42:14,681 ARE JUST CONFUSING AT THE LEVEL 7733 05:42:14,681 --> 05:42:17,617 OF DETAIL THAT WE DON'T NEED. 7734 05:42:17,617 --> 05:42:19,753 SIMILARLY, I AGREE WITH BLAKE 7735 05:42:19,753 --> 05:42:21,588 ABOUT THE IMPORTANCE OF HAVING 7736 05:42:21,588 --> 05:42:22,122 METHODICAL FORMALISM FOR 7737 05:42:22,122 --> 05:42:23,523 UNDERSTANDING WHAT THE BRAIN IS 7738 05:42:23,523 --> 05:42:25,191 DOING, WHATEVER IT'S DOING IT'S 7739 05:42:25,191 --> 05:42:26,826 NOT SOMETHING THAT WE CAN DRAW A 7740 05:42:26,826 --> 05:42:28,161 CARTOON PICTURE. 7741 05:42:28,161 --> 05:42:30,663 IT'S SOMETHING WE HAVE TO WRITE 7742 05:42:30,663 --> 05:42:31,431 DOWN AS MATHEMATICS. 7743 05:42:31,431 --> 05:42:32,365 HOPEFULLY IT'S NOT SOMETHING 7744 05:42:32,365 --> 05:42:35,301 THAT IS COMES AS A HUGE SURPRISE 7745 05:42:35,301 --> 05:42:37,337 TO PEOPLE THAT MY VERSION OF 7746 05:42:37,337 --> 05:42:39,305 THAT IS MORE THAN CALCULUS AND 7747 05:42:39,305 --> 05:42:42,041 LINEAR ALGEBRA. 7748 05:42:42,041 --> 05:42:44,277 IT REALLY IS FEEDBACK CONTROL, 7749 05:42:44,277 --> 05:42:45,111 DYNAMICS CONTROL. 7750 05:42:45,111 --> 05:42:47,480 AND THE REASON THAT I FEEL LIKE 7751 05:42:47,480 --> 05:42:49,115 THIS IS A FUNDAMENTAL PIECE OF 7752 05:42:49,115 --> 05:42:51,050 FORMALISM WE NEED TO UNDERSTAND 7753 05:42:51,050 --> 05:42:53,153 THE BRAIN IS BECAUSE THE BRAIN 7754 05:42:53,153 --> 05:42:54,521 NEVER REALLY JUST SITS THERE AND 7755 05:42:54,521 --> 05:42:55,388 PERCEIVES ANYTHING. 7756 05:42:55,388 --> 05:42:58,425 MOST OF US ARE WALKING AROUND 7757 05:42:58,425 --> 05:42:59,492 THE WORLD MAKING DECISIONS 7758 05:42:59,492 --> 05:43:02,128 MOMENT TO MOMENT, WHAT TO 7759 05:43:02,128 --> 05:43:03,196 PERCEIVE NEXT, SENSING IT, 7760 05:43:03,196 --> 05:43:06,433 SEEKING OUT MORE INFORMATION, 7761 05:43:06,433 --> 05:43:08,768 THAT'S SOMETHING WE GET OUT OF 7762 05:43:08,768 --> 05:43:10,003 EMBODIED INTELLIGENCE THAT AN 7763 05:43:10,003 --> 05:43:11,738 ARTIFICIAL NETWORK SITTING THERE 7764 05:43:11,738 --> 05:43:13,706 TALKING DOESN'T REALLY HAVE. 7765 05:43:13,706 --> 05:43:16,810 AND THIS ABILITY TO SEEK OUT 7766 05:43:16,810 --> 05:43:18,211 NEW INFORMATION IS EXPRESSED IN 7767 05:43:18,211 --> 05:43:20,213 THE LANGUAGE OF CONTROL THEORY 7768 05:43:20,213 --> 05:43:21,915 AND LANGUAGE OF REINFORCEMENT 7769 05:43:21,915 --> 05:43:23,116 AND LANGUAGE OF LEARNING. 7770 05:43:23,116 --> 05:43:25,485 AND IF IT'S ONE THING I DO KNOW 7771 05:43:25,485 --> 05:43:28,655 ABOUT THAT TYPE OF MATHEMATICS, 7772 05:43:28,655 --> 05:43:30,090 IT'S HORRENDOUSLY UNINTUITIVE. 7773 05:43:30,090 --> 05:43:32,392 I CAN SHOW SMALL DIAGRAM AND ASK 7774 05:43:32,392 --> 05:43:34,194 PEOPLE TO PREDICT WHAT'S GOING 7775 05:43:34,194 --> 05:43:38,465 TO HAPPEN, AND VERY FEW PEOPLE 7776 05:43:38,465 --> 05:43:40,033 CAN, IT'S NOT SOMETHING HUMANS 7777 05:43:40,033 --> 05:43:41,468 ARE GOOD AT. 7778 05:43:41,468 --> 05:43:42,235 FORTUNATELY WE HAVE MATHEMATICS 7779 05:43:42,235 --> 05:43:42,969 OF DOING THAT. 7780 05:43:42,969 --> 05:43:46,406 I'LL ADD THAT TO THE LIST OF 7781 05:43:46,406 --> 05:43:52,545 THINGS THAT ALL NEUROSCIENTISTS 7782 05:43:52,545 --> 05:43:53,213 SHOULD BE CONVERSANT AT - THAT 7783 05:43:53,213 --> 05:43:54,414 TYPE OF MATHEMATICS. THANK YOU. 7784 05:43:54,414 --> 05:43:57,150 >> THE LAST POINT ABOUT HOW 7785 05:43:57,150 --> 05:43:57,984 UNINTUITIVE A SMALL NETWORK CAN 7786 05:43:57,984 --> 05:44:01,087 BE THAT'S IN LARGE PART BECAUSE 7787 05:44:01,087 --> 05:44:02,021 THERE'S MORE MATH TO BE 7788 05:44:02,021 --> 05:44:04,591 DEVELOPED. I THINK ONE OF THE 7789 05:44:04,591 --> 05:44:05,592 GOALS FOR MATHEMATICAL 7790 05:44:05,592 --> 05:44:07,827 DEVELOPMENT IS TO LIKE IMPROVE 7791 05:44:07,827 --> 05:44:09,462 RESULT SO THAT THERE IS 7792 05:44:09,462 --> 05:44:12,165 INTUITION YOU CAN BUILD VIA THE 7793 05:44:12,165 --> 05:44:12,932 MATH. 7794 05:44:12,932 --> 05:44:15,468 INTUITION IS VALUABLE IN BOTH 7795 05:44:15,468 --> 05:44:16,436 DIRECTION, BIOLOGIST INTUITION 7796 05:44:16,436 --> 05:44:19,439 FROM YEARS OF EXPERIMENTS IS 7797 05:44:19,439 --> 05:44:20,106 VALUABLE. 7798 05:44:20,106 --> 05:44:20,940 THE MATHEMATICIAN'S INTUITION 7799 05:44:20,940 --> 05:44:22,275 FROM STUDYING THE MODEL SERIOUSLY 7800 05:44:22,275 --> 05:44:24,344 ON THEIR OWN IS VERY VALUABLE, 7801 05:44:24,344 --> 05:44:26,880 IF YOU PUT THOSE TOGETHER A LOT 7802 05:44:26,880 --> 05:44:29,048 OF PROGRESS CAN BE MADE. 7803 05:44:29,048 --> 05:44:31,251 >> IF I CAN JUST MAKE A POINT ON 7804 05:44:31,251 --> 05:44:41,794 THE ISSUE BEING DISCUSSED HERE, 7805 05:44:42,362 --> 05:44:44,430 I THINK IT'S FASCINATING, THE CONVERGENCE 7806 05:44:44,430 --> 05:44:49,702 BETWEEN THE REPRESENTATION FOUND 7807 05:44:49,702 --> 05:44:52,906 BY A.I. AND THE BRAIN. 7808 05:44:52,906 --> 05:44:57,410 BUT YOU HAVE TO REMEMBER AT 7809 05:44:57,410 --> 05:45:01,014 LEAST IN THE REAL NATURAL WORLD 7810 05:45:01,014 --> 05:45:11,291 INTELLIGENCE IS FOR LIVING INTELLIGENTLY 7811 05:45:11,291 --> 05:45:22,402 TO BE PRODUCTIVE,SUCCESSFUL, AND REPRODUCE. 7812 05:45:22,402 --> 05:45:26,591 INTELLIGENCE IS RELATED TO EMBODIMENT. 7813 05:45:26,591 --> 05:45:31,309 EMBODIED ORGANISM, NOT JUSTTHE FUNCTION OF 7814 05:45:31,309 --> 05:45:32,545 THE BRAIN. 7815 05:45:32,545 --> 05:45:35,415 OUR DISCUSSION KEEPS THE BRAIN AS A 7816 05:45:35,415 --> 05:45:38,918 COMPUTATIONAL OBJECT. LET’S SEEHOW WELL IT 7817 05:45:38,918 --> 05:45:40,253 CAN WORK TO OUR 7818 05:45:40,253 --> 05:45:45,892 OTHER COMPUTATIONAL OBJECTS, A.I. SYSTEMS. 7819 05:45:45,892 --> 05:45:52,165 BUT THIS IS NOT THE ENTIRETY OF INTELLIGENCE. 7820 05:45:52,165 --> 05:45:56,803 IT IS A FUNCTION OF THE WHOLE 7821 05:45:56,803 --> 05:46:07,213 ORGANISM, WHOLE ANIMAL. 7822 05:46:10,850 --> 05:46:16,823 INTELLIGENCE IS ALWAYS IN TERMS OF 7823 05:46:16,823 --> 05:46:18,825 AFFORDANCES - PERCEPTUAL AND 7824 05:46:18,825 --> 05:46:21,527 BEHAVIORAL. WILL BE THE SAME WITH 7825 05:46:21,527 --> 05:46:25,832 A.I. SYSTEMS AS WE MOVE TOWARDS 7826 05:46:25,832 --> 05:46:30,169 TRULY GENERAL A.I., A.I. THAT 7827 05:46:30,169 --> 05:46:38,444 SITS THERE ABOUT LANGUAGE AND 7828 05:46:38,444 --> 05:46:43,483 MATH IS NOT GENERAL A.I. 7829 05:46:43,483 --> 05:46:48,755 MATH IS NOT GENERAL A.I. 7830 05:46:48,755 --> 05:46:53,516 >>BING HAS DONE WORK MOVING AWAY FROM 7831 05:46:53,516 --> 05:46:58,040 BRAINS AS BAGS OFNEURONS AND IS BUILDING 7832 05:46:58,040 --> 05:46:58,831 MODELS 7833 05:46:58,831 --> 05:47:02,101 OF AGENTS AND EMBODIMENT. 7834 05:47:02,101 --> 05:47:08,457 WE CAN EMBODY AGENTS WITH PHYSICS, 7835 05:47:08,457 --> 05:47:16,647 BIOMECHANICS, KINEMATICS. WHEN DOWE GO WITH 7836 05:47:16,647 --> 05:47:18,551 ALL THAT? 7837 05:47:33,733 --> 05:47:40,606 EVOLUTION IS NOT YET DONE. 7838 05:47:40,606 --> 05:47:51,584 I DON’T THINK WE WILL BE DONE EITHER BECAUSE WE ARE DOING 7839 05:47:51,584 --> 05:48:00,493 WHAT EVOLUTION HAS DONE: CREATING ARTIFICIAL ORGANISMS. 7840 05:48:04,897 --> 05:48:09,257 ; IN VISION IF WE BUILD A SYSTEM THAT DOESN’T 7841 05:48:09,257 --> 05:48:13,806 HAVE TO REPLICATE ALL THE DETAILS OF THE BRAIN 7842 05:48:14,173 --> 05:48:23,082 BUT CRITICALLY CAPTURES ITS INFORMATION PROCESSING ALGORITHMS ANDACHIEVES 7843 05:48:23,516 --> 05:48:35,728 HUMAN-LEVEL ROBUST VISUAL PERCEPTION WITH WELL-DEFINED PERFORMANCE BENCHMARKS. 7844 05:48:35,728 --> 05:48:37,497 >> I WOULD SAY, YEAH, FOR 7845 05:48:37,497 --> 05:48:39,232 VISION, THAT'S AN AMAZING 7846 05:48:39,232 --> 05:48:39,632 SUCCESS. 7847 05:48:39,632 --> 05:48:42,401 YOU KNOW, I THINK THAT'S A 7848 05:48:42,401 --> 05:48:43,936 CLEARLY WELL-DEFINED BENCHMARKS 7849 05:48:43,936 --> 05:48:46,172 THAT WE CAN ACHIEVE. 7850 05:48:46,172 --> 05:48:49,475 THE SAME FOR LIKE ROBOTICS, YOU 7851 05:48:49,475 --> 05:48:53,646 CREATE A ROBOT THAT CAN GO AND 7852 05:48:53,646 --> 05:48:55,915 LOAD YOUR DISHWASHER AND HAS 7853 05:48:55,915 --> 05:48:58,751 DEXTERITY AND GRASPING IN A WAY 7854 05:48:58,751 --> 05:49:01,420 THAT IS INSPIRED BY 7855 05:49:01,420 --> 05:49:02,388 UNDERSTANDING OF MOTOR CONTROL 7856 05:49:02,388 --> 05:49:05,091 IN HUMANS OR OTHER ANIMALS, HAS 7857 05:49:05,091 --> 05:49:08,961 VISUAL SYSTEM, THAT'S A 7858 05:49:08,961 --> 05:49:13,232 WELL-DEFINED METRIC AND WE'RE 7859 05:49:13,232 --> 05:49:13,466 DONE. 7860 05:49:13,466 --> 05:49:18,671 >> ALSO ANDREAS USED THE WORD 7861 05:49:18,671 --> 05:49:19,605 "BEAUTY" EARLIER, THAT'S A BIG 7862 05:49:19,605 --> 05:49:20,306 PART OF IT. 7863 05:49:20,306 --> 05:49:22,642 THEN WE'LL HAVE A SENSE THAT 7864 05:49:22,642 --> 05:49:30,249 WE'RE GETTING THERE. 7865 05:49:30,249 --> 05:49:38,991 >> I CAN'T RESIST MURRAY GELL-MANN'S 7866 05:49:38,991 --> 05:49:40,193 REMARK IT'S SO BEAUTIFUL IT 7867 05:49:40,193 --> 05:49:43,296 SHOULD BE TAKEN AS EXPERIMENTAL 7868 05:49:43,296 --> 05:49:43,496 FACT. 7869 05:49:43,496 --> 05:49:47,767 >> ALONG THOSE LINES, HOW MUCH 7870 05:49:47,767 --> 05:49:50,636 IS VISUAL RECOGNITION A MODEL 7871 05:49:50,636 --> 05:49:52,772 FOR COGNITION? 7872 05:49:52,772 --> 05:49:56,409 IF WE GOT VISUAL RECOGNITION FOR 7873 05:49:56,409 --> 05:49:59,979 VIDEOS AS GOOD AS WE HAVE VISUAL 7874 05:49:59,979 --> 05:50:01,848 RECOGNITION FOR ImageNet, 7875 05:50:01,848 --> 05:50:02,982 WILL WE THEN HAVE ENOUGH VISION 7876 05:50:02,982 --> 05:50:05,251 TO SAY THIS IS THE FOUNDATION ON 7877 05:50:05,251 --> 05:50:11,858 WHICH WE SHOULD BUILD COGNITION? 7878 05:50:11,858 --> 05:50:13,059 >> YEAH, PEOPLE THINK ABOUT 7879 05:50:13,059 --> 05:50:17,997 RECOGNITION AS A FUNCTION OF THE 7880 05:50:17,997 --> 05:50:18,965 VENTRAL STREAM, THAT'S EVEN HALF 7881 05:50:18,965 --> 05:50:21,400 THE STORY OF VISION LET ALONE 7882 05:50:21,400 --> 05:50:23,202 COGNITION. 7883 05:50:23,202 --> 05:50:24,737 BUT I THINK THAT, YOU KNOW, LIKE 7884 05:50:24,737 --> 05:50:26,772 PROBABLY 80% OF THE BRAIN IS 7885 05:50:26,772 --> 05:50:28,774 VISUAL IN SOME SENSE, LIKE WILL 7886 05:50:28,774 --> 05:50:34,914 RESPOND TO VISUAL STIMULUS, EVEN 7887 05:50:34,914 --> 05:50:36,682 PRE-FRONTAL CORTEX, THERE'S 7888 05:50:36,682 --> 05:50:37,450 STRUCTURES THAT ARE MULTI-MODAL, 7889 05:50:37,450 --> 05:50:38,718 RESPONSIVE TO SOUND AND SO ON. 7890 05:50:38,718 --> 05:50:41,287 I FEEL LIKE WE'RE MISSING AN 7891 05:50:41,287 --> 05:50:42,788 UNDERSTANDING OF THE BASIC DATA 7892 05:50:42,788 --> 05:50:45,124 STRUCTURE OF HOW THE BRAIN IS 7893 05:50:45,124 --> 05:50:46,859 ORGANIZING THE WORLD, RIGHT? 7894 05:50:46,859 --> 05:50:48,327 WE JUST DON'T HAVE THAT RIGHT 7895 05:50:48,327 --> 05:50:48,494 NOW. 7896 05:50:48,494 --> 05:50:49,595 AND ONCE WE GET THAT I THINK 7897 05:50:49,595 --> 05:50:52,064 THAT'S GOING TO BE A FOUNDATION 7898 05:50:52,064 --> 05:50:54,800 FOR COGNITION, LIKE THOSE ARE 7899 05:50:54,800 --> 05:50:55,635 THE OBJECTS WE'RE THINKING 7900 05:50:55,635 --> 05:50:57,837 ABOUT, WHAT IS AN OBJECT, HOW IS 7901 05:50:57,837 --> 05:50:59,305 IT REPRESENTED IN THE BRAIN? 7902 05:50:59,305 --> 05:51:03,276 GOES FAR BEYOND THE QUESTION OF 7903 05:51:03,276 --> 05:51:03,743 RECOGNITION. 7904 05:51:03,743 --> 05:51:07,146 >> ONE WAY I LIKE TO THINK ABOUT 7905 05:51:07,146 --> 05:51:10,082 IS COGNITION IN HUMANS, YOU 7906 05:51:10,082 --> 05:51:12,084 KNOW, A LOT IS ABOUT PLANNING 7907 05:51:12,084 --> 05:51:14,787 INTO THE FUTURE, AND HAVING LIKE 7908 05:51:14,787 --> 05:51:16,255 A MODEL OF THE WORLD THAT 7909 05:51:16,255 --> 05:51:18,324 EXTENDS ALL THE WAY BACK TO THE 7910 05:51:18,324 --> 05:51:20,059 BIG BANG AND PREDICTING WHAT'S 7911 05:51:20,059 --> 05:51:21,060 GOING TO HAPPEN, YOU KNOW, 7912 05:51:21,060 --> 05:51:22,728 TOMORROW, A YEAR FROM NOW, TWO 7913 05:51:22,728 --> 05:51:24,263 YEARS FROM NOW. 7914 05:51:24,263 --> 05:51:27,500 AND I THINK IN PRIMATES VISION 7915 05:51:27,500 --> 05:51:33,306 IS SO DOMINANT OF A SENSE IT'S 7916 05:51:33,306 --> 05:51:34,307 ENABLES US TO BUILD COMPLEX 7917 05:51:34,307 --> 05:51:35,942 MODELS OF THE WORLD, A VIDEO IS 7918 05:51:35,942 --> 05:51:37,310 SHORT SPACE AND TIME BUT IF YOU 7919 05:51:37,310 --> 05:51:41,981 GO TO MORE COGNITIVE BEHAVIORS, 7920 05:51:41,981 --> 05:51:43,749 VISION OR LANGUAGE, IT'S ABOUT 7921 05:51:43,749 --> 05:51:44,583 EXTENDING THE WORLD 7922 05:51:44,583 --> 05:51:48,721 UNDERSTANDING IN SPACE AND TIME. 7923 05:51:48,721 --> 05:51:50,623 AND I THINK VISUAL WON'T GET US 7924 05:51:50,623 --> 05:51:52,792 ALL THE WAY THERE BUT I THINK 7925 05:51:52,792 --> 05:51:54,860 IT'S A HUGE OPPORTUNITY TO HAVE 7926 05:51:54,860 --> 05:51:57,396 A VERY GOOD START BY USING 7927 05:51:57,396 --> 05:52:00,499 VISUALLY GUIDED COGNITION AND 7928 05:52:00,499 --> 05:52:05,471 GOING INTO MORE COMPLEX 7929 05:52:05,471 --> 05:52:06,572 COGNITIVE BEHAVIORS. 7930 05:52:06,572 --> 05:52:11,143 >> I THINK THE ONLY THING I WILL 7931 05:52:11,143 --> 05:52:13,512 ADD THOUGH IS THERE IS AN 7932 05:52:13,512 --> 05:52:14,847 INTERESTING ISSUE FOR EMBODIMENT 7933 05:52:14,847 --> 05:52:16,182 AND ROBOTICS THAT IS OFTEN 7934 05:52:16,182 --> 05:52:17,950 OVERLOOKED WHICH IS THAT VISION 7935 05:52:17,950 --> 05:52:20,786 IS NICE AND ALL, BUT IF YOU'VE 7936 05:52:20,786 --> 05:52:22,355 GOT A BODY MOVING THROUGH SPACE, 7937 05:52:22,355 --> 05:52:24,490 I WOULD TAKE THE SENSE OF TOUCH 7938 05:52:24,490 --> 05:52:27,693 OVER VISION ANY DAY. 7939 05:52:27,693 --> 05:52:29,128 AND SO I THINK FOR ACTUALLY 7940 05:52:29,128 --> 05:52:30,763 UNDERSTANDING THE HUMAN BRAIN 7941 05:52:30,763 --> 05:52:33,232 AND HOW WE MODEL THE PHYSICAL 7942 05:52:33,232 --> 05:52:34,266 WORLD, PROBABLY A SENSE OF TOUCH 7943 05:52:34,266 --> 05:52:36,335 IS GOING TO BE CRITICAL IN THE 7944 05:52:36,335 --> 05:52:36,502 END. 7945 05:52:36,502 --> 05:52:39,739 AND IT'S SOMETHING THAT WE DON'T 7946 05:52:39,739 --> 05:52:41,374 YET HAVE ACCORDING TO MY 7947 05:52:41,374 --> 05:52:44,510 KNOWLEDGE A GOOD MODEL OF AT 7948 05:52:44,510 --> 05:52:45,011 ALL. 7949 05:52:45,011 --> 05:52:47,780 >> I WANT TO MAKE A COMMENT THAT 7950 05:52:47,780 --> 05:52:49,882 SO VISION OF COURSE IS A SENSORY 7951 05:52:49,882 --> 05:52:51,650 PERCEPTION, RIGHT? 7952 05:52:51,650 --> 05:52:54,453 BUT THERE'S ALSO THE WHOLE, YOU 7953 05:52:54,453 --> 05:52:55,187 KNOW, SPATIAL ORGANIZATION, 7954 05:52:55,187 --> 05:52:56,222 REPRESENTATION OF SPACE AND 7955 05:52:56,222 --> 05:52:57,490 OBJECTS IN SPACE. 7956 05:52:57,490 --> 05:53:00,793 AND THAT SEEMS TO BE OBVIOUSLY 7957 05:53:00,793 --> 05:53:01,761 VERY USEFUL FOR PERCEIVING 7958 05:53:01,761 --> 05:53:02,428 VISUAL WORLD BUT ALSO SOMETHING 7959 05:53:02,428 --> 05:53:04,230 YOU CAN USE FOR OTHER THINGS. 7960 05:53:04,230 --> 05:53:06,599 THERE ARE THESE EXAMPLES OF LIKE 7961 05:53:06,599 --> 05:53:08,734 BLIND MATHEMATICIANS WHO LIKE 7962 05:53:08,734 --> 05:53:10,336 HAVE AMAZING SPATIAL 7963 05:53:10,336 --> 05:53:10,936 UNDERSTANDING AND DO GEOMETRY 7964 05:53:10,936 --> 05:53:13,172 AND SO ON. 7965 05:53:13,172 --> 05:53:15,908 THERE'S SOMETHING ABOUT THAT RAW 7966 05:53:15,908 --> 05:53:17,109 REPRESENTATION OF SPACE IN TWO 7967 05:53:17,109 --> 05:53:18,611 OR THREE DIMENSIONS, WHATEVER IT 7968 05:53:18,611 --> 05:53:20,312 IS, I CAN'T GO BEYOND THREE 7969 05:53:20,312 --> 05:53:21,881 DIMENSIONS, BUT SOMETHING ABOUT 7970 05:53:21,881 --> 05:53:23,683 THAT WHICH I THINK IS DEVELOPED 7971 05:53:23,683 --> 05:53:27,653 TO AID THE VISUAL SYSTEM BUT 7972 05:53:27,653 --> 05:53:28,754 ENORMOUSLY USEFUL IN 7973 05:53:28,754 --> 05:53:30,923 UNDERSTANDING, YOU KNOW, 7974 05:53:30,923 --> 05:53:32,124 COGNITION AND OTHER MODALITIES. 7975 05:53:32,124 --> 05:53:33,426 >> PART OF THE MOTIVATION FOR MY 7976 05:53:33,426 --> 05:53:35,861 QUESTION WAS OF COURSE THAT WE 7977 05:53:35,861 --> 05:53:42,968 NOW KNOW THERE ARE VERY RICH 7978 05:53:42,968 --> 05:53:43,736 MOTOR SIGNALS, AUDITORY SIGNALS 7979 05:53:43,736 --> 05:53:46,138 IN THE CONTEXT, WHETHER YOU CAN 7980 05:53:46,138 --> 05:53:49,909 FACTORIZE VISION AWAY FROM ALL 7981 05:53:49,909 --> 05:53:54,380 THESE OTHER SENSES AND ACTION IS 7982 05:53:54,380 --> 05:53:54,980 THE QUESTION. 7983 05:53:54,980 --> 05:53:57,683 >> DEFINITELY YOU ARE CANNOT BUT 7984 05:53:57,683 --> 05:54:00,753 E.G., IN PRIMATES EYE MOVEMENT 7985 05:54:00,753 --> 05:54:04,223 IT'S DOMINANT MOTOR SYSTEM 7986 05:54:04,223 --> 05:54:06,625 IF YOU STUDY VISION FREE-VIEWING, 7987 05:54:06,625 --> 05:54:09,795 YOU GO DOWN TO V1, YOU HAVE 7988 05:54:09,795 --> 05:54:15,901 ISSUES OF INSTABILITY. HOW DO YOU 7989 05:54:15,901 --> 05:54:16,302 GET A STABLE ALLOCENTRIC 7990 05:54:16,302 --> 05:54:17,770 REPRESENTATION OF THE WORLD, AND 7991 05:54:17,770 --> 05:54:20,339 IF YOU HAVE BODIES MOVING IN IT 7992 05:54:20,339 --> 05:54:21,507 YOU GET COMPLEX DYNAMICS. 7993 05:54:21,507 --> 05:54:25,511 THIS STUFF HAS TO BE EVENTUALLY 7994 05:54:25,511 --> 05:54:28,814 INCORPORATED IN, BUT, YOU KNOW, 7995 05:54:28,814 --> 05:54:30,049 EXPERIMENTALLY IT'S FAIRLY 7996 05:54:30,049 --> 05:54:33,419 STRAIGHTFORWARD, IT'S EASIER TO 7997 05:54:33,419 --> 05:54:34,620 GET THIS DATA FREE VIEWING THAN 7998 05:54:34,620 --> 05:54:36,789 TRAINING ANIMALS TO FIXATE, 7999 05:54:36,789 --> 05:54:40,326 WE CAN EMBRACE AND DO 8000 05:54:40,326 --> 05:54:40,993 IMMEDIATELY. 8001 05:54:40,993 --> 05:54:43,195 >> SO DOES PANEL HAVE CONSENSUS 8002 05:54:43,195 --> 05:54:45,598 THAT THE NEXT GENERATIONS OF 8003 05:54:45,598 --> 05:54:49,001 NEUROAI SYSTEMS SHOULD BE 8004 05:54:49,001 --> 05:54:52,772 MULTIMODAL AT LEAST? 8005 05:54:52,772 --> 05:54:54,807 >> YEAH, AND BE MULTIPLE 8006 05:54:54,807 --> 05:54:56,008 MODALITIES, THEY WILL DEPEND ON 8007 05:54:56,008 --> 05:54:59,712 WHAT KIND OF SYSTEM IT IS, 8008 05:54:59,712 --> 05:55:00,079 RIGHT? 8009 05:55:00,079 --> 05:55:03,582 THE WORRY THAT A BAT INHABITS, 8010 05:55:03,582 --> 05:55:05,451 THE FAMOUS EXAMPLE, DIFFERENT 8011 05:55:05,451 --> 05:55:08,087 THAN THE WORLD THE HUMAN 8012 05:55:08,087 --> 05:55:10,656 INHABITS, PARTLY BECAUSE OF 8013 05:55:10,656 --> 05:55:13,359 SENSORY MODALITIES, IT'S ALSO IN 8014 05:55:13,359 --> 05:55:15,761 THE CONTEXT OF THE EMBODIMENT. 8015 05:55:15,761 --> 05:55:17,363 AND TOGETHER THEY HAVE A MODEL 8016 05:55:17,363 --> 05:55:18,497 OF THE WORLD. 8017 05:55:18,497 --> 05:55:19,965 AND I THINK THAT IS GOING TO -- 8018 05:55:19,965 --> 05:55:21,834 WHEN WE BUILD THESE A.I. SYSTEMS 8019 05:55:21,834 --> 05:55:23,569 THERE WILL BE A DIVERSITY OF THE 8020 05:55:23,569 --> 05:55:25,337 A.I. SYSTEM PROBABLY. 8021 05:55:25,337 --> 05:55:26,505 WHEN I SAY EMBODIMENT I DON'T 8022 05:55:26,505 --> 05:55:29,108 JUST MEAN IN THE PHYSICAL WORLD. 8023 05:55:29,108 --> 05:55:31,310 WE'LL HAVE TO DEFINE A 8024 05:55:31,310 --> 05:55:33,212 CORRESPONDING THING IN VIRTUAL 8025 05:55:33,212 --> 05:55:33,546 WORLD. 8026 05:55:33,546 --> 05:55:40,786 WHERE A.I. WILL ALSO LIVE. 8027 05:55:40,786 --> 05:55:43,189 >> CAN I RETURN TO THE QUESTION 8028 05:55:43,189 --> 05:55:44,857 ABOUT VISION AND COGNITION? 8029 05:55:44,857 --> 05:55:49,295 ONE OF THE HARDEST ASPECTS IS 8030 05:55:49,295 --> 05:55:50,729 PERCEIVING INVARIANCE. 8031 05:55:50,729 --> 05:55:53,299 MANY PEOPLE SAY HALLMARK OF 8032 05:55:53,299 --> 05:55:56,535 INTELLIGENCE IS GENERALIZATION. 8033 05:55:56,535 --> 05:55:59,772 CURIOUS IF CARINA CAN SPEAK TO 8034 05:55:59,772 --> 05:56:01,373 MATHEMATICAL THEORIES TO TIE 8035 05:56:01,373 --> 05:56:02,608 TOGETHER COMPUTING, ROTATION 8036 05:56:02,608 --> 05:56:05,778 VARIANCE, TO MORE GENERAL 8037 05:56:05,778 --> 05:56:06,812 PATTERN INVARIANCE. 8038 05:56:06,812 --> 05:56:08,914 >> YEAH, I THINK THAT'S REALLY 8039 05:56:08,914 --> 05:56:09,615 INTERESTING. 8040 05:56:09,615 --> 05:56:11,417 I THINK THERE'S IN COMPUTER 8041 05:56:11,417 --> 05:56:12,751 VISION LIKE THE DEEP NETWORKS 8042 05:56:12,751 --> 05:56:14,954 AND EVERYTHING THERE'S A RICH 8043 05:56:14,954 --> 05:56:15,955 HISTORY OF ACTUALLY 8044 05:56:15,955 --> 05:56:17,056 MATHEMATICIANS WORKING IN TRYING 8045 05:56:17,056 --> 05:56:22,127 TO BUILD THAT SORT OF 8046 05:56:22,127 --> 05:56:25,030 EXPLICITLY, BUILD UNION GROUP 8047 05:56:25,030 --> 05:56:25,564 THEORETIC TRANSFORMATIONS, 8048 05:56:25,564 --> 05:56:27,299 REPRESENTATION THEORY AND SO ON. 8049 05:56:27,299 --> 05:56:28,968 AND WHAT'S INTERESTING IS THAT 8050 05:56:28,968 --> 05:56:32,104 THE WAY THE BRAIN PROCESSES 8051 05:56:32,104 --> 05:56:33,706 VISUAL INFORMATION SEEMS TO HAVE 8052 05:56:33,706 --> 05:56:34,974 SOME OF THOSE INVARIANCES BUT 8053 05:56:34,974 --> 05:56:37,076 NOT ALL, SO YOU HAVE THIS GROUP 8054 05:56:37,076 --> 05:56:39,812 OF RIGID MOTIONS IN 3D, RIGHT? 8055 05:56:39,812 --> 05:56:41,780 OR SOMETHING LIKE THAT. 8056 05:56:41,780 --> 05:56:43,549 AND SOMEHOW WE DO PERCEIVE UP 8057 05:56:43,549 --> 05:56:48,354 AND DOWN SLIGHTLY DIFFERENTLY, 8058 05:56:48,354 --> 05:56:48,554 RIGHT? 8059 05:56:48,554 --> 05:56:49,321 INVARIANCES ARE NOT QUITE THE 8060 05:56:49,321 --> 05:56:49,755 SAME. THEY'RE THERE BUT 8061 05:56:49,755 --> 05:56:51,323 NOT QUITE, SORT OF THE 8062 05:56:51,323 --> 05:56:54,360 PLATONIC GEOMETRIC IDEAL 8063 05:56:54,360 --> 05:56:58,163 SHOULD BE, SO I 8064 05:56:58,163 --> 05:56:59,365 THINK THERE'S ACTUALLY VERY 8065 05:56:59,365 --> 05:57:01,066 INTERESTING QUESTIONS THERE. 8066 05:57:01,066 --> 05:57:03,202 YOU KNOW, FOR MATHEMATICIANS IN 8067 05:57:03,202 --> 05:57:03,736 COLLABORATION WITH VISION 8068 05:57:03,736 --> 05:57:06,105 EXPERTS TO WORK ON TO TRY TO 8069 05:57:06,105 --> 05:57:11,010 UNDERSTAND WHAT ARE THE BASIC, 8070 05:57:11,010 --> 05:57:15,714 YOU KNOW, GEOMETRIC INVARIANCES 8071 05:57:15,714 --> 05:57:17,449 OF OUR PERCEPTUAL WORLD. 8072 05:57:17,449 --> 05:57:19,418 I THINK THAT'S VERY INTERESTING. 8073 05:57:19,418 --> 05:57:22,488 >> CAN I ASK BLAKE A QUESTION 8074 05:57:22,488 --> 05:57:24,890 WHICH IS ONLY TANGENTALLY 8075 05:57:24,890 --> 05:57:27,326 RELATED TO THIS QUESTION? 8076 05:57:27,326 --> 05:57:28,060 INTERESTING CARINA SAYINGS 8077 05:57:28,060 --> 05:57:29,295 CERTAIN TYPES OF MATH MAY BE 8078 05:57:29,295 --> 05:57:32,097 NEEDED THAT WE CAN'T WRAP OUR 8079 05:57:32,097 --> 05:57:34,400 HEADS AROUND TODAY. 8080 05:57:34,400 --> 05:57:43,709 AND I WAS ADVOCATING FOR MATH 8081 05:57:43,709 --> 05:57:43,976 TO TRANSLATE. 8082 05:57:43,976 --> 05:57:45,577 MAYBE MOVING INFORMATION FROM 8083 05:57:45,577 --> 05:57:47,279 NEUROSCIENCE DATA INTO THE A.I. 8084 05:57:47,279 --> 05:57:48,013 MODELS. 8085 05:57:48,013 --> 05:57:49,982 SO, WOULD YOU SAY THAT MATH IS 8086 05:57:49,982 --> 05:57:54,453 THE RATE LIMITING STEP OR WOULD 8087 05:57:54,453 --> 05:57:56,255 YOU SAY THAT ARE WE STUCK 8088 05:57:56,255 --> 05:57:57,589 BECAUSE THE MATH DOESN'T EXIST? 8089 05:57:57,589 --> 05:58:00,326 OR ARE WE GOING TO BECOME STUCK 8090 05:58:00,326 --> 05:58:04,330 BECAUSE MAYBE NOT THAT MANY 8091 05:58:04,330 --> 05:58:05,464 MATHEMATICIANS -- MAYBE NOT MANY 8092 05:58:05,464 --> 05:58:08,767 HUMANS CAN HANDLE THAT MATH? 8093 05:58:08,767 --> 05:58:10,302 >> I'LL START, YEAH. 8094 05:58:10,302 --> 05:58:11,904 SO, MY BIASED ANSWER IS THAT 8095 05:58:11,904 --> 05:58:13,906 MATH IS ONE OF THE RATE 8096 05:58:13,906 --> 05:58:14,340 LIMITING STEPS. 8097 05:58:14,340 --> 05:58:15,774 I DON'T KNOW IF IT'S THE ONLY 8098 05:58:15,774 --> 05:58:18,444 ONE, BUT I DO THINK THAT THERE 8099 05:58:18,444 --> 05:58:21,380 IS THIS PROBLEM WHERE IT'S LIKE 8100 05:58:21,380 --> 05:58:23,182 YOU HAVE EVEN TAKE A DEEP 8101 05:58:23,182 --> 05:58:24,350 LEARNING NETWORK, RIGHT, THERE'S 8102 05:58:24,350 --> 05:58:25,551 NO DYNAMICS. 8103 05:58:25,551 --> 05:58:28,253 IT'S JUST A COMPOSITION OF MATH. 8104 05:58:28,253 --> 05:58:28,454 RIGHT? 8105 05:58:28,454 --> 05:58:29,555 AS A MATHEMATICIAN I WOULD 8106 05:58:29,555 --> 05:58:31,357 DESCRIBE THAT AS A 8107 05:58:31,357 --> 05:58:32,524 PARAMETERIZATION OF THE FAMILY 8108 05:58:32,524 --> 05:58:33,959 OF FUNCTIONS, RIGHT? 8109 05:58:33,959 --> 05:58:35,461 WEIGHTS ARE YOUR PARAMETERS, OR 8110 05:58:35,461 --> 05:58:37,996 YOUR WEIGHTS AND BIASES ARE 8111 05:58:37,996 --> 05:58:39,865 YOUR PARAMETERS, AND WE DON'T 8112 05:58:39,865 --> 05:58:41,433 UNDERSTAND THAT FAMILY WELL. 8113 05:58:41,433 --> 05:58:43,168 SO RIGHT NOW ACTUALLY IN THIS 8114 05:58:43,168 --> 05:58:44,670 MOMENT PARALLEL TO THIS THERE'S 8115 05:58:44,670 --> 05:58:47,773 A WHOLE LIKE PUSH IN MATH AND 8116 05:58:47,773 --> 05:58:50,175 A.I., RIGHT, WHERE THERE'S AN 8117 05:58:50,175 --> 05:58:52,411 EFFORT TO GET MATHEMATICIANS TO 8118 05:58:52,411 --> 05:58:54,346 LOOK VERY SERIOUSLY AT DEEP 8119 05:58:54,346 --> 05:58:55,013 LEARNING. 8120 05:58:55,013 --> 05:58:57,116 I THINK THEY SHOULD ALSO LOOK 8121 05:58:57,116 --> 05:58:58,183 SERIOUSLY AS NEUROAI, THAT'S 8122 05:58:58,183 --> 05:58:59,918 MY PUSH HERE, BUT 8123 05:58:59,918 --> 05:59:00,853 CERTAINLY THAT'S HAPPENING IN 8124 05:59:00,853 --> 05:59:01,320 DEEP LEARNING. 8125 05:59:01,320 --> 05:59:03,455 BECAUSE OF THE SUCCESS OF THOSE 8126 05:59:03,455 --> 05:59:05,457 MODELS, THERE'S NOW KIND OF A 8127 05:59:05,457 --> 05:59:05,958 GROWING COMMUNITY OF 8128 05:59:05,958 --> 05:59:08,360 MATHEMATICIANS WHO ARE TRYING TO 8129 05:59:08,360 --> 05:59:09,027 UNDERSTAND MATHEMATICALLY 8130 05:59:09,027 --> 05:59:12,231 WHAT'S GOING ON THERE, WHAT 8131 05:59:12,231 --> 05:59:13,966 HAPPENS IN TRAINING. 8132 05:59:13,966 --> 05:59:16,168 VIEW A DEEP NETWORK AS A FAMILY 8133 05:59:16,168 --> 05:59:18,337 OF FUNCTIONS, WITH SOME GIANT 8134 05:59:18,337 --> 05:59:18,937 PARAMETER SPACE, WHEN YOU'RE 8135 05:59:18,937 --> 05:59:20,072 TRAINING YOU'RE MOVING IN A 8136 05:59:20,072 --> 05:59:21,740 PARAMETER SPACE. 8137 05:59:21,740 --> 05:59:22,541 THERE ARE SYMMETRIES, DIRECTIONS 8138 05:59:22,541 --> 05:59:25,744 WHERE YOU CAN MOVE AND NOT ALTER 8139 05:59:25,744 --> 05:59:27,212 THE FUNCTION OR WHERE YOU ALTER 8140 05:59:27,212 --> 05:59:28,881 THE FUNCTION A LOT. A LOT OF 8141 05:59:28,881 --> 05:59:32,484 GEOMETRY TO THE PARAMETER SPACE, 8142 05:59:32,484 --> 05:59:37,656 THE WEIGHT SPACE, AND I THINK 8143 05:59:37,656 --> 05:59:38,057 THAT'S COMING. 8144 05:59:38,057 --> 05:59:39,491 BECAUSE OF A.I. THE MATH 8145 05:59:39,491 --> 05:59:40,459 COMMUNITY IS WAKING UP TO THE 8146 05:59:40,459 --> 05:59:42,127 FACT THAT THIS IS IMPORTANT AND 8147 05:59:42,127 --> 05:59:44,630 MAKES ME A LITTLE SAD, RIGHT, 8148 05:59:44,630 --> 05:59:46,832 THAT THEY ARE NOT MORE FOCUSED 8149 05:59:46,832 --> 05:59:47,499 ON THE NEUROSCIENCE SIDE WHICH I 8150 05:59:47,499 --> 05:59:49,468 THINK WOULD ALSO BE VERY, VERY 8151 05:59:49,468 --> 05:59:50,102 FRUITFUL. 8152 05:59:50,102 --> 05:59:51,904 BECAUSE I THINK THERE'S EVEN 8153 05:59:51,904 --> 05:59:53,472 MORE INTEREST MATH TO BE DONE 8154 05:59:53,472 --> 05:59:55,574 WHEN YOU LOOK AT DYNAMICS AND 8155 05:59:55,574 --> 05:59:56,909 DIFFERENT ARCHITECTURES WE SEE 8156 05:59:56,909 --> 06:00:00,279 IN NEUROSCIENCE AND SO ON. 8157 06:00:00,279 --> 06:00:02,915 SO AND THEN JUST TO KIND OF GO 8158 06:00:02,915 --> 06:00:04,583 BACK AND ANSWER YOUR ACTUAL 8159 06:00:04,583 --> 06:00:07,619 QUESTION, I DO THINK IF WE 8160 06:00:07,619 --> 06:00:09,421 UNDERSTAND THE MATH BETTER IT 8161 06:00:09,421 --> 06:00:11,890 WILL GUIDE US INTO BETTER 8162 06:00:11,890 --> 06:00:13,859 LEARNING RULES, YOU KNOW, 8163 06:00:13,859 --> 06:00:14,960 TRAINING PARADIGMS AND SO ON AT 8164 06:00:14,960 --> 06:00:17,329 LEAST FOR THOSE FAMILIES OF 8165 06:00:17,329 --> 06:00:17,796 MODELS. 8166 06:00:17,796 --> 06:00:19,932 I DO THINK I OPTIMISTICALLY 8167 06:00:19,932 --> 06:00:21,633 WOULD LOVE TO BE IN A WORLD 8168 06:00:21,633 --> 06:00:25,471 WHERE WE UNDERSTAND THE MATH OF 8169 06:00:25,471 --> 06:00:26,538 SIMPLE NONLINEAR MODELS AS WELL 8170 06:00:26,538 --> 06:00:28,774 AS WE UNDERSTAND THE MATH OF 8171 06:00:28,774 --> 06:00:29,374 LINEAR SYSTEMS. 8172 06:00:29,374 --> 06:00:30,776 IF WE WERE IN THAT WORLD WE 8173 06:00:30,776 --> 06:00:34,980 WOULD MAKE A LOT OF PROGRESS, I 8174 06:00:34,980 --> 06:00:35,447 THINK. 8175 06:00:35,447 --> 06:00:37,149 >> WELL, BING HAS HAD A HAND 8176 06:00:37,149 --> 06:00:39,051 RAISED FOR A WHILE. 8177 06:00:39,051 --> 06:00:39,785 >> GO AHEAD. 8178 06:00:39,785 --> 06:00:41,453 >> I'LL LET HER UNMUTE. 8179 06:00:41,453 --> 06:00:46,492 SHE'S SWOOPING IN ON THE SCREEN. 8180 06:00:46,492 --> 06:00:47,459 >> HI. 8181 06:00:47,459 --> 06:00:52,764 CAN YOU HEAR ME OKAY? 8182 06:00:52,764 --> 06:00:55,701 JOE, ARE YOU ABLE TO HEAR ME 8183 06:00:55,701 --> 06:00:56,602 OKAY? 8184 06:00:56,602 --> 06:00:57,436 >> EVERYTHING IS GOOD. 8185 06:00:57,436 --> 06:00:59,304 >> WE HEAR YOU. 8186 06:00:59,304 --> 06:00:59,571 >> OKAY. 8187 06:00:59,571 --> 06:01:00,506 THANKS. 8188 06:01:00,506 --> 06:01:02,274 I WANT TO QUICKLY RESPOND TO 8189 06:01:02,274 --> 06:01:02,908 STEVE'S QUESTION EARLIER WHETHER 8190 06:01:02,908 --> 06:01:05,744 OR NOT WE AGREE THAT THE NEXT 8191 06:01:05,744 --> 06:01:11,450 GENERATION OF MODELS SHOULD BE 8192 06:01:11,450 --> 06:01:12,684 MULTIMODAL. BLAKE'S ANALOGY 8193 06:01:12,684 --> 06:01:16,788 TALKING ABOUT ROBOTICS IS APT HERE, 8194 06:01:16,788 --> 06:01:19,224 AND ALI TALKED ABOUT HOW 8195 06:01:19,224 --> 06:01:19,791 FUNDAMENTALLY NEUROSCIENCE, 8196 06:01:19,791 --> 06:01:20,559 WE'RE STUDYING BIOLOGY, BIOLOGY 8197 06:01:20,559 --> 06:01:21,793 IS ABOUT ANIMALS, THAT IS THE 8198 06:01:21,793 --> 06:01:21,994 UNIT. 8199 06:01:21,994 --> 06:01:24,263 SO THE UNIT IS THE ANIMAL, THE 8200 06:01:24,263 --> 06:01:25,964 UNIT IS THE ROBOT, THAT MEANS 8201 06:01:25,964 --> 06:01:28,066 MUCH MORE THAN HAVING 8202 06:01:28,066 --> 06:01:31,570 MULTI-MODAL MODELS WE NEED 8203 06:01:31,570 --> 06:01:32,471 AGENT-BASED MODELS, INTEGRATION 8204 06:01:32,471 --> 06:01:33,505 AND PLANNING AND HAS MEMORY AND 8205 06:01:33,505 --> 06:01:35,807 PLANS FOR THE FUTURE. 8206 06:01:35,807 --> 06:01:36,909 NOT JUST LIKE FIGURING OUT WHAT 8207 06:01:36,909 --> 06:01:41,113 WE'LL DO IN A COUPLE YEARS, IN 8208 06:01:41,113 --> 06:01:42,181 EVERY TIME SCALE MILLISECONDS TO 8209 06:01:42,181 --> 06:01:43,482 YEAR IS THE THING WE SHOULD BE 8210 06:01:43,482 --> 06:01:44,616 DOING, I THINK. 8211 06:01:44,616 --> 06:01:48,520 AND I THINK IN ORDER TO BE ABLE 8212 06:01:48,520 --> 06:01:50,956 TO DO SOMETHING LIKE THAT IS 8213 06:01:50,956 --> 06:01:53,225 GOING TO, I DON'T KNOW, IT'S 8214 06:01:53,225 --> 06:01:55,627 LIKE AGAIN I'M GOING TO MAKE 8215 06:01:55,627 --> 06:01:56,128 ROBOTICS ANALOGIES HERE. 8216 06:01:56,128 --> 06:01:57,396 WE'RE NOT TRYING TO UNDERSTAND 8217 06:01:57,396 --> 06:01:57,696 THE WHEELS. 8218 06:01:57,696 --> 06:02:00,332 WE'RE NOT TRYING TO UNDERSTAND 8219 06:02:00,332 --> 06:02:00,899 THE WINGS. 8220 06:02:00,899 --> 06:02:01,934 WE NEED TO UNDERSTAND THOSE 8221 06:02:01,934 --> 06:02:03,502 THINGS BUT WE NEED TO UNDERSTAND 8222 06:02:03,502 --> 06:02:07,239 THE CONTEXT WHAT IT TAKES FOR 8223 06:02:07,239 --> 06:02:09,641 THE THING TO DO THE THING IT 8224 06:02:09,641 --> 06:02:09,841 ACTUALLY DOES. 8225 06:02:09,841 --> 06:02:11,343 IN ANIMALS, IT'S ESSENTIALLY TO 8226 06:02:11,343 --> 06:02:12,077 SURVIVE. 8227 06:02:12,077 --> 06:02:14,313 AND SO I THINK LIKE IF WE DON'T 8228 06:02:14,313 --> 06:02:16,215 KEEP IN MIND THAT'S THE ULTIMATE 8229 06:02:16,215 --> 06:02:18,650 GOAL, THAT WE ARE TRYING TO 8230 06:02:18,650 --> 06:02:20,285 ACCOMPLISH, IT'S GOING TO BE 8231 06:02:20,285 --> 06:02:22,821 EASY TO GET STUCK IN THESE 8232 06:02:22,821 --> 06:02:23,889 PIECEMEAL SOLUTIONS. 8233 06:02:23,889 --> 06:02:25,691 AND THAT LEVEL OF UNDERSTANDING 8234 06:02:25,691 --> 06:02:27,626 IS GOING TO BE LIKE EVEN IF WE 8235 06:02:27,626 --> 06:02:28,860 DO SUCCEED IN UNDERSTANDING 8236 06:02:28,860 --> 06:02:31,330 VISUAL SYSTEM THAT'S NOT WHAT 8237 06:02:31,330 --> 06:02:32,864 ANIMALS DO, EVEN HUMANS, AND 8238 06:02:32,864 --> 06:02:34,199 PRIMATES, WHO ARE VISUAL 8239 06:02:34,199 --> 06:02:34,833 CREATURES. 8240 06:02:34,833 --> 06:02:36,401 WE DON'T JUST SEE, WE DO A LOT 8241 06:02:36,401 --> 06:02:36,969 MORE THAN THAT. 8242 06:02:36,969 --> 06:02:39,671 I THINK THAT'S SOMETHING THAT WE 8243 06:02:39,671 --> 06:02:41,673 EXPERIENCE EVEN THE CONTEXT OF 8244 06:02:41,673 --> 06:02:43,609 SOCIAL INTERACTIONS LIKE WAS 8245 06:02:43,609 --> 06:02:46,445 BROUGHT UP EARLIER, LIKE WE'RE 8246 06:02:46,445 --> 06:02:48,447 VERY SOCIAL CREATURES. 8247 06:02:48,447 --> 06:02:50,616 THE REASON WE HAVE BODY 8248 06:02:50,616 --> 06:02:51,917 LANGUAGE, IF I WAS IN THE ROOM I 8249 06:02:51,917 --> 06:02:54,753 WOULD SEE YOUR BODY LANGUAGE 8250 06:02:54,753 --> 06:02:57,956 BETTER, AND SO THAT'S NOT 8251 06:02:57,956 --> 06:02:58,590 EXACTLY VISION, SORT OF 8252 06:02:58,590 --> 06:02:59,825 PROJECTING YOUR -- THE WAY 8253 06:02:59,825 --> 06:03:01,493 YOU'RE PROJECTING THE WAY YOU'RE 8254 06:03:01,493 --> 06:03:03,495 SITTING, LOOKING AT ME, YOU 8255 06:03:03,495 --> 06:03:04,963 MIGHT BE MAKING FACES, I CAN'T 8256 06:03:04,963 --> 06:03:06,431 REALLY SEE BUT BEING ABLE TO 8257 06:03:06,431 --> 06:03:07,833 INTERPRET IN THE CONTEXT HOW I 8258 06:03:07,833 --> 06:03:09,034 WOULD HOLD MY FACE AND CONTROL 8259 06:03:09,034 --> 06:03:11,103 THE FACIAL MUSCLES IN MY FACE, 8260 06:03:11,103 --> 06:03:13,538 WOULD BE TO INTERPRET THAT, 8261 06:03:13,538 --> 06:03:17,476 THAT'S THE REPRESENTATION, THE 8262 06:03:17,476 --> 06:03:17,776 THE 8263 06:03:17,776 --> 06:03:18,944 INVARIANCE IN MY HEAD. 8264 06:03:18,944 --> 06:03:21,780 IT'S MORE THAN VISUAL. 8265 06:03:21,780 --> 06:03:23,815 YOU HAVE TO IMAGINE WHAT I WOULD 8266 06:03:23,815 --> 06:03:29,855 DO IF I WERE THE AGENT TO INTERPRET 8267 06:03:29,855 --> 06:03:31,723 WHAT YOU ARE DOING MAKING A FACE. 8268 06:03:31,723 --> 06:03:32,624 >> SO I AGREE WITH CARINA, 8269 06:03:32,624 --> 06:03:35,293 THERE'S A LOT OF WORK TO DO IN 8270 06:03:35,293 --> 06:03:37,696 MATH TO UNDERSTAND ARTIFICIAL 8271 06:03:37,696 --> 06:03:39,631 NEURAL NETWORKS, MY GUESS WOULD 8272 06:03:39,631 --> 06:03:41,967 BE, WELL, AND TO UNDERSTAND 8273 06:03:41,967 --> 06:03:42,634 BIOLOGICAL NEURAL NETWORKS, A 8274 06:03:42,634 --> 06:03:44,836 LOT OF WORK IN ARTIFICIAL NEURAL 8275 06:03:44,836 --> 06:03:48,774 NETWORKS WILL PORT OVER TO 8276 06:03:48,774 --> 06:03:49,975 BIOLOGICAL NEURAL NETWORKS BUT 8277 06:03:49,975 --> 06:03:51,009 WORTH EMPHASIZING THERE ARE 8278 06:03:51,009 --> 06:03:52,611 ALREADY A LOT OF EXCELLENT 8279 06:03:52,611 --> 06:03:54,880 MATHEMATICAL TOOLS FOR DOING THE 8280 06:03:54,880 --> 06:03:57,616 COMPARISONS THAT MAYBE ARE NOT 8281 06:03:57,616 --> 06:03:59,217 FULLY UTILIZED BY THE FIELD OF 8282 06:03:59,217 --> 06:04:01,620 NEUROSCIENCE OR A.I. IN GENERAL, 8283 06:04:01,620 --> 06:04:04,189 TO THE EXTENT THEY COULD BE. 8284 06:04:04,189 --> 06:04:08,560 SO, I'LL GIVE CONCRETE EXAMPLE. 8285 06:04:08,560 --> 06:04:12,164 YOU KNOW, THERE WAS A BEAUTIFUL 8286 06:04:12,164 --> 06:04:14,633 PAPER, CARSEN STRINGER AND KEN 8287 06:04:14,633 --> 06:04:17,469 HARRIS WHERE THEY EXAMINE 8288 06:04:17,469 --> 06:04:22,774 GEOMETRY OF REPRESENTATIONS, 8289 06:04:22,774 --> 06:04:24,710 PRIMARY VISUAL CORTEX IN MICE, 8290 06:04:24,710 --> 06:04:30,082 HOW THE DECAY OF THE EIGEN- 8291 06:04:30,082 --> 06:04:31,616 SPECTRUM OF THE COVARIANCE 8292 06:04:31,616 --> 06:04:33,819 MATRIX TELLS YOU SOMETHING ABOUT 8293 06:04:33,819 --> 06:04:37,823 STRUCTURE OF THE MANIFOLD. A 8294 06:04:37,823 --> 06:04:39,157 GIFTED STUDENT WHO BASICALLY WAS 8295 06:04:39,157 --> 06:04:42,060 INSPIRED BY THIS WORK DIRECTLY 8296 06:04:42,060 --> 06:04:46,531 FROM NEUROSCIENCE, AND WENT AND 8297 06:04:46,531 --> 06:04:48,200 ANALYZED ALL OF THE MOST 8298 06:04:48,200 --> 06:04:49,801 ADVANCED A.I. MODELS, DISCOVERED 8299 06:04:49,801 --> 06:04:53,105 THAT THE A.I. MODELS THAT 8300 06:04:53,105 --> 06:04:54,439 PERFORMED BEST OUT OF 8301 06:04:54,439 --> 06:04:55,941 DISTRIBUTION GENERALIZATION ARE 8302 06:04:55,941 --> 06:05:03,115 THOSE THAT HAVE THE GEOMETRICAL 8303 06:05:03,115 --> 06:05:05,350 AND MANIFOLD PROPERTIES SIMILAR 8304 06:05:05,350 --> 06:05:06,351 TO THOSE SEEN THE BRAIN. COURTESY 8305 06:05:06,351 --> 06:05:09,721 OF MATHEMATICAL TOOLS THAT KEN 8306 06:05:09,721 --> 06:05:10,388 AND CARSON USED IN THEIR PAPER. 8307 06:05:10,388 --> 06:05:12,557 I THINK IT'S A BEAUTIFUL EXAMPLE 8308 06:05:12,557 --> 06:05:18,930 WHERE THANKS TO THE FACT THERE 8309 06:05:18,930 --> 06:05:20,766 WERE EXPERIMENTALISTS WHO DIDN'T 8310 06:05:20,766 --> 06:05:21,833 JUST PUBLISH DATA THAT DESCRIBED 8311 06:05:21,833 --> 06:05:23,468 THINGS WITHOUT THE MATH BUT THE 8312 06:05:23,468 --> 06:05:25,137 MATH WAS INTEGRAL TO THE ENTIRE 8313 06:05:25,137 --> 06:05:27,139 ANALYSIS THEY DID IN THEIR 8314 06:05:27,139 --> 06:05:29,841 PAPER, THAT MADE IT A LOT EASIER 8315 06:05:29,841 --> 06:05:31,309 TO HAVE A DIALOGUE AND TO 8316 06:05:31,309 --> 06:05:34,012 EXPLORE THINGS IN A.I. MODELS AS 8317 06:05:34,012 --> 06:05:35,480 WELL. 8318 06:05:35,480 --> 06:05:36,882 SO, I SUSPECT THERE'S A LOT OF 8319 06:05:36,882 --> 06:05:37,883 MATHEMATICAL TOOLS THAT COULD 8320 06:05:37,883 --> 06:05:40,786 LET US DO MORE LIKE THAT, IN 8321 06:05:40,786 --> 06:05:42,020 ADDITION TO THE FACT WE COULD 8322 06:05:42,020 --> 06:05:44,890 HAVE A BETTER DESCRIPTION AND 8323 06:05:44,890 --> 06:05:45,991 UNDERSTANDING OF NONLINEAR 8324 06:05:45,991 --> 06:05:47,959 DYNAMICS, ABSOLUTELY. 8325 06:05:47,959 --> 06:05:49,661 >> A QUICK COMMENT, KEN HARRIS 8326 06:05:49,661 --> 06:05:55,734 WAS AN UNDERGRAD MATH MAJOR, AND 8327 06:05:55,734 --> 06:05:59,471 HE WAS MY POSTDOC ADVISOR. WAS 8328 06:05:59,471 --> 06:06:00,372 TRAINED VERY MUCH IN MATH. 8329 06:06:00,372 --> 06:06:00,772 SO 8330 06:06:00,772 --> 06:06:03,175 THIS IS ONE OF THE EXAMPLES. 8331 06:06:03,175 --> 06:06:06,645 >> AT THIS POINT WE HAVE TO 8332 06:06:06,645 --> 06:06:10,348 SWITCH GEARS, I'M BEING TOLD. 8333 06:06:10,348 --> 06:06:14,286 I'D LIKE TO INTRODUCE THE SWITCH 8334 06:06:14,286 --> 06:06:19,891 THROUGH A STRONG NONLINEARITY. 8335 06:06:19,891 --> 06:06:21,426 ANDREAS WILL GIVE US A SHORT 8336 06:06:21,426 --> 06:06:24,796 PRESENTATION TO PERTURB US FROM 8337 06:06:24,796 --> 06:06:26,865 THE CURRENT DISCUSSION INTO THE 8338 06:06:26,865 --> 06:06:29,968 NEXT ROUND OF GENERAL 8339 06:06:29,968 --> 06:06:34,139 DISCUSSIONS, AND I'M TOLD 8340 06:06:34,139 --> 06:06:41,146 ANDREAS IS THE TWIN OF THE MODEL. 8341 06:06:41,146 --> 06:06:43,348 >> THIS IS JUST THE FOUR 8342 06:06:43,348 --> 06:06:46,017 SLIDES, JUST TO SPEAR HEAD THE 8343 06:06:46,017 --> 06:06:47,586 DISCUSSION AND GET EVERYBODY'S 8344 06:06:47,586 --> 06:06:50,155 INVOLVEMENT HERE OF WHAT ARE THE 8345 06:06:50,155 --> 06:06:51,690 POSSIBILITIES, WHAT WE WOULD 8346 06:06:51,690 --> 06:06:53,758 LIKE OR WOULD THE COMMUNITY LIKE 8347 06:06:53,758 --> 06:06:58,463 IN THE NEXT FEW YEARS TO REALLY 8348 06:06:58,463 --> 06:06:59,831 BRING NEUROSCIENCE AND A.I. IN 8349 06:06:59,831 --> 06:07:01,566 ALL THESE FORMS WE'VE BEEN 8350 06:07:01,566 --> 06:07:03,635 DISCUSSING TODAY THROUGHOUT THIS 8351 06:07:03,635 --> 06:07:08,773 MORNING, DISCUSSING MORE 8352 06:07:08,773 --> 06:07:09,040 TOMORROW. 8353 06:07:09,040 --> 06:07:09,741 SO WHAT DOES POTENTIAL ROAD 8354 06:07:09,741 --> 06:07:10,508 AHEAD LOOK LIKE? 8355 06:07:10,508 --> 06:07:14,813 HERE IS A SLIDE WE PUT TOGETHER 8356 06:07:14,813 --> 06:07:17,782 WITH, YOU KNOW, BLAKE RICHARDS, 8357 06:07:17,782 --> 06:07:19,517 XAQ, TONY ZADOR, ME AND A FEW 8358 06:07:19,517 --> 06:07:22,621 OTHER COLLEAGUES A COUPLE YEARS 8359 06:07:22,621 --> 06:07:23,154 AGO. 8360 06:07:23,154 --> 06:07:26,791 YOU KNOW, ESSENTIALLY LAYING OUT 8361 06:07:26,791 --> 06:07:28,526 WHAT WE THINK IS A GOOD WAY TO 8362 06:07:28,526 --> 06:07:29,761 THINK ABOUT FOUNDATION MODELS. 8363 06:07:29,761 --> 06:07:33,832 SO THE IDEA HERE IS THAT, YOU 8364 06:07:33,832 --> 06:07:36,401 KNOW, YOU CAN THINK OF NEURAL 8365 06:07:36,401 --> 06:07:39,271 NETWORKS AND A.I. AS A TOOL NOT 8366 06:07:39,271 --> 06:07:42,707 AS A MODEL OF THE BRAIN, IN A 8367 06:07:42,707 --> 06:07:45,443 WAY WE THINK IS TRYING TO LIKE 8368 06:07:45,443 --> 06:07:46,711 REPRESENT OR DO WHAT THE BRAIN 8369 06:07:46,711 --> 06:07:48,780 DOES, THE WAY IT DOES IT, BUT IS 8370 06:07:48,780 --> 06:07:51,716 A WAY TO FIND STATISTICAL 8371 06:07:51,716 --> 06:07:52,751 RELATIONSHIPS BETWEEN DOMAINS 8372 06:07:52,751 --> 06:07:54,886 WHERE WE HAVE A LOT OF DATA. 8373 06:07:54,886 --> 06:07:57,188 FOR EXAMPLE, HERE YOU COULD 8374 06:07:57,188 --> 06:07:59,558 IMAGINE YOU HAVE STIMULI, AND 8375 06:07:59,558 --> 06:08:01,393 THIS COULD INVOLVE VIDEO, THESE 8376 06:08:01,393 --> 06:08:05,764 MODELS COULD BE LIKE STANDARD 8377 06:08:05,764 --> 06:08:09,034 EVEN OFF-THE-SHELF MODELS, 8378 06:08:09,034 --> 06:08:10,235 PARSING VIDEO, LEARNING THE 8379 06:08:10,235 --> 06:08:12,637 STRUCTURE OF VIDEO, NEURAL 8380 06:08:12,637 --> 06:08:14,239 ACTIVITY, TASK SIMULATIONS, OR 8381 06:08:14,239 --> 06:08:15,373 THEORY. 8382 06:08:15,373 --> 06:08:16,775 ON THE OUTPUT YOU CAN GET, YOU 8383 06:08:16,775 --> 06:08:20,779 KNOW, YOU CAN DECODE FROM THESE 8384 06:08:20,779 --> 06:08:23,815 MODELS, USE IT FOR BRAIN-MACHINE 8385 06:08:23,815 --> 06:08:24,482 INTERFACE, PREDICT BEHAVIORS, 8386 06:08:24,482 --> 06:08:27,585 PREDICT ACTIVITY, HOW TO CONTROL 8387 06:08:27,585 --> 06:08:28,720 ACTIVITY, DIAGNOSE DISEASES, AND 8388 06:08:28,720 --> 06:08:29,988 SO ON. 8389 06:08:29,988 --> 06:08:32,791 AND TEST THEORIES. 8390 06:08:32,791 --> 06:08:35,193 THE IDEA IS THAT IF YOU THINK OF 8391 06:08:35,193 --> 06:08:36,661 THE -- YOU KNOW IS THERE GOING 8392 06:08:36,661 --> 06:08:40,565 TO BE A MOMENT LIKE ALPHAFOLD, 8393 06:08:40,565 --> 06:08:44,169 WHICH HAS BEEN SEMINAL DISCOVERY 8394 06:08:44,169 --> 06:08:45,704 OR ADVANCEMENT IN BIOLOGY, WHERE 8395 06:08:45,704 --> 06:08:48,273 THEY BROUGHT IN LARGE DATA THAT 8396 06:08:48,273 --> 06:08:50,709 HUMANS COLLECTED FOR MANY, MANY 8397 06:08:50,709 --> 06:08:55,547 YEARS, A LOT OF PhDs THESES, 8398 06:08:55,547 --> 06:08:58,516 MONEY SPENT DERIVING THE STRUCTURE 8399 06:08:58,516 --> 06:09:00,051 OF PROTEINS, WITH NEURAL 8400 06:09:00,051 --> 06:09:03,588 NETWORKS TO PRODUCE A PRODUCT 8401 06:09:03,588 --> 06:09:04,889 THAT IMPACTED BIOLOGY IN A VERY 8402 06:09:04,889 --> 06:09:06,324 BIG WAY. 8403 06:09:06,324 --> 06:09:07,592 AND IN MOLECULAR BIOLOGY THERE'S 8404 06:09:07,592 --> 06:09:10,929 OTHER KINDS OF IDEAS LIKE THAT, 8405 06:09:10,929 --> 06:09:13,498 YOU KNOW, LOOKING AT, YOU KNOW, 8406 06:09:13,498 --> 06:09:15,900 DNA, HOW TO SYNTHESIZE NEW 8407 06:09:15,900 --> 06:09:21,373 PROTEINS, HOW TO STUDY 8408 06:09:21,373 --> 06:09:22,841 PROTEIN-PROTEIN INTERACTION, 8409 06:09:22,841 --> 06:09:24,776 THIS MOMENT WHERE WE BRING LARGE 8410 06:09:24,776 --> 06:09:25,977 DATA AND LARGE COMPUTE TOGETHER. 8411 06:09:25,977 --> 06:09:29,347 THIS IS ONE IDEA. 8412 06:09:29,347 --> 06:09:32,917 NOW, IN ORDER TO DO THESE, I 8413 06:09:32,917 --> 06:09:35,153 SAID IT EARLIER ON, A COUPLE 8414 06:09:35,153 --> 06:09:36,254 OTHER SPEAKERS, CURRENT DATA WE 8415 06:09:36,254 --> 06:09:38,590 HAVE ARE GOING TO BE EXTREMELY 8416 06:09:38,590 --> 06:09:40,859 USEFUL IN MANY WAYS TO TEST 8417 06:09:40,859 --> 06:09:43,328 THESE MODELS BUT WE THINK THAT 8418 06:09:43,328 --> 06:09:46,431 TO BUILD THIS FOUNDATION MODEL 8419 06:09:46,431 --> 06:09:47,632 IN OUR EXPERIENCE YOU NEED TO 8420 06:09:47,632 --> 06:09:51,369 COLLECT IN A VERY SYSTEMATIC WAY 8421 06:09:51,369 --> 06:09:59,644 VERY LARGE DATASETS IN THESE 8422 06:09:59,644 --> 06:10:00,879 DIFFERENT MODALITIES, FOR 8423 06:10:00,879 --> 06:10:02,547 EXAMPLE, BRAIN HAS ALREADY DONE 8424 06:10:02,547 --> 06:10:04,416 AMAZING JOB COLLECTING THIS 8425 06:10:04,416 --> 06:10:04,983 ACTIVITY. 8426 06:10:04,983 --> 06:10:05,717 LOOKING AT NEURAL ACTIVITY, WE 8427 06:10:05,717 --> 06:10:08,453 NEED TO COME UP WITH 8428 06:10:08,453 --> 06:10:10,889 METHODOLOGIES TO COLLECT, YOU 8429 06:10:10,889 --> 06:10:14,059 KNOW, HIGH ENTROPIC DATA IN 8430 06:10:14,059 --> 06:10:15,727 NATURALISTIC CONDITIONS, IDEALLY 8431 06:10:15,727 --> 06:10:17,162 HAVING ANIMALS INTERACT FREELY 8432 06:10:17,162 --> 06:10:18,897 WITH EACH OTHER, SO-AND-SO 8433 06:10:18,897 --> 06:10:20,765 FORTH, TO HAVE ALSO EMBODIMENT. 8434 06:10:20,765 --> 06:10:22,600 THE OTHER THING THAT IS VERY 8435 06:10:22,600 --> 06:10:25,403 IMPORTANT HERE IS LARGE COMPUTE. 8436 06:10:25,403 --> 06:10:28,206 IF YOU CALCULATE, IF DO YOU BACK 8437 06:10:28,206 --> 06:10:29,207 OF THE ENVELOPE CALCULATIONS AND 8438 06:10:29,207 --> 06:10:31,342 PROJECT LET'S SAY THE AMOUNT OF 8439 06:10:31,342 --> 06:10:34,345 DATA THAT WE AS NEUROSCIENTISTS 8440 06:10:34,345 --> 06:10:35,413 MAY BE COLLECTING IN THE NEXT 8441 06:10:35,413 --> 06:10:38,550 TWO OR THREE YEARS, IF WE START 8442 06:10:38,550 --> 06:10:41,619 TO LIKE USE, YOU KNOW, EVEN LIKE 8443 06:10:41,619 --> 06:10:44,522 STANDARD TRANSFORMER TYPE OF 8444 06:10:44,522 --> 06:10:45,156 ARCHITECTURES, TOKENIZE THESE 8445 06:10:45,156 --> 06:10:47,459 THINGS, STARTING TO GET IN THE 8446 06:10:47,459 --> 06:10:52,230 BALL PARK OF QUITE LARGE SCALE 8447 06:10:52,230 --> 06:10:53,498 COMPUTE THAT RIGHT NOW IS NOT 8448 06:10:53,498 --> 06:10:55,033 CLEAR IF IT'S AVAILABLE AT THE 8449 06:10:55,033 --> 06:10:58,570 LEVEL OF WHAT WE HAVE 8450 06:10:58,570 --> 06:10:58,903 ACCESSIBILITY. 8451 06:10:58,903 --> 06:11:00,772 NOW, THE GOOD NEWS HERE, 8452 06:11:00,772 --> 06:11:01,639 ALTHOUGH THIS MARKET IS 8453 06:11:01,639 --> 06:11:06,744 FLUCTUATING A LOT, THERE IS SOME 8454 06:11:06,744 --> 06:11:09,380 EVIDENCE THAT MAYBE SILICON 8455 06:11:09,380 --> 06:11:11,249 VALLEY AND OTHERS OVERESTIMATED 8456 06:11:11,249 --> 06:11:13,752 NEED TO COMPUTE AND SOME NUMBERS 8457 06:11:13,752 --> 06:11:15,086 THAT YOU HEAR AROUND FOR EXAMPLE 8458 06:11:15,086 --> 06:11:18,890 THERE'S LIKE FIVE TO TEN MILLION 8459 06:11:18,890 --> 06:11:20,625 H100s OUT THERE, 30% BEING USED 8460 06:11:20,625 --> 06:11:23,228 AT ANY GIVEN POINT IN TIME. 8461 06:11:23,228 --> 06:11:25,597 COULD BE AN OPPORTUNITY HERE FOR 8462 06:11:25,597 --> 06:11:29,434 AN INITIVE TO PROVIDE LARGE SCALE 8463 06:11:29,434 --> 06:11:31,069 CLOUD COMPUTE FOR PEOPLE THAT 8464 06:11:31,069 --> 06:11:33,171 ARE BUILDING THIS KIND OF LARGE 8465 06:11:33,171 --> 06:11:34,139 SCALE MODEL. 8466 06:11:34,139 --> 06:11:36,774 THE OTHER THING THAT IS VERY 8467 06:11:36,774 --> 06:11:37,876 EXCITING IS THAT THESE MODELS, 8468 06:11:37,876 --> 06:11:39,344 THE WAY THEY CAN BE BUILT, THEY 8469 06:11:39,344 --> 06:11:40,912 ARE NOT JUST GOING TO BE -- THEY 8470 06:11:40,912 --> 06:11:43,982 ARE GOING TO BE MULTI-MODAL 8471 06:11:43,982 --> 06:11:45,383 GOING BEYOND, YOU KNOW, THE WAY 8472 06:11:45,383 --> 06:11:46,618 THAT MAYBE WHEN WE'VE BEEN 8473 06:11:46,618 --> 06:11:49,921 DISCUSSING TODAY IN TERMS OF 8474 06:11:49,921 --> 06:11:53,424 MOTOR THEY CAN'T COMBINE DATA 8475 06:11:53,424 --> 06:11:54,526 FROM DIFFERENT MODALITIES EVEN 8476 06:11:54,526 --> 06:11:56,027 WITHOUT HAVING THE SAME DATA, 8477 06:11:56,027 --> 06:12:00,365 THE SAME SPECIES, FOR EXAMPLE YOU 8478 06:12:00,365 --> 06:12:01,132 COULD COMBINE TRANSCRIPTOMIC DATA 8479 06:12:01,132 --> 06:12:09,140 FROM ECOLOGICAL MANIPULATIONS, 8480 06:12:09,140 --> 06:12:10,175 OTHER FIELDS ARE LEARNING 8481 06:12:10,175 --> 06:12:12,944 TRANSCRIPTOMIC LANDSCAPE OF BRAINS 8482 06:12:12,944 --> 06:12:16,381 OR OTHER ORGANS, OR PERFORMING 8483 06:12:16,381 --> 06:12:17,348 ECOLOGICAL MANIPULATIONS AND 8484 06:12:17,348 --> 06:12:18,149 DOING HIGH-THROUGHPUT 8485 06:12:18,149 --> 06:12:19,817 MEASUREMENTS, ONE COULD IN 8486 06:12:19,817 --> 06:12:20,952 THEORY BRING ALL THIS TOGETHER 8487 06:12:20,952 --> 06:12:25,023 IN THE FUTURE AND THEN HAVE A 8488 06:12:25,023 --> 06:12:26,991 LOT OF APPLICATIONS IN DISEASE 8489 06:12:26,991 --> 06:12:32,063 DIAGNOSIS FOR EXAMPLE OR EVEN 8490 06:12:32,063 --> 06:12:32,797 PREDICTING TREATMENTS FOR 8491 06:12:32,797 --> 06:12:33,431 DISEASES. 8492 06:12:33,431 --> 06:12:35,567 SO I JUST WANTED TO SHOW YOU 8493 06:12:35,567 --> 06:12:39,404 HERE ARE TWO EXAMPLES OF SOME 8494 06:12:39,404 --> 06:12:43,641 RECENT FOUNDATION MODELS, ONE BY 8495 06:12:43,641 --> 06:12:49,581 BLAKE RICHARD, EVA DYER, AND 8496 06:12:49,581 --> 06:12:50,882 LAJOLE LABS, WHERE THEY TOOK 8497 06:12:50,882 --> 06:12:55,086 DATA, LIKE MOUSE DATA, 8498 06:12:55,086 --> 06:12:55,687 BASICALLY USED TRANSFORMER 8499 06:12:55,687 --> 06:12:58,022 ARCHITECTURES, THEY HAVE BUILT 8500 06:12:58,022 --> 06:12:58,756 EARLY VERSIONS OF FOUNDATION 8501 06:12:58,756 --> 06:12:58,990 MODELS. 8502 06:12:58,990 --> 06:13:00,992 HERE ON THE LEFT IS ANOTHER 8503 06:13:00,992 --> 06:13:03,695 MODEL THAT IS BASED ON MICrONS 8504 06:13:03,695 --> 06:13:06,231 DATA THAT WE BUILT, THAT 8505 06:13:06,231 --> 06:13:08,833 ESSENTIALLY TAKES -- IS BUILDING 8506 06:13:08,833 --> 06:13:13,004 THESE MODEL, BUILT BY PUBLIC 8507 06:13:13,004 --> 06:13:14,772 DATA FROM MANY MICE, TRAINED ON 8508 06:13:14,772 --> 06:13:16,241 HUNDRED THOUSAND NEURONS, TRYING 8509 06:13:16,241 --> 06:13:18,810 TO LEARN RELATIONSHIPS BETWEEN 8510 06:13:18,810 --> 06:13:20,778 THE STRUCTURAL OF THE NEURAL 8511 06:13:20,778 --> 06:13:22,447 ACTIVITY. WHAT'S INTERESTING IS 8512 06:13:22,447 --> 06:13:26,517 MODELS ARE NOT JUST CAPABLE OF 8513 06:13:26,517 --> 06:13:28,253 PREDICTING NEURAL ACTIVITY, BUT 8514 06:13:28,253 --> 06:13:30,321 FOR EXAMPLE BECAUSE IN MICRONS 8515 06:13:30,321 --> 06:13:32,991 WE HAD FUNCTIONAL AND STRUCTURAL 8516 06:13:32,991 --> 06:13:36,127 DATA TOGETHER THE MODEL WAS 8517 06:13:36,127 --> 06:13:41,399 NEVER TRAINED ON ANATOMICAL DATA 8518 06:13:41,399 --> 06:13:49,874 COULD PREDICT MORPHOLOGICAL CELL 8519 06:13:49,874 --> 06:13:51,309 TYPES. CAN ACT AS MULTI-MODAL, 8520 06:13:51,309 --> 06:13:54,012 NOT JUST IN THE DOMAIN THAT THEY 8521 06:13:54,012 --> 06:13:57,515 ARE TRAINED, THAT'S SORT OF THE 8522 06:13:57,515 --> 06:13:59,517 ORIGINAL INSPIRATION OF CALLING 8523 06:13:59,517 --> 06:14:00,385 THEM FOUNDATION MODELS. 8524 06:14:00,385 --> 06:14:03,121 I THINK WE JUST LIKE TO CLOSE 8525 06:14:03,121 --> 06:14:04,956 WITH THIS SLIDE HERE, WHERE I 8526 06:14:04,956 --> 06:14:08,359 THINK SOME PEOPLE TALK ABOUT 8527 06:14:08,359 --> 06:14:09,327 PLATONIC SYMMETRY AND BEAUTY BUT 8528 06:14:09,327 --> 06:14:11,796 IF YOU THINK ABOUT IT WE HAVE 8529 06:14:11,796 --> 06:14:13,298 BRAINS WITH THE BODY OF COURSE, 8530 06:14:13,298 --> 06:14:16,567 WE HAVE THE WORLD, AND THEN WE 8531 06:14:16,567 --> 06:14:17,869 HAVE VISUAL NEURAL NETWORKS, ONE 8532 06:14:17,869 --> 06:14:21,339 FORM OR ANOTHER, SOME SORT OF IN 8533 06:14:21,339 --> 06:14:22,373 SILICO DEVICES NOW. 8534 06:14:22,373 --> 06:14:26,244 AND WE'RE TRYING TO FIND LIKE 8535 06:14:26,244 --> 06:14:26,944 RELATIONSHIPS, CORRESPONDENCES 8536 06:14:26,944 --> 06:14:30,415 BETWEEN THE THREE THINGS, HOW 8537 06:14:30,415 --> 06:14:31,950 THE WORLD IS ORGANIZED, HOW 8538 06:14:31,950 --> 06:14:34,052 BRAINS MODEL THE WORLD, HOW A.I. 8539 06:14:34,052 --> 06:14:36,020 MODELS THE WORLD, AND THE 8540 06:14:36,020 --> 06:14:37,855 TRIANGLE IS WHAT IT'S ALL ABOUT. 8541 06:14:37,855 --> 06:14:42,627 AND AT LEAST I BELIEVE AND MANY 8542 06:14:42,627 --> 06:14:45,163 OTHERS THAT USING DATA-DRIVEN 8543 06:14:45,163 --> 06:14:48,800 APPROACHES TO BUILD DIGITAL 8544 06:14:48,800 --> 06:14:50,868 TWINS OF BRAINS AND, YOU KNOW, 8545 06:14:50,868 --> 06:14:57,408 IT'S AN IMPORTANT ROADMAP TO 8546 06:14:57,408 --> 06:14:58,643 GET TO ISOMORPHISM BETWEEN BRAIN 8547 06:14:58,643 --> 06:14:58,876 AND THE WORLD. 8548 06:14:58,876 --> 06:15:08,319 I WOULD LIKE TO LEAVE IT THERE. 8549 06:15:08,319 --> 06:15:09,253 >> BLAKE? 8550 06:15:09,253 --> 06:15:10,888 >> WE'RE TAKING VIDEOCAST 8551 06:15:10,888 --> 06:15:12,123 QUESTIONS, WE WANT THIS RAPID 8552 06:15:12,123 --> 06:15:13,958 SESSION TO BE FORWARD LOOKING, 8553 06:15:13,958 --> 06:15:17,028 INTO THE NEXT TEN YEARS, SO 8554 06:15:17,028 --> 06:15:18,262 THINK BIG. 8555 06:15:18,262 --> 06:15:19,864 WHAT'S NEEDED, WHAT ARE GAPS, 8556 06:15:19,864 --> 06:15:30,375 WHAT DO WE NEED TO GET THERE? 8557 06:15:31,008 --> 06:15:33,277 >> HOW MUCH DATA COULD WE 8558 06:15:33,277 --> 06:15:34,545 POSSIBLY GET AND HOW THAT 8559 06:15:34,545 --> 06:15:40,418 COMPARE NOW TO LARGE LANGUAGE 8560 06:15:40,418 --> 06:15:43,054 MODELS? 8561 06:15:43,054 --> 06:15:43,187 8562 06:15:43,187 --> 06:15:47,492 >> YEAH, I THINK WE'RE NOT CRAZY 8563 06:15:47,492 --> 06:15:48,993 SCALABILITY WE CAN REACH IN 8564 06:15:48,993 --> 06:15:51,462 TERMS OF IN CLOSE TO TWO 8565 06:15:51,462 --> 06:15:55,266 PETABYTES OF DATA PER DAY FOR 8566 06:15:55,266 --> 06:15:56,901 EXAMPLE JUST ELECTROPHYSIOLOGY 8567 06:15:56,901 --> 06:15:59,737 ALONE, NON-HUMAN PRIMATES. 8568 06:15:59,737 --> 06:16:02,273 AND THEN, YOU CAN IMAGINE 8569 06:16:02,273 --> 06:16:03,941 HOW DO YOU TOKENIZE 8570 06:16:03,941 --> 06:16:05,676 THESE THINGS. 8571 06:16:05,676 --> 06:16:09,614 I THINK IN THE NEXT LET'S SAY 8572 06:16:09,614 --> 06:16:11,582 WITHIN THREE, FOUR, TO FIVE 8573 06:16:11,582 --> 06:16:14,786 YEARS, YOU KNOW, YOU CAN REACH, 8574 06:16:14,786 --> 06:16:16,421 YOU MAY HAVE NEEDS, 8575 06:16:16,421 --> 06:16:17,255 COMPUTATIONAL NEEDS OF THE 8576 06:16:17,255 --> 06:16:19,590 COMMUNITY THAT ARE AT THE SCALE 8577 06:16:19,590 --> 06:16:21,926 OF LET'S SAY LLAMA4 IS SUPPOSED 8578 06:16:21,926 --> 06:16:26,364 TO BE TRAINED ON. 8579 06:16:26,364 --> 06:16:31,803 >> YEAH, I'D AGREE WITH THAT. 8580 06:16:31,803 --> 06:16:34,205 I THINK ANDREAS HINTED, DEPENDS 8581 06:16:34,205 --> 06:16:36,507 HOW YOU TOKENNIZE AND CUT UP THE 8582 06:16:36,507 --> 06:16:36,941 DATA. 8583 06:16:36,941 --> 06:16:41,646 IF YOU RUN A BACK OF THE 8584 06:16:41,646 --> 06:16:43,514 ENVELOPE CALCULATION IT'S 8585 06:16:43,514 --> 06:16:45,316 POSSIBLE TO REQUIRE IN THE NEXT 8586 06:16:45,316 --> 06:16:46,050 FEW YEARS WITH CONCERTED EFFORT 8587 06:16:46,050 --> 06:16:48,219 AT THE SCALE OF DATA THAT THE 8588 06:16:48,219 --> 06:16:49,454 FIRST SETS OF LARGE LANGUAGE 8589 06:16:49,454 --> 06:16:55,193 MODELS WERE TRAINED ON. 8590 06:16:55,193 --> 06:16:57,128 >> IF I MAY PIVOT A TINY BIT 8591 06:16:57,128 --> 06:17:02,533 BEFORE YOUR QUESTION, MAYBE 8592 06:17:02,533 --> 06:17:04,936 AFTER YOUR QUESTION. 8593 06:17:04,936 --> 06:17:06,003 THE PURPOSE IS FORWARD LOOKING 8594 06:17:06,003 --> 06:17:07,872 TO SAY MAYBE IN TEN YEARS WILL 8595 06:17:07,872 --> 06:17:09,740 WE BE IN A POSITION TO 8596 06:17:09,740 --> 06:17:11,809 DISTINGUISH MODELS OF THE BRAIN 8597 06:17:11,809 --> 06:17:13,845 FROM REALLY GREAT ARCHITECTURES 8598 06:17:13,845 --> 06:17:23,387 THAT ALSO DO THINGS LIKE THE 8599 06:17:23,387 --> 06:17:24,689 BRAIN DOES? 8600 06:17:24,689 --> 06:17:26,224 >> IF I CAN ADD TO THE QUESTION, 8601 06:17:26,224 --> 06:17:27,925 THIS IS A QUESTION FOR THE 8602 06:17:27,925 --> 06:17:30,828 PEOPLE WORKING TOWARDS A.I., SO 8603 06:17:30,828 --> 06:17:32,363 PROBABLY BLAKE AND ALI. 8604 06:17:32,363 --> 06:17:34,765 A.I. IS A VERY 8605 06:17:34,765 --> 06:17:36,734 PERFORMANCE-DRIVEN FIELD. 8606 06:17:36,734 --> 06:17:39,904 SO, WHAT HAPPENS IF PUSHING 8607 06:17:39,904 --> 06:17:41,272 TOWARDS THE MODEL OF THE BRAIN 8608 06:17:41,272 --> 06:17:44,542 REQUIRES A STEP BACK IN 8609 06:17:44,542 --> 06:17:47,445 PERFORMANCE? 8610 06:17:47,445 --> 06:17:49,547 >> YEAH, SO I THINK THIS COMES 8611 06:17:49,547 --> 06:17:52,583 BACK TO THE DISCUSSION AROUND 8612 06:17:52,583 --> 06:17:55,052 METRICS OF COMPARISON, RIGHT? 8613 06:17:55,052 --> 06:17:58,756 SO INDEED I SUSPECT THAT ON SOME 8614 06:17:58,756 --> 06:18:00,758 METRIC MODELING THE BRAIN WOULD 8615 06:18:00,758 --> 06:18:02,026 REQUIRE A STEP BACKWARDS. 8616 06:18:02,026 --> 06:18:04,529 I THINK THOUGH IT'S SAFE TO SAY 8617 06:18:04,529 --> 06:18:07,732 THAT IN A LOT OF THE AREAS THAT 8618 06:18:07,732 --> 06:18:10,768 A.I. IS CURRENTLY POOR AT, WHAT 8619 06:18:10,768 --> 06:18:11,636 WE WANT IS MORE HUMAN-LIKE 8620 06:18:11,636 --> 06:18:14,338 BEHAVIOR FROM THE A.I. SYSTEMS. 8621 06:18:14,338 --> 06:18:16,774 SO, WE'RE NOT YET -- MARTIN WAS 8622 06:18:16,774 --> 06:18:18,276 SHOWING SOME OF THAT PLATEAU, 8623 06:18:18,276 --> 06:18:20,945 WITH SOME VISION MODELS, EVEN IN 8624 06:18:20,945 --> 06:18:22,847 VISION THE REALITY IS THAT ANY 8625 06:18:22,847 --> 06:18:24,248 A.I. RESEARCHER WOULD LOVE TO 8626 06:18:24,248 --> 06:18:26,450 HAVE A MODEL THAT'S MORE LIKE 8627 06:18:26,450 --> 06:18:28,653 HUMAN VISION THAN CURRENT A.I. 8628 06:18:28,653 --> 06:18:29,186 MODELS. 8629 06:18:29,186 --> 06:18:31,389 SO, I DON'T THINK WE'RE QUITE AT 8630 06:18:31,389 --> 06:18:33,024 THE POINT YET WHERE WE HAVE TO 8631 06:18:33,024 --> 06:18:38,396 WORRY ABOUT THAT TOO MUCH BUT 8632 06:18:38,396 --> 06:18:39,797 FORWARD THINKING MIGHT BE TEN 8633 06:18:39,797 --> 06:18:44,201 YEARS FROM NOW, AND I THINK THAT 8634 06:18:44,201 --> 06:18:45,703 THE INTERESTING QUESTION WILL 8635 06:18:45,703 --> 06:18:49,774 BE, YOU KNOW, TO WHAT EXTENT IS 8636 06:18:49,774 --> 06:18:52,009 THERE ACTUALLY LIKE A REAL 8637 06:18:52,009 --> 06:18:54,612 DIVERGENCE BETWEEN OUR A.I. 8638 06:18:54,612 --> 06:18:57,882 SYSTEMS AND HUMAN INTELLIGENCE 8639 06:18:57,882 --> 06:19:00,451 THAT'S JUST LIKE RADICALLY 8640 06:19:00,451 --> 06:19:00,885 DIFFERENT? 8641 06:19:00,885 --> 06:19:02,687 I THINK MY GUESS IS THAT IT 8642 06:19:02,687 --> 06:19:06,457 MIGHT NOT ACTUALLY END UP BEING 8643 06:19:06,457 --> 06:19:07,525 SO RADICALLY DIFFERENT INSOFAR 8644 06:19:07,525 --> 06:19:11,629 AS THE ECONOMIC IMPERATIVES 8645 06:19:11,629 --> 06:19:13,798 ARE TO CREATE A.I. SYSTEMS 8646 06:19:13,798 --> 06:19:16,100 TO MIMIC HUMAN INTELLIGENCE BUT 8647 06:19:16,100 --> 06:19:16,701 MAYBE NOT IMAGINING POTENTIAL 8648 06:19:16,701 --> 06:19:18,903 APPLICATIONS FOR A.I. TO DO 8649 06:19:18,903 --> 06:19:21,138 TOTALLY NON-HUMAN THINGS. 8650 06:19:21,138 --> 06:19:23,374 AND IF THAT EMERGES, IF THAT 8651 06:19:23,374 --> 06:19:25,676 CRYSTALLIZES, THEN I THINK YOU'LL 8652 06:19:25,676 --> 06:19:27,979 SEE THAT DIVERGENCE STRONGLY. 8653 06:19:27,979 --> 06:19:31,048 >> SO WHAT A POSSIBLE 10-YEAR 8654 06:19:31,048 --> 06:19:39,023 MOONSHOT GOAL BE AN IN SILICO 8655 06:19:39,023 --> 06:19:43,694 TRANSPLANT LIKE KIDNEYS OR 8656 06:19:43,694 --> 06:19:45,863 HEARING, MALADAPTIVE PROBLEM IN 8657 06:19:45,863 --> 06:19:47,531 THE BRAIN, WOULD YOU ABLE TO 8658 06:19:47,531 --> 06:19:49,233 RESTORE SIGHT, WOULD THAT BE A 8659 06:19:49,233 --> 06:19:50,901 HOLY GRAIL RATHER THAN METRICS 8660 06:19:50,901 --> 06:19:54,872 IN THE SHORT TERM? 8661 06:19:54,872 --> 06:19:57,341 IT'S A GREAT EXAMPLE OF A HOLY 8662 06:19:57,341 --> 06:19:57,908 GRAIL. 8663 06:19:57,908 --> 06:20:01,078 IF SOMEONE STUFFERS A STROKE IF 8664 06:20:01,078 --> 06:20:03,514 YOU COULD REPLACE THAT PART OF 8665 06:20:03,514 --> 06:20:05,716 THE BRAIN WITH A DIGITAL TWIN, 8666 06:20:05,716 --> 06:20:07,551 YES, A TOTAL AMAZING MOONSHOT TO 8667 06:20:07,551 --> 06:20:10,655 AIM FOR OR TO HAVE, YOU KNOW, 8668 06:20:10,655 --> 06:20:12,690 ADDITIONAL MODIFICATIONS OF YOUR 8669 06:20:12,690 --> 06:20:13,491 COGNITION. 8670 06:20:13,491 --> 06:20:15,259 A.I. ALREADY DOES THAT, BUT THEN 8671 06:20:15,259 --> 06:20:16,727 VIA YOUR HANDS AND THE SCREEN 8672 06:20:16,727 --> 06:20:20,264 BUT YOU CAN IMAGINE HAVING IT BE 8673 06:20:20,264 --> 06:20:22,033 MORE SEAMLESS WITH APPROPRIATE 8674 06:20:22,033 --> 06:20:28,272 BRAIN-COMPUTER INTERFACES. 8675 06:20:28,272 --> 06:20:31,242 >> I AGREE. 8676 06:20:31,242 --> 06:20:31,876 NICE INITIATIVE WELL DEFINED. 8677 06:20:31,876 --> 06:20:32,777 >> SO -- OH. 8678 06:20:32,777 --> 06:20:33,911 >> GO AHEAD. 8679 06:20:33,911 --> 06:20:37,048 >> JOE IS POINTING AT ME. 8680 06:20:37,048 --> 06:20:40,418 WANT TO PUSH ON HEALTH. WE'RE AT 8681 06:20:40,418 --> 06:20:41,686 THE NIH. LOT OF CONVERSATIONS 8682 06:20:41,686 --> 06:20:44,155 THAT I FIND LIKE HOW ARE WE 8683 06:20:44,155 --> 06:20:45,356 GOING TO GET NEUROSCIENCE DATA, 8684 06:20:45,356 --> 06:20:48,359 HAVING TRIED TO DO THIS FOR TEN 8685 06:20:48,359 --> 06:20:49,293 YEARS, LIKE RIDING A BIKE AND 8686 06:20:49,293 --> 06:20:50,728 TRYING TO JUMP ON A MOVING 8687 06:20:50,728 --> 06:20:52,763 TRAIN, THE FIELD IS MOVING SO 8688 06:20:52,763 --> 06:20:56,634 FAST WE'VE ALL TALKED ABOUT, 8689 06:20:56,634 --> 06:20:58,135 NEUROSCIENCE GOES SLOW. 8690 06:20:58,135 --> 06:21:01,338 THERE'S A BIGGER GOAL HERE, 8691 06:21:01,338 --> 06:21:02,340 MENTAL HEALTH, NEUROLOGICAL 8692 06:21:02,340 --> 06:21:04,175 DISORDERS, THIS IS HUGE. 8693 06:21:04,175 --> 06:21:05,476 THIS IS WHY WE'RE HERE. 8694 06:21:05,476 --> 06:21:08,579 IT'S WHY THE BRAIN INITIATIVE 8695 06:21:08,579 --> 06:21:08,879 EXISTS. 8696 06:21:08,879 --> 06:21:11,082 AND WHEN I HEAR THE VISION OF 8697 06:21:11,082 --> 06:21:14,251 THE FOUNDATION MODEL, I GUESS 8698 06:21:14,251 --> 06:21:18,556 THERE'S THIS -- SEEMS LIKE A BIG 8699 06:21:18,556 --> 06:21:20,624 PILL TO SWALLOW IF WE CAN THROW 8700 06:21:20,624 --> 06:21:22,460 THIS TOGETHER, DO A DIGITAL TWIN 8701 06:21:22,460 --> 06:21:29,133 OF BRAIN IN A.I. FORM, GREAT IF 8702 06:21:29,133 --> 06:21:31,769 THAT WORKS BUT TO TARGET 8703 06:21:31,769 --> 06:21:35,239 SOMETHING LESS MOONSHOTS OR 8704 06:21:35,239 --> 06:21:38,309 MAYBE MORE REALISTIC IS HOW ARE 8705 06:21:38,309 --> 06:21:40,978 WE GOING TO TAKE THIS SORT OF 8706 06:21:40,978 --> 06:21:43,681 STRUCTURE AND UNDERSTAND 8707 06:21:43,681 --> 06:21:51,322 SOMETHING, JUST UNDERSTAND THE 8708 06:21:51,322 --> 06:21:52,223 NEUROBIOLOGICAL UNDERPINNINGS OF 8709 06:21:52,223 --> 06:21:53,758 DEPRESSION OR SCHIZOPHRENIA TO 8710 06:21:53,758 --> 06:21:54,358 FIX IT. INSTEAD OF REPLACE, 8711 06:21:54,358 --> 06:21:56,193 STROKE IS LIKE A GROSS 8712 06:21:56,193 --> 06:21:57,795 IMPAIRMENT OF NEURAL 8713 06:21:57,795 --> 06:21:59,330 CIRCUITS, WE CAN REPLACE THAT 8714 06:21:59,330 --> 06:21:59,797 MAYBE. 8715 06:21:59,797 --> 06:22:04,201 BUT, YOU KNOW, WE HAVE MAJOR 8716 06:22:04,201 --> 06:22:05,035 NEUROLOGICAL DISORDERS NOT 8717 06:22:05,035 --> 06:22:07,037 MACROSCOPIC DAMAGE, RIGHT? 8718 06:22:07,037 --> 06:22:08,439 AND WE'RE NO WHERE NEAR 8719 06:22:08,439 --> 06:22:09,039 COMPUTATIONALLY UNDERSTANDING 8720 06:22:09,039 --> 06:22:11,375 HOW THEY WORK. 8721 06:22:11,375 --> 06:22:14,178 SO, MAYBE LIKE THE QUESTION IS 8722 06:22:14,178 --> 06:22:17,515 HOW CAN WE USE A.I. TO -- THE 8723 06:22:17,515 --> 06:22:21,118 SORT OF VISION THAT YOU'RE 8724 06:22:21,118 --> 06:22:23,687 INTRODUCING TO STRUCTURE AND 8725 06:22:23,687 --> 06:22:27,024 TARGET OUR COMPUTATIONAL 8726 06:22:27,024 --> 06:22:28,092 UNDERSTANDING OF THESE 8727 06:22:28,092 --> 06:22:29,326 DISORDERS? 8728 06:22:29,326 --> 06:22:30,361 >> I CAN ANSWER THAT. 8729 06:22:30,361 --> 06:22:33,831 THERE IS A VERY DIRECT PATH. 8730 06:22:33,831 --> 06:22:38,869 IN FACT THE BRAIN INITATIVE IS 8731 06:22:38,869 --> 06:22:40,571 FUNDING SOME OF THIS RESEARCH. 8732 06:22:40,571 --> 06:22:42,706 HERE IS THE PATH. 8733 06:22:42,706 --> 06:22:47,444 IF YOU HAVE A FOUNDATION MODEL 8734 06:22:47,444 --> 06:22:49,980 OF THE BRAIN WITH DEPRESSION, 8735 06:22:49,980 --> 06:22:52,016 YOU CAN STIMULATE, ONE OF THE 8736 06:22:52,016 --> 06:22:53,818 WAYS WE TRY AND TREAT DEPRESSION, 8737 06:22:53,818 --> 06:22:57,054 IN FACT WE HAVE EXPERTS IN 8738 06:22:57,054 --> 06:23:01,025 AUDIENCE, YOU CAN HAVE OPTIMIZE 8739 06:23:01,025 --> 06:23:02,726 -- BECAUSE BASICALLY YOU USE A 8740 06:23:02,726 --> 06:23:05,162 FOUNDATION MODEL TO FINE TUNE 8741 06:23:05,162 --> 06:23:08,032 FOR THE PARTICULAR PATIENT AND 8742 06:23:08,032 --> 06:23:08,632 USE ELECTRICAL STIMULATION TO 8743 06:23:08,632 --> 06:23:12,169 OPTIMIZE IN SILICO TO FIND THE 8744 06:23:12,169 --> 06:23:14,038 RIGHT STIMULATION PATTERN TO GET 8745 06:23:14,038 --> 06:23:15,673 THE ACTIVITY PATTERN TO THE 8746 06:23:15,673 --> 06:23:16,473 REGIME TO CURE DEPRESSION. 8747 06:23:16,473 --> 06:23:18,542 SO THAT'S A CONCRETE WAY OF AT 8748 06:23:18,542 --> 06:23:21,178 LEAST TRYING IN THE FUTURE TO 8749 06:23:21,178 --> 06:23:22,246 BUILD THESE KIND OF THINGS. 8750 06:23:22,246 --> 06:23:24,682 THE IDEA YOU HAVE A BRAIN STATE 8751 06:23:24,682 --> 06:23:25,683 THAT IS DEPRESSED, WE'RE GOING 8752 06:23:25,683 --> 06:23:28,352 TO GET THE BRAIN AWAY FROM THAT 8753 06:23:28,352 --> 06:23:35,025 BRAIN STATE, FIND THE OPTIMAL 8754 06:23:35,025 --> 06:23:37,628 STIMULATION PATTERN, HOW DO YOU 8755 06:23:37,628 --> 06:23:43,834 OPTIMIZE? CAN DO IT IN SILIO, 8756 06:23:43,834 --> 06:23:47,404 THROUGH DIGITAL TWIN, AND THEN 8757 06:23:47,404 --> 06:23:48,839 TEST IT CLOSED LOOP IN PATIENT. 8758 06:23:48,839 --> 06:23:53,711 WE HAVE A PROJECT LIKE THIS WITH 8759 06:23:53,711 --> 06:23:54,144 SAMEER SHETH AND XAQ PITKOW. 8760 06:23:54,144 --> 06:24:00,584 >> I'LL MAKE ANOTHER COMMENT, 8761 06:24:00,584 --> 06:24:02,520 WHICH WILL BELONG TOMORROW WHEN 8762 06:24:02,520 --> 06:24:04,855 WILL TALK ABOUT NEUROMORPHIC. 8763 06:24:04,855 --> 06:24:07,925 HOW TO CAPTURE ASPECTS OF BRAIN? 8764 06:24:07,925 --> 06:24:08,759 WE HAVE AN OPPORTUNITY TO 8765 06:24:08,759 --> 06:24:11,662 LOOK AT HOW SMALL CHANGES IN 8766 06:24:11,662 --> 06:24:15,065 PHYSICAL SYSTEMS AFFECT MAYBE 8767 06:24:15,065 --> 06:24:17,167 MORE MACRO BEHAVIOR. AND 8768 06:24:17,167 --> 06:24:19,436 UNDERSTANDING HOW TINY CHANGES 8769 06:24:19,436 --> 06:24:21,739 LIKE ION CHANNEL TURN INTO MENTAL 8770 06:24:21,739 --> 06:24:31,081 DISEASES IS SOMETHING I DON'T 8771 06:24:31,081 --> 06:24:37,087 NEUROSCIENCE HAS A HANDLE ON. 8772 06:24:37,087 --> 06:24:40,791 NEUROMORPHIC IS A PATHWAY TO COMPARE CHANGES THAT CAN BE TESTED IN A.I. 8773 06:24:40,791 --> 06:24:41,091 QUANTIFIABLY AND IN THE BRAIN. 8774 06:24:41,091 --> 06:24:41,992 >> THAT'S SOMETHING THAT IS 8775 06:24:41,992 --> 06:24:43,961 SOMETHING WE DON'T HAVE A GOOD 8776 06:24:43,961 --> 06:24:44,962 MATHEMATICAL HANDLE ON EITHER. 8777 06:24:44,962 --> 06:24:47,231 IT'S LIKE IF YOU GIVE ME A 8778 06:24:47,231 --> 06:24:49,433 MODEL, I KNOW EVERYTHING, EVERY 8779 06:24:49,433 --> 06:24:51,835 PARAMETER, IT'S AN ARTIFICIAL 8780 06:24:51,835 --> 06:24:53,938 INTELLIGENCE MODEL, WHAT'S GOING 8781 06:24:53,938 --> 06:24:56,874 TO HAPPEN WHEN I CHANGE 8782 06:24:56,874 --> 06:24:57,975 PARAMETERS, A BETTER 8783 06:24:57,975 --> 06:25:00,411 UNDERSTANDING OF HOW MOVING IN 8784 06:25:00,411 --> 06:25:01,812 THAT PARAMETER SPACE AFFECT, THE 8785 06:25:01,812 --> 06:25:04,248 DYNAMIC WILL HELP US TO 8786 06:25:04,248 --> 06:25:05,716 UNDERSTAND WHAT'S GOING WRONG, 8787 06:25:05,716 --> 06:25:08,819 BECAUSE A LOT OF THESE TRICKY 8788 06:25:08,819 --> 06:25:10,654 DISORDERS ARE DYNAMICS THAT ARE 8789 06:25:10,654 --> 06:25:12,323 GOING WRONG, BECAUSE SOMETHING'S 8790 06:25:12,323 --> 06:25:15,492 HAPPENING IN A MICROSCOPIC LEVEL, 8791 06:25:15,492 --> 06:25:18,062 YOU SEE THIS MACROSCOPIC ISSUE, 8792 06:25:18,062 --> 06:25:21,131 I THINK ABSOLUTELY WE NEED A 8793 06:25:21,131 --> 06:25:22,533 BETTER UNDERSTANDING THAT A.I. 8794 06:25:22,533 --> 06:25:32,643 QUESTIONS CAN HELP. 8795 06:25:32,643 --> 06:25:34,845 >> I'LL GO NEXT. 8796 06:25:34,845 --> 06:25:36,580 I REALLY APPRECIATED BLAKE'S 8797 06:25:36,580 --> 06:25:37,848 POINTS ABOUT IMAGE DISCUSSION 8798 06:25:37,848 --> 06:25:39,850 BASICALLY SAYING, HEY, TOUCH IS 8799 06:25:39,850 --> 06:25:40,784 KIND OF IMPORTANT. 8800 06:25:40,784 --> 06:25:45,389 AND SO IF WE'RE GOING TO BE 8801 06:25:45,389 --> 06:25:47,891 LOWER, LOWER LIFE FORM 8802 06:25:47,891 --> 06:25:50,027 REPRESENTATION FOR THE MOMENT, 8803 06:25:50,027 --> 06:25:54,932 COCKROACHES HAVE REALLY GREAT 8804 06:25:54,932 --> 06:25:55,165 FEELERS. 8805 06:25:55,165 --> 06:25:58,335 ONLY A MILLION NEURONS. 8806 06:25:58,335 --> 06:26:00,471 TO YOUR POINT WHERE IS THE 8807 06:26:00,471 --> 06:26:03,941 PROJECT TO EMULATE A COCKROACH 8808 06:26:03,941 --> 06:26:04,308 ANTENNA TODAY? 8809 06:26:04,308 --> 06:26:08,445 THERE'S A LOT OF WORK ON IMAGE 8810 06:26:08,445 --> 06:26:10,381 SENSORS, DELTA FRAME SENSORS, 8811 06:26:10,381 --> 06:26:12,182 HIGH ORDERS OF COMPLEXITY IN IMAGE 8812 06:26:12,182 --> 06:26:15,452 SENSING PIXEL, BUT WHAT ABOUT 8813 06:26:15,452 --> 06:26:16,453 THE ANTENNA? 8814 06:26:16,453 --> 06:26:18,288 SO SHOULDN'T WE BE FUNDING A 8815 06:26:18,288 --> 06:26:21,125 PROJECT TO WORK BACKWARDS FROM 8816 06:26:21,125 --> 06:26:27,264 THE SENSE, UNDERSTAND HOW 8817 06:26:27,264 --> 06:26:28,465 NEUROMORPHIC SIGNALS WORK, AND GO 8818 06:26:28,465 --> 06:26:32,770 BACKWARDS RATHER THAN FROM THE 8819 06:26:32,770 --> 06:26:34,705 BRAIN FORWARDS? 8820 06:26:34,705 --> 06:26:36,340 >> SO I'LL JUST SAY I MEAN I'M 8821 06:26:36,340 --> 06:26:37,207 FULLY ON BOARD. 8822 06:26:37,207 --> 06:26:38,409 I THINK IT WOULD BE FASCINATING 8823 06:26:38,409 --> 06:26:42,513 TO UNDERSTAND HOW A COCKROACH'S 8824 06:26:42,513 --> 06:26:44,014 ANTENNA WORKS. 8825 06:26:44,014 --> 06:26:47,785 SOME PEOPLE ARE TRYING TO DO 8826 06:26:47,785 --> 06:26:49,920 THIS KIND OF WORK, SO A 8827 06:26:49,920 --> 06:26:52,389 COLLEAGUE OF MINE AT HARVARD IS 8828 06:26:52,389 --> 06:26:54,691 DOING WORK ON BUILDING NOT 8829 06:26:54,691 --> 06:26:55,526 COCKROACH ANTENNA BUT VIRTUAL 8830 06:26:55,526 --> 06:26:59,329 RAT WITH A BODY THAT MOVES WITH 8831 06:26:59,329 --> 06:26:59,897 PROPRIOCEPTIVE SIGNALS, YOU 8832 06:26:59,897 --> 06:27:01,365 COULD IMAGINE STARTING TO 8833 06:27:01,365 --> 06:27:02,399 INCORPORATE TOUCH SIGNALS. 8834 06:27:02,399 --> 06:27:06,003 I'D LOVE TO SEE MORE OF THIS 8835 06:27:06,003 --> 06:27:06,437 WORK. 8836 06:27:06,437 --> 06:27:07,671 VIRTUAL BODIES, ALL SORTS OF 8837 06:27:07,671 --> 06:27:09,540 SPECIES, WOULD BE A WONDERFUL 8838 06:27:09,540 --> 06:27:11,475 THING, A WONDERFUL LIKE 8839 06:27:11,475 --> 06:27:15,112 LONG-TERM PROJECT FOR NEUROAI 8840 06:27:15,112 --> 06:27:16,447 AS WELL. 8841 06:27:16,447 --> 06:27:17,314 >> I MIGHT ALSO COMMENT THAT 8842 06:27:17,314 --> 06:27:19,283 TOMORROW WILL BE A GREAT TIME TO 8843 06:27:19,283 --> 06:27:23,220 BRING THAT CONVERSATION UP 8844 06:27:23,220 --> 06:27:24,121 AGAIN. 8845 06:27:24,121 --> 06:27:25,989 I THINK INVERTEBRATE SYSTEMS ARE 8846 06:27:25,989 --> 06:27:27,724 SMALL ENOUGH WE CAN BUILD THEM 8847 06:27:27,724 --> 06:27:28,692 WITH TECHNOLOGY AVAILABLE TODAY, 8848 06:27:28,692 --> 06:27:31,428 START TO LOOK AT WHAT HAPPENS IF 8849 06:27:31,428 --> 06:27:33,530 WE BUILD THE NEURO-INSPIRED, 8850 06:27:33,530 --> 06:27:35,199 ATTACH NEURO-INSPIRED SYSTEM, 8851 06:27:35,199 --> 06:27:37,167 SORRY, NEURO-INSPIRED SENSOR TO 8852 06:27:37,167 --> 06:27:37,968 THE NEURO-INSPIRED BRAIN SYSTEM. 8853 06:27:37,968 --> 06:27:41,538 SO MAYBE PUT A PIN IN THAT FOR 8854 06:27:41,538 --> 06:27:48,178 TOMORROW. 8855 06:27:48,178 --> 06:27:48,612 >> DR. HARTMAN? 8856 06:27:48,612 --> 06:27:49,947 >> NORTHWESTERN UNIVERSITY. 8857 06:27:49,947 --> 06:27:52,449 AND I STUDY SENSE OF TOUCH IN 8858 06:27:52,449 --> 06:27:53,016 RAT WHISKERS. 8859 06:27:53,016 --> 06:27:55,319 SO I WON'T BE TALKING ABOUT 8860 06:27:55,319 --> 06:27:56,186 COCKROACH ANTENNA BUT TOMORROW 8861 06:27:56,186 --> 06:27:58,789 I'LL TALK ABOUT MECHANICS OF RAT 8862 06:27:58,789 --> 06:27:59,256 WHISKERS. 8863 06:27:59,256 --> 06:28:00,791 I DID APPRECIATE THE COMMENT 8864 06:28:00,791 --> 06:28:01,792 ABOUT THE IMPORTANCE OF THE 8865 06:28:01,792 --> 06:28:04,761 SENSE OF TOUCH FOR ROBOTS. 8866 06:28:04,761 --> 06:28:06,797 AND WHILE I AGREE TOUCH IS 8867 06:28:06,797 --> 06:28:08,999 IMPORTANT FOR PROBLEMS LIKE 8868 06:28:08,999 --> 06:28:09,733 MANIPULATION OR OBJECT 8869 06:28:09,733 --> 06:28:10,834 IDENTIFICATION, I THINK THAT 8870 06:28:10,834 --> 06:28:12,603 WHAT IS REALLY IMPORTANT FOR 8871 06:28:12,603 --> 06:28:13,837 MAKING ROBOT BODIES WALK AROUND 8872 06:28:13,837 --> 06:28:16,006 IN THE WORLD, MOVE AROUND IN THE 8873 06:28:16,006 --> 06:28:19,943 WORLD, IS SENSE OF 8874 06:28:19,943 --> 06:28:20,677 PROPRIOCEPTION INTEGRATED WITH 8875 06:28:20,677 --> 06:28:24,381 MOTOR CONTROL, STILL A HARD 8876 06:28:24,381 --> 06:28:25,549 PROBLEM FOR ROBOTICISTS. 8877 06:28:25,549 --> 06:28:28,485 AND THE ISSUE REALLY ISN'T THE 8878 06:28:28,485 --> 06:28:29,553 AVAILABILITY OF SENSORY DATA, 8879 06:28:29,553 --> 06:28:37,427 FOR RIGID BODIES, WE KNOW 8880 06:28:37,427 --> 06:28:41,398 POSITIONS AND VELOCITIES. IT'S 8881 06:28:41,398 --> 06:28:43,834 DIFFERENT STORY FOR SOFT ROBOTS. 8882 06:28:43,834 --> 06:28:45,602 THE PROBLEM IS DEVELOPING 8883 06:28:45,602 --> 06:28:46,069 ADAPTIVE CONTROLLERS. 8884 06:28:46,069 --> 06:28:49,706 AND ONE OF THE BIG PROBLEMS WE 8885 06:28:49,706 --> 06:28:51,775 FACE IN ROBOTICS IS ROBOTS BREAK 8886 06:28:51,775 --> 06:28:54,945 EASILY, IT'S HARD TO GET DATA 8887 06:28:54,945 --> 06:28:57,614 FROM ROBOTS AS IT IS FROM 8888 06:28:57,614 --> 06:28:58,749 ANIMALS. 8889 06:28:58,749 --> 06:29:00,017 ROBOTICISTS OFTEN TURN TO 8890 06:29:00,017 --> 06:29:02,152 SIMULATIONS AS WE'RE TURNING TO 8891 06:29:02,152 --> 06:29:03,453 SIMULATIONS IN NEUROSCIENCE. 8892 06:29:03,453 --> 06:29:07,191 AND THEN THE PROBLEM YOU RUN 8893 06:29:07,191 --> 06:29:08,759 INTO, YOU CAN CREATE A 8894 06:29:08,759 --> 06:29:11,728 CONTROLLER THAT WORKS PERFECTLY 8895 06:29:11,728 --> 06:29:15,832 ON A SIMULATION BUT FAILS IN THE 8896 06:29:15,832 --> 06:29:16,767 REAL WORLD. 8897 06:29:16,767 --> 06:29:19,703 ONE OF THE TOOLS WE'LL NEED IN 8898 06:29:19,703 --> 06:29:23,173 THE FUTURE ARE BETTER 8899 06:29:23,173 --> 06:29:25,943 UNDERSTANDINGS OF CONTACT 8900 06:29:25,943 --> 06:29:27,711 MECHANICS, COLLISIONS, FUNDAMENTAL 8901 06:29:27,711 --> 06:29:31,148 PHYSICAL PRINCIPLES, FRICTION, 8902 06:29:31,148 --> 06:29:34,518 FLUID MECHANICS, TRIBOLOGY. WE 8903 06:29:34,518 --> 06:29:35,552 NEED MECHANICAL UNDERSTANDING 8904 06:29:35,552 --> 06:29:37,187 TO UNDERSTAND THE KINDS OF 8905 06:29:37,187 --> 06:29:39,256 SENSORY DATA AND MOTOR DATA THAT 8906 06:29:39,256 --> 06:29:45,395 THE BRAIN IS RECEIVING AND 8907 06:29:45,395 --> 06:29:46,096 CONTROLLING. 8908 06:29:46,096 --> 06:29:56,673 >> HARD TO DISAGREE WITH THAT. 8909 06:30:02,079 --> 06:30:02,713 >> YOU WERE DISCUSSING WHAT 8910 06:30:02,713 --> 06:30:04,147 WOULD BE REQUIRED WHAT WOULD BE 8911 06:30:04,147 --> 06:30:05,582 REQUIRED TO TRAIN THE NEXT 8912 06:30:05,582 --> 06:30:08,185 GENERATION OF A.I. STUDENTS, 8913 06:30:08,185 --> 06:30:09,586 PART WAS BLAKE SAYING MAYBE 8914 06:30:09,586 --> 06:30:14,591 STUDENTS NEED TO BE TRAINED UP 8915 06:30:14,591 --> 06:30:15,425 TO MATHEMATICAL FORMULAS, 8916 06:30:15,425 --> 06:30:17,761 TRAINED IN MAYBE ADVANCED 8917 06:30:17,761 --> 06:30:18,762 SYSTEMS AND COGNITIVE 8918 06:30:18,762 --> 06:30:19,129 NEUROSCIENCE. 8919 06:30:19,129 --> 06:30:21,265 RIGHT NOW LIKE THE MODEL OF HOW 8920 06:30:21,265 --> 06:30:25,402 WE DO THAT IS AD HOC. 8921 06:30:25,402 --> 06:30:26,803 UNDERGRAD TRAINING IS INFORMAL, 8922 06:30:26,803 --> 06:30:28,005 MARRY NEUROSCIENCE RELEVANT TO 8923 06:30:28,005 --> 06:30:29,106 YOUR PROJECT LATER, IF DO YOU 8924 06:30:29,106 --> 06:30:30,774 THAT PhD, if it's neuroscience 8925 06:30:30,774 --> 06:30:32,843 you catch up on formalisms later 8926 06:30:32,843 --> 06:30:33,243 on. 8927 06:30:33,243 --> 06:30:35,779 HOW DO YOU GO ABOUT TRYING TO DO 8928 06:30:35,779 --> 06:30:36,179 THAT? 8929 06:30:36,179 --> 06:30:38,215 HOW DO YOU EDUCATE A STUDENT IN 8930 06:30:38,215 --> 06:30:42,819 BOTH MATHEMATICS UP TO NONLINEAR 8931 06:30:42,819 --> 06:30:44,688 FEED BACK CONTROL BUT KNOW SUB- 8932 06:30:44,688 --> 06:30:46,657 STANTIVE NEUROSCIENCE, IF LIKE 8933 06:30:46,657 --> 06:30:48,558 UNDERGRADUATE CURRICULUM OR DO 8934 06:30:48,558 --> 06:30:50,961 WE KEEP WORKING MODEL OF TRAIN 8935 06:30:50,961 --> 06:30:54,998 EVERYBODY FAST IN Ph.D. WHAT 8936 06:30:54,998 --> 06:30:57,801 THEY ARE MISSING, OR CREATE LIKE 8937 06:30:57,801 --> 06:31:01,204 COLD SPRING HARBOR THAT IS A 8938 06:31:01,204 --> 06:31:02,139 THREE-MONTH INTERNSHIP, WHAT DO 8939 06:31:02,139 --> 06:31:04,574 YOU THINK ABOUT TRYING ACHIEVE 8940 06:31:04,574 --> 06:31:08,578 THAT KIND OF TRAINING? 8941 06:31:08,578 --> 06:31:10,314 >> I GUESS I DON'T HAVE A GOOD 8942 06:31:10,314 --> 06:31:17,421 ANSWER FOR YOUR TYPE OF TRAINING 8943 06:31:17,421 --> 06:31:20,223 BUT I WOULD, SAY I'M NOT SURE WE 8944 06:31:20,223 --> 06:31:20,857 CAN CREATE SOMEONE WELL VERSED 8945 06:31:20,857 --> 06:31:23,160 IN ALL FIELDS NECESSARY. 8946 06:31:23,160 --> 06:31:26,096 I WOULD ADVOCATE FOR MAKING SURE 8947 06:31:26,096 --> 06:31:27,331 PEOPLE HAVE BROAD ENOUGH 8948 06:31:27,331 --> 06:31:29,199 TRAINING SO THEY CAN SPECIALIZE 8949 06:31:29,199 --> 06:31:31,735 BUT NOT LOSE ABILITY TO 8950 06:31:31,735 --> 06:31:33,437 COMMUNICATE WITH OTHER FIELDS. 8951 06:31:33,437 --> 06:31:34,538 WHETHER THAT IS SOMETHING THAT 8952 06:31:34,538 --> 06:31:38,041 HAS TO BE INTERNSHIP OR GOOD 8953 06:31:38,041 --> 06:31:39,476 COLLABORATION, OR ENCOURAGING 8954 06:31:39,476 --> 06:31:40,877 MORE COLLABORATIONS OR EVEN 8955 06:31:40,877 --> 06:31:42,079 ENCOURAGING MORE WORKSHOPS LIKE 8956 06:31:42,079 --> 06:31:43,413 THIS IS AN OPEN-ENDED QUESTION. 8957 06:31:43,413 --> 06:31:45,482 MAYBE WE DON'T KNOW WHAT THE 8958 06:31:45,482 --> 06:31:46,917 RIGHT APPROACH IS. 8959 06:31:46,917 --> 06:31:50,754 BUT I DO THINK THIS ABILITY TO 8960 06:31:50,754 --> 06:31:52,589 COLLABORATE AND SEEK FREELY WITH 8961 06:31:52,589 --> 06:31:54,157 EACH OTHER AND TRADE IDEAS 8962 06:31:54,157 --> 06:31:55,325 COMING FROM DIFFERENT FIELDS, 8963 06:31:55,325 --> 06:31:57,294 MAYBE WE'RE SAYING THE SAME 8964 06:31:57,294 --> 06:31:58,528 THING WITH DIFFERENT LANGUAGE, 8965 06:31:58,528 --> 06:32:01,898 THAT'S AN IMPORTANT QUESTION TO 8966 06:32:01,898 --> 06:32:03,233 ADDRESS. 8967 06:32:03,233 --> 06:32:05,135 >> I HAVE ONE SMALL THOUGHT AS 8968 06:32:05,135 --> 06:32:08,939 WELL, JUST SORT OF RULE OF 8969 06:32:08,939 --> 06:32:09,473 THUMB. 8970 06:32:09,473 --> 06:32:11,742 IN MY EXPERIENCE, THE EARLIER 8971 06:32:11,742 --> 06:32:13,810 YOU LEARN FORMAL THEORY THE 8972 06:32:13,810 --> 06:32:14,044 BETTER. 8973 06:32:14,044 --> 06:32:15,412 NUMBER ONE. 8974 06:32:15,412 --> 06:32:19,049 NUMBER TWO, THE MORE THAT YOU 8975 06:32:19,049 --> 06:32:20,751 LEARN BIOLOGICAL THEORY FROM 8976 06:32:20,751 --> 06:32:21,852 ACTUALLY RUNNING EXPERIMENTS, 8977 06:32:21,852 --> 06:32:23,854 THE BETTER. 8978 06:32:23,854 --> 06:32:25,589 SO, I SUPPOSE MY INSTINCT FOR 8979 06:32:25,589 --> 06:32:28,759 HOW YOU WANT TO DO THIS IS YOU 8980 06:32:28,759 --> 06:32:31,828 MAKE IT SO THAT EVERYONE TAKES 8981 06:32:31,828 --> 06:32:34,264 THE CORE MATHEMATICAL COURSES IN 8982 06:32:34,264 --> 06:32:35,098 THEIR UNDERGRAD, AND EVERYONE 8983 06:32:35,098 --> 06:32:37,701 HAS TO AT SOME POINT IN THEIR 8984 06:32:37,701 --> 06:32:46,209 CAREER ACTUALLY TRY RUNNING AN 8985 06:32:46,209 --> 06:32:47,611 EXPERIMENT. 8986 06:32:47,611 --> 06:32:49,546 >> ON THE SUBJECT OF TRAINING, 8987 06:32:49,546 --> 06:32:51,581 I'VE WORKED ON THIS A LOT AND 8988 06:32:51,581 --> 06:32:53,984 THOUGHT ABOUT IT QUITE A LOT OF 8989 06:32:53,984 --> 06:32:54,317 DETAIL. 8990 06:32:54,317 --> 06:32:56,319 AND MOST RECENTLY I'VE HAD THE 8991 06:32:56,319 --> 06:33:04,461 PLEASURE OF TALKING TO ASHLEY 8992 06:33:04,461 --> 06:33:06,897 JUAVINET AT UCSD, AND I THINK WE 8993 06:33:06,897 --> 06:33:09,199 SHOULD BE HONEST WITH OURSELVES, 8994 06:33:09,199 --> 06:33:10,934 WE CAN'T JUST EXPECT STUDENTS TO 8995 06:33:10,934 --> 06:33:12,836 LEARN MORE OF EVERYTHING ALL THE 8996 06:33:12,836 --> 06:33:13,036 TIME. 8997 06:33:13,036 --> 06:33:13,437 IT'S UNREASONABLE. 8998 06:33:13,437 --> 06:33:16,339 I THINK THERE ARE SOME PARTS 8999 06:33:16,339 --> 06:33:23,713 OF CLASSICAL CURRICULUM THAT CAN 9000 06:33:23,713 --> 06:33:25,048 GO. DIFFERNTIAL EQUATIONS, ONE OF 9001 06:33:25,048 --> 06:33:26,950 MY FAVORITE PARTS OF MATHEMATICS 9002 06:33:26,950 --> 06:33:30,454 BUT WE DO NOT NEED TO TEACH 9003 06:33:30,454 --> 06:33:31,621 STUDENTS HOW TO SOLVE DERIVATIONS 9004 06:33:31,621 --> 06:33:35,492 ANALYTICALLY IN THE SAME DEPTH 9005 06:33:35,492 --> 06:33:35,859 THAT I LEARNED. 9006 06:33:35,859 --> 06:33:37,093 WE CAN TEACH PROGRAMMING IN THE 9007 06:33:37,093 --> 06:33:38,094 CONTEXT OF BIOLOGY WITHOUT 9008 06:33:38,094 --> 06:33:40,363 HAVING IT BE A COMPLETELY 9009 06:33:40,363 --> 06:33:42,399 SEPARATE CS CLASS PEOPLE HAVE TO 9010 06:33:42,399 --> 06:33:44,267 GO TO, ROLL DICE OR WHATEVER. 9011 06:33:44,267 --> 06:33:46,303 WE CAN DO THAT AND GENETICS AT 9012 06:33:46,303 --> 06:33:47,604 THE SAME TIME. 9013 06:33:47,604 --> 06:33:48,839 IT IS POSSIBLE TO NOT JUST 9014 06:33:48,839 --> 06:33:49,906 EXPECT STUDENTS TO LEARN MORE OF 9015 06:33:49,906 --> 06:33:53,743 EVERYTHING BUT TEACH IT IN A 9016 06:33:53,743 --> 06:33:54,377 FUNDAMENTALLY INTERDISCIPLINARY 9017 06:33:54,377 --> 06:33:55,312 WAY, AND THAT DOES INVOLVE 9018 06:33:55,312 --> 06:33:58,148 GETTING RID OF SOME PARTS OF THE 9019 06:33:58,148 --> 06:34:02,085 CURRICULUM THAT WE MORE 9020 06:34:02,085 --> 06:34:05,789 CLASSICICALLY TEACH, BUT IT'S 9021 06:34:05,789 --> 06:34:07,324 DOABLE. 9022 06:34:07,324 --> 06:34:14,097 I BELIEVE IT TO BE DOABLE. 9023 06:34:14,097 --> 06:34:15,232 9024 06:34:15,232 --> 06:34:17,701 >> I WANT TO ADD LIKE I'M ALL 9025 06:34:17,701 --> 06:34:19,069 FOR CORE CURRICULUM BUT WE DON'T 9026 06:34:19,069 --> 06:34:20,770 KNOW ALL THE TOOLS NEEDED TO 9027 06:34:20,770 --> 06:34:23,740 SOLVE THE BRAIN. MAYBE YOU NEED 9028 06:34:23,740 --> 06:34:27,911 TO KNOW SOME CATEGORY THEORY, 9029 06:34:27,911 --> 06:34:28,578 OTHER AREA OF KNOWLEDGE. 9030 06:34:28,578 --> 06:34:30,213 I DON'T THINK WE ALL NEED TO 9031 06:34:30,213 --> 06:34:33,283 TAKE THE SAME SET OF CLASSES. 9032 06:34:33,283 --> 06:34:35,685 >> YEAH, I ECHO THAT. 9033 06:34:35,685 --> 06:34:36,686 I THINK IT'S GOING -- I DON'T 9034 06:34:36,686 --> 06:34:41,224 THINK WE SHOULD GO FOR A ONE 9035 06:34:41,224 --> 06:34:41,725 A 9036 06:34:41,725 --> 06:34:43,260 ONE-SIZE-FITS-ALL MODEL OF 9037 06:34:43,260 --> 06:34:43,560 TRAINING. 9038 06:34:43,560 --> 06:34:44,628 PEOPLE DO HAVE DIFFERENT 9039 06:34:44,628 --> 06:34:47,864 TRAINING AND CAN ALSO WORK IN 9040 06:34:47,864 --> 06:34:49,699 THE SPACE, AND COLLABORATE AND, 9041 06:34:49,699 --> 06:34:51,268 YOU KNOW, IT'S VERY 9042 06:34:51,268 --> 06:34:55,105 INTERDISCIPLINARY FUNDAMENTALLY. 9043 06:34:55,105 --> 06:34:57,107 I WOULD ALSO SAY TO LOOK AT 9044 06:34:57,107 --> 06:34:59,609 PHYSICS, PHYSICS IS A GOOD 9045 06:34:59,609 --> 06:34:59,910 MODEL. 9046 06:34:59,910 --> 06:35:02,546 IT'S AN EXPERIMENTAL, AND, YOU 9047 06:35:02,546 --> 06:35:03,180 KNOW, THEORETICAL, COMPUTATIONAL 9048 06:35:03,180 --> 06:35:06,116 SCIENCE, SO THERE IS LIKE A CORE 9049 06:35:06,116 --> 06:35:09,252 IN SOME SENSE LIKE EVERYBODY 9050 06:35:09,252 --> 06:35:11,988 LEARNS SOME BASIC MATH, SOME 9051 06:35:11,988 --> 06:35:13,356 BASIC PHYSICS, AND THEN PEOPLE 9052 06:35:13,356 --> 06:35:14,824 GO OFF IN DIFFERENT DIRECTIONS 9053 06:35:14,824 --> 06:35:18,762 BUT THERE IS THAT, BECAUSE THERE 9054 06:35:18,762 --> 06:35:20,297 IS A SIGNIFICANT AMOUNT OF 9055 06:35:20,297 --> 06:35:21,164 MATHEMATICAL TRAINING I GUESS, 9056 06:35:21,164 --> 06:35:21,798 RIGHT? 9057 06:35:21,798 --> 06:35:23,033 WE DO HAVE THIS TRADITION THAT 9058 06:35:23,033 --> 06:35:25,302 GOES BACK A FEW HUNDREDS YEARS, 9059 06:35:25,302 --> 06:35:25,936 PHYSICISTS FORMULATING THE 9060 06:35:25,936 --> 06:35:28,872 PROBLEMS THAT THEY ARE THINKING 9061 06:35:28,872 --> 06:35:29,773 HARD ABOUT MATHEMATICALLY AND 9062 06:35:29,773 --> 06:35:31,041 ONCE YOU DO THAT LOTS OF PEOPLE 9063 06:35:31,041 --> 06:35:32,776 CAN WORK ON THEM. 9064 06:35:32,776 --> 06:35:34,644 SO I THINK THAT -- YOU KNOW, I 9065 06:35:34,644 --> 06:35:36,379 SORT OF AGREE WITH THE IDEA TO 9066 06:35:36,379 --> 06:35:37,514 HAVE A LITTLE MORE IN THE 9067 06:35:37,514 --> 06:35:39,482 BIOLOGY TRAINING, JUST TO BE 9068 06:35:39,482 --> 06:35:42,519 ABLE TO SORT OF WRITE DOWN YOUR 9069 06:35:42,519 --> 06:35:44,187 IDEAS OR FORMULATE PROBLEMS, 9070 06:35:44,187 --> 06:35:47,557 QUESTIONS, IN A WAY, IT'S REALLY 9071 06:35:47,557 --> 06:35:48,224 A COMMUNICATION THING, IRONIC 9072 06:35:48,224 --> 06:35:50,727 BECAUSE PEOPLE THINK OF MATH AS 9073 06:35:50,727 --> 06:35:52,963 A BARRIER TO COMMUNICATION BUT I 9074 06:35:52,963 --> 06:35:55,599 OBVIOUSLY THINK ABOUT IT AS A 9075 06:35:55,599 --> 06:35:59,436 WAY OF CLARIFYING AND IMPROVING 9076 06:35:59,436 --> 06:36:00,103 COMMUNICATION. 9077 06:36:00,103 --> 06:36:01,905 >> I HAVE SOME EXPERIENCE IN 9078 06:36:01,905 --> 06:36:03,473 THIS AREA ALSO. 9079 06:36:03,473 --> 06:36:06,376 AND HAVE DISCOVERED THAT TO 9080 06:36:06,376 --> 06:36:09,412 TEACH MATHEMATICS TO BIOLOGY, 9081 06:36:09,412 --> 06:36:10,847 NEUROSCIENCE, COGNITIVE SCIENCE, 9082 06:36:10,847 --> 06:36:13,183 PSYCHOLOGY STUDENTS AND 9083 06:36:13,183 --> 06:36:13,783 INCREASINGLY COMPUTER SCIENCE 9084 06:36:13,783 --> 06:36:15,752 STUDENTS, THEY TURN OFF IF YOU 9085 06:36:15,752 --> 06:36:18,555 GIVE THEM A CLASSICAL 9086 06:36:18,555 --> 06:36:20,457 PRESENTATION. 9087 06:36:20,457 --> 06:36:23,693 SO, I'VE CREATED A SERIES OF 9088 06:36:23,693 --> 06:36:26,863 ARTIFICIAL ORGANISMS LIKE 9089 06:36:26,863 --> 06:36:28,164 ARTIFICIAL COLI TO TEACH 9090 06:36:28,164 --> 06:36:32,769 GRADIENT DESCENT. MY FAVORITE 9091 06:36:32,769 --> 06:36:35,939 CREATURE IS LIMULUS LINEARIS, 9092 06:36:35,939 --> 06:36:38,208 PROVIDES A GLIMPSE OF LINEAR 9093 06:36:38,208 --> 06:36:39,576 OPERATOR THEORY IN A WAY THAT IF 9094 06:36:39,576 --> 06:36:41,911 YOU TOLD THEM THIS IS THE 9095 06:36:41,911 --> 06:36:46,750 CONVOLUTION THEOREM THIS IS 9096 06:36:46,750 --> 06:36:48,818 HILBERT SPACE, THEY ARE GONE BUT 9097 06:36:48,818 --> 06:36:50,687 STUDENTS LOVE PLAYING WITH THESE 9098 06:36:50,687 --> 06:36:51,054 CREATURES. 9099 06:36:51,054 --> 06:36:53,790 >> THE LESSON IS IT'S NOT JUST 9100 06:36:53,790 --> 06:36:56,026 PUTTING ONUS ON STUDENTS TO SEEK 9101 06:36:56,026 --> 06:36:57,160 ALL THINGS AND BECOME WORLD 9102 06:36:57,160 --> 06:36:58,595 EXPERTS ON ALL THINGS BUT ALSO 9103 06:36:58,595 --> 06:37:01,364 ON INSTRUCTORS TO COME UP WITH 9104 06:37:01,364 --> 06:37:06,703 CREATIVE SOLUTIONS PERHAPS 9105 06:37:06,703 --> 06:37:11,107 LEVERAGING NEW A.I. TOOLS. 9106 06:37:11,107 --> 06:37:16,112 PLEASE PLUG THE NIH TO DO THIS 9107 06:37:16,112 --> 06:37:18,782 ANNUALLY ALSO. 9108 06:37:18,782 --> 06:37:20,950 >> YEAH, I WANT TO ECHO BEFORE I 9109 06:37:20,950 --> 06:37:25,789 MOVE ON BACK TO THE QUESTION, I 9110 06:37:25,789 --> 06:37:29,125 THINK ABOUT MAYBE SOME OF THE 9111 06:37:29,125 --> 06:37:35,532 PROGRAMS BY THE WAY, MINDY SHI, 9112 06:37:35,532 --> 06:37:35,932 TEMPLE UNIVERSITY. 9113 06:37:35,932 --> 06:37:37,200 MAYBE MOVING FORWARD IN FIVE, 9114 06:37:37,200 --> 06:37:39,069 TEN YEARS IN THIS COMMUNITY, WE 9115 06:37:39,069 --> 06:37:41,905 CAN MOVE FORWARD WITH SOMETHING 9116 06:37:41,905 --> 06:37:42,772 LIKE COMPUTATION NEUROSCIENCE. 9117 06:37:42,772 --> 06:37:47,210 I THINK THERE'S SOMETHING THERE. 9118 06:37:47,210 --> 06:37:49,679 ALSO COMBINING DATA SCIENCE, 9119 06:37:49,679 --> 06:37:53,683 A.I., SOMETHING LIKE KIND OF 9120 06:37:53,683 --> 06:37:55,552 JOINT TRAINING PROGRAMS WOULD BE 9121 06:37:55,552 --> 06:37:56,753 HELPFUL, THINK ABOUT THEORIES OF 9122 06:37:56,753 --> 06:37:59,422 CURRICULUM THAT WE CAN PUT 9123 06:37:59,422 --> 06:38:00,790 TOGETHER COMBINING WHAT WE HAVE 9124 06:38:00,790 --> 06:38:02,325 JUST TALK ABOUT, I THINK THERE 9125 06:38:02,325 --> 06:38:07,397 MIGHT BE SOMETHING THAT WE CAN 9126 06:38:07,397 --> 06:38:10,533 PLAN AS COMMUNITY TO MOVE THIS 9127 06:38:10,533 --> 06:38:12,669 MOST SYSTEMATICALLY TOWARD THE 9128 06:38:12,669 --> 06:38:14,738 TRAINING AND MENTORING AND ALSO 9129 06:38:14,738 --> 06:38:17,240 COLLABORATIONS AS WELL. 9130 06:38:17,240 --> 06:38:18,908 SO GETTING BACK TO THE QUESTION 9131 06:38:18,908 --> 06:38:22,879 THAT I HAVE, LISTENING THROUGH 9132 06:38:22,879 --> 06:38:25,381 ALL THE SESSIONS, I JUST WANT TO 9133 06:38:25,381 --> 06:38:29,953 KIND OF LIKE COMMENT, ALSO LOOK 9134 06:38:29,953 --> 06:38:32,789 FOR INPUT FOR THIS IDEA IN MY 9135 06:38:32,789 --> 06:38:34,758 RESEARCH AND ALSO MY GROUP AND 9136 06:38:34,758 --> 06:38:38,895 MANY OTHERS WORK IN MACHINE 9137 06:38:38,895 --> 06:38:39,562 LEARNING AND FUNDAMENTAL 9138 06:38:39,562 --> 06:38:41,030 OPTIMIZATION, FUNDAMENTAL A.I. 9139 06:38:41,030 --> 06:38:44,400 I THINK IT'S NOT NEW IN THE 9140 06:38:44,400 --> 06:38:45,869 SENSE THAT WE'RE TALKING ABOUT 9141 06:38:45,869 --> 06:38:47,137 DATA COLLECTION. 9142 06:38:47,137 --> 06:38:49,873 WE'RE TALKING ABOUT EVALUATION 9143 06:38:49,873 --> 06:38:50,240 METRICS. 9144 06:38:50,240 --> 06:38:51,274 BENCHMARKS. 9145 06:38:51,274 --> 06:38:56,346 WE'RE ALSO TALKING ABOUT MODELS 9146 06:38:56,346 --> 06:38:58,715 MOVING FORWARD FROM LIKE UNI- 9147 06:38:58,715 --> 06:38:59,482 MODAL, MULTI-MODAL, EVEN JUST 9148 06:38:59,482 --> 06:39:01,484 MULTI-AGENT, SO ALL THESE IDEAS 9149 06:39:01,484 --> 06:39:02,452 ARE GREAT. 9150 06:39:02,452 --> 06:39:04,654 SO I'M JUST THINKING ABOUT ONE 9151 06:39:04,654 --> 06:39:07,056 THING THAT WE ACTUALLY STRUCK, 9152 06:39:07,056 --> 06:39:11,194 SOMETHING I THINK THAT HAS A 9153 06:39:11,194 --> 06:39:14,264 CONNECTION ABOUT BIAS AND NO 9154 06:39:14,264 --> 06:39:16,499 MATTER WHAT YOU PUT INTO IT, 9155 06:39:16,499 --> 06:39:18,568 HEALTH DISPARITIES, BIAS IN YOUR 9156 06:39:18,568 --> 06:39:20,970 MODELS, AND IF WE -- FOR FOLKS 9157 06:39:20,970 --> 06:39:23,139 WHO WORK ON FUNDAMENTAL A.I. 9158 06:39:23,139 --> 06:39:27,177 LIKE THE GENERATION GAP THAT YOU 9159 06:39:27,177 --> 06:39:31,080 SEE, IN YOUR TRAINING DATA, 9160 06:39:31,080 --> 06:39:31,314 RIGHT? 9161 06:39:31,314 --> 06:39:33,650 SO IF YOU THINK ABOUT THIS 9162 06:39:33,650 --> 06:39:35,385 COMMUNITY, YOU HAVE VERY NICE 9163 06:39:35,385 --> 06:39:37,487 FRAMEWORK MAYBE TO TOUCH UPON 9164 06:39:37,487 --> 06:39:39,522 THE ISSUES THAT WOULD HELP A LOT 9165 06:39:39,522 --> 06:39:41,858 OF FOLKS WORK ON METHOD 9166 06:39:41,858 --> 06:39:44,360 DEVELOPMENT AND THEN FEEDBACK TO 9167 06:39:44,360 --> 06:39:47,664 BETTER UNDERSTAND THE BRAIN, THE 9168 06:39:47,664 --> 06:39:49,532 BIOLOGY, WOULD -- - HOW DO YOU 9169 06:39:49,532 --> 06:39:52,101 THINK ABOUT CREATING DATASETS 9170 06:39:52,101 --> 06:39:57,707 THAT YOU KIND OF LIKE SAMPLE THE 9171 06:39:57,707 --> 06:39:58,741 BRAIN ACTIVITIES, WHATEVER DATA 9172 06:39:58,741 --> 06:40:01,411 ENTITIES THAT WE CAN COLLECT AND 9173 06:40:01,411 --> 06:40:03,112 FEATURES WE CAN COLLECT SO THAT 9174 06:40:03,112 --> 06:40:07,717 WE CAN CREATE A BENCHMARK, 9175 06:40:07,717 --> 06:40:10,253 CREATE DATASETS THAT WE CAN FIND 9176 06:40:10,253 --> 06:40:11,588 THE DISTRIBUTIONS OF THE DATA OR 9177 06:40:11,588 --> 06:40:14,524 PARAMETER SPACE THAT WE HAVE ON 9178 06:40:14,524 --> 06:40:17,260 THE TRAINING DATASETS WOULD BE 9179 06:40:17,260 --> 06:40:18,862 CLOSE ENOUGH OR TO THE 9180 06:40:18,862 --> 06:40:21,297 SIMULATION DATA, TO THE TESTING 9181 06:40:21,297 --> 06:40:22,198 DATASET, WHATEVER WE HAVE, SO 9182 06:40:22,198 --> 06:40:24,934 THAT WE CAN BUILD SOME MODELS 9183 06:40:24,934 --> 06:40:26,603 THAT ACTUALLY ARE USEFUL AND 9184 06:40:26,603 --> 06:40:29,105 WHEN WE PUT THEM INTO 9185 06:40:29,105 --> 06:40:30,807 IMPLEMENTATION, INTO PRACTICE, 9186 06:40:30,807 --> 06:40:32,408 THAT WE DON'T SEE THAT MUCH GAP 9187 06:40:32,408 --> 06:40:35,478 IN THE PERFORMANCE OF THE 9188 06:40:35,478 --> 06:40:37,146 MODELS, NO MATTER WHAT YOU ARE 9189 06:40:37,146 --> 06:40:39,082 DEALING WITH PREDICTION, 9190 06:40:39,082 --> 06:40:40,783 FOUNDATION MODELS, TRANSFER OF 9191 06:40:40,783 --> 06:40:43,887 LEARNING, SO ON AND SO FORTH, TO 9192 06:40:43,887 --> 06:40:48,558 MY -- FROM MY PERSPECTIVE VERY 9193 06:40:48,558 --> 06:40:50,526 FUNDAMENTAL PROBLEMS IN MANY 9194 06:40:50,526 --> 06:40:52,595 FIELDS, NEUROSCIENCE, NEUROAI 9195 06:40:52,595 --> 06:40:54,397 HOPEFULLY THAT YOU HAVE SOME 9196 06:40:54,397 --> 06:40:56,933 INPUT OR FEEDBACK CAN IMPROVE 9197 06:40:56,933 --> 06:40:58,434 THIS. 9198 06:40:58,434 --> 06:41:07,744 YEAH, THANKS. 9199 06:41:07,744 --> 06:41:08,645 9200 06:41:08,645 --> 06:41:11,347 >> YEAH, FULLY AGREE. 9201 06:41:11,347 --> 06:41:13,216 I THINK THIS WOULD WORK, I THINK 9202 06:41:13,216 --> 06:41:16,552 ONE CLEAR PATH FOR THE FUTURE IS 9203 06:41:16,552 --> 06:41:17,854 TO BRING DIFFERENT ALSO 9204 06:41:17,854 --> 06:41:20,657 FOUNDATION MODELS FROM DIFFERENT 9205 06:41:20,657 --> 06:41:23,526 FIELDS, COMBINED TOGETHER, AND 9206 06:41:23,526 --> 06:41:24,327 THEN ALSO HAVE APPLICATIONS 9207 06:41:24,327 --> 06:41:26,362 BEYOND JUST THE THEME HERE IS 9208 06:41:26,362 --> 06:41:30,733 ABOUT NEUROSCIENCE FOR A.I. BUT 9209 06:41:30,733 --> 06:41:32,602 ALSO OTHER APPLICATIONS. 9210 06:41:32,602 --> 06:41:35,071 >> IF I MAY, PART OF IT IS ALSO 9211 06:41:35,071 --> 06:41:35,338 CULTURAL. 9212 06:41:35,338 --> 06:41:37,206 WE'RE NOT REALLY USED TO 9213 06:41:37,206 --> 06:41:40,643 THINKING LIKE YOU ALL DO, AS 9214 06:41:40,643 --> 06:41:41,778 NEUROSCIENTISTS, TO CREATE 9215 06:41:41,778 --> 06:41:42,078 BENCHMARKS. 9216 06:41:42,078 --> 06:41:47,183 IT'S STILL NEW -- STILL A NEW 9217 06:41:47,183 --> 06:41:47,383 THING. 9218 06:41:47,383 --> 06:41:50,019 THANKS TO THE BRAIN INITIATIVE 9219 06:41:50,019 --> 06:41:54,624 WE'VE BEEN THINKING ABOUT 9220 06:41:54,624 --> 06:41:56,793 DISSEMINATION, SO RIGHT NOW IS 9221 06:41:56,793 --> 06:41:58,861 EXACTLY THE KIND OF THING 9222 06:41:58,861 --> 06:42:00,930 ANDREAS IS TALKING ABOUT, 9223 06:42:00,930 --> 06:42:02,665 FOUNDATION MODELS BLAKE AND 9224 06:42:02,665 --> 06:42:03,900 GROUP ALSO, AND PATRICK TALKED 9225 06:42:03,900 --> 06:42:05,802 ABOUT IN THE MORNING. 9226 06:42:05,802 --> 06:42:07,103 ALL OF THOSE ARE IN THE 9227 06:42:07,103 --> 06:42:09,505 DIRECTION THAT YOU WERE TALKING 9228 06:42:09,505 --> 06:42:14,644 ABOUT SO KEEPING AN EYE ON THOSE 9229 06:42:14,644 --> 06:42:15,311 PAPERS. 9230 06:42:15,311 --> 06:42:15,878 >> PAMELA ABSHIRE, UNIVERSITY 9231 06:42:15,878 --> 06:42:17,313 OF MARYLAND. AS WE START 9232 06:42:17,313 --> 06:42:20,717 THINKING HOW TO TRAIN A NEW 9233 06:42:20,717 --> 06:42:22,185 GENERATION, AND THERE'S SO MUCH 9234 06:42:22,185 --> 06:42:23,853 MORE STUFF THAT WE NEED TO PILE 9235 06:42:23,853 --> 06:42:26,990 INTO PEOPLE'S HEADS. 9236 06:42:26,990 --> 06:42:28,725 LET'S NOT LOSE THE JOY OF 9237 06:42:28,725 --> 06:42:30,259 EXPLORATION THAT DRIVES ALL THE 9238 06:42:30,259 --> 06:42:31,894 PEOPLE IN THIS ROOM SO LET'S 9239 06:42:31,894 --> 06:42:35,431 LEAVE SOME SPACE WITHIN THAT 9240 06:42:35,431 --> 06:42:37,233 EDUCATIONAL PROCESS FOR 9241 06:42:37,233 --> 06:42:38,601 EXPLORATION AND DISCOVERY AND 9242 06:42:38,601 --> 06:42:39,535 EXPERIENTIAL LEARNING SO PEOPLE 9243 06:42:39,535 --> 06:42:40,770 CAN WORK ON SOME OF THESE 9244 06:42:40,770 --> 06:42:42,739 PROBLEMS AND GET A REAL SENSE OF 9245 06:42:42,739 --> 06:42:43,139 IT. 9246 06:42:43,139 --> 06:42:48,211 MAN, I WANT TO PLAY WITH STEVE'S 9247 06:42:48,211 --> 06:42:48,811 BUGS. 9248 06:42:48,811 --> 06:42:51,347 SO THE COMMENT IS SIMILAR TO THE 9249 06:42:51,347 --> 06:42:54,083 LAST ONE, HOW DO WE -- I HAVE A 9250 06:42:54,083 --> 06:42:54,717 QUESTION. 9251 06:42:54,717 --> 06:42:57,820 THE QUESTION IS HOW DO WE AVOID 9252 06:42:57,820 --> 06:43:00,490 GETTING TRAPPED IN OUR OWN 9253 06:43:00,490 --> 06:43:01,391 EXPERIMENTAL BIAS, BOTH IN 9254 06:43:01,391 --> 06:43:02,859 TERMS -- AS WE BUILD THESE 9255 06:43:02,859 --> 06:43:04,861 MODELS HOW DO WE AVOID GETTING 9256 06:43:04,861 --> 06:43:06,429 TRAPPED IN BIAS OF OUR 9257 06:43:06,429 --> 06:43:08,197 EXPERIMENTAL CONSTRUCT IN TERMS 9258 06:43:08,197 --> 06:43:09,332 OF THE NEUROPHYSIOLOGICAL 9259 06:43:09,332 --> 06:43:10,967 EXPERIMENTS AND ALSO IN TERMS OF 9260 06:43:10,967 --> 06:43:12,402 THE A.I. MODELS THAT WE HAVE TO 9261 06:43:12,402 --> 06:43:13,136 DESCRIBE THEM. 9262 06:43:13,136 --> 06:43:15,204 MAYBE THIS HARKENS BACK TO WE 9263 06:43:15,204 --> 06:43:18,241 NEED BETTER MATH BUT I THINK 9264 06:43:18,241 --> 06:43:19,175 IT'S BIGGER THAN THAT BECAUSE 9265 06:43:19,175 --> 06:43:21,477 LIKE THE REAL ISSUE IS GOING 9266 06:43:21,477 --> 06:43:22,712 BACK TO THE FIRST PRESENTATION 9267 06:43:22,712 --> 06:43:26,315 TODAY, WE DON'T WANT TO MISTAKE 9268 06:43:26,315 --> 06:43:27,350 CHIHUAHUAS FOR BLUEBERRY 9269 06:43:27,350 --> 06:43:29,719 MUFFINS, GET TRAPPED INTO THESE 9270 06:43:29,719 --> 06:43:30,953 ARTIFACTS AND JUST LOOKING AT 9271 06:43:30,953 --> 06:43:34,090 THE TEXTURE OF THE SKIN OF THE 9272 06:43:34,090 --> 06:43:36,059 ELEPHANT, NOT TELLING IT'S A 9273 06:43:36,059 --> 06:43:39,462 TAIL OR TRUNK OR WHATEVER, 9274 06:43:39,462 --> 06:43:40,763 RIGHT? 9275 06:43:40,763 --> 06:43:43,533 NO MATTER -- GIVEN THE SECOND 9276 06:43:43,533 --> 06:43:45,401 PRESENTATION HE MADE, NO MATTER 9277 06:43:45,401 --> 06:43:46,803 HOW MANY PETABYTES OF DATA YOU 9278 06:43:46,803 --> 06:43:48,704 COLLECT YOU'RE NOT GOING TO 9279 06:43:48,704 --> 06:43:50,139 EVER -- YOU'RE ALWAYS GOING TO 9280 06:43:50,139 --> 06:43:52,108 BE IN A DATA-POOR SITUATION FOR 9281 06:43:52,108 --> 06:43:53,876 LOOKING AT THE CORRELATIONS 9282 06:43:53,876 --> 06:43:56,245 BETWEEN ALL THOSE DIFFERENT 9283 06:43:56,245 --> 06:43:57,046 VARIABLES, YOU'LL HAVE 9284 06:43:57,046 --> 06:43:58,214 RELATIVELY FEW EXAMPLES FOR 9285 06:43:58,214 --> 06:43:59,515 LOOKING FOR THE REAL 9286 06:43:59,515 --> 06:44:00,950 CORRELATIONS IN THE REAL 9287 06:44:00,950 --> 06:44:04,687 UNDERLYING FACTORS WHEN WE PUT 9288 06:44:04,687 --> 06:44:05,688 IT TOGETHER. 9289 06:44:05,688 --> 06:44:09,625 HOW DO WE AVOID GETTING TRAPPED 9290 06:44:09,625 --> 06:44:11,861 INTO OUR OWN EXPERIMENTAL BIAS 9291 06:44:11,861 --> 06:44:13,696 FOR TRYING TO REALLY FIND HOW 9292 06:44:13,696 --> 06:44:16,766 THESE THINGS ARE WORKING, YOU 9293 06:44:16,766 --> 06:44:18,901 KNOW, I'LL END WITH WHAT I THINK 9294 06:44:18,901 --> 06:44:20,369 MAY BE CONTROVERSIAL, I DON'T 9295 06:44:20,369 --> 06:44:21,204 KNOW. 9296 06:44:21,204 --> 06:44:24,173 WHAT I'M TAKING AWAY FROM LARGE 9297 06:44:24,173 --> 06:44:25,374 LANGUAGE MODEL LEARNING AND 9298 06:44:25,374 --> 06:44:26,876 PERFORMANCE THAT WE'VE SEEN 9299 06:44:26,876 --> 06:44:28,711 RECENTLY ISN'T SO MUCH THAT 9300 06:44:28,711 --> 06:44:30,580 ARTIFICIAL INTELLIGENCE IS 9301 06:44:30,580 --> 06:44:31,047 GREAT. 9302 06:44:31,047 --> 06:44:32,381 IT'S, WOW, WHY DIDN'T WE 9303 06:44:32,381 --> 06:44:34,350 UNDERSTAND EARLIER THAT LANGUAGE 9304 06:44:34,350 --> 06:44:36,285 IS A LOOKUP TABLE, RIGHT? 9305 06:44:36,285 --> 06:44:38,054 I THINK WE'RE LEARNING MORE 9306 06:44:38,054 --> 06:44:39,255 ABOUT THE NATURE OF THE PROBLEM 9307 06:44:39,255 --> 06:44:40,890 THAN WE ARE ABOUT THE NATURE OF 9308 06:44:40,890 --> 06:44:43,626 INTELLIGENCE WITH THE TOOLS THAT 9309 06:44:43,626 --> 06:44:49,699 WE HAVE. 9310 06:44:49,699 --> 06:44:53,636 >> I SAW DORIS SMALLING, ANDREAS 9311 06:44:53,636 --> 06:44:55,972 SMILING AT DIFFERENT PARTS OF 9312 06:44:55,972 --> 06:45:02,645 THE QUESTION. 9313 06:45:02,645 --> 06:45:07,984 >> EVERYTHING NON-TRIVIAL. 9314 06:45:07,984 --> 06:45:09,519 YOU GO IN WITH HYPOTHESIS, 9315 06:45:09,519 --> 06:45:14,991 YOU'RE PROVED WRONG BY THE DATA. 9316 06:45:14,991 --> 06:45:17,493 >> AND TO ADD TO THAT WHEN YOU 9317 06:45:17,493 --> 06:45:19,829 BUILD THESE MODELS, YOU DO 9318 06:45:19,829 --> 06:45:20,763 CLOSED LOOP EXPERIMENTS, SO THE 9319 06:45:20,763 --> 06:45:24,834 IDEA IS THAT YOU WILL GENERATE 9320 06:45:24,834 --> 06:45:25,968 SOME HYPOTHESES, SOME STIMULI, 9321 06:45:25,968 --> 06:45:27,436 TEST THEM BACK. 9322 06:45:27,436 --> 06:45:29,105 YOU MAY BE RIGHT THAT HOW MUCH 9323 06:45:29,105 --> 06:45:32,041 DATA YOU NEED MAYBE WE DON'T 9324 06:45:32,041 --> 06:45:33,576 HAVE ENOUGH DATA. 9325 06:45:33,576 --> 06:45:36,979 YOU CAN TEST, YOU CAN BUILD THE 9326 06:45:36,979 --> 06:45:38,614 MODEL, DO SOME DISCOVERY OR 9327 06:45:38,614 --> 06:45:40,816 ANALYZE THE MODEL, YOU FIND 9328 06:45:40,816 --> 06:45:45,354 SOMETHING, TEST IT BACK TO 9329 06:45:45,354 --> 06:45:45,922 CLOSED LOOP EXPERIMENTS. 9330 06:45:45,922 --> 06:45:47,256 WITHOUT THAT YOU MAY BE DOING 9331 06:45:47,256 --> 06:45:56,032 SOME KIND OF FANCY LOOKUP TABLE. 9332 06:45:56,032 --> 06:45:57,466 BETTER IN LARGE LANGUAGE MODEL 9333 06:45:57,466 --> 06:46:00,403 BUT YEAH, I THINK IT'S 9334 06:46:00,403 --> 06:46:04,774 THROUGH -- YOU NEED THIS TIGHT 9335 06:46:04,774 --> 06:46:05,174 LINK WITH EXPERIMENTS. 9336 06:46:05,174 --> 06:46:07,076 >> I MIGHT MAKE A COMMENT. 9337 06:46:07,076 --> 06:46:09,245 MAYBE I HEARD YOUR QUESTION 9338 06:46:09,245 --> 06:46:09,679 DIFFERENTLY. 9339 06:46:09,679 --> 06:46:12,582 YOU'RE ALMOST ASKING HOW DO YOU 9340 06:46:12,582 --> 06:46:14,517 AVOID GETTING OVERFIT TO YOUR 9341 06:46:14,517 --> 06:46:15,718 SPECIFIC EXPERIMENTAL SYSTEM. 9342 06:46:15,718 --> 06:46:16,686 THE ANSWER IS COLLABORATIONS 9343 06:46:16,686 --> 06:46:18,888 WITH PEOPLE WHO AREN'T IN YOUR 9344 06:46:18,888 --> 06:46:19,589 SPECIFIC SYSTEM. 9345 06:46:19,589 --> 06:46:22,792 IT MAY BE THAT YOUR SYSTEM HAS A 9346 06:46:22,792 --> 06:46:24,527 PARTICULAR ANSWER, BUT IN TERMS 9347 06:46:24,527 --> 06:46:25,461 OF IMPACTING A.I., TALKING TO 9348 06:46:25,461 --> 06:46:27,763 THE A.I. PEOPLE AND SAYING MAYBE 9349 06:46:27,763 --> 06:46:29,332 THIS ISN'T THE MOST USEFUL 9350 06:46:29,332 --> 06:46:35,104 SITUATION FOR YOUR SPECIFIC 9351 06:46:35,104 --> 06:46:35,504 MODEL. 9352 06:46:35,504 --> 06:46:41,010 WE CAN LEARN FROM THOSE 9353 06:46:41,010 --> 06:46:41,644 COLLABORATIONS AND DISCUSSIONS. 9354 06:46:41,644 --> 06:46:46,616 >> SO, I HAVE A GENERAL COMMENT. 9355 06:46:46,616 --> 06:46:52,788 RAMANI DURAISWAMI, UNIVERSITY OF 9356 06:46:52,788 --> 06:46:55,524 MARYLAND. SO, WHEN A.I. -- THE 9357 06:46:55,524 --> 06:46:59,028 FASHION IN A.I. KEEPS CHANGING. 9358 06:46:59,028 --> 06:47:02,331 WE HAD MLPs, CONVOLUTIONAL 9359 06:47:02,331 --> 06:47:03,065 NEURAL NETWORKS, TRANSFORMERS, 9360 06:47:03,065 --> 06:47:05,034 NOW PERHAPS STATE-SPACE MODELS, 9361 06:47:05,034 --> 06:47:06,235 ET CETERA. 9362 06:47:06,235 --> 06:47:07,303 EVERYBODY GOES ALONG WITH THAT. 9363 06:47:07,303 --> 06:47:10,606 ON THE OTHER HAND NEURAL SYSTEMS 9364 06:47:10,606 --> 06:47:14,610 SEEM TO BE SYSTEMS DEVELOPED FOR 9365 06:47:14,610 --> 06:47:17,413 TASK-BASED SOLVING OF PROBLEMS. 9366 06:47:17,413 --> 06:47:20,049 YOU WANT TO NAVIGATE THROUGH THE 9367 06:47:20,049 --> 06:47:20,850 WORLD, RECOGNIZE SOMETHING, ET 9368 06:47:20,850 --> 06:47:24,053 CETERA, ET CETERA. 9369 06:47:24,053 --> 06:47:28,724 AND THE ARCHITECTURES ARE 9370 06:47:28,724 --> 06:47:30,259 PERHAPS NOT UNIFORM ACROSS THE 9371 06:47:30,259 --> 06:47:32,361 DIFFERENT SYSTEMS, THERE MIGHT 9372 06:47:32,361 --> 06:47:33,763 BE SOME COMMONALITY BUT THERE 9373 06:47:33,763 --> 06:47:35,531 ARE ALSO SIGNIFICANT 9374 06:47:35,531 --> 06:47:35,865 DIFFERENCES. 9375 06:47:35,865 --> 06:47:39,335 AND IN THE TREATMENT OF DISEASE 9376 06:47:39,335 --> 06:47:42,672 WHAT HAPPENS IS THERE ARE 9377 06:47:42,672 --> 06:47:44,307 SIGNIFICANT VARIATIONS AT THE 9378 06:47:44,307 --> 06:47:49,945 INDIVIDUAL LEVEL, SO IN SOME 9379 06:47:49,945 --> 06:47:53,883 SENSE, THE TWO SYSTEMS WHILE 9380 06:47:53,883 --> 06:47:56,652 HAVING SOME ERROR-DRIVEN 9381 06:47:56,652 --> 06:47:57,787 OPTIMIZATION, WILL BE DIFFERENT 9382 06:47:57,787 --> 06:48:00,022 ENOUGH ALL THE TIME IN THE SENSE 9383 06:48:00,022 --> 06:48:04,527 THAT THE NEURAL SYSTEMS WILL 9384 06:48:04,527 --> 06:48:06,128 HAVE VARIATIONS WHICH ARE -- AND 9385 06:48:06,128 --> 06:48:08,130 VARIATIONS WILL GO DOWN TO THE 9386 06:48:08,130 --> 06:48:10,032 INDIVIDUAL LEVEL. 9387 06:48:10,032 --> 06:48:12,935 SO, ANY SORT OF TREATMENT, 9388 06:48:12,935 --> 06:48:15,871 ANYTHING ELSE WOULD REQUIRE 9389 06:48:15,871 --> 06:48:17,840 ALMOST PERSONALIZED MODELS FOR 9390 06:48:17,840 --> 06:48:22,078 EVERY ONE OF THE 7.5 BILLION 9391 06:48:22,078 --> 06:48:23,646 PEOPLE OR WHATEVER, AND SO, IN 9392 06:48:23,646 --> 06:48:25,715 SOME SENSE I THINK WE SHOULD -- 9393 06:48:25,715 --> 06:48:27,283 AND SOMEONE ELSE HAD ASKED THE 9394 06:48:27,283 --> 06:48:28,884 QUESTION WE SHOULDN'T LOSE SIGHT 9395 06:48:28,884 --> 06:48:31,354 OF DISEASE AND SO FORTH. 9396 06:48:31,354 --> 06:48:35,891 I THINK PART OF THAT COMES FROM 9397 06:48:35,891 --> 06:48:39,261 HAVING THE ABILITY IN MODELS TO 9398 06:48:39,261 --> 06:48:41,397 HAVE THIS KIND OF VARIABILITY 9399 06:48:41,397 --> 06:48:42,932 AND ABILITY TO FIT THE 9400 06:48:42,932 --> 06:48:46,168 INDIVIDUAL KIND OF A THING. 9401 06:48:46,168 --> 06:48:47,403 >> IF I MIGHT COMMENT ON THAT SO 9402 06:48:47,403 --> 06:48:50,506 ONE OF THE REASONS I'M A LARGE 9403 06:48:50,506 --> 06:48:51,907 ADVOCATE FOR FOUNDATION MODEL 9404 06:48:51,907 --> 06:48:53,909 APPROACH IN NEUROSCIENCE IS 9405 06:48:53,909 --> 06:48:55,644 BECAUSE WHAT FOUNDATION MODELS 9406 06:48:55,644 --> 06:48:58,647 BRING YOU IS THE ABILITY TO FINE 9407 06:48:58,647 --> 06:49:00,116 TUNE VARIOUS SMALL AMOUNTS OF 9408 06:49:00,116 --> 06:49:00,316 DATA. 9409 06:49:00,316 --> 06:49:01,183 I AGREE WITH YOU. 9410 06:49:01,183 --> 06:49:04,787 YOU WILL NEED PERSONALIZED 9411 06:49:04,787 --> 06:49:06,355 MODELS TO EXPLAIN WHAT'S GOING 9412 06:49:06,355 --> 06:49:07,690 ON FOR SOMEONE WHO IS SUFFERING 9413 06:49:07,690 --> 06:49:10,059 FROM, SAY, A MENTAL HEALTH 9414 06:49:10,059 --> 06:49:10,793 DISORDER. 9415 06:49:10,793 --> 06:49:14,630 BUT THE WHOLE POINT IS THAT IF 9416 06:49:14,630 --> 06:49:15,731 YOU DO SUFFICIENTLY POWERFUL 9417 06:49:15,731 --> 06:49:17,233 PRE-TRAINING ON A LARGE ENOUGH 9418 06:49:17,233 --> 06:49:19,835 DATASET, YOU WILL THEN BE ABLE 9419 06:49:19,835 --> 06:49:21,404 TO FINE TUNE AND INDIVIDUALIZE 9420 06:49:21,404 --> 06:49:23,606 THAT MODEL FOR YOUR INDIVIDUAL 9421 06:49:23,606 --> 06:49:25,508 WITH ACTUALLY REALISTIC AMOUNT 9422 06:49:25,508 --> 06:49:29,779 OF RECORDING FROM THEM. 9423 06:49:29,779 --> 06:49:32,848 >> IF I MAY, SO ONE THOUGHT, THE 9424 06:49:32,848 --> 06:49:36,585 FOUNDATIONAL MODELS BY THE VERY 9425 06:49:36,585 --> 06:49:39,622 DEFINITION ARE EXTREMELY LARGE. 9426 06:49:39,622 --> 06:49:43,692 AND THAT IN SOME SENSE YOU'RE 9427 06:49:43,692 --> 06:49:46,662 SORT OF HAVING TO OVERCOMPLICATE 9428 06:49:46,662 --> 06:49:48,964 THINGS TO SIMPLIFY LATER. 9429 06:49:48,964 --> 06:49:49,999 ANYWAY, THAT'S JUST -- 9430 06:49:49,999 --> 06:49:51,801 >> YEAH, THAT MIGHT JUST BE THE 9431 06:49:51,801 --> 06:49:58,140 SITUATION WE FIND OURSELVES IN. 9432 06:49:58,140 --> 06:49:59,141 >> IF WE COULD HAVE A QUESTION 9433 06:49:59,141 --> 06:50:02,545 FROM THE OTHER SIDE. 9434 06:50:02,545 --> 06:50:02,812 >> YES. 9435 06:50:02,812 --> 06:50:03,646 THANK YOU. AM I CLEARLY AUDIBLE? 9436 06:50:03,646 --> 06:50:07,950 I'M RUCHIRA DATTA. 9437 06:50:07,950 --> 06:50:08,684 YES, I'M A MATHEMATICIAN, AND 9438 06:50:08,684 --> 06:50:10,886 I'VE GOT QUITE A BIT OF 9439 06:50:10,886 --> 06:50:11,454 EXPERIENCE IN COMPUTATIONAL 9440 06:50:11,454 --> 06:50:14,423 BIOLOGY, QUITE A BIT OF 9441 06:50:14,423 --> 06:50:15,624 EXPERIENCE IN MACHINE LEARNING 9442 06:50:15,624 --> 06:50:17,226 AND A.I., MOSTLY HAVEN'T 9443 06:50:17,226 --> 06:50:20,296 INTERSECTED, THIS IS MY FIRST 9444 06:50:20,296 --> 06:50:20,963 INTRODUCTION TO COMPUTATIONAL 9445 06:50:20,963 --> 06:50:26,502 NEUROSCIENCE AT THIS CONFERENCE 9446 06:50:26,502 --> 06:50:28,871 I HEARD FROM CARINA CURTO THE 9447 06:50:28,871 --> 06:50:31,574 VISION OF MATH FOR COMPUTATIONAL 9448 06:50:31,574 --> 06:50:33,209 NEUROSCIENCE OF BEING ABLE TO 9449 06:50:33,209 --> 06:50:35,578 UNDERSTAND THE NATURE OF THE 9450 06:50:35,578 --> 06:50:39,315 PARAMETER SPACE AND, YOU KNOW, 9451 06:50:39,315 --> 06:50:45,421 THE INVARIANCE AND VARIANCE AND 9452 06:50:45,421 --> 06:50:48,190 I HEARD FROM ANDREAS DID THIS 9453 06:50:48,190 --> 06:50:51,827 BROAD VISION OF HAVING THESE 9454 06:50:51,827 --> 06:50:53,295 NATURALISTIC DATASETS, HIGH 9455 06:50:53,295 --> 06:50:54,396 ENTROPY DATASETS, LARGE AMOUNTS 9456 06:50:54,396 --> 06:50:55,598 OF COMPUTE. 9457 06:50:55,598 --> 06:50:58,000 AND POSSIBLY BOTH OF YOU HAD IN 9458 06:50:58,000 --> 06:50:59,134 MIND THAT INNER VISION FOR THE 9459 06:50:59,134 --> 06:51:04,573 FUTURE WHAT THE OTHER PERSON WAS 9460 06:51:04,573 --> 06:51:05,774 DESCRIBING WOULD HAPPEN BUT IF 9461 06:51:05,774 --> 06:51:09,879 YOU COULD SAY HOW DO YOU SEE 9462 06:51:09,879 --> 06:51:13,115 THESE TWO THINGS INTERFITTING IN 9463 06:51:13,115 --> 06:51:13,916 THE FUTURE? 9464 06:51:13,916 --> 06:51:16,085 THE MATH AND THE NATURALISTIC 9465 06:51:16,085 --> 06:51:19,989 DATASETS, LIKE DO YOU HAVE LIKE 9466 06:51:19,989 --> 06:51:21,023 A PROGRAM IN MIND? 9467 06:51:21,023 --> 06:51:23,259 >> WE DON'T HAVE A PROGRAM. 9468 06:51:23,259 --> 06:51:27,062 BUT I MEAN I THINK THE WAY I 9469 06:51:27,062 --> 06:51:28,297 WAS -- WHAT I WAS TRYING TO SAY 9470 06:51:28,297 --> 06:51:30,699 IS THAT WHEN YOU FIT MODELS, THE 9471 06:51:30,699 --> 06:51:33,969 MODELS COME FROM A MODEL FAMILY, 9472 06:51:33,969 --> 06:51:35,271 WITH SOME PARAMETERIZED MODEL 9473 06:51:35,271 --> 06:51:38,807 FAMILY SO UNDERSTANDING THAT 9474 06:51:38,807 --> 06:51:40,509 FAMILY BETTER IS VALUABLE IN 9475 06:51:40,509 --> 06:51:41,277 UNDERSTANDING WHAT KINDS OF 9476 06:51:41,277 --> 06:51:44,914 THINGS CAN BE TRAINED EVEN IN 9477 06:51:44,914 --> 06:51:45,381 PRINCIPLE. 9478 06:51:45,381 --> 06:51:47,983 MAYBE YOUR MODEL FAMILY DOESN'T 9479 06:51:47,983 --> 06:51:54,089 EVEN HAVE THE CAPACITY TO DO -- 9480 06:51:54,089 --> 06:51:55,558 TO CONTAIN SOME NETWORKS THAT DO 9481 06:51:55,558 --> 06:51:58,260 ALL THE TESTS YOU WANT IT TO DO. 9482 06:51:58,260 --> 06:51:59,461 SOMEHOW HAVING A BETTER 9483 06:51:59,461 --> 06:52:01,864 UNDERSTANDING OF THE FAMILY OF 9484 06:52:01,864 --> 06:52:03,365 OBJECTS, RIGHT, THAT YOU'RE 9485 06:52:03,365 --> 06:52:05,834 LIVING IN THE PARAMETER SPACE 9486 06:52:05,834 --> 06:52:06,635 WHEN YOU'RE TRAINING YOU'RE 9487 06:52:06,635 --> 06:52:09,672 STUCK IN THE PARAMETER SPACE, 9488 06:52:09,672 --> 06:52:11,874 MOVING AROUND, UNDERSTANDING 9489 06:52:11,874 --> 06:52:14,410 THAT BETTER CAN HELP DEVELOP 9490 06:52:14,410 --> 06:52:17,446 BETTER TRAINING PROTOCOLS BUT 9491 06:52:17,446 --> 06:52:20,349 ALSO CAN HELP PICK OUT DIFFERENT 9492 06:52:20,349 --> 06:52:22,384 ARCHITECTURES, SO SOME 9493 06:52:22,384 --> 06:52:23,352 ARCHITECTURES ARE FUNDAMENTALLY 9494 06:52:23,352 --> 06:52:25,087 EASIER TO TRAIN FOR CERTAIN 9495 06:52:25,087 --> 06:52:26,155 TASKS THAN OTHERS AND WE DON'T 9496 06:52:26,155 --> 06:52:27,389 KNOW WHAT THEY ARE. 9497 06:52:27,389 --> 06:52:28,724 MAYBE THERE'S CERTAIN KIND OF 9498 06:52:28,724 --> 06:52:30,426 TASKS WELL THIS IS THE 9499 06:52:30,426 --> 06:52:31,727 ARCHITECTURE YOU SHOULD START 9500 06:52:31,727 --> 06:52:33,295 WITH, RIGHT, AND NOT BOTHER 9501 06:52:33,295 --> 06:52:35,664 WITH, YOU KNOW, THIS GIANT -- 9502 06:52:35,664 --> 06:52:36,899 THESE NETWORKS, REALLY YOU WANT 9503 06:52:36,899 --> 06:52:38,434 TO START WITH SOMETHING SMALLER 9504 06:52:38,434 --> 06:52:40,603 WITH A CERTAIN STRUCTURE AND 9505 06:52:40,603 --> 06:52:41,770 YOU'LL TRAIN IT QUICKLY. 9506 06:52:41,770 --> 06:52:43,839 IN SOME SENSE I LIKE TO THINK 9507 06:52:43,839 --> 06:52:46,008 EVOLUTION HAS DONE THAT FOR LOTS 9508 06:52:46,008 --> 06:52:47,343 OF CREATURES FOR DIFFERENT 9509 06:52:47,343 --> 06:52:48,010 TASKS. 9510 06:52:48,010 --> 06:52:50,379 THEY STARTED YOU OFF WITH LIKE 9511 06:52:50,379 --> 06:52:51,847 AN ARCHITECTURE THAT'S BETTER 9512 06:52:51,847 --> 06:52:54,049 AND THEN, YOU KNOW, THROUGH 9513 06:52:54,049 --> 06:52:55,684 TRAINING, THROUGH EXPERIENCE, 9514 06:52:55,684 --> 06:52:56,652 THE ANIMAL FINE TUNES. 9515 06:52:56,652 --> 06:53:00,289 AND SO I THINK THERE'S THIS 9516 06:53:00,289 --> 06:53:01,323 INTERPLAY BETWEEN MODEL FAMILY, 9517 06:53:01,323 --> 06:53:04,126 YOU KNOW, WHICH CAN BE DEFINED 9518 06:53:04,126 --> 06:53:05,894 VIA ARCHITECTURE AND A CERTAIN 9519 06:53:05,894 --> 06:53:10,032 SHAPE TO THE PARAMETER SPACE, 9520 06:53:10,032 --> 06:53:12,201 AND THEN THE DATA, SO YOU WANT 9521 06:53:12,201 --> 06:53:13,602 TO FIT THOSE THINGS TOGETHER, 9522 06:53:13,602 --> 06:53:14,837 DEPENDING ON THE KIND OF DATA 9523 06:53:14,837 --> 06:53:16,538 THAT YOU'RE GOING TO TRAIN WITH 9524 06:53:16,538 --> 06:53:19,575 THERE MIGHT BE BETTER OR WORSE 9525 06:53:19,575 --> 06:53:20,776 ARCHITECTURES, BETTER OR WORSE 9526 06:53:20,776 --> 06:53:27,883 NETWORK FAMILIES TO START WITH. 9527 06:53:27,883 --> 06:53:29,685 AND SO ANDREAS WILL SAY MORE BUT 9528 06:53:29,685 --> 06:53:34,890 HIS POINT IS LIKE LET'S LOOK AT 9529 06:53:34,890 --> 06:53:35,824 DATA THAT ISN'T BIASED TOWARDS 9530 06:53:35,824 --> 06:53:39,395 CERTAIN KINDS OF MODELS SO WE 9531 06:53:39,395 --> 06:53:42,131 GET A MORE ROBUST -- SO WE'RE 9532 06:53:42,131 --> 06:53:43,866 FITTING MORE ROBUST NETWORKS IS 9533 06:53:43,866 --> 06:53:53,242 MY UNDERSTANDING BUT I'LL LET 9534 06:53:53,242 --> 06:53:54,843 HIM ELABORATE. 9535 06:53:54,843 --> 06:53:56,879 >> ULTIMATELY IF WE HAVE THESE 9536 06:53:56,879 --> 06:53:59,882 MODELS, GENERATE DATA FROM THEM, 9537 06:53:59,882 --> 06:54:04,386 UNDER CONDITIONS THAT WE CAN 9538 06:54:04,386 --> 06:54:07,523 DESCRIBE THE SYMMETRIES, IN THE 9539 06:54:07,523 --> 06:54:08,724 NEURAL DATA, ALONG THE LINES OF 9540 06:54:08,724 --> 06:54:15,497 WHAT DORIS WAS TALKING ABOUT, 9541 06:54:15,497 --> 06:54:16,765 FUNDAMENTAL INVARIANCE IN 9542 06:54:16,765 --> 06:54:19,702 PHYSICS AND STUFF LIKE THAT, SO 9543 06:54:19,702 --> 06:54:20,969 PLAYS A CRITICAL ROLE. 9544 06:54:20,969 --> 06:54:22,838 PROBLEMS ARE DEALING WITH DATA 9545 06:54:22,838 --> 06:54:25,374 MORE NOISY, ONE ADVANTAGE OF 9546 06:54:25,374 --> 06:54:27,543 LARGE SCALE MODELS, FOUNDATIONAL 9547 06:54:27,543 --> 06:54:30,179 MODELS CAN RUN LIKE BASICALLY ON 9548 06:54:30,179 --> 06:54:31,714 UNLIMITED NUMBER OF EXPERIMENTS 9549 06:54:31,714 --> 06:54:36,885 SO WE CAN GET A BETTER SAMPLE 9550 06:54:36,885 --> 06:54:37,486 COMPLEXITY OF THE MODEL. 9551 06:54:37,486 --> 06:54:40,756 AT THE END OF THE DAY WE HAVE TO 9552 06:54:40,756 --> 06:54:44,760 LINK IT MATHEMATICAL, TO 9553 06:54:44,760 --> 06:54:46,528 MATHEMATICS, FIELDS LIKE 9554 06:54:46,528 --> 06:54:47,463 TOPOLOGY, GEOMETRY, AND THINGS 9555 06:54:47,463 --> 06:54:49,531 LIKE THAT. 9556 06:54:49,531 --> 06:54:52,501 GROUP THEORY. 9557 06:54:52,501 --> 06:54:56,572 YOU'RE THE EXPERT TOO. 9558 06:54:56,572 --> 06:54:57,272 >> I AGREE. 9559 06:54:57,272 --> 06:54:59,174 GRACE, CAN WE DO ONE MORE? 9560 06:54:59,174 --> 06:55:03,779 ONE LAST QUESTION. 9561 06:55:03,779 --> 06:55:04,313 >> THANK YOU. 9562 06:55:04,313 --> 06:55:04,947 >> OKAY. JIM ZHENG 9563 06:55:04,947 --> 06:55:06,982 UNIVERSITY OF TEXAS HEALTH 9564 06:55:06,982 --> 06:55:09,384 CENTER IN HOUSTON. 9565 06:55:09,384 --> 06:55:11,386 WONDERING WHETHER THE PANEL CAN 9566 06:55:11,386 --> 06:55:13,689 ADDRESS THE ISSUE OF A.I. 9567 06:55:13,689 --> 06:55:15,124 DISPARITY LIKE MANY RESEARCH 9568 06:55:15,124 --> 06:55:17,226 GROUPS THEY, YOU KNOW, DON'T 9569 06:55:17,226 --> 06:55:20,028 HAVE ACCESS TO A.I. TECHNOLOGY 9570 06:55:20,028 --> 06:55:21,864 OR COMPUTING RESOURCE OR THEY 9571 06:55:21,864 --> 06:55:24,533 REALLY DON'T HAVE ANY A.I. 9572 06:55:24,533 --> 06:55:26,301 EXPERTS TO COLLABORATE WITH. 9573 06:55:26,301 --> 06:55:29,638 SOME GROUPS, THEY HAVE ALL THESE 9574 06:55:29,638 --> 06:55:35,077 RESOURCES, THEY REALLY EXCEL IN 9575 06:55:35,077 --> 06:55:37,713 THE CURRENT ENVIRONMENT. 9576 06:55:37,713 --> 06:55:41,550 >> I THINK THAT'S A VALID POINT. 9577 06:55:41,550 --> 06:55:43,018 NOT JUST VALID BUT VERY 9578 06:55:43,018 --> 06:55:44,153 IMPORTANT POINT. 9579 06:55:44,153 --> 06:55:45,654 MY OWN PERSONAL ANSWER TO IT IS 9580 06:55:45,654 --> 06:55:51,593 THAT I WOULD REALLY LIKE TO SEE 9581 06:55:51,593 --> 06:55:54,096 GOVERNMENT INVESTMENT IN COMPUTE 9582 06:55:54,096 --> 06:55:58,066 SO THAT THERE CAN BE MORE 9583 06:55:58,066 --> 06:56:01,970 EQUITABLE ACCESS TO A.I. 9584 06:56:01,970 --> 06:56:02,237 RESOURCES. 9585 06:56:02,237 --> 06:56:04,606 THERE MIGHT BE OTHER SOLUTIONS, 9586 06:56:04,606 --> 06:56:05,707 BUT THAT SEEMS LIKE THE MOST 9587 06:56:05,707 --> 06:56:11,213 OBVIOUS ONE TO ME. 9588 06:56:11,213 --> 06:56:12,781 >> I FULLY AGREE, THIS IS A 9589 06:56:12,781 --> 06:56:13,715 MAJOR ISSUE. 9590 06:56:13,715 --> 06:56:16,785 I WOULD SAY THE FIRST STEP IS TO 9591 06:56:16,785 --> 06:56:20,222 REACH OUT TO FIND COLLABORATORS 9592 06:56:20,222 --> 06:56:23,458 THAT ARE DOING -- THERE'S A LOT 9593 06:56:23,458 --> 06:56:24,927 OF -- THESE DAYS THERE'S A LOT 9594 06:56:24,927 --> 06:56:27,396 OF PEOPLE THAT ARE DOING A.I. 9595 06:56:27,396 --> 06:56:29,131 RESEARCH AND ARE VERY 9596 06:56:29,131 --> 06:56:32,167 INTERESTED, STUDENTS TO DO 9597 06:56:32,167 --> 06:56:33,969 NEUROSCIENCE RESEARCH, BECAUSE, 9598 06:56:33,969 --> 06:56:41,543 YOU KNOW, THAT FIELD IS SO 9599 06:56:41,543 --> 06:56:42,911 COMPETITIVE DOMINATED BY 9600 06:56:42,911 --> 06:56:45,681 INDUSTRY, THE ISSUE BECOMES 9601 06:56:45,681 --> 06:56:47,115 FINDING ACCESS TO GPUs, AND 9602 06:56:47,115 --> 06:56:48,383 BUT THERE ARE SOME RESOURCES BUT 9603 06:56:48,383 --> 06:56:51,887 THIS IS A BIG PROBLEM FOR 9604 06:56:51,887 --> 06:56:52,421 EVERYBODY. 9605 06:56:52,421 --> 06:56:53,222 IN ACADEMIA. 9606 06:56:53,222 --> 06:56:56,158 BECAUSE WE CANNOT COMPETE WITH 9607 06:56:56,158 --> 06:56:59,828 RESOURCES THAT INDUSTRY GETS, 9608 06:56:59,828 --> 06:57:01,396 RIGHT? 9609 06:57:01,396 --> 06:57:04,633 THAT'S FUNDAMENTAL PROBLEM. 9610 06:57:04,633 --> 06:57:05,968 >> YEAH. 9611 06:57:05,968 --> 06:57:07,970 >> THERE'S A LOT OF RESOURCES 9612 06:57:07,970 --> 06:57:09,338 THAT OF COURSE SHOULD BE 9613 06:57:09,338 --> 06:57:10,706 STRENGTH BIND MORE GOVERNMENT 9614 06:57:10,706 --> 06:57:12,341 INVESTMENT BUT THERE'S A LOT OF 9615 06:57:12,341 --> 06:57:13,208 FEDERALLY PROVIDED COMPUTE 9616 06:57:13,208 --> 06:57:14,343 CLUSTERS THAT VERY FEW PEOPLE 9617 06:57:14,343 --> 06:57:16,211 SEEM TO KNOW ABOUT FOR REASONS 9618 06:57:16,211 --> 06:57:18,180 THAT ARE STILL MYSTERIOUS TO ME. 9619 06:57:18,180 --> 06:57:20,449 NSF HAS ITS OWN CLUSTERS 9620 06:57:20,449 --> 06:57:23,318 ESSENTIALLY COST FREE FOR 9621 06:57:23,318 --> 06:57:24,753 STUDENTS AND TRAINEES, WRITE A 9622 06:57:24,753 --> 06:57:27,789 PROPOSAL, AND THEY GIVE YOU. 9623 06:57:27,789 --> 06:57:28,891 NATIONAL LAB OPENS COMPUTE 9624 06:57:28,891 --> 06:57:30,325 CLUSTERS TO PEOPLE, THE KIND OF 9625 06:57:30,325 --> 06:57:33,395 THING THAT REQUIRES A BETTER 9626 06:57:33,395 --> 06:57:36,431 WEBSITE OR MORE PRO-ACTIVE 9627 06:57:36,431 --> 06:57:37,499 E-MAILING. 9628 06:57:37,499 --> 06:57:39,635 >> I WANT TO ADD THAT AS WE LOOK 9629 06:57:39,635 --> 06:57:41,937 TO PROVIDE MORE COMPUTE WE HAVE 9630 06:57:41,937 --> 06:57:45,307 TO MAKE COMPUTE MORE ENERGY 9631 06:57:45,307 --> 06:57:53,849 EFFICIENT. 9632 06:57:53,849 --> 06:57:56,118 >> KEN WHANG WANT TO COMMENT 9633 06:57:56,118 --> 06:57:58,620 ABOUT THE NSF COMPUTE RESOURCES? 9634 06:57:58,620 --> 06:57:59,187 >> SURE. 9635 06:57:59,187 --> 06:58:01,189 CAN YOU HEAR ME? 9636 06:58:01,189 --> 06:58:06,862 SO, ONE THING, THE ISSUE OF A.I. 9637 06:58:06,862 --> 06:58:08,964 RESEARCH RESOURCES IS A TOUGH 9638 06:58:08,964 --> 06:58:11,400 ONE BUT THERE IS -- YOU KNOW, 9639 06:58:11,400 --> 06:58:12,534 SOMETHING THAT IS GETTING OFF 9640 06:58:12,534 --> 06:58:14,569 THE GROUND CALLED THE NATIONAL 9641 06:58:14,569 --> 06:58:17,706 A.I. RESEARCH RESOURCE. 9642 06:58:17,706 --> 06:58:20,542 WHICH IS A COLLECTION OF 9643 06:58:20,542 --> 06:58:21,443 RESOURCES THAT ARE HONESTLY 9644 06:58:21,443 --> 06:58:26,949 CURRENTLY A LITTLE BIT TRICKY TO 9645 06:58:26,949 --> 06:58:27,582 NAVIGATE, BUT SPECIFICALLY THEY 9646 06:58:27,582 --> 06:58:29,851 ARE TO TRY TO GET THE SAME KINDS 9647 06:58:29,851 --> 06:58:35,991 OF OR SIMILAR KINDS OF RESOURCES 9648 06:58:35,991 --> 06:58:38,360 FOR RESEARCHERS AS ARE AVAILABLE 9649 06:58:38,360 --> 06:58:43,665 TO FOLKS IN INDUSTRY. 9650 06:58:43,665 --> 06:58:48,470 THE OTHER THING, ACTUALLY WE'LL 9651 06:58:48,470 --> 06:58:50,038 TALK TOMORROW ABOUT NSF-FUNDED 9652 06:58:50,038 --> 06:58:52,274 RESOURCE FOR FOLKS WHO ARE 9653 06:58:52,274 --> 06:59:02,384 WORKING ON NEUROMORPHIC 9654 06:59:02,384 --> 06:59:03,418 HARDWARE. 9655 06:59:03,418 --> 06:59:05,687 TAKE A LOOK AT RESOURCES 9656 06:59:05,687 --> 06:59:06,621 AVAILABLE FROM NSF INVESTMENT. 9657 06:59:06,621 --> 06:59:08,023 THERE'S A LOT OUT THERE AND 9658 06:59:08,023 --> 06:59:10,626 SYSTEMS THAT YOU CAN GO TO THAT 9659 06:59:10,626 --> 06:59:11,927 WILL HELP YOU RECOMMEND WHAT 9660 06:59:11,927 --> 06:59:13,261 KIND OF RESOURCE YOU SHOULD BE 9661 06:59:13,261 --> 06:59:14,629 USING BASED ON THE KINDS OF 9662 06:59:14,629 --> 06:59:15,998 COMPUTE JOBS YOU HAVE. 9663 06:59:15,998 --> 06:59:19,601 SO THERE IS ACTUALLY A LOT OUT 9664 06:59:19,601 --> 06:59:19,801 THERE. 9665 06:59:19,801 --> 06:59:22,504 AND YOU SHOULD TAKE ADVANTAGE OF 9666 06:59:22,504 --> 06:59:23,105 IT. 9667 06:59:23,105 --> 06:59:25,107 YOU ALSO DON'T NEED TO BE NSF 9668 06:59:25,107 --> 06:59:35,650 FUNDED TO USE THESE RESOURCES. 9669 06:59:37,586 --> 06:59:43,658 >> I WANTED TO MENTION TOO THAT 9670 06:59:43,658 --> 06:59:45,160 HIGH PERFORMANCE COMPUTING 9671 06:59:45,160 --> 06:59:46,294 RESOURCES ARE PROVIDED TO PEOPLE 9672 06:59:46,294 --> 06:59:48,430 WHO HAVE ACTIVE DOD FUNDING, 9673 06:59:48,430 --> 06:59:52,401 COULD BE A CONTRACT, A GRANT, 9674 06:59:52,401 --> 06:59:55,237 BUT YOU -- THERE ARE 9675 06:59:55,237 --> 06:59:56,772 RESTRICTIONS. 9676 06:59:56,772 --> 07:00:00,675 USERS HAVE TO BE 9677 07:00:00,675 --> 07:00:01,910 PERMANENT RESIDENTS OR U.S. 9678 07:00:01,910 --> 07:00:03,779 CITIZENS, A BACKGROUND CHECK IS 9679 07:00:03,779 --> 07:00:04,513 REQUIRED. SO THERE IS A BAR. 9680 07:00:04,513 --> 07:00:06,882 BUT BEYOND THAT, 9681 07:00:06,882 --> 07:00:10,285 THE RESOURCES ARE AVAILABLE AT 9682 07:00:10,285 --> 07:00:12,454 NO COST. 9683 07:00:12,454 --> 07:00:13,388 >> THANK YOU, HAL. 9684 07:00:13,388 --> 07:00:16,591 WITH THAT, I WANT TO EXPRESS A 9685 07:00:16,591 --> 07:00:17,192 TREMENDOUS GRATITUDE FOR 9686 07:00:17,192 --> 07:00:19,227 EVERYONE WHO HAS STAYED WITH US 9687 07:00:19,227 --> 07:00:22,330 THIS LONG AND TO A JAM-PACKED 9688 07:00:22,330 --> 07:00:23,532 AGENDA, THANK YOU FOR CONCLUDING 9689 07:00:23,532 --> 07:00:28,270 THE FIRST DAY OF THE BRAIN 9690 07:00:28,270 --> 07:00:29,371 NEUROAI WORKSHOP. THANK YOU 9691 07:00:29,371 --> 07:00:31,039 KANAKA AND STEVE FOR CO-MODERATING 9692 07:00:31,039 --> 07:00:33,075 AND EVERYONE HERE AND GAVE US 9693 07:00:33,075 --> 07:00:34,242 EXCELLENT PRESENTATIONS AND 9694 07:00:34,242 --> 07:00:35,710 DISCUSSIONS THROUGHOUT THE DAY 9695 07:00:35,710 --> 07:00:36,812 WHICH WE'LL CONTINUE TOMORROW. 9696 07:00:36,812 --> 07:00:38,947 SO I WAS GOING TO GIVE A RECAP 9697 07:00:38,947 --> 07:00:41,149 AT 8:30 IN THE MORNING BUT LET'S 9698 07:00:41,149 --> 07:00:42,584 CALL THAT 8:40-ISH. 9699 07:00:42,584 --> 07:00:43,552 GET HERE EARLY ENOUGH. 9700 07:00:43,552 --> 07:00:45,987 THIS IS THE POSTER BLITZ THAT'S 9701 07:00:45,987 --> 07:00:47,389 REALLY IMPORTANT. 9702 07:00:47,389 --> 07:00:49,524 8:45 TOMORROW MORNING. 9703 07:00:49,524 --> 07:00:50,959 YOU'LL HEAR FROM EXCELLENT 9704 07:00:50,959 --> 07:00:52,761 POSTER PRESENTERS AND THEN CAN 9705 07:00:52,761 --> 07:00:57,999 GO OUT AT 9:00 AND HEAR SOME 9706 07:00:57,999 --> 07:00:58,333 GREAT SCIENCE. 9707 07:00:58,333 --> 07:00:59,801 SO THAT IS MOST OF EVERYTHING I 9708 07:00:59,801 --> 07:01:00,769 NEED TO TELL YOU. 9709 07:01:00,769 --> 07:01:03,105 I THINK YOU KNOW WHERE THE EXITS 9710 07:01:03,105 --> 07:01:03,472 ARE. 9711 07:01:03,472 --> 07:01:06,441 DON'T GET LOST IN THE DARK 9712 07:01:06,441 --> 07:01:06,675 PLEASE. 9713 07:01:06,675 --> 07:01:08,310 THERE'S SOME TURNSTILES ON THE 9714 07:01:08,310 --> 07:01:09,611 WAY TO THE METRO CENTER IF 9715 07:01:09,611 --> 07:01:10,679 THAT'S THE DIRECTION YOU'RE 9716 07:01:10,679 --> 07:01:11,980 GOING. 9717 07:01:11,980 --> 07:01:16,251 BUT AGAIN, THANK YOU TO 9718 07:01:16,251 --> 07:01:17,986 EVERYONE. 9719 07:01:17,986 --> 07:01:21,223 WE'LL SEE YOU TOMORROW. 9720 07:01:21,223 --> 07:01:31,199 [APPLAUSE]