1 00:00:05,772 --> 00:00:07,241 ALL RIGHT, WELCOME, EVERYONE, 2 00:00:07,241 --> 00:00:08,809 I THINK WE'RE READY TO GET 3 00:00:08,809 --> 00:00:11,712 STARTED WITH DAY TWO OF THE 4 00:00:11,712 --> 00:00:21,388 WORKSHOP. I THINK WE HAD AN 5 00:00:21,388 --> 00:00:22,022 EXCELLENT DAY YESTERDAY AND I AM 6 00:00:22,022 --> 00:00:22,656 EXCITED TO HAVE PEOPLE HERE IN 7 00:00:22,656 --> 00:00:23,257 THE ROOM AND WATCHING ON ZOOM 8 00:00:23,257 --> 00:00:23,857 AND ON THE VIDEO CAST. I AM 9 00:00:23,857 --> 00:00:24,458 BRIEFLY GOING TO GIVE YOU A, 10 00:00:24,458 --> 00:00:25,125 WELL, I'M SUPPOSED TO GIVE YOU A 11 00:00:25,125 --> 00:00:25,659 RECAP OF WHAT HAPPENED 12 00:00:25,659 --> 00:00:26,260 YESTERDAY. A LOT HAPPENED 13 00:00:26,260 --> 00:00:27,327 YESTERDAY AS YOU KNOW IF YOU 14 00:00:27,327 --> 00:00:30,264 LOOKED AT THE AGENDA AND YOU 15 00:00:30,264 --> 00:00:31,598 HAVE BEEN PARTICIPATING WE HAVE 16 00:00:31,598 --> 00:00:33,033 A DENSE SCHEDULE FULL OF 17 00:00:33,033 --> 00:00:36,537 EXCELLENT TALKS AND DISCUSSIONS 18 00:00:36,537 --> 00:00:37,604 AND YESTERDAY CERTAINLY AFTER 19 00:00:37,604 --> 00:00:39,873 THAT AND IT WAS A GREAT START TO 20 00:00:39,873 --> 00:00:41,675 THE WORKSHOP. WE WERE EXCITED 21 00:00:41,675 --> 00:00:46,280 TO HEAR VERY SUBSTANTIVE DEEP 22 00:00:46,280 --> 00:00:49,116 CONVERSATIONS AROUND IMPORTANT 23 00:00:49,116 --> 00:00:49,783 CONVERSATIONS AROUND NEUROAI AND 24 00:00:49,783 --> 00:00:51,952 WHAT TYPES OF DATA ARE NEEDED 25 00:00:51,952 --> 00:00:55,856 FOR DIFFERENT NEUROAI 26 00:00:55,856 --> 00:00:58,125 APPROACHES. I AM SHOWING THE 27 00:00:58,125 --> 00:00:59,860 AGENDA FOR DAY ONE RIGHT NOW 28 00:00:59,860 --> 00:01:00,727 JUST TO REMIND US OF THE SESSION 29 00:01:00,727 --> 00:01:02,429 STRUCTURE SO WE HAVE TWO OF 30 00:01:02,429 --> 00:01:05,666 THESE MAIN SESSIONS PER DAY. WE 31 00:01:05,666 --> 00:01:06,233 HAVE SESSION THREE AND FOUR 32 00:01:06,233 --> 00:01:07,868 COMING UP TODAY. YESTERDAY 33 00:01:07,868 --> 00:01:09,636 SESSION ONE WAS KIND OF BRINGING 34 00:01:09,636 --> 00:01:11,939 TOGETHER A BROAD OVERVIEW OR 35 00:01:11,939 --> 00:01:13,240 SURVEY OF DIFFERENT WAYS THAT 36 00:01:13,240 --> 00:01:14,741 PEOPLE ARE THINKING ABOUT 37 00:01:14,741 --> 00:01:16,810 NATURAL INTELLIGENCE AND AI AND 38 00:01:16,810 --> 00:01:18,612 THAT REALLY IS THE CORE THEME OF 39 00:01:18,612 --> 00:01:21,982 THIS WORKSHOP. I GAVE A STAGE 40 00:01:21,982 --> 00:01:23,951 SETTING SET OR REMARKS 41 00:01:23,951 --> 00:01:25,118 ACKNOWLEDGING THIS KIND OF VERY 42 00:01:25,118 --> 00:01:27,955 LONG HISTORY OF COMPUTING BEING 43 00:01:27,955 --> 00:01:29,856 THOUGHT OF AS HOW WE CAN USE 44 00:01:29,856 --> 00:01:31,358 MACHINES, COMPUTING MACHINES TO 45 00:01:31,358 --> 00:01:33,894 LEARN FROM DATA. AND IT'S TAKEN 46 00:01:33,894 --> 00:01:35,796 A LOT OF DIFFERENT FORMS AND 47 00:01:35,796 --> 00:01:38,198 ESPECIALLY IN THE LAST TEN YEARS 48 00:01:38,198 --> 00:01:40,334 OR SO. WE HAVE CLEARLY REACHED 49 00:01:40,334 --> 00:01:42,669 SOME SORT OF CONVERGENCE POINT. 50 00:01:42,669 --> 00:01:45,272 WE'RE ONE DECADE INTO THE NIH 51 00:01:45,272 --> 00:01:47,908 BRAIN INITIATIVE. WE HAD AN 52 00:01:47,908 --> 00:01:49,843 OVERVIEW OF MANY OF THE BRAIN 53 00:01:49,843 --> 00:01:54,414 DATA RESOURCES. WE HAVE A TRUE 54 00:01:54,414 --> 00:01:57,417 WEALTH OF NEURAL BEHAVIORAL, 55 00:01:57,417 --> 00:01:59,119 PHYSIOLOGICAL, MOLECULAR DATA 56 00:01:59,119 --> 00:02:03,724 OUT THERE AND FROM OUR 57 00:02:03,724 --> 00:02:05,792 LARGE-SCALE PROJECTS INCLUDING BICAN 58 00:02:05,792 --> 00:02:08,095 MAKING MOLECULAR AND CELL-TYPE ATLASES 59 00:02:08,095 --> 00:02:10,097 OF BRAINS OF ACROSS SPECIES FROM 60 00:02:10,097 --> 00:02:14,201 FLIES TO HUMANS. AND WE SAW 61 00:02:14,201 --> 00:02:15,869 EXAMPLES OF THAT AND HOW THAT'S 62 00:02:15,869 --> 00:02:19,006 COMING TOGETHER AND HOW THAT CAN 63 00:02:19,006 --> 00:02:20,140 POTENTIALLY BE USED TO 64 00:02:20,140 --> 00:02:22,476 ACCELERATE A LOT OF THESE KIND 65 00:02:22,476 --> 00:02:28,448 OF IDEAS THAT PEOPLE 66 00:02:28,448 --> 00:02:31,852 HAVE BEEN THROWING AROUND IN THE 67 00:02:31,852 --> 00:02:37,457 FIELD UNDER THE UMBRELLA OF 68 00:02:37,457 --> 00:02:37,657 NEUROAI. 69 00:02:37,657 --> 00:02:38,191 WE HEARD TALKS OF LIVING 70 00:02:38,191 --> 00:02:38,725 SYSTEMS. BIO COMPUTING. 71 00:02:38,725 --> 00:02:39,793 WE HEARD ABOUT DIFFERENT WAYS OF 72 00:02:39,793 --> 00:02:43,630 THINKING ABOUT, YOU KNOW, 73 00:02:43,630 --> 00:02:49,036 POTENTIAL USE CASES FOR 74 00:02:49,036 --> 00:02:49,836 LARGE-SCALE AND SMALL SCALE 75 00:02:49,836 --> 00:02:51,738 NEUROAI MODELS. IN THE 76 00:02:51,738 --> 00:02:53,840 AFTERNOON IN SESSION TWO WE GOT 77 00:02:53,840 --> 00:02:55,142 INTO THE DEEPER QUESTIONS ABOUT 78 00:02:55,142 --> 00:02:56,676 METRICS THE SPACES OVER WHICH WE 79 00:02:56,676 --> 00:02:59,846 DEFINE METRICS HOW AND WHEN THEY 80 00:02:59,846 --> 00:03:02,916 ARE USEFUL. AND HOW TO THINK 81 00:03:02,916 --> 00:03:03,683 ABOUT THE, YOU KNOW, 82 00:03:03,683 --> 00:03:06,153 THERE'S A LOT OF COMPLEXITY IN 83 00:03:06,153 --> 00:03:10,123 HOW WE MAP FROM ONE TYPE OF 84 00:03:10,123 --> 00:03:11,858 MODEL TO OBSERVATIONS IN BRAINS 85 00:03:11,858 --> 00:03:14,361 AND THAT WAS AN EXCELLENT 86 00:03:14,361 --> 00:03:15,495 CONVERSATION I THINK. AND THIS 87 00:03:15,495 --> 00:03:16,496 IS A CONVERSATION THAT'S BEEN 88 00:03:16,496 --> 00:03:18,832 HAPPENING FOR A LONG TIME IN THE 89 00:03:18,832 --> 00:03:20,434 FIELD. BUT THIS WAS I THINK 90 00:03:20,434 --> 00:03:22,602 REALLY PULLING A LOT OF THREADS 91 00:03:22,602 --> 00:03:25,138 TOGETHER. SO TODAY, KIND OF ON 92 00:03:25,138 --> 00:03:26,640 THAT FOUNDATION WE'RE LOOKING 93 00:03:26,640 --> 00:03:29,376 FORWARD TO EXPANDING THE SCOPE 94 00:03:29,376 --> 00:03:31,778 IN SESSION THREE TO INCLUDE NEW 95 00:03:31,778 --> 00:03:34,014 COMPLEMENTARY MODEL PLATFORMS 96 00:03:34,014 --> 00:03:37,617 INCLUDING A FOCUS ON 97 00:03:37,617 --> 00:03:39,119 NEUROMORPHIC COMPUTING, 98 00:03:39,119 --> 00:03:40,487 NEUROMORPHIC ENGINEERING, BRINGING 99 00:03:40,487 --> 00:03:41,788 PHYSICAL DEVICES INTO THE LOOP WITH 100 00:03:41,788 --> 00:03:45,025 OUR SCIENTIFIC, EXPERIMENTAL PLATFORMS AND 101 00:03:45,025 --> 00:03:45,992 THINKING MORE ABOUT 102 00:03:45,992 --> 00:03:46,560 ENVIRONMENTAL AND SENSORY 103 00:03:46,560 --> 00:03:48,295 FEEDBACK AS A KEY COMPONENT OF 104 00:03:48,295 --> 00:03:50,330 NATURAL INTELLIGENCE AND SO THAT 105 00:03:50,330 --> 00:03:51,998 WILL BE A GREAT SET OF 106 00:03:51,998 --> 00:03:53,867 DISCUSSIONS TO OPEN UP THE SCOPE 107 00:03:53,867 --> 00:03:56,336 OF NEUROAI. AND IN THE 108 00:03:56,336 --> 00:03:57,904 AFTERNOON SESSION FOUR, WE'LL 109 00:03:57,904 --> 00:03:59,706 START LOOKING AT TRANSLATION AND 110 00:03:59,706 --> 00:04:01,041 HOW TO TRANSITION THESE IDEAS 111 00:04:01,041 --> 00:04:03,343 AND THEORIES AND MODELS INTO 112 00:04:03,343 --> 00:04:05,011 TECHNOLOGIES THAT CAN ADVANCE 113 00:04:05,011 --> 00:04:06,980 OUR SCIENCE AND OUR TECHNOLOGIES 114 00:04:06,980 --> 00:04:08,915 FOR BRAIN HEALTH. AT THE END OF 115 00:04:08,915 --> 00:04:11,818 THE DAY WE'LL HAVE A WRAP UP OF 116 00:04:11,818 --> 00:04:14,321 BOTH THE DAYS SESSIONS AS WELL 117 00:04:14,321 --> 00:04:18,592 AS A WORKSHOP SYNTHESIS PRIOR TO 118 00:04:18,592 --> 00:04:20,260 CLOSING REMARKS. AND SO I JUST 119 00:04:20,260 --> 00:04:22,395 WANT TO QUICKLY REMIND FOLKS 120 00:04:22,395 --> 00:04:24,498 THAT WHEN WE HAVE THE DISCUSSION 121 00:04:24,498 --> 00:04:25,732 PANELS WE'LL HAVE A SERIES OF 122 00:04:25,732 --> 00:04:27,467 SHORT TALKS THEN WE'LL HAVE 123 00:04:27,467 --> 00:04:31,304 PANEL DISCUSSION DURING THE 124 00:04:31,304 --> 00:04:31,805 DURING THE PANEL DISCUSSION 125 00:04:31,805 --> 00:04:33,640 THERE'LL BE A SLIDE LIKE THIS WITH 126 00:04:33,640 --> 00:04:34,641 A QR CODE ON THE LEFT AND YOU CAN 127 00:04:34,641 --> 00:04:36,409 E-MAIL A RESPONSE THAT WE'LL SEE 128 00:04:36,409 --> 00:04:38,111 IMMEDIATELY. LIKE I SAID, SO, 129 00:04:38,111 --> 00:04:39,846 WE HAVE SHORT TALKS, PANEL 130 00:04:39,846 --> 00:04:42,716 DISCUSSION AND THEN OPEN UP TO 131 00:04:42,716 --> 00:04:44,718 AUDIENCE Q&A AND IF YOU HAVE A 132 00:04:44,718 --> 00:04:47,387 BROAD FUTURE LOOKING OR 133 00:04:47,387 --> 00:04:48,088 OPEN-ENDED QUESTION PLEASE SAVE 134 00:04:48,088 --> 00:04:50,056 THAT FOR THE AUDIENCE Q&A. IF 135 00:04:50,056 --> 00:04:51,858 YOU HAVE A VERY QUICK ONE 136 00:04:51,858 --> 00:04:54,861 BETWEEN SPEAKERS, YOU CAN BE 137 00:04:54,861 --> 00:04:56,830 WAITING AT AN AISLE MIC AND WE 138 00:04:56,830 --> 00:04:59,766 CAN CALL ON YOU. IF YOU ARE ON 139 00:04:59,766 --> 00:05:02,502 ZOOM, IF YOU'RE A ZOOM PANELIST 140 00:05:02,502 --> 00:05:04,871 PLEASE REMAIN MUTED AND DISABLE 141 00:05:04,871 --> 00:05:06,106 YOUR VIDEO UNLESS YOU'VE BEEN 142 00:05:06,106 --> 00:05:08,208 PROMOTED TO SPEAK. OTHER ZOOM 143 00:05:08,208 --> 00:05:11,678 PARTICIPANTS, JUST SUBMIT A 144 00:05:11,678 --> 00:05:13,079 QUESTION, PLEASE USE THE Q&A 145 00:05:13,079 --> 00:05:15,282 BUTTON AND IF YOU'RE ON VIDEOCAST 146 00:05:15,282 --> 00:05:17,184 YOU CAN SEND A QUESTION 147 00:05:17,184 --> 00:05:19,152 THROUGH THE SEND LIVE FEEDBACK 148 00:05:19,152 --> 00:05:20,253 FORM AND WE WILL GET THAT 149 00:05:20,253 --> 00:05:22,222 IMMEDIATELY. JUST TO REMIND YOU 150 00:05:22,222 --> 00:05:24,658 WHERE YOU ARE. YOU ARE SEATED 151 00:05:24,658 --> 00:05:26,259 IN NATCHER AUDITORIUM. I AM 152 00:05:26,259 --> 00:05:28,562 TOLD TO REMIND EVERYONE THAT NO 153 00:05:28,562 --> 00:05:30,797 FOOD OR BEVERAGE IS TO BE 154 00:05:30,797 --> 00:05:33,633 PERMITTED INTO THE AUDITORIUM 155 00:05:33,633 --> 00:05:35,869 AND MOST IMPORTANTLY, I AM ABOUT 156 00:05:35,869 --> 00:05:41,575 TO HAND OFF TO COURTNEY PINARD TO 157 00:05:41,575 --> 00:05:43,743 INTRODUCE THE EARLY-CAREER POSTER 158 00:05:43,743 --> 00:05:44,845 BLITZES AND SO THOSE WILL BE UP 159 00:05:44,845 --> 00:05:47,047 AT THE ATRIUM. I AM SURE YOU'VE 160 00:05:47,047 --> 00:05:48,381 SEEN WHERE THOSE ARE SET UP. 161 00:05:48,381 --> 00:05:54,721 BUT I WILL HAND IT OVER TO OUR POSTER 162 00:05:54,721 --> 00:06:05,265 BLITZ AWARDEES. WELCOME TO THE STAGE. 163 00:06:42,068 --> 00:06:43,703 MORNING, MY NAME IS COURTNEY 164 00:06:43,703 --> 00:06:46,373 PINARD. I WORK IN THE DIVISION 165 00:06:46,373 --> 00:06:48,608 OF NEUROSCIENCE AND I'M JOINED 166 00:06:48,608 --> 00:06:51,177 BY THE POSTER BLITZ AWARDEES AND 167 00:06:51,177 --> 00:06:53,647 MY COLLEAGUE JESSICA MOLLICK. 168 00:06:53,647 --> 00:06:57,017 SHE IS A PROGRAM OFFICER 169 00:06:57,017 --> 00:06:58,618 AT THE NATIONAL INSTITUTE 170 00:06:58,618 --> 00:07:01,454 OF DRUG ABUSE IN THE BEHAVIORAL 171 00:07:01,454 --> 00:07:02,922 AND COGNITIVE BRANCH. 172 00:07:02,922 --> 00:07:07,827 AND SO I WANTED TO START WITH 173 00:07:07,827 --> 00:07:11,865 JUST INTRODUCING THE BRAIN 174 00:07:11,865 --> 00:07:13,900 NEUROAI TRAINING SUBCOMMITTEE 175 00:07:13,900 --> 00:07:15,235 WHOSE MEMBERSHIP SPANS ACROSS AGENCIES, 176 00:07:15,235 --> 00:07:18,905 NIH INSTITUDES AND TEAMS. SO THIS 177 00:07:18,905 --> 00:07:20,206 SUBCOMMITTEE WAS RESPONSIBLE FOR 178 00:07:20,206 --> 00:07:22,242 REVIEWING ABSTRACTS AND JESSICA 179 00:07:22,242 --> 00:07:24,577 AND I LED THAT EFFORT. THE 180 00:07:24,577 --> 00:07:26,446 POSTER BLITZ COMPETITION WAS A 181 00:07:26,446 --> 00:07:29,516 GREAT SUCCESS. WE HAD OVER ONE 182 00:07:29,516 --> 00:07:31,785 HUNDRED ABSTRACT SUBMISSIONS 183 00:07:31,785 --> 00:07:33,953 FROM EARLY CAREER SCHOLARS FROM 184 00:07:33,953 --> 00:07:37,357 HIGH SCHOOL ALL THE WAY UP UNTIL 185 00:07:37,357 --> 00:07:38,591 POST-DOCTORAL CAREER STAGES. 186 00:07:38,591 --> 00:07:40,727 AND I WANTED TO THANK EVERYONE 187 00:07:40,727 --> 00:07:42,696 WHO SUBMITTED AN ABSTRACT. 188 00:07:42,696 --> 00:07:44,764 THERE ARE MANY TALENTED 189 00:07:44,764 --> 00:07:47,434 TRAINEES. AND AS YOU'LL SEE IN 190 00:07:47,434 --> 00:07:48,568 THE PRESENTATIONS THIS MORNING 191 00:07:48,568 --> 00:07:53,073 THE SUBMISSION TOPICS RANGE FROM 192 00:07:53,073 --> 00:07:55,408 THEORY, MODELING AND PRINCIPLES 193 00:07:55,408 --> 00:07:56,643 OF INTELLIGENCE TO BRAIN 194 00:07:56,643 --> 00:07:58,745 INSPIRED ALGORITHMS AND 195 00:07:58,745 --> 00:08:01,147 ROBOTICS, METRICS FOR MAPPING 196 00:08:01,147 --> 00:08:03,516 MODELS TO BRAINS AND 197 00:08:03,516 --> 00:08:05,218 NEUROMORPHIC COMPUTING SO 198 00:08:05,218 --> 00:08:07,454 PLEASE, PLEASE ATTEND THIS 199 00:08:07,454 --> 00:08:09,189 POSTER SESSION AFTER THE BLITZ 200 00:08:09,189 --> 00:08:10,490 TO LEARN MORE. 201 00:08:10,490 --> 00:08:11,691 FOR THE BLITZ COMPETITION 202 00:08:11,691 --> 00:08:14,461 SCHOLARS NEEDED TO DEMONSTRATE 203 00:08:14,461 --> 00:08:16,629 HOW THEIR RESEARCH CONSIDERS THE RECIPROCAL 204 00:08:16,629 --> 00:08:18,798 INTERACTION BETWEEN NEUROSCIENCE AND AI RESEARCH. AND HERE 205 00:08:18,798 --> 00:08:21,901 SCHOLARS WERE ASKED TO PREPARE A 206 00:08:21,901 --> 00:08:23,503 ONE MINUTE PRESENTATION TO 207 00:08:23,503 --> 00:08:25,004 SHOWCASE THEIR POSTERS DURING THE 208 00:08:25,004 --> 00:08:26,740 SESSION WHICH FOLLOWS THE BLITZ 209 00:08:26,740 --> 00:08:29,943 AS I SAID. SO I WOULD LIKE TO 210 00:08:29,943 --> 00:08:31,945 START WITH JUST CONGRATULATING 211 00:08:31,945 --> 00:08:42,455 THE BLITZ HONOREES. CIANA DEVEAU, 212 00:08:43,456 --> 00:08:46,493 HARRISON ESPINO, LESLIE FAMULARO, 213 00:08:46,493 --> 00:08:49,629 SEUNG JU KIM, DANEIL LEVENSTEIN, 214 00:08:49,629 --> 00:08:51,197 JULIANA LONDONO ALVAREZ, 215 00:08:51,197 --> 00:09:01,741 BRADLEY THIELMAN, 216 00:09:02,542 --> 00:09:07,781 ERIC WANG FROM COLLEGE STATION 217 00:09:07,781 --> 00:09:08,815 HIGH SCHOOL IN TEXAS, 218 00:09:08,815 --> 00:09:15,188 QINGYANG ALICE WANG, AND 219 00:09:15,188 --> 00:09:19,626 ALEXANDER WHITE FROM NATIONAL 220 00:09:19,626 --> 00:09:21,461 TSINGHUA UNIVERSITY IN TAIWAN. 221 00:09:21,461 --> 00:09:32,005 SO FIRST UP WE HAVE CIANA DEVEAU. 222 00:09:32,305 --> 00:09:33,406 >> GOOD MORNING, EVERYONE, IN 223 00:09:33,406 --> 00:09:36,409 OUR WORK WE SHOW A NEW ROLE FOR 224 00:09:36,409 --> 00:09:38,545 THE DENSE LOCAL RECURRENT 225 00:09:38,545 --> 00:09:39,512 CONNECTIONS IN LAYER TWO THREE 226 00:09:39,512 --> 00:09:42,315 OF THE VISUAL CORTEX. HERE THEY 227 00:09:42,315 --> 00:09:46,052 FILTER SEQUENCES OF INPUT. USING 228 00:09:46,052 --> 00:09:48,154 TWO-PHOTON OPTOGENETIC 229 00:09:48,154 --> 00:09:49,656 STIMULATION, WE CAN MIMIC NATURAL 230 00:09:49,656 --> 00:09:51,724 MOVIE, PRESENTING IT IN THE PLACE OF THE IMAGE EITHER 231 00:09:51,724 --> 00:09:53,326 AT THE CORRECT TIME POINT IN THE 232 00:09:53,326 --> 00:09:55,795 MOVIE OR A RANDOM UNMATCHED TIME 233 00:09:55,795 --> 00:09:59,532 POINT. WE FIND AMPLIFICATION OF 234 00:09:59,532 --> 00:10:01,701 THE OPTO PATTERN ONLY IN THE CORRECT 235 00:10:01,701 --> 00:10:03,903 CONTEXT OF THE MOVIE, WHILE IT IS 236 00:10:03,903 --> 00:10:06,272 ATTENUATED IN THE UNMATCHED 237 00:10:06,272 --> 00:10:07,440 CONTEXT. WE THEN EXAMINE IN 238 00:10:07,440 --> 00:10:08,775 ARTIFICIAL RECURRENT NETWORK 239 00:10:08,775 --> 00:10:11,077 TRAINED TO AMPLIFY SPECIFIC 240 00:10:11,077 --> 00:10:12,312 INPUT SEQUENCES. AND FIND THAT 241 00:10:12,312 --> 00:10:15,515 IT PRODUCES THE SAME CONTEXT- 242 00:10:15,515 --> 00:10:18,852 DEPENDENT EFFECT. SO THE 243 00:10:18,852 --> 00:10:20,553 BIOLOGICAL AND MODEL DATA ALIGN SHOWING 244 00:10:20,553 --> 00:10:22,222 THAT THE SEQUENCE FILTERING IS 245 00:10:22,222 --> 00:10:25,091 SPECIFIC, AMPLIFYING NATURAL 246 00:10:25,091 --> 00:10:28,261 SEQUENCES WHILE SUPPRESSING OTHERS. 247 00:10:28,261 --> 00:10:31,164 ADDITIONAL EXPERIMENTS EXPLORE MECHANSISM 248 00:10:31,164 --> 00:10:34,234 UNDERLYING THIS FILTERING EFFECT. 249 00:10:34,234 --> 00:10:38,771 SO PLEASE STOP BY THE POSTERS FOR 250 00:10:38,771 --> 00:10:39,572 MORE DETAIL AND DISCUSSION. 251 00:10:39,572 --> 00:10:44,677 THANK YOU. 252 00:10:44,677 --> 00:10:46,713 THANK YOU. 253 00:10:46,713 --> 00:10:49,282 >> HELLO, EVERYONE, I WILL 254 00:10:49,282 --> 00:10:50,850 SHOW THIS BRAIN INSPIRED ALGORITHM FOR 255 00:10:50,850 --> 00:10:52,519 AUTOOMOUS ROBOTIC NAVIGATION WHICH 256 00:10:52,519 --> 00:10:54,888 SIMULTANEOUSLY MAPS THE ENVIRONMENT 257 00:10:54,888 --> 00:10:56,322 AND PLANS COST-OPTIMAL PASSTHROUGH. 258 00:10:56,322 --> 00:10:57,891 THIS IS A QUICK OVERVIEW OF THE 259 00:10:57,891 --> 00:11:00,727 MODEL. AS WE FIRST FIND A 260 00:11:00,727 --> 00:11:02,395 PATH BY PROPAGATING THROUGH A 261 00:11:02,395 --> 00:11:05,064 GRID OF PLACE CELLS FROM START TO END AND THIS 262 00:11:05,064 --> 00:11:09,135 IS SUPPOSED TO MIMIC HIPPOCAMPAL 263 00:11:09,135 --> 00:11:10,403 PREPLAY. AFTER WE HAVE THE PATH, WE 264 00:11:10,403 --> 00:11:12,872 COLLECT COST DATA, SO THINGS LIKE THE 265 00:11:12,872 --> 00:11:14,107 ENERGY EXPENDED AROUND THE WHEELS, 266 00:11:14,107 --> 00:11:16,242 TIME SPENT AROUND OBSTACLES, STEEPNESS 267 00:11:16,242 --> 00:11:18,578 OF THE TERRAIN. AND ONCE WE HAVE THAT 268 00:11:18,578 --> 00:11:19,879 INFORMATION WE CAN UPDATE THE PLANNER 269 00:11:19,879 --> 00:11:21,814 WEIGHTS USING LEARING RULE CALLED E-PROP. 270 00:11:21,814 --> 00:11:22,749 THIS CHANGES THE SPEED OF THE 271 00:11:22,749 --> 00:11:24,918 CONNECTION PROPAGATED BETWEEN NEURONS, SIMILAR 272 00:11:24,918 --> 00:11:26,886 TO ANOTHER INSPIRATION WHICH WAS 273 00:11:26,886 --> 00:11:28,821 AXONAL-DEPENDENT PLASTICITY WHERE CHANGES 274 00:11:28,821 --> 00:11:30,423 IN THE WHITE MATTER WILL CONNECT AFFECT 275 00:11:30,423 --> 00:11:32,392 CONDUCTION VELOCITY. WE PUT THIS ON 276 00:11:32,392 --> 00:11:35,595 A ROBOT AND ALLOW IT TO AUTONOMOUSLY 277 00:11:35,595 --> 00:11:36,963 NAVIGATE FOR TWELVE HOURS. 278 00:11:36,963 --> 00:11:39,532 WE FOUND TWO THINGS, ONE, WE CREATED 279 00:11:39,532 --> 00:11:42,068 INFORMATIVE MAPS OF THE ENVIRONMENT AND 280 00:11:42,068 --> 00:11:43,636 COMPARED TO SOTA PLANNING ALGORITHMS 281 00:11:43,636 --> 00:11:45,872 LIKE A*, RRT*, IT ACTUALLY PERFORMED 282 00:11:45,872 --> 00:11:48,274 BETTER AND PLANNED LOWER-COST 283 00:11:48,274 --> 00:11:52,879 PATHS THAN TO ADAPTIVE PLANNERS LIKE D*LITE. 284 00:11:52,879 --> 00:11:53,580 IF YOU WANT MORE INFORMATION 285 00:11:53,580 --> 00:11:56,749 PLEASE STOP BY THE POSTER. 286 00:11:56,749 --> 00:12:06,960 THANK YOU. 287 00:12:07,927 --> 00:12:09,762 >> GOOD MORNING, EVERYONE, MY 288 00:12:09,762 --> 00:12:13,967 MY NAME IS LESLIE FAMULARO, THIS 289 00:12:13,967 --> 00:12:18,204 WORKSHOP A RECURRING THEME IS TO 290 00:12:18,204 --> 00:12:20,340 INTRODUCE NEUROMORPHIC OR BIOINSPIRED 291 00:12:20,340 --> 00:12:22,609 COMPUTING IN DEEP LEARNING. WE MADE 292 00:12:22,609 --> 00:12:25,078 SUCH AN ATTEMPT. SPECIFICALLY 293 00:12:25,078 --> 00:12:28,481 WE FOCUS ON AUDITORY NEUROSCIENCE. WE 294 00:12:28,481 --> 00:12:30,149 TAKE A FORWARD MODEL FROM THE 295 00:12:30,149 --> 00:12:31,784 COCHLEA TO THE BRAIN AND WE 296 00:12:31,784 --> 00:12:33,953 COMBINE THAT WITH THE DEEP 297 00:12:33,953 --> 00:12:35,021 NEURAL NETWORK THROUGH 298 00:12:35,021 --> 00:12:38,324 DIFFERENTIALABILITY SO 299 00:12:38,324 --> 00:12:40,660 THAT IT CAN BE UPDATED FROM THE 300 00:12:40,660 --> 00:12:42,595 FRONT END WHICH IS OUR BIOINSPIRED 301 00:12:42,595 --> 00:12:44,931 ARCHITECTURE AND A DEEP NEURAL 302 00:12:44,931 --> 00:12:45,665 NETWORK AT THE SAME TIME. WE 303 00:12:45,665 --> 00:12:47,734 USE THIS TO TRAIN MODELS FOR 304 00:12:47,734 --> 00:12:50,003 ENGINEERING BASED TASKS LIKE 305 00:12:50,003 --> 00:12:50,937 SPEECH RECOGNITION AND 306 00:12:50,937 --> 00:12:52,071 ENHANCEMENT AND WE CONSISTENTLY 307 00:12:52,071 --> 00:12:54,607 FIND OUR MODEL TO OUTPERFORM 308 00:12:54,607 --> 00:12:55,775 DEEP LEARNING ALONE BOTH 309 00:12:55,775 --> 00:12:57,610 PERFORMANCE AND ALSO ROBUSTNESS. 310 00:12:57,610 --> 00:13:01,514 OUR PARAMETERS ARE ALSO 311 00:13:01,514 --> 00:13:03,182 INTERPRETABLE BECAUSE THEY'RE 312 00:13:03,182 --> 00:13:07,420 BASED ON NEUROSCIENCE AND WE'RE 313 00:13:07,420 --> 00:13:09,756 EXPLORING APPLICATIONS ON PERSONALIZED HEARING, 314 00:13:09,756 --> 00:13:12,158 CLINICAL APPLICATIONS HEARING AID FITTING. IF YOU 315 00:13:12,158 --> 00:13:19,065 WANT TO HEAR ME STOP BY POSTER 316 00:13:19,065 --> 00:13:29,242 THREE, THANK YOU. 317 00:13:29,242 --> 00:13:35,214 >> OKAY. HELLO, EVERYONE. I AM 318 00:13:35,214 --> 00:13:36,883 SEUNG JU KIM, MATERIALS SCIENTIST. MY RESEARCH 319 00:13:36,883 --> 00:13:39,752 DEVELOPING NEUROMORPHIC HARDWARE 320 00:13:39,752 --> 00:13:41,554 USING PULSE CMOS TECHNOLOGY. TRADITIONAL 321 00:13:41,554 --> 00:13:44,724 TRANSISTOR-BASED COMPUTING IS 322 00:13:44,724 --> 00:13:47,860 DIGITAL PROCESS-CENTRIC, RESULTING 323 00:13:47,860 --> 00:13:49,295 IN MASSIVE ENERGY CONSUMPTION AND 324 00:13:49,295 --> 00:13:51,864 INEFFICIENT COMPUTING. HOWEVER 325 00:13:51,864 --> 00:13:54,033 NEUROMORPHIC COMPUTING USING 326 00:13:54,033 --> 00:13:57,036 MEMRISTOR TECHNOLOGY CAN 327 00:13:57,036 --> 00:13:59,839 COMPUTE A.I. EFFICIENTLY BY 328 00:13:59,839 --> 00:14:03,176 USING ION MIGRATIONS TO UPDATE 329 00:14:03,176 --> 00:14:05,244 THE SYNAPTIC WEIGHT, MUCH BETTER 330 00:14:05,244 --> 00:14:09,082 THAN THE DIGITAL PROCESS 331 00:14:09,082 --> 00:14:13,886 COMPUTINGS. SO MEMRISTORS CAN 332 00:14:13,886 --> 00:14:15,021 REPRESENT MULTI-INFO WITHIN SINGLE 333 00:14:15,021 --> 00:14:16,656 DEVICES. WE SHOWED THAT USING 334 00:14:16,656 --> 00:14:21,260 HALIDE IONS, INSTEAD OF OXYGEN IONS 335 00:14:21,260 --> 00:14:29,602 CAN, MIGRATE ION IN THE ENTIRE CELL 336 00:14:29,602 --> 00:14:31,104 ACHIEVING LINAR SYNAPTIC UPDATING. 337 00:14:31,104 --> 00:14:31,804 FOR MORE CHECK OUT MY POSTER. 338 00:14:31,804 --> 00:14:42,014 THANK YOU. 339 00:15:00,500 --> 00:15:01,934 >> YEAH SO IF YOU WERE TO ASK ME 340 00:15:01,934 --> 00:15:04,137 WHAT DOES THE HIPPOCAMPUS DO. I 341 00:15:04,137 --> 00:15:06,873 WOULD TELL YOU IT DOES TWO 342 00:15:06,873 --> 00:15:09,609 THINGS. ONE AN ONLINE MODE 343 00:15:09,609 --> 00:15:11,844 THAT REPRESENTS WHERE AN AGENT 344 00:15:11,844 --> 00:15:13,980 OR ANIMAL IS IN THE WORLD AND 345 00:15:13,980 --> 00:15:15,014 THIS IS WHERE YOU SEE PLACE 346 00:15:15,014 --> 00:15:16,816 CELLS AND AN OFFLINE MODE IN 347 00:15:16,816 --> 00:15:19,685 WHICH IT SIMULATES TRAJECTORIES 348 00:15:19,685 --> 00:15:21,788 THROUGH THE WORLD WHICH WE OFTEN 349 00:15:21,788 --> 00:15:23,589 CALL REPLAY. SO THE QUESTION IS 350 00:15:23,589 --> 00:15:25,691 HOW DOES THE HIPPOCAMPUS DEVELOP COGNTIVE MAP 351 00:15:25,691 --> 00:15:27,527 AND GENERATE REPLAY? SOMETHING THAT 352 00:15:27,527 --> 00:15:30,763 CAN DO BOTH OF THESE THINGS FROM 353 00:15:30,763 --> 00:15:31,664 EGOCENTRIC INFORMATION AVAILABLE 354 00:15:31,664 --> 00:15:34,133 TO THE ANIMAL AND SO TO STUDY 355 00:15:34,133 --> 00:15:36,202 THIS WE BUILT THIS GRID WORLD 356 00:15:36,202 --> 00:15:37,370 ENVIRONMENT IN WHICH AN AGENT 357 00:15:37,370 --> 00:15:38,805 MOVES AROUND AND COLLECTS 358 00:15:38,805 --> 00:15:39,639 SENSORY INFORMATION FROM THE 359 00:15:39,639 --> 00:15:41,707 VISUAL FIELD IN FRONT OF IT. 360 00:15:41,707 --> 00:15:44,210 AND WE DREW FROM SOME RECENT 361 00:15:44,210 --> 00:15:47,280 RESULTS IN A.I. AND COMPUTATIONAL 362 00:15:47,280 --> 00:15:49,348 NEUROSCIENCE. SHOWING THAT WHEN 363 00:15:49,348 --> 00:15:50,650 RECURRENT NEURAL NETWORKS LEARN 364 00:15:50,650 --> 00:15:52,718 TO PREDICT SENSORY INFORMATION 365 00:15:52,718 --> 00:15:53,653 THEY DEVELOP PLACE CELLS AND SO 366 00:15:53,653 --> 00:15:55,588 THE FIRST THING WE DID WAS 367 00:15:55,588 --> 00:15:56,823 REPLICATE THIS WORK. WE SHOW 368 00:15:56,823 --> 00:15:58,224 THAT INDEED WE END UP WITH 369 00:15:58,224 --> 00:16:00,293 THE EMERGENCE OF PLACE CELLS IN 370 00:16:00,293 --> 00:16:02,128 OUR NETWORK BUT MORE IMPORTANTLY 371 00:16:02,128 --> 00:16:03,863 WHEN WE ACTUALLY TURN OFF THE 372 00:16:03,863 --> 00:16:06,065 SENSORY INPUT WE END UP WITH A 373 00:16:06,065 --> 00:16:08,701 REPRESENTATION OF THE WORLD THAT 374 00:16:08,701 --> 00:16:10,403 SIMULATES PLAUSIBLE TRAJECTORIES. 375 00:16:10,403 --> 00:16:13,472 THIS SHOWS THE DECODED POSTERIOR 376 00:16:13,472 --> 00:16:16,142 OF THE NETWORK DURING SLEEP WHICH REPRESENTS 377 00:16:16,142 --> 00:16:17,276 A COHERENT LOCATION IN THE 378 00:16:17,276 --> 00:16:18,277 ENVIRONMENT. AND THIS CAN MOVE 379 00:16:18,277 --> 00:16:19,812 AROUND IN THE ENVIRONMENT 380 00:16:19,812 --> 00:16:20,980 PRODUCING TRAJECTORIES THROUGH 381 00:16:20,980 --> 00:16:23,049 THE WORLD INCLUDING THOSE THAT 382 00:16:23,049 --> 00:16:24,350 THE AGENT HAS NEVER SEEN BEFORE 383 00:16:24,350 --> 00:16:27,854 AS WELL AS OUTPUTS THAT LOOK LIKE 384 00:16:27,854 --> 00:16:28,487 SENSORY INFORMATION THAT THE 385 00:16:28,487 --> 00:16:30,823 AGENT WOULD HAVE SEEN IN THE 386 00:16:30,823 --> 00:16:31,924 ENVIRONMENT. AND SO WE WANTED 387 00:16:31,924 --> 00:16:33,426 TO KNOW HOW THE NETWORK IS 388 00:16:33,426 --> 00:16:36,729 ACTUALLY DOING THIS, WE USE 389 00:16:36,729 --> 00:16:38,431 TOOLS FROM NEUROSCIENCE TO LOOK 390 00:16:38,431 --> 00:16:39,565 AT THE NEURAL MANIFOLD AND WE 391 00:16:39,565 --> 00:16:41,534 FOUND THIS WAS HAPPENING BECAUSE 392 00:16:41,534 --> 00:16:44,103 THE NETWORK LEARNED AN ATTRACTIVE 393 00:16:44,103 --> 00:16:45,238 MANIFOLD OF SPACE AND THIS ONLY 394 00:16:45,238 --> 00:16:47,139 EMERGED WHEN THE NETWORK LEARN TO PREDICT 395 00:16:47,139 --> 00:16:48,975 SEQUENCES OF SENSORY INPUT NOT 396 00:16:48,975 --> 00:16:51,210 SAY SINGLE OR NEXT STEP INPUT 397 00:16:51,210 --> 00:16:53,479 DESPITE THE FACT THAT ALL OF 398 00:16:53,479 --> 00:16:54,580 THESE NETWORKS LEARN SPATIALLY 399 00:16:54,580 --> 00:16:56,282 TUNED CELLS AND THIS CORRESPONDED 400 00:16:56,282 --> 00:16:57,683 TO THE EMERGENCE OF 401 00:16:57,683 --> 00:16:59,252 REPLAY ONLY IN THE SEQUENTIAL 402 00:16:59,252 --> 00:17:00,720 NETWORKS. SO THE ANSWER WE COME 403 00:17:00,720 --> 00:17:03,623 UP WITH IS BASICALLY THAT THE 404 00:17:03,623 --> 00:17:05,157 HIPPOCAMPUS COULD LEARN A 405 00:17:05,157 --> 00:17:06,158 COGNITIVE MAP THRU LEARNING TO 406 00:17:06,158 --> 00:17:07,326 PREDICT A SEQUENCE OF SENSORY 407 00:17:07,326 --> 00:17:10,730 INFORMATION AS AN AGENT MOVES 408 00:17:10,730 --> 00:17:20,973 THROUGH THE 409 00:17:29,415 --> 00:17:29,649 ENVIRONMENT. 410 00:17:29,649 --> 00:17:33,686 >> OKAY SO ATTRACTOR- BASED 411 00:17:33,686 --> 00:17:35,688 MODELS USES ATTRACTORS TO 412 00:17:35,688 --> 00:17:37,323 REPRESENT PROCESSES SO LIKE 413 00:17:37,323 --> 00:17:40,693 MEMORIES CAN BE ENCODED AS FIX 414 00:17:40,693 --> 00:17:43,062 POINTS WHILE LOCOMOTION HAS 415 00:17:43,062 --> 00:17:44,864 TRADITIONALLY BEEN MODELLED 416 00:17:44,864 --> 00:17:48,100 USING COUPLED OSCILLATORS. I THINK THIS 417 00:17:48,100 --> 00:17:50,469 COULD OFFER AN ATTRACTIVE 418 00:17:50,469 --> 00:17:52,405 UNIFYING NETWORK WHERE MULTIPLE 419 00:17:52,405 --> 00:17:55,675 GAITS CAN BE ENCODED AS CO-EXISTING 420 00:17:55,675 --> 00:17:57,109 LIMIT CYCLES. HOWEVER, THIS IS HARD 421 00:17:57,109 --> 00:17:58,844 TO IMPLEMENT. BUT USING 422 00:17:58,844 --> 00:18:00,613 MATHEMATICAL INSIGHTS FROM A 423 00:18:00,613 --> 00:18:02,748 VERY SPECIAL FAMILY OF THRESHOLD 424 00:18:02,748 --> 00:18:04,784 NEURAL NETWORKS, WE DEVELOP 425 00:18:04,784 --> 00:18:07,653 SEVERAL ATTRACTOR-BASE MODELS 426 00:18:07,653 --> 00:18:15,227 INCLUDING A 43 NEURAL NETWORK THAT 427 00:18:15,227 --> 00:18:16,862 ENCODES A SEQUENCE OF QUADRUPED GAITS. 428 00:18:16,862 --> 00:18:19,732 THE NETWORK TRANSITIONS TO THE NEXT GAIT WITHOUT 429 00:18:19,732 --> 00:18:21,167 ANY INFORMATION ABOUT WHICH GAIT 430 00:18:21,167 --> 00:18:22,601 COMES NEXT. AND I WILL BE VERY 431 00:18:22,601 --> 00:18:24,904 HAPPY TO SHOW YOU THE 432 00:18:24,904 --> 00:18:28,274 MATHEMATICS BEHIND IT AT MY 433 00:18:28,274 --> 00:18:38,617 POSTER. THANK YOU. 434 00:18:45,024 --> 00:18:47,293 >> MORNING, I'M BRAD, SO SCIENCE 435 00:18:47,293 --> 00:18:49,595 AND COMPUTING BOTH NEED 436 00:18:49,595 --> 00:18:51,797 ABSTRACTIONS AND WE BUILD THESE BY 437 00:18:51,797 --> 00:18:54,867 FIRST PLACING SPIKES INTO TIME 438 00:18:54,867 --> 00:18:57,236 BINS TO FORM POPULATION ACTIVITY 439 00:18:57,236 --> 00:18:57,970 VECTORS BUT THERE'S NO REASON 440 00:18:57,970 --> 00:19:00,006 FOR THE BRAIN TO RESPECT THESE 441 00:19:00,006 --> 00:19:03,242 TIME BINS. TIME IN THE BRAIN IS 442 00:19:03,242 --> 00:19:06,312 RELATIVE AND SPIKES ONLY MAKE 443 00:19:06,312 --> 00:19:07,847 SENSE TO THE NETWORK THAT 444 00:19:07,847 --> 00:19:10,182 GENERATED THEM. WE KNOW THAT 445 00:19:10,182 --> 00:19:11,050 SYNAPTIC INTERACTION IS THE BASIS FOR NEURAL COMPUTATION. 446 00:19:11,050 --> 00:19:12,852 MY POSTER PRESENTS A PROCESS THAT 447 00:19:12,852 --> 00:19:14,920 UNIFIES ACTIVITIES FROM THE 448 00:19:14,920 --> 00:19:17,390 CONNECTOME INTO A SINGLE 449 00:19:17,390 --> 00:19:19,859 MATHEMATICAL OBJECT, THIS DIRECTED GRAPH 450 00:19:19,859 --> 00:19:24,030 AND I WILL PRESENT AN EFFICIENT ALGORITHM 451 00:19:24,030 --> 00:19:25,097 TO IDENTIFY ISOMORPHIC SUBGRAPHS TO THIS 452 00:19:25,097 --> 00:19:27,066 DIRECTED GRAPH WHICH IS A HARD PROBLEM IN 453 00:19:27,066 --> 00:19:28,467 GENERAL BUT THIS IS NOT THE 454 00:19:28,467 --> 00:19:30,970 GENERAL CASE BUT I WILL ARGUE 455 00:19:30,970 --> 00:19:32,605 THAT THESE ISOMORPHIC SUBGRAPHS ARE 456 00:19:32,605 --> 00:19:37,009 CANDIDATES FOR ISOMORPHIC COMPUTATIONS 457 00:19:37,009 --> 00:19:47,386 SO SEE ME AT POSTER 14. 458 00:19:52,425 --> 00:19:54,427 >> HELLO, EVERYONE, MY NAME IS 459 00:19:54,427 --> 00:19:56,896 ERIC WANG, ASSOCIATIVE LEARNING 460 00:19:56,896 --> 00:19:58,664 IN WHICH THE CONNECTIONS BETWEEN 461 00:19:58,664 --> 00:20:00,066 DIFFERENT STIMULI ADAPT TO THE 462 00:20:00,066 --> 00:20:02,168 ENVIRONMENT AND OCCURS THROUGH 463 00:20:02,168 --> 00:20:04,637 TWO WAYS OF VOLUME TRANSMISSION 464 00:20:04,637 --> 00:20:06,739 AND WIRE TRANSMISSION. TO MIMIC 465 00:20:06,739 --> 00:20:08,474 ASSOCIATIVE LEARNING I USE 466 00:20:08,474 --> 00:20:11,444 ELECTRIC STIMULI TO EMULATE WIRE 467 00:20:11,444 --> 00:20:12,244 TRANSMISSION AND PHOTO STIMULUS TO EMULATE 468 00:20:12,244 --> 00:20:13,846 VOLUME TRANSMISSION. SO I WANT TO 469 00:20:13,846 --> 00:20:15,081 ENHANCE THE EFFICIENCY OF THE 470 00:20:15,081 --> 00:20:18,217 PROCESS SO I INCORPORATED 471 00:20:18,217 --> 00:20:20,720 QUANTUM DOTS WITH A PVP FILM WHICH 472 00:20:20,720 --> 00:20:24,290 WAS BETWEEN SILVER AND INDIUM TIN OXIDE. 473 00:20:24,290 --> 00:20:25,424 I THEN MEASURED THE CURRENT OF 474 00:20:25,424 --> 00:20:27,059 STIMULUS COMPARED TO BEFORE THE ASSOCIATION AND 475 00:20:27,059 --> 00:20:31,864 IT SHOWED A SIGNIFICANT INCREASE 476 00:20:31,864 --> 00:20:33,699 IN RESPONSE. THE RESISTIVE SWITCHING 477 00:20:33,699 --> 00:20:39,038 ALSO SHOWED THE ASSOCIATION BETWEEN THE 478 00:20:39,038 --> 00:20:40,506 ELECTRIC AND PHOTO STIMULI. I USED 479 00:20:40,506 --> 00:20:43,275 CLASSICAL CONDITIONING AND IT SHOWED THE ELECTRIC AND PHOTO 480 00:20:43,275 --> 00:20:44,510 STIMULI HAVE QUICK TRAINING 481 00:20:44,510 --> 00:20:47,079 AND LONG RETENTION FOR 482 00:20:47,079 --> 00:20:50,649 ASSOCIATIVE LEARNING. FINALLY 483 00:20:50,649 --> 00:20:53,319 PUT THIS IN A NEURAL NETWORK AND 484 00:20:53,319 --> 00:20:55,087 PERFORMED TESTS. AFTER THE ASSOCIATION, THE ACCURACY OF 485 00:20:55,087 --> 00:20:57,189 THE RECOGNITION OF NUMBERS 0-9 486 00:20:57,189 --> 00:20:59,225 WERE DRASTICALLY INCREASED USING A 487 00:20:59,225 --> 00:20:59,859 SMALL DATA TRAINING SET. THANK 488 00:20:59,859 --> 00:21:09,969 YOU. 489 00:21:20,179 --> 00:21:25,885 >> HI, EVERYONE, SO HOW DOES THE 490 00:21:25,885 --> 00:21:29,021 EXCITATORY-INHIBITORY STRUCTURE 491 00:21:29,021 --> 00:21:29,955 CHANGES FUNCTIONAL COMPLEXITY OF 492 00:21:29,955 --> 00:21:31,824 NETWORKS? I AM SHOWING 493 00:21:31,824 --> 00:21:33,058 YOU TWO STORIES HERE WITH ONE 494 00:21:33,058 --> 00:21:36,529 ON LEARNING IN THE POSTER. SO 495 00:21:36,529 --> 00:21:39,131 THE UNIVERSAL APPROXIMATE THEOREM IS THE GUARANTEE 496 00:21:39,131 --> 00:21:41,634 FOR THE SUCCESS OF DEEP LEARNING 497 00:21:41,634 --> 00:21:44,003 TODAY. BUT IT WAS KNOWN THAT 498 00:21:44,003 --> 00:21:46,605 HAVING INHIBITORY CONNECTIONS 499 00:21:46,605 --> 00:21:49,108 IN THE NETWORK IS A REQUIREMENT 500 00:21:49,108 --> 00:21:51,243 FOR THAT TO HOLD. HOW MANY 501 00:21:51,243 --> 00:21:54,413 INHIBITORY NEURONS DO WE NEED? THAT IS 502 00:21:54,413 --> 00:21:56,482 SHOWN ON THE RIGHT. WHERE WE 503 00:21:56,482 --> 00:21:59,618 LEVERAGE THE RECENTLY AVAILABLE 504 00:21:59,618 --> 00:22:02,588 LARVA DROSOPHILA EM CONNECTOME SET. THE HIGHER 505 00:22:02,588 --> 00:22:04,823 FUNCTIONAL COMPLEXITY SIMULATED 506 00:22:04,823 --> 00:22:06,492 BRAIN NETWORKS ARE COLORED BY 507 00:22:06,492 --> 00:22:09,929 LIGHTER DOTS. AND THE MAIN TAKE 508 00:22:09,929 --> 00:22:11,830 HOME HERE IS THE GREEN STAR 509 00:22:11,830 --> 00:22:13,933 WHICH MARKS THE BRAIN E-I AS 510 00:22:13,933 --> 00:22:14,967 STRUCTURED PROPERTIES THAT 511 00:22:14,967 --> 00:22:17,503 MATCHES TO THE HIGHER FUNCTIONAL 512 00:22:17,503 --> 00:22:19,672 COMPLEXITY REGION. I'M ALICE WANG 513 00:22:19,672 --> 00:22:21,974 FROM HOPKINS, LOOKING FORWARD TO 514 00:22:21,974 --> 00:22:25,411 SEEING YOU AT THE POSTER. THANK 515 00:22:25,411 --> 00:22:35,521 YOU 516 00:22:38,991 --> 00:22:39,091 YOU. 517 00:22:39,091 --> 00:22:41,160 >> HOW IS THE BRAIN SO FLEXIBLE? 518 00:22:41,160 --> 00:22:43,462 IT IS ABLE TO RESPOND RAPIDLY TO 519 00:22:43,462 --> 00:22:45,764 UNEXPECTED STIMULI WITHOUT THE 520 00:22:45,764 --> 00:22:46,599 NEED 521 00:22:46,599 --> 00:22:50,569 >> SYNAPTIC PLASTICITY OR 522 00:22:50,569 --> 00:22:52,671 REWIRING. THE -- THIS -- THE 523 00:22:52,671 --> 00:22:54,807 HYPER FLEXIBILITY HINGES ON THE 524 00:22:54,807 --> 00:22:56,609 RECURRENT STRUCTURE OF THE BRAIN. 525 00:22:56,609 --> 00:22:59,578 THE -- ESPECIALLY MUTUAL 526 00:22:59,578 --> 00:23:00,879 INHIBITION. THESE CONNECTIONS 527 00:23:00,879 --> 00:23:03,249 ADD BIFURCATIONS SUCH THAT SMALL 528 00:23:03,249 --> 00:23:06,685 NUDGES AND CONTEXTUAL INPUT CAN 529 00:23:06,685 --> 00:23:08,420 SEAMLESSLY TRANSITION THE NETWORK INTO 530 00:23:08,420 --> 00:23:09,822 COMPUTATIONAL MODES. HERE WE PUSH THIS TO 531 00:23:09,822 --> 00:23:12,224 ITS LIMIT WE CONSTRUCT A FOUR 532 00:23:12,224 --> 00:23:14,393 NEURON CIRCUIT CAPABLE OF 533 00:23:14,393 --> 00:23:19,665 COMPRESSING 24 UNIQUE FUNCTIONS INTO A SINGLE CIRCUIT WITH A SET OFSYNPATIC WEIGHTS. 534 00:23:19,665 --> 00:23:20,566 SIMPLY BY ADJUSTING THE BIAS CURRENTS OF 535 00:23:20,566 --> 00:23:24,870 NEURONS, WE ARE ABLE TO SEAMLESSLY 536 00:23:24,870 --> 00:23:26,272 TRANSITION BETWEEN ALL THESE. WE 537 00:23:26,272 --> 00:23:28,107 ARE ABLE TO COMPUTE LOGIC FROM 538 00:23:28,107 --> 00:23:30,476 THREE DIFFERENT TYPES OF INPUTS: 539 00:23:30,476 --> 00:23:33,245 TEMPORAL, MAGNITUDE AND PHASE. 540 00:23:33,245 --> 00:23:35,281 THIS IS A HARDWARE AGNOSTIC 541 00:23:35,281 --> 00:23:38,417 APPROACH AND DOES NOT REQUIRE 542 00:23:38,417 --> 00:23:39,885 ANY EXOTIC MATERIALS SO IF YOUR 543 00:23:39,885 --> 00:23:41,320 INTERESTED IN LEARNING MORE 544 00:23:41,320 --> 00:23:43,923 PLEASE STOP BY MY POSTER. THANK 545 00:23:43,923 --> 00:23:52,598 YOU. 546 00:23:52,598 --> 00:23:54,066 >> THANKS FOR YOUR ATTENTION 547 00:23:54,066 --> 00:23:56,035 THIS MORNING. LET'S GIVE A BIG 548 00:23:56,035 --> 00:23:59,972 ROUND OF APPLAUSE TO OUR POSTER 549 00:23:59,972 --> 00:24:04,410 BLITZ HONOREES. I ALSO WANTED 550 00:24:04,410 --> 00:24:07,680 TO THANK OUR SPONSORS FOR THEIR 551 00:24:07,680 --> 00:24:08,814 CONTRIBUTION TO THE EARLY CAREER 552 00:24:08,814 --> 00:24:12,284 SCHOLAR ACTIVITIES DURING THIS 553 00:24:12,284 --> 00:24:13,919 WORKSHOP. IF EVERYONE WHO IS 554 00:24:13,919 --> 00:24:15,387 PRESENTING A POSTER COULD JOIN 555 00:24:15,387 --> 00:24:19,224 US BY THE STAIRWELL WE'RE GOING 556 00:24:19,224 --> 00:24:21,093 TAKE A PHOTO AND PLEASE JOIN US 557 00:24:21,093 --> 00:24:23,329 AT THE POSTER SESSION. 558 00:24:23,329 --> 00:24:24,263 OKAY, WELCOME, 559 00:24:24,263 --> 00:24:27,399 EVERYONE TO SESSION THREE. OF 560 00:24:27,399 --> 00:24:30,336 THE WORKSHOP THIS SESSION IS 561 00:24:30,336 --> 00:24:36,375 CHAIRED BY DR. BRAD WHO IS A DISTINGUISHED 562 00:24:36,375 --> 00:24:38,777 MEMBER OF TECHNICAL STAFF AT SANDIA NATIONAL 563 00:24:38,777 --> 00:24:42,147 LABS WHERE HE WORKS IN 564 00:24:42,147 --> 00:24:43,582 NEUROSCIENCE TO ADVANCE 565 00:24:43,582 --> 00:24:45,250 ARTIFICIAL INTELLIGENCE FOR 566 00:24:45,250 --> 00:24:51,357 FUTURE SCIENTIFIC COMPUTING 567 00:24:51,357 --> 00:24:52,491 APPLICATIONS. PRIOR TO JOINING SANDIA 568 00:24:52,491 --> 00:24:55,728 IN 2011 HE WAS A POST-DOCTORAL 569 00:24:55,728 --> 00:24:57,796 RESEARCH ASSOCIATE AT THE SALK 570 00:24:57,796 --> 00:25:03,235 INSTITUTE FOR BIOLOGICAL STUDIES 571 00:25:03,235 --> 00:25:07,239 WITH A PH.D IN COMPNEURO, MS AND BS FROM 572 00:25:07,239 --> 00:25:07,973 FROM RICE UNIVERSITY AND I WELCOME 573 00:25:07,973 --> 00:25:09,942 BRAD TO THE STAGE AND TO GIVE 574 00:25:09,942 --> 00:25:19,218 OUR FIRST PRESENTATION. WELCOME, BRAD 575 00:25:19,218 --> 00:25:20,886 >> BRAD: THANK YOU, JOE AND 576 00:25:20,886 --> 00:25:23,288 THANK YOU, GRACE FOR HELPING 577 00:25:23,288 --> 00:25:24,757 ORGANIZE THIS WONDERFUL WORKSHOP 578 00:25:24,757 --> 00:25:25,457 THIS SESSION IS GOING TO BE A 579 00:25:25,457 --> 00:25:27,926 LITTLE BIT OF A CHANGE FROM THE 580 00:25:27,926 --> 00:25:29,595 DISCUSSIONS YESTERDAY WHICH WERE 581 00:25:29,595 --> 00:25:33,031 LARGELY FOCUSED ON KIND OF MORE 582 00:25:33,031 --> 00:25:34,233 STANDARD A.I. APPROACHES AND HOW 583 00:25:34,233 --> 00:25:35,601 IT RELATE TO NEUROSCIENCE THIS 584 00:25:35,601 --> 00:25:38,470 IS GOING TO FOCUS ON 585 00:25:38,470 --> 00:25:42,241 NEUROMORPHIC COMPUTING AND WE'LL 586 00:25:42,241 --> 00:25:44,576 LATER TALK ABOUT WHAT 587 00:25:44,576 --> 00:25:45,944 NEUROMORPHIC COMPUTING LATER 588 00:25:45,944 --> 00:25:48,380 THAN I WILL. I WILL TALK ABOUT 589 00:25:48,380 --> 00:25:49,982 WHERE I THINK NEUROMORPHIC 590 00:25:49,982 --> 00:25:51,383 COMPUTING CAN HELP US UNDERSTAND 591 00:25:51,383 --> 00:25:52,918 THE BRAIN WHY WOULD THIS 592 00:25:52,918 --> 00:25:54,787 TECHNOLOGY TODAY BE SOMETHING WE 593 00:25:54,787 --> 00:25:56,922 WOULD WANT TO LOOK TO IN THE 594 00:25:56,922 --> 00:26:01,126 NEAR FUTURE? SO I'M GOING TO, 595 00:26:01,126 --> 00:26:05,998 LET'S SEE HERE -- SO I WANT TO 596 00:26:05,998 --> 00:26:07,833 MOTIVATE THIS BY MENTIONING HOW 597 00:26:07,833 --> 00:26:11,937 MODELING AND SIMULATION 598 00:26:11,937 --> 00:26:14,006 HAS THE ABILITY TO PROJECT THE 599 00:26:14,006 --> 00:26:16,041 FUTURE IF YOU WILL IS SOMETHING 600 00:26:16,041 --> 00:26:17,910 THAT HAS ADVANCED TREMENDOUSLY 601 00:26:17,910 --> 00:26:19,611 OVER THE PAST FEW DECADES. 602 00:26:19,611 --> 00:26:21,580 TO THE POINT WHERE TODAY AS MANY 603 00:26:21,580 --> 00:26:22,848 OF YOU ARE FAMILIAR WITH 604 00:26:22,848 --> 00:26:25,684 HURRICANE PREDICTION. THIS WAS 605 00:26:25,684 --> 00:26:27,319 HURRICANE MILTON THAT HIT TAMPA 606 00:26:27,319 --> 00:26:29,955 A FEW WEEKS AGO. 607 00:26:29,955 --> 00:26:32,357 WE HAVE MODSIM CAPABILITIES 608 00:26:32,357 --> 00:26:35,461 THAT ARE TREMENDOUSLY EFFECTIVE 609 00:26:35,461 --> 00:26:36,762 AT PROJECTING SEVERAL DAYS OUT 610 00:26:36,762 --> 00:26:38,330 IN THE CASE OF A HURRICANE, 611 00:26:38,330 --> 00:26:40,332 WHERE THINGS MAY HIT AND IN THIS 612 00:26:40,332 --> 00:26:41,767 CASE IF YOU ARE FAMILIAR WITH 613 00:26:41,767 --> 00:26:43,602 THE STORY THIS HURRICANE WENT 614 00:26:43,602 --> 00:26:44,670 THE WRONG DIRECTION NORMALLY 615 00:26:44,670 --> 00:26:46,939 THEY GO EAST TO WEST THIS WENT 616 00:26:46,939 --> 00:26:48,674 WEST TO EAST AND YET THEY 617 00:26:48,674 --> 00:26:50,509 PREDICTED LIKE FOUR DAYS IN 618 00:26:50,509 --> 00:26:51,810 ADVANCE ALMOST EXACTLY WHERE THE 619 00:26:51,810 --> 00:26:53,378 STORM WAS GOING TO HIT. AND 620 00:26:53,378 --> 00:26:55,614 THIS IS AMAZING, IT SAVES MANY 621 00:26:55,614 --> 00:26:58,083 LIVES, THIS IS A TREMENDOUSLY 622 00:26:58,083 --> 00:26:59,651 IMPORTANT ADVANCE FOR OUR 623 00:26:59,651 --> 00:27:01,887 SOCIETY. AND THIS LEVERAGES 624 00:27:01,887 --> 00:27:04,156 LARGE SUPERCOMPUTERS. SOME OF 625 00:27:04,156 --> 00:27:06,024 THE BIGGEST SUPERCOMPUTERS IN 626 00:27:06,024 --> 00:27:07,993 THE WORLD ARE USED TO MAKE THESE 627 00:27:07,993 --> 00:27:10,128 WEATHER PREDICTION SO THE 628 00:27:10,128 --> 00:27:12,164 QUESTION IS WHY CAN'T WE DO THIS 629 00:27:12,164 --> 00:27:15,801 FOR NEUROLOGICAL HEALTH? WE'RE AT NIH. 630 00:27:15,801 --> 00:27:17,402 ARE WE NOT ABLE TO TAKE 631 00:27:17,402 --> 00:27:19,004 ADVANTAGE OF LARGE-SCALE 632 00:27:19,004 --> 00:27:20,472 SUPERCOMPUTERS IN A WAY THAT 633 00:27:20,472 --> 00:27:23,108 WE CAN PREDICT A HURRICANE, USE 634 00:27:23,108 --> 00:27:24,810 THAT TO PREDICT THE FUTURE 635 00:27:24,810 --> 00:27:26,011 OUTCOME OF DISEASE AND HEALTH? 636 00:27:26,011 --> 00:27:28,280 WE DO THIS FOR OTHER PARTS OF 637 00:27:28,280 --> 00:27:30,115 BIOLOGY BUT FOR THE BRAIN IT'S 638 00:27:30,115 --> 00:27:35,120 NOT QUITE AS EASY. SO, JUST 639 00:27:35,120 --> 00:27:36,822 KEEPING THAT IN MIND THIS SORT 640 00:27:36,822 --> 00:27:39,291 OF VISION OF HOW DO WE BRING 641 00:27:39,291 --> 00:27:39,958 THIS SORT OF KNOWLEDGE THAT WE 642 00:27:39,958 --> 00:27:42,294 HAVE OF THE NEURON AND CIRCUITS THAT THE BRAIN 643 00:27:42,294 --> 00:27:43,529 INITIATIVE HAS BEEN INVESTING 644 00:27:43,529 --> 00:27:45,497 WITH GREAT SUCCESS OVER THE LAST 645 00:27:45,497 --> 00:27:47,399 DECADE AND ACTUALLY USE IT TO 646 00:27:47,399 --> 00:27:49,301 POTENTIALLY MODEL THE BRAIN'S 647 00:27:49,301 --> 00:27:52,004 FUTURE. WE HAD A DISCUSSION ON 648 00:27:52,004 --> 00:27:53,605 FOUNDATION MODELS. I AM GOING TO 649 00:27:53,605 --> 00:27:55,340 TALK ABOUT A MORE CLASSIC WAY OF 650 00:27:55,340 --> 00:27:56,808 DOING THIS WHICH IS SIMULATIONS 651 00:27:56,808 --> 00:27:58,243 AND THEN KIND OF SAY WHAT THE 652 00:27:58,243 --> 00:28:00,445 CHALLENGES ARE WITH THAT AND 653 00:28:00,445 --> 00:28:00,913 THEN LOOKING TOWARDS 654 00:28:00,913 --> 00:28:02,915 NEUROMORPHIC COMPUTING AS A 655 00:28:02,915 --> 00:28:06,118 SOLUTION BUT THEN ALSO KIND OF 656 00:28:06,118 --> 00:28:07,085 INTRODUCING HOW NEUROMORPHIC 657 00:28:07,085 --> 00:28:09,221 COMPUTING WHICH IS NEW HARDWARE 658 00:28:09,221 --> 00:28:10,889 INSPIRED BY HOW THE BRAIN IS 659 00:28:10,889 --> 00:28:12,991 CONSTRUCTED TO GIVE US A DEEPER 660 00:28:12,991 --> 00:28:16,461 UNDERSTANDING OF HOW THE BRAIN IS 661 00:28:16,461 --> 00:28:17,462 FUNCTIONING. SO GOING -- 662 00:28:17,462 --> 00:28:21,300 I CAN GIVE A FULL TALK ABOUT THE 663 00:28:21,300 --> 00:28:23,001 CHALLENGES OF MODSIM OF THE BRAIN. WE HAD 664 00:28:23,001 --> 00:28:24,736 A FEW DISCUSSIONS ABOUT THIS 665 00:28:24,736 --> 00:28:26,972 BEFORE. I THINK IT'S, YOU KNOW, 666 00:28:26,972 --> 00:28:27,973 THERE HAVE BEEN SOME VERY 667 00:28:27,973 --> 00:28:29,174 NOTABLE EFFORTS IN THIS 668 00:28:29,174 --> 00:28:31,476 DIRECTION IN THE HUMAN BRAIN PROJECT 669 00:28:31,476 --> 00:28:34,179 FOR ONE, BEEN A LOT OF EFFORT IN 670 00:28:34,179 --> 00:28:35,447 TRYING TO BUILD SIMULATIONS AND 671 00:28:35,447 --> 00:28:37,416 I THINK QUITE SUCCESSFUL IN WHAT 672 00:28:37,416 --> 00:28:38,784 THEY WERE TARGETING. AND THE 673 00:28:38,784 --> 00:28:40,085 LESSONS THAT COME OUT OF THIS IS 674 00:28:40,085 --> 00:28:42,087 THAT WE ABSOLUTELY HAVE TO HAVE 675 00:28:42,087 --> 00:28:42,854 DATA AND MY PERSONAL EXPERIENCE 676 00:28:42,854 --> 00:28:46,825 IN THIS AND LOOKING AT THE ADULT NEUROGENESIS 677 00:28:46,825 --> 00:28:48,260 AND HIPPOCAMPUS, THE LIMITING FACTOR 678 00:28:48,260 --> 00:28:50,629 VERY OFTEN IS THAT WE DON'T HAVE 679 00:28:50,629 --> 00:28:53,966 A CONNECTOME FOR REGIONS WE WANT 680 00:28:53,966 --> 00:28:55,400 MIGHT TO MODEL OR DETAIL ON ALL 681 00:28:55,400 --> 00:28:58,270 THE CELL TYPES. MOST DISEASES 682 00:28:58,270 --> 00:29:00,372 IN MENTAL HEALTH ESPECIALLY 683 00:29:00,372 --> 00:29:02,674 RELATE TO MODULATORY 684 00:29:02,674 --> 00:29:03,775 NEUROTRANSMITTERS WHICH ARE 685 00:29:03,775 --> 00:29:05,310 DIFFICULT TO MEASURE IN VIVO, 686 00:29:05,310 --> 00:29:06,745 THEIR BEHAVIOR AND THEIR 687 00:29:06,745 --> 00:29:10,248 EFFECTS. WE NEED A LOT OF DATA 688 00:29:10,248 --> 00:29:12,384 AND WE NEED MORE THAN JUST THE 689 00:29:12,384 --> 00:29:13,785 CONNECTOME ALTHOUGH THAT IS A 690 00:29:13,785 --> 00:29:14,486 TREMENDOUS START AND THERE'S 691 00:29:14,486 --> 00:29:16,622 ANOTHER PROBLEM, THE BRAIN IS 692 00:29:16,622 --> 00:29:18,657 REALLY BIG AND IS VERY COMPLEX. 693 00:29:18,657 --> 00:29:20,258 AND I WOULD PROBABLY SAY AS MUCH 694 00:29:20,258 --> 00:29:21,793 A CHALLENGE FOR THIS COMMUNITY 695 00:29:21,793 --> 00:29:25,097 AS WELL AS FOR THE HIGH 696 00:29:25,097 --> 00:29:26,565 PERFORMANCE COMPUTING COMMUNITY 697 00:29:26,565 --> 00:29:28,667 THIS PROBLEM IS ONE EXTREMELY 698 00:29:28,667 --> 00:29:30,135 CHALLENGING, RIGHT, SO IF YOU 699 00:29:30,135 --> 00:29:34,973 LOOK AT THE TIMESCALES. OUR BRAINS 700 00:29:34,973 --> 00:29:37,309 HAVE TEN TO THE FIFTEEN SYNAPSES GIVE 701 00:29:37,309 --> 00:29:40,078 OR TAKE. OUR BRAIN IS FIRING AT 702 00:29:40,078 --> 00:29:42,381 A HERTZ MAYBE MORE. SO THAT'S A 703 00:29:42,381 --> 00:29:44,516 LOT PER SECOND. BUT, MOST 704 00:29:44,516 --> 00:29:45,817 NEUROLOGICAL DISORDERS OCCUR 705 00:29:45,817 --> 00:29:48,754 OVER WEEKS, MONTHS, EVEN YEARS 706 00:29:48,754 --> 00:29:52,124 AND SO WE'RE TALKING ABOUT 707 00:29:52,124 --> 00:29:53,258 TIMESCALES AND SIZE SCALES THAT 708 00:29:53,258 --> 00:29:56,194 PUT IT ALTOGETHER. I NEED, NOT 709 00:29:56,194 --> 00:29:59,631 JUST AN EXASCALE MACHINE, I NEED 710 00:29:59,631 --> 00:30:03,602 MANY EXASCALE MACHINES, MANY TIMES A 711 00:30:03,602 --> 00:30:05,604 MACHINE RUNNING FOR MANY MONTHS 712 00:30:05,604 --> 00:30:07,706 AND MANY YEARS. THAT'S JUST 713 00:30:07,706 --> 00:30:09,341 INTRACTABLE IF WE WANT TO DO 714 00:30:09,341 --> 00:30:12,711 EVEN A SYNAPSE-LEVEL SIMULATION OF 715 00:30:12,711 --> 00:30:13,779 MENTAL HEALTH DISORDER SO SOMETHING HAS 716 00:30:13,779 --> 00:30:15,647 TO CHANGE HERE. AND THEN THE 717 00:30:15,647 --> 00:30:16,682 FINAL THING IS WE DON'T HAVE A 718 00:30:16,682 --> 00:30:19,418 GREAT UNDERSTANDING OF THE 719 00:30:19,418 --> 00:30:21,753 BEHAVIORS OF THE BRAIN SO WITH 720 00:30:21,753 --> 00:30:23,755 THIS IN MIND I WANT TO TALK 721 00:30:23,755 --> 00:30:26,725 BRIEFLY ABOUT WHAT NEUROMORPHIC 722 00:30:26,725 --> 00:30:28,827 IS. AND WE WILL HAVE MORE TALKS 723 00:30:28,827 --> 00:30:31,530 THAT WILL GO INTO MORE DEPTH IN 724 00:30:31,530 --> 00:30:34,032 THIS. NEUROMORPHIC CAN MEAN DIGITAL 725 00:30:34,032 --> 00:30:34,800 IMPLEMENTATIONS THAT KIND OF USE 726 00:30:34,800 --> 00:30:36,601 STANDARD FABRICATION 727 00:30:36,601 --> 00:30:37,536 TECHNOLOGIES, WE CAN GET 728 00:30:37,536 --> 00:30:40,338 AT SCALE. WE HAVE A SYSTEM AT 729 00:30:40,338 --> 00:30:42,441 SANDIA THAT'S OVER A BILLION 730 00:30:42,441 --> 00:30:44,810 NEURONS FROM INTEL AND WE'RE 731 00:30:44,810 --> 00:30:47,579 GOING TO GET ONE SOON FROM SPINNAKER 732 00:30:47,579 --> 00:30:49,881 , FROM THE SPINNAKER 733 00:30:49,881 --> 00:30:51,717 PROJECT. WE ALSO HAVE ANALOG 734 00:30:51,717 --> 00:30:53,018 SYSTEMS WHICH YOU'LL HEAR ABOUT 735 00:30:53,018 --> 00:30:54,386 IN A LITTLE BIT. SO I WILL 736 00:30:54,386 --> 00:30:55,854 FOCUS ON THE DIGITAL LARGE-SCALE 737 00:30:55,854 --> 00:30:56,788 SYSTEMS. THIS IS WHAT WE LOOK 738 00:30:56,788 --> 00:30:58,857 AT THE MOST. IT'S A LITTLE MORE 739 00:30:58,857 --> 00:31:01,927 NEAR TERM BUT BOTH ARE 740 00:31:01,927 --> 00:31:04,196 IMPORTANT. EVENTUALLY WE'LL 741 00:31:04,196 --> 00:31:08,133 PROBABLY HAVE NON-SILICON 742 00:31:08,133 --> 00:31:09,534 TECHNOLOGIES. WE'LL HAVE TO GO 743 00:31:09,534 --> 00:31:11,536 INTO ANALOG SCALING, MEMRISTORS. 744 00:31:11,536 --> 00:31:13,071 I WILL NOT TALK MUCH ABOUT THAT 745 00:31:13,071 --> 00:31:14,473 BUT THAT IS THE FUTURE OF THE 746 00:31:14,473 --> 00:31:16,074 FIELD AND SOMETHING WE WANT TO 747 00:31:16,074 --> 00:31:18,110 KEEP IN MIND. SO WHAT CAN 748 00:31:18,110 --> 00:31:20,812 NEUROMORPHIC COMPUTING DO? FOR 749 00:31:20,812 --> 00:31:21,546 THE LAST FEW MINUTES I WANT TO 750 00:31:21,546 --> 00:31:24,649 THROUGHOUT OUT A COUPLE OF THESE 751 00:31:24,649 --> 00:31:25,283 IDEAS. 752 00:31:25,283 --> 00:31:26,651 THERE ARE THREE THINGS I THINK 753 00:31:26,651 --> 00:31:28,954 WE ARE INTERESTED IN. ONE IS 754 00:31:28,954 --> 00:31:31,056 SIMULATE AT SCALE. THE SECOND IS 755 00:31:31,056 --> 00:31:32,390 GUIDE HOW TO ANALYZE DATA, I'M 756 00:31:32,390 --> 00:31:33,825 NOT GOING TO TALK AT ALL ABOUT 757 00:31:33,825 --> 00:31:37,095 THIS. BRAD THEILMAN WHO HAS A POSTER 758 00:31:37,095 --> 00:31:38,897 UPSTAIRS HAS A GREAT STORY ABOUT 759 00:31:38,897 --> 00:31:41,266 THAT I ENCOURAGE YOU TO TALK TO 760 00:31:41,266 --> 00:31:42,934 HIM. THE THIRD I WILL TOUCH ON 761 00:31:42,934 --> 00:31:44,469 BRIEFLY IS MAYBE WE CAN GET A 762 00:31:44,469 --> 00:31:46,037 BETTER INSIGHT INTO WHAT THE 763 00:31:46,037 --> 00:31:47,372 COMPUTING OF THE BRAIN ACTUALLY 764 00:31:47,372 --> 00:31:49,307 IS. SO VERY QUICKLY JUST AN 765 00:31:49,307 --> 00:31:50,976 EXAMPLE OF SIMULATING SYSTEMS AT 766 00:31:50,976 --> 00:31:53,278 SCALE. SO I MENTIONED, YOU 767 00:31:53,278 --> 00:31:57,182 KNOW, HUMAN SCALE IS TEN TO THE 768 00:31:57,182 --> 00:32:00,185 FIFTEENTHS SYNAPSES THAT IS NOT 769 00:32:00,185 --> 00:32:01,319 GOING TO HAPPEN ANYTIME SOON BUT 770 00:32:01,319 --> 00:32:03,522 FLYBRAIN, WE HAVE THE 771 00:32:03,522 --> 00:32:06,057 CONNECTOME, THE BRAIN INITIATIVE 772 00:32:06,057 --> 00:32:07,893 WAS INSTRUMENTAL IN SEEING THIS 773 00:32:07,893 --> 00:32:10,061 COME OUT. THIS IS ONE OF THE 774 00:32:10,061 --> 00:32:11,863 NINE NATURE PAPERS THEY HAVE 775 00:32:11,863 --> 00:32:13,532 THAT TALK ABOUT THE MODEL OF THE CONNECTOME. 776 00:32:13,532 --> 00:32:15,100 SO WE ASK THE QUESTION AND THIS IS 777 00:32:15,100 --> 00:32:17,502 JUST KIND OF A SHOT IN THE DARK, 778 00:32:17,502 --> 00:32:18,503 CAN OUR NEUROMORPHIC SYSTEMS 779 00:32:18,503 --> 00:32:20,438 JUST OUT OF THE BOX JUST TAKE 780 00:32:20,438 --> 00:32:21,940 THIS MODEL THAT WAS PUT UP 781 00:32:21,940 --> 00:32:23,208 THERE, DATA SHARING, YOU KNOW, 782 00:32:23,208 --> 00:32:26,011 THE WHOLE THING, CAN WE PUT THIS 783 00:32:26,011 --> 00:32:28,146 ON THE LOIHI CHIPS WHAT WE HAVE? AND THE 784 00:32:28,146 --> 00:32:29,681 ANSWER IS YES. USING SOME OF 785 00:32:29,681 --> 00:32:31,449 THE SOFTWARE TOOLS THAT WE'VE 786 00:32:31,449 --> 00:32:32,450 BEEN DEVELOPING FOR THIS IN 787 00:32:32,450 --> 00:32:33,618 MIND, YOU KNOW, KIND OF THE WAY 788 00:32:33,618 --> 00:32:35,687 OF STRESS TESTING THIS, WE WERE 789 00:32:35,687 --> 00:32:37,522 ABLE TO TAKE THIS CONNECTOME AND 790 00:32:37,522 --> 00:32:40,759 JUST, YOU KNOW, THEY PUT THEIR 791 00:32:40,759 --> 00:32:42,594 MODEL ONLINE. WE SET THIS 792 00:32:42,594 --> 00:32:49,234 THROUGH OUR TOOL CALLED FUGU AND 793 00:32:49,234 --> 00:32:50,302 STACS. AND THERE'S A LITTLE BIT 794 00:32:50,302 --> 00:32:51,636 OF WORK THAT GOES INTO THIS. I 795 00:32:51,636 --> 00:32:54,739 WILL NOT GO INTO THE 796 00:32:54,739 --> 00:32:55,774 NITTY-GRITTY DETAILS. THESE CHIPS ARE 797 00:32:55,774 --> 00:32:57,309 NOT MADE TO HAVE BIOLOGICAL CONNECTOMES ON THEM. 798 00:32:57,309 --> 00:32:59,211 WE'RE LEARNING A LOT ABOUT WHAT THE 799 00:32:59,211 --> 00:33:01,680 MAPPING IS AND IT WORKS AND THIS 800 00:33:01,680 --> 00:33:03,014 IS A BLOWUP OF THE PLOT. WE 801 00:33:03,014 --> 00:33:05,617 STILL NEED TO DIVE INTO WHAT IS 802 00:33:05,617 --> 00:33:08,787 ACTUALLY GOING ON HERE BUT THIS 803 00:33:08,787 --> 00:33:12,257 IS ENCOURAGING. THIS IS A REAL 804 00:33:12,257 --> 00:33:13,391 BRAIN, FULL BRAIN, ON A LOIHI CHIP WHICH 805 00:33:13,391 --> 00:33:16,161 IS PRETTY EXCITING. SO THE FINAL 806 00:33:16,161 --> 00:33:17,495 THING IS THIS IDEA OF CAN YOU 807 00:33:17,495 --> 00:33:21,666 ACTUALLY USE NEUROMORPHIC TO GET 808 00:33:21,666 --> 00:33:22,367 MORE EXPRESSIVE DESCRIPTIONS OF 809 00:33:22,367 --> 00:33:22,667 COMPUTATION? 810 00:33:22,667 --> 00:33:24,236 SO CAN WE GET AWAY FROM SIMPLE 811 00:33:24,236 --> 00:33:25,704 BEHAVIORS I USE THIS IDEA OF 812 00:33:25,704 --> 00:33:28,907 ANIMAL BEHAVIOR KIND OF LIMITS, 813 00:33:28,907 --> 00:33:32,010 TO DISCRIMINATE BETWEEN TWO 814 00:33:32,010 --> 00:33:33,144 OBJECTS. I WANT TO, YOU KNOW I 815 00:33:33,144 --> 00:33:34,679 DON'T HAVE A LOT OF TIME BUT I 816 00:33:34,679 --> 00:33:36,882 WANT TO SAY THAT A.I. ALONE IS NOT 817 00:33:36,882 --> 00:33:38,750 GOING TO DO THIS. IF YOU'RE 818 00:33:38,750 --> 00:33:40,585 FAMILIAR WITH WHY A.I. IS SO 819 00:33:40,585 --> 00:33:42,387 POWERFUL THERE'S A LOT OF THINGS 820 00:33:42,387 --> 00:33:45,657 LIKE UNIVERSAL FUNCTION APPROXIMATIONS, 821 00:33:45,657 --> 00:33:48,026 THE TURING COMPLETENESS NEURON. CAN 822 00:33:48,026 --> 00:33:49,728 DO ANYTHING SO THE FACT THAT A 823 00:33:49,728 --> 00:33:50,662 NEURAL NETWORK CAN DO SOMETHING 824 00:33:50,662 --> 00:33:51,963 DOESN'T MEAN THE BRAIN CAN DO 825 00:33:51,963 --> 00:33:53,231 IT. THAT'S A FUNDAMENTAL THING 826 00:33:53,231 --> 00:33:55,667 YOU HAVE TO KIND OF APPRECIATE 827 00:33:55,667 --> 00:33:58,169 HERE. NEUROMORPHIC HARDWARE ON 828 00:33:58,169 --> 00:33:59,137 THE OTHER HAND HAS REAL 829 00:33:59,137 --> 00:34:00,605 CONSTRAINTS. IT HAS TO PERFORM A 830 00:34:00,605 --> 00:34:03,575 COMPUTATION UNDER A FINITE SPACE 831 00:34:03,575 --> 00:34:09,080 AND ENERGY BUDGET AND THESE 832 00:34:09,080 --> 00:34:09,981 RESTRICTIONS ARE CRITICAL IN 833 00:34:09,981 --> 00:34:11,683 MOVING FORWARD. AGAIN, I'M NOT 834 00:34:11,683 --> 00:34:14,219 GOING TO GO INTO THE FULL DEPTH 835 00:34:14,219 --> 00:34:16,922 OF THIS BUT THESE CHIPS HAVE NEURONS 836 00:34:16,922 --> 00:34:18,256 THAT EXIST IN SPACE WITH A GRAPH AND 837 00:34:18,256 --> 00:34:19,958 THAT IS SOMETHING THAT WE HAVE 838 00:34:19,958 --> 00:34:23,461 TO FIT OUR ALGORITHMS TO TO BE 839 00:34:23,461 --> 00:34:24,863 IMPACTFUL. SO WHAT WE HAVE 840 00:34:24,863 --> 00:34:26,565 OBSERVED IS JUST A FINAL TEASE, 841 00:34:26,565 --> 00:34:28,800 I'M HAPPY TO TALK ABOUT THIS AT 842 00:34:28,800 --> 00:34:30,769 LENGTH IN DISCUSSION OR AT THE 843 00:34:30,769 --> 00:34:33,238 BREAK WE HAVE SEEN REAL CASES 844 00:34:33,238 --> 00:34:35,307 THAT NEUROMORPHIC HARDWARE SUCH 845 00:34:35,307 --> 00:34:36,408 AS THESE LOIHI PLATFORMS CAN DO 846 00:34:36,408 --> 00:34:39,544 PROBLEMS THAT REQUIRE SAMPLING 847 00:34:39,544 --> 00:34:43,715 LIKE MONTE CARLO SIMULATIONS 848 00:34:43,715 --> 00:34:48,019 OR MONITORING PROBLEMS THAT ARE 849 00:34:48,019 --> 00:34:49,054 SPATIALLY CONSTRAINED AND LOCALIZED 850 00:34:49,054 --> 00:34:51,289 OVER SPATIAL REGIONS. WE HAVE PAPERS 851 00:34:51,289 --> 00:34:54,326 THAT HAVE LOOKED AT WHERE WE CAN LOOK 852 00:34:54,326 --> 00:34:56,428 AT THIS PROCESS AS WELL AS 853 00:34:56,428 --> 00:35:00,498 FINITE ELEMENT SIMULATIONS. 854 00:35:00,498 --> 00:35:02,901 THESE MAP EXTREMELY WELL, FAR 855 00:35:02,901 --> 00:35:04,302 MORE THAN ANYONE WOULD HAVE 856 00:35:04,302 --> 00:35:05,770 GUESSED 10 YEARS AGO AND THEY 857 00:35:05,770 --> 00:35:07,973 SUGGEST THAT THE BRAIN COULD 858 00:35:07,973 --> 00:35:09,874 ACTUALLY VERY WELL BE DOING 859 00:35:09,874 --> 00:35:11,109 COMPUTATIONS THAT ARE ANALOGOUS 860 00:35:11,109 --> 00:35:13,144 TO THIS AND WE KNOW THEY DO 861 00:35:13,144 --> 00:35:15,647 ESPECIALLY WHEN IT COMES TO STOCHASTIC 862 00:35:15,647 --> 00:35:17,082 SAMPLING JUST TO KIND OF WRAP UP 863 00:35:17,082 --> 00:35:21,386 THIS CHALLENGE OF COMPLEXITY, 864 00:35:21,386 --> 00:35:21,953 TIMESCALES AND BEHAVIORS. I 865 00:35:21,953 --> 00:35:23,221 BELIEVE NEUROMORPHIC COMPUTING 866 00:35:23,221 --> 00:35:24,956 HAS A REAL POSSIBILITY OF, YOU 867 00:35:24,956 --> 00:35:26,891 KNOW, CONTRIBUTING TO MOVING 868 00:35:26,891 --> 00:35:29,761 FORWARD ALONG ALL THREE OF THESE 869 00:35:29,761 --> 00:35:31,563 SORT OF CHALLENGES. ESPECIALLY 870 00:35:31,563 --> 00:35:34,032 AS DATA STARTS BECOMING MORE 871 00:35:34,032 --> 00:35:37,335 VIABLE GOING FORWARD. SO, WITH 872 00:35:37,335 --> 00:35:39,838 THAT, I WOULD LIKE TO THANK YOU 873 00:35:39,838 --> 00:35:42,107 VERY MUCH AND I -- HOPEFULLY THE 874 00:35:42,107 --> 00:35:44,609 REMAINDER OF THE SESSION IS 875 00:35:44,609 --> 00:35:46,745 GOING TO HELP GO INTO THIS IN 876 00:35:46,745 --> 00:35:56,955 MORE DETAIL. THANK YOU. 877 00:35:56,955 --> 00:36:01,993 >> PLEASE WELCOME KWABENA 878 00:36:01,993 --> 00:36:03,762 BOAHEN FROM STANFORD UNIVERSITY. 879 00:36:03,762 --> 00:36:06,765 >> YES, MICS ARE WORKING, DO I 880 00:36:06,765 --> 00:36:08,933 ADVANCE OVER HERE? OH, I SEE, 881 00:36:08,933 --> 00:36:10,335 YEAH, SO IT'S GREAT TO BE HERE 882 00:36:10,335 --> 00:36:12,470 AND THANKS FOR THE INVITATION, 883 00:36:12,470 --> 00:36:13,972 EVERYONE AND I'M LOOKING FORWARD 884 00:36:13,972 --> 00:36:15,974 TO SOME MORE LIVELY DEBATE TODAY 885 00:36:15,974 --> 00:36:17,509 LIKE WE HAD YESTERDAY. I AM 886 00:36:17,509 --> 00:36:28,053 GOING TO TALK MORE ABOUT YEAH, 887 00:36:32,657 --> 00:36:34,359 SO I'M GOING TO TALK MORE ABOUT 888 00:36:34,359 --> 00:36:36,261 THE SUSTAINABILITY CRISIS THAT 889 00:36:36,261 --> 00:36:39,397 A.I. IS FACING RIGHT NOW. AND ONE 890 00:36:39,397 --> 00:36:40,999 OF THE PROMISES OF NEUROMORPHIC 891 00:36:40,999 --> 00:36:43,101 ENGINEERING IS TO TRY AND MATCH 892 00:36:43,101 --> 00:36:45,804 THE EFFICIENCY OF YOUR -- THE 893 00:36:45,804 --> 00:36:47,205 HUMAN BRAIN. AND ONE OF THE 894 00:36:47,205 --> 00:36:50,909 GOALS OF A.I. IS TO RECREATE HUMAN 895 00:36:50,909 --> 00:36:51,576 INTELLIGENCE, RIGHT? SO THESE 896 00:36:51,576 --> 00:36:54,012 TWO THINGS ARE RELATED. AND SO 897 00:36:54,012 --> 00:36:56,614 WE ALL KNOW THAT, YOU KNOW, 898 00:36:56,614 --> 00:36:59,651 MICROSOFT AND OPENAI ARE 899 00:36:59,651 --> 00:37:02,353 PLOTTING THIS HUNDRED BILLION 900 00:37:02,353 --> 00:37:04,456 DOLLAR A.I. SUPERCOMPUTER LAUNCHED 901 00:37:04,456 --> 00:37:07,692 AS SOON AS 2038 AND EXPANDING 902 00:37:07,692 --> 00:37:12,030 THROUGH 2030. WILL BE AS MUCH 903 00:37:12,030 --> 00:37:17,802 AS 5 GIGAWATTS OF ELECTRICITY SUPPLIED BY FIVE NUCLEAR POWER PLANTSTO RUN MILLIONS OF GPUS. 904 00:37:17,802 --> 00:37:20,171 SO WE HAVE TO ASK OURSELVES IS THIS 905 00:37:20,171 --> 00:37:21,306 SUSTAINABLE AND IF WE WANT TO 906 00:37:21,306 --> 00:37:24,342 FIX IT, WHAT DO WE HAVE TO DO? 907 00:37:24,342 --> 00:37:26,478 A LOT OF PEOPLE FOCUS ON THE 908 00:37:26,478 --> 00:37:27,712 AMOUNT OF ENERGY THAT GOES TO 909 00:37:27,712 --> 00:37:29,347 COMPUTE BUT IT TURNS OUT THAT 910 00:37:29,347 --> 00:37:31,783 MOST OF THIS ENERGY IS GOING TO 911 00:37:31,783 --> 00:37:32,717 COMMUNICATION. AND TO 912 00:37:32,717 --> 00:37:34,486 ILLUSTRATE THAT, I'M GOING TO 913 00:37:34,486 --> 00:37:38,890 LIKE SHOW YOU A LITTLE TOY 914 00:37:38,890 --> 00:37:41,359 NEURAL NETWORK. THIS ONE IS 915 00:37:41,359 --> 00:37:43,695 A DEEP NET WITH TEN LAYERS. 916 00:37:43,695 --> 00:37:46,965 EACH ONE OF THEM HAS TEN NEURONS 917 00:37:46,965 --> 00:37:49,134 AND YOU ALL KNOW YOU COME IN 918 00:37:49,134 --> 00:37:50,902 WITH THESE 10 INPUTS. YOU DISTRIBUTE THEM 919 00:37:50,902 --> 00:37:52,937 TO THE TEN NEURONS IN THAT LAYER, 920 00:37:52,937 --> 00:37:56,541 AND THEN YOU APPLY SOME 921 00:37:56,541 --> 00:37:56,875 WEIGHTS. AND THEN YOU SUM THOSE 922 00:37:56,875 --> 00:37:58,877 WEIGHTS, WEIGHTED INPUTS. 923 00:37:58,877 --> 00:38:00,378 AND YOU THRESHOLD AND 924 00:38:00,378 --> 00:38:02,213 THOSE ARE YOUR ACTIVATIONS THAT 925 00:38:02,213 --> 00:38:04,682 YOU SEND TO THE NEXT LAYER. SO 926 00:38:04,682 --> 00:38:07,418 THIS VECTOR IS YOUR MATRIX MULTIPLIER 927 00:38:07,418 --> 00:38:09,120 USED TO DO THIS EFFECTIVELY AND 928 00:38:09,120 --> 00:38:10,355 NORMALLY WE THINK OF THESE 929 00:38:10,355 --> 00:38:12,157 NEURAL NETWORKS AS FEED FORWARD 930 00:38:12,157 --> 00:38:13,858 BUT ACTUALLY MOST OF THE STATE 931 00:38:13,858 --> 00:38:16,561 OF THE ART NETWORKS WE'RE USING 932 00:38:16,561 --> 00:38:18,530 ARE NOT JUST FEED FORWARD THEY 933 00:38:18,530 --> 00:38:21,132 HAVE SKIP CONNECTIONS, RESIDUAL 934 00:38:21,132 --> 00:38:23,334 CONNECTIONS, SOMETIMES TOP-DOWN 935 00:38:23,334 --> 00:38:24,969 FEEDBACK LIKE IN AN ENCODER DECODER 936 00:38:24,969 --> 00:38:27,772 STACK OF A TRANSFORMER. SO YOU 937 00:38:27,772 --> 00:38:29,140 HAVE TO COMMUNICATE THESE 938 00:38:29,140 --> 00:38:31,509 SIGNALS AS WELL ACROSS SEVERAL 939 00:38:31,509 --> 00:38:34,179 LAYERS. NOW, PEOPLE ARE FOCUSED 940 00:38:34,179 --> 00:38:35,046 ON REALLY DOING THE COMPUTE 941 00:38:35,046 --> 00:38:36,181 EFFICIENTLY AND THEY HAVE COME 942 00:38:36,181 --> 00:38:38,149 UP WITH THIS CONCEPT OF 943 00:38:38,149 --> 00:38:40,885 COMPUTE-IN-MEMORY. AND IN THIS CASE, 944 00:38:40,885 --> 00:38:43,087 YOUR TEN BY TEN MATRIX IS STORED 945 00:38:43,087 --> 00:38:47,225 IN A TEN BY TEN MEMORY ARRAY. 946 00:38:47,225 --> 00:38:50,695 AND IN THIS CASE, YOUR 947 00:38:50,695 --> 00:38:53,464 ACTIVATIONS ARE APPLIED TO WHAT 948 00:38:53,464 --> 00:38:55,867 IS CALLED A WORD LINE OF THIS MEMORY. 949 00:38:55,867 --> 00:38:57,335 VOLTAGE CHANGES THE 950 00:38:57,335 --> 00:38:58,803 PROPERTIES OF THE MEMORY CELLS 951 00:38:58,803 --> 00:39:00,605 SUCH THAT IT MAKES A CURRENT 952 00:39:00,605 --> 00:39:02,273 THAT'S PROPORTIONAL TO SOME 953 00:39:02,273 --> 00:39:03,508 PHYSICAL STATE IN THE CELL SO 954 00:39:03,508 --> 00:39:05,143 THAT'S HOW YOU STORE YOUR WEIGHT 955 00:39:05,143 --> 00:39:07,412 AND DO YOUR MULTIPLICATION AND 956 00:39:07,412 --> 00:39:08,680 THEN YOU CAN BASICALLY SUM THESE 957 00:39:08,680 --> 00:39:10,114 CURRENTS ALONG WHAT IS CALLED 958 00:39:10,114 --> 00:39:11,749 THE BIT LINE AND YOU GOT YOUR 959 00:39:11,749 --> 00:39:13,685 DOT PRODUCT BETWEEN A VECTOR OF 960 00:39:13,685 --> 00:39:15,086 ACTIVATIONS AND A VECTOR OF 961 00:39:15,086 --> 00:39:16,487 WEIGHTS FOR A PARTICULAR UNIT OR NEURON. 962 00:39:16,487 --> 00:39:17,722 AND THEN YOU THRESHOLD AND YOU 963 00:39:17,722 --> 00:39:19,757 SEND THESE ACTIVATIONS TO SOME 964 00:39:19,757 --> 00:39:21,426 OTHER CROSSBAR. AND TO DO THAT 965 00:39:21,426 --> 00:39:23,895 YOU USE THESE INTERCONNECTS 966 00:39:23,895 --> 00:39:25,797 HERE. AND IF YOU HAD TO RUN 967 00:39:25,797 --> 00:39:27,265 THESE OUTPUT FROM THIS CROSSBAR 968 00:39:27,265 --> 00:39:29,968 , THIS MEMORY ARRAY TO ONE 969 00:39:29,968 --> 00:39:33,638 SEVERAL LAYERS AWAY THEN YOU CAN 970 00:39:33,638 --> 00:39:34,806 END UP RUNNING THESE WIRES AS 971 00:39:34,806 --> 00:39:39,110 FAR AS ABOUT, YOU KNOW, TEN BY 972 00:39:39,110 --> 00:39:41,980 TIMES TEN BECAUSE THEY RUN TEN 973 00:39:41,980 --> 00:39:44,115 BY TEN AND THERE ARE TEN OF THEM 974 00:39:44,115 --> 00:39:47,085 AND IF A FIXED FRACTION OF NEURONS 975 00:39:47,085 --> 00:39:48,386 SIGNAL EACH INFERENCE, THEN YOU HAVE 976 00:39:48,386 --> 00:39:51,055 A SET OF SIGNALS YOU'RE SENDING 977 00:39:51,055 --> 00:39:53,191 OUT PROPORTIONAL TO THE NUMBER 978 00:39:53,191 --> 00:39:54,058 OF NEURONS. SO THIS WILL SCALE IN ENERGY 979 00:39:54,058 --> 00:39:57,028 BECAUSE THE ENERGY IS PROPORTIONAL TO 980 00:39:57,028 --> 00:39:58,997 THE AMOUNT OF WIRE THAT YOU CHARGE 981 00:39:58,997 --> 00:40:00,431 OR DISCHARGE AS THE NUMBER OF 982 00:40:00,431 --> 00:40:03,368 NEURONS SQUARED. OF COURSE, YOU 983 00:40:03,368 --> 00:40:04,269 HAVE MORE LAYERS OF WIRING 984 00:40:04,269 --> 00:40:05,737 AND SO ON AND SO FORTH. 985 00:40:05,737 --> 00:40:06,804 YOU CAN IMPROVE THIS BUT THE 986 00:40:06,804 --> 00:40:08,806 IDEA IS THAT IF YOU HAVE A FIXED 987 00:40:08,806 --> 00:40:11,209 FRACTION OF SIGNALS THAT ARE 988 00:40:11,209 --> 00:40:13,978 TRAVELING A DISTANCE, THAT SCALES 989 00:40:13,978 --> 00:40:15,580 LINEARLY WITH THE NUMBER OF NEURONS, 990 00:40:15,580 --> 00:40:17,682 THEN YOU GET THIS QUADRATIC SCALE. 991 00:40:17,682 --> 00:40:19,050 AND IT'S ALL COMING FROM THE 992 00:40:19,050 --> 00:40:20,785 COMMUNICATION. NOT FROM THE 993 00:40:20,785 --> 00:40:21,352 COMPUTER. THIS COMPUTE IS DONE 994 00:40:21,352 --> 00:40:23,254 EFFICIENTLY IN THE CROSSBARS OR MEMORY ARRAYS. 995 00:40:23,254 --> 00:40:25,356 NOW IF YOU LOOK AT THE BRAIN 996 00:40:25,356 --> 00:40:27,058 IT'S MANAGED TO DO THIS TO 997 00:40:27,058 --> 00:40:28,993 COMMUNICATE THESE SIGNALS 998 00:40:28,993 --> 00:40:31,195 WITHOUT SCALING IN ENERGY QUADRATICALLY. 999 00:40:31,195 --> 00:40:33,064 PEOPLE HAVE ACTUALLY MEASURED THIS 1000 00:40:33,064 --> 00:40:37,168 FROM A MOUSE'S 70 MILLION NEURON BRAIN 1001 00:40:37,168 --> 00:40:40,238 TO A HUMAN'S 86 BILLIONG NEURON BRAIN. 1002 00:40:40,238 --> 00:40:41,773 THE ENERGY IS JUST A THOUSAND FOLD LARGER. 1003 00:40:41,773 --> 00:40:44,242 SO THIS IS A SLOPE OF ONE ON A 1004 00:40:44,242 --> 00:40:46,711 LOG-LOG SCALE OF NUMBER OF 1005 00:40:46,711 --> 00:40:50,548 NEURONS VERSUS ENERGY USE IN 1006 00:40:50,548 --> 00:40:52,517 GLUCOSE PER MINUTE. SO, YOU CAN 1007 00:40:52,517 --> 00:40:55,420 ASK YOURSELF, WHAT WOULD HAPPEN 1008 00:40:55,420 --> 00:41:00,358 IF BIOLOGICAL BRAIN SCALED QUADRATICALLY 1009 00:41:00,358 --> 00:41:02,126 INSTEAD OF LINEARLY? THAT 1010 00:41:02,126 --> 00:41:04,329 MEANS YOU EXTRAPOLATE THIS DATA 1011 00:41:04,329 --> 00:41:05,296 DOWN TO ONE NEURON, CHANGE THE 1012 00:41:05,296 --> 00:41:06,631 SLOPE TO TWO AND EXTRAPOLATE 1013 00:41:06,631 --> 00:41:08,199 BACK UP SO THAT'S WHAT YOU GET 1014 00:41:08,199 --> 00:41:10,468 AND SO IN THAT CASE, BY THE TIME 1015 00:41:10,468 --> 00:41:12,136 YOU GET TO 86 BILLION NEURONS 1016 00:41:12,136 --> 00:41:15,239 INSTEAD OF USING 25 WATTS YOU 1017 00:41:15,239 --> 00:41:17,475 WILL BE USING THREE TERAWATTS, 1018 00:41:17,475 --> 00:41:19,777 THAT'S THE ENTIRE U.S. ENERGY 1019 00:41:19,777 --> 00:41:21,279 CONSUMPTION. NOW THIS NUMBER 1020 00:41:21,279 --> 00:41:23,581 MAY SEEM OUTLANDISH, BUT, YOU 1021 00:41:23,581 --> 00:41:26,784 KNOW, STATE OF THE ART TRANSFORM 1022 00:41:26,784 --> 00:41:28,853 LARGE LANGUAGE MODELS OF ABOUT A 1023 00:41:28,853 --> 00:41:31,989 TRILLION PARAMETERS RIGHT NOW. 1024 00:41:31,989 --> 00:41:34,625 THAT'S LIKE A MOUSE-SCALE BRAIN AND 1025 00:41:34,625 --> 00:41:38,696 THEY WANT TO GET TO A 1026 00:41:38,696 --> 00:41:41,866 QUADRILLION WHICH IS A HUMAN- 1027 00:41:41,866 --> 00:41:46,471 SCALE BRAIN - A FACTOR OF 1000 - 1028 00:41:46,471 --> 00:41:52,844 TO TAKE YOU FROM FIVE 1029 00:41:52,844 --> 00:41:53,244 GIGAWATTS TO FIVE TERAWATTS. 1030 00:41:53,244 --> 00:41:54,579 GETTING THIS EXPONENT RIGHT, CHANGING 1031 00:41:54,579 --> 00:41:57,582 FROM QUADRATIC TO LINEAR IS CRITICAL 1032 00:41:57,582 --> 00:41:58,316 TO SUSTAINABLY FROM MOUSE SCALE TO 1033 00:41:58,316 --> 00:42:00,852 HUMAN SCALE. AND IT'S ALL COMING FROM 1034 00:42:00,852 --> 00:42:02,754 THE COMMUNICATION AND THIS IS TOUGH 1035 00:42:02,754 --> 00:42:05,323 BECAUSE, HOW DO YOU 1036 00:42:05,323 --> 00:42:07,892 REDUCE THE ENERGY OF 1037 00:42:07,892 --> 00:42:08,993 COMMUNICATION, YOU'RE GOING NEED 1038 00:42:08,993 --> 00:42:11,162 TO DO TWO THINGS. YOU SHORTEN THE 1039 00:42:11,162 --> 00:42:13,431 WIRES WHICH WILL USE LESS ENERGY 1040 00:42:13,431 --> 00:42:14,665 BECAUSE YOU USE LESS CHARGE TO 1041 00:42:14,665 --> 00:42:16,901 CHARGE THEM UP AND DOWN. AND YOU 1042 00:42:16,901 --> 00:42:19,103 REDUCE THE NUMBER OF SIGNALS, 1043 00:42:19,103 --> 00:42:20,271 LET'S LOOK AT THESE TWO THINGS 1044 00:42:20,271 --> 00:42:21,973 IF YOU DO BOTH OF THOSE THINGS, 1045 00:42:21,973 --> 00:42:24,742 YOU SHORTEN THE WIRES BY 1046 00:42:24,742 --> 00:42:25,910 BASICALLY STACKING THESE CROSSBARS 1047 00:42:25,910 --> 00:42:29,414 IN 3D SO THAT THOSE LONG 1048 00:42:29,414 --> 00:42:30,815 INTERCONNECTS ARE NOW GOING 1049 00:42:30,815 --> 00:42:32,950 VERTICALLY TO THE STACK THAT'S 1050 00:42:32,950 --> 00:42:34,719 GOING TO BE MUCH SHORTER AND BY 1051 00:42:34,719 --> 00:42:36,454 A SQUARE ROOT OF N IF YOU HAVE A 1052 00:42:36,454 --> 00:42:37,688 SYSTEM OF N NEURONS THAT YOU 1053 00:42:37,688 --> 00:42:40,058 ARRANGE IN THIS SQUARE ROOT OF N BY 1054 00:42:40,058 --> 00:42:45,630 SQUARE ROOT OF N BY SQUARE ROOT OF N CUBED. 1055 00:42:45,630 --> 00:42:47,198 THAT GETS YOU FROM QUADRATIC TO THREE 1056 00:42:47,198 --> 00:42:49,367 HALVES. TO GET FROM THREE HALVES TO 1057 00:42:49,367 --> 00:42:50,468 LINEAR YOU HAVE TO REDUCE THE 1058 00:42:50,468 --> 00:42:52,203 NUMBER OF SIGNALS YOU'RE SENDING BY 1059 00:42:52,203 --> 00:42:55,606 SQUARE ROOT OF N. SO INSTEAD OF A 1060 00:42:55,606 --> 00:42:59,377 LAYER WITH SQUARE ROOT OF N NEURONS 1061 00:42:59,377 --> 00:43:00,578 SENDING SIGNALS IT HAS TO SEND A 1062 00:43:00,578 --> 00:43:02,447 FIXED NUMBER OF SIGNALS, LET'S 1063 00:43:02,447 --> 00:43:07,652 SAY D. AS YOU MAKE THE NETWORK WIDER 1064 00:43:07,652 --> 00:43:08,953 AND DEEPER,YOU SEND THAT SAME D 1065 00:43:08,953 --> 00:43:10,521 SIGNAL BETWEEN LAYERS. THAT GET YOU 1066 00:43:10,521 --> 00:43:14,725 THE OTHER SQUARE ROOT OF N BUT YOU 1067 00:43:14,725 --> 00:43:17,328 CAN BASICALLY CHOOSE SPARSITY, RIGHT? 1068 00:43:17,328 --> 00:43:20,531 AND WITH THAT YOU CAN SCALE 1069 00:43:20,531 --> 00:43:21,799 LINEARLY INSTEAD OF QUADRATICALLY. THE FLIP SIDE OF 1070 00:43:21,799 --> 00:43:24,836 SENDING D SIGNALS IS WHEN A 1071 00:43:24,836 --> 00:43:26,871 LAYER OF SQUARE ROOT OF N NEURONS 1072 00:43:26,871 --> 00:43:29,507 RECEIVE D SIGNALS IT ALSO HAS TO 1073 00:43:29,507 --> 00:43:32,276 PRODUCE D SIGNALS. SO YOU NEED 1074 00:43:32,276 --> 00:43:34,512 NEURON MODELS THAT RESPOND MORE 1075 00:43:34,512 --> 00:43:35,613 SELECTIVELY THAN THEY DO RIGHT 1076 00:43:35,613 --> 00:43:37,648 NOW. NOW WITH THESE RELUS, HALF OF 1077 00:43:37,648 --> 00:43:39,283 YOUR ITEMS IN THE DATASET ARE 1078 00:43:39,283 --> 00:43:41,085 GOING TO ACTIVATE A NEURON. 1079 00:43:41,085 --> 00:43:43,688 ABOUT HALF OF THE NEURONS WILL 1080 00:43:43,688 --> 00:43:45,556 BE ADDED. WE'RE SAYING THAT 1081 00:43:45,556 --> 00:43:46,724 BASICALLY IT SHOULD BE MUCH 1082 00:43:46,724 --> 00:43:48,626 SMALLER THAN THAT AND SHOULD GET 1083 00:43:48,626 --> 00:43:50,862 SMALLER AND SMALLER FRACTION AS 1084 00:43:50,862 --> 00:43:54,699 YOU SCALE THESE SYSTEMS. AND SO 1085 00:43:54,699 --> 00:43:57,034 THAT'S THE SELECTIVITY SIDE. I AM 1086 00:43:57,034 --> 00:43:59,237 GOING TO WRAP THIS UP WITH I 1087 00:43:59,237 --> 00:44:01,205 THINK IN ORDER TO REALLY SCALE 1088 00:44:01,205 --> 00:44:02,874 THINGS SUSTAINABLY AND MATCH THE 1089 00:44:02,874 --> 00:44:04,442 ENERGY EFFICIENCY OF THE BRAIN, 1090 00:44:04,442 --> 00:44:08,012 WE HAVE TO SORT OF REVISIT WHAT 1091 00:44:08,012 --> 00:44:12,350 THESE ATEMS OF COMPUTATION AND 1092 00:44:12,350 --> 00:44:13,918 COMMUNICATION ARE BASICALLY. 1093 00:44:13,918 --> 00:44:16,320 AND I'M GOING TO REVIEW HOW OUR 1094 00:44:16,320 --> 00:44:17,355 CONCEPTIONS OF HOW THE BRAIN 1095 00:44:17,355 --> 00:44:19,090 LEARNS HAS EVOLVED OVER THE LAST 1096 00:44:19,090 --> 00:44:20,858 CENTURY, RIGHT? MOST OF THE 1097 00:44:20,858 --> 00:44:21,926 MODELS WE ARE BUILDING RIGHT NOW 1098 00:44:21,926 --> 00:44:24,862 ARE BASED ON THIS SYNAPTOCENTRIC 1099 00:44:24,862 --> 00:44:26,831 CONCEPT. AND THE THING 1100 00:44:26,831 --> 00:44:28,766 YOU COMMUNICATE IS THE 1101 00:44:28,766 --> 00:44:30,635 NONNEGATIVE PART OF THE 1102 00:44:30,635 --> 00:44:32,837 ACTIVATION AND THIS RELEASE TO 1103 00:44:32,837 --> 00:44:36,841 THESE DENSE SIGNALS AND THAT IS 1104 00:44:36,841 --> 00:44:39,944 BASICALLY ENERGETICALLY 1105 00:44:39,944 --> 00:44:41,913 UNSUSTAINABLE. PEOPLE IN THE 1106 00:44:41,913 --> 00:44:43,681 NEUROMORPHIC COMMUNITY, DECADES AGO, 1107 00:44:43,681 --> 00:44:44,982 MADE THE OBSERVATION THAT ACTUALLY 1108 00:44:44,982 --> 00:44:46,584 NEURONS COMMUNICATE WITH SPIKES. 1109 00:44:46,584 --> 00:44:48,619 OKAY. SO THE IDEA WAS TO GO TO 1110 00:44:48,619 --> 00:44:51,322 SPIKING TO SPECIFY THE ACTIVITY 1111 00:44:51,322 --> 00:44:53,991 BUT ACTUALLY THE THRESHOLD IN 1112 00:44:53,991 --> 00:44:56,394 YOUR RELU ACTIVATION COMES 1113 00:44:56,394 --> 00:44:58,229 FROM THAT THRESHOLD OF THE SPIKE 1114 00:44:58,229 --> 00:44:59,530 CONCEPT. IT DOESN'T CHANGE THE 1115 00:44:59,530 --> 00:45:02,366 FRACTION OF THE NEURONS THAT ARE 1116 00:45:02,366 --> 00:45:06,470 ACTIVATED IF YOU USE THE WEIGHTED SUM 1117 00:45:06,470 --> 00:45:08,673 EXTRACTION FOR YOUR BASIC ATOM OF COMPUTATION. 1118 00:45:08,673 --> 00:45:09,473 WE HAD TO THINK ABOUT HOW WE CAN 1119 00:45:09,473 --> 00:45:11,876 WITH A FIXED NUMBER OF SIGNALS, 1120 00:45:11,876 --> 00:45:12,944 ENCODE THE MOST NUMBER OF MESSAGES. 1121 00:45:12,944 --> 00:45:14,445 IF I HAVE THESE D SIGNALS 1122 00:45:14,445 --> 00:45:16,681 IN A LAYER. I CAN ACTUALLY 1123 00:45:16,681 --> 00:45:19,350 ENCODE A LOT MORE MESSAGES THAT 1124 00:45:19,350 --> 00:45:22,153 PAY ATTENTION TO WHAT 1125 00:45:22,153 --> 00:45:24,021 COMPUTATIONAL NEURONS IS ACTIVE 1126 00:45:24,021 --> 00:45:24,855 AS WELL AS THE ORDER IN WHICH 1127 00:45:24,855 --> 00:45:26,490 THOSE D NEURONS FIRE. THIS 1128 00:45:26,490 --> 00:45:28,826 LEADS TO THIS CONCEPT OF SPIKE 1129 00:45:28,826 --> 00:45:32,063 SEQUENCES AS THE ATOM OF 1130 00:45:32,063 --> 00:45:33,230 COMMUNICATION IN THE NERVOUS SYSTEM 1131 00:45:33,230 --> 00:45:34,832 AND OF COURSE THERE'S A LOT OF 1132 00:45:34,832 --> 00:45:36,934 INFORMATION OR A LOT OF DATA 1133 00:45:36,934 --> 00:45:38,102 SUPPORTING THIS NOW. 1134 00:45:38,102 --> 00:45:40,605 A LOT FROM THE HIPPOCAMPUS AND 1135 00:45:40,605 --> 00:45:41,906 ALSO, BUT THERE'S NOT MUCH 1136 00:45:41,906 --> 00:45:43,307 INFORMATION ABOUT HOW THESE 1137 00:45:43,307 --> 00:45:44,842 SEQUENCES ACTUALLY SORT OF 1138 00:45:44,842 --> 00:45:46,811 DECODED BY DOWNSTREAM NEURONS AND 1139 00:45:46,811 --> 00:45:48,412 SO ON AND SO FORTH AND THIS MAY 1140 00:45:48,412 --> 00:45:50,681 BE SOMETHING THAT THE DENDRITES 1141 00:45:50,681 --> 00:45:52,350 WITH ALL THE ACTIVE PROPERTIES 1142 00:45:52,350 --> 00:45:53,551 THAT YOU'VE HEARD ABOUT FROM YIOTA COULD 1143 00:45:53,551 --> 00:45:56,854 CONTRIBUTE TO SO THIS, THESE 1144 00:45:56,854 --> 00:45:58,022 ELEMENTS ARE REALLY REQUIRED FOR 1145 00:45:58,022 --> 00:46:00,925 US TO MATCH THAT SCALING, LINEAR 1146 00:46:00,925 --> 00:46:02,860 SCALING THAT THE BRAIN IS ABLE 1147 00:46:02,860 --> 00:46:05,196 TO ACHIEVE. AND I WROTE A 1148 00:46:05,196 --> 00:46:07,398 PERSPECTIVE OF THESE IDEAS AND 1149 00:46:07,398 --> 00:46:08,766 HOW DENDRITES COULD DO THIS 1150 00:46:08,766 --> 00:46:15,439 AND HOW THE WHOLE THING COULD 1151 00:46:15,439 --> 00:46:17,241 WORK IN THIS PUBLICATION, THIS IS THE 1152 00:46:17,241 --> 00:46:18,709 EIGHT MINUTE VERSION AND I'M OUT 1153 00:46:18,709 --> 00:46:20,344 OF TIME I GUESS AND I WOULD LIKE 1154 00:46:20,344 --> 00:46:22,146 TO THANK -- WE ARE MISSING THAT 1155 00:46:22,146 --> 00:46:24,949 LAST SLIDE. I UPDATED THE 1156 00:46:24,949 --> 00:46:27,752 SLIDES. I WOULD LIKE TO THANK 1157 00:46:27,752 --> 00:46:30,855 NSL FOR FUNDING THIS AND FOR THE 1158 00:46:30,855 --> 00:46:32,356 EFRI BRAID PROGRAM THAT GRACE SET 1159 00:46:32,356 --> 00:46:33,524 UP A COUPLE YEARS AGO. THANK 1160 00:46:33,524 --> 00:46:34,458 YOU VERY MUCH FOR YOUR 1161 00:46:34,458 --> 00:46:43,668 TIME. 1162 00:46:43,668 --> 00:46:48,873 >> PLEASE WELCOME JENNIFER HASLER 1163 00:46:48,873 --> 00:46:56,213 FROM GEORGIA TECH. 1164 00:46:56,213 --> 00:46:58,416 >> SO THANK YOU FOR HAVING ME 1165 00:46:58,416 --> 00:47:01,752 ME HERE AND THANK YOU FOR JUST THIS 1166 00:47:01,752 --> 00:47:05,623 WONDERFUL WORKSHOP SO THIS HAS 1167 00:47:05,623 --> 00:47:06,824 BEEN REALLY INTERESTING 1168 00:47:06,824 --> 00:47:07,625 DISCUSSIONS WE'RE HAVING HERE I 1169 00:47:07,625 --> 00:47:09,827 KNOW A LOT OF TIMES WHEN PEOPLE 1170 00:47:09,827 --> 00:47:11,028 SIT AND LISTEN USUALLY HEAR ME 1171 00:47:11,028 --> 00:47:13,064 TALK ABOUT STUFF LIKE OH, THERE 1172 00:47:13,064 --> 00:47:15,700 MUST BE SOME FLOATING GATES OR 1173 00:47:15,700 --> 00:47:17,234 DEVICES. WE ALWAYS DROP A 1174 00:47:17,234 --> 00:47:18,502 LITTLE BIT BUT I ACTUALLY WANT 1175 00:47:18,502 --> 00:47:19,804 TO TALK A LITTLE BIT MORE FROM A 1176 00:47:19,804 --> 00:47:22,473 BIGGER PICTURE ASIDE FROM THE 1177 00:47:22,473 --> 00:47:23,741 NEUROMORPHIC PERSPECTIVE AND IN 1178 00:47:23,741 --> 00:47:24,675 PARTICULAR, YOU KNOW, IF WE 1179 00:47:24,675 --> 00:47:26,310 WANTED TO START TALKING ABOUT 1180 00:47:26,310 --> 00:47:27,611 BUILDING BRAIN SCALE KIND OF 1181 00:47:27,611 --> 00:47:30,348 THINGS WHAT THAT WILL LOOK LIKE. 1182 00:47:30,348 --> 00:47:32,450 AND SO I WANT TO START FROM WHAT 1183 00:47:32,450 --> 00:47:35,152 WAS A PAPER THAT NOW IS ABOUT 1184 00:47:35,152 --> 00:47:36,821 ELEVEN YEARS OLD WAS SORT OF THE 1185 00:47:36,821 --> 00:47:39,757 ONE OF THE FIRST MAJOR SORT OF 1186 00:47:39,757 --> 00:47:41,292 NEUROMORPHIC ROADMAPS SAYING 1187 00:47:41,292 --> 00:47:44,128 COULD WE ACTUALLY TRY TO BUILD A 1188 00:47:44,128 --> 00:47:46,464 SILICON CORTEX? AND COULD WE DO 1189 00:47:46,464 --> 00:47:47,932 IT IN THAT SPACE AND THIS WAS 1190 00:47:47,932 --> 00:47:49,600 ACTUALLY A REALLY ROUGH THING 1191 00:47:49,600 --> 00:47:50,801 BECAUSE WITHIN NEUROMORPHIC 1192 00:47:50,801 --> 00:47:52,169 COMMUNITY AT THE TIME MULTIPLE 1193 00:47:52,169 --> 00:47:53,904 PEOPLE TRIED AND THAT TURNED OUT 1194 00:47:53,904 --> 00:47:55,673 TO BE THE PROBLEM EVERYONE WAS 1195 00:47:55,673 --> 00:47:59,110 KIND OF TERRIFIED BECAUSE IT 1196 00:47:59,110 --> 00:48:00,811 WOULD BE -- CREATE QUITE A LOT 1197 00:48:00,811 --> 00:48:02,813 OF FIRE SHALL WE SAY AND SO WITH 1198 00:48:02,813 --> 00:48:04,915 THIS KIND OF -- WAS ONE OF THOSE 1199 00:48:04,915 --> 00:48:06,050 FIRST DISCUSSIONS AND IT WAS 1200 00:48:06,050 --> 00:48:09,053 KIND OF IN THE AFFIRMATIVE AND 1201 00:48:09,053 --> 00:48:15,726 ENTIRELY IN SILICON. SO 7 TO 10 1202 00:48:15,726 --> 00:48:17,628 NANOMETER TECHNOLOGY. VERY 1203 00:48:17,628 --> 00:48:18,863 SPARSE CONNECTIVITY. EVERYTHING 1204 00:48:18,863 --> 00:48:21,098 THAT WE UNDERSTOOD AT THE POINT 1205 00:48:21,098 --> 00:48:22,032 LOOKED LIKE YOU COULD PROBABLY 1206 00:48:22,032 --> 00:48:23,067 DO THIS AND SOMETHING MIGHT BE 1207 00:48:23,067 --> 00:48:27,772 ABOUT A METER OR SO CUBED AND 1208 00:48:27,772 --> 00:48:29,440 50-100 WATTS OF POWER WHICH WAS 1209 00:48:29,440 --> 00:48:30,808 KIND OF AN INTERESTING 1210 00:48:30,808 --> 00:48:32,710 PERSPECTIVE OF IT AND ACTUALLY 1211 00:48:32,710 --> 00:48:33,744 THE FIGURES HERE ARE ACTUALLY 1212 00:48:33,744 --> 00:48:35,279 THE FIGURES FROM THAT. THERE 1213 00:48:35,279 --> 00:48:36,881 WAS A SIMILAR PAPER AFTERWARDS 1214 00:48:36,881 --> 00:48:38,582 FOR MORE PUBLIC CONSUMPTION BUT 1215 00:48:38,582 --> 00:48:39,917 THIS WAS AN INTERESTING THING. 1216 00:48:39,917 --> 00:48:42,052 ONE OF THE KEY CRITERIA WAS THE 1217 00:48:42,052 --> 00:48:44,388 ENERGY AND POWER WAS INCREDIBLY 1218 00:48:44,388 --> 00:48:46,190 IMPORTANT QUESTION HERE ALL THE 1219 00:48:46,190 --> 00:48:47,691 WAY THROUGH. IT'S ARGUABLY NOT 1220 00:48:47,691 --> 00:48:49,527 JUST IMPORTANT ON THE SILICON 1221 00:48:49,527 --> 00:48:51,562 SIDE AS A CONSTRAINT BUT ALSO IN 1222 00:48:51,562 --> 00:48:53,731 THE NEUROSCIENCE ITSELF AND SO 1223 00:48:53,731 --> 00:48:54,532 THERE'S VERY KEY PART. THE 1224 00:48:54,532 --> 00:48:56,467 INTERESTING THING IS I LOOK AT 1225 00:48:56,467 --> 00:48:58,636 IT ELEVEN YEARS AGO NOW AND 1226 00:48:58,636 --> 00:48:59,370 REALIZE THERE'S NOTHING I WOULD 1227 00:48:59,370 --> 00:49:00,704 CHANGE IN THE CONVERSATION WHICH 1228 00:49:00,704 --> 00:49:02,840 IS KIND OF GOOD OR TERRIFYING, I 1229 00:49:02,840 --> 00:49:04,809 DON'T KNOW. BUT IT'S ALSO A 1230 00:49:04,809 --> 00:49:07,411 CASE OF THE PROGRESS ON THIS 1231 00:49:07,411 --> 00:49:08,479 HASN'T BEEN ALL THAT FAST AND 1232 00:49:08,479 --> 00:49:10,181 THAT'S ONE OF THE QUESTIONS WE 1233 00:49:10,181 --> 00:49:11,682 MAY WANT TO ASK IN A 1234 00:49:11,682 --> 00:49:12,750 CONVERSATION LIKE THIS. HOW DO WE 1235 00:49:12,750 --> 00:49:14,819 START TO LOOK AT THAT? I WANT 1236 00:49:14,819 --> 00:49:16,353 TO WALK THROUGH THIS A LITTLE 1237 00:49:16,353 --> 00:49:17,488 BIT. I WOULD SAY ONE STARTS TO 1238 00:49:17,488 --> 00:49:19,356 GET INTO A CONVERSATION LIKE 1239 00:49:19,356 --> 00:49:21,025 THIS BECAUSE THERE'S BEEN AN 1240 00:49:21,025 --> 00:49:22,593 UNBELIEVABLE WORK DONE TO 1241 00:49:22,593 --> 00:49:23,994 UNDERSTAND THE NEUROMORPHIC 1242 00:49:23,994 --> 00:49:25,796 SPACE OBVIOUSLY LOTS OF THIS HAS 1243 00:49:25,796 --> 00:49:28,232 ITS ROOTS FROM PEOPLE LIKE 1244 00:49:28,232 --> 00:49:30,701 CARVER MEAD AND THE TEXTBOOKS 1245 00:49:30,701 --> 00:49:32,603 THAT WERE THERE AND THAT GENERAL 1246 00:49:32,603 --> 00:49:34,138 VIEW THAT WE SAW BOTH FROM 1247 00:49:34,138 --> 00:49:38,442 FEYNMAN BUT ALSO HOW CARVER 1248 00:49:38,442 --> 00:49:40,644 WOULD WRITE IT: IF YOU UNDERSTAND 1249 00:49:40,644 --> 00:49:42,746 IT YOU CAN BUILD IT AND I THINK 1250 00:49:42,746 --> 00:49:47,852 THAT IS HUGE. THIS IS A WHOLE 1251 00:49:47,852 --> 00:49:50,654 COMMUNITY, MULTIPLE COMMUNITIES 1252 00:49:50,654 --> 00:49:51,622 THAT SPAWNED. THERE'S SEVERAL 1253 00:49:51,622 --> 00:49:52,656 INDIVIDUALS THAT HAVE DONE 1254 00:49:52,656 --> 00:49:54,725 AMAZING WORK. AMAZING THINGS. 1255 00:49:54,725 --> 00:49:56,160 GERT AND KWABENA AND SEVERAL OTHER 1256 00:49:56,160 --> 00:49:57,962 PEOPLE IN THE ROOM HERE. 1257 00:49:57,962 --> 00:49:59,163 THERE'S BEEN A LOT 1258 00:49:59,163 --> 00:50:00,865 OF THINGS THAT HAVE BEEN BUILT 1259 00:50:00,865 --> 00:50:03,400 UP OVER DECADES ON THIS. A FEW 1260 00:50:03,400 --> 00:50:04,468 THINGS THAT DEFINITELY WE'VE 1261 00:50:04,468 --> 00:50:06,871 SPENT SOME TIME LOOKING AT, 1262 00:50:06,871 --> 00:50:08,939 OBVIOUSLY I'VE MENTIONED ALL OF 1263 00:50:08,939 --> 00:50:11,508 THE SINGLE TRANSISTOR SYNAPSE 1264 00:50:11,508 --> 00:50:13,644 CONCEPT WHICH IS THE FIRST KIND OF 1265 00:50:13,644 --> 00:50:14,945 NONVOLATILE CROSSBARS, ACTUALLY LED 1266 00:50:14,945 --> 00:50:17,681 TO THE NAMING OF COMPUTE-IN-MEMORY 1267 00:50:17,681 --> 00:50:19,550 MORE THAN TWENTY YEARS AGO. THE 1268 00:50:19,550 --> 00:50:21,185 CONCEPT OF TRANSISTOR CHANNEL 1269 00:50:21,185 --> 00:50:24,121 MODELS WHICH ALLOWS YOU TO HAVE 1270 00:50:24,121 --> 00:50:26,257 VERY TIGHT CONNECTIONS OF 1271 00:50:26,257 --> 00:50:29,059 SILICON. AND DISCUSSIONS ON 1272 00:50:29,059 --> 00:50:30,461 DENDRITES, ANALOG COMPUTING THEORY. 1273 00:50:30,461 --> 00:50:32,096 GIVE YOU AN OPPORTUNITY TO SAY, 1274 00:50:32,096 --> 00:50:33,264 COOL, WE CAN REALLY START TO 1275 00:50:33,264 --> 00:50:35,099 LOOK AT A MAP ON THE ENERGY 1276 00:50:35,099 --> 00:50:36,934 CONSTRAINT WHICH WAS ACTUALLY IN 1277 00:50:36,934 --> 00:50:37,902 THE ORIGINAL PAPER TALKING ABOUT 1278 00:50:37,902 --> 00:50:40,037 IF YOU LOOK AT THINGS IN TERMS 1279 00:50:40,037 --> 00:50:43,741 OF MAC: MULTIPLY-ACCUMULATE 1280 00:50:43,741 --> 00:50:44,575 YOU WANT TO BE CAREFUL OF 1281 00:50:44,575 --> 00:50:46,810 APPLES TO APPLES COMPARISON HERE 1282 00:50:46,810 --> 00:50:48,913 BUT IF YOU LOOK THE FACTOR OF 1000 1283 00:50:48,913 --> 00:50:50,047 IMPROVEMENT IN ENERGY EFFICIENCY FROM 1284 00:50:50,047 --> 00:50:52,783 THE FIRST DSPS IN THE LATE 70S 1285 00:50:52,783 --> 00:50:54,118 TO TODAY. THERE'S ABOUT A FACTOR 1286 00:50:54,118 --> 00:50:55,819 A THOUSAND OR MORE IMPROVEMENT. 1287 00:50:55,819 --> 00:50:58,856 WHAT WAS ORIGINALLY HYPOTHESIZED 1288 00:50:58,856 --> 00:51:01,058 BY CARVER MEAD IS JUST THE 1289 00:51:01,058 --> 00:51:03,761 ANALOG COMPUTING ITSELF WILL GET 1290 00:51:03,761 --> 00:51:05,129 YOU ANOTHER FACTOR OF A THOUSAND AND 1291 00:51:05,129 --> 00:51:06,263 NEUROMORPHIC COULD GO PAST THAT. 1292 00:51:06,263 --> 00:51:09,400 IT TURNS OUT THAT NOT IS THAT 1293 00:51:09,400 --> 00:51:11,902 ONLY TRUE EXPERIMENTALLY FOR NOW 1294 00:51:11,902 --> 00:51:13,203 TWENTY YEARS BUT NOW ALSO 1295 00:51:13,203 --> 00:51:14,471 THERE'S OTHER OPPORTUNITIES 1296 00:51:14,471 --> 00:51:16,807 PARTICULARLY AS YOU GO INTO MORE 1297 00:51:16,807 --> 00:51:18,575 NEUROMORPHIC ALGORITHMS, DENDRITES 1298 00:51:18,575 --> 00:51:21,412 AND SO FORTH THAT CAN ACTUALLY GET 1299 00:51:21,412 --> 00:51:23,948 MORE EFFICIENT. THAT EFFICIENCY THEN 1300 00:51:23,948 --> 00:51:27,451 STARTS TO LEAD YOU OUT INTO A 1301 00:51:27,451 --> 00:51:28,352 WHOLE RANGE OF QUESTIONS 1302 00:51:28,352 --> 00:51:30,354 PARTICULARLY IF YOU LOOK AT THE 1303 00:51:30,354 --> 00:51:31,855 DENDRITIC STRUCTURE. ONE OF THE 1304 00:51:31,855 --> 00:51:33,424 THINGS CLOSE TO MY HEART WHERE 1305 00:51:33,424 --> 00:51:34,792 WE STARTED TO REALIZE THAT YOU 1306 00:51:34,792 --> 00:51:38,429 COULD ACTUALLY DO THINGS WITH 1307 00:51:38,429 --> 00:51:42,466 DENDRITES THAT LOOK LIKE HIDDEN 1308 00:51:42,466 --> 00:51:43,534 MARKOV MODEL CLASSIFIERS AND THERE'S 1309 00:51:43,534 --> 00:51:44,401 A DIRECTION BETWEEN THESE 1310 00:51:44,401 --> 00:51:45,269 TECHNIQUES THAT EVEN PEOPLE IN 1311 00:51:45,269 --> 00:51:47,604 THE SPEECH COMMUNITY GO, THOSE 1312 00:51:47,604 --> 00:51:49,073 ARE HARD TO DO IN FULL 1313 00:51:49,073 --> 00:51:50,808 CAPABILITY AND WE REALIZED, 1314 00:51:50,808 --> 00:51:55,079 WAIT, THE NEURONS WITH THE 1315 00:51:55,079 --> 00:51:55,412 DENDRITES ARE DOING SUCH 1316 00:51:55,412 --> 00:51:57,881 AMAZING COMPUTATION AND ACTUALLY 1317 00:51:57,881 --> 00:51:58,515 BE ABLE TO DEMONSTRATE ENTIRELY 1318 00:51:58,515 --> 00:52:02,453 IN NEUROMORPHIC STRUCTURES. IT LEADS 1319 00:52:02,453 --> 00:52:03,053 YOU TO KNOW THAT ENCODING IS 1320 00:52:03,053 --> 00:52:07,424 CRITICAL. EVERY SPIKE MATTERS. 1321 00:52:07,424 --> 00:52:10,427 SPIKE RATES ARE QUITE LOW AND 1322 00:52:10,427 --> 00:52:11,762 THAT THERE'S A LOT OF IMPORTANCE 1323 00:52:11,762 --> 00:52:14,298 IN THAT. YOU SEE THAT IN THE 1324 00:52:14,298 --> 00:52:16,934 MODELS. I LOVE THE STUFF THAT 1325 00:52:16,934 --> 00:52:19,403 CAME OUT OF MARK TILDEN'S GROUP AND 1326 00:52:19,403 --> 00:52:22,439 ALSO TALKING ABOUT OPTIMAL PATH 1327 00:52:22,439 --> 00:52:23,340 PLANNING AND DOING THIS WITH HODGKIN- 1328 00:52:23,340 --> 00:52:24,875 HOXLEY NEURONS, AND THEY 1329 00:52:24,875 --> 00:52:28,645 DO MATTER FROM AN 1330 00:52:28,645 --> 00:52:30,147 ALGORITHMIC SENSE AND ALL THESE 1331 00:52:30,147 --> 00:52:31,815 THINGS CREATE OPPORTUNITIES BOTH 1332 00:52:31,815 --> 00:52:33,717 IN ENERGY EFFICIENCY AND MORE 1333 00:52:33,717 --> 00:52:34,618 IMPROVED ALGORITHMS ALL THE WAY 1334 00:52:34,618 --> 00:52:36,053 THROUGH. THE PROBLEM IS WE ONLY 1335 00:52:36,053 --> 00:52:37,388 HAVE A FEW OF THESE OVER TIME. 1336 00:52:37,388 --> 00:52:40,257 AND SO I THINK THE SPACE IS -- 1337 00:52:40,257 --> 00:52:41,392 IN TERMS OF CONCEPTS IS REALLY 1338 00:52:41,392 --> 00:52:43,494 WHERE THERE'S A LOT FOR US TO 1339 00:52:43,494 --> 00:52:45,496 GROW ON. AND SO IF I WAS 1340 00:52:45,496 --> 00:52:46,964 LOOKING AT THIS, ON THE ROADMAP, 1341 00:52:46,964 --> 00:52:48,832 I MEAN, LIKE I SAID, THERE'S 1342 00:52:48,832 --> 00:52:51,935 STILL A REALLY GOOD MAP GOING 1343 00:52:51,935 --> 00:52:54,071 FORWARD. AND I THINK THE THINGS 1344 00:52:54,071 --> 00:52:56,607 THAT CARVER TALKED ABOUT MAKE 1345 00:52:56,607 --> 00:53:02,613 SENSE. I THINK WHAT'S REALLY 1346 00:53:02,613 --> 00:53:03,881 HELPED IS OUR INFRASTRUCTURE HAS 1347 00:53:03,881 --> 00:53:08,052 DONE WELL ON THE ANALOG AND MIXED 1348 00:53:08,052 --> 00:53:13,457 SIGNAL SIDE. WE HAVE A CONCEPTUAL IDEA 1349 00:53:13,457 --> 00:53:15,426 WE ARE IN GOOD SHAPE. EVEN IN THE 1350 00:53:15,426 --> 00:53:16,126 EARLIEST OF NEUROMORPHIC ENGINEERING 1351 00:53:16,126 --> 00:53:20,898 THERE WAS A SENSE THAT NEUROBIOLOGY 1352 00:53:20,898 --> 00:53:22,433 AND ELECTRONIC COMPUTATION SHOULD 1353 00:53:22,433 --> 00:53:23,167 BE BIDIRECTIONALLY 1354 00:53:23,167 --> 00:53:24,435 HELPING EACH OTHER. I KNOW 1355 00:53:24,435 --> 00:53:25,936 WE HEARD A LOT OF THAT ALREADY 1356 00:53:25,936 --> 00:53:28,272 AND SO I AM GLAD THIS IS ALL 1357 00:53:28,272 --> 00:53:29,640 CONSISTENT. THE ISSUE THEN 1358 00:53:29,640 --> 00:53:31,341 BECOMES THOUGH IS THAT THE 1359 00:53:31,341 --> 00:53:31,975 PROGRESS HAS ACTUALLY BEEN 1360 00:53:31,975 --> 00:53:33,710 RATHER SLOW AND HOW DO WE JUST 1361 00:53:33,710 --> 00:53:34,812 KIND OF KEEP PROGRESSING THIS? 1362 00:53:34,812 --> 00:53:37,081 AND I THINK THIS WAS A 1363 00:53:37,081 --> 00:53:37,714 CONCEPTUAL QUESTION. THERE'S 1364 00:53:37,714 --> 00:53:39,983 SOME THINGS THAT START TO MAKE 1365 00:53:39,983 --> 00:53:42,619 SENSE AND FROM SOME SORT OF 1366 00:53:42,619 --> 00:53:45,255 SMALL CONCEPTS AND SOME LARGER 1367 00:53:45,255 --> 00:53:46,023 CONCEPTS ON WHAT WE CAN BUILD 1368 00:53:46,023 --> 00:53:49,193 ONE OF THE FIRST THINGS I CAN 1369 00:53:49,193 --> 00:53:50,494 SAY RIGHT OFF THE BAT IS WE WANT 1370 00:53:50,494 --> 00:53:54,965 TO FOCUS ON WHAT WE CAN DO WITH 1371 00:53:54,965 --> 00:53:56,700 SILICON FABRICATION. IF WE WANT 1372 00:53:56,700 --> 00:53:58,836 TO PUSH THE ALGORITHMS WE WANT 1373 00:53:58,836 --> 00:54:01,238 TO FOCUS ON THE SPACES AND SEE 1374 00:54:01,238 --> 00:54:03,207 HOW MUCH WE CAN DRIVE IN THAT 1375 00:54:03,207 --> 00:54:03,507 DIRECTION. 1376 00:54:03,507 --> 00:54:04,475 WE NEED TO THINK ABOUT WHAT WE 1377 00:54:04,475 --> 00:54:07,111 CAN DO WITH MORE COMPLEX NEURON 1378 00:54:07,111 --> 00:54:09,012 MODELS, PARTICULARLY WITH DENDRITES, 1379 00:54:09,012 --> 00:54:10,781 WITH CHANNEL MODELS AND ALSO BE 1380 00:54:10,781 --> 00:54:12,983 CAREFUL WITH THE ENCODING. AS 1381 00:54:12,983 --> 00:54:14,551 MUCH AS WE CAN WE NEED TO BE 1382 00:54:14,551 --> 00:54:17,421 AVOIDING RATE ENCODING STRUCTURES, 1383 00:54:17,421 --> 00:54:18,989 INTEGRATE-AND-FIRE NEURONS, ONLY ASSUME THOSE 1384 00:54:18,989 --> 00:54:22,226 WOULD DEFAULT IN HIGH LEVEL 1385 00:54:22,226 --> 00:54:23,060 STIMULATION. THE 1386 00:54:23,060 --> 00:54:24,094 PROGRAMMABILITY IS ESSENTIAL. IF 1387 00:54:24,094 --> 00:54:25,395 YOU'RE GOING TO BUILD AN ANALOG 1388 00:54:25,395 --> 00:54:27,998 SYSTEM I WILL TELL YOU THAT YOU 1389 00:54:27,998 --> 00:54:30,467 ABSOLUTELY HAVE TO HAVE THE 1390 00:54:30,467 --> 00:54:32,436 ACCURACY SOMEWHERE. YOU WILL 1391 00:54:32,436 --> 00:54:34,571 JUST NOT LEARN ABOUT EVERY 1392 00:54:34,571 --> 00:54:36,907 ASPECT OF MISMATCH. AS MUCH AS 1393 00:54:36,907 --> 00:54:40,944 I'VE HOPED THAT I'VE NEVER 1394 00:54:40,944 --> 00:54:42,446 SEEN IT. THERE'S NEED TO TRAINE 1395 00:54:42,446 --> 00:54:43,647 STUDENTS AND HAVE IC FAB 1396 00:54:43,647 --> 00:54:46,250 AVAILABLE AND ALL THE TOOLS 1397 00:54:46,250 --> 00:54:47,718 AROUND IT AND WHAT I WOULD WANT 1398 00:54:47,718 --> 00:54:49,253 TO PROPOSE AROUND THIS IS 1399 00:54:49,253 --> 00:54:50,921 IMAGINE, YOU CAN START TALKING 1400 00:54:50,921 --> 00:54:58,795 ABOUT A NEUROBIOLOGICAL ALGORITHM OF 1401 00:54:58,795 --> 00:55:00,564 A CORTICAL COLUMN THAT YOU CAN USE 1402 00:55:00,564 --> 00:55:03,333 FOR AN ENGINEERING APPLICATION. TAKE THESE IDEAS OF A WHOLE 1403 00:55:03,333 --> 00:55:06,036 BUNCH OF PYRAMIDAL CELL-TYPE NEURONS 1404 00:55:06,036 --> 00:55:06,670 INTERNEURONS AND ACTUALLY WORK AT 1405 00:55:06,670 --> 00:55:09,273 THIS TAKING THE CORTICAL 1406 00:55:09,273 --> 00:55:10,641 COLUMNS WITH ALL THAT COMPLEXITY 1407 00:55:10,641 --> 00:55:12,543 AND KIND OF BUILDING THAT UP 1408 00:55:12,543 --> 00:55:13,977 INTO MULTIPLE HIGHER AND HIGHER 1409 00:55:13,977 --> 00:55:15,779 LAYERS OF COMPLEXITY AS WE BUILD 1410 00:55:15,779 --> 00:55:19,516 UP. I THINK THE OPPORTUNITY IS 1411 00:55:19,516 --> 00:55:20,517 HUGE I THINK THE FUTURE IS 1412 00:55:20,517 --> 00:55:21,952 BRIGHT AND I THINK IT'S NOW, YOU 1413 00:55:21,952 --> 00:55:23,153 KNOW, THE OPPORTUNITY TO REALLY 1414 00:55:23,153 --> 00:55:27,224 TRY SOME OF THESE IDEAS. THANK 1415 00:55:27,224 --> 00:55:32,729 YOU 1416 00:55:32,729 --> 00:55:32,963 YOU. 1417 00:55:32,963 --> 00:55:35,265 >> PLEASE WELCOME FRANCIS CHANCE 1418 00:55:35,265 --> 00:55:36,533 FROM SANDIA NATIONAL LABS. 1419 00:55:36,533 --> 00:55:40,237 >> HI, WE'VE BEEN TALKING ABOUT 1420 00:55:40,237 --> 00:55:43,173 THE VIRTUOUS CYCLE BETWEEN 1421 00:55:43,173 --> 00:55:45,375 NEUROSCIENCE AND ARTIFICIAL 1422 00:55:45,375 --> 00:55:47,878 INTELLIGENCE. I WANTED TO TALK 1423 00:55:47,878 --> 00:55:49,379 ABOUT THE CYCLE BETWEEN SCIENCE 1424 00:55:49,379 --> 00:55:54,318 AND COMPUTING AND WHAT KIND OF 1425 00:55:54,318 --> 00:55:55,352 QUESTIONS WE CAN ASK AND THEN I 1426 00:55:55,352 --> 00:55:58,088 WILL TALK ABOUT HOW THIS CAN 1427 00:55:58,088 --> 00:56:01,959 IMPACT NEUROAI. 1428 00:56:01,959 --> 00:56:05,195 SO FOR ME I THINK WE SHOULD BE 1429 00:56:05,195 --> 00:56:06,863 FOCUSING OR WE COULD ESTABLISH A 1430 00:56:06,863 --> 00:56:09,733 LOT OF SYNERGY BY FOCUSING ON 1431 00:56:09,733 --> 00:56:10,767 NEURAL PRIMITIVES. 1432 00:56:10,767 --> 00:56:13,704 AND I KNOW THIS WORD WILL GET 1433 00:56:13,704 --> 00:56:15,472 OVERLOADED AS WE GET INTO THIS 1434 00:56:15,472 --> 00:56:18,308 SESSION. I WANT TO KNOW WHAT 1435 00:56:18,308 --> 00:56:20,544 KINDS OF COMPUTATIONS, 1436 00:56:20,544 --> 00:56:24,982 BIOLOGICAL CIRCUITS COMPUTE 1437 00:56:24,982 --> 00:56:27,951 OVER AND OVER AGAIN ACROSS 1438 00:56:27,951 --> 00:56:28,218 DIFFERENT BIOLOGICAL SYSTEMS 1439 00:56:28,218 --> 00:56:31,321 TO ME, THEY ARE SLIGHTLY A LEVEL 1440 00:56:31,321 --> 00:56:33,657 HIGHER THAN BASIC OPERATIONS SO 1441 00:56:33,657 --> 00:56:35,659 I WOULD SAY THIS IS A LITTLE BIT 1442 00:56:35,659 --> 00:56:38,161 HIGHER OF THE ATOMS OF COMPUTATION 1443 00:56:38,161 --> 00:56:39,930 THAT KWABENA JUST TALKED ABOUT. 1444 00:56:39,930 --> 00:56:41,565 THEY'RE NOT THE BEHAVIOR 1445 00:56:41,565 --> 00:56:42,899 ITSELF. WE WANT TO UNDERSTAND 1446 00:56:42,899 --> 00:56:44,401 HOW THESE NEURAL PRIMITIVES 1447 00:56:44,401 --> 00:56:45,969 SUPPORT ANIMAL BEHAVIORS SO WE 1448 00:56:45,969 --> 00:56:47,871 UNDERSTAND HOW TO IMPLEMENT THEM 1449 00:56:47,871 --> 00:56:49,439 IN NEUROMORPHIC SYSTEMS TO SUPPORT 1450 00:56:49,439 --> 00:56:50,841 DIFFERENT TYPES OF TASKS. 1451 00:56:50,841 --> 00:56:52,409 ALL RIGHT, SO I'M GOING TO GIVE 1452 00:56:52,409 --> 00:56:53,944 YOU AN EXAMPLE OF MY OWN 1453 00:56:53,944 --> 00:56:56,480 RESEARCH TO GIVE YOU SOME IDEA 1454 00:56:56,480 --> 00:56:58,548 OF WHAT I MEAN BY PRIMITIVES AND 1455 00:56:58,548 --> 00:57:00,183 HOW WE USE THAT TO DRIVE 1456 00:57:00,183 --> 00:57:01,351 NEUROMORPHIC DEVELOPMENT AS WELL 1457 00:57:01,351 --> 00:57:03,553 AS THINK ABOUT HOW WE CAN IMPLEMENT 1458 00:57:03,553 --> 00:57:05,022 THESE NEURAL PRIMITIVES IN DIFFERENT 1459 00:57:05,022 --> 00:57:07,190 ALGORITHMS FOR NEUROMORPHIC TASKS. 1460 00:57:07,190 --> 00:57:08,325 I'VE BEEN INTERESTED IN COORDINATE 1461 00:57:08,325 --> 00:57:09,359 TRANSFORMATIONS. HERE IS AN 1462 00:57:09,359 --> 00:57:11,495 EXAMPLE OF THE BEHAVIOR THAT IS 1463 00:57:11,495 --> 00:57:13,397 SUPPORTED BY A TRANSFORMATION. 1464 00:57:13,397 --> 00:57:15,499 I WILL TALK ABOUT WHAT THE 1465 00:57:15,499 --> 00:57:16,833 COORDINATE TRANSFORMATION IS IN 1466 00:57:16,833 --> 00:57:18,535 A SECOND BUT WE HAVE A MONKEY, 1467 00:57:18,535 --> 00:57:22,939 IT'S TRAIN TO FIXATE AT A SPOT 1468 00:57:22,939 --> 00:57:26,143 IN FRONT OF IT. IT THEN NEEDS TO 1469 00:57:26,143 --> 00:57:27,711 REACH A TARGET THAT APPEARS AT SOME 1470 00:57:27,711 --> 00:57:29,212 ANGLE AND DISTANCE AWAY AS THE 1471 00:57:29,212 --> 00:57:32,182 FIXATION POINT IS MOVED TO 1472 00:57:32,182 --> 00:57:34,318 THE RIGHT AND THE MONKEY 1473 00:57:34,318 --> 00:57:36,853 TURNS ITS HEAD BUT NOT ITS BODY 1474 00:57:36,853 --> 00:57:39,122 TO MAINTAIN FIXATION AND THE 1475 00:57:39,122 --> 00:57:40,724 MONKEY NEEDS TO REACH OUT AND 1476 00:57:40,724 --> 00:57:42,459 TOUCH THAT TARGET AGAIN. THE 1477 00:57:42,459 --> 00:57:44,161 ARM'S TAKING A DIFFERENT 1478 00:57:44,161 --> 00:57:44,995 TRAJECTORY EVEN THOUGH THE SAME 1479 00:57:44,995 --> 00:57:46,630 IMAGE IS FALLING ON THE MONKEY'S 1480 00:57:46,630 --> 00:57:48,732 EYES SO THE MONKEY'S NERVOUS 1481 00:57:48,732 --> 00:57:50,100 SYSTEM NEEDS TO PERFORM 1482 00:57:50,100 --> 00:57:51,535 COORDINATE TRANSFORMATION FROM 1483 00:57:51,535 --> 00:57:53,103 THE VISUAL INFORMATION THAT'S 1484 00:57:53,103 --> 00:57:55,005 COMING IN WHICH IS GOING TO BE 1485 00:57:55,005 --> 00:57:58,742 EYE-CENTRIC COORDINATES AND 1486 00:57:58,742 --> 00:58:00,110 TRANSFORM THAT INFORMATION INTO 1487 00:58:00,110 --> 00:58:01,511 MOTOR COMMAND THAT'S GOING TO BE 1488 00:58:01,511 --> 00:58:03,680 BODY-CENTRIC COORDINATES OR 1489 00:58:03,680 --> 00:58:06,483 WHAT TRAJECTORY DOES THE ARM 1490 00:58:06,483 --> 00:58:08,051 NEED TO MAKE TO REACH THAT 1491 00:58:08,051 --> 00:58:09,419 TARGET. IF WE LOOK AT THE 1492 00:58:09,419 --> 00:58:11,321 LITERATURE THE MODELS GENERALLY 1493 00:58:11,321 --> 00:58:13,924 COMBINE TWO DIFFERENT TYPES OF 1494 00:58:13,924 --> 00:58:16,927 INFORMATION. VISUAL OR SENSORY 1495 00:58:16,927 --> 00:58:23,300 AND PROPRIOCEPTIVE INFORMATION. 1496 00:58:23,300 --> 00:58:24,768 AND THESE ARE COMBINED INTO BASIS FUNCTIONS 1497 00:58:24,768 --> 00:58:25,802 THAT REQUIRE INTERACTION AND 1498 00:58:25,802 --> 00:58:29,973 THEN THE OUTPUT IS READ OUT IN 1499 00:58:29,973 --> 00:58:31,475 THE DESIRED REFERENCE FRAME BY 1500 00:58:31,475 --> 00:58:33,377 SOME CONFIGURATION OF SYNAPTIC 1501 00:58:33,377 --> 00:58:36,279 WEIGHTS. I WILL TRY THIS ONCE 1502 00:58:36,279 --> 00:58:36,980 WHICH CAN BE CONFIGURED HERE 1503 00:58:36,980 --> 00:58:38,648 BETWEEN THE BASIS FUNCTIONS AND 1504 00:58:38,648 --> 00:58:45,689 THE OUTPUT LAYER. NOW THERE ARE 1505 00:58:45,689 --> 00:58:47,657 MANY MODELS OF THIS. I'VE BEEN 1506 00:58:47,657 --> 00:58:49,025 SPECIFICALLY INTERESTED IN 1507 00:58:49,025 --> 00:58:51,061 MODELS IN THE PARIETAL CORTEX, I 1508 00:58:51,061 --> 00:58:52,229 HAVE THREE REFERENCES HERE AND 1509 00:58:52,229 --> 00:58:54,331 I'VE BEEN INTERESTED IN THESE 1510 00:58:54,331 --> 00:58:55,399 BECAUSE THEY SPECIFICALLY 1511 00:58:55,399 --> 00:58:57,701 PREDICT THAT THE BASIS FUNCTIONS 1512 00:58:57,701 --> 00:59:00,337 ARE FORMED BY MULTIPLICATION 1513 00:59:00,337 --> 00:59:02,205 OF THE VISUAL INPUT BY THE 1514 00:59:02,205 --> 00:59:04,741 PROPRIOCEPTIVE INPUT AND THERE'S 1515 00:59:04,741 --> 00:59:06,843 EVIDENCE OF THIS IN THE 1516 00:59:06,843 --> 00:59:07,044 NEUROBIOLOGICAL LITERATURE THROUGH 1517 00:59:07,044 --> 00:59:07,911 THE GAIN FIELDS DISCOVERED BY 1518 00:59:07,911 --> 00:59:10,180 ANDERSON AND MOUNTCASTLE IN THE 1980S 1519 00:59:10,180 --> 00:59:12,282 WHAT WE HAVE HERE IS AN EXAMPLE 1520 00:59:12,282 --> 00:59:14,117 OF A VISUAL NEURON. THE FIRING 1521 00:59:14,117 --> 00:59:17,120 RATES ON THE Y AXIS. THE X AXIS IS 1522 00:59:17,120 --> 00:59:19,923 THE POSITION OF THE VISUAL STIMULUS 1523 00:59:19,923 --> 00:59:21,591 THE DIFFERENCE BETWEEN THE TWO TUNING CURVES IS THE DIFFERENCE INTHE HEAD POSITION 1524 00:59:21,591 --> 00:59:23,894 RELATIVE TO THE BODY. AND WHAT YOU 1525 00:59:23,894 --> 00:59:25,228 SEE IS IT'S MULTIPLICATIVE. THE 1526 00:59:25,228 --> 00:59:27,964 EFFECTIVE HEAD POSITION ON THE 1527 00:59:27,964 --> 00:59:30,634 VISUAL RESPONSE IS 1528 00:59:30,634 --> 00:59:32,335 MULTIPLICATIVE SCALING SO WE'RE 1529 00:59:32,335 --> 00:59:33,503 SEEING EVIDENCE OF THIS 1530 00:59:33,503 --> 00:59:34,271 MULTIPLICATION OF THE TWO 1531 00:59:34,271 --> 00:59:36,239 DIFFERENT MODALITIES OF INPUT. 1532 00:59:36,239 --> 00:59:37,574 ALL RIGHT, NOW I THINK FOR 1533 00:59:37,574 --> 00:59:38,708 NEURAL PRIMITIVES IT'S IMPORTANT 1534 00:59:38,708 --> 00:59:40,210 TO LOOK FOR COMPUTATIONS THAT 1535 00:59:40,210 --> 00:59:42,979 ARE RELEVANT FOR A WIDE VARIETY 1536 00:59:42,979 --> 00:59:46,883 OF BIOLOGICAL SYSTEMS. SO HERE'S 1537 00:59:46,883 --> 00:59:48,285 AN EXAMPLE FROM DROSOPHILA. I 1538 00:59:48,285 --> 00:59:49,853 WILL NOT GO THROUGH ALL THE 1539 00:59:49,853 --> 00:59:51,121 EXPERIMENTAL DETAILS BUT I WILL 1540 00:59:51,121 --> 00:59:52,856 NOTE IT'S A LINEAR RESPONSE 1541 00:59:52,856 --> 00:59:54,191 CURVE. AND THE DIFFERENCE 1542 00:59:54,191 --> 00:59:57,060 BETWEEN -- AND IT'S A LINEAR 1543 00:59:57,060 --> 00:59:59,162 RESPONSE CURVE NOT A GAUSSIAN TUNING 1544 00:59:59,162 --> 01:00:00,330 CURVE, IT'S A NAVIGATION TASK 1545 01:00:00,330 --> 01:00:02,132 NOT A REACHING TASK. AND THE 1546 01:00:02,132 --> 01:00:04,401 DIFFERENCE BETWEEN THE TWO 1547 01:00:04,401 --> 01:00:07,170 COLORED CURVES IS HEADING 1548 01:00:07,170 --> 01:00:08,472 DIRECTION RATHER THAN HEAD 1549 01:00:08,472 --> 01:00:09,773 POSITION. BUT AGAIN WHAT WE SEE 1550 01:00:09,773 --> 01:00:13,176 IS THE EFFECTIVE HEADING IS A 1551 01:00:13,176 --> 01:00:16,279 MULTIPLICATIVE SCALING OF THE 1552 01:00:16,279 --> 01:00:20,684 RESPONSE CURVE. SO IN MY OWN 1553 01:00:20,684 --> 01:00:23,820 RESEARCH WE HAVE FOCUSED ON WHAT 1554 01:00:23,820 --> 01:00:28,425 DRAGONFLIES NEED TO DO TO CATCH 1555 01:00:28,425 --> 01:00:32,629 PREY. WORK DONE BY CLAIRE PLUNKETT 1556 01:00:32,629 --> 01:00:36,800 FUNDED BY THE DOE THROUGH NSF CRCNS 1557 01:00:36,800 --> 01:00:38,902 POINTED OUT IN THE FUNDERING PANEL 1558 01:00:38,902 --> 01:00:39,536 YESTERDAY. GRATEFUL FOR OPPORTUNITIES 1559 01:00:39,536 --> 01:00:40,937 THAT THE CRCNS PROGRAM SUPPORTS. 1560 01:00:40,937 --> 01:00:44,341 IN THIS MODEL WE HAVE A COMBINATION 1561 01:00:44,341 --> 01:00:46,443 OF VISUAL INPUT AND HEAD POSITION AND 1562 01:00:46,443 --> 01:00:50,313 MULTIPLICATIVE INPUTS TO FORM BASIS 1563 01:00:50,313 --> 01:00:51,882 FUNCTIONS AND THE TURN DIRECTION 1564 01:00:51,882 --> 01:00:53,250 NEEDED TO INTERCEPT THE PREY IS 1565 01:00:53,250 --> 01:00:55,051 CALCULATED IN THIS OUTPUT. NOW WHAT 1566 01:00:55,051 --> 01:00:56,786 WAS INTERESTING WHEN I STARTED 1567 01:00:56,786 --> 01:00:59,222 WORKING WITH THE NEUROMORPHIC 1568 01:00:59,222 --> 01:01:05,495 ENGINEERS IS IMPLEMENTING -- YOU 1569 01:01:05,495 --> 01:01:07,631 THINK THE HEAVY LIFTING IS DONE 1570 01:01:07,631 --> 01:01:09,666 BY THE SYNAPTIC CONFIGURATION 1571 01:01:09,666 --> 01:01:11,201 BETWEEN THE BASIS FUNCTIONS AND 1572 01:01:11,201 --> 01:01:13,003 THE OUTPUT LAYER THAT'S 1573 01:01:13,003 --> 01:01:14,504 SOMETHING WE KIND OF KNOW HOW TO 1574 01:01:14,504 --> 01:01:15,772 DO. WHAT HAS BEEN ACTUALLY MORE 1575 01:01:15,772 --> 01:01:16,907 CHALLENGING IS THE 1576 01:01:16,907 --> 01:01:18,275 MULTIPLICATIONS HAPPENING IN THIS 1577 01:01:18,275 --> 01:01:22,779 TO FORM THESE BASIS FUNCTIONS. 1578 01:01:22,779 --> 01:01:25,415 IT'S NOT KNOWN HOW THIS IS DONE BY 1579 01:01:25,415 --> 01:01:27,851 BIOLOGICAL NEURONS. AND 1580 01:01:27,851 --> 01:01:29,319 IMPLEMENTING THIS IN THE 1581 01:01:29,319 --> 01:01:30,220 NEUROMORPHIC SYSTEM HAS ALSO 1582 01:01:30,220 --> 01:01:31,321 BEEN AN INTERESTING PROBLEM FOR 1583 01:01:31,321 --> 01:01:33,390 US. I WOULD LIKE TO TALK ABOUT 1584 01:01:33,390 --> 01:01:34,791 -- WE'VE BEEN EXPLORING DIFFERENT 1585 01:01:34,791 --> 01:01:36,393 APPROACHES. 1586 01:01:36,393 --> 01:01:37,961 HERE'S ONE EXAMPLE, THIS WAS 1587 01:01:37,961 --> 01:01:41,131 DEVELOPED BY SUMA CARDWELL AT 1588 01:01:41,131 --> 01:01:43,833 SANDIA NATIONAL LABS. I HOPE 1589 01:01:43,833 --> 01:01:45,368 SHE'S OUT THERE LISTENING. SHE'S 1590 01:01:45,368 --> 01:01:47,671 IMPLEMENTING SHUNTING INHIBITION 1591 01:01:47,671 --> 01:01:48,972 IN AN ANALOG NEUROMORPHIC DENDRITE. 1592 01:01:48,972 --> 01:01:50,674 YOU CAN THINK OF THIS AS THE OUTPUT 1593 01:01:50,674 --> 01:01:52,776 OF ONE OF THESE -- OH, ONE OF 1594 01:01:52,776 --> 01:01:55,912 THESE BASIS FUNCTION NEURONS AND 1595 01:01:55,912 --> 01:01:58,715 THE OUTPUT VOLTAGE WOULD BE THE 1596 01:01:58,715 --> 01:02:00,083 RESPONSE OF ONE OF THESE NEURONS 1597 01:02:00,083 --> 01:02:01,885 THE PEAK INPUT WOULD BE THE 1598 01:02:01,885 --> 01:02:04,120 VISUAL INPUT AND THE HEAD 1599 01:02:04,120 --> 01:02:06,256 POSITION WOULD BE CONTROLLING 1600 01:02:06,256 --> 01:02:07,757 THE OVERALL CONDUCTANCE OF THE 1601 01:02:07,757 --> 01:02:09,593 NEURON. WE'RE EXPLORING OTHER 1602 01:02:09,593 --> 01:02:10,427 APPROACHES SO WE CAN UNDERSTAND 1603 01:02:10,427 --> 01:02:11,728 THE ADVANTAGES AND DISADVANTAGES 1604 01:02:11,728 --> 01:02:12,729 OF THESE DIFFERENT APPROACHES 1605 01:02:12,729 --> 01:02:13,897 BUT THIS IS THE IDEA OF WHAT 1606 01:02:13,897 --> 01:02:15,599 WE'RE TRYING TO DO. LOOK TO THE 1607 01:02:15,599 --> 01:02:17,334 NEURAL SYSTEM, UNDERSTAND WHAT 1608 01:02:17,334 --> 01:02:18,101 COMPUTATIONS ARE IMPORTANT AND 1609 01:02:18,101 --> 01:02:19,769 THEN SEE IF WE CAN GET CREATIVE 1610 01:02:19,769 --> 01:02:22,238 ABOUT HOW TO IMPLEMENT THIS IN 1611 01:02:22,238 --> 01:02:25,275 NEUROMORPHIC HARDWARE. NOW I 1612 01:02:25,275 --> 01:02:27,644 STARTED BY SAYING I WANTED TO 1613 01:02:27,644 --> 01:02:30,914 TALK ABOUT THIS VIRTUOUS CIRCLE 1614 01:02:30,914 --> 01:02:32,282 BETWEEN NEUROSCIENCE AND NEUROMORPHIC COMPUTATION WHEN WE 1615 01:02:32,282 --> 01:02:33,283 TALK ABOUT THIS BACK AND FORTH 1616 01:02:33,283 --> 01:02:34,751 BETWEEN THE TWO FIELDS THERE'S 1617 01:02:34,751 --> 01:02:36,386 ALWAYS AN OPEN CHALLENGE OR AN 1618 01:02:36,386 --> 01:02:38,054 OPEN RESEARCH QUESTION OF WHERE 1619 01:02:38,054 --> 01:02:40,256 DO YOU LOOK IN BIOLOGY FOR 1620 01:02:40,256 --> 01:02:41,157 INSPIRATION AND I THINK WE 1621 01:02:41,157 --> 01:02:41,925 SHOULD START HAVING MORE 1622 01:02:41,925 --> 01:02:43,994 COMMUNICATION BETWEEN THE FIELDS 1623 01:02:43,994 --> 01:02:45,362 BECAUSE WE NEED TO BE 1624 01:02:45,362 --> 01:02:47,430 INTENTIONAL ABOUT WHY WE CHOOSE 1625 01:02:47,430 --> 01:02:49,265 THAT SPECIFIC FUNCTION, SORRY 1626 01:02:49,265 --> 01:02:51,267 THAT SPECIFIC BIOLOGICAL SYSTEM 1627 01:02:51,267 --> 01:02:53,837 FOR INSPIRATION. AND IN NEUROAI I 1628 01:02:53,837 --> 01:02:55,639 SEE AN OPPORTUNITY NOT JUST FOR 1629 01:02:55,639 --> 01:02:57,040 DEVELOPING ENERGY EFFICIENT 1630 01:02:57,040 --> 01:02:57,440 IMPLEMENTATIONS. 1631 01:02:57,440 --> 01:02:59,242 WHEN I STARTED PREPARING THIS 1632 01:02:59,242 --> 01:03:01,044 TALK I WAS EXPECTING TO SAY IF 1633 01:03:01,044 --> 01:03:03,013 WE HAVE A NEURAL-INSPIRED A.I. 1634 01:03:03,013 --> 01:03:06,116 MODEL IT MAKES SENSE TO HAVE 1635 01:03:06,116 --> 01:03:07,283 NEURAL-INSPIRED HARDWARE TO RUN 1636 01:03:07,283 --> 01:03:08,918 IT ON. BUT I THINK THERE'S AN 1637 01:03:08,918 --> 01:03:09,886 OPPORTUNITY HERE AS WE START 1638 01:03:09,886 --> 01:03:12,589 TALKING ABOUT HOW WE COMPARE 1639 01:03:12,589 --> 01:03:14,391 DIFFERENT SOLUTION IN BIOLOGICAL 1640 01:03:14,391 --> 01:03:15,258 SYSTEMS WITH DIFFERENT 1641 01:03:15,258 --> 01:03:18,294 APPROACHES THAT WE CAN USE IN 1642 01:03:18,294 --> 01:03:19,663 NEUROMORPHIC SYSTEMS, THERE'S AN 1643 01:03:19,663 --> 01:03:22,065 OPPORTUNITY TO CREATE A 1644 01:03:22,065 --> 01:03:23,566 FRAMEWORK FOR FACILITATING 1645 01:03:23,566 --> 01:03:25,835 COMMUNICATION BETWEEN THE 1646 01:03:25,835 --> 01:03:26,436 DIFFERENT FIELDS AND 1647 01:03:26,436 --> 01:03:28,071 SPECIFICALLY IF WE'RE TALKING 1648 01:03:28,071 --> 01:03:30,407 ABOUT HOW TO ABSTRACT BIOLOGY 1649 01:03:30,407 --> 01:03:31,441 FOR TRANSLATION INTO MODELS IN 1650 01:03:31,441 --> 01:03:34,511 OTHER COMMUNITIES WHETHER 1651 01:03:34,511 --> 01:03:36,546 IT BE NEUROMORPHIC OR 1652 01:03:36,546 --> 01:03:37,180 ARTIFICIAL INTELLIGENCE AND I 1653 01:03:37,180 --> 01:03:38,448 WILL LEAVE YOU WITH THAT. 1654 01:03:38,448 --> 01:03:39,249 THAT'S WHERE I THINK WE SHOULD 1655 01:03:39,249 --> 01:03:49,726 BE LOOKING FOR THE FUTURE. 1656 01:03:51,227 --> 01:03:56,499 >> PLEASE WELCOME SUEYEON CHUNG 1657 01:03:56,499 --> 01:04:05,909 FROM NYU AND THE FLAT IRON INSTITUTE. 1658 01:04:05,909 --> 01:04:08,044 >> OKAY. THANK YOU, EVERYONE. TODAY 1659 01:04:08,044 --> 01:04:10,947 I WILL DISCUSS HOW UNDERSTANDING 1660 01:04:10,947 --> 01:04:13,416 WHAT'S KNOWN AS NEURAL MANIFOLDS 1661 01:04:13,416 --> 01:04:15,719 CAN BRIDGE THE GAP BETWEEN 1662 01:04:15,719 --> 01:04:18,521 NEURAL SCIENCE AND A.I. LEADING TO 1663 01:04:18,521 --> 01:04:19,556 MORE EFFICIENT AND BIOLOGICALLY 1664 01:04:19,556 --> 01:04:24,627 ALIGNED A.I. SYSTEMS. WE ARE AT A 1665 01:04:24,627 --> 01:04:25,495 EXCITING MOMENT WHERE NEUROSCIENCE 1666 01:04:25,495 --> 01:04:27,664 AND A.I. ARE CONVERGING. 1667 01:04:27,664 --> 01:04:31,634 DEEP NETWORKS CAN PREDICT NEURAL 1668 01:04:31,634 --> 01:04:33,002 ACTIVITIES IN PRIMATE BRAINS 1669 01:04:33,002 --> 01:04:36,973 WITH REMARKABLE ACCURACY AND THIS 1670 01:04:36,973 --> 01:04:37,941 CONVERGENCE PRESENTS AN 1671 01:04:37,941 --> 01:04:39,609 OPPORTUNITY TO IDENTIFY A NEURAL 1672 01:04:39,609 --> 01:04:41,745 FRAMEWORK. FIRST I WOULD LIKE 1673 01:04:41,745 --> 01:04:44,080 TO PAUSE AND ASK WHAT ARE THE 1674 01:04:44,080 --> 01:04:45,515 GOALS OF THEORY AT THIS 1675 01:04:45,515 --> 01:04:47,317 INTERSECTION? WE ALL KNOW THAT 1676 01:04:47,317 --> 01:04:48,752 DEEP NETWORKS WORK VERY WELL 1677 01:04:48,752 --> 01:04:52,489 BOTH AS A MODEL OF THE BRAIN BUT 1678 01:04:52,489 --> 01:04:57,393 ALSO AS AN ENGINEERING SYSTEM. 1679 01:04:57,393 --> 01:04:59,596 ANOTHER WAY TO SEE THIS IS THEY 1680 01:04:59,596 --> 01:05:00,964 ARE ACTUALLY FULLY TRANSPARENT 1681 01:05:00,964 --> 01:05:03,032 AND WE HAVE ACCESS TO ALL OF 1682 01:05:03,032 --> 01:05:04,434 THEIR INTERNAL REPRESENTATIONS, 1683 01:05:04,434 --> 01:05:05,935 THE ARCHITECTURE THAT WAS USED 1684 01:05:05,935 --> 01:05:07,871 AND ALSO THE LEARNING RULES THAT 1685 01:05:07,871 --> 01:05:10,106 HAVE BEEN USED. SO MY VIEW FOR 1686 01:05:10,106 --> 01:05:12,408 THE ROLE OF THEORY AT THIS 1687 01:05:12,408 --> 01:05:14,410 INTERSECTION HAS TO DO WITH THE 1688 01:05:14,410 --> 01:05:16,613 IDEA OF THE EMERGENCE. AND ALSO 1689 01:05:16,613 --> 01:05:18,448 THERE'S MULTIPLE LEVELS OF 1690 01:05:18,448 --> 01:05:19,682 ABSTRACTION THAT WAS PRESENTED 1691 01:05:19,682 --> 01:05:23,153 EARLY -- PRESENTED EARLIER IN 1692 01:05:23,153 --> 01:05:23,419 THE TALK. 1693 01:05:23,419 --> 01:05:26,456 THERE'S A LOT OF COMPLEX INTERACTIONS 1694 01:05:26,456 --> 01:05:27,390 IN DEEP NETWORKS THAT WE DON'T 1695 01:05:27,390 --> 01:05:28,458 UNDERSTAND. THE CONNECTION 1696 01:05:28,458 --> 01:05:29,959 BETWEEN FEATURES AND UNITS AND 1697 01:05:29,959 --> 01:05:35,398 THE STRUCTURE OF THEIR INTERNAL 1698 01:05:35,398 --> 01:05:36,833 REPRESENTATIONS IN EACH LAYER 1699 01:05:36,833 --> 01:05:41,237 AND HOW THAT CONTRIBUTES TO THE 1700 01:05:41,237 --> 01:05:41,871 PERFORMANCE AFTER TRAINING. THIS IS 1701 01:05:41,871 --> 01:05:44,440 SIMILAR TO THE CHALLENGES THAT 1702 01:05:44,440 --> 01:05:46,209 WE FACE IN MODERN NEUROSCIENCE 1703 01:05:46,209 --> 01:05:47,977 IN INCREASING ACCESS THAT WE 1704 01:05:47,977 --> 01:05:49,612 HAVE WITH THE LARGE NUMBER OF 1705 01:05:49,612 --> 01:05:50,647 NEURONS AT THE SAME TIME HOW 1706 01:05:50,647 --> 01:05:52,148 FIRING RATES AND INDIVIDUAL 1707 01:05:52,148 --> 01:05:53,383 NEURONS AND THEIR INTERACTIONS 1708 01:05:53,383 --> 01:05:56,219 GIVE RISE TO REPRESENTATIONS AND 1709 01:05:56,219 --> 01:05:57,720 HOW THESE REPRESENTATIONS 1710 01:05:57,720 --> 01:05:59,122 IMPLEMENT TASK FUNCTIONS LIKE 1711 01:05:59,122 --> 01:06:00,824 SENSORY AND MOTOR FUNCTIONS. 1712 01:06:00,824 --> 01:06:02,926 THIS IS A SHARED CHALLENGE 1713 01:06:02,926 --> 01:06:04,894 BETWEEN A.I. AND NEUROSCIENCE AND 1714 01:06:04,894 --> 01:06:07,730 THE TASK FOR THEORY IS TO 1715 01:06:07,730 --> 01:06:08,965 FORMALIZE AND CONNECT THE 1716 01:06:08,965 --> 01:06:11,034 RELATIONSHIP BETWEEN THESE 1717 01:06:11,034 --> 01:06:13,269 PROPERTIES AND PHENOMENA AT 1718 01:06:13,269 --> 01:06:15,505 DIFFERENT SCALES AND PROVIDE 1719 01:06:15,505 --> 01:06:19,008 INTERPRETABILITY IN DESIGN 1720 01:06:19,008 --> 01:06:23,847 PRINCIPLES. ONE THEORETICAL 1721 01:06:23,847 --> 01:06:25,815 FRAMEWORK GAINING TRACTION IS CALLED 1722 01:06:25,815 --> 01:06:27,917 NEURAL MANIFOLDS FRAMEWORK. LET ME 1723 01:06:27,917 --> 01:06:29,319 EXPLAIN WHAT THAT IS. LET'S SAY 1724 01:06:29,319 --> 01:06:32,689 YOU'RE GIVEN AN IMAGE OF A DOG 1725 01:06:32,689 --> 01:06:35,425 AND THEN A POPULATION OF NEURONS 1726 01:06:35,425 --> 01:06:36,926 WILL RESPOND AND THAT BECOMES A 1727 01:06:36,926 --> 01:06:39,796 POINT IN THE SPACE IN THE NEURAL 1728 01:06:39,796 --> 01:06:41,297 ACTIVITIES AND THIS KIND OF 1729 01:06:41,297 --> 01:06:49,606 SPACE IS IN GENERAL HIGH 1730 01:06:49,606 --> 01:06:51,908 DIMENSIONAL. NOW IN REAL LIFE AN 1731 01:06:51,908 --> 01:06:54,210 IMAGE IS NOT REALLY STATIC. 1732 01:06:54,210 --> 01:06:56,112 LET'S SAY A DOG THAT YOU'RE 1733 01:06:56,112 --> 01:06:57,680 LOOKING AT IS RUNNING AND MOVING 1734 01:06:57,680 --> 01:06:59,816 AROUND THEN ALL OF THESE NEURONS 1735 01:06:59,816 --> 01:07:01,584 WILL RESPOND A LITTLE BIT 1736 01:07:01,584 --> 01:07:03,820 DIFFERENTLY. AND THIS PICTURE 1737 01:07:03,820 --> 01:07:05,922 OF NEURAL REPRESENTATION FOR AN 1738 01:07:05,922 --> 01:07:08,625 IMAGE OF A DOG ALSO WILL MOVE 1739 01:07:08,625 --> 01:07:09,926 AROUND IN THE NEURAL STATE SPACE 1740 01:07:09,926 --> 01:07:11,594 AND WE END UP GETTING A SET OF 1741 01:07:11,594 --> 01:07:14,430 NEURAL ACTIVITIES WITH AN 1742 01:07:14,430 --> 01:07:17,367 UNDERLYING GEOMETRY WHICH IS 1743 01:07:17,367 --> 01:07:20,803 WHAT WE CALL NEURAL MANIFOLDS 1744 01:07:20,803 --> 01:07:21,771 NOW WHEN YOU DISTINGUISH BETWEEN 1745 01:07:21,771 --> 01:07:23,673 TWO DIFFERENT OBJECTS OR 1746 01:07:23,673 --> 01:07:25,441 CONCEPTS LET'S SAY BETWEEN DOGS 1747 01:07:25,441 --> 01:07:28,511 FROM CATS THE PROBLEM OF THE 1748 01:07:28,511 --> 01:07:30,179 DISCRIMINATING BETWEEN DOGS FROM 1749 01:07:30,179 --> 01:07:33,616 CATS BECOMES A SEPARATING A DOG 1750 01:07:33,616 --> 01:07:36,686 MANIFOLD FROM A CAT MANIFOLD. WHAT 1751 01:07:36,686 --> 01:07:38,021 WE HAVE FOUND IN OUR WORK IS 1752 01:07:38,021 --> 01:07:40,523 THAT THE GEOMETRY OF THESE 1753 01:07:40,523 --> 01:07:42,158 NEURAL MANIFOLDS IS DIRECTLY 1754 01:07:42,158 --> 01:07:43,793 RELATED TO THE SYSTEMS CAPACITY 1755 01:07:43,793 --> 01:07:45,461 TO REPRESENT AND RECOGNIZE 1756 01:07:45,461 --> 01:07:47,096 OBJECTS. AND THIS WAS A 1757 01:07:47,096 --> 01:07:48,831 STARTING POINT FOR OUR NEURAL 1758 01:07:48,831 --> 01:07:51,534 MANIFOLDS FRAMEWORK AND IT GIVES 1759 01:07:51,534 --> 01:07:53,970 US A FEW KEY METRICS FOR 1760 01:07:53,970 --> 01:07:55,905 EFFICIENCY AND INTERPRETABILITY 1761 01:07:55,905 --> 01:07:57,006 FIRST, THE CAPACITY OF NEURAL 1762 01:07:57,006 --> 01:08:03,980 MANIFOLDS IS DEFINED AS 1763 01:08:03,980 --> 01:08:05,214 CATEGORIES, CAN BE REPRESENTED 1764 01:08:05,214 --> 01:08:07,083 PER GIVEN NEURAL POPULATION 1765 01:08:07,083 --> 01:08:09,452 SIZE. AND IT MEASURES THE 1766 01:08:09,452 --> 01:08:10,253 EFFICIENCY OF THE 1767 01:08:10,253 --> 01:08:11,888 REPRESENTATIONS OF THE BRAIN AND 1768 01:08:11,888 --> 01:08:18,995 A.I. SYSTEMS. THEN, THE GEOMETRICAL 1769 01:08:18,995 --> 01:08:21,397 PROPERTIES OF THESE MANIFOLDS 1770 01:08:21,397 --> 01:08:23,833 SUCH AS RADIUS, WHICH HAS TO DO 1771 01:08:23,833 --> 01:08:25,301 WITH SIZE AND ORGANIZATION OF 1772 01:08:25,301 --> 01:08:27,837 THESE MANIFOLDS CAPTURES A 1773 01:08:27,837 --> 01:08:29,272 STRUCTURE OF THEIR INTERNAL 1774 01:08:29,272 --> 01:08:29,672 REPRESENTATIONS. 1775 01:08:29,672 --> 01:08:31,774 SO IF OBJECT MANIFOLDS IN THE 1776 01:08:31,774 --> 01:08:34,243 NEURAL ACTIVITY OR 1777 01:08:34,243 --> 01:08:36,079 REPRESENTATIONS IN A.I. SYSTEM ARE 1778 01:08:36,079 --> 01:08:39,449 FLAT, SMALL, AND WELL SPREAD 1779 01:08:39,449 --> 01:08:41,284 OUT, AND NICELY STACKED OR 1780 01:08:41,284 --> 01:08:44,821 LINED, MEANING THAT THE OBJECT 1781 01:08:44,821 --> 01:08:47,123 MANIFOLD'S DIMENSIONALITY, 1782 01:08:47,123 --> 01:08:49,058 RADIUS AND LOCATION CORRELATIONS 1783 01:08:49,058 --> 01:08:51,294 ARE SMALL WHILE AXIS 1784 01:08:51,294 --> 01:08:52,061 CORRELATIONS ARE HIGH, THEN YOU 1785 01:08:52,061 --> 01:08:53,997 CAN REPRESENT MORE OF THESE 1786 01:08:53,997 --> 01:08:55,031 OBJECT MANIFOLDS WITH THE SAME 1787 01:08:55,031 --> 01:08:57,734 NUMBER OF NEURONS IN A GIVEN 1788 01:08:57,734 --> 01:08:59,802 NEURAL REPRESENTATION SPACE, 1789 01:08:59,802 --> 01:09:00,436 SUCH THAT YOU CAN DISTINGUISH 1790 01:09:00,436 --> 01:09:03,072 BETWEEN THEM. AND THAT RESULTS 1791 01:09:03,072 --> 01:09:04,941 IN HIGH CAPACITY OF OBJECT 1792 01:09:04,941 --> 01:09:06,676 MANIFOLDS SO THIS IS ACTUALLY 1793 01:09:06,676 --> 01:09:08,678 JUST LIKE PACKING THINGS IN THE 1794 01:09:08,678 --> 01:09:10,046 PHYSICAL SPACE EXCEPT THAT WE'RE 1795 01:09:10,046 --> 01:09:11,647 TALKING ABOUT NEURAL POPULATION 1796 01:09:11,647 --> 01:09:13,349 RESPONSES, CORRESPONDING TO 1797 01:09:13,349 --> 01:09:16,686 DIFFERENT OBJECT OR CONCEPTS. 1798 01:09:16,686 --> 01:09:19,055 AND THE IMPORTANTLY THIS ALLOWS 1799 01:09:19,055 --> 01:09:21,524 US TO DISTINGUISH BETWEEN 1800 01:09:21,524 --> 01:09:24,127 EFFICIENT AND UNTANGLED 1801 01:09:24,127 --> 01:09:24,627 REPRESENTATIONS VERSUS 1802 01:09:24,627 --> 01:09:25,895 INEFFICIENT AND ENTANGLED 1803 01:09:25,895 --> 01:09:27,130 REPRESENTATIONS AS MEASURED BY 1804 01:09:27,130 --> 01:09:30,466 THE MANIFOLD CAPACITY. WHILE 1805 01:09:30,466 --> 01:09:32,802 THE GEOMETRICAL MEASURES PROVIDE 1806 01:09:32,802 --> 01:09:34,437 INTERPRETABILITY SUCH AS THE 1807 01:09:34,437 --> 01:09:35,571 DATA CLOUDS THAT YOU'RE LOOKING 1808 01:09:35,571 --> 01:09:37,573 AT FROM THE RECORDED DATA IN THE 1809 01:09:37,573 --> 01:09:38,975 THE BRAIN OR THE -- FROM THE 1810 01:09:38,975 --> 01:09:41,344 FEATURES OR REPRESENTATIONS FROM 1811 01:09:41,344 --> 01:09:44,447 A.I. SYSTEMS, THEY ARE FLAT 1812 01:09:44,447 --> 01:09:45,715 OR COMPRESSED AND PROVIDES A 1813 01:09:45,715 --> 01:09:48,117 LANGUAGE AND METRIC TO TALK 1814 01:09:48,117 --> 01:09:50,620 ABOUT THE STRUCTURE IN THE 1815 01:09:50,620 --> 01:09:51,587 REPRESENTATION. SO WHEN WE 1816 01:09:51,587 --> 01:09:53,990 APPLY THIS FRAMEWORK TO VISUAL 1817 01:09:53,990 --> 01:09:56,225 DEEP NETWORKS, WE ACTUALLY END UP 1818 01:09:56,225 --> 01:09:57,760 SEEING A VERY CLEAR PATTERN. SO 1819 01:09:57,760 --> 01:10:00,363 AS WE MOVE THROUGH THE NETWORK 1820 01:10:00,363 --> 01:10:01,964 LAYERS, THE REPRESENTATIONS 1821 01:10:01,964 --> 01:10:02,999 BECOME MORE EFFICIENT. AS YOU 1822 01:10:02,999 --> 01:10:07,036 CAN SEE ON THE TOP LEFT, THE 1823 01:10:07,036 --> 01:10:08,304 MANIFOLD CAPACITY INCREASES 1824 01:10:08,304 --> 01:10:10,506 WHILE THE DIMENSIONALITY ON THE 1825 01:10:10,506 --> 01:10:13,376 TOP RIGHT AND RADIUS ON THE 1826 01:10:13,376 --> 01:10:14,710 BOTTOM LEFT AND THE CORRELATIONS 1827 01:10:14,710 --> 01:10:16,879 ON THE BOTTOM RIGHT, THEY END UP 1828 01:10:16,879 --> 01:10:18,347 DECREASING SO IT TELLS YOU THAT 1829 01:10:18,347 --> 01:10:20,817 THE REPRESENTATIONS IN A.I. 1830 01:10:20,817 --> 01:10:22,285 SYSTEMS ARE GETTING MORE 1831 01:10:22,285 --> 01:10:22,852 EFFICIENT AS A RESULT OF 1832 01:10:22,852 --> 01:10:25,288 CHANGING STRUCTURES IN THE 1833 01:10:25,288 --> 01:10:28,291 DISTRIBUTED REPRESENTATIONS. 1834 01:10:28,291 --> 01:10:31,627 AND THIS ALSO MATCHES WHAT WE 1835 01:10:31,627 --> 01:10:33,763 OBSERVE IN THE BIOLOGICAL VISUAL 1836 01:10:33,763 --> 01:10:35,765 SYSTEMS IN MONKEYS. SO NOW, 1837 01:10:35,765 --> 01:10:37,600 WE'RE APPLYING THIS FRAMEWORK ACROSS 1838 01:10:37,600 --> 01:10:39,368 MULTIPLE DOMAINS LIKE VISION, 1839 01:10:39,368 --> 01:10:42,538 LANGUAGE AND SPEECH, A.I. SYSTEMS 1840 01:10:42,538 --> 01:10:45,708 FOR INTERPRETABILITY AS WELL 1841 01:10:45,708 --> 01:10:47,009 AS INCREASING LIST OF EXPERIMENTAL 1842 01:10:47,009 --> 01:10:48,778 NEURAL DATASETS GOING FROM 1843 01:10:48,778 --> 01:10:51,514 MONKEYS TO HUMANS TO MICE AND 1844 01:10:51,514 --> 01:10:53,816 ALSO FROM VISION TO ADDITION TO 1845 01:10:53,816 --> 01:10:55,485 MEMORY AREAS LIKE HIPPOCAMPUS. 1846 01:10:55,485 --> 01:10:56,853 AND THESE APPLICATIONS ARE 1847 01:10:56,853 --> 01:11:01,090 REVEALING COMMON PRINCIPLES OF 1848 01:11:01,090 --> 01:11:06,562 HOW BOTH ARTIFICIAL SYSTEMS AND 1849 01:11:06,562 --> 01:11:10,166 BIOLOGICAL SYSTEMS DO THIS IN AN 1850 01:11:10,166 --> 01:11:11,100 EFFICIENT MANNER. AND THIS IS THE 1851 01:11:11,100 --> 01:11:14,470 BRAIN AND A.I. SYSTEMS USING 1852 01:11:14,470 --> 01:11:16,839 MANIFOLDS THEORY BUT IT ALSO 1853 01:11:16,839 --> 01:11:19,308 TURNS OUT YOU CAN USE THIS TO 1854 01:11:19,308 --> 01:11:22,545 BUILD MORE EFFICIENT A.I. SYSTEMS. 1855 01:11:22,545 --> 01:11:25,915 THE CLASSIC EFFICIENT CODING THEORY FROM HORACE BARLOW FOCUSED ON 1856 01:11:25,915 --> 01:11:26,115 HOW 1857 01:11:26,115 --> 01:11:36,793 THE BRAIN COMPRESSES RAW SENSORY SIGNALS. 1858 01:11:36,793 --> 01:11:38,094 MODERN FRAMEWORK EXTENDS THIS TO EFFICIENT REPRESENTATIONS FOR 1859 01:11:38,094 --> 01:11:39,562 COMPLEX TASKS FOCUSED ON DISCARDING IRRELEVANT INFORMATION. AND THIS 1860 01:11:39,562 --> 01:11:39,996 SHIFT FROM STIMULUS 1861 01:11:39,996 --> 01:11:41,164 EFFICIENCY TO TASK EFFICIENCY 1862 01:11:41,164 --> 01:11:43,332 NOT ONLY HELPS EXPLAIN NEURAL 1863 01:11:43,332 --> 01:11:45,768 RESPONSES IN HIGHER SENSORY 1864 01:11:45,768 --> 01:11:48,304 AREAS THAT ARE CLOSER TO COMPLEX 1865 01:11:48,304 --> 01:11:49,939 TASKS BUT ALSO IT ALLOWS FOR 1866 01:11:49,939 --> 01:11:52,175 TRAINING A.I. SYSTEMS WITH 1867 01:11:52,175 --> 01:11:54,343 EFFICIENCY PRINCIPLES AS AN 1868 01:11:54,343 --> 01:11:56,312 OBJECTIVE FUNCTION. SO BY 1869 01:11:56,312 --> 01:11:58,147 OPTIMIZING NEURAL NETWORKS TO 1870 01:11:58,147 --> 01:11:59,816 MAXIMIZE THE MANIFOLD CAPACITY 1871 01:11:59,816 --> 01:12:02,485 WHICH IS THEIR ABILITY TO 1872 01:12:02,485 --> 01:12:03,886 EFFICIENTLY REPRESENT OBJECT 1873 01:12:03,886 --> 01:12:05,288 VARIATIONS, WE ACHIEVE BOTH 1874 01:12:05,288 --> 01:12:07,156 STATE-OF-THE-ART CLASSIFICATION 1875 01:12:07,156 --> 01:12:10,993 PERFORMANCE AND BETTER PREDICTION OF 1876 01:12:10,993 --> 01:12:15,698 MACAQUE VISUAL CORTEX RESPONSES. 1877 01:12:15,698 --> 01:12:16,365 THIS SUGGESTS THAT MAXIMIZING 1878 01:12:16,365 --> 01:12:18,834 REPRESENTATIONAL EFFICIENCY IS A KEY 1879 01:12:18,834 --> 01:12:20,903 PRINCIPLE SHARED BY BOTH BIOLOGICAL 1880 01:12:20,903 --> 01:12:22,839 AND ARTIFICIAL VISUAL SYSTEMS. SO TO 1881 01:12:22,839 --> 01:12:26,275 SUMMARZE, THE THEORY OF NEURAL 1882 01:12:26,275 --> 01:12:31,113 MANIFOLDS OPENS THE BLACK BOX OF 1883 01:12:31,113 --> 01:12:35,451 BRAIN AND A.I. WITH INTERPRETABLE METRICS. 1884 01:12:35,451 --> 01:12:37,653 AND CAN ALSO EXTEND THIS TO COMPLEX 1885 01:12:37,653 --> 01:12:39,388 COGNITIVE TASKS, AND GUIDE DESIGN 1886 01:12:39,388 --> 01:12:42,391 OF BRAIN-ALIGNED HIGH PERFORMING 1887 01:12:42,391 --> 01:12:44,327 A.I. SYSTEMS. LOOKING AHEAD, 1888 01:12:44,327 --> 01:12:46,562 WE -- IN MY VIEW, WE NEED TWO 1889 01:12:46,562 --> 01:12:48,531 KEY ADVANCES. FIRST, WE SHOULD 1890 01:12:48,531 --> 01:12:50,933 BUILD ON OUR RECENT ADVANCES IN 1891 01:12:50,933 --> 01:12:53,202 NEURAL MANIFOLDS FRAMEWORK TO 1892 01:12:53,202 --> 01:12:55,137 BOTH THE THEORIES AND MODELS FOR 1893 01:12:55,137 --> 01:12:57,673 BROADER RANGE OF TASKS INCLUDING 1894 01:12:57,673 --> 01:13:00,142 CONTINUAL LEARNING AND EMBODIED 1895 01:13:00,142 --> 01:13:02,111 INTELLIGENCE AND SECOND I THINK 1896 01:13:02,111 --> 01:13:04,480 WE NEED TO BETTER BENCHMARK 1897 01:13:04,480 --> 01:13:06,883 DATASETS FOR NATURALISTIC 1898 01:13:06,883 --> 01:13:07,383 BEHAVIORS AND PHYSICAL 1899 01:13:07,383 --> 01:13:09,185 INTELLIGENCE. WHILE A.I. SYSTEMS 1900 01:13:09,185 --> 01:13:11,354 HAVE BENEFITED FROM DATASETS 1901 01:13:11,354 --> 01:13:13,055 LIKE IMAGENET AND COMPUTER VISION 1902 01:13:13,055 --> 01:13:16,225 I THINK NEUROSCIENCE CAN OFFER 1903 01:13:16,225 --> 01:13:17,260 UNIQUE ADVANTAGES IN THESE DOMAINS, 1904 01:13:17,260 --> 01:13:20,897 IN PHYSICAL INTELLIGENCE, THROUGH 1905 01:13:20,897 --> 01:13:22,465 ANIMAL EXPERIMENTS AND THESE ADVANCES 1906 01:13:22,465 --> 01:13:24,667 SUPPORTED BY SOPHISTICATED THEORETICAL 1907 01:13:24,667 --> 01:13:26,836 ANIMAL WILL PUSH OUR UNDERSTANDING 1908 01:13:26,836 --> 01:13:28,437 FORWARD. AND I HOPE I WAS ABLE TO 1909 01:13:28,437 --> 01:13:30,606 CONVINCE YOU THAT BY USING THE 1910 01:13:30,606 --> 01:13:33,542 THEORY OF NEURAL MANIFOLDS AS AN 1911 01:13:33,542 --> 01:13:35,144 INTERPRETABILITY TOOL WE CAN 1912 01:13:35,144 --> 01:13:37,313 OPEN THE BLACK BOX OF BOTH BRAIN 1913 01:13:37,313 --> 01:13:39,615 AND A.I. SYSTEMS AND BUILD 1914 01:13:39,615 --> 01:13:42,919 EFFICIENT BRAIN-ALIGNED AND SAFE 1915 01:13:42,919 --> 01:13:45,254 AND HIGHLY PERFORMANT A.I. 1916 01:13:45,254 --> 01:13:46,689 SYSTEMS. AND OF COURSE THIS 1917 01:13:46,689 --> 01:13:48,190 RESEARCH WAS POSSIBLE THANKS TO 1918 01:13:48,190 --> 01:13:49,558 THE SUPPORT OF BRAIN INITIATIVE, 1919 01:13:49,558 --> 01:13:51,794 SIMON'S FOUNDATION AND OTHER 1920 01:13:51,794 --> 01:13:53,796 FUNDERS AND THE COLLABORATION OF 1921 01:13:53,796 --> 01:13:54,997 EXCELLENT STUDENTS AND POSTDOCS 1922 01:13:54,997 --> 01:14:05,474 AND COLLEAGUES, THANK YOU. 1923 01:14:06,509 --> 01:14:10,479 >> PLEASE WELCOME OUR NEXT 1924 01:14:10,479 --> 01:14:14,250 SPEAKER DHIREESHA KUDITHIPUDI FROM UT SAN ANTONIO. 1925 01:14:14,250 --> 01:14:17,753 >> GOOD MORNING, EVERYONE. IT'S 1926 01:14:17,753 --> 01:14:20,156 BEEN AN INCREDIBLE TIME TO BE A 1927 01:14:20,156 --> 01:14:21,390 PART OF THIS NEUROMORPHIC 1928 01:14:21,390 --> 01:14:22,458 COMPUTING COMMUNITY BECAUSE WE 1929 01:14:22,458 --> 01:14:23,526 HAVE THIS OPPORTUNITY TO CHANGE 1930 01:14:23,526 --> 01:14:25,895 THE WAY WE ARE THINKING OF HOW 1931 01:14:25,895 --> 01:14:29,131 WE ARE DOING SYSTEM OR HARDWARE 1932 01:14:29,131 --> 01:14:31,734 DESIGN. AND IN THAT FRONT, OUR 1933 01:14:31,734 --> 01:14:33,035 OVERARCHING GOAL FROM OUR LAB IS 1934 01:14:33,035 --> 01:14:34,637 TO CREATE THE NEXT GENERATION OF 1935 01:14:34,637 --> 01:14:38,374 A.I. SYSTEMS THAT ARE CAPABLE OF 1936 01:14:38,374 --> 01:14:40,076 PERFORMING NATURAL TASKS BY WITH 1937 01:14:40,076 --> 01:14:43,212 REALLY HIGH ENERGY EFFICIENCY. 1938 01:14:43,212 --> 01:14:48,250 AND TO ALIGN THIS GOAL, OR TO 1939 01:14:48,250 --> 01:14:49,885 KIND OF HAVE A GRAND CHALLENGE 1940 01:14:49,885 --> 01:14:52,488 THAT FRAMES THOSE AND MAKES US 1941 01:14:52,488 --> 01:14:54,056 THINK HARD ABOUT THIS PROBLEM OF 1942 01:14:54,056 --> 01:14:55,391 BUILDING THE NEXT GENERATION 1943 01:14:55,391 --> 01:14:57,526 SYSTEMS WE'LL LOOK AT THE 1944 01:14:57,526 --> 01:14:58,861 PROBLEM OF CONTINUAL LEARNING 1945 01:14:58,861 --> 01:15:04,266 AND CONTINUAL LEARNING OR OFTEN 1946 01:15:04,266 --> 01:15:05,267 REFERRED TO AS LIFELONG LEARNING 1947 01:15:05,267 --> 01:15:07,103 IS A CHALLENGE IN WHICH AN AGENT 1948 01:15:07,103 --> 01:15:10,272 MUST LEARN FROM SEQUENTIAL OR 1949 01:15:10,272 --> 01:15:13,075 NONSTATIONARY DATA FROM 1950 01:15:13,075 --> 01:15:13,776 CHANGING DISTRIBUTIONS. 1951 01:15:13,776 --> 01:15:16,278 IF YOU LOOK AT THIS CARTOON 1952 01:15:16,278 --> 01:15:17,480 HERE, YOU HAVE THE MODEL WHICH 1953 01:15:17,480 --> 01:15:20,383 IS LEARNING A SEQUENCE OF TASKS. 1954 01:15:20,383 --> 01:15:23,052 TASK A, TASK B AND THEN IF 1955 01:15:23,052 --> 01:15:24,854 IT'S LEARNED-- IT COULD BE LEARNING 1956 01:15:24,854 --> 01:15:27,590 TASK C OR IF IT -- IT COULD HAVE 1957 01:15:27,590 --> 01:15:30,626 TASK A REINTRODUCED BUT QUITE 1958 01:15:30,626 --> 01:15:34,964 OFTEN CURRENT MODELS DO NOT DO 1959 01:15:34,964 --> 01:15:36,799 WELL WHEN THEY -- WHEN SUCH TASK 1960 01:15:36,799 --> 01:15:38,834 IS REINTRODUCED AND THERE IS 1961 01:15:38,834 --> 01:15:42,805 THIS PROBLEM REFERRED TO AS 1962 01:15:42,805 --> 01:15:44,073 CATASTROPHICIC FORGOTING AND 1963 01:15:44,073 --> 01:15:45,708 THIS IS NOT THE ONLY CHALLENGE 1964 01:15:45,708 --> 01:15:47,910 IN TERMS OF CONTINUAL LEARNING 1965 01:15:47,910 --> 01:15:51,113 THIS IS ONE OF THE MOST STUDIED 1966 01:15:51,113 --> 01:15:51,514 PROBLEMS. 1967 01:15:51,514 --> 01:15:53,716 SO HOW DO YOU ENSURE THAT THE 1968 01:15:53,716 --> 01:15:55,151 MODEL IS CONTINUALLY LEARNING 1969 01:15:55,151 --> 01:15:57,686 THROUGH ITS LIFETIME? SO IF YOU 1970 01:15:57,686 --> 01:15:58,921 LOOK AT THIS CARTOON AGAIN, 1971 01:15:58,921 --> 01:16:00,823 HERE, IF YOU LOOK AT THE 1972 01:16:00,823 --> 01:16:06,629 CONVENTIONAL A.I. MODEL, WHEN A 1973 01:16:06,629 --> 01:16:08,030 NEW TASK IS INTRODUCED TO THE 1974 01:16:08,030 --> 01:16:11,400 MODEL, IT IS NOT ABLE TO ADOPT 1975 01:16:11,400 --> 01:16:14,837 THAT TASK OR WHEN AN OLD TASK IS 1976 01:16:14,837 --> 01:16:17,673 REINTRODUCED IT IS 1977 01:16:17,673 --> 01:16:18,274 CATASTROPHICALLY FAILING. 1978 01:16:18,274 --> 01:16:19,742 WHEREAS WHAT WE WOULD LIKE TO 1979 01:16:19,742 --> 01:16:22,478 SEE IS WHEN THESE NEW TASKS ARE 1980 01:16:22,478 --> 01:16:26,749 INTRODUCED, THE MODEL IS ABLE TO 1981 01:16:26,749 --> 01:16:29,051 PERFORM FAIRLY WELL WITH PERHAPS 1982 01:16:29,051 --> 01:16:31,887 A SMALL REFRACTORY PERIOD OF 1983 01:16:31,887 --> 01:16:33,756 LEARNING, BUT ALSO IF A PREVIOUS 1984 01:16:33,756 --> 01:16:39,762 TASK IS RESPRINTRODUCED TO THE 1985 01:16:39,762 --> 01:16:41,063 MODEL WE WOULD LIKE TO IMPROVE 1986 01:16:41,063 --> 01:16:42,164 PERFORMANCE OVER THE PAST 1987 01:16:42,164 --> 01:16:43,098 PERFORMANCE. SO THAT'S A 1988 01:16:43,098 --> 01:16:43,999 DAUNTING TASK AND THERE ARE 1989 01:16:43,999 --> 01:16:45,801 SEVERAL STUDIES IN THE MACHINE 1990 01:16:45,801 --> 01:16:48,737 LEARNING COMMUNITY ON HOW TO 1991 01:16:48,737 --> 01:16:51,474 ADDRESS THIS PROBLEM. BUT WE 1992 01:16:51,474 --> 01:16:54,477 THINK THAT IS A HUGE OPPORTUNITY 1993 01:16:54,477 --> 01:16:59,949 HERE IN LOOKING AT THE LEARNING OR 1994 01:16:59,949 --> 01:17:01,016 PLASTICITY MECHANISMS THAT ARE 1995 01:17:01,016 --> 01:17:02,451 STUDIED IN THE NEUROSCIENCE 1996 01:17:02,451 --> 01:17:06,322 COMMUNITY AND HELP US BUILD THIS 1997 01:17:06,322 --> 01:17:07,289 HIGHLY FUNCTIONAL CONTINUAL 1998 01:17:07,289 --> 01:17:09,959 LEARNING MACHINES AND ONE OF THE 1999 01:17:09,959 --> 01:17:10,793 REASONS WE ARE HEAVILY 2000 01:17:10,793 --> 01:17:12,361 INTERESTED IN BUILDING THESE 2001 01:17:12,361 --> 01:17:14,063 SYSTEMS THAT ARE HIGHLY ENERGY 2002 01:17:14,063 --> 01:17:16,732 EFFICIENT. IF YOU LOOK AT THE 2003 01:17:16,732 --> 01:17:17,867 BIOLOGICAL BRAINS WE COULD SOLVE 2004 01:17:17,867 --> 01:17:22,972 SOME OF THESE PROBLEMS WITH VERY 2005 01:17:22,972 --> 01:17:26,075 FEW POWER. LIKE THE COMPUTATION 2006 01:17:26,075 --> 01:17:29,879 IS SHOWING JUST .1 WATTS OF ATP 2007 01:17:29,879 --> 01:17:33,315 REQUIRED. WHEREAS TO SOLVE 2008 01:17:33,315 --> 01:17:34,316 PROBLEMS OF SIMILAR SCALE OR -- 2009 01:17:34,316 --> 01:17:36,151 A SCALE THAT IS OF INTEREST TO 2010 01:17:36,151 --> 01:17:37,853 THE CONTINUAL LEARNING COMMUNITY 2011 01:17:37,853 --> 01:17:40,856 HERE, YOU WOULD SEE THAT THERE 2012 01:17:40,856 --> 01:17:42,925 IS A SIGNIFICANT TRAINING COST 2013 01:17:42,925 --> 01:17:46,395 AND A SIGNIFICANT ENERGY COST SO 2014 01:17:46,395 --> 01:17:47,763 PERHAPS INSTEAD OF LOOKING AT 2015 01:17:47,763 --> 01:17:50,032 TRADITIONAL APPROACHES OF TAKING 2016 01:17:50,032 --> 01:17:54,136 THESE MODELS AND MAPPING ONTO AN 2017 01:17:54,136 --> 01:17:55,237 A.I. ACCELERATORS, WOULD THESE 2018 01:17:55,237 --> 01:17:57,273 NEUROMORPHIC SUBSTRATES OR 2019 01:17:57,273 --> 01:17:59,575 NEUROMORPHIC PLATFORMS SERVE AS 2020 01:17:59,575 --> 01:18:00,843 A NATURAL SUBSTRATE FOR US? WE 2021 01:18:00,843 --> 01:18:04,580 ARE LOOKING AT THAT AS AN 2022 01:18:04,580 --> 01:18:06,348 OPPORTUNITY. SO PERHAPS I KNOW 2023 01:18:06,348 --> 01:18:08,784 WE HAVE BEEN TALKING YESTERDAY 2024 01:18:08,784 --> 01:18:12,922 ABOUT THIS RECIPROCITY BETWEEN 2025 01:18:12,922 --> 01:18:13,889 NEUROSCIENCE AND A.I. BUT THERE IS 2026 01:18:13,889 --> 01:18:16,058 THIS OPPORTUNITY OF LOOKING AT 2027 01:18:16,058 --> 01:18:18,561 WHAT THESE MODELS BRING IN FROM 2028 01:18:18,561 --> 01:18:20,129 THE NEUROSCIENCE COMMUNITY, THE 2029 01:18:20,129 --> 01:18:25,200 LEARNING AND PLASTICITY MECHANISMS 2030 01:18:25,200 --> 01:18:28,103 OR THE HIERARCHICAL STRUCTURE OR 2031 01:18:28,103 --> 01:18:29,371 SPARSE, DISTRIBUTED REPRESENTATIONS. 2032 01:18:29,371 --> 01:18:30,739 THERE'S THESE FEATURES IN THE 2033 01:18:30,739 --> 01:18:32,808 NEUROMORPHIC COMPUTING SYSTEMS 2034 01:18:32,808 --> 01:18:34,343 THAT EARLIER SPEAKERS HAVE 2035 01:18:34,343 --> 01:18:35,444 MENTIONED ABOUT WHICH ARE 2036 01:18:35,444 --> 01:18:38,047 NATURALLY TUNED TO COMPUTE IN 2037 01:18:38,047 --> 01:18:40,416 MEMORY AND THEY CAN WORK WITH 2038 01:18:40,416 --> 01:18:42,017 PROBLEMISTIC LEARNING OR ON- 2039 01:18:42,017 --> 01:18:44,486 DEVICE AND LOCAL LEARNING. SO IF 2040 01:18:44,486 --> 01:18:46,455 YOU CAN KIND OF LOOK AT THIS AS 2041 01:18:46,455 --> 01:18:48,290 A NATURAL SUBSTRATE MAYBE THIS 2042 01:18:48,290 --> 01:18:49,925 IS AN OPPORTUNITY TO BUILD THESE 2043 01:18:49,925 --> 01:18:51,961 CONTINUAL LEARNING MODELS IN A 2044 01:18:51,961 --> 01:18:52,995 VERY EFFICIENT WAY AND DEPLOY 2045 01:18:52,995 --> 01:18:56,298 THEM IN THIS HARDWARE. SO THIS 2046 01:18:56,298 --> 01:19:00,769 SPACE IS REALLY A MULTISCALE 2047 01:19:00,769 --> 01:19:02,438 OPTIMIZATION PROBLEM. BECAUSE 2048 01:19:02,438 --> 01:19:03,739 IF WE GROUP THE DIFFERENT 2049 01:19:03,739 --> 01:19:06,075 MECHANISMS IN WHICH WE'RE 2050 01:19:06,075 --> 01:19:06,842 ADDRESSING CONTINUAL LEARNING 2051 01:19:06,842 --> 01:19:09,111 PARTICULARLY NEUROINSPIRED 2052 01:19:09,111 --> 01:19:10,813 MECHANISMS, YOU CAN GROUP THEM 2053 01:19:10,813 --> 01:19:14,717 INTO DYNAMIC ARCHITECTURES, 2054 01:19:14,717 --> 01:19:16,318 SYNAPTIC CONSOLIDATION OR REPLAY 2055 01:19:16,318 --> 01:19:17,886 MECHANISMS. AND EACH OF THESE 2056 01:19:17,886 --> 01:19:19,054 REQUIRES EITHER SIGNIFICANTLY 2057 01:19:19,054 --> 01:19:20,789 RESOURCE-GROWING MECHANISMS 2058 01:19:20,789 --> 01:19:22,625 LIKE YOU MIGHT NEED LOT OF 2059 01:19:22,625 --> 01:19:24,793 MEMORY IN ORDER TO STORE ALL OF 2060 01:19:24,793 --> 01:19:26,495 YOUR PAST EXPERIENCES: AND WHAT 2061 01:19:26,495 --> 01:19:28,097 TYPES OF MEMORY ARE THE RIGHT 2062 01:19:28,097 --> 01:19:29,498 ONES TO DO THIS? 2063 01:19:29,498 --> 01:19:35,938 AND YOU MIGHT NEED TO TRACK YOUR 2064 01:19:35,938 --> 01:19:37,606 SYNAPTIC TRACES FOR SOME OF 2065 01:19:37,606 --> 01:19:38,641 THESE CONSOLIDATION MECHANISMS 2066 01:19:38,641 --> 01:19:41,610 AND HOW DO I REPRESENT A 2067 01:19:41,610 --> 01:19:43,512 SYNAPTIC- OR NEUROGENESIS ON 2068 01:19:43,512 --> 01:19:45,247 HARDWARE? THE SUBSTRATE IS VERY 2069 01:19:45,247 --> 01:19:47,416 DIFFERENT COMPARED TO THE 2070 01:19:47,416 --> 01:19:49,184 BIOLOGICAL SUBSTRATE. THERE IS A 2071 01:19:49,184 --> 01:19:50,252 HUGE OPPORTUNITY HERE IN HOW WE 2072 01:19:50,252 --> 01:19:52,187 ARE THINKING ABOUT DESIGNING THE 2073 01:19:52,187 --> 01:19:52,955 HARDWARE SYSTEMS. 2074 01:19:52,955 --> 01:19:54,556 WE LOOKED AT SOME OF THE 2075 01:19:54,556 --> 01:19:57,026 CONSOLIDATION MECHANISMS LIKE, 2076 01:19:57,026 --> 01:20:00,963 FOR EXAMPLE, ACTIVITY DEPENDENT 2077 01:20:00,963 --> 01:20:01,997 METAPLASTICITY WHERE WE ARE 2078 01:20:01,997 --> 01:20:03,632 TRYING TO PRESERVE THE 2079 01:20:03,632 --> 01:20:06,468 PLASTICITY OF PREVIOUSLY ACTIVE 2080 01:20:06,468 --> 01:20:07,903 NEURONS AND IS REDUCED TO 2081 01:20:07,903 --> 01:20:11,006 PRESERVE OLD KNOWLEDGE AND WE 2082 01:20:11,006 --> 01:20:14,343 KIND OF REALIZED THIS IN A 2083 01:20:14,343 --> 01:20:15,411 HARDWARE SYSTEM AND THE 2084 01:20:15,411 --> 01:20:17,179 CHALLENGE HERE IS THAT THIS ONE 2085 01:20:17,179 --> 01:20:19,448 TECHNIQUE IS NOT ENOUGH TO 2086 01:20:19,448 --> 01:20:20,983 DESIGN A CONTINUAL LEARNING 2087 01:20:20,983 --> 01:20:23,052 MECHANISM. IT HELPS US IN GETTING 2088 01:20:23,052 --> 01:20:25,220 TO SOLVING SMALLER PROBLEMS IN 2089 01:20:25,220 --> 01:20:27,723 THE CONTINUAL LEARNING DOMAIN 2090 01:20:27,723 --> 01:20:28,590 AND THERE IS -- ANOTHER 2091 01:20:28,590 --> 01:20:31,060 TECHNIQUE WE LOOKED AT, FOR 2092 01:20:31,060 --> 01:20:35,397 EXAMPLE, CALLED SYNAPTIC 2093 01:20:35,397 --> 01:20:37,733 CONSOLIDATION WHERE YOU TRY TO 2094 01:20:37,733 --> 01:20:39,768 ACCUMULATE KNOWLEDGE OVER MULTIPLE 2095 01:20:39,768 --> 01:20:40,769 TIMESCALES SO THIS IS DIFFICULT 2096 01:20:40,769 --> 01:20:42,971 TO TRANSFORM FROM SHORT-TERM TO 2097 01:20:42,971 --> 01:20:46,008 LONG-TERM MEMORY AND TO ENSURE 2098 01:20:46,008 --> 01:20:48,877 THAT INFORMATION IS RETAINED. REQUIRE 2099 01:20:48,877 --> 01:20:50,279 MULTIPLE PLASTICITY EVENTS WHEN YOU 2100 01:20:50,279 --> 01:20:51,146 TRY TO REALIZE SOMETHING LIKE 2101 01:20:51,146 --> 01:20:53,482 THIS AND WHAT MY STUDENT FATIMA 2102 01:20:53,482 --> 01:20:56,318 HERE IS WHEN SHE LOOKED AT 2103 01:20:56,318 --> 01:20:57,920 REALIZING THIS IN A HARDWARE 2104 01:20:57,920 --> 01:20:59,455 SYSTEM, SHE REALIZED THERE IS 2105 01:20:59,455 --> 01:21:01,123 PERHAPS A NATURAL FEATURE IN 2106 01:21:01,123 --> 01:21:02,257 SOME OF THESE NONVOLATILE 2107 01:21:02,257 --> 01:21:04,393 MEMORY DEVICES. WHERE YOU COULD 2108 01:21:04,393 --> 01:21:08,030 UPDATE THE PROBABILISTIC 2109 01:21:08,030 --> 01:21:09,264 METAPLASTICITY INSTEAD OF 2110 01:21:09,264 --> 01:21:11,467 UPDATING THE WEIGHTS YOU ARE TRYING 2111 01:21:11,467 --> 01:21:12,634 TO UPDATE THE PROBABILITY OF THE 2112 01:21:12,634 --> 01:21:15,571 WEIGHT BEING UPDATED. SO IT COULD 2113 01:21:15,571 --> 01:21:17,506 REDUCE A COMPLEXITY OF YOUR 2114 01:21:17,506 --> 01:21:21,910 HARDWARE AND ALSO SAVE SOME POWER 2115 01:21:21,910 --> 01:21:23,278 AND ENERGY. SO CURRENTLY OUR 2116 01:21:23,278 --> 01:21:29,752 GROUP IS REALLY LOOKING AT 2117 01:21:29,752 --> 01:21:30,953 DIFFERENT SPACES OR PROBLEMS 2118 01:21:30,953 --> 01:21:32,354 WITHIN THE CONTINUAL LEARNING 2119 01:21:32,354 --> 01:21:34,857 DOMAIN. ONE OF THE BIG PROJECTS 2120 01:21:34,857 --> 01:21:37,159 WE ARE WORKING ON IS HOW DO YOU 2121 01:21:37,159 --> 01:21:40,629 ACTUALLY DESIGN THESE SYSTEMS IN 2122 01:21:40,629 --> 01:21:43,031 THE TEMPORAL DOMAIN? SO WE ARE 2123 01:21:43,031 --> 01:21:45,601 LOOKING -- WORKING CLOSELY WITH 2124 01:21:45,601 --> 01:21:48,203 A NEUROSCIENTIST IN 2125 01:21:48,203 --> 01:21:49,705 UNDERSTANDING THE TEMPORAL 2126 01:21:49,705 --> 01:21:51,373 SCAFFOLDING HYPOTHESIS THAT 2127 01:21:51,373 --> 01:21:53,375 HAPPENS DURING SLEEP AND HOW DO 2128 01:21:53,375 --> 01:21:55,911 WE BUILD NEW TYPES OF MODELS 2129 01:21:55,911 --> 01:21:58,147 USING THIS KIND OF HYPOTHESIS SO 2130 01:21:58,147 --> 01:21:59,081 THAT IT IS REALLY ENERGY 2131 01:21:59,081 --> 01:22:00,482 EFFICIENT IN LEARNING IN THE 2132 01:22:00,482 --> 01:22:03,986 TEMPORAL DOMAIN AND THANKS TO 2133 01:22:03,986 --> 01:22:05,154 GRACE'S BRAID PROJECT WE WERE 2134 01:22:05,154 --> 01:22:06,421 SUPPORTED THROUGH THAT EFFORT. WE 2135 01:22:06,421 --> 01:22:08,524 ARE ALSO LOOKING AT COMBINING 2136 01:22:08,524 --> 01:22:12,628 MULTIPLE OF THESE PLASTICITY 2137 01:22:12,628 --> 01:22:14,463 MECHANISMS AND ALSO LOOKING AT 2138 01:22:14,463 --> 01:22:15,631 LOCALITY AND PLASTICITY WITHIN 2139 01:22:15,631 --> 01:22:18,267 THE SPIKING NEURAL NETWORK 2140 01:22:18,267 --> 01:22:20,369 DOMAIN. WELL IN ORDER TO 2141 01:22:20,369 --> 01:22:22,337 ACHIEVE THIS, I THINK ONE THING 2142 01:22:22,337 --> 01:22:23,806 WE NEED IN THE STOPGAP IS -- 2143 01:22:23,806 --> 01:22:26,475 >> ONE MINUTE WARNING. 2144 01:22:26,475 --> 01:22:29,144 >> LARGE-SCALE SPIKING NEURAL 2145 01:22:29,144 --> 01:22:30,479 NETWORKS AND PERHAPS A PLUG IN 2146 01:22:30,479 --> 01:22:32,381 PLAY APPROACH TO INTEGRATE THESE 2147 01:22:32,381 --> 01:22:33,849 KINDS OF MECHANISMS DIRECTLY 2148 01:22:33,849 --> 01:22:36,952 INTO OUR MODELS AND HARDWARE. 2149 01:22:36,952 --> 01:22:41,523 AND MAYBE WE COULD EMBED THIS IN 2150 01:22:41,523 --> 01:22:43,492 THESE KINDS OF PRECISION A.I. HEALTH 2151 01:22:43,492 --> 01:22:45,294 ASSISTANCE WHETHER IT'S THE 2152 01:22:45,294 --> 01:22:46,428 TWINS OR KINS I THINK THAT WILL 2153 01:22:46,428 --> 01:22:48,430 BE A HUGE OPPORTUNITY. 2154 01:22:48,430 --> 01:22:49,798 BUT WAIT, WHILE THIS SOUNDS 2155 01:22:49,798 --> 01:22:53,202 GREAT, BUT WE DO NOT HAVE ACCESS 2156 01:22:53,202 --> 01:22:54,603 TO THIS NEUROMORPHIC 2157 01:22:54,603 --> 01:22:55,237 INFRASTRUCTURE WHERE WE COULD 2158 01:22:55,237 --> 01:22:58,473 REALLY SIMULATE THIS LARGE-SCALE 2159 01:22:58,473 --> 01:23:00,542 MODELS. AND THIS IS WHERE WE 2160 01:23:00,542 --> 01:23:03,445 RECENTLY LAUNCHED A PROJECT CALLED 2161 01:23:03,445 --> 01:23:04,847 THE NEUROMORPHIC COMMONS. THIS IS 2162 01:23:04,847 --> 01:23:11,787 SUPPORTED BY THE NSF CIRC GRANT AND 2163 01:23:11,787 --> 01:23:14,656 THE IDEA IS TO BUILD THIS OPEN 2164 01:23:14,656 --> 01:23:16,792 SOURCE NATIONAL SCALE HUB AND 2165 01:23:16,792 --> 01:23:18,994 WORKING VERY CLOSELY WITH 2166 01:23:18,994 --> 01:23:20,896 INDUSTRIAL PARTNERS IN TRYING TO 2167 01:23:20,896 --> 01:23:23,732 GET ACCESS TO THE COMMUNITY. 2168 01:23:23,732 --> 01:23:27,035 BECAUSE RIGHT NOW, ACCESS TO THE 2169 01:23:27,035 --> 01:23:29,137 NEUROMORPHIC HARDWARE 2170 01:23:29,137 --> 01:23:30,172 INFRASTRUCTURE IS A BARRIER, 2171 01:23:30,172 --> 01:23:32,107 ENTRY BARRIER FOR A LOT OF SMALL 2172 01:23:32,107 --> 01:23:34,209 RESEARCH TEAMS OR ACADEMIC LABS 2173 01:23:34,209 --> 01:23:37,279 SO WE WANT TO REDUCE THOSE 2174 01:23:37,279 --> 01:23:40,115 BARRIERS AND HAVE PEOPLE 2175 01:23:40,115 --> 01:23:46,321 DISCOVER OR UNDERSTAND THESE 2176 01:23:46,321 --> 01:23:49,224 COMPUTATIONAL MODELS BETTER AND 2177 01:23:49,224 --> 01:23:50,259 HOPEFULLY HELP THE NEUROSCIENCE 2178 01:23:50,259 --> 01:23:51,260 COMMUNITY AND OTHER COMMUNITIES 2179 01:23:51,260 --> 01:23:52,628 IN THE LIFE SCIENCES AND THIS IS 2180 01:23:52,628 --> 01:23:55,097 A COLLABORATION BETWEEN 2181 01:23:55,097 --> 01:23:57,032 UNIVERSITY OF TENNESSEE, SAN 2182 01:23:57,032 --> 01:23:58,667 DIEGO, HARVARD AND SEVERAL OTHER 2183 01:23:58,667 --> 01:23:59,868 INDUSTRY PARTNERS BUT MORE 2184 01:23:59,868 --> 01:24:02,204 IMPORTANTLY THIS IS A COMMUNITY 2185 01:24:02,204 --> 01:24:03,205 DRIVEN EFFORT SO IF YOU'RE 2186 01:24:03,205 --> 01:24:04,907 INTERESTED PLEASE REACH OUT TO 2187 01:24:04,907 --> 01:24:07,276 US THROUGH THIS URL THAT'S 2188 01:24:07,276 --> 01:24:08,810 POSTED HERE. I WILL BE HAPPY 2189 01:24:08,810 --> 01:24:10,946 TO CHAT MORE AND THANKS TO THIS 2190 01:24:10,946 --> 01:24:12,848 AMAZING GROUP OF RESEARCHERS I 2191 01:24:12,848 --> 01:24:14,583 GET TO WORK WITH WHO ENTRUST ME 2192 01:24:14,583 --> 01:24:17,252 IN DOING THIS RESEARCH AND 2193 01:24:17,252 --> 01:24:18,954 THANKS TO OUR SPONSORS FOR 2194 01:24:18,954 --> 01:24:29,197 MAKING THIS HAPPEN. THANK YOU 2195 01:24:29,197 --> 01:24:30,766 >> PLEASE WELCOME OUR LAST 2196 01:24:30,766 --> 01:24:31,300 SPEAKER MITRA HARTMANN FROM 2197 01:24:31,300 --> 01:24:34,536 NORTHWESTERN UNIVERSITY. 2198 01:24:34,536 --> 01:24:37,005 >> THANKS VERY MUCH FOR THE 2199 01:24:37,005 --> 01:24:38,640 INVITATION, I NEED TO WARN THE 2200 01:24:38,640 --> 01:24:41,076 AUDIENCE WHEN WE TESTED SOME 2201 01:24:41,076 --> 01:24:46,148 VIDEOS THIS MORNING WE FOUND THE 2202 01:24:46,148 --> 01:24:48,584 COMPUTER WAS NOT ABLE TO KEEP UP 2203 01:24:48,584 --> 01:24:51,219 SO I MAY HAVE TO COUNT ON YOUR 2204 01:24:51,219 --> 01:24:52,087 GENERALIZED INTELLIGENCE TO 2205 01:24:52,087 --> 01:24:54,289 FOLLOW ALONG. WE THINK ABOUT 2206 01:24:54,289 --> 01:24:57,059 THE BRAIN AS A CLOSED LOOP 2207 01:24:57,059 --> 01:24:57,893 SYSTEM AND THIS PHOTOGRAPH 2208 01:24:57,893 --> 01:24:59,261 INDICATES WHY WE NEED TO TAKE 2209 01:24:59,261 --> 01:25:00,595 THIS VERY SERIOUSLY. BOTH OF 2210 01:25:00,595 --> 01:25:02,798 THESE ARE MAMMALS BUT THEIR 2211 01:25:02,798 --> 01:25:06,234 NEURONS ARE TUNED FOR COMPLETELY 2212 01:25:06,234 --> 01:25:07,135 SENSORIMOTOR SYSTEMS FOR THEIR NICHE 2213 01:25:07,135 --> 01:25:11,940 I AM GOING TO TELL YOU ABOUT THE 2214 01:25:11,940 --> 01:25:14,776 RAT VIBRISSAL SYSTEM AS A MODEL OF 2215 01:25:14,776 --> 01:25:16,244 EMBODIED INTELLIGENCE. AND WE CAN 2216 01:25:16,244 --> 01:25:20,282 SEE SOME NEW IDEAS AND RESOURCES 2217 01:25:20,282 --> 01:25:21,783 THAT THE SCIENTIFIC COMMUNITY 2218 01:25:21,783 --> 01:25:23,418 WILL NEED OVER THE NEXT FEW 2219 01:25:23,418 --> 01:25:25,687 YEARS. RATS DO SOME FAIRLY 2220 01:25:25,687 --> 01:25:26,989 COMPLICATED THINGS WITH THEIR 2221 01:25:26,989 --> 01:25:28,657 WHISKERS YOU CAN SEE RIGHT NOW 2222 01:25:28,657 --> 01:25:31,526 IT'S ABOUT TO PUT ITS PAW RIGHT 2223 01:25:31,526 --> 01:25:33,996 ON THE PLACE ON THE TINKERTOY 2224 01:25:33,996 --> 01:25:36,531 WHERE THE WHISKERS WERE 2225 01:25:36,531 --> 01:25:37,866 IMMEDIATELY BEFORE. AND THE 2226 01:25:37,866 --> 01:25:41,470 WHISKERS ARE ARRANGED IN VERY 2227 01:25:41,470 --> 01:25:43,672 REGULAR ROWS AND COLUMNS. 2228 01:25:43,672 --> 01:25:46,408 THEY ARE GIVEN LETTERS AND 2229 01:25:46,408 --> 01:25:53,882 NUMBERS. A-E,1-5 THE WAY THAT 2230 01:25:53,882 --> 01:25:57,219 THE WHISKERS ARE ACTUATED THEY 2231 01:25:57,219 --> 01:25:59,154 INSERT INTO FOLLICLES AND THESE 2232 01:25:59,154 --> 01:26:01,456 SLING MUSCLES PULL BACK. THE SKIN SERVES AS A FULCRUM AND THE 2233 01:26:01,456 --> 01:26:03,091 WHISKERS ARE PROTRACTED FORWARD. 2234 01:26:03,091 --> 01:26:04,526 THIS IS A CARTOON VERSION. 2235 01:26:04,526 --> 01:26:07,029 THERE'S ALSO A LOT OF EXTRINSIC 2236 01:26:07,029 --> 01:26:13,068 MUSCLES THAT ATTACH THE WHISKER TO 2237 01:26:13,068 --> 01:26:14,236 THE SKULL. NO SENSORS ALONG THE WHISKER LENGTH 2238 01:26:14,236 --> 01:26:17,005 BUT THE FOLLICLE IS REPLETE WITH 2239 01:26:17,005 --> 01:26:18,740 MECHANORECEPTORS. SO ALL SENSING 2240 01:26:18,740 --> 01:26:21,143 OCCURS AT THE WHISKER BASE AND THERE 2241 01:26:21,143 --> 01:26:23,645 ARE NO PROPRIOCEPTORS IN THE 2242 01:26:23,645 --> 01:26:24,513 WHISKER MUSCLES. 2243 01:26:24,513 --> 01:26:26,481 I'M GOING TO WALK YOU THROUGH 2244 01:26:26,481 --> 01:26:28,116 THREE STAGES OF THE SENSORY-MOTOR- 2245 01:26:28,116 --> 01:26:28,517 BRAIN ENVIRONMENT LOOP. 2246 01:26:28,517 --> 01:26:30,352 THE FIRST IS MUSCLES. THE SECOND 2247 01:26:30,352 --> 01:26:33,088 IS TACTILE. THE THIRD IS ENCODING 2248 01:26:33,088 --> 01:26:35,390 BY THE PRIMARY SENSORY NEURONS. AND 2249 01:26:35,390 --> 01:26:36,925 AT THE LAST BIT WE'LL LOOK TOWARDS 2250 01:26:36,925 --> 01:26:38,627 THE FUTURE, SO FOR MUSCLES THE 2251 01:26:38,627 --> 01:26:40,662 MOST I WILL HAVE A CHANCE TO 2252 01:26:40,662 --> 01:26:42,597 TELL YOU ABOUT IS WE'RE BUILDING 2253 01:26:42,597 --> 01:26:46,268 BIOMECHANICAL MODELS OF ALL THE 2254 01:26:46,268 --> 01:26:49,104 COMPLEX MUSCULATURE THAT MOVES 2255 01:26:49,104 --> 01:26:51,073 THE FOLLICLES AND IT'S A LOT 2256 01:26:51,073 --> 01:26:54,476 MORE COMPLICATED THAN THE ORIGINAL 2257 01:26:54,476 --> 01:26:55,310 DIAGRAM WILL SUGGEST. THAT'S 2258 01:26:55,310 --> 01:26:56,645 ALL I HAVE TIME FOR TODAY BUT 2259 01:26:56,645 --> 01:27:02,384 FOR THE TACTILE-MECHANICAL SIGNALS, 2260 01:27:02,384 --> 01:27:03,819 I WILL TELL YOU MORE. 2261 01:27:03,819 --> 01:27:05,153 TO UNDERSTAND THE SIGNALS IS TO 2262 01:27:05,153 --> 01:27:06,188 GET THE MORPHOLOGY RIGHT. 2263 01:27:06,188 --> 01:27:08,123 AND THIS INVOLVED A LOT MORE 2264 01:27:08,123 --> 01:27:09,958 THAN TAKING A THREE-DIMENSIONAL 2265 01:27:09,958 --> 01:27:11,526 SCAN. WE STARTED THIS WORK WAY 2266 01:27:11,526 --> 01:27:16,164 BACK IN 2011. OUR MODEL INCLUDES 2267 01:27:16,164 --> 01:27:16,865 WHISKERS AND SKULL FEATURES. AND 2268 01:27:16,865 --> 01:27:19,501 WE HAVE DONE COMPARATIVE 2269 01:27:19,501 --> 01:27:21,036 STUDIES ACROSS THE MOUSE, RAT 2270 01:27:21,036 --> 01:27:26,475 AND HARBOR SEAL. SO NEXT IS THE 2271 01:27:26,475 --> 01:27:28,877 MECHANICS. MY STUDENT NADINA ZWEIFEL 2272 01:27:28,877 --> 01:27:33,248 CREATED THE WHISKIT PHYSICS 2273 01:27:33,248 --> 01:27:36,985 SIMULATOR. THAT IS ALL THE TORQUES 2274 01:27:36,985 --> 01:27:40,155 MOMENTS AND FORCES AT THE WHISKER 2275 01:27:40,155 --> 01:27:42,424 BASE. EXAMPLE OF WHISKIT IN ACTION. 2276 01:27:42,424 --> 01:27:45,460 THIS IS A TYPICAL LABORATORY 2277 01:27:45,460 --> 01:27:47,896 EXPERIMENT WHERE IT IS WHISKING 2278 01:27:47,896 --> 01:27:50,031 AGAINST TWO PEGS AND THE MECHANICAL 2279 01:27:50,031 --> 01:27:51,767 SIGNALS FOR ONE SIDE OF THE FACE 2280 01:27:51,767 --> 01:27:53,969 IS IN THE NEXT VIDEO. MOMENTS IN RED, 2281 01:27:53,969 --> 01:27:55,871 FORCES IN BLUE. EACH CIRCLE REPRESENTS 2282 01:27:55,871 --> 01:28:04,646 ONE OF THE FOLLICLES. WHISKERS DOWN BELOW SHOW VECTOR MAGNITUDE. 2283 01:28:04,646 --> 01:28:08,250 SO HERE ARE THE SIGNALS ASSOCIATED 2284 01:28:08,250 --> 01:28:09,484 WITH THE RAT WHISKING AGAINST PEGS. 2285 01:28:09,484 --> 01:28:13,989 THESE ARE DATA THAT A BIOLOGIST WOULD NEVER BE ABLE TO OBTAINBECAUSE YOU CAN'T RECORD 2286 01:28:13,989 --> 01:28:15,257 THE MECHANICAL SIGNALS WITHOUT 2287 01:28:15,257 --> 01:28:17,392 INTERFERING WITH THE SIGNALS. SO 2288 01:28:17,392 --> 01:28:19,528 MORE INTERESTING IN A LABORATORY 2289 01:28:19,528 --> 01:28:21,363 EXPERIMENT IS THAT WE CAN GO OUT 2290 01:28:21,363 --> 01:28:25,534 INTO A DARK CHICAGO ALLEYWAY 2291 01:28:25,534 --> 01:28:29,538 AND LOOK AT A SIMULATED RAT 2292 01:28:29,538 --> 01:28:30,605 EXPLORING A DRAINPIPE. WE CAN 2293 01:28:30,605 --> 01:28:32,807 PUT THE ANIMAL IN ITS MORE 2294 01:28:32,807 --> 01:28:34,476 NATURAL ENVIRONMENT AND LOOK AT 2295 01:28:34,476 --> 01:28:35,310 THE SIGNALS BEING GENERATED 2296 01:28:35,310 --> 01:28:36,845 DURING THIS TYPE OF BEHAVIOR. 2297 01:28:36,845 --> 01:28:38,246 THEN WE CAN RETURN TO THE 2298 01:28:38,246 --> 01:28:40,382 LABORATORY AND LOOK AT SIGNALS 2299 01:28:40,382 --> 01:28:41,016 ASSOCIATED WITH A RAT THAT'S 2300 01:28:41,016 --> 01:28:42,784 BEEN TRAINED TO SEARCH FOR A 2301 01:28:42,784 --> 01:28:45,120 WATER REWARD. AND IN THIS CASE 2302 01:28:45,120 --> 01:28:46,388 WE HAVE THREE CAMERA VIEWS SO 2303 01:28:46,388 --> 01:28:48,356 WE'RE ABLE TO SUPERIMPOSE THE 2304 01:28:48,356 --> 01:28:51,726 MODEL ON TOP OF THE RAT IN THESE 2305 01:28:51,726 --> 01:28:53,295 CAMERA VIEWS AND TRACK THE 2306 01:28:53,295 --> 01:28:54,663 WHISKERS AND HEAD IN EACH FRAME. 2307 01:28:54,663 --> 01:28:56,631 AND THAT GIVES US -- THEN ALLOWS 2308 01:28:56,631 --> 01:28:58,700 US TO LOOK AT THE MECHANICAL 2309 01:28:58,700 --> 01:28:59,901 SIGNALS HERE ARE EXAMPLES JUST 2310 01:28:59,901 --> 01:29:04,472 FOR THE C1 WHISKER. WE HAVE ALL 2311 01:29:04,472 --> 01:29:06,408 THREE MOMENTS PLOTTED IN TIME. 2312 01:29:06,408 --> 01:29:08,009 THE IMPORTANT THING HERE IS 2313 01:29:08,009 --> 01:29:09,644 THESE ARE THE INPUT SIGNAL TO 2314 01:29:09,644 --> 01:29:12,314 THE PRIMARY SENSORY NEURONS THAT 2315 01:29:12,314 --> 01:29:13,815 ENCODE WHISKER-BASED TOUCH. 2316 01:29:13,815 --> 01:29:16,017 SO MOVING ONTO CODING BY PRIMARY 2317 01:29:16,017 --> 01:29:18,286 SENSORY NEURONS I WISH I COULD 2318 01:29:18,286 --> 01:29:19,554 PRESENT DATA FROM THE AWAKE 2319 01:29:19,554 --> 01:29:21,089 ANIMAL BUT WE'RE GOING RETURN TO 2320 01:29:21,089 --> 01:29:22,857 THE ANESTHETIZED ANIMAL. (VIDEO) 2321 01:29:22,857 --> 01:29:24,626 HERE WE HAVE TWO CAMERA VIEWS. WHAT WE CAN DO 2322 01:29:24,626 --> 01:29:27,028 THOUGH IS MANUALLY DEFLECT THE WHISKER AT 2323 01:29:27,028 --> 01:29:28,930 DIFFERENT ARCLENGTHS, SPEEDS, AND DIRECTIONS USING 2324 01:29:28,930 --> 01:29:30,899 A THIN 0.5 MM PENCIL LEAD.  WE CAN THEN MERGE 2325 01:29:30,899 --> 01:29:32,734 THE TWO CAMERA VIEWS TO TRACK THE 3D MOTIONS 2326 01:29:32,734 --> 01:29:33,835 OF THE FLEXIBLE WHISKER TO LOOK AT HOW IT 2327 01:29:33,835 --> 01:29:35,303 MECHANICALLY RESPONDS DURING EACH TOUCH. 2328 01:29:35,303 --> 01:29:36,271 BECAUSE WE HAVE 2329 01:29:36,271 --> 01:29:37,772 THREE-DIMENSIONAL MODEL -- 2330 01:29:37,772 --> 01:29:39,407 BECAUSE WE HAVE THESE MECHANICAL 2331 01:29:39,407 --> 01:29:40,675 MODELS WE CAN TRACK AND 2332 01:29:40,675 --> 01:29:42,510 RECONSTRUCT THE 2333 01:29:42,510 --> 01:29:43,812 THREE-DIMENSIONAL SHAPE OF THE 2334 01:29:43,812 --> 01:29:45,347 WHISKER, COMPUTE MECHANICAL 2335 01:29:45,347 --> 01:29:47,215 SIGNALS AT THE WHISKER BASE AND 2336 01:29:47,215 --> 01:29:50,819 ARE ABLE TO CHARACTERIZE THE 2337 01:29:50,819 --> 01:29:54,456 TUNING SPACES AND THEIR DISTRIBUTION. 2338 01:29:54,456 --> 01:29:56,258 SO THE UPSHOT, THE FINAL RESULT, THE 2339 01:29:56,258 --> 01:29:57,325 TAKE HOME FROM THESE EXPERIMENTS 2340 01:29:57,325 --> 01:29:59,327 THAT WE DID, IS THAT EACH 2341 01:29:59,327 --> 01:30:01,296 PRIMARY SENSORY NEURON IS 2342 01:30:01,296 --> 01:30:03,265 REPRESENTING MANY MECHANICAL 2343 01:30:03,265 --> 01:30:05,700 FEATURES SIMULTANEOUSLY AND THAT 2344 01:30:05,700 --> 01:30:08,236 RESULT CONTRASTS DIRECTLY WITH 2345 01:30:08,236 --> 01:30:11,706 THE TRADITIONAL VIEW THAT EACH NEURON 2346 01:30:11,706 --> 01:30:13,141 REPRESENTS A SINGLE KINEMATIC STIMULUS FEATURE. OUR 2347 01:30:13,141 --> 01:30:14,142 PRELIMINARY DATA SUGGESTS THAT 2348 01:30:14,142 --> 01:30:16,911 THIS WILL RESULT WILL GENERALIZE TO 2349 01:30:16,911 --> 01:30:19,948 AWAKE ANIMAL. THIS REPRESENTATION IN 2350 01:30:19,948 --> 01:30:21,583 PRIMARY SENSORY NEURON IS GOING TO 2351 01:30:21,583 --> 01:30:23,018 DIRECTLY CONSTRAIN THE COMPUTATIONS 2352 01:30:23,018 --> 01:30:24,119 PERFORMED BY MORE CENTRAL 2353 01:30:24,119 --> 01:30:26,221 NEURONS. SO I WALKED YOU THROUGH 2354 01:30:26,221 --> 01:30:27,489 THE THREE STEPS HERE. NOW WE'RE 2355 01:30:27,489 --> 01:30:29,090 GOING TO TAKE A LOOK AT THE 2356 01:30:29,090 --> 01:30:31,159 FUTURE. WHAT IS THIS FUTURE? 2357 01:30:31,159 --> 01:30:33,028 WELL THESE PRIMARY SENSORY 2358 01:30:33,028 --> 01:30:34,396 NEURONS ARE OF COURSE PROJECTING 2359 01:30:34,396 --> 01:30:35,830 TO SECONDARY NEURONS THAT WILL 2360 01:30:35,830 --> 01:30:37,966 COMPLETE THE LOOP. THEY PROJECT 2361 01:30:37,966 --> 01:30:40,435 BACK OUT TO FACIAL MOTOR NUCLEUS 2362 01:30:40,435 --> 01:30:41,936 NEURONS THAT CONTROL THE 2363 01:30:41,936 --> 01:30:43,238 WHISKERS BUT ALSO PROJECT TO ALL 2364 01:30:43,238 --> 01:30:44,873 OF YOUR FAVORITE MORE CENTRAL 2365 01:30:44,873 --> 01:30:46,941 STRUCTURES AND SO WHAT I WOULD 2366 01:30:46,941 --> 01:30:49,144 LIKE YOU TO IMAGINE IS HAVING 2367 01:30:49,144 --> 01:30:50,779 ACCESS TO THE RESPONSES OF THESE 2368 01:30:50,779 --> 01:30:53,081 SECONDARY SENSORY NEURONS, OVER 2369 01:30:53,081 --> 01:30:55,050 MILLIONS OF SIMULATIONS WITH 2370 01:30:55,050 --> 01:30:57,519 REALISTIC CLOSED-LOOP 2371 01:30:57,519 --> 01:30:58,219 BIOMECHANICS. 2372 01:30:58,219 --> 01:30:58,887 IMPORTANTLY WE DON'T REALLY NEED 2373 01:30:58,887 --> 01:31:01,122 THIS TO BE THE PERFECT RAT. WE 2374 01:31:01,122 --> 01:31:02,123 JUST NEED ITS PARAMETERS TO BE 2375 01:31:02,123 --> 01:31:03,892 IN THE RANGE OF MECHANICALLY 2376 01:31:03,892 --> 01:31:05,827 REASONABLE RATS. AND HERE WE 2377 01:31:05,827 --> 01:31:07,429 HAVE TAKEN INSPIRATION FROM THE 2378 01:31:07,429 --> 01:31:10,565 FIELD OF EVOLUTIONARY ROBOTICS. 2379 01:31:10,565 --> 01:31:12,300 SO LET'S LOOK AT FUTURE TOOLS, 2380 01:31:12,300 --> 01:31:14,169 RESOURCES AND APPROACHES. AS I 2381 01:31:14,169 --> 01:31:15,870 THINK ABOUT WHAT WE'RE GOING TO 2382 01:31:15,870 --> 01:31:18,173 NEED AS WE START TO TACKLE THIS 2383 01:31:18,173 --> 01:31:20,141 REALLY DIFFICULT PROBLEM OF THE 2384 01:31:20,141 --> 01:31:21,876 SECONDARY STAGE NEURONS, WE'RE 2385 01:31:21,876 --> 01:31:24,646 GOING TO NEED A LOT OF CIRCUIT 2386 01:31:24,646 --> 01:31:26,414 BREAKING. WE NEED FUNCTIONAL ANATOMY. 2387 01:31:26,414 --> 01:31:28,249 WE NEED TO CHARACTERIZE THESE 2388 01:31:28,249 --> 01:31:29,617 DESCENDING CONNECTIONS. WE NEED 2389 01:31:29,617 --> 01:31:30,552 RIGOROUS BEHAVIORAL QUANTIFICATION 2390 01:31:30,552 --> 01:31:31,986 BECAUSE SENSORY ACQUISITION 2391 01:31:31,986 --> 01:31:34,322 BEHAVIORS ARE GOING TO ESTABLISH 2392 01:31:34,322 --> 01:31:35,423 THESE MECHANICAL REGULARITIES 2393 01:31:35,423 --> 01:31:36,825 THAT THE BRAIN CAN EXPLOIT. WE 2394 01:31:36,825 --> 01:31:38,226 NEED TO TAKE A COMPARATIVE - I 2395 01:31:38,226 --> 01:31:39,427 THINK WE NEED TO TAKE A 2396 01:31:39,427 --> 01:31:41,229 COMPARATIVE APPROACH SO WE NEED 2397 01:31:41,229 --> 01:31:44,032 TO GO CROSS-MODAL. SO VISION IS 2398 01:31:44,032 --> 01:31:45,967 AN ALLOCTIVE SYSTEM BUT TOUCH 2399 01:31:45,967 --> 01:31:47,969 IS HOMEOACTIVE. AND THOSE ARE 2400 01:31:47,969 --> 01:31:49,437 IMPORTANT DISTINCTIONS THAT WILL 2401 01:31:49,437 --> 01:31:52,974 HAVE IMPLICATIONS FOR EFFERENCE 2402 01:31:52,974 --> 01:31:54,576 COPY AND THE SELF-PREDICTION OF 2403 01:31:54,576 --> 01:31:56,277 SELF-GENERATED MOTION. 2404 01:31:56,277 --> 01:31:59,981 WE ALSO NEED TO LOOK 2405 01:31:59,981 --> 01:32:00,949 ACROSS SPECIES. INVERTEBRATES 2406 01:32:00,949 --> 01:32:03,184 HAVE BEEN A BIG MODEL - THE RESULTS 2407 01:32:03,184 --> 01:32:05,153 FROM INVERTEBRATESHAVE BEEN A BIG 2408 01:32:05,153 --> 01:32:07,522 MODEL FOR MY WORK. WE NEED A.I. TO 2409 01:32:07,522 --> 01:32:11,359 PREDICT FAMILIES OF CIRCUITS TO 2410 01:32:11,359 --> 01:32:12,853 GENERATE TESTABLE HYPOTHESES. NEED 2411 01:32:12,853 --> 01:32:14,631 RIGOROUS APPROACHES FORVALIDATING MODELS 2412 01:32:14,631 --> 01:32:16,265 OF NEURAL POPULATIONS ACROSS DIVERSE 2413 01:32:16,265 --> 01:32:16,765 BEHAVIORS. 2414 01:32:16,765 --> 01:32:18,433 AND IMPORTANTLY THESE SOLUTIONS 2415 01:32:18,433 --> 01:32:20,402 MUST WORK ON HARDWARE THAT IS ON 2416 01:32:20,402 --> 01:32:21,436 THE ANIMAL NOT JUST IN 2417 01:32:21,436 --> 01:32:23,671 SIMULATION THAT MEANS WE NEED 2418 01:32:23,671 --> 01:32:26,040 MORE BIOMECHANICALLY REALISTIC 2419 01:32:26,040 --> 01:32:27,375 MODELS OF ALL OF THE DIFFERENT 2420 01:32:27,375 --> 01:32:29,077 STRUCTURES AND THEIR ROLE IN THE 2421 01:32:29,077 --> 01:32:29,911 SENSORY DATA ACQUISITION PROCESS. 2422 01:32:29,911 --> 01:32:31,513 BUT AS I MENTIONED YESTERDAY IN 2423 01:32:31,513 --> 01:32:32,614 ONE OF THE QUESTIONS, THE 2424 01:32:32,614 --> 01:32:37,185 SOLUTION THAT WE NEED BETTER 2425 01:32:37,185 --> 01:32:39,687 SOLUTIONS FOR THE SIM2REAL 2426 01:32:39,687 --> 01:32:43,291 PROBLEMS. ONCE WE START TO ADD REAL SENSORS, MOTORS WITH NOISE, ALLTHESE 2427 01:32:43,291 --> 01:32:46,194 NONSTATIONARITIES, NONLINEARITIES 2428 01:32:46,194 --> 01:32:47,862 WE MAY FIND WE NEED DIFFERENT CONTROL CIRCUITRY. 2429 01:32:47,862 --> 01:32:49,130 WE NEED BETTER MODELS OF COLLISIONS, CONTACT MECHANICS, FRICTION ANDFLUID FLOW. 2430 01:32:49,130 --> 01:32:50,732 HARDWARE MODELS HAVE BEEN AN 2431 01:32:50,732 --> 01:32:52,767 ESSENTIAL COMPONENT OF OUR WORK, 2432 01:32:52,767 --> 01:32:54,736 AND I WOULD LIKE TO SAY THAT -- 2433 01:32:54,736 --> 01:32:56,738 SO ON THE LEFT HERE WE'RE 2434 01:32:56,738 --> 01:33:01,443 SHOWING AN ARTIFICIAL WHISKER 2435 01:33:01,443 --> 01:33:03,344 COMPARED WITH THE REAL WHISKER. 2436 01:33:03,344 --> 01:33:05,213 WE'RE DOING EXPERIMENTS IN SHAPE 2437 01:33:05,213 --> 01:33:06,347 EXTRACTION AND THE WAY I LIKE TO 2438 01:33:06,347 --> 01:33:07,916 THINK ABOUT IT IS THAT 2439 01:33:07,916 --> 01:33:08,983 NEUROSCIENCE HAS BEEN THE 2440 01:33:08,983 --> 01:33:13,088 INSPIRATION FOR ARTIFICIAL 2441 01:33:13,088 --> 01:33:14,589 INTELLIGENCE BUT LINKING BIO- 2442 01:33:14,589 --> 01:33:16,891 MECHANICS IS THE INSPIRATION FOR 2443 01:33:16,891 --> 01:33:18,059 EMBODIED A.I. SO WITH THAT I WOULD 2444 01:33:18,059 --> 01:33:19,627 LIKE TO THANK THE PEOPLE WHO DID 2445 01:33:19,627 --> 01:33:21,729 THE WORK SHOWN IN THE TOP ROW 2446 01:33:21,729 --> 01:33:23,598 AND MY CURRENT LABORATORY SHOWN IN 2447 01:33:23,598 --> 01:33:25,433 THIS THE BOTTOM ROW PARTICULARLY 2448 01:33:25,433 --> 01:33:30,071 THANKS TO BOTH NSF AND NIH BOTH 2449 01:33:30,071 --> 01:33:31,973 OF WHICH HAVE BEEN INCREDIBLY 2450 01:33:31,973 --> 01:33:35,043 IMPORTANT FOR DIFFERENT STAGES 2451 01:33:35,043 --> 01:33:39,914 OF THIS WORK. THANK YOU. 2452 01:33:39,914 --> 01:33:41,516 >> THANK YOU TO ALL OF OUR 2453 01:33:41,516 --> 01:33:43,751 SPEAKERS AND AT THIS TIME I 2454 01:33:43,751 --> 01:33:45,887 WOULD LIKE TO INVITE ALL FOUR OF 2455 01:33:45,887 --> 01:33:56,431 OUR DISCUSSANTS TO THE STAGE. I 2456 01:34:16,384 --> 01:34:18,486 WOULD LIKE TO REMIND FOLKS WHO 2457 01:34:18,486 --> 01:34:21,823 ARE WATCHING ON THE WEBINAR, 2458 01:34:21,823 --> 01:34:25,226 ZOOM OR THE WEBCAST YOU CAN SEND 2459 01:34:25,226 --> 01:34:27,996 QUESTIONS THROUGH THE ZOOM Q&A 2460 01:34:27,996 --> 01:34:29,831 OR THE VIEW CAST LIVE FEEDBACK 2461 01:34:29,831 --> 01:34:32,901 FORM OR IF YOU SEE THE QR CODE 2462 01:34:32,901 --> 01:34:34,068 ON THE SCREEN NOW YOU CAN SCAN 2463 01:34:34,068 --> 01:34:36,271 THAT AND SEND AN E-MAIL FOR 2464 01:34:36,271 --> 01:34:37,805 DISCUSSION. LET'S KICK OFF THE 2465 01:34:37,805 --> 01:34:41,809 PANEL DISCUSSION RIGHT NOW WITH 2466 01:34:41,809 --> 01:34:42,477 CHIARA, OKAY. 2467 01:34:42,477 --> 01:34:43,912 >> CAN YOU HEAR ME? FIRST OF 2468 01:34:43,912 --> 01:34:44,946 ALL, THANKS TO ALL OF YOU FOR 2469 01:34:44,946 --> 01:34:47,248 THE AMAZING TALKS, THEY WERE ALL 2470 01:34:47,248 --> 01:34:50,618 DIFFERENT AND ALL INSPIRING. SO 2471 01:34:50,618 --> 01:34:54,956 I HAVE A QUESTION THAT IS MOSTLY 2472 01:34:54,956 --> 01:35:02,363 FOR THOSE WHO DEVELOP 2473 01:35:02,363 --> 01:35:06,134 NEUROMORPHIC CIRCUITS AND 2474 01:35:06,134 --> 01:35:08,570 HARDWARE. THAT IS, VERY OLD 2475 01:35:08,570 --> 01:35:10,939 QUESTION BUT I AM TRYING TO 2476 01:35:10,939 --> 01:35:14,175 REPHRASE IT. SO WHICH CRITERIA 2477 01:35:14,175 --> 01:35:18,980 DO YOU USE TO CHOOSE WHICH 2478 01:35:18,980 --> 01:35:20,214 NEUROCOMPUTATION OR WHICH 2479 01:35:20,214 --> 01:35:21,115 NEUROPRIMITIVES YOU WILL 2480 01:35:21,115 --> 01:35:22,450 IMPLEMENT ON YOUR HARDWARE OR ON 2481 01:35:22,450 --> 01:35:24,118 YOUR MODEL AND THIS RELATES A 2482 01:35:24,118 --> 01:35:27,322 LOT TO THE OTHER QUESTION THAT 2483 01:35:27,322 --> 01:35:29,657 IS WHICH IS THE CRITERIA TO 2484 01:35:29,657 --> 01:35:33,828 CHOOSE THE RIGHT LEVEL OF 2485 01:35:33,828 --> 01:35:34,529 ABSTRACTION 2486 01:35:34,529 --> 01:35:45,039 ABSTRACTION IN THESE MODELS? 2487 01:35:47,742 --> 01:35:51,179 IT'S NOT JUST ABOUT THE HARDWARE 2488 01:35:51,179 --> 01:35:56,718 BUT WHAT YOU WERE USING, 2489 01:35:56,718 --> 01:35:57,452 SHUNTING INHIBITION. WHY 2490 01:35:57,452 --> 01:35:59,454 SHUNTING INHIBITION, WHY NOT 2491 01:35:59,454 --> 01:36:00,655 SOMETHING DIFFERENT? 2492 01:36:00,655 --> 01:36:02,090 YOU COULD HAVE USE OTHER 2493 01:36:02,090 --> 01:36:02,657 MECHANISMS? I DON'T KNOW. 2494 01:36:02,657 --> 01:36:06,594 >> I THINK FOR ME, WHEN 2495 01:36:06,594 --> 01:36:08,196 I CHOOSE WHAT -- I'M JUST 2496 01:36:08,196 --> 01:36:12,066 GOING TO CALL IT A PRIMITIVE, 2497 01:36:12,066 --> 01:36:14,068 WHAT PRIMITIVE I WANT TO USE. 2498 01:36:14,068 --> 01:36:15,803 OH MY GOSH, SEEING MYSELF DOWN 2499 01:36:15,803 --> 01:36:17,705 THERE IS REALLY SCARY. I WAS 2500 01:36:17,705 --> 01:36:19,173 LOOKING FOR A COMPUTATION THAT I 2501 01:36:19,173 --> 01:36:20,808 COULD LINK TO THE BEHAVIOR. 2502 01:36:20,808 --> 01:36:23,378 BECAUSE I WAS INTERESTED IN 2503 01:36:23,378 --> 01:36:24,912 HOW -- HOW THE CONSTRAINTS OF 2504 01:36:24,912 --> 01:36:27,482 THE BIOLOGICAL SYSTEM AFFECTED 2505 01:36:27,482 --> 01:36:29,384 THE MECHANISM. SO FOR ME, I SEE 2506 01:36:29,384 --> 01:36:33,121 A LOT OF VALUE IN INVERTEBRATES 2507 01:36:33,121 --> 01:36:36,257 WHICH I THINK MITRA ALSO ALLUDED 2508 01:36:36,257 --> 01:36:38,526 TO. BECAUSE IN INVERTEBRATES 2509 01:36:38,526 --> 01:36:40,395 THE CIRCUIT IS SMALL SO WE CAN 2510 01:36:40,395 --> 01:36:41,529 CHARACTERIZE THE COMPLETE 2511 01:36:41,529 --> 01:36:46,200 CIRCUIT AND SEE HOW THE 2512 01:36:46,200 --> 01:36:47,035 INFORMATION PROCESSING IS LINKED 2513 01:36:47,035 --> 01:36:49,971 FROM HARDWARE ALL THE WAY TO 2514 01:36:49,971 --> 01:36:53,274 BEHAVIOR. I THINK THAT WE NEED 2515 01:36:53,274 --> 01:36:55,009 TO BE CAUTIOUS NOT MIMICKING 2516 01:36:55,009 --> 01:36:56,944 EVERY ASPECT OF THE BIOLOGICAL 2517 01:36:56,944 --> 01:36:58,413 CIRCUIT JUST BECAUSE IT'S THERE. 2518 01:36:58,413 --> 01:36:59,881 SO I THINK FOR ME IT'S IMPORTANT 2519 01:36:59,881 --> 01:37:01,949 TO LOOK ACROSS DIFFERENT 2520 01:37:01,949 --> 01:37:03,484 SYSTEMS. WE DON'T KNOW WHAT THE 2521 01:37:03,484 --> 01:37:09,323 MECHANISM IS IN THE DRAGONFLY. 2522 01:37:09,323 --> 01:37:10,892 SHUNTING INHIBITION WE HAVE SEEN 2523 01:37:10,892 --> 01:37:11,826 TO DO MULTIPLICATION IN THE DRAGONFLY. 2524 01:37:11,826 --> 01:37:14,162 AND IT WAS SOMETHING THAT WE 2525 01:37:14,162 --> 01:37:15,697 REALIZED WE COULD IMPLEMENT IN 2526 01:37:15,697 --> 01:37:16,831 AN EXISTING NEUROMORPHIC 2527 01:37:16,831 --> 01:37:18,499 APPROACH SO THAT IS HOW THAT -- 2528 01:37:18,499 --> 01:37:20,868 THOSE TWO PIECES CAME TOGETHER. 2529 01:37:20,868 --> 01:37:21,569 I THINK IT WILL BE IMPORTANT 2530 01:37:21,569 --> 01:37:23,871 GOING FORWARD TO UNDERSTAND HOW 2531 01:37:23,871 --> 01:37:26,474 THE CONSTRAINTS OF THE SYSTEM OF 2532 01:37:26,474 --> 01:37:27,408 THE BIOLOGICAL SYSTEM HAVE 2533 01:37:27,408 --> 01:37:29,577 AFFECTED THE SOLUTION THAT IS 2534 01:37:29,577 --> 01:37:31,746 IMPLEMENTED IN BIOLOGY AND 2535 01:37:31,746 --> 01:37:32,880 MAPPING THE CONSTRAINTS FACING 2536 01:37:32,880 --> 01:37:34,482 THE BIOLOGICAL SYSTEM AND THE 2537 01:37:34,482 --> 01:37:37,251 TASK TO WHAT WE WANT TO USE THE 2538 01:37:37,251 --> 01:37:39,220 NEUROMORPHIC SYSTEM FOR. SO FOR 2539 01:37:39,220 --> 01:37:42,323 EXAMPLE IN DROSOPHILA WITH THE 2540 01:37:42,323 --> 01:37:43,658 NAVIGATION TASKS THE BASIS FUNCTIONS 2541 01:37:43,658 --> 01:37:46,461 ARE DIFFERENT. IT MAKES THE SYSTEM 2542 01:37:46,461 --> 01:37:47,929 SMALLER BUT MAYBE LESS FLEXIBLE 2543 01:37:47,929 --> 01:37:49,964 THAN YOU MIGHT SEE IN MAMMALIAN 2544 01:37:49,964 --> 01:37:52,567 SYSTEM SO DEPENDING ON WHAT 2545 01:37:52,567 --> 01:37:54,469 WE WANT TO USE THE NEUROMORPHIC 2546 01:37:54,469 --> 01:37:56,037 SYSTEM FOR THAT CAN AFFECT WHICH 2547 01:37:56,037 --> 01:37:59,006 WHETHER WE LOOK TO MAMMALS OR 2548 01:37:59,006 --> 01:38:01,242 VERTEBRAE OR INVERTEBRATES FOR 2549 01:38:01,242 --> 01:38:02,910 THE SOLUTION. HOW ABOUT YOU, 2550 01:38:02,910 --> 01:38:06,447 DOES ANYONE ELSE WANT TO 2551 01:38:06,447 --> 01:38:11,652 ADD 2552 01:38:11,652 --> 01:38:11,919 ANYTHING? 2553 01:38:11,919 --> 01:38:12,787 >> I APPRECIATE THE QUESTION 2554 01:38:12,787 --> 01:38:14,555 BECAUSE IT'S A QUESTION THAT HAS 2555 01:38:14,555 --> 01:38:18,025 COME UP FOR OVER DECADES OF HOW 2556 01:38:18,025 --> 01:38:21,229 DO YOU CHOOSE OR STRATEGIZE. 2557 01:38:21,229 --> 01:38:22,864 THAT'S FAIR. I THINK FROM MY 2558 01:38:22,864 --> 01:38:25,767 PERSPECTIVE IS I SEE THIS AS A 2559 01:38:25,767 --> 01:38:27,769 CONTINUUM IN TERMS OF HOW 2560 01:38:27,769 --> 01:38:29,036 NEUROMORPHIC OR HOW SORT OF 2561 01:38:29,036 --> 01:38:31,372 ENGINEERING AND I THINK THIS 2562 01:38:31,372 --> 01:38:33,441 COMES WITH THE FACT THAT DOING 2563 01:38:33,441 --> 01:38:34,842 THIS EFFECTIVELY USUALLY MEANS 2564 01:38:34,842 --> 01:38:36,677 YOU'RE DOING SOME LEVEL OF 2565 01:38:36,677 --> 01:38:38,446 THINGS THAT ARE BOTH VERY 2566 01:38:38,446 --> 01:38:39,947 SERIOUS ENGINEERING THAT HAVE 2567 01:38:39,947 --> 01:38:41,716 ENGINEERING GOALS AND SOME 2568 01:38:41,716 --> 01:38:45,319 BECAUSE IT'S THE LOVE OF THE 2569 01:38:45,319 --> 01:38:47,421 NEUROSCIENCE AND SO WE'VE DONE 2570 01:38:47,421 --> 01:38:48,322 MANY THINGS THAT ARE VERY 2571 01:38:48,322 --> 01:38:50,458 CAREFUL ENGINEERING IC DESIGN 2572 01:38:50,458 --> 01:38:51,626 AND SO FORTH AND THERE'S ALSO 2573 01:38:51,626 --> 01:38:53,494 BEEN CASES WHERE WE LOOKED AT 2574 01:38:53,494 --> 01:38:55,096 THE NEUROSCIENCE AND SAID HOW 2575 01:38:55,096 --> 01:38:59,867 CLOSE CAN I MODEL CIRCUITS FOR 2576 01:38:59,867 --> 01:39:00,067 THIS? 2577 01:39:00,067 --> 01:39:02,537 TO BE FAIR PART OF THE REASON WE 2578 01:39:02,537 --> 01:39:05,072 DO IS A LITTLE WORD CALLED FUN. 2579 01:39:05,072 --> 01:39:07,308 WE ENJOY IT. WE DON'T TALK 2580 01:39:07,308 --> 01:39:09,644 ABOUT IT AS MUCH. BUT THAT THEN 2581 01:39:09,644 --> 01:39:11,679 YIELDS THE FRAMEWORK AND THE 2582 01:39:11,679 --> 01:39:12,380 PRIMITIVES THAT WE CAN WORK WITH 2583 01:39:12,380 --> 01:39:13,881 THAT WE SAY WHEN AN APPLICATION 2584 01:39:13,881 --> 01:39:16,751 SHOWS UP, IT VERY MUCH NATURALLY 2585 01:39:16,751 --> 01:39:18,486 FITS INTO THE CONVERSATION. AND 2586 01:39:18,486 --> 01:39:19,821 SO I THINK THAT CONTINUUM AND 2587 01:39:19,821 --> 01:39:22,223 THAT BALANCE IS IMPORTANT. 2588 01:39:22,223 --> 01:39:23,691 BECAUSE I DON'T THINK YOU CAN 2589 01:39:23,691 --> 01:39:25,793 JUST PICK ONE SPOT. YOU HAVE TO 2590 01:39:25,793 --> 01:39:28,396 KIND OF GET AN INTUITION OF THAT 2591 01:39:28,396 --> 01:39:31,332 WHOLE RANGE. OTHERWISE YOU'RE 2592 01:39:31,332 --> 01:39:32,934 USUALLY MISSING SOMETHING AND 2593 01:39:32,934 --> 01:39:34,068 MISSING AN OPPORTUNITY AND I CAN 2594 01:39:34,068 --> 01:39:36,771 PROBABLY TELL MULTIPLE STORIES 2595 01:39:36,771 --> 01:39:38,472 WHERE WE FOUND BY ACTUALLY 2596 01:39:38,472 --> 01:39:39,907 DOING -- BE DOING THAT IN TWO 2597 01:39:39,907 --> 01:39:41,709 DIFFERENT SPACES WE ACTUALLY 2598 01:39:41,709 --> 01:39:42,677 FOUND CONNECTIONS BETWEEN THEM. 2599 01:39:42,677 --> 01:39:44,278 AND THOSE WERE THE PLACES WHERE 2600 01:39:44,278 --> 01:39:46,647 WE HAD THE MOST INTERESTING 2601 01:39:46,647 --> 01:39:49,917 FEEDBACK INTO THE NEUROSCIENCE 2602 01:39:49,917 --> 01:39:54,488 QUESTIONS. BECAUSE OF THAT. 2603 01:39:54,488 --> 01:39:56,157 >> LET ME ADD TO THAT I THINK 2604 01:39:56,157 --> 01:39:58,459 EARLY ON WE WERE LOOKING AT THIS 2605 01:39:58,459 --> 01:40:00,928 PRIMITIVES HOW WOULD THEY HELP 2606 01:40:00,928 --> 01:40:01,529 US WHETHER THEY'RE NEURON 2607 01:40:01,529 --> 01:40:04,999 SYNAPSES OR DENDRITES OR ANY OF 2608 01:40:04,999 --> 01:40:06,467 THESE PRIMITIVES HOW WOULD THEY 2609 01:40:06,467 --> 01:40:08,035 HELP US GET TO AN ENERGY 2610 01:40:08,035 --> 01:40:09,203 EFFICIENT SOLUTION BUT I THINK 2611 01:40:09,203 --> 01:40:13,007 WE SHIFTED FROM THAT APPROACH 2612 01:40:13,007 --> 01:40:15,109 INTO LOOKING AT THIS GRANDER 2613 01:40:15,109 --> 01:40:17,578 PROBLEM AS A GUIDELINE FOR US 2614 01:40:17,578 --> 01:40:19,046 LIKE FOR EXAMPLE THIS CONTINUAL 2615 01:40:19,046 --> 01:40:20,348 LEARNING CHANNEL THAT WE ARE 2616 01:40:20,348 --> 01:40:23,117 LOOKING AT REQUIRES OR FORCES US 2617 01:40:23,117 --> 01:40:25,519 TO NOT JUST LOOK AT A SINGLE 2618 01:40:25,519 --> 01:40:28,356 PRIMITIVE OR AN ABSTRACTION THAT 2619 01:40:28,356 --> 01:40:29,891 WORK AT A CERTAIN LEVEL WHICH 2620 01:40:29,891 --> 01:40:31,158 YOU HAVE TO LOOK AT WHAT 2621 01:40:31,158 --> 01:40:34,929 JENNIFER WAS SAYING, THIS IS A 2622 01:40:34,929 --> 01:40:36,764 CONTINUUM. FOR EXAMPLE, YOU 2623 01:40:36,764 --> 01:40:38,733 COULD -- WE WERE THINKING LIKE 2624 01:40:38,733 --> 01:40:41,869 SYNAPSES AND NEURONS AS 2625 01:40:41,869 --> 01:40:44,005 PRIMITIVES BUT THEN WHEN YOU'RE 2626 01:40:44,005 --> 01:40:48,876 THINKING ABOUT ADDRESSING 2627 01:40:48,876 --> 01:40:51,279 CATASTROPHIC FORGETTING PROBLEM 2628 01:40:51,279 --> 01:40:52,647 YOU NEED SYNAPTOGENESIS OR PRUNING 2629 01:40:52,647 --> 01:40:54,615 IN A DYNAMIC FASHION AND YOU NEED 2630 01:40:54,615 --> 01:40:57,018 TO BE ABLE TO DYNAMICALLY 2631 01:40:57,018 --> 01:40:58,619 RECONFIGURE THESE UNITS. SO IF I 2632 01:40:58,619 --> 01:41:01,122 GO TO A LOWER LEVEL OF 2633 01:41:01,122 --> 01:41:02,023 GRANULARITY WOULD IT BE 2634 01:41:02,023 --> 01:41:03,324 BENEFICIAL OR NOT? I DO NOT 2635 01:41:03,324 --> 01:41:07,428 KNOW AT THIS POINT SO IT IS VERY 2636 01:41:07,428 --> 01:41:10,131 BENEFICIAL TO LOOK AT THIS AS A 2637 01:41:10,131 --> 01:41:12,166 PROBLEM YOU'RE TRYING TO SOLVE. 2638 01:41:12,166 --> 01:41:14,902 AND WHAT WOULD BE THE GUIDING 2639 01:41:14,902 --> 01:41:16,070 PRINCIPLES THAT NEUROSCIENCE 2640 01:41:16,070 --> 01:41:19,640 COULD OFFER AND GIVE US AN 2641 01:41:19,640 --> 01:41:27,815 AN 2642 01:41:27,815 --> 01:41:28,082 ADVANTAGE? 2643 01:41:28,082 --> 01:41:29,350 >> I THINK WHAT YOU'RE GETTING 2644 01:41:29,350 --> 01:41:31,052 AT IS REALLY IMPORTANT AND WHEN 2645 01:41:31,052 --> 01:41:33,688 YOU TALK ABOUT ABSTRACTION BOTH 2646 01:41:33,688 --> 01:41:37,558 AS A NEUROMORPHIC RESEARCH AS 2647 01:41:37,558 --> 01:41:39,527 WELL AS DOING MODELING BEFORE 2648 01:41:39,527 --> 01:41:41,562 ONCE YOU MAKE A DECISION ON 2649 01:41:41,562 --> 01:41:43,531 ABSTRACTION YOU PUT BARRIERS IS 2650 01:41:43,531 --> 01:41:45,566 HOW FAR YOU CAN GO AND I THINK 2651 01:41:45,566 --> 01:41:47,335 MITRA'S TALK WAS FANTASTIC BECAUSE 2652 01:41:47,335 --> 01:41:50,404 IT'S SHOWING THE RICHNESS OF THE 2653 01:41:50,404 --> 01:41:53,140 BEHAVIOR THAT NEURAL CIRCUITS 2654 01:41:53,140 --> 01:41:56,143 PERFORM. AND IF YOU MAKE AN 2655 01:41:56,143 --> 01:41:58,079 ABSTRACTED MODEL OF THIS SENSORY 2656 01:41:58,079 --> 01:42:00,247 CORTEX AND THE WHISKER PROCESSING 2657 01:42:00,247 --> 01:42:01,782 SYSTEM AND IT'S ON A NARROW 2658 01:42:01,782 --> 01:42:03,217 INTERPRETATION OF WHAT THE 2659 01:42:03,217 --> 01:42:05,319 WHISKERS ARE DOING, YOU ABSTRACT 2660 01:42:05,319 --> 01:42:08,322 IT AND A SIMPLE MODEL CAN DO X, 2661 01:42:08,322 --> 01:42:09,724 Y AND Z BUT YOU'RE NEVER GOING 2662 01:42:09,724 --> 01:42:12,426 TO GO FROM THAT ABSTRACTED MODEL 2663 01:42:12,426 --> 01:42:13,761 TO SOMETHING MORE RICH AND 2664 01:42:13,761 --> 01:42:14,862 COMPLEX. IT'S VERY HARD TO GO 2665 01:42:14,862 --> 01:42:17,031 BACK. I FEEL THE ABSTRACTION IS 2666 01:42:17,031 --> 01:42:18,466 ABSOLUTELY CRITICAL. 2667 01:42:18,466 --> 01:42:20,001 NEUROMORPHIC YOU PHYSICALLY HAVE 2668 01:42:20,001 --> 01:42:21,669 TO ABSTRACT BUT WITH MODELING 2669 01:42:21,669 --> 01:42:23,204 YOU MAKE AN INTENTIONAL DECISION 2670 01:42:23,204 --> 01:42:26,440 AS WELL. WE WANT TO BE 2671 01:42:26,440 --> 01:42:27,475 APPRECIATIVE OF THE FACT THAT WE 2672 01:42:27,475 --> 01:42:29,110 DON'T UNDERSTAND WHAT THE 2673 01:42:29,110 --> 01:42:30,277 COMPUTATIONS THE BRAIN IS DOING 2674 01:42:30,277 --> 01:42:32,013 MOST OF THE TIME. RIGHT? SO 2675 01:42:32,013 --> 01:42:34,648 THE BOX THAT WE'RE TRYING TO 2676 01:42:34,648 --> 01:42:36,550 CAPTURE IS KIND OF ILL-DEFINED 2677 01:42:36,550 --> 01:42:38,953 SO WE NEED TO BE -- AS YOU WERE 2678 01:42:38,953 --> 01:42:40,287 SAYING VERY SPECIFIC FOR THIS 2679 01:42:40,287 --> 01:42:41,322 TASK, FOR THIS THING WE'RE 2680 01:42:41,322 --> 01:42:43,257 TRYING TO DO, THIS IS THE 2681 01:42:43,257 --> 01:42:45,259 ABSTRACTION THAT WE HAVE TO DO. 2682 01:42:45,259 --> 01:42:47,661 BUT WE HAVE TO BE AWARE THAT IF 2683 01:42:47,661 --> 01:42:50,531 I WANT TO GO TO SOMETHING ELSE, 2684 01:42:50,531 --> 01:42:52,166 THAT ABSTRACTION MAY NOT BE 2685 01:42:52,166 --> 01:42:54,935 VALID AND I MIGHT BE MISGUIDED 2686 01:42:54,935 --> 01:42:57,371 BY USING AN ILL-FITTED 2687 01:42:57,371 --> 01:42:59,273 ABSTRACTION TO SOLVE A NEW 2688 01:42:59,273 --> 01:43:01,675 PROBLEM AND I DON'T THINK -- AT 2689 01:43:01,675 --> 01:43:05,880 LEAST IN THE COMP NEURO COMMUNITY, 2690 01:43:05,880 --> 01:43:07,481 WE DON'T TALK ABOUT IT NEARLY ENOUGH. 2691 01:43:07,481 --> 01:43:08,416 NEUROMORPHIC IS EARLY ENOUGH. IT'S 2692 01:43:08,416 --> 01:43:09,350 A GOOD THING TO TALK ABOUT NOW. 2693 01:43:09,350 --> 01:43:11,852 >> IT'S NOT ONLY IMPORTANT TO 2694 01:43:11,852 --> 01:43:15,022 CHOOSE THE RIGHT LEVEL OF 2695 01:43:15,022 --> 01:43:16,157 ABSTRACTION BUT ALSO TO CHOOSE 2696 01:43:16,157 --> 01:43:18,459 WHICH IS THE TASK IN A WAY THAT 2697 01:43:18,459 --> 01:43:22,596 IT'S NOT TOO SMALL, NOT TOO 2698 01:43:22,596 --> 01:43:24,298 LARGE AND TO FIND SOLUTIONS THAT 2699 01:43:24,298 --> 01:43:32,606 CAN THEN BE APPLIED AS WELL. 2700 01:43:32,606 --> 01:43:34,508 >> CARINA, YOUR QUESTION. 2701 01:43:34,508 --> 01:43:37,845 >> I AM GOING TO TRY TO SAY THIS 2702 01:43:37,845 --> 01:43:39,380 CONCISELY. I WAS STRUCK ACROSS 2703 01:43:39,380 --> 01:43:41,782 MANY TALKS ABOUT THE INTERPLAY 2704 01:43:41,782 --> 01:43:42,983 BETWEEN NEURAL REPRESENTATIONS, 2705 01:43:42,983 --> 01:43:46,253 FOR INSTANCE, THE MANIFOLDS OR 2706 01:43:46,253 --> 01:43:49,056 THE, YOU KNOW, SORT OF 2707 01:43:49,056 --> 01:43:51,959 REPRESENTATIONS THAT ARE MAYBE 2708 01:43:51,959 --> 01:43:58,232 PLAY WELL WITH GEOMETRIC 2709 01:43:58,232 --> 01:43:59,800 TRANSFORMATIONS AND THIS NEED TO 2710 01:43:59,800 --> 01:44:02,103 SPARSIFY THE NUMBER OF ACTIVE 2711 01:44:02,103 --> 01:44:02,369 NEURONS. I WAS STRUCK 2712 01:44:02,369 --> 01:44:03,404 BY KWABENA'S TALK ABOUT 2713 01:44:03,404 --> 01:44:06,340 HOW TO GET THAT LINEAR SCALING 2714 01:44:06,340 --> 01:44:08,008 AND SO I HAVE SORT OF A COUPLE 2715 01:44:08,008 --> 01:44:09,977 OF QUESTIONS IN THIS REGARD BUT 2716 01:44:09,977 --> 01:44:12,947 BROADLY SPEAKING HOW DO YOU SEE 2717 01:44:12,947 --> 01:44:15,649 THE, WHAT WE LEARN ABOUT NEURAL 2718 01:44:15,649 --> 01:44:18,152 REPRESENTATIONS INFLUENCING, YOU 2719 01:44:18,152 --> 01:44:19,820 KNOW, THE STYLE OF NEUROMORPHIC 2720 01:44:19,820 --> 01:44:21,388 HARDWARE THAT WE'RE COMING UP 2721 01:44:21,388 --> 01:44:24,291 WITH AND THEN SORT OF 2722 01:44:24,291 --> 01:44:25,993 SPECIFICALLY FOR SUEYEON, I WAS 2723 01:44:25,993 --> 01:44:28,395 CURIOUS IN YOUR NEURAL MANIFOLDS 2724 01:44:28,395 --> 01:44:31,632 HOW MANY NEURONS DOES A TYPICAL 2725 01:44:31,632 --> 01:44:32,766 MANIFOLD INVOLVE? IS IT HALF 2726 01:44:32,766 --> 01:44:35,603 THE NEURONS OR IS IT MUCH LESS? 2727 01:44:35,603 --> 01:44:40,141 ARE THE MANIFOLDS SOMEHOW -- 2728 01:44:40,141 --> 01:44:42,343 SOMEHOW IMPLEMENTING A KIND OF 2729 01:44:42,343 --> 01:44:44,445 SPARSE CODING, COMBINATORIAL CODING 2730 01:44:44,445 --> 01:44:44,812 EFFECTIVELY IS THE QUESTION. 2731 01:44:44,812 --> 01:44:46,213 >> THANK YOU FOR THE QUESTION. 2732 01:44:46,213 --> 01:44:48,749 I THINK, I WAS ACTUALLY GETTING 2733 01:44:48,749 --> 01:44:50,451 INSPIRED BY A LOT OF 2734 01:44:50,451 --> 01:44:52,086 NEUROMORPHIC TALKS AND THINKING 2735 01:44:52,086 --> 01:44:53,420 ABOUT THE SIMILAR TYPE OF 2736 01:44:53,420 --> 01:44:55,489 QUESTIONS. TO ANSWER THE FIRST 2737 01:44:55,489 --> 01:44:56,690 QUESTION ABOUT THE CONNECTION 2738 01:44:56,690 --> 01:45:00,528 BETWEEN THE EFFICIENCY OF 2739 01:45:00,528 --> 01:45:01,262 REPRESENTATIONS AND NEUROMORPHIC 2740 01:45:01,262 --> 01:45:04,298 COMPUTATION SO I THINK THERE'S 2741 01:45:04,298 --> 01:45:05,466 SEVERAL INTERESTING DIRECTIONS 2742 01:45:05,466 --> 01:45:07,935 AT THE CONVERGENCE OF THESE 2743 01:45:07,935 --> 01:45:08,869 IDEAS. FOR EXAMPLE, WHAT COULD 2744 01:45:08,869 --> 01:45:13,240 BE TYPES OF REPRESENTATIONS THAT 2745 01:45:13,240 --> 01:45:15,309 EMERGE FROM COMPUTATIONS THAT 2746 01:45:15,309 --> 01:45:18,012 ARE IMPLEMENTED BY NEUROMORPHIC 2747 01:45:18,012 --> 01:45:21,682 COMPUTING. SO IF THEY PROVIDE 2748 01:45:21,682 --> 01:45:26,587 SOME SORT OF BIAS OR PREFERENCE 2749 01:45:26,587 --> 01:45:30,024 WHAT IS THE RESULT OF THOSE AND 2750 01:45:30,024 --> 01:45:33,027 THOSE QUESTIONS ARE INTERESTING AND 2751 01:45:33,027 --> 01:45:34,695 NOT ASKED AS MUCH. AND THE OTHER 2752 01:45:34,695 --> 01:45:37,698 CONNECTION IS, SO THE -- THE 2753 01:45:37,698 --> 01:45:39,633 IDEA OF MANIFOLD CAPACITY 2754 01:45:39,633 --> 01:45:41,635 DEFINITELY HAS TO DO WITH THE 2755 01:45:41,635 --> 01:45:43,037 REPRESENTATIONAL EFFICIENCY 2756 01:45:43,037 --> 01:45:45,306 MEANING THAT IT GIVES SOME SENSE 2757 01:45:45,306 --> 01:45:48,175 OF HOW MANY NEURONS OR UNITS ARE 2758 01:45:48,175 --> 01:45:49,009 REQUIRED TO REPRESENT THE GIVEN 2759 01:45:49,009 --> 01:45:51,645 NUMBER OF CONCEPTS WITH 2760 01:45:51,645 --> 01:45:53,514 CERTAIN GEOMETRY THAT'S MEASURED 2761 01:45:53,514 --> 01:45:55,316 FROM THE DATA AND AS WE SAW IN 2762 01:45:55,316 --> 01:45:57,785 EARLIER TALKS, THE NUMBER OF 2763 01:45:57,785 --> 01:46:00,221 UNITS, THE NUMBER OF NEURONS 2764 01:46:00,221 --> 01:46:02,790 REQUIRED HAS TO ALSO DO WITH 2765 01:46:02,790 --> 01:46:04,091 ENERGY EFFICIENCY OR ENERGY 2766 01:46:04,091 --> 01:46:08,229 THAT'S NEEDED SO I THINK THERE'S 2767 01:46:08,229 --> 01:46:09,663 ALSO DIRECT CONNECTION TO ENERGY 2768 01:46:09,663 --> 01:46:11,131 EFFICIENCY FROM THE 2769 01:46:11,131 --> 01:46:14,134 REPRESENTATIONAL EFFICIENCY AND 2770 01:46:14,134 --> 01:46:16,770 WHAT'S THE SECOND QUESTION 2771 01:46:16,770 --> 01:46:17,371 WAS.... 2772 01:46:17,371 --> 01:46:18,639 >> I WAS JUST CURIOUS, THE 2773 01:46:18,639 --> 01:46:21,275 NEURON -- THERE'S A LOT OF ECHO. 2774 01:46:21,275 --> 01:46:23,344 I WAS CURIOUS OF THE NEURAL 2775 01:46:23,344 --> 01:46:25,312 MANIFOLDS THAT YOU SEE THAT YOU 2776 01:46:25,312 --> 01:46:27,648 OBSERVE WHEN YOU APPLY YOUR 2777 01:46:27,648 --> 01:46:32,119 FRAMEWORK TO DATA IF THEY TEND 2778 01:46:32,119 --> 01:46:36,890 TO BE SUPPORTED ON FEW NEURONS 2779 01:46:36,890 --> 01:46:41,829 IS THAT A KIND OF COMBINATORIAL, SPARSE? 2780 01:46:41,829 --> 01:46:46,467 >> IT ACTUALLY DEPENDS -- IT'S 2781 01:46:46,467 --> 01:46:48,802 KIND OF HARD TO ANSWER WHAT ARE 2782 01:46:48,802 --> 01:46:50,671 THE TYPICAL NEURONS THAT IS 2783 01:46:50,671 --> 01:46:51,772 REQUIRED FOR COMPUTATION BECAUSE 2784 01:46:51,772 --> 01:46:53,641 IT IS A FUNCTION OF THE GEOMETRY 2785 01:46:53,641 --> 01:46:57,111 OF THE DATA THAT WE SEE. AND 2786 01:46:57,111 --> 01:46:58,445 IT -- PROBABLY ALSO -- ACTUALLY 2787 01:46:58,445 --> 01:47:00,180 IT DEPENDS ON THE REGION IN THE 2788 01:47:00,180 --> 01:47:01,448 BRAIN THAT YOU'RE LOOKING AT. 2789 01:47:01,448 --> 01:47:03,350 IF YOU'RE LOOKING AT EARLIER 2790 01:47:03,350 --> 01:47:05,419 SENSORY REGION WHERE IS THE 2791 01:47:05,419 --> 01:47:06,420 INFORMATION HAS NOT BEEN 2792 01:47:06,420 --> 01:47:08,922 TRANSFORMED AS MUCH THEN THE 2793 01:47:08,922 --> 01:47:10,257 REPRESENTATIONS TEND TO BE 2794 01:47:10,257 --> 01:47:11,592 INEFFICIENT SO IN ORDER TO 2795 01:47:11,592 --> 01:47:12,526 REPRESENT SAME NUMBER OF 2796 01:47:12,526 --> 01:47:13,527 CATEGORIES YOU WOULD NEED A LOT 2797 01:47:13,527 --> 01:47:15,829 OF NEURONS WHILE IF YOU GO ALL 2798 01:47:15,829 --> 01:47:19,266 THE WAY DOWN TO I.T. OR THE 2799 01:47:19,266 --> 01:47:20,668 PREFRONTAL CORTEX THE 2800 01:47:20,668 --> 01:47:22,469 REPRESENTATIONS TEND TO BE MORE 2801 01:47:22,469 --> 01:47:23,904 ABSTRACTED WHICH WE KNOW FROM 2802 01:47:23,904 --> 01:47:26,440 THE GEOMETRY OF MANIFOLDS BEING 2803 01:47:26,440 --> 01:47:27,941 SMALLER AND SO ON THEN THE 2804 01:47:27,941 --> 01:47:31,078 MEASURED CAPACITY TENDS TO BE 2805 01:47:31,078 --> 01:47:34,982 HIGHER WHICH MEANS THAT YOU 2806 01:47:34,982 --> 01:47:37,151 WOULD NEED SMALLER NUMBER OF NEURONS 2807 01:47:37,151 --> 01:47:38,619 TO REPRESENT THE SAME NUMBER OF 2808 01:47:38,619 --> 01:47:41,355 CATEGORIES. THIS REALLY IS A 2809 01:47:41,355 --> 01:47:42,122 FUNCTION OF HOW MANY - THE ACTUAL 2810 01:47:42,122 --> 01:47:44,391 STRUCTURE IN THE DATA. 2811 01:47:44,391 --> 01:47:50,831 >> YIOTA, YOUR QUESTION, THANK YOU. 2812 01:47:50,831 --> 01:47:53,000 >> THANK YOU ALL FOR A WONDERFUL 2813 01:47:53,000 --> 01:47:54,268 SET OF PRESENTATIONS. I 2814 01:47:54,268 --> 01:47:56,370 APPRECIATE THAT WE WENT FROM 2815 01:47:56,370 --> 01:47:57,738 SUBCELLULAR MECHANISMS WITH 2816 01:47:57,738 --> 01:48:00,607 JENNIFER TO THE PRIMITIVES TO 2817 01:48:00,607 --> 01:48:01,342 MANIFOLDS TO CONTINUAL LEARNING 2818 01:48:01,342 --> 01:48:03,477 ALL THE WAY TO EMBODIMENT THAT 2819 01:48:03,477 --> 01:48:05,913 WAS REALLY FASCINATING FOR ME TO 2820 01:48:05,913 --> 01:48:07,781 SEE. AND I AM WONDERING 2821 01:48:07,781 --> 01:48:10,651 FOLLOWING UP WITH CARINA'S 2822 01:48:10,651 --> 01:48:12,853 QUESTIONS HOW YOU SELECT THE 2823 01:48:12,853 --> 01:48:16,090 RIGHT LEVEL OF ABSTRACTION TO 2824 01:48:16,090 --> 01:48:18,459 ADDRESS IN HARDWARE DOES THE 2825 01:48:18,459 --> 01:48:19,893 MATERIAL PLAY A ROLE? WE 2826 01:48:19,893 --> 01:48:21,261 HAVEN'T REALLY TALKED ABOUT WHAT 2827 01:48:21,261 --> 01:48:25,499 KIND OF HARDWARE DEVICES ARE YOU 2828 01:48:25,499 --> 01:48:27,401 USING AND WHETHER THE PLUSES AND 2829 01:48:27,401 --> 01:48:29,970 MINUSES OF THE MATERIAL DICTATE 2830 01:48:29,970 --> 01:48:32,339 ALSO THE LEVEL OF ABSTRACTION 2831 01:48:32,339 --> 01:48:34,508 THAT IS APPROPRIATE. THAT'S ONE 2832 01:48:34,508 --> 01:48:40,414 QUESTION AND THE SECOND ONE IS 2833 01:48:40,414 --> 01:48:42,383 WHAT DO YOU FEEL THAT YOU ARE 2834 01:48:42,383 --> 01:48:43,517 MISSING FROM THIS INTERACTION 2835 01:48:43,517 --> 01:48:45,719 THAT WOULD FACILITATE THE 2836 01:48:45,719 --> 01:48:47,621 DISCOVERY PROCESS FURTHER I CAN 2837 01:48:47,621 --> 01:48:49,289 SAY FOR SURE I DON'T KNOW THE 2838 01:48:49,289 --> 01:48:50,424 LIMITATIONS OF THE MATERIAL FOR 2839 01:48:50,424 --> 01:48:52,626 EXAMPLE, THAT'S MY FIRST 2840 01:48:52,626 --> 01:48:53,560 QUESTION, BUT IT COULD BE 2841 01:48:53,560 --> 01:48:54,695 SOMETHING ELSE. THOSE ARE THE 2842 01:48:54,695 --> 01:49:05,239 TWO THINGS I WOULD LIKE TO KNOW 2843 01:49:10,978 --> 01:49:13,680 >> I CAN TRY TO ANSWER BOTH 2844 01:49:13,680 --> 01:49:14,681 QUESTIONS. MATERIAL SCIENCE HAS 2845 01:49:14,681 --> 01:49:16,350 BEEN AN IMPORTANT PART OF OUR 2846 01:49:16,350 --> 01:49:18,452 WORK AND I'VE BEEN FORTUNATE TO 2847 01:49:18,452 --> 01:49:22,556 WORK WITH SCIENTISTS WHO HELP 2848 01:49:22,556 --> 01:49:24,558 DESIGN FOR US SOME ARTIFICIAL 2849 01:49:24,558 --> 01:49:28,529 WHISKERS THAT WERE DYNAMICALLY 2850 01:49:28,529 --> 01:49:29,997 SIMILAR TO REAL WHISKERS SO THAT WAS 2851 01:49:29,997 --> 01:49:30,831 AN IMPORTANT PART OF OUR WORK AND 2852 01:49:30,831 --> 01:49:33,267 TO THE LEVEL OF ABSTRACTION I 2853 01:49:33,267 --> 01:49:34,601 WAS STRUCK BY SOMETHING THAT 2854 01:49:34,601 --> 01:49:36,770 FRANCES SAID EARLIER THAT YOU 2855 01:49:36,770 --> 01:49:39,706 DON'T WANT TO REPLICATE ALL THE 2856 01:49:39,706 --> 01:49:41,542 BIOLOGICAL DETAIL AND I AGREE IN 2857 01:49:41,542 --> 01:49:44,278 MANY RESPECTS ALTHOUGH WE HAVE 2858 01:49:44,278 --> 01:49:45,646 PUT IN QUITE A LOT INTO OUR 2859 01:49:45,646 --> 01:49:47,648 MODELS I THINK THE QUESTION IS, 2860 01:49:47,648 --> 01:49:49,183 AGAIN, WHAT SORT OF TASK OR 2861 01:49:49,183 --> 01:49:50,984 BEHAVIOR ARE YOU REALLY 2862 01:49:50,984 --> 01:49:52,953 INTERESTED IN? GOING BACK TO 2863 01:49:52,953 --> 01:49:55,055 THE PICTURE OF THE CAT AND THE 2864 01:49:55,055 --> 01:49:57,090 RABBIT. ARE YOU INTERESTED IN 2865 01:49:57,090 --> 01:50:01,161 HOW THEY HOP AND HOW THEY -- 2866 01:50:01,161 --> 01:50:02,496 STALK? OR ARE YOU INTERESTED IN 2867 01:50:02,496 --> 01:50:04,965 HOW THEY LEARN TO DO THAT? AND 2868 01:50:04,965 --> 01:50:06,633 IF YOU'RE INTERESTED IN ONE OF 2869 01:50:06,633 --> 01:50:09,670 THEM YOU NEED MORE BIOLOGICAL 2870 01:50:09,670 --> 01:50:11,104 DETAIL THAN THE OTHER AND WITH 2871 01:50:11,104 --> 01:50:12,272 THIS AUDIENCE IF YOU'RE 2872 01:50:12,272 --> 01:50:13,740 INTERESTED IN THE MORE GENERAL 2873 01:50:13,740 --> 01:50:15,809 ALGORITHMS LIKE HOW DO ANIMALS 2874 01:50:15,809 --> 01:50:17,044 LEARN WHAT I THINK MIGHT BE 2875 01:50:17,044 --> 01:50:22,049 HELPFUL IS A FEW REALLY GOOD 2876 01:50:22,049 --> 01:50:24,351 SYSTEMS THAT ARE IN REAL WORLD 2877 01:50:24,351 --> 01:50:25,219 INSTANTIATED SYSTEMS WHERE YOU 2878 01:50:25,219 --> 01:50:27,287 CAN TEST OUT THESE GENERAL 2879 01:50:27,287 --> 01:50:28,489 ALGORITHMS ON THESE PARTICULAR 2880 01:50:28,489 --> 01:50:30,457 SPECIES AND BEHAVIORS. SO 2881 01:50:30,457 --> 01:50:31,058 THAT'S HOW I SEE THE TWO 2882 01:50:31,058 --> 01:50:41,235 INTERACTING. 2883 01:50:41,235 --> 01:50:44,037 >> I MIGHT TAKE THE BAIT ON THE 2884 01:50:44,037 --> 01:50:45,839 NEUROMORPHIC HARDWARE MATERIALS. 2885 01:50:45,839 --> 01:50:47,374 SO THE NEUROMORPHIC FIELD HAS 2886 01:50:47,374 --> 01:50:49,343 LOOKED A LOT AT DIFFERENT 2887 01:50:49,343 --> 01:50:53,914 MATERIALS FOR POST-MOORE'S LAW, 2888 01:50:53,914 --> 01:50:57,918 NON-CMOS, AND MY OPINION IS I 2889 01:50:57,918 --> 01:50:58,785 THINK WE SHOULD LOOK AT 2890 01:50:58,785 --> 01:51:00,354 EVERYTHING. RIGHT? I MEAN, ON 2891 01:51:00,354 --> 01:51:02,155 THE MATERIAL SIDE, RIGHT, FOR 2892 01:51:02,155 --> 01:51:04,691 SCALING PURPOSES, FOR, YOU KNOW, 2893 01:51:04,691 --> 01:51:07,928 FUTURE OF COMPUTING AND SO FORTH 2894 01:51:07,928 --> 01:51:10,430 I THINK THAT AT SOME LEVEL THE 2895 01:51:10,430 --> 01:51:11,965 VISION AS I UNDERSTAND IT AND 2896 01:51:11,965 --> 01:51:15,435 YOU ALL LIVED IT MORE OF CARVER 2897 01:51:15,435 --> 01:51:18,572 MEAD IS THAT, YOU KNOW, BRAINS 2898 01:51:18,572 --> 01:51:22,075 ARE MADE OF CARBON, FATTY 2899 01:51:22,075 --> 01:51:23,744 MEMBRANES AND WATER AND SO 2900 01:51:23,744 --> 01:51:24,711 FORTH, RIGHT? 2901 01:51:24,711 --> 01:51:26,480 BUT THE COMPUTATION CAN BE 2902 01:51:26,480 --> 01:51:28,916 REALIZED IN SILICON, THAT 2903 01:51:28,916 --> 01:51:32,553 COMPUTATION CAN BE REALIZED IN 2904 01:51:32,553 --> 01:51:35,756 TANTALUM OXIDE MEMRISTORS, IN CARBON 2905 01:51:35,756 --> 01:51:37,057 NANOTUBE, IT CAN BE REALIZED IN ANY 2906 01:51:37,057 --> 01:51:38,692 NUMBER OF THINGS. IT'S GOING TO 2907 01:51:38,692 --> 01:51:40,193 MAYBE CHANGE THE CIRCUIT, MAYBE 2908 01:51:40,193 --> 01:51:41,895 CHANGE THE DEVICE PROPERTIES BUT 2909 01:51:41,895 --> 01:51:44,565 AT SOME LEVEL THE MATH, THE BIO- 2910 01:51:44,565 --> 01:51:45,766 PHYSICS BECOMES SOME MATERIAL 2911 01:51:45,766 --> 01:51:48,335 PHYSICS IN WHATEVER SUBSTRATE WE 2912 01:51:48,335 --> 01:51:49,736 USE AND BY MAKING THAT MAPPING 2913 01:51:49,736 --> 01:51:52,139 DOWN TO THAT -- WHETHER IT'S THE 2914 01:51:52,139 --> 01:51:55,309 NEURAL DYNAMICS, STOCHASTIC SYNAPSE, 2915 01:51:55,309 --> 01:52:01,348 OR THE NOISY WHATEVER IT IS THE 2916 01:52:01,348 --> 01:52:03,650 PROCESS OF THAT MAPPING THAT YOU 2917 01:52:03,650 --> 01:52:05,519 UNDERSTAND WHAT THAT PHYSICS IS 2918 01:52:05,519 --> 01:52:07,154 GIVING YOU. BUT I DON'T THINK 2919 01:52:07,154 --> 01:52:09,556 THE MATERIAL MATTERS TOO MUCH. 2920 01:52:09,556 --> 01:52:11,325 THIS MAY BE CONTROVERSIAL IT 2921 01:52:11,325 --> 01:52:13,360 MATTERS FROM A SCALING POINT OF 2922 01:52:13,360 --> 01:52:15,395 VIEW BUT FOR OUR PURPOSES HERE I 2923 01:52:15,395 --> 01:52:19,600 DON'T THINK IT MATTERS TOO MUCH. 2924 01:52:19,600 --> 01:52:22,469 CMOS IS GREAT BECAUSE WE CAN 2925 01:52:22,469 --> 01:52:24,504 BUILD IT REALLY EFFICIENTLY. 2926 01:52:24,504 --> 01:52:26,173 THAT'S WHY WE'RE FOCUSED ON IT A 2927 01:52:26,173 --> 01:52:27,941 LOT ON THIS STAGE BUT IN THE 2928 01:52:27,941 --> 01:52:29,276 FUTURE IT MAY VERY WELL CHANGE 2929 01:52:29,276 --> 01:52:31,111 TEN YEARS FROM NOW WE'LL 2930 01:52:31,111 --> 01:52:33,547 PROBABLY BE USING ALL SORTS OF 2931 01:52:33,547 --> 01:52:34,448 MATERIALS THAT WE DON'T USE 2932 01:52:34,448 --> 01:52:38,919 TODAY AND I DON'T THINK IT WILL 2933 01:52:38,919 --> 01:52:40,253 CHANGE MUCH WHAT THE QUESTIONS 2934 01:52:40,253 --> 01:52:41,088 ARE BUT I MIGHT HAVE A 2935 01:52:41,088 --> 01:52:48,028 DISAGREEMENT OVER HERE. 2936 01:52:48,028 --> 01:52:56,136 >> WE HAVE ONE MORE DISCUS ANT 2937 01:52:56,136 --> 01:52:56,403 QUESTION. 2938 01:52:56,403 --> 01:53:02,342 >> IS THERE -- YOU WANT TO -- IF 2939 01:53:02,342 --> 01:53:03,844 YOU HAVE -- LET'S FINISH QUICKLY 2940 01:53:03,844 --> 01:53:04,511 OR -- 2941 01:53:04,511 --> 01:53:09,916 >> YEAH, SORRY. NO, I JUST 2942 01:53:09,916 --> 01:53:11,785 THINK JUST A COUPLE COMMENTS 2943 01:53:11,785 --> 01:53:13,553 THERE. YEAH, THE -- I DO THINK 2944 01:53:13,553 --> 01:53:14,788 THERE'S A LOT OF INTERESTING 2945 01:53:14,788 --> 01:53:19,259 THINGS WITH THE MATERIALS, LIKE 2946 01:53:19,259 --> 01:53:19,860 YOUR EXAMPLE MITRA. 2947 01:53:19,860 --> 01:53:21,962 I DO THINK THAT HOW WE IMPLEMENT 2948 01:53:21,962 --> 01:53:24,598 IS PROBABLY NOT TOO CRITICAL. I 2949 01:53:24,598 --> 01:53:25,799 WOULD AGREE WITH THAT STATEMENT 2950 01:53:25,799 --> 01:53:28,035 I DO THINK ON THE FLIP SIDE THE 2951 01:53:28,035 --> 01:53:29,236 SILICON SIDE IS PROBABLY OFTEN 2952 01:53:29,236 --> 01:53:31,304 BEEN UNDERLOOKED IN THE LAST 2953 01:53:31,304 --> 01:53:33,273 DECADE AND A HALF. AND I THINK 2954 01:53:33,273 --> 01:53:37,311 THAT IS A VERY SERIOUS PROBLEM. 2955 01:53:37,311 --> 01:53:38,912 AND PROBABLY HAS -- IS PROBABLY 2956 01:53:38,912 --> 01:53:39,913 SOMETHING WE DON'T WANT TO DO 2957 01:53:39,913 --> 01:53:42,449 BECAUSE I THINK THERE'S A LOT 2958 01:53:42,449 --> 01:53:43,717 THAT ACTUALLY FITS REALLY WELL 2959 01:53:43,717 --> 01:53:45,519 THERE. I THINK IT'S ALSO 2960 01:53:45,519 --> 01:53:46,820 INTERESTING BECAUSE THE SILICON 2961 01:53:46,820 --> 01:53:48,889 SIDE STILL, THE CAPABILITY THAT 2962 01:53:48,889 --> 01:53:50,757 YOU CAN GET OUT OF THE SILICON 2963 01:53:50,757 --> 01:53:52,559 SIDE IS STILL THE LEADING 2964 01:53:52,559 --> 01:53:53,694 IMPLEMENTATION BY FAR OF 2965 01:53:53,694 --> 01:53:55,762 ANYTHING ELSE THAT'S BEEN DONE. 2966 01:53:55,762 --> 01:53:57,230 SO I THINK IF WE WANT TO MOVE 2967 01:53:57,230 --> 01:53:59,232 FORWARD I THINK THAT'S GOING TO 2968 01:53:59,232 --> 01:54:01,435 BE -- THAT NEEDS TO BE IN OUR 2969 01:54:01,435 --> 01:54:04,671 CONVERSATION AND IN OUR 2970 01:54:04,671 --> 01:54:07,541 APPROACH. 2971 01:54:07,541 --> 01:54:10,777 >> OKAY. SO WHAT IS THE MINIMUM 2972 01:54:10,777 --> 01:54:13,313 SET OF COMPONENT THAT YOU 2973 01:54:13,313 --> 01:54:17,484 ABSOLUTELY HAVE TO HAVE IN YOUR 2974 01:54:17,484 --> 01:54:20,353 COMPUTING? SO NEURON OR SYNAPSE 2975 01:54:20,353 --> 01:54:26,059 OR LEARNING OR DENDRITE OR NOT? 2976 01:54:26,059 --> 01:54:28,929 >> PASSIVE DYNAMIC WALKERS. YOU 2977 01:54:28,929 --> 01:54:33,600 DON'T NEED ANY BRAIN AT ALL. 2978 01:54:33,600 --> 01:54:36,336 >> I WILL TRY AND TAKE THAT 2979 01:54:36,336 --> 01:54:39,740 QUESTION. I MEAN, IF YOU LOOK 2980 01:54:39,740 --> 01:54:41,742 ACROSS, RIGHT NOW WE HAVE THE 2981 01:54:41,742 --> 01:54:43,844 FLY CONNECTOME, SOMEBODY 2982 01:54:43,844 --> 01:54:45,912 MENTIONED THEY RECONSTRUCTED A 2983 01:54:45,912 --> 01:54:48,749 SYNAPSE EVERY NEURON, IN FACT 2984 01:54:48,749 --> 01:54:50,717 IT'S NANOMETER RESOLUTION, WE 2985 01:54:50,717 --> 01:54:54,454 GOT LIKE A CUBIC MILLIMETER OF 2986 01:54:54,454 --> 01:54:57,290 MOUSE AND OF HUMAN CONNECTOME AND THE 2987 01:54:57,290 --> 01:54:59,025 THING THAT STRUCK ME IN ANSWERING 2988 01:54:59,025 --> 01:55:00,894 YOUR QUESTION IS, THE FLY CONNECTOME 2989 01:55:00,894 --> 01:55:03,196 IS VERY DIFFERENT FROM THE HUMAN 2990 01:55:03,196 --> 01:55:04,898 CONNECTOME. A LOT OF US, YOU 2991 01:55:04,898 --> 01:55:08,335 KNOW, ME, I HAD NO IDEA THAT 2992 01:55:08,335 --> 01:55:10,604 FLIES WERE THIS DIFFERENT. MOST 2993 01:55:10,604 --> 01:55:12,272 OF THE NEURONS IN A FLY ARE NOT 2994 01:55:12,272 --> 01:55:14,241 POLARIZED SO THEY DON'T HAVE A 2995 01:55:14,241 --> 01:55:16,276 DENDRITE. THEY DON'T HAVE AN 2996 01:55:16,276 --> 01:55:17,511 AXON THEY HAVE SOMETHING CALLED 2997 01:55:17,511 --> 01:55:19,346 A NEURITE. IN OTHER WORDS YOU 2998 01:55:19,346 --> 01:55:20,313 CAN'T SEE WHICH DIRECTION THE 2999 01:55:20,313 --> 01:55:21,915 INFORMATION IS FLOWING. THE 3000 01:55:21,915 --> 01:55:24,284 MOST TYPICAL THING YOU FIND IN A 3001 01:55:24,284 --> 01:55:26,419 FLY IS YOU HAVE ALL TEN MICRONS 3002 01:55:26,419 --> 01:55:28,488 OF NEURITE THAT IS RECEIVING 3003 01:55:28,488 --> 01:55:32,192 FIVE TO TEN SYNAPSES, INPUT 3004 01:55:32,192 --> 01:55:34,261 SYNAPSES AND IT'S TURNING AROUND 3005 01:55:34,261 --> 01:55:36,863 AND MAKING AN OUTPUT SYNAPSE FROM 3006 01:55:36,863 --> 01:55:38,431 THAT SAME 10 MICRONS OF NEURON. 3007 01:55:38,431 --> 01:55:42,135 IT'S VERY CLEAR IN A FLY THE ATOM 3008 01:55:42,135 --> 01:55:44,371 OF COMPUTATION IS TEN MICRONS OF 3009 01:55:44,371 --> 01:55:46,740 NEURITE. INPUTS AND OUTPUTS ARE ALL THERE. 3010 01:55:46,740 --> 01:55:50,443 AND SO, WHY IS A FLY DOING THAT? 3011 01:55:50,443 --> 01:55:52,312 HOW HAVE WE EVOLVED TO DO THIS 3012 01:55:52,312 --> 01:55:53,980 POLARIZATION AND IS IT STILL THE 3013 01:55:53,980 --> 01:55:57,117 CASE THAT THESE TEN MICRONS 3014 01:55:57,117 --> 01:55:59,419 OF DENDRITE OR STILL THE ATOM OF 3015 01:55:59,419 --> 01:56:00,453 COMPUTATION BUT WE ORGANIZE THEM 3016 01:56:00,453 --> 01:56:03,723 IN A DIFFERENT WAY. AND OUR 3017 01:56:03,723 --> 01:56:05,525 COMMUNICATION BECOMES LONGER SO 3018 01:56:05,525 --> 01:56:07,294 WE HAVE SPECIALIZED AXONS TO DO 3019 01:56:07,294 --> 01:56:09,462 THAT AND SO ON AND SO FORTH. 3020 01:56:09,462 --> 01:56:12,132 BUT, YOU KNOW, SO THE QUESTION, 3021 01:56:12,132 --> 01:56:14,134 YOU KNOW, SOMEBODY MENTIONED 3022 01:56:14,134 --> 01:56:15,702 THIS, LOOKING ACROSS DIFFERENT 3023 01:56:15,702 --> 01:56:18,104 SYSTEMS AND DIFFERENT SOLUTIONS, 3024 01:56:18,104 --> 01:56:19,639 YOU KNOW, I THINK THAT THE 3025 01:56:19,639 --> 01:56:23,076 ANSWER TO THAT SHOULD BE COMMON 3026 01:56:23,076 --> 01:56:24,511 THE FLY AND THE HUMAN AND ALL 3027 01:56:24,511 --> 01:56:25,712 THIS STUFF. SO THAT'S TELLING 3028 01:56:25,712 --> 01:56:28,281 YOU SOMETHING. BRAD WHEN YOU SAY 3029 01:56:28,281 --> 01:56:30,851 YOU SIMULATED THE WHOLE FLY I 3030 01:56:30,851 --> 01:56:32,652 CAN'T IMAGINE HOW YOU CAN DO 3031 01:56:32,652 --> 01:56:34,254 THAT BECAUSE THERE'S NO SUCH 3032 01:56:34,254 --> 01:56:37,290 THING AS NEURONS WITH AXONS AND 3033 01:56:37,290 --> 01:56:41,862 DENDRITES IN A FLY. 3034 01:56:41,862 --> 01:56:44,497 THE ONLY THING I WILL SAY IS WE 3035 01:56:44,497 --> 01:56:47,234 HAVE TO BE HUMBLE ABOUT THE 3036 01:56:47,234 --> 01:56:48,568 ANSWERS TO THESE QUESTIONS AND 3037 01:56:48,568 --> 01:56:50,070 NOW WE HAVE DATA, IN FACT THE 3038 01:56:50,070 --> 01:56:53,106 ANSWER TO THIS QUESTION IS 3039 01:56:53,106 --> 01:56:54,541 SITTING ON SOMEBODY'S DESK. THERE'S 3040 01:56:54,541 --> 01:56:57,377 A PETABYTE DATASET THERE IF WE 3041 01:56:57,377 --> 01:57:00,280 EXAMINE IT CLOSELY WE REALLY 3042 01:57:00,280 --> 01:57:01,047 WILL BE ABLE TO ANSWER THESE 3043 01:57:01,047 --> 01:57:02,182 QUESTIONS AND COMPARE THEM. 3044 01:57:02,182 --> 01:57:03,817 THAT'S WHY I'VE BEEN QUIET. I 3045 01:57:03,817 --> 01:57:06,786 REALLY DON'T HAVE -- I HAVEN'T 3046 01:57:06,786 --> 01:57:08,488 EXAMINED THAT PETABYTE OF 3047 01:57:08,488 --> 01:57:10,423 DATA BUT THERE'S A GREAT 3048 01:57:10,423 --> 01:57:11,825 OPPORTUNITY HERE TO BE VERY 3049 01:57:11,825 --> 01:57:13,326 OPEN-MINDED ABOUT THIS, TO BE 3050 01:57:13,326 --> 01:57:14,527 VERY COMPARATIVE AND TO REALLY 3051 01:57:14,527 --> 01:57:16,463 GO FOR WHAT GIVES YOU THE 3052 01:57:16,463 --> 01:57:18,465 BIGGEST BANG FOR YOUR BUCK AND I 3053 01:57:18,465 --> 01:57:21,434 THINK THAT'S CHANGING THE WAY OF 3054 01:57:21,434 --> 01:57:23,970 THE EXPONENT WITH WHICH THINGS SCALE. 3055 01:57:23,970 --> 01:57:24,971 SO OVERALL, IF YOU BELIEVE THAT, 3056 01:57:24,971 --> 01:57:28,475 NEUROMORPHIC COMPUTING HAS SOMETHING 3057 01:57:28,475 --> 01:57:31,511 TO OFFER, THAT IS FUNDAMENTALLY 3058 01:57:31,511 --> 01:57:33,213 SUPERIOR TO THE KIND OF 3059 01:57:33,213 --> 01:57:34,648 COMPUTING WE'RE DOING NOW IT'S 3060 01:57:34,648 --> 01:57:37,417 LIKE GOING FROM CLASSICAL 3061 01:57:37,417 --> 01:57:39,352 COMPUTERS TO QUANTUM COMPUTERS 3062 01:57:39,352 --> 01:57:42,889 THAT AMOUNT OF CHANGING THE EXPONENT. 3063 01:57:42,889 --> 01:57:44,524 THE RIGHT INGREDIENTS WILL GIVE YOU 3064 01:57:44,524 --> 01:57:45,825 THAT. THE WRONG INGREDIENTS YOU 3065 01:57:45,825 --> 01:57:46,726 CAN DO WHATEVER YOU WANT YOU'LL 3066 01:57:46,726 --> 01:57:51,531 NEVER GET THAT. SO. YEAH. 3067 01:57:51,531 --> 01:57:55,302 >> ARE WE READY TO OPEN UP TO 3068 01:57:55,302 --> 01:57:57,604 AUDIENCE QUESTIONS? I WILL ASK 3069 01:57:57,604 --> 01:57:59,572 PEOPLE TO THE IN ROOM 3070 01:57:59,572 --> 01:58:00,907 MICROPHONES BUT I WOULD QUICKLY 3071 01:58:00,907 --> 01:58:02,776 LIKE TO BRING IN A QUESTION OVER 3072 01:58:02,776 --> 01:58:05,011 ZOOM ABOUT NEUROMORPHIC HARDWARE 3073 01:58:05,011 --> 01:58:07,580 VERSUS POTENTIALLY NEUROMORPHIC 3074 01:58:07,580 --> 01:58:11,017 NEUROAI SOFTWARE SO THIS QUESTION 3075 01:58:11,017 --> 01:58:12,252 MENTIIONS 3076 01:58:12,252 --> 01:58:14,854 HYPERDIMENSIONAL COMPUTING 3077 01:58:14,854 --> 01:58:15,989 THAT CAN POTENTIALLY 3078 01:58:15,989 --> 01:58:18,258 COMPLEMENT HARDWARE SOLUTIONS, 3079 01:58:18,258 --> 01:58:19,559 JENNIFER TALKED ABOUT HOW GOING 3080 01:58:19,559 --> 01:58:22,395 ANALOG GETS YOU A LOT OF THE WAY 3081 01:58:22,395 --> 01:58:24,297 THERE BUT NEURAL STRUCTURE GETS 3082 01:58:24,297 --> 01:58:27,968 YOU A LOT FARTHER. WHERE IS THE 3083 01:58:27,968 --> 01:58:30,170 BALANCE BETWEEN HYPERDIMENSIONAL 3084 01:58:30,170 --> 01:58:35,141 COMPUTING VERSUS HARDWARE OR OTHER 3085 01:58:35,141 --> 01:58:37,310 APPROACHES TO NEUROMORPHIC AS FAR AS 3086 01:58:37,310 --> 01:58:37,944 ATTAINING ENERGY EFFICIENCY? 3087 01:58:37,944 --> 01:58:40,613 WHERE DO YOU SEE IT AT THAT 3088 01:58:40,613 --> 01:58:41,181 BALANCE POINT? 3089 01:58:41,181 --> 01:58:43,717 >> I THINK SOMETHING LIKE HYPER- 3090 01:58:43,717 --> 01:58:45,752 DIMENSIONAL COMPUTING WHICH GOES 3091 01:58:45,752 --> 01:58:47,954 BY A FEW NAMES, VECTOR SYMBOLIC 3092 01:58:47,954 --> 01:58:49,589 ARCHITECTURE. IT'S AN ALGORITHM 3093 01:58:49,589 --> 01:58:51,691 FRAMEWORK. RIGHT? AND SO IT'S 3094 01:58:51,691 --> 01:58:55,095 JUST A MATHEMATICAL DESCRIPTION, 3095 01:58:55,095 --> 01:58:57,731 A WAY THAT FORMULATES THIS. 3096 01:58:57,731 --> 01:58:59,065 DEEP NETWORK IS ANOTHER ONE. I 3097 01:58:59,065 --> 01:59:01,701 THINK THERE'S A CHALLENGE OF HOW 3098 01:59:01,701 --> 01:59:05,305 DO YOU REPRESENT WHATEVER YOUR 3099 01:59:05,305 --> 01:59:06,339 MATHEMATICAL FRAMEWORK IS INTO 3100 01:59:06,339 --> 01:59:09,809 PRIMITIVES THAT UNDERLY 3101 01:59:09,809 --> 01:59:12,178 REPRESENTATION KERNEL 3102 01:59:12,178 --> 01:59:13,446 OF HARDWARE PERFORMANCE. 3103 01:59:13,446 --> 01:59:15,115 THE 3104 01:59:15,115 --> 01:59:16,016 HARDWARE HAS SOME SORT OF LEVEL 3105 01:59:16,016 --> 01:59:18,818 OF WHICH YOU PROGRAM IT. BE IT 3106 01:59:18,818 --> 01:59:20,453 NEURONS AND SYNAPSES OR 3107 01:59:20,453 --> 01:59:22,355 SOMETHING ELSE AND THERE'S A 3108 01:59:22,355 --> 01:59:23,623 COMPILATION CHALLENGE FOR ANY 3109 01:59:23,623 --> 01:59:25,325 MATH THAT WE WANT TO DEAL WITH 3110 01:59:25,325 --> 01:59:27,761 IN TRYING TO REPRESENT, YOU 3111 01:59:27,761 --> 01:59:30,463 KNOW, SAY, HYPERDIMENSIONAL 3112 01:59:30,463 --> 01:59:32,899 COMPUTING INTO SOMETHING THAT'S 3113 01:59:32,899 --> 01:59:33,700 COMPATIBLE WITH NEUROMORPHIC 3114 01:59:33,700 --> 01:59:35,602 HARDWARE AND THAT'S AN 3115 01:59:35,602 --> 01:59:37,504 INTERESTING AND OPEN QUESTION. 3116 01:59:37,504 --> 01:59:39,439 IT'S A LITTLE MORE OF A 3117 01:59:39,439 --> 01:59:41,007 CS MATH QUESTION BUT IT HAS A 3118 01:59:41,007 --> 01:59:43,710 LOT OF IMPLICATIONS ON, YOU 3119 01:59:43,710 --> 01:59:46,679 KNOW, KIND OF THE QUESTIONS WE 3120 01:59:46,679 --> 01:59:53,486 CAN ASK WITH IT. 3121 01:59:53,486 --> 01:59:55,388 >> QUICKLY ADD TO THAT. I THINK 3122 01:59:55,388 --> 01:59:57,891 A LOT OF THE LEADING A.I. SYSTEMS 3123 01:59:57,891 --> 02:00:01,361 ARE BASED ON FEEDFORWARD NETWORKS 3124 02:00:01,361 --> 02:00:09,269 AND TIME-DEPENDENT COMPUTATIONS TEND 3125 02:00:09,269 --> 02:00:11,838 TO BE SOLVED BY TRANSFORMER-LIKE 3126 02:00:11,838 --> 02:00:13,573 ARCHICTURES. COMPUTATIONS THAT ARE 3127 02:00:13,573 --> 02:00:15,942 CONTINUOUS, TIME DEPENDENT, OR 3128 02:00:15,942 --> 02:00:18,344 RECURRENT MIGHT BE REALLY WELL 3129 02:00:18,344 --> 02:00:18,978 IMPLEMENTED WITH NEUROMORPHIC 3130 02:00:18,978 --> 02:00:21,614 COMPUTING, JUST A SPECULATION. 3131 02:00:21,614 --> 02:00:22,749 >> GREAT. THANK YOU. GOING TO 3132 02:00:22,749 --> 02:00:25,852 JOE AT THE MIC. 3133 02:00:25,852 --> 02:00:31,024 >> HI. NO ONE REALLY OWNS THE 3134 02:00:31,024 --> 02:00:34,861 DEFINITION OF NEUROMORPHIC. BUT 3135 02:00:34,861 --> 02:00:36,663 THERE ARE A LOT OF DIFFERENT 3136 02:00:36,663 --> 02:00:39,332 DEFINITIONS OUT THERE. AND SO 3137 02:00:39,332 --> 02:00:42,402 I'M CURIOUS TO HEAR WHAT THE 3138 02:00:42,402 --> 02:00:44,170 PANEL WOULD DEFINE AS 3139 02:00:44,170 --> 02:00:48,241 NEUROMORPHIC COMPUTATION 3140 02:00:48,241 --> 02:00:50,577 AND WHAT ARE THE 3141 02:00:50,577 --> 02:00:52,412 ELEMENTS THAT MUST BE PRESENT 3142 02:00:52,412 --> 02:00:53,847 FOR IT TO BE CONSIDERED 3143 02:00:53,847 --> 02:00:55,949 NEUROMORPHIC VERSUS TRADITIONAL? 3144 02:00:55,949 --> 02:00:57,417 >> I NOTE THAT THE PARTICIPANT 3145 02:00:57,417 --> 02:01:01,387 GUIDANCE INCLUDED LET'S NOT GET 3146 02:01:01,387 --> 02:01:04,023 INTO TERMINOLOGICAL DEBATES BUT 3147 02:01:04,023 --> 02:01:05,024 IF THERE ARE QUICK ANSWERS THAT 3148 02:01:05,024 --> 02:01:06,059 ARE USEFUL AND GIVE US USEFUL 3149 02:01:06,059 --> 02:01:11,164 BOUNDARIES. 3150 02:01:11,164 --> 02:01:14,367 >> I WOULD SAY IT'S COMPUTING 3151 02:01:14,367 --> 02:01:20,573 THAT SCALES, ITS ENERGY 3152 02:01:20,573 --> 02:01:29,616 EFFICIENCY AS EFFICIENTLY 3153 02:01:29,616 --> 02:01:29,916 AS THE BRAIN. 3154 02:01:29,916 --> 02:01:30,416 >> THAT'S A VERY SHORT 3155 02:01:30,416 --> 02:01:39,292 DEBATE. [LAUGHTER] 3156 02:01:39,292 --> 02:01:43,396 OKAY. I APPRECIATE THAT. YOUR 3157 02:01:43,396 --> 02:01:53,673 MIC IS NOT ON. 3158 02:02:05,552 --> 02:02:07,487 >> HI, I'M HISHAM FROM THE UNIVERSITY 3159 02:02:07,487 --> 02:02:09,389 OF MICHIGAN SO MY QUESTION IS, 3160 02:02:09,389 --> 02:02:10,757 MAYBE A LITTLE BIT TANGENTIAL 3161 02:02:10,757 --> 02:02:13,426 BUT WE ARE AT THE NIH SO MY 3162 02:02:13,426 --> 02:02:14,894 QUESTION IS, WHAT DO YOU GUYS 3163 02:02:14,894 --> 02:02:17,463 THINK SOME OF THE CLOSEST I 3164 02:02:17,463 --> 02:02:19,566 GUESS HEALTH GOALS OR 3165 02:02:19,566 --> 02:02:21,601 DISCOVERIES COULD BE WITH 3166 02:02:21,601 --> 02:02:25,071 SOMETHING LIKE NEUROMORPHIC 3167 02:02:25,071 --> 02:02:34,681 COMPUTING? SORRY, I CAN'T HEAR 3168 02:02:34,681 --> 02:02:34,847 YOU. 3169 02:02:34,847 --> 02:02:36,849 >> IS THE QUESTION MORE ABOUT 3170 02:02:36,849 --> 02:02:38,218 HOW NEUROMORPHIC COMPUTING WOULD 3171 02:02:38,218 --> 02:02:40,520 BE BENEFICIAL IN THE HEALTHCARE 3172 02:02:40,520 --> 02:02:41,921 DOMAIN? 3173 02:02:41,921 --> 02:02:42,255 >> YEAH. 3174 02:02:42,255 --> 02:02:45,458 >> I THINK THERE ARE A COUPLE 3175 02:02:45,458 --> 02:02:46,459 OF EXAMPLES THAT I -- WE ARE 3176 02:02:46,459 --> 02:02:48,628 WORKING ON, LIKE, FOR EXAMPLE, 3177 02:02:48,628 --> 02:02:51,731 IN THE TRAUMA RESEARCH ACADEMY 3178 02:02:51,731 --> 02:02:52,932 WE HAVE SEEN LIKE, YOU KNOW, 3179 02:02:52,932 --> 02:02:54,601 THERE ARE A LOT OF THESE DATA 3180 02:02:54,601 --> 02:02:57,170 POINTS THAT ARE COLLECTED AND 3181 02:02:57,170 --> 02:02:58,771 YOU NEED TO BE ABLE TO MAKE 3182 02:02:58,771 --> 02:03:00,340 REALTIME DECISION MAKING IN A 3183 02:03:00,340 --> 02:03:03,243 MATTER OF 30 SECONDS. AND YOU 3184 02:03:03,243 --> 02:03:05,078 NEED TO BE ABLE TO DO SOME FORM 3185 02:03:05,078 --> 02:03:07,013 OF PREDICTION, SOME FORM OF 3186 02:03:07,013 --> 02:03:08,481 CLASSIFICATION IN A VERY DYNAMIC 3187 02:03:08,481 --> 02:03:11,417 ENVIRONMENT AND DATA IS SLIGHTLY 3188 02:03:11,417 --> 02:03:13,519 SPARSE. SO WE ARE SEEING THAT 3189 02:03:13,519 --> 02:03:14,754 THERE COULD BE AN OPPORTUNITY 3190 02:03:14,754 --> 02:03:16,222 HERE WHERE YOU COULD HAVE THIS 3191 02:03:16,222 --> 02:03:21,227 AS AN A.I. HEALTH ASSISTANT. AND 3192 02:03:21,227 --> 02:03:23,796 THERE ARE SEVERAL OTHER SMALLER 3193 02:03:23,796 --> 02:03:26,966 EXAMPLES THAT ARE SHOWN IN THE 3194 02:03:26,966 --> 02:03:29,302 COMMUNITY WHETHER THEY'RE USING 3195 02:03:29,302 --> 02:03:32,272 LIKE EEG DATA, EMG DATA, OR OTHER 3196 02:03:32,272 --> 02:03:34,474 FORMS OF DATA TO DO DIFFERENT 3197 02:03:34,474 --> 02:03:39,479 TYPES OF CLASSIFICATION OR SOME 3198 02:03:39,479 --> 02:03:40,580 FORM OF PREDICTION IN DIFFERENT 3199 02:03:40,580 --> 02:03:44,083 TYPES OF DISEASES, FOR EXAMPLE 3200 02:03:44,083 --> 02:03:45,685 ONE DOMAIN THEY ARE ALSO LOOKING 3201 02:03:45,685 --> 02:03:48,221 AT IS IN HOW YOU COULD CLASSIFY 3202 02:03:48,221 --> 02:03:50,857 DIFFERENT SUBTYPES OF 3203 02:03:50,857 --> 02:03:52,258 NEURODEGENERATIVE DISEASES. 3204 02:03:52,258 --> 02:03:53,726 THAT IS PROBABLY NOT AT A 3205 02:03:53,726 --> 02:03:57,397 REALTIME CLASSIFICATION PROBLEM. 3206 02:03:57,397 --> 02:03:59,265 THAT IS -- BUT THERE IS A LOT OF 3207 02:03:59,265 --> 02:04:02,435 MISSING DATA POINTS IN THERE AND 3208 02:04:02,435 --> 02:04:08,708 IT'S UNDERREPRESENTED OR 3209 02:04:08,708 --> 02:04:11,244 UNDERSAMPLED OR IMBALANCED. WE 3210 02:04:11,244 --> 02:04:12,912 SEE SOME OF THESE SPIKING 3211 02:04:12,912 --> 02:04:13,880 ALGORITHMS ARE REALLY GOOD AT 3212 02:04:13,880 --> 02:04:15,982 THAT AND MAYBE THE NEXT STAGE 3213 02:04:15,982 --> 02:04:18,184 WILL BE CONSIDERING DEPLOYMENT 3214 02:04:18,184 --> 02:04:19,719 ON NEUROMORPHIC HARDWARE. THESE 3215 02:04:19,719 --> 02:04:21,087 ARE SOME EXAMPLES WE ARE 3216 02:04:21,087 --> 02:04:21,421 THINKING OF. 3217 02:04:21,421 --> 02:04:24,057 >> WE DO HAVE A WHOLE SESSION 3218 02:04:24,057 --> 02:04:25,124 DEDICATED TO THESE KINDS OF 3219 02:04:25,124 --> 02:04:25,725 QUESTIONS THIS AFTERNOON, 3220 02:04:25,725 --> 02:04:27,160 SESSION FOUR. AND GIVEN THAT WE 3221 02:04:27,160 --> 02:04:28,494 ONLY HAVE A MINUTE LEFT AND I 3222 02:04:28,494 --> 02:04:30,129 SEE THREE PEOPLE STANDING, I 3223 02:04:30,129 --> 02:04:37,437 JUST WANT TO KEEP IT GOING. 3224 02:04:37,437 --> 02:04:41,541 >> I'M RUHIRDATTA, I'M NEW TO 3225 02:04:41,541 --> 02:04:43,309 NEUROMORPHIC COMPUTING. I HAVE A 3226 02:04:43,309 --> 02:04:47,947 QUESTION ABOUT THE THOR COMMONS. 3227 02:04:47,947 --> 02:04:50,450 IF YOU COULD ELABORATE A LITTLE 3228 02:04:50,450 --> 02:04:52,151 BIT. I HAVE SEEN THINGS BEING 3229 02:04:52,151 --> 02:04:53,453 IN SILICON 3230 02:04:53,453 --> 02:04:54,887 THAT ARE SPECIALIZED 3231 02:04:54,887 --> 02:04:56,456 NEUROMORPHIC HARDWARE BUT I'VE 3232 02:04:56,456 --> 02:04:59,292 ALSO SEEN ELSEWHERE THAT THERE'S 3233 02:04:59,292 --> 02:05:00,460 SOMETIMES SPIKING NEURAL 3234 02:05:00,460 --> 02:05:03,062 NETWORKS ARE IMPLEMENTED IN GPUS 3235 02:05:03,062 --> 02:05:03,996 THEMSELVES. AND SOMETIMES 3236 02:05:03,996 --> 02:05:06,699 THERE'S FPGAS. SO COULD YOU SAY A 3237 02:05:06,699 --> 02:05:10,069 LITTLE BIT MORE ABOUT WHAT THE 3238 02:05:10,069 --> 02:05:13,473 THOR COMMONS IS GOING TO BE 3239 02:05:13,473 --> 02:05:15,308 LIKE? PROFESSOR KUDITHIPUDI, 3240 02:05:15,308 --> 02:05:18,111 IF I SAID THAT RIGHT. 3241 02:05:18,111 --> 02:05:20,079 >> WAS THE QUESTION ABOUT HOW 3242 02:05:20,079 --> 02:05:21,280 HOW PERFORMANCE VARIES ACROSS 3243 02:05:21,280 --> 02:05:24,050 THESE PLATFORMS OR WHAT IS 3244 02:05:24,050 --> 02:05:25,418 THE -- 3245 02:05:25,418 --> 02:05:28,654 >> OR WHAT KINDS OF PLATFORMS 3246 02:05:28,654 --> 02:05:31,424 ARE GOING TO BE AVAILABLE IN THE 3247 02:05:31,424 --> 02:05:31,657 THOR COMMONS? 3248 02:05:31,657 --> 02:05:32,558 >> OH, IN THE NEUROMORPHIC 3249 02:05:32,558 --> 02:05:34,394 COMMONS, SORRY. 3250 02:05:34,394 --> 02:05:36,062 >> YEAH. 3251 02:05:36,062 --> 02:05:38,631 >> THANK YOU. SO THE 3252 02:05:38,631 --> 02:05:39,866 NEUROMORPHIC COMMONS IS GOING TO 3253 02:05:39,866 --> 02:05:46,739 BE PROVIDING ACCESS TO PRIMARILY 3254 02:05:46,739 --> 02:05:48,641 NEUROMORPHIC HARDWARE NOT THE 3255 02:05:48,641 --> 02:05:52,979 GPU PLATFORMS. WE WOULD PERHAPS 3256 02:05:52,979 --> 02:05:54,814 BE OPEN TO THAT IDEA AT A LATER 3257 02:05:54,814 --> 02:05:56,482 PHASE. WE ARE IN CONVERSATIONS 3258 02:05:56,482 --> 02:05:58,284 WITH DIFFERENT INDUSTRY PARTNERS 3259 02:05:58,284 --> 02:05:59,952 BUT THE IDEA IS TO KIND OF 3260 02:05:59,952 --> 02:06:03,623 PROVIDE ACCESS TO LARGE-SCALE 3261 02:06:03,623 --> 02:06:05,958 SPIKING NEUROMORPHIC HARDWARE 3262 02:06:05,958 --> 02:06:07,293 WHERE PEOPLE CURRENTLY ARE NOT 3263 02:06:07,293 --> 02:06:09,495 ABLE TO REALIZE SOME OF THE 3264 02:06:09,495 --> 02:06:11,964 SPIKING NEURAL NETWORKS. IN 3265 02:06:11,964 --> 02:06:15,201 THIS LARGE-SCALE HARDWARE 3266 02:06:15,201 --> 02:06:16,836 PLATFORM SO WE ARE GOING TO WORK 3267 02:06:16,836 --> 02:06:18,438 WITH INDUSTRY PARTNER WHO CAN 3268 02:06:18,438 --> 02:06:21,808 GIVE US ACCESS TO THAT KIND OF 3269 02:06:21,808 --> 02:06:22,175 INFRASTRUCTURE. 3270 02:06:22,175 --> 02:06:23,609 >> AND WOULD THOSE BE 3271 02:06:23,609 --> 02:06:28,848 SPECIALIZED INDUSTRY PARTNER 3272 02:06:28,848 --> 02:06:29,248 LIKE BRAINCHIP? 3273 02:06:29,248 --> 02:06:31,884 OR WOULD THOSE BE LARGE-SCALE 3274 02:06:31,884 --> 02:06:36,022 TECH LIKE INTEL OR IBM? 3275 02:06:36,022 --> 02:06:37,824 >> WE ARE TALKING TO THE 3276 02:06:37,824 --> 02:06:39,625 START-UP ECOSYSTEM AS WELL AS THE 3277 02:06:39,625 --> 02:06:41,661 LARGER INDUSTRY PARTNERS. 3278 02:06:41,661 --> 02:06:44,130 BECAUSE THERE IS A LOT OF 3279 02:06:44,130 --> 02:06:45,465 HETEROGENEITY IN THESE PLATFORMS 3280 02:06:45,465 --> 02:06:48,434 AND WHAT THEY CAN PROVIDE IN 3281 02:06:48,434 --> 02:06:51,904 TERMS OF THE SIMULATION 3282 02:06:51,904 --> 02:06:52,805 CAPABILITIES AND THE SCALE TO 3283 02:06:52,805 --> 02:06:54,173 WHICH WE CAN SIMULATE. SO WE ARE 3284 02:06:54,173 --> 02:06:57,310 GOING TO GET ACCESS TO MULTIPLE 3285 02:06:57,310 --> 02:06:59,445 OF THOSE PLATFORMS AND SOME OF 3286 02:06:59,445 --> 02:07:01,447 THOSE COULD BE COMING FROM 3287 02:07:01,447 --> 02:07:09,555 ACADEMIC LABS AS 3288 02:07:09,555 --> 02:07:09,755 WELL. 3289 02:07:09,755 --> 02:07:11,090 >> GREAT. THANK YOU. 3290 02:07:11,090 --> 02:07:12,158 >> AND THANKS TO THE SPEAKERS 3291 02:07:12,158 --> 02:07:19,365 AND THE PANEL ISSIST -- 3292 02:07:19,365 --> 02:07:20,099 PANELISTS AND TO EVERYONE FOR 3293 02:07:20,099 --> 02:07:22,535 THEIR QUESTIONS. I WILL RELEASE 3294 02:07:22,535 --> 02:07:23,836 EVERYONE TO LUNCH NOW BUT I WANT 3295 02:07:23,836 --> 02:07:29,108 TO MAKE A FEW QUICK 3296 02:07:29,108 --> 02:07:29,575 ANNOUNCEMENTS. 3297 02:07:29,575 --> 02:07:31,711 THE POSTER SESSION IS STILL 3298 02:07:31,711 --> 02:07:34,313 AVAILABLE DURING LUNCH. WE ARE 3299 02:07:34,313 --> 02:07:40,853 RESUMING AT 1:30 FOR SESSION FOUR. 3300 02:07:40,853 --> 02:07:44,223 I WOULD LIKE TO INTRODUCE OUR 3301 02:07:44,223 --> 02:07:47,393 SESSION CHAIR, GINA ADAM. 3302 02:07:47,393 --> 02:07:49,662 SHE'S AN ASSOCIATE PROFESSOR IN 3303 02:07:49,662 --> 02:07:52,865 THE ELECTRICAL AND COMPUTER 3304 02:07:52,865 --> 02:07:54,066 ENGINEERING DEPARTMENT AT GEORGE 3305 02:07:54,066 --> 02:07:56,903 MASON UNIVERSITY. LET'S WELCOME HER 3306 02:07:56,903 --> 02:08:00,940 TO GIVE US OUR FIRST TALK. 3307 02:08:00,940 --> 02:08:03,042 >> ESTEEMED COLLEAGUES. 3308 02:08:03,042 --> 02:08:05,611 IT'S MY PLEASURE TO WELCOME YOU 3309 02:08:05,611 --> 02:08:10,483 TO SESSION FOUR TOWARDS RECIPROCAL 3310 02:08:10,483 --> 02:08:12,585 BRAIN NEUROAI ADVANCES FOR 3311 02:08:12,585 --> 02:08:14,787 INTELLIGENT COMPUTING, ROBOTICS 3312 02:08:14,787 --> 02:08:17,924 AND NEUROTECHNOLOGY. 3313 02:08:17,924 --> 02:08:21,694 WE HAVE AN ESTEEMED SET OF 3314 02:08:21,694 --> 02:08:23,963 PRESENTERS AND DISCUSSANTS. 3315 02:08:23,963 --> 02:08:25,998 IT'S REALLY, REALLY GREAT TO 3316 02:08:25,998 --> 02:08:27,900 HAVE THE DISCUSSION AT THE END 3317 02:08:27,900 --> 02:08:29,435 OF THE WORKSHOP. 3318 02:08:29,435 --> 02:08:32,838 IT IS ALSO VERY NICE TO SEE THE 3319 02:08:32,838 --> 02:08:39,178 INTEREST THAT NIH HAS IN 3320 02:08:39,178 --> 02:08:45,051 COMPUTING AND HOW TO BRIDGE THE 3321 02:08:45,051 --> 02:08:45,851 GAP TOWARD NEUROAI. 3322 02:08:45,851 --> 02:08:47,987 THIS IS A TIMELY MOMENT BECAUSE 3323 02:08:47,987 --> 02:08:51,490 I THINK WE'RE IN A HISTORIC TIME 3324 02:08:51,490 --> 02:08:54,961 IN OUR DEVELOPMENT OF A.I. 3325 02:08:54,961 --> 02:08:58,764 THE CURRENT A.I. FRAMEWORK AS 3326 02:08:58,764 --> 02:09:04,303 DISCUSSED IN SESSION THREE ARE 3327 02:09:04,303 --> 02:09:05,204 ARE DISEMBODIED AND LARGE SCALE. 3328 02:09:05,204 --> 02:09:07,707 THEY STRUGGLE WITH THE SEVERE 3329 02:09:07,707 --> 02:09:09,008 ENERGY REQUIREMENTS. 3330 02:09:09,008 --> 02:09:11,877 A LOT OF REQUIREMENTS IN TERMS 3331 02:09:11,877 --> 02:09:15,047 OF HARDWARE FOR THE SERVER FARMS TO 3332 02:09:15,047 --> 02:09:17,950 RUN THESE VERY LARGE-SCALE 3333 02:09:17,950 --> 02:09:19,986 MODELS. 3334 02:09:19,986 --> 02:09:25,024 CURRENTLY THE A.I. FRAMEWORK HAS 3335 02:09:25,024 --> 02:09:26,459 DONE SIGNIFICANT ADVANCES TO 3336 02:09:26,459 --> 02:09:30,062 SUPPORT THE SCIENTIFIC RESEARCH. 3337 02:09:30,062 --> 02:09:34,800 BUT THEY ARE RESOURCE OBLIVIOUS 3338 02:09:34,800 --> 02:09:36,736 ALGORITHMS. 3339 02:09:36,736 --> 02:09:40,239 IT IS TIME TO TAKE THAT KNOWLEDGE, 3340 02:09:40,239 --> 02:09:43,109 BRING NEW KNOWLEDGE FROM NEUROSCIENCE 3341 02:09:43,109 --> 02:09:46,112 AND BIOMECHANICS AS DISCUSSED 3342 02:09:46,112 --> 02:09:48,214 IN THE PRIOR SESSION AND OTHER 3343 02:09:48,214 --> 02:09:49,882 TYPES OF DISCIPLINES AND THINK 3344 02:09:49,882 --> 02:09:55,187 HOW WE CAN DEVELOP PHYSICAL A.I. 3345 02:09:55,187 --> 02:09:57,523 FRAMEWORKS THAT CAN SUPPORT 3346 02:09:57,523 --> 02:10:01,027 ADAPTIVE COMPUTING, CONTINUAL 3347 02:10:01,027 --> 02:10:05,297 LEARNING IN A WAY THAT'S OPTIMIZED 3348 02:10:05,297 --> 02:10:07,967 TO PHYSICAL CONSTRAINTS. 3349 02:10:07,967 --> 02:10:10,169 THAT IS KEY TO NEUROAI WHEN WE 3350 02:10:10,169 --> 02:10:14,807 THINK ABOUT APPLICATIONS IN 3351 02:10:14,807 --> 02:10:18,010 ROBOTIC, NEURO-CONTROL, PROSTHETICS. 3352 02:10:18,010 --> 02:10:22,348 DISCUSSANTS AND PRESENTERS IN THIS 3353 02:10:22,348 --> 02:10:24,050 SESSION WILL TALK IN MORE DEPTH 3354 02:10:24,050 --> 02:10:26,752 ABOUT THEIR WORK. 3355 02:10:26,752 --> 02:10:31,490 SO, IN ORDER TO ACHIEVE THAT, WE 3356 02:10:31,490 --> 02:10:33,492 HAVE TO DRAW INSPIRATION FROM 3357 02:10:33,492 --> 02:10:41,600 THE BRAIN FOR THE DEVELOPMENT OF 3358 02:10:41,600 --> 02:10:43,969 NEUROMORPHIC COMPUTING SYSTEMS. 3359 02:10:43,969 --> 02:10:46,872 A GRAND CHALLENGE IN OUR TIME IS TO 3360 02:10:46,872 --> 02:10:49,175 DEVELOP THE ARTIFICAL NEURAL SYSTEMS 3361 02:10:49,175 --> 02:10:52,378 THAT PERFORM AT THE COMPLEXITY AND 3362 02:10:52,378 --> 02:10:54,747 ENERGY EFFICIENCY OF THEIR BIOLOGICAL 3363 02:10:54,747 --> 02:10:55,181 COUNTERPARTS. 3364 02:10:55,181 --> 02:11:02,722 WE HAD DR. JENNIFER HASLER TALKING 3365 02:11:02,722 --> 02:11:06,025 ABOUT A ROADMAP TO DEVELOP NEUROMORPHIC 3366 02:11:06,025 --> 02:11:08,394 SYSTEMS AND HOW ANALOG COMPUTING 3367 02:11:08,394 --> 02:11:10,796 AND NEUROMORPHIC TECHNIQUES 3368 02:11:10,796 --> 02:11:13,966 AND PRINCIPLES CAN BRIDGE THE GAP IN 3369 02:11:13,966 --> 02:11:15,868 TERMS OF POWER EFFICIENCY. 3370 02:11:15,868 --> 02:11:21,073 WHERE WE ARE TODAY IS THE RESULT 3371 02:11:21,073 --> 02:11:23,109 OF THE DEVELOPMENT IN THE PAST 3372 02:11:23,109 --> 02:11:29,048 FOUR DECADES IN NEUROMORPHIC 3373 02:11:29,048 --> 02:11:29,782 HARDWARE. 3374 02:11:29,782 --> 02:11:35,087 I WANT TO HIGHLIGHT SOME OF 3375 02:11:35,087 --> 02:11:36,689 THESE CONTRIBUTIONS. WE HAVE 3376 02:11:36,689 --> 02:11:42,561 CARVER MEAD IN THE '80S WITH WORK 3377 02:11:42,561 --> 02:11:46,232 ON TRANSLATING NEUROMORPHIC 3378 02:11:46,232 --> 02:11:51,437 PRINCIPLES IN ANALOGUE VLSI AND THE 3379 02:11:51,437 --> 02:11:54,607 DEVELOPMENT OF THE SILICON RETINA 3380 02:11:54,607 --> 02:11:56,308 FOLLOWING SIMILAR PRINCIPLES. 3381 02:11:56,308 --> 02:11:58,911 I WANT TO GIVE A SHOUT OUT 3382 02:11:58,911 --> 02:12:02,248 BECAUSE WE HAVE DR. 3383 02:12:02,248 --> 02:12:03,282 GIACOMO INDIVERI FROM ZURICH. 3384 02:12:03,282 --> 02:12:06,185 I WANT TO GIVE A SHOUT OUT TO 3385 02:12:06,185 --> 02:12:10,990 THE INSTITUTE OF NEUROINFORMATICS 3386 02:12:10,990 --> 02:12:13,793 ESTABLISHED IN THE MID '90S. 3387 02:12:13,793 --> 02:12:18,497 FROM THAT POINT ON, YOU SEE WE MOVE 3388 02:12:18,497 --> 02:12:24,603 FROM THE ERA OF SMALL-SCALE TO LARGE- 3389 02:12:24,603 --> 02:12:33,045 SCALE INTEGRATION OF NEUROMORPHIC 3390 02:12:33,045 --> 02:12:34,013 HARDWARE USING CMOS TECHNOLOGIES. 3391 02:12:34,013 --> 02:12:35,648 DIFFERENT FUNDERS AND COUNTRIES ARE 3392 02:12:35,648 --> 02:12:38,484 SUPPORTING THESE DEVELOPMENTS 3393 02:12:38,484 --> 02:12:41,220 WITH THE SPINNAKER IN EUROPE. 3394 02:12:41,220 --> 02:12:47,459 WE HAVE BRAINSCALES WAFERS, THE 3395 02:12:47,459 --> 02:12:58,237 SYNAPSE PROJECT FUNDED BY DARPA 3396 02:12:59,205 --> 02:13:01,373 AND NOW WE LOOK AT HOW TO 3397 02:13:01,373 --> 02:13:02,041 COMMERCIALIZE NEUROMORPHIC HARDWARE 3398 02:13:02,041 --> 02:13:03,042 DEVELOPMENTS. 3399 02:13:03,042 --> 02:13:09,348 I WANT TO POINT OUT THE LOIHI CHIPS 3400 02:13:09,348 --> 02:13:13,619 BY INTEL AND IT WAS DISCUSSED 3401 02:13:13,619 --> 02:13:15,754 IN SESSION THREE. 3402 02:13:15,754 --> 02:13:20,226 THINKING ABOUT CREATING THE 3403 02:13:20,226 --> 02:13:22,828 NEUROMORPHIC SYSTEMS AT SANDIA. 3404 02:13:22,828 --> 02:13:29,702 OTHER TYPES OF NEUROMORPHIC CHIP 3405 02:13:29,702 --> 02:13:30,536 ARE ALSO BEING DEVELOPED. 3406 02:13:30,536 --> 02:13:35,074 I WANTED TO GIVE A SHOUT OUT TO 3407 02:13:35,074 --> 02:13:37,877 GIACOMO INDIVERI AND THEIR 3408 02:13:37,877 --> 02:13:47,086 DEVELOPMENT OF THE DYNAP 3409 02:13:47,086 --> 02:13:47,953 CHIPS. SO WHAT'S NEXT? WHAT IS 3410 02:13:47,953 --> 02:13:49,788 COMING NEXT FOR PHYSICL A.I.? 3411 02:13:49,788 --> 02:13:54,360 WE HAVE THE NOBEL PRIZE IN 3412 02:13:54,360 --> 02:13:56,028 PHYSICS GIVEN FOR A.I. 3413 02:13:56,028 --> 02:13:58,030 DEVELOPMENT, WE THINK ABOUT WHAT 3414 02:13:58,030 --> 02:14:00,633 IS NEXT FOR THE PHYSICAL A.I.? 3415 02:14:00,633 --> 02:14:04,136 AND HOW DO WE CONTINUE THIS 3416 02:14:04,136 --> 02:14:07,339 DEVELOPMENT OF NEUROMORPHIC 3417 02:14:07,339 --> 02:14:09,775 HARDWARE FOR TRANSLATION TO 3418 02:14:09,775 --> 02:14:15,814 INTELLIGENCE AND COMPUTING, ROBOTICS, 3419 02:14:15,814 --> 02:14:21,820 AND NEUROTECHNOLOGIES? I 3420 02:14:21,820 --> 02:14:25,824 THINK THE EMERGENT DEVICES WILL 3421 02:14:25,824 --> 02:14:28,127 PLAY A KEY ROLE IN THE 3422 02:14:28,127 --> 02:14:32,031 DEVELOPMENT OF ENERGY EFFICIENT 3423 02:14:32,031 --> 02:14:35,768 AND COMPACT NEUROAI SYSTEMS. 3424 02:14:35,768 --> 02:14:39,371 IN ORDER TO INCORPORATE EMERGING 3425 02:14:39,371 --> 02:14:42,474 NEUROMORPHIC TECHNOLOGIES, WE 3426 02:14:42,474 --> 02:14:44,643 NEED TO HAVE A VISION FOR 3427 02:14:44,643 --> 02:14:46,078 VERTICAL INTEGRATION. 3428 02:14:46,078 --> 02:14:47,646 WE NEED TO THINK ABOUT THE STACK 3429 02:14:47,646 --> 02:14:57,022 AND BE ABLE TO HAVE CODESIGN 3430 02:14:57,022 --> 02:14:59,058 APPROACHES TO OPTIMIZE ANYWHERE 3431 02:14:59,058 --> 02:15:03,395 FROM MATERIALS TO DEVICES, TO CIRCUIT 3432 02:15:03,395 --> 02:15:13,872 DESIGN, PROTOTYPING, TO ALGORITHMS 3433 02:15:13,872 --> 02:15:14,873 DRIVEN BY BIOINSPIRATION. 3434 02:15:14,873 --> 02:15:16,342 I SHOW HERE THE DIFFERENT DEVICES 3435 02:15:16,342 --> 02:15:18,143 THAT WE'RE WORKING ON. WE'RE WORKING 3436 02:15:18,143 --> 02:15:22,648 ON MEMRISTIVE DEVICES BASED ON 3437 02:15:22,648 --> 02:15:26,585 OXIDE MATERIALS THAT HAVE SHOWN 3438 02:15:26,585 --> 02:15:37,129 SYNAPTIC PROPERTIES, SUCH AS STDP BEHAVIOR, SPIKING BEHAVIOR. 3439 02:15:40,366 --> 02:15:43,068 THIS IS A SINGLE DEVICE AND WE 3440 02:15:43,068 --> 02:15:46,672 HAVE HERE 20,000 DEVICES INTEGRATED 3441 02:15:46,672 --> 02:15:49,274 ON TOP OF TRANSISTOR CIRCUITRY 3442 02:15:49,274 --> 02:15:55,180 WHICH IS PACKAGED AND CONNECTED 3443 02:15:55,180 --> 02:15:58,817 TO CUSTOM PROTOTYPING BOARDS AND 3444 02:15:58,817 --> 02:16:05,457 INTEGRATED HERE WITH FPGA BOARD. 3445 02:16:05,457 --> 02:16:07,960 USER AT THE TOP LEVEL OF THE STACK 3446 02:16:07,960 --> 02:16:10,429 CAN PROGRAM THESE DEVICES ALL THE 3447 02:16:10,429 --> 02:16:13,932 WAY DOWN WITHOUT KNOWING HOW TO 3448 02:16:13,932 --> 02:16:17,102 DO THE MEASUREMENTS ON THE DEVICE 3449 02:16:17,102 --> 02:16:18,737 ITSELF. BUT BE ABLE TO ACCESS 3450 02:16:18,737 --> 02:16:24,309 EVERYTHING AT THE SYSTEM LEVEL. 3451 02:16:24,309 --> 02:16:27,546 THIS IDEA OF HAVING A USER 3452 02:16:27,546 --> 02:16:29,181 FRIENDLY NEUROMORPHIC SYSTEM 3453 02:16:29,181 --> 02:16:31,383 THAT WE CAN USE FOR PROTOTYPING 3454 02:16:31,383 --> 02:16:32,651 IS KEY. 3455 02:16:32,651 --> 02:16:36,655 I'M GLAD TO SEE THE INITIATIVES 3456 02:16:36,655 --> 02:16:38,490 IN BUILDING COMMUNITIES FOR 3457 02:16:38,490 --> 02:16:41,593 THIS NEUROMORPHIC HARDWARE AND 3458 02:16:41,593 --> 02:16:44,096 THIS WAS BROUGHT UP IN SESSION 3459 02:16:44,096 --> 02:16:48,067 THREE WITH THE THOR INITIATIVE. 3460 02:16:48,067 --> 02:16:50,803 AND OUR GOAL IS TO CONTRIBUTE TO 3461 02:16:50,803 --> 02:16:56,508 THAT AND BE ABLE TO HAVE DIVERSE 3462 02:16:56,508 --> 02:16:59,545 USERS AND STAFF BE ABLE TO 3463 02:16:59,545 --> 02:17:01,447 UTILIZE THE PLATFORMS AND 3464 02:17:01,447 --> 02:17:04,149 INTEGRATE THEIR OWN DEVICES AND 3465 02:17:04,149 --> 02:17:06,752 TECHNOLOGIES INTO THE SYSTEM. 3466 02:17:06,752 --> 02:17:09,455 FOR THIS, I WANT TO GIVE A SHOUT 3467 02:17:09,455 --> 02:17:11,790 OUT TO ALL OF MY STUDENTS WHO 3468 02:17:11,790 --> 02:17:15,994 HELPED OUT WITH THIS AS WELL AS 3469 02:17:15,994 --> 02:17:17,129 THE DIFFERENT FUNDING AGENCIES. 3470 02:17:17,129 --> 02:17:21,567 AND I WANT TO TAKE A FEW MINUTES 3471 02:17:21,567 --> 02:17:25,270 TO DISCUSS ANOTHER EFFORT IN 3472 02:17:25,270 --> 02:17:28,273 COMMUNITY BUILDING. THE 2024 3473 02:17:28,273 --> 02:17:36,081 NEUROMORPHIC COMPUTING FOR SCIENCES 3474 02:17:36,081 --> 02:17:38,617 WORKSHOP WAS SPONSORED BY THE DOE. 3475 02:17:38,617 --> 02:17:40,419 THIS WAS TWO MONTHS AGO AND WE HAD 3476 02:17:40,419 --> 02:17:42,654 ABOUT 70 PARTICIPANTS. THE VISION 3477 02:17:42,654 --> 02:17:45,758 WAS TO DISCUSS KEY RESEARCH NEEDS, 3478 02:17:45,758 --> 02:17:47,593 CHALLENGES, AND NEXT STEPS IN 3479 02:17:47,593 --> 02:17:51,029 BIO-REALISTIC NEUROMORPHIC CIRCUITS. 3480 02:17:51,029 --> 02:17:54,133 THREE THEMES WERE PRESENTED: 3481 02:17:54,133 --> 02:17:56,702 NEUROSCIENCE, MICROELECTRONICS, 3482 02:17:56,702 --> 02:18:05,010 LARGE-SCALE MODELING AND SIMULATION. 3483 02:18:05,010 --> 02:18:07,846 WE STRUCTURED OUR DISCUSSIONS 3484 02:18:07,846 --> 02:18:10,782 ACROSS FOUR KEY QUESTIONS, WHAT 3485 02:18:10,782 --> 02:18:13,185 ARE THE PRIMITIVES NEEDED TO CAPTURE 3486 02:18:13,185 --> 02:18:17,156 CRITICAL, BIOLOGICAL COMPUTING 3487 02:18:17,156 --> 02:18:20,292 MECHANISMS? WHAT ARE TECHNOLOGIES 3488 02:18:20,292 --> 02:18:21,693 NEEDED FOR DEMONSTRATION AND PROTOTYPING? 3489 02:18:21,693 --> 02:18:26,198 WHAT ARE THE CRITICAL CHARACTERISTICS 3490 02:18:26,198 --> 02:18:29,067 FOR SCALING SIMULATIONS? A CROSS 3491 02:18:29,067 --> 02:18:30,269 CUTTING DISCUSSION HAPPENED 3492 02:18:30,269 --> 02:18:33,071 REGARDING NEUROSCIENCE-BASED 3493 02:18:33,071 --> 02:18:36,708 BENCH MARK AND DATA SETS TO 3494 02:18:36,708 --> 02:18:39,645 EFFECTIVELY CHARACTERIZE THESE 3495 02:18:39,645 --> 02:18:48,420 NEUROMORPHIC CIRCUITRY AND SIMULATION. 3496 02:18:48,420 --> 02:18:52,591 WE CAME UP THIS SLIDE AT THE 3497 02:18:52,591 --> 02:18:58,197 CONCLUSION OF THE WORKSHOP FOR 3498 02:18:58,197 --> 02:19:00,465 ACCELERATING THE DESIGN, PROTOTYPING, 3499 02:19:00,465 --> 02:19:04,670 OF PRIMITIVES, TRANSFORMING THE 3500 02:19:04,670 --> 02:19:07,539 CONDUCTIVITY AND PIONEERING AN 3501 02:19:07,539 --> 02:19:10,776 INTEGRATED ECOSYSTEMS FOR SCALABLE 3502 02:19:10,776 --> 02:19:14,513 NEUROMORPHIC CODESIGN, LEVERAGING THE 3503 02:19:14,513 --> 02:19:18,183 NEUROSCIENCE-INSPIRED DYNAMICS AND 3504 02:19:18,183 --> 02:19:19,084 ALGORITHMS. 3505 02:19:19,084 --> 02:19:22,120 THE REPORT WILL COME UP SOON. 3506 02:19:22,120 --> 02:19:26,592 THANK YOU VERY MUCH. 3507 02:19:26,592 --> 02:19:31,697 >> THANK YOU VERY MUCH GINA FOR 3508 02:19:31,697 --> 02:19:35,867 SHARING THIS DOE WORKSHOP WITH US. 3509 02:19:35,867 --> 02:19:38,337 WE'LL INVITE CHIARA BARTOLOZZI 3510 02:19:38,337 --> 02:19:44,409 FROM IIT TO TELL US ABOUT EMBODIED 3511 02:19:44,409 --> 02:19:44,943 NEUROMORPHIC INTELLIGENCE. 3512 02:19:44,943 --> 02:19:51,049 >> THANK YOU FOR INVITING ME. 3513 02:19:51,049 --> 02:19:57,022 SO, EMBODIED IS A WORD THAT IS 3514 02:19:57,022 --> 02:19:59,524 RESONATING IN THIS MEETING AND 3515 02:19:59,524 --> 02:20:04,763 ALSO NEUROMORPHIC. 3516 02:20:04,763 --> 02:20:09,568 SO AS YOU UNDERSTOOD 3517 02:20:09,568 --> 02:20:11,603 FROM ONE OF THE LATEST QUESTIONS 3518 02:20:11,603 --> 02:20:14,573 HERE IN THE PREVIOUS SESSION, 3519 02:20:14,573 --> 02:20:18,243 THERE IS NO CONSENSUS ON WHAT 3520 02:20:18,243 --> 02:20:19,211 NEUROMORPHIC IS. 3521 02:20:19,211 --> 02:20:25,884 I WANTED TO START BY GIVING MY 3522 02:20:25,884 --> 02:20:27,019 OWN DEFINITION. I WANT TO UNDERSTAND 3523 02:20:27,019 --> 02:20:30,055 BIOLOGY AND THE COMPUTATION DONE IN 3524 02:20:30,055 --> 02:20:33,258 BIOLOGY TO BUILD BETTER TECHNOLOGY. 3525 02:20:33,258 --> 02:20:36,762 WHAT DOES IT MEAN AND IN WHICH 3526 02:20:36,762 --> 02:20:43,936 DOMAIN IS BIOLOGY BETTER THAN 3527 02:20:43,936 --> 02:20:46,838 TRADITIONAL DIGITAL COMPUTATION? 3528 02:20:46,838 --> 02:20:52,811 I WANT TO POINT TO THE TWO 3529 02:20:52,811 --> 02:20:54,646 FIGURES HERE THAT SHOW 3530 02:20:54,646 --> 02:20:58,917 THAT BASICALLY THE BRAIN IS NOT 3531 02:20:58,917 --> 02:21:02,921 A BUBBLE THERE, BUT EMBODIED IN A 3532 02:21:02,921 --> 02:21:06,258 BODY, CONNECTED TO A BODY THAT 3533 02:21:06,258 --> 02:21:06,992 ACTS. 3534 02:21:06,992 --> 02:21:08,126 SO. 3535 02:21:08,126 --> 02:21:10,329 THE BRAIN IS PART OF A SYSTEM 3536 02:21:10,329 --> 02:21:12,230 WITH BRAIN AND BODY. WHEN WE 3537 02:21:12,230 --> 02:21:19,538 THINK ABOUT NEUROMORPHIC TASKS, 3538 02:21:19,538 --> 02:21:23,675 THEY INTERACT WITH THE WORLD IN 3539 02:21:23,675 --> 02:21:26,545 REAL TIME. THIS IS WHAT TONY ZADOR 3540 02:21:26,545 --> 02:21:35,454 IN THEIR WORK ON THE EMBODIED TURING 3541 02:21:35,454 --> 02:21:40,325 TEST MENTIONED AS ANIMAL-LEVEL COMMON 3542 02:21:40,325 --> 02:21:42,327 SENSE SENSORIMOTOR INTELLIGENCE. 3543 02:21:42,327 --> 02:21:45,597 ONCE WE HAVE THAT, WE CAN USE 3544 02:21:45,597 --> 02:21:49,735 NEUROMORPHIC TECHNOLOGY TO 3545 02:21:49,735 --> 02:21:52,037 IMPLEMENT THIS, THEN THE OTHER 3546 02:21:52,037 --> 02:21:54,106 TASKS THAT ARE MORE 3547 02:21:54,106 --> 02:21:58,977 HUMAN-LIKE IN TERMS OF ABSTRACT 3548 02:21:58,977 --> 02:22:00,212 REASONING, LANGUAGE, THAT WOULD 3549 02:22:00,212 --> 02:22:04,116 BE ON TOP OF THIS SENSORIMOTOR 3550 02:22:04,116 --> 02:22:04,750 INTELLIGENCE. 3551 02:22:04,750 --> 02:22:08,920 WHAT DO WE NEED TO DO FOR INTERACTING 3552 02:22:08,920 --> 02:22:12,524 WITH REAL WORLD IN REAL TIME? 3553 02:22:12,524 --> 02:22:15,460 THAT'S WHAT ANIMALS DO, THEY GATHER 3554 02:22:15,460 --> 02:22:17,462 INFORMATION FROM THE EXTERNAL WORLD 3555 02:22:17,462 --> 02:22:20,332 AND THAT MEANS INTERPRET THE SENSORY 3556 02:22:20,332 --> 02:22:21,466 INFORMATION TO MAKE A DECISION 3557 02:22:21,466 --> 02:22:25,003 ON WHAT TO DO, WHICH ACTION TO 3558 02:22:25,003 --> 02:22:27,305 PERFORM IN ORDER TO ACHIEVE 3559 02:22:27,305 --> 02:22:28,407 THEIR OWN GOAL. 3560 02:22:28,407 --> 02:22:30,475 THEN THEY NEED TO CONTROL THEIR 3561 02:22:30,475 --> 02:22:34,312 BODY AND THERE WAS A BEAUTIFUL 3562 02:22:34,312 --> 02:22:36,014 EXAMPLE OF WHAT MITRA WAS 3563 02:22:36,014 --> 02:22:37,582 TALKING ABOUT TODAY. SO IF WE USE 3564 02:22:37,582 --> 02:22:45,090 NEUROCOMPUTATIONAL PRIMITIVES AND 3565 02:22:45,090 --> 02:22:46,291 NEUROMORPHIC, THE IDEA THAT WE HAVE 3566 02:22:46,291 --> 02:22:53,832 A SYSTEM THAT CAN INHERITS THE 3567 02:22:53,832 --> 02:22:57,969 QUALITIES OF BIOLOGY SYSTEMS TO 3568 02:22:57,969 --> 02:23:00,071 BE MORE EFFICIENT IN THE USE OF 3569 02:23:00,071 --> 02:23:00,806 RESOURCES. 3570 02:23:00,806 --> 02:23:03,308 SO IN THIS PRESENTATION, I'M 3571 02:23:03,308 --> 02:23:06,845 GOING TO SKIM THROUGH A FEW 3572 02:23:06,845 --> 02:23:11,550 EXAMPLES AND I WILL GO VERY 3573 02:23:11,550 --> 02:23:13,418 FAST, OF WHAT WE DID IN OUR LAB 3574 02:23:13,418 --> 02:23:16,488 IN FOUR DIFFERENT DOMAINS. 3575 02:23:16,488 --> 02:23:20,125 WE DON'T WANT TO REPRESENT A 3576 02:23:20,125 --> 02:23:27,866 FULL STATE-OF-THE-ART BUT IN 3577 02:23:27,866 --> 02:23:33,338 TERMS OF SENSING, I WAS HAPPY 3578 02:23:33,338 --> 02:23:35,874 YESTERDAY TO HERE ABOUT TOUCH 3579 02:23:35,874 --> 02:23:40,679 IT'S A SENSORY MODALITY THAT IS 3580 02:23:40,679 --> 02:23:48,954 DISCARDED BY MOST OF THE PEOPLE 3581 02:23:48,954 --> 02:23:49,287 IN VISION, BUT IT IS IMPORTANT. 3582 02:23:49,287 --> 02:23:52,591 WHAT WE WANT TO DO HERE WAS TO 3583 02:23:52,591 --> 02:24:00,265 DEVELOP CIRCUITS FOR ENCODING 3584 02:24:00,265 --> 02:24:06,471 IN THE HUMAN SKIN. WHAT YOU SEE 3585 02:24:06,471 --> 02:24:09,374 HERE IS A TEXTBOOK DESCRIPTION 3586 02:24:09,374 --> 02:24:14,846 OF MECHANORECEPTORS IN THE SKIN. 3587 02:24:14,846 --> 02:24:22,988 THEY ARE EVENT DRIVEN, PRODUCE SPIKES 3588 02:24:22,988 --> 02:24:29,728 UPON CONTACT. WE WANTED TO DO THIS 3589 02:24:29,728 --> 02:24:31,463 IN FAST, ADAPTIVE TYPE OF CELLS THAT 3590 02:24:31,463 --> 02:24:32,898 RESPOND TO THE CHANGE IN PRESSURE. 3591 02:24:32,898 --> 02:24:37,402 WE DID THAT WITH CMOS CIRCUITS 3592 02:24:37,402 --> 02:24:39,404 AND INTERFACE WITH PIEZOELECTRIC 3593 02:24:39,404 --> 02:24:42,641 SENSORS IN A WAY THAT IS SIMILAR 3594 02:24:42,641 --> 02:24:45,911 TO EVENT-DRIVEN SENSORS THAT MOST 3595 02:24:45,911 --> 02:24:49,047 OF US HAVE HEARD ABOUT. 3596 02:24:49,047 --> 02:24:51,783 HERE WE SHOW THE RESULTS WHERE 3597 02:24:51,783 --> 02:24:57,489 IF WE APPLY A CHANGE IN FORCE, 3598 02:24:57,489 --> 02:25:01,960 WE HAVE A DIFFERENT BEHAVIOR 3599 02:25:01,960 --> 02:25:03,595 OF DIFFERENT NEURONS THAT FIRE 3600 02:25:03,595 --> 02:25:04,462 ACTION POTENTIALS. 3601 02:25:04,462 --> 02:25:06,665 THIS IS TO SHOW YOU THAT THIS 3602 02:25:06,665 --> 02:25:09,100 WORKS AND WHEN THERE IS A CHANGE 3603 02:25:09,100 --> 02:25:11,670 IN THE OUTPUT, THERE IS A CHANGE 3604 02:25:11,670 --> 02:25:16,074 IN THE INSTANTANEOUS FIRING RATE. 3605 02:25:16,074 --> 02:25:18,410 SO THIS IS THE SENSING PART, 3606 02:25:18,410 --> 02:25:19,377 HOW WE GATHER INFORMATION. 3607 02:25:19,377 --> 02:25:23,481 IN TERMS OF PERCEPTION, IT'S 3608 02:25:23,481 --> 02:25:27,619 NICE BECAUSE WE BASICALLY 3609 02:25:27,619 --> 02:25:36,027 IMPLEMENT A MODEL THAT WAS CITED 3610 02:25:36,027 --> 02:25:40,865 YESTERDAY. IT'S THE ORIGINAL 3611 02:25:40,865 --> 02:25:41,967 ITTI-KOCK MODEL OF SELECTIVE 3612 02:25:41,967 --> 02:25:43,001 ATTENTION THAT SELECTS WHAT'S 3613 02:25:43,001 --> 02:25:48,607 INTERESTING IN THE VISUAL FIELD. 3614 02:25:48,607 --> 02:25:52,877 THIS IS ITTI-KOCK 2.0 BY ERNST 3615 02:25:52,877 --> 02:25:54,746 NIEBUR, HE USES BORDER OWNERSHIP AND 3616 02:25:54,746 --> 02:25:57,549 GROUPING CELLS TO USE THE CONCEPT 3617 02:25:57,549 --> 02:26:04,022 OF GESTALT OBJECTS OR PROTO-OBJECTS 3618 02:26:04,022 --> 02:26:06,358 TO ATTRACT ATTENTION. NOT ONLY IN 3619 02:26:06,358 --> 02:26:08,426 FEATURE SPACE THERE IS SOMETHING THAT 3620 02:26:08,426 --> 02:26:11,029 CAN ATTRACT ATTENTION, BUT ALSO, IF 3621 02:26:11,029 --> 02:26:12,998 THERE IS SOMETHING THAT COULD 3622 02:26:12,998 --> 02:26:14,566 REPRESENT AN OBJECT. 3623 02:26:14,566 --> 02:26:21,473 WE USED LEAKY INTEGRATE-AND FIRE 3624 02:26:21,473 --> 02:26:24,075 NEURONS AND SPIKING NETWORKS TO 3625 02:26:24,075 --> 02:26:25,877 IMPLEMENT THIS MODEL USING EVENT- 3626 02:26:25,877 --> 02:26:31,016 BASED VISION SENSORS. WE ALSO USED 3627 02:26:31,016 --> 02:26:40,759 COMPETITIVE STEREO MATCHING TO ADD 3628 02:26:40,759 --> 02:26:41,660 FEATURE SPACE FOR DEPTH PERCEPTION 3629 02:26:41,660 --> 02:26:45,830 SO AN OBJECT CLOSE TO THE CAMERA OR 3630 02:26:45,830 --> 02:26:48,833 ROBOT IS SOMETHING INTERESTING. 3631 02:26:48,833 --> 02:26:51,102 LONG STORY SHORT: THIS IS THE VIDEO 3632 02:26:51,102 --> 02:26:56,441 AND THE ROBOT CAN FIND THE 3633 02:26:56,441 --> 02:27:01,212 MOST INTERESTING PART IN THE 3634 02:27:01,212 --> 02:27:01,946 VISUAL INPUT. 3635 02:27:01,946 --> 02:27:04,049 I DON'T THINK I HAVE MUCH TIME, 3636 02:27:04,049 --> 02:27:05,717 BUT BASICALLY WE WORKED ON 3637 02:27:05,717 --> 02:27:08,653 CONTROL. 3638 02:27:08,653 --> 02:27:11,790 AND HERE WE WANTED TO SOLVE THE 3639 02:27:11,790 --> 02:27:13,958 PROBLEM OF INVERSE KINEMATICS. 3640 02:27:13,958 --> 02:27:16,261 WHEN THERE IS SOMETHING IN THE 3641 02:27:16,261 --> 02:27:18,296 SPACE, WHAT IS THE CONFIGURATION 3642 02:27:18,296 --> 02:27:18,830 OR PROPRIOCEPTION OF THE ROBOT 3643 02:27:18,830 --> 02:27:21,666 TO GET TO THAT POSITION. 3644 02:27:21,666 --> 02:27:22,701 WHAT I WANTED TO SHOW YOU HERE 3645 02:27:22,701 --> 02:27:25,337 IS THAT WHEN WE WORK IN ROBOTICS, 3646 02:27:25,337 --> 02:27:28,206 WE KEEP EXCHANGING INFORMATION 3647 02:27:28,206 --> 02:27:30,475 BETWEEN THE CLOCK-BASED DOMAIN 3648 02:27:30,475 --> 02:27:34,179 NON-NEUROMORPHIC AND THE 3649 02:27:34,179 --> 02:27:35,547 SPIKING DOMAIN THAT IS 3650 02:27:35,547 --> 02:27:38,516 NEUROMORPHIC. WE BUILD THIS SMALL 3651 02:27:38,516 --> 02:27:41,352 MODULE THAT HAS TO COMMUNICATE WITH 3652 02:27:41,352 --> 02:27:50,128 THE ROBOT THAT IS NOT NEUROMORPHIC. 3653 02:27:50,128 --> 02:27:53,732 IMPLEMENTING REINFORCEMENT 3654 02:27:53,732 --> 02:27:59,471 LEARNING WITH SPATIAL SEMANTIC 3655 02:27:59,471 --> 02:28:07,112 POINTERS FOR DECIDING WHEN THE 3656 02:28:07,112 --> 02:28:09,881 ROBOT HAD TO MOVE TO HIT DYNAMIC 3657 02:28:09,881 --> 02:28:12,217 TASK - HIERARCHICAL TASK. WE SHOW 3658 02:28:12,217 --> 02:28:13,651 THAT DEPENDING ON INITIAL VELOCITY 3659 02:28:13,651 --> 02:28:16,721 OF THE TASK, THE ROBOT IS 3660 02:28:16,721 --> 02:28:19,357 CAPABLE OF WAITING UNTIL THE 3661 02:28:19,357 --> 02:28:21,192 THE RIGHT TIME TO MOVE AND HIT 3662 02:28:21,192 --> 02:28:21,893 THE TARGET. 3663 02:28:21,893 --> 02:28:28,566 THESE ARE ALL SMALL THINGS IN 3664 02:28:28,566 --> 02:28:28,967 THE FOUR DOMAINS. 3665 02:28:28,967 --> 02:28:30,468 WHAT I THINK IS IMPORTANT IS 3666 02:28:30,468 --> 02:28:32,170 THAT WE PUT EVERYTHING TOGETHER. 3667 02:28:32,170 --> 02:28:37,208 WE HAVE END-TO-END SPIKING SYSTEM 3668 02:28:37,208 --> 02:28:37,876 WITH NEUROMORPHIC TECHNOLOGY TO 3669 02:28:37,876 --> 02:28:40,612 HAVE THE FULL STACK OF 3670 02:28:40,612 --> 02:28:42,347 BEHAVIOR BECAUSE IT'S COSTLY TO 3671 02:28:42,347 --> 02:28:44,582 MOVE FROM ONE DOMAIN TO THE 3672 02:28:44,582 --> 02:28:48,920 OTHER EVERYTIME AND OF COURSE WE 3673 02:28:48,920 --> 02:28:51,990 NEED TO SCALE UP THE TYPE OF 3674 02:28:51,990 --> 02:28:54,292 TASK WE ARE TRYING TO SOLVE HERE. 3675 02:28:54,292 --> 02:28:56,461 A COUPLE OF WORDS ON WHAT IS 3676 02:28:56,461 --> 02:28:58,563 NEXT. I THINK WE NEED TO 3677 02:28:58,563 --> 02:29:01,132 TAKE THE EMBODIMENT KEYWORD 3678 02:29:01,132 --> 02:29:01,833 TO THE NEXT LEVEL. 3679 02:29:01,833 --> 02:29:04,369 IT'S NOT ONLY A SENSORIMOTOR 3680 02:29:04,369 --> 02:29:09,941 LOOP, BU BUT WE NEED TO LOOK INO 3681 02:29:09,941 --> 02:29:13,311 ACTIONS THAT ARE PURPUSEFUL TO GATHER 3682 02:29:13,311 --> 02:29:19,751 INFORMATION. NOT ONLY OBSERVING 3683 02:29:19,751 --> 02:29:22,854 SOMETHING AND THEN PLAN THE ACTION 3684 02:29:22,854 --> 02:29:25,190 BUT ALSO PLANNING THE ACTION 3685 02:29:25,190 --> 02:29:27,826 TO OBSERVE SOMETHING. THIS PICTURE 3686 02:29:27,826 --> 02:29:31,463 OF THE ICUB ROBOT IS INTERESTING. 3687 02:29:31,463 --> 02:29:33,231 WHEN YOU NEED TO CLASSIFY SOMETHING 3688 02:29:33,231 --> 02:29:35,233 YOU CAN TAKE IT, MOVE IT, USE YOU 3689 02:29:35,233 --> 02:29:37,235 INFORMATION GATHERED FROM 3690 02:29:37,235 --> 02:29:39,204 DIFFERENT SENSORY MODALITIES TO 3691 02:29:39,204 --> 02:29:41,172 GATHER INFORMATION ON WHAT 3692 02:29:41,172 --> 02:29:43,241 YOU'RE HE DOING IN RELATION TO 3693 02:29:43,241 --> 02:29:45,810 YOUR OWN BODY. THIS CAN SIMPLIFY 3694 02:29:45,810 --> 02:29:49,147 THE PROBLEM OF CLASSIFICATION. 3695 02:29:49,147 --> 02:29:53,751 WHILE IN TRADITIONAL A.I. SYSTEMS, 3696 02:29:53,751 --> 02:29:56,287 YOU GET TONS OF DATA THAT YOU 3697 02:29:56,287 --> 02:29:57,789 DON'T KNOW HOW THEY ARE EVEN 3698 02:29:57,789 --> 02:29:58,990 ACQUIRED. BUT HERE YOU CAN USE YOUR 3699 02:29:58,990 --> 02:30:02,293 OWN ACTIONS TO ACQUIRE DATA AND 3700 02:30:02,293 --> 02:30:04,262 PROCESS THEM. THIS IS ACTIVE SENSING. 3701 02:30:04,262 --> 02:30:06,030 WE CAN ALSO THINK ABOUT 3702 02:30:06,030 --> 02:30:16,541 MORPHOLOGICAL COMPUTATION, POSITION 3703 02:30:16,741 --> 02:30:18,409 OF SENSORS CAN SIMPLIFY COMPUTATION. 3704 02:30:18,409 --> 02:30:21,246 ATTENTION ALLOWS YOU TO CALCULATE 3705 02:30:21,246 --> 02:30:23,181 WITH FULL RATE OF VISION, USE HIGHER 3706 02:30:23,181 --> 02:30:26,284 RESOLUTION IN THE CENTER AND LOWER 3707 02:30:26,284 --> 02:30:27,819 IN PERIPHERY TO MAKE COMPUTATION MORE 3708 02:30:27,819 --> 02:30:29,854 EFFICIENT AND NOT HAVING TO COMPUTE 3709 02:30:29,854 --> 02:30:35,660 A LOT OF DATA AT THE SAME TIME. BUT 3710 02:30:35,660 --> 02:30:38,663 THIS HAS TO BE COUPLED TO ATTENTION 3711 02:30:38,663 --> 02:30:45,003 TO WORK. WE SHOULD MOVE TO OTHER 3712 02:30:45,003 --> 02:30:46,671 PLATFORMS, WHERE THE BODY IS 3713 02:30:46,671 --> 02:30:50,808 PERFORMING COMPUTATION AND COUPLE 3714 02:30:50,808 --> 02:30:53,745 IT WITH THE NEUROMORPHIC COMPUTATION 3715 02:30:53,745 --> 02:31:00,451 TO HAVE REAL EFFICIENT SYSTEMS. 3716 02:31:00,451 --> 02:31:02,220 THESE ARE THE PEOPLE THAT DID THE 3717 02:31:02,220 --> 02:31:04,789 WORK AND I'M FUNDED BY EUROPEAN 3718 02:31:04,789 --> 02:31:09,928 AGENCIES. 3719 02:31:09,928 --> 02:31:12,363 >> THANK YOU SO MUCH. 3720 02:31:12,363 --> 02:31:18,736 [APPLAUSE] 3721 02:31:18,736 --> 02:31:25,009 IT'S GREAT -- THANK YOU SO MUCH 3722 02:31:25,009 --> 02:31:27,845 FOR THAT INFORMATION ON THE 3723 02:31:27,845 --> 02:31:33,584 EMBODIED COMPUTATION SETTING IT APART 3724 02:31:33,584 --> 02:31:36,287 FROM DISEMBODIED COMPUTATION. 3725 02:31:36,287 --> 02:31:44,796 WE'LL NOW BE JOINED BY JOE HAYS. 3726 02:31:44,796 --> 02:31:49,100 >> I'M JOE HAYS FROM THE 3727 02:31:49,100 --> 02:31:51,302 NAVAL RESEARCH LABORATORY. 3728 02:31:51,302 --> 02:31:53,738 GIVEN MY ATTENDANCE HERE, THERE 3729 02:31:53,738 --> 02:31:57,375 ARE A NUMBER OF THINGS I'LL BE 3730 02:31:57,375 --> 02:31:58,810 MESSAGES THAT YOU'VE ALREADY 3731 02:31:58,810 --> 02:31:59,110 HEARD. 3732 02:31:59,110 --> 02:32:01,512 BUT I THINK MY PRESENTATION WILL 3733 02:32:01,512 --> 02:32:03,047 BE A LITTLE BIT DIFFERENT 3734 02:32:03,047 --> 02:32:05,116 BECAUSE I'M GOING TO FOCUS ON 3735 02:32:05,116 --> 02:32:07,118 THE APPLICATION, I WON'T TALK 3736 02:32:07,118 --> 02:32:09,354 TOO MUCH ABOUT THE DETAILS OF 3737 02:32:09,354 --> 02:32:15,326 NEUROSCIENCE OR THE NEUROMORPHIC 3738 02:32:15,326 --> 02:32:16,294 IMPLEMENTATION. 3739 02:32:16,294 --> 02:32:19,731 HOPEFULLY THERE IS VALUE HERE. 3740 02:32:19,731 --> 02:32:22,567 OKAY, SO I COME FROM THE 3741 02:32:22,567 --> 02:32:24,268 SPACECRAFT ENGINEERING DIVISION 3742 02:32:24,268 --> 02:32:27,238 AT NRL. 3743 02:32:27,238 --> 02:32:30,074 OUR GROUP HAS DONE A LOT. 3744 02:32:30,074 --> 02:32:32,276 WE ONLY BUILD ONE-OF-A-KIND 3745 02:32:32,276 --> 02:32:33,211 SYSTEMS. 3746 02:32:33,211 --> 02:32:35,747 WE LAUNCHED OVER A HUNDRED 3747 02:32:35,747 --> 02:32:39,984 SATELLITES INTO ORBIT AND 3748 02:32:39,984 --> 02:32:40,218 EACH ONE IS UNIQUE. 3749 02:32:40,218 --> 02:32:44,222 WE HAVE A ROADMAP FOR OUR 3750 02:32:44,222 --> 02:32:46,224 ROBOTICS WORK. 3751 02:32:46,224 --> 02:32:48,526 THE LAST TWO DECADES OUR 3752 02:32:48,526 --> 02:32:49,927 DIVISION HAS BEEN FOCUSING ON 3753 02:32:49,927 --> 02:32:54,499 THE ABILITY TO ROBOTICALLY SERVICE 3754 02:32:54,499 --> 02:32:57,168 A SATELLITE WHILE IT'S IN ORBIT. 3755 02:32:57,168 --> 02:32:59,404 THE NAVY HAS A MAINTENANCE PLAN 3756 02:32:59,404 --> 02:33:02,407 FOR EVERYTHING THEY BUY EXCEPT 3757 02:33:02,407 --> 02:33:04,075 FOR SATELLITES. 3758 02:33:04,075 --> 02:33:05,843 SOME OF THE SATELLITES ARE A 3759 02:33:05,843 --> 02:33:06,744 BILLION DOLLARS. 3760 02:33:06,744 --> 02:33:08,946 TO NOT HAVE A MAINTENANCE PLAN 3761 02:33:08,946 --> 02:33:11,249 IS NOT A GOOD PLAN. 3762 02:33:11,249 --> 02:33:12,884 WE'VE BEEN DEVELOPING THE 3763 02:33:12,884 --> 02:33:13,685 CAPABILITY TO PERFORM MAINTENANCE 3764 02:33:13,685 --> 02:33:15,253 ON THESE SYSTEMS WHILE IN ORBIT. 3765 02:33:15,253 --> 02:33:16,187 THIS IS USING OLD TECHNOLOGY. 3766 02:33:16,187 --> 02:33:21,659 NO NEUROMORPHICS OR MACHINE LEARNING. 3767 02:33:21,659 --> 02:33:23,628 WE ARE SPRINGBOARDING OFF OF 3768 02:33:23,628 --> 02:33:25,797 THIS CAPABILITY AND I'LL TALK 3769 02:33:25,797 --> 02:33:26,998 ABOUT THAT. 3770 02:33:26,998 --> 02:33:29,033 BUT THIS SATELLITE WAS JUST 3771 02:33:29,033 --> 02:33:32,036 DELIVERED TO THE LAUNCH VEHICLE 3772 02:33:32,036 --> 02:33:35,073 AND IN ABOUT 12 MONTHS, IT 3773 02:33:35,073 --> 02:33:36,307 SHOULD BE LAUNCHED INTO ORBIT 3774 02:33:36,307 --> 02:33:38,443 AND FOR THE FIRST TIME WE'LL 3775 02:33:38,443 --> 02:33:42,647 SHOW THE WORLD HOW TO SERVICE A 3776 02:33:42,647 --> 02:33:44,082 SATELLITE WHILE IN ORBIT. 3777 02:33:44,082 --> 02:33:46,484 THERE ARE A NUMBER OF THINGS THE 3778 02:33:46,484 --> 02:33:48,519 SATELLITE CAN DO BUT I'M NOT 3779 02:33:48,519 --> 02:33:50,621 GOING TO SPEND TIME ON THAT. 3780 02:33:50,621 --> 02:33:55,393 I AM GOING TO TALK ABOUT HOW 3781 02:33:55,393 --> 02:33:56,494 WE'RE SPRINGBOARDING OFF OF THAT 3782 02:33:56,494 --> 02:34:00,631 WORK AND BUILDING NEUROMORPHICS 3783 02:34:00,631 --> 02:34:02,967 INTO THE ROBOTICS EFFORTS. 3784 02:34:02,967 --> 02:34:06,504 ONE AREA WE SPENT A LOT OF TIME 3785 02:34:06,504 --> 02:34:08,506 ON IS INVESTIGATING HOW TO 3786 02:34:08,506 --> 02:34:10,208 PERFORM MAINTENANCE ON A SHIP 3787 02:34:10,208 --> 02:34:11,809 WITH ROBOTS. 3788 02:34:11,809 --> 02:34:20,318 WE TAKE LIVE -- SOLDIERS OR 3789 02:34:20,318 --> 02:34:23,554 SAILORS OUT OF HARM'S WAY. 3790 02:34:23,554 --> 02:34:27,558 LIKEWISE WE LOOK TO CHANGE THE 3791 02:34:27,558 --> 02:34:30,595 PARADIGM ON HOW SATELLITES ARE 3792 02:34:30,595 --> 02:34:30,795 BUILT. AS SOON AS THE JAMES 3793 02:34:30,795 --> 02:34:33,397 WEB TELECOPE LARGEST SATELLITE 3794 02:34:33,397 --> 02:34:36,868 WAS PUT INTO SPACE, PEOPLE WERE 3795 02:34:36,868 --> 02:34:41,139 ASKING FOR A BIGGER VERSION. 3796 02:34:41,139 --> 02:34:42,940 AND WE'RE NOW INVESTIGATING TO 3797 02:34:42,940 --> 02:34:45,076 BUILD THOSE ROBOTICALLY IN SPACE 3798 02:34:45,076 --> 02:34:49,046 WHERE THEY'RE TOO BIG TO MASS 3799 02:34:49,046 --> 02:34:50,181 LIFT OFF OF THE EARTH. 3800 02:34:50,181 --> 02:34:53,818 YOU SEND UP AN AMAZON TRUCK AND 3801 02:34:53,818 --> 02:34:59,257 YOU HAVE A ROBOT IN ORBIT 3802 02:34:59,257 --> 02:35:01,325 CONSTRUCTING A STRUCTURE THAT IS 3803 02:35:01,325 --> 02:35:03,928 MUCH LARGER THAN IT COULD BE 3804 02:35:03,928 --> 02:35:05,163 BUILT AND LAUNCHED OFF THE 3805 02:35:05,163 --> 02:35:08,733 EARTH. 3806 02:35:08,733 --> 02:35:14,205 WE ARE VERY INVESTED IN OUR 3807 02:35:14,205 --> 02:35:17,074 RESEARCH RELATED TO CONTINUAL 3808 02:35:17,074 --> 02:35:19,911 LEARNING AGENTS AND WE HAVE TO 3809 02:35:19,911 --> 02:35:25,149 FOCUS ON POWER AND THEREFORE THE 3810 02:35:25,149 --> 02:35:28,719 EDGE INTELLIGENCE BEING DESIGNED 3811 02:35:28,719 --> 02:35:30,354 INTO THESE SYSTEMS IS VERY 3812 02:35:30,354 --> 02:35:31,322 IMPORTANT. 3813 02:35:31,322 --> 02:35:32,957 IN SPACE, POWER IS KEY. 3814 02:35:32,957 --> 02:35:35,593 MOST OF YOUR DECISIONS ARE MADE 3815 02:35:35,593 --> 02:35:37,195 OFF OF WHAT POWER IS AVAILABLE 3816 02:35:37,195 --> 02:35:40,198 TO YOU. 3817 02:35:40,198 --> 02:35:44,001 LIKEWISE, OUR SHIP MECHANIC ALSO 3818 02:35:44,001 --> 02:35:45,970 HAS LIMITED -- ACCESS TO LIMITED 3819 02:35:45,970 --> 02:35:47,104 POWER. 3820 02:35:47,104 --> 02:35:49,373 AS WE PUSH MORE AND MORE 3821 02:35:49,373 --> 02:35:51,542 INTELLIGENCE CLOSER TO THE 3822 02:35:51,542 --> 02:35:54,245 SENSOR AND ACTUATOR, THAT'S WHAT 3823 02:35:54,245 --> 02:35:56,747 WE CALL EDGE INTELLIGENCE, AS WE 3824 02:35:56,747 --> 02:36:00,284 PUSH THAT MORE AND MORE OF THAT 3825 02:36:00,284 --> 02:36:01,586 CAPABILITY TOWARDS THE SENSORS 3826 02:36:01,586 --> 02:36:04,188 AND ACTUATORS USING TRADITIONAL 3827 02:36:04,188 --> 02:36:06,457 COMPUTING, IT MAKES OUR POWER 3828 02:36:06,457 --> 02:36:08,092 CONSUMPTION AND OUR SIZE, WEIGHT 3829 02:36:08,092 --> 02:36:10,127 AND POWER OF OUR SYSTEMS GO UP. 3830 02:36:10,127 --> 02:36:11,462 THAT'S THE WRONG DIRECTION. 3831 02:36:11,462 --> 02:36:12,964 WE NEED TO GO DOWN. 3832 02:36:12,964 --> 02:36:16,133 SO WE'RE INVESTING A LOT OF TIME 3833 02:36:16,133 --> 02:36:19,503 AND ENERGY INTO NEUROMORPHICS 3834 02:36:19,503 --> 02:36:20,605 BECAUSE WE BELIEVE THAT TO BE 3835 02:36:20,605 --> 02:36:23,341 OUR BEST CANDIDATE TO BE ABLE TO 3836 02:36:23,341 --> 02:36:27,311 REALIZE EDGE INTELLIGENCE, BUT 3837 02:36:27,311 --> 02:36:37,889 STILL REDUCE OUR SWAP. 3838 02:36:37,889 --> 02:36:39,557 NEUROMORPHIC IS AN OVERLOADED 3839 02:36:39,557 --> 02:36:40,258 TERM. THERE IS NEUROMORPHIC 3840 02:36:40,258 --> 02:36:42,627 COMPUTATION AND CHIARA JUST 3841 02:36:42,627 --> 02:36:46,230 DID A GREAT JOB OF INTRODUCING 3842 02:36:46,230 --> 02:36:48,666 NEUROMORPHIC SENSING. 3843 02:36:48,666 --> 02:36:50,001 THEY ARE DISTINCT AND 3844 02:36:50,001 --> 02:36:52,436 DIFFERENT BUT HIGHLY RELATED. 3845 02:36:52,436 --> 02:36:55,773 THERE ARE MANY PEOPLE WHO DO 3846 02:36:55,773 --> 02:36:57,608 NEUROMORPHIC SENSING WITHOUT 3847 02:36:57,608 --> 02:37:00,278 COMPUTING AND VICE VERSA. 3848 02:37:00,278 --> 02:37:02,947 FROM A RESEARCH PERSPECTIVE, 3849 02:37:02,947 --> 02:37:04,749 OUR GROUP IS ENGAGED IN 3850 02:37:04,749 --> 02:37:05,616 A LOT OF THINGS. 3851 02:37:05,616 --> 02:37:10,087 I'M NOT GOING TO FOCUS ON ANY 3852 02:37:10,087 --> 02:37:13,824 RESEARCH TOPIC, BUT WHEN WE DO 3853 02:37:13,824 --> 02:37:16,827 BASIC RESEARCH, IT'S KNOWN AS 3854 02:37:16,827 --> 02:37:17,561 6-1 RESEARCH. 3855 02:37:17,561 --> 02:37:23,834 WE DO A LOT OF 6-1 RESEARCH IN 3856 02:37:23,834 --> 02:37:25,036 NEUROMORPHICS AND CONTINUAL LEARNING. 3857 02:37:25,036 --> 02:37:28,406 WE TRY TO BRING THAT BASIC RESEARCH 3858 02:37:28,406 --> 02:37:30,574 TO 6-2 RESEARCH AS WE'VE BEEN SHOWN 3859 02:37:30,574 --> 02:37:32,143 HERE. 3860 02:37:32,143 --> 02:37:33,911 AGAIN, CONTINUAL LEARNING. 3861 02:37:33,911 --> 02:37:36,213 WE'VE A TALKED ABOUT HERE IN THE 3862 02:37:36,213 --> 02:37:42,053 LAST DAY OR TWO. 3863 02:37:42,053 --> 02:37:45,056 SIMPLY DEFINED, IT'S THE ABILITY 3864 02:37:45,056 --> 02:37:46,590 TO INTEGRATE ANY CAPABILITIES 3865 02:37:46,590 --> 02:37:50,127 AND KNOWLEDGE WITHOUT LOSING 3866 02:37:50,127 --> 02:37:54,432 PREVIOUSLY LEARNED CAPABILITIES 3867 02:37:54,432 --> 02:37:54,765 AND KNOWLEDGE. 3868 02:37:54,765 --> 02:37:57,001 I LIKE TO POINT OUT IT'S DONE IN 3869 02:37:57,001 --> 02:37:58,436 A SINGLE LIFETIME. 3870 02:37:58,436 --> 02:38:02,106 WHEN WE TRAIN AGENTS TODAY, WE 3871 02:38:02,106 --> 02:38:06,944 DO IT IN BATCH MODE. TAKE A GRADIENT 3872 02:38:06,944 --> 02:38:09,246 STEP, RESET, COLLECT DATA, BATCH IT, TAKE ANOTHER STEP. 3873 02:38:09,246 --> 02:38:11,749 TRUE CONTINUAL LIFELONG LEARNING 3874 02:38:11,749 --> 02:38:19,890 DOES NOT BEHAVE THAT WAY. 3875 02:38:19,890 --> 02:38:22,793 I WOULD LIKE TO GIVE A SHOUT OUT 3876 02:38:22,793 --> 02:38:27,832 TO DHIREESHA. SHE LED A TEAM 3877 02:38:27,832 --> 02:38:30,468 THAT PUT OUT A GREAT 3878 02:38:30,468 --> 02:38:31,836 PAPER. HIGHLY RECOMMEND IT. 3879 02:38:31,836 --> 02:38:35,006 THEY DOCUMENT A NUMBER OF 3880 02:38:35,006 --> 02:38:36,273 BIOLOGICAL MECHANISMS THAT ARE 3881 02:38:36,273 --> 02:38:38,376 BELIEVED TO BE INVOLVED IN 3882 02:38:38,376 --> 02:38:39,176 CONTINUAL LEARNING AS WELL AS 3883 02:38:39,176 --> 02:38:41,645 THE KEY INDICATORS THAT YOU CAN 3884 02:38:41,645 --> 02:38:45,683 SEE ALONG THE TOP THAT DEFINE 3885 02:38:45,683 --> 02:38:46,884 CONTINUAL LEARNING AND 3886 02:38:46,884 --> 02:38:52,023 THE INTERSECTION OF THOSE 3887 02:38:52,023 --> 02:38:54,191 MECHANISMS WITH KEY INDICATORS 3888 02:38:54,191 --> 02:38:58,896 OR FEATURES. 3889 02:38:58,896 --> 02:39:03,267 AS A GROUP, WE ARE INVESTED IN 3890 02:39:03,267 --> 02:39:07,004 INVESTIGATING A NUMBER OF THESE 3891 02:39:07,004 --> 02:39:07,605 MECHANISMS. 3892 02:39:07,605 --> 02:39:11,509 AS A HIGH LEVEL, WHY DO WE CARE 3893 02:39:11,509 --> 02:39:13,110 ABOUT CONTINUAL LEARNING? 3894 02:39:13,110 --> 02:39:16,881 WE HAVE AGENTS LIKE A ROBOTIC 3895 02:39:16,881 --> 02:39:18,315 DOG OR MAINTENANCE ON THE SHIP, 3896 02:39:18,315 --> 02:39:20,184 WE DON'T HAVE TIME TO BRING IT 3897 02:39:20,184 --> 02:39:21,819 BACK INTO THE LAB TO TEACH 3898 02:39:21,819 --> 02:39:24,855 IT THE NEW TOOL OR TEACH IT A 3899 02:39:24,855 --> 02:39:26,257 NEW OBJECT THAT IT HAS TO 3900 02:39:26,257 --> 02:39:26,991 IDENTIFY. 3901 02:39:26,991 --> 02:39:30,561 WE NEED TO ALLOW A NON-PHD SAILOR 3902 02:39:30,561 --> 02:39:34,532 TO BE ABLE TO POINT A NEW OBJECT 3903 02:39:34,532 --> 02:39:38,169 OR TOOL AT OUR ROBOT AND LEARN 3904 02:39:38,169 --> 02:39:40,304 HOW TO IDENTIFY OR MANIPULATE IT 3905 02:39:40,304 --> 02:39:41,939 WITHOUT BRINGING IT BACK TO THE LAB 3906 02:39:41,939 --> 02:39:47,645 TO TEACH IT THIS NEW THING. 3907 02:39:47,645 --> 02:39:51,549 THOSE ARE CAPABILITIES THAT WE 3908 02:39:51,549 --> 02:39:53,551 NEED DRASTICALLY AND HENCE 3909 02:39:53,551 --> 02:39:54,151 MOTIVATION FOR INVESTIGATING THOSE 3910 02:39:54,151 --> 02:39:57,288 CAPABILITIES. 3911 02:39:57,288 --> 02:39:59,089 AGAIN, WE'VE HAD A LOT OF 3912 02:39:59,089 --> 02:40:01,125 DISCUSSION ABOUT EMBODIMENT 3913 02:40:01,125 --> 02:40:04,161 WHICH HAS MADE ME VERY HAPPY THE 3914 02:40:04,161 --> 02:40:05,596 LAST COUPLE OF DAYS. 3915 02:40:05,596 --> 02:40:09,500 I GO TO A LOT OF 3916 02:40:09,500 --> 02:40:11,969 MACHINE LEARNING CONFERENCES 3917 02:40:11,969 --> 02:40:19,343 AND THERE IS NO DISCUSSION ON 3918 02:40:19,343 --> 02:40:21,412 EMBODIMENT. WHEN WE GROUND ALGORITHMS 3919 02:40:21,412 --> 02:40:23,914 IN CONSTRAINTS OF OUR PLATFORM, WE 3920 02:40:23,914 --> 02:40:26,817 FACE ISSUES WE NEED TO SOLVE TO MAKE 3921 02:40:26,817 --> 02:40:30,488 TRUE CONTINUOUSLY LEARNING AGENTS. 3922 02:40:30,488 --> 02:40:34,792 AND JUST LIKE EVERYONE ELSE, 3923 02:40:34,792 --> 02:40:36,227 I'LL MESSAGE THE SAME THING. 3924 02:40:36,227 --> 02:40:39,363 THIS IS AN INTERDISCIPLINARY 3925 02:40:39,363 --> 02:40:42,133 EFFORT. 3926 02:40:42,133 --> 02:40:44,235 NEUROAI IS A WONDERFUL 3927 02:40:44,235 --> 02:40:46,737 COMMUNITY OF EXPERTISE AND WE 3928 02:40:46,737 --> 02:40:47,872 NEED TO COLLABORATE. 3929 02:40:47,872 --> 02:40:52,009 I'M NOT AN EXPERT IN 3930 02:40:52,009 --> 02:40:54,345 NEUROSCIENCE OR MACHINE LEARNING. 3931 02:40:54,345 --> 02:40:54,612 I'M A ROBOTICIST. I CONSUME WORK 3932 02:40:54,612 --> 02:40:58,983 THAT YOU ALL ARE DEVELOPING. 3933 02:40:58,983 --> 02:41:00,885 WE COULDN'T BE SUCCESSFUL 3934 02:41:00,885 --> 02:41:03,020 WITHOUT YOUR HELP. 3935 02:41:03,020 --> 02:41:04,655 HOPEFULLY WE'LL BE GIVING BACK 3936 02:41:04,655 --> 02:41:06,924 TO YOU AS WE APPLY THESE SYSTEMS 3937 02:41:06,924 --> 02:41:10,494 THAT WE'LL BE ABLE TO PROVE OUT 3938 02:41:10,494 --> 02:41:13,297 WHICH ALGORITHMS THAT WE USE BUBBLE 3939 02:41:13,297 --> 02:41:14,565 UP TO SOLVE OUR PROBLEMS AND THAT 3940 02:41:14,565 --> 02:41:16,834 WILL BE FEEDBACK FOR YOU ALL. 3941 02:41:16,834 --> 02:41:19,937 AGAIN, THE SAME DIAGRAM WE'VE 3942 02:41:19,937 --> 02:41:22,172 SEEN A HUNDRED TIMES IN THE LAST 3943 02:41:22,172 --> 02:41:24,241 TWO DAYS, BUT THESE TWO 3944 02:41:24,241 --> 02:41:25,142 COMMUNITIES ARE SUPPORTING ONE 3945 02:41:25,142 --> 02:41:26,110 ANOTHER. 3946 02:41:26,110 --> 02:41:28,879 AND I JUST ENCOURAGE THIS 3947 02:41:28,879 --> 02:41:31,582 CONTINUING COLLABORATION. 3948 02:41:31,582 --> 02:41:34,318 AND WE'LL DO GREAT THINGS GOING 3949 02:41:34,318 --> 02:41:40,491 FORWARD. 3950 02:41:40,491 --> 02:41:42,126 >> THANK YOU FOR A WONDERFUL 3951 02:41:42,126 --> 02:41:42,893 TALK, JOE. 3952 02:41:42,893 --> 02:41:49,833 I WOULD LIKE TO INVITE WILL 3953 02:41:49,833 --> 02:41:55,272 NOURSE TO TALK TO US ALSO ON 3954 02:41:55,272 --> 02:41:55,539 ROBOTICS. 3955 02:41:55,539 --> 02:41:58,642 >> THANK YOU FOR INVITING ME. 3956 02:41:58,642 --> 02:42:00,611 IT'S AN HONOR TO BE ON THE SAME 3957 02:42:00,611 --> 02:42:07,885 STAGE AS ALL THESE KEY SPEAKERS. 3958 02:42:07,885 --> 02:42:10,588 I'M WILL NOURSE AND I SERVE AS PART 3959 02:42:10,588 --> 02:42:14,558 OF A LARGE-SCALE NSF GRANT C3NS 3960 02:42:14,558 --> 02:42:18,829 COMMUNICATION, COORDINATION AND 3961 02:42:18,829 --> 02:42:19,430 CONTROL OF NEUROMECHANICAL SYSTEMS 3962 02:42:19,430 --> 02:42:28,906 I'M GOING TO TALK ABOUT TODAY IS 3963 02:42:28,906 --> 02:42:32,309 WHAT IT WOULD MEAN FOR ROBOTS TO 3964 02:42:32,309 --> 02:42:36,480 HAVE INSECT-SCALE INTELLIGENCE. 3965 02:42:36,480 --> 02:42:39,249 I ARGUE WE DON'T HAVE THAT YET AND 3966 02:42:39,249 --> 02:42:40,484 WHAT IT WILL TAKE FOR US TO GET THERE 3967 02:42:40,484 --> 02:42:43,253 FIRST I HAVE TO TALK ABOUT WHAT IS 3968 02:42:43,253 --> 02:42:48,292 THE CURRENT STATE-OF-THE-ART IN 3969 02:42:48,292 --> 02:42:48,826 ROBOTICS. THROUGH MODERN A.I., 3970 02:42:48,826 --> 02:42:52,262 WE HAVE INTELLIGENT ROBOTIC 3971 02:42:52,262 --> 02:42:54,732 SYSTEMS THAT CAN DO ALMOST ANY 3972 02:42:54,732 --> 02:42:57,901 GENERAL TASK, RIGHT? THE WORK OF 3973 02:42:57,901 --> 02:43:06,276 MARCO HUTTER HAS BEEN INSTRUMENTAL. 3974 02:43:06,276 --> 02:43:07,811 ONE OF THE EXAMPLES OFTEN USED 3975 02:43:07,811 --> 02:43:10,514 AS A HARD PROBLEM FOR ROBOTICS 3976 02:43:10,514 --> 02:43:13,350 IS FOLDING LAUNDRY. 3977 02:43:13,350 --> 02:43:18,022 IN THIS LAST MONTH, THEY 3978 02:43:18,022 --> 02:43:19,857 DEMONSTRATE FOLDING UNSEEN 3979 02:43:19,857 --> 02:43:20,624 LAUNDRY. 3980 02:43:20,624 --> 02:43:22,493 THESE PROBLEMS ARE NOW BASICALLY 3981 02:43:22,493 --> 02:43:23,427 SOLVED. 3982 02:43:23,427 --> 02:43:25,029 WHAT I'M GOING TO TALK ABOUT 3983 02:43:25,029 --> 02:43:27,998 NEXT IS AUTONOMY. 3984 02:43:27,998 --> 02:43:29,867 THESE ARE GENERAL TASKS. 3985 02:43:29,867 --> 02:43:33,704 HOW AUTONOMOUS ARE THESE ROBOTS 3986 02:43:33,704 --> 02:43:34,972 TODAY? I'LL USE LENS OF MANAGEMENT. 3987 02:43:34,972 --> 02:43:36,940 IDEALLY WE WANT THE ROBOTS TO 3988 02:43:36,940 --> 02:43:37,808 WORK FOR US. 3989 02:43:37,808 --> 02:43:40,844 WE HAVE IT MANAGE THEM IN SOME 3990 02:43:40,844 --> 02:43:41,311 WAY. 3991 02:43:41,311 --> 02:43:44,848 THE LOWEST LEVEL OF MANAGEMENT 3992 02:43:44,848 --> 02:43:50,154 IS STEP-BY-STEP CONTROL OR 3993 02:43:50,154 --> 02:43:58,062 TELEOPERATION. WHERE THE USER 3994 02:43:58,062 --> 02:43:58,829 SPECIFY EXACT TRAJECTORY OF EVERY 3995 02:43:58,829 --> 02:44:01,065 INDIVIDUAL JOINT OR ROBOT SURGERY MACHINE. 3996 02:44:01,065 --> 02:44:05,569 THE NEXT LEVEL UP OF MANAGEMENT 3997 02:44:05,569 --> 02:44:06,570 IS GIVING AN EXACT SET OF STEP 3998 02:44:06,570 --> 02:44:07,304 THAT THEY CAN EXECUTE EVERYTIME. 3999 02:44:07,304 --> 02:44:09,640 THIS IS ROBOTIC MANUFACTURING 4000 02:44:09,640 --> 02:44:13,110 AND ASSEMBLY, ET CETERA. 4001 02:44:13,110 --> 02:44:14,478 WHERE ROBOTS ARE IN RESEARCH 4002 02:44:14,478 --> 02:44:16,213 RIGHT NOW AND WHERE I THINK 4003 02:44:16,213 --> 02:44:18,282 THEY'RE GOING TO BE IN PRODUCTS 4004 02:44:18,282 --> 02:44:19,717 SOON IS SOME KIND OF GENERAL 4005 02:44:19,717 --> 02:44:24,388 TASK. 4006 02:44:24,388 --> 02:44:27,891 YOU CAN SAY ASSEMBLE THIS PART 4007 02:44:27,891 --> 02:44:31,328 OR FOLD THIS PILE OF LAUNDRY. 4008 02:44:31,328 --> 02:44:34,231 ALL THESE DIFFERENT LEVELS OF 4009 02:44:34,231 --> 02:44:35,699 MANAGEMENT ARE USEFUL AND 4010 02:44:35,699 --> 02:44:37,101 IMPORTANT BUT THERE IS ANOTHER 4011 02:44:37,101 --> 02:44:39,303 FAMILY OF TASKS THAT WE NEED TO 4012 02:44:39,303 --> 02:44:43,841 GET INTO IF WE WANT ROBOTS TO 4013 02:44:43,841 --> 02:44:46,677 ACHIEVE INTELLIGENCE. 4014 02:44:46,677 --> 02:44:48,178 HERE IS A LIST OF RESPONSIBILITIES 4015 02:44:48,178 --> 02:44:49,246 AND I WANT YOU TO DO EACH OF THEM 4016 02:44:49,246 --> 02:44:51,582 AS THE NEED ARISES. THIS IS THE 4017 02:44:51,582 --> 02:44:52,583 IDEA OF HAVING AUTONOMOUS 4018 02:44:52,583 --> 02:44:54,451 CONTEXT-DEPENDENT BEHAVIOR 4019 02:44:54,451 --> 02:44:54,718 SELECTION. 4020 02:44:54,718 --> 02:44:57,554 THIS IS A LONG-TIME HORIZON 4021 02:44:57,554 --> 02:45:00,190 PROBLEM THAT WE DON'T SEE SOLVED 4022 02:45:00,190 --> 02:45:02,526 WITH ROBOTICS AT THE MOMENT. 4023 02:45:02,526 --> 02:45:04,962 IT'S WHERE WE WANT ROBOTS TO BE 4024 02:45:04,962 --> 02:45:07,865 AND ANIMALS DO THIS ALL THE TIME. 4025 02:45:07,865 --> 02:45:10,033 I'M NOT GOING TO TALK ABOUT HOW 4026 02:45:10,033 --> 02:45:13,137 YOU COULD SCALE TRADITIONAL A.I. 4027 02:45:13,137 --> 02:45:14,271 SOLUTIONS BECAUSE I THINK 4028 02:45:14,271 --> 02:45:16,874 PEOPLE IN THIS ROOM AGREE THAT 4029 02:45:16,874 --> 02:45:19,977 PURE SCALING IS NOT GOING TO SOLVE 4030 02:45:19,977 --> 02:45:21,044 ALL THE PROBLEMS. WHY DO I 4031 02:45:21,044 --> 02:45:26,183 SAY INSECT-SCALE INTELLIGENCE? 4032 02:45:26,183 --> 02:45:30,220 INSECTS ARE NOT JUST A STOP ON 4033 02:45:30,220 --> 02:45:30,420 THE WAY TO MAMMAL-TYPE INTELLIGENCE. 4034 02:45:30,420 --> 02:45:32,856 IT'S AN INCREDIBLE FEAT AND WILL 4035 02:45:32,856 --> 02:45:36,660 BE AN INCREDIBLE ACHIEVEMENT FOR 4036 02:45:36,660 --> 02:45:37,494 ROBOTS. 4037 02:45:37,494 --> 02:45:41,131 INSECTS HAVE GOAL-DIRECTED 4038 02:45:41,131 --> 02:45:42,533 MOTION. FRUIT FLIES SMELL ROTTING 4039 02:45:42,533 --> 02:45:51,942 STRAWBERRY IN THE BREEZE AND IS 4040 02:45:51,942 --> 02:45:54,044 ABLE TO FIND THAT SPECIFIC GOAL. 4041 02:45:54,044 --> 02:45:55,546 THEY HAVE LONG-TERM NAVIGATION. ABLE 4042 02:45:55,546 --> 02:45:57,481 TO TRAVEL LARGE DISTANCES FOR FOOD. 4043 02:45:57,481 --> 02:45:59,817 AS SOON AS THEY FIND THEIR FOOD, 4044 02:45:59,817 --> 02:46:03,554 THEY TAKE A STRAIGHT LINE BACK 4045 02:46:03,554 --> 02:46:06,690 TO THEIR HOME NEST WITH LIMITED 4046 02:46:06,690 --> 02:46:08,792 COMPUTATION. THEY DO COMPLEX MOTOR 4047 02:46:08,792 --> 02:46:12,663 BEHAVIOR. THIS IS A FRUIT FLY 4048 02:46:12,663 --> 02:46:19,937 CROSSING A CHASM LARGER THAN ITSELF. 4049 02:46:19,937 --> 02:46:21,271 COORDINATING ALL OF IT'S MUSCLES. 4050 02:46:21,271 --> 02:46:24,875 THEY'RE ADAPTIVE. YOU CAN TAKE A FLY 4051 02:46:24,875 --> 02:46:26,977 CUT OFF ONE LEG AND IT'S ABLE TO 4052 02:46:26,977 --> 02:46:31,548 IMMEDIATELY ADJUST ITSELF SO IT 4053 02:46:31,548 --> 02:46:33,650 DOESN'T LOSE STABILITY AND STILL 4054 02:46:33,650 --> 02:46:38,589 MAINTAIN NORMAL LOCOMOTION. 4055 02:46:38,589 --> 02:46:41,458 ANTS CAN ACHIEVE AND FORM LARGE 4056 02:46:41,458 --> 02:46:44,995 STRUCTURES AND WORK AS A LARGE 4057 02:46:44,995 --> 02:46:45,796 SCALE SYSTEM. 4058 02:46:45,796 --> 02:46:47,731 PUTTING THIS ALL TOGETHER, 4059 02:46:47,731 --> 02:46:50,300 INSECTS HAVE THE CAPABILITY OF 4060 02:46:50,300 --> 02:46:52,336 AUTONOMOUSLY SWITCHING AMONG THE 4061 02:46:52,336 --> 02:46:53,737 TASKS DEPENDING ON THE CONTEXT 4062 02:46:53,737 --> 02:46:54,805 OF THE ENVIRONMENT THEY'RE 4063 02:46:54,805 --> 02:46:55,806 LOOKING AT. 4064 02:46:55,806 --> 02:46:58,175 LOOKING AT INSECTS, ONE OF THE 4065 02:46:58,175 --> 02:47:02,512 PLUSES IS WE HAVE ALL 140,000 4066 02:47:02,512 --> 02:47:07,684 NEURONS IN THE BRAIN OF A FRUIT FLY, 4067 02:47:07,684 --> 02:47:11,088 23,000 VENTRAL NERVE CORDS 4068 02:47:11,088 --> 02:47:11,989 ANALOGOUS TO THE SPINE. WHAT'S 4069 02:47:11,989 --> 02:47:13,257 EXCITING ABOUT CONTEXT-DEPENDENT 4070 02:47:13,257 --> 02:47:14,091 ACTION SELECTION? WE KNOW THE BRAIN 4071 02:47:14,091 --> 02:47:18,262 REGION INVOLVED: THE CENTRAL COMPLEX. 4072 02:47:18,262 --> 02:47:21,198 IT HAS 3000 NEURONS AND BEEN STUDIED 4073 02:47:21,198 --> 02:47:23,600 FOR ITS ROLE IN PATH PLANNING AND 4074 02:47:23,600 --> 02:47:24,368 NAVIGATION. 4075 02:47:24,368 --> 02:47:27,371 ANOTHER OVERLOOKED ASPECT IS THE 4076 02:47:27,371 --> 02:47:29,106 CENTER OF THE BRAIN WHERE ALL 4077 02:47:29,106 --> 02:47:30,841 SENSOR INFORMATION CONVERGES 4078 02:47:30,841 --> 02:47:33,977 INTO THE CENTRAL COMPLEX AND 4079 02:47:33,977 --> 02:47:35,412 MOTOR COMMANDS COME OUT. 4080 02:47:35,412 --> 02:47:36,346 THIS REGION TAKES IN THE STATE 4081 02:47:36,346 --> 02:47:37,915 OF EVERYTHING IN THE WORLD AND 4082 02:47:37,915 --> 02:47:39,816 DECIDES WHAT THE ANIMAL HAS TO 4083 02:47:39,816 --> 02:47:41,151 DO NEXT. IT'S ONLY 3,000 NEURONS. 4084 02:47:41,151 --> 02:47:50,694 WHY CANT' WE TAKE THE CONNECTOME 4085 02:47:50,694 --> 02:47:54,831 PLOP IT ON NEUROMORPHIC HARDWARE 4086 02:47:54,831 --> 02:47:58,035 AND RUN IT ON A ROBOT AUTONOMOUSLY? 4087 02:47:58,035 --> 02:47:59,870 FUTURE FUNDING AND OPPORTUNITIES 4088 02:47:59,870 --> 02:48:01,038 CAN HELP US WITH. EVEN THOUGH 4089 02:48:01,038 --> 02:48:03,774 WE HAVE A CONNECTOME OF THE BRAIN 4090 02:48:03,774 --> 02:48:07,077 AND VENTRAL NERVE CORD, WE DON'T 4091 02:48:07,077 --> 02:48:08,679 HAVE A CONNECTOME OF ALL OF THAT IN 4092 02:48:08,679 --> 02:48:18,622 ONE ANIMAL. THIS IS IMPORTANT FOR 4093 02:48:18,622 --> 02:48:23,794 DESCENDING INFORMATION. COMING SOON. 4094 02:48:23,794 --> 02:48:26,263 ADDITIONALLY, WE NEED STATISTICALLY 4095 02:48:26,263 --> 02:48:30,100 SIGNIFICANT CONNECTOMES. TO 4096 02:48:30,100 --> 02:48:36,540 UNDERSTAND WHAT IS A COMPONENT 4097 02:48:36,540 --> 02:48:41,178 OF THIS INTELLIGENCE AND 4098 02:48:41,178 --> 02:48:41,478 INTERINDIVIDUAL VARIATION, WE 4099 02:48:41,478 --> 02:48:45,215 NEED LOT MORE CONNECTOMES TO LOOK AT 4100 02:48:45,215 --> 02:48:47,284 THE STATISTICS OF NEURAL MAPS. 4101 02:48:47,284 --> 02:48:50,954 ADDITIONALLY, SOME SMALL ELEMENTS 4102 02:48:50,954 --> 02:48:53,991 ARE MISSING BECAUSE THE VISION 4103 02:48:53,991 --> 02:48:55,359 ALGORITHMS CAN'T DETECT THEM. 4104 02:48:55,359 --> 02:48:57,995 THESE ARE ELECTRICAL OR EPHAPTIC 4105 02:48:57,995 --> 02:48:58,562 OR GAP JUNCTIONS. THEY ALLOW 4106 02:48:58,562 --> 02:49:01,999 BIDIRECTIONAL COMMUNICATIONS 4107 02:49:01,999 --> 02:49:03,600 OVER THE SAME CHANNEL. 4108 02:49:03,600 --> 02:49:06,536 THESE ARE EXPERIMENTALLY FOUND IN 4109 02:49:06,536 --> 02:49:08,639 DIFFERENT PATHWAYS IN INSECTS BUT 4110 02:49:08,639 --> 02:49:12,175 YOU CAN'T SEE THEM IN THE CONNECTOME. 4111 02:49:12,175 --> 02:49:16,546 IN MOTOR CONTROL, FIBERS BETWEEN THE 4112 02:49:16,546 --> 02:49:19,516 PERIPHERAL SENSORY SYSTEMS AND 4113 02:49:19,516 --> 02:49:22,152 VENTRAL NERVE CORD ARE TOO SMALL TO 4114 02:49:22,152 --> 02:49:26,556 BE PICKED UP SO THAT DATA IS NOT 4115 02:49:26,556 --> 02:49:28,992 THERE OR IT'S SPORADIC. 4116 02:49:28,992 --> 02:49:33,330 WE CAN DO COMPLEMENTARY ORGANISM 4117 02:49:33,330 --> 02:49:35,832 CONNECTOMES. LOOK FOR SIMILAR NERVOUS 4118 02:49:35,832 --> 02:49:40,971 SYSTEM BUT DIFFERENT BEHAVIOR. THE 4119 02:49:40,971 --> 02:49:43,940 PRAYING MANTIS IS A PREDATOR AND IT 4120 02:49:43,940 --> 02:49:44,574 HAS DIFFERENT PRIORITIES IT NEEDS 4121 02:49:44,574 --> 02:49:46,443 ITS CENTRAL COMPLEX TO DO. ONE OF 4122 02:49:46,443 --> 02:49:46,777 THOSE BIG ONES BEING PREDICTION. 4123 02:49:46,777 --> 02:49:51,048 IT HAS TO BE ABLE TO PREDICT THE 4124 02:49:51,048 --> 02:49:53,784 ACTIONS OF ITS PREY. HOW THE 4125 02:49:53,784 --> 02:49:56,720 CENTRAL COMPLEX DIFFERS ACROSS 4126 02:49:56,720 --> 02:50:01,625 DIFFERENT LEVELS OF ANIMAL BEHAVIOR, 4127 02:50:01,625 --> 02:50:03,660 MAY REVEAL CORE CONCEPTS OF INTELLIGENCE 4128 02:50:03,660 --> 02:50:09,466 ON THE HARDWARE SIDE, THE ENTIRE 4129 02:50:09,466 --> 02:50:11,101 CONNECTOME CAN BE PORTED INTO 4130 02:50:11,101 --> 02:50:12,069 NEUROMORPHIC HARDWARE - EXCITING. 4131 02:50:12,069 --> 02:50:14,371 WE STILL HAVE LIMITATIONS BEFORE 4132 02:50:14,371 --> 02:50:16,406 WE CAN RUN THIS ON PHYSICAL 4133 02:50:16,406 --> 02:50:20,343 SYSTEMS AND INTELLIGENT ROBOTS. 4134 02:50:20,343 --> 02:50:23,013 DUE TO THE FAN IN AND OUT CONSTRAINTS 4135 02:50:23,013 --> 02:50:24,648 OF CURRENT GENERATION NEUROMORPHIC 4136 02:50:24,648 --> 02:50:26,416 HARDWARE, IT REQUIRES MORE CHIPS THAN 4137 02:50:26,416 --> 02:50:31,455 THAN YOU WOULD EXPECT BASED ON NEURON 4138 02:50:31,455 --> 02:50:42,566 COUNTS. A CHIP WITH MILLION NEURONS 4139 02:50:47,471 --> 02:50:49,539 CAN'T FIT 140,000 CONNECTOME. 4140 02:50:49,539 --> 02:50:58,682 WE LACK EASILY AVAILABLE SYSTEMS FOR 4141 02:50:58,682 --> 02:50:59,082 RESEARCHERS INSPITE OF CLOUD-SCALE 4142 02:50:59,082 --> 02:51:02,018 SYSTEMS. AT THE END OF THE DAY, TO 4143 02:51:02,018 --> 02:51:06,323 PUT IT ON A ROBOTIC SYSTEM, WE NEED 4144 02:51:06,323 --> 02:51:09,960 TO HAVE PHYSICAL SYSTEMS IN HAND. 4145 02:51:09,960 --> 02:51:11,428 FINALLY, THE LAST THING I WANT 4146 02:51:11,428 --> 02:51:13,964 TO MENTION IS NEURAL COMPLEXITY. 4147 02:51:13,964 --> 02:51:16,266 EVEN THOUGH WE HAVE CONNECTOME 4148 02:51:16,266 --> 02:51:19,769 AND A POINT-TO-POINT MAP OF NEURONS, 4149 02:51:19,769 --> 02:51:24,841 A LOT OF NEURONS DON'T BEHAVE 4150 02:51:24,841 --> 02:51:28,245 AS POINT-TO-POINT. THIS IS A 4151 02:51:28,245 --> 02:51:30,780 PROJECTION PATTERN OF ONE SINGLE 4152 02:51:30,780 --> 02:51:38,655 NEURON IN THE FRUIT FLY, RING NEURON, 4153 02:51:38,655 --> 02:51:40,157 INNERVATING ENTIRE CENTRAL COMPLEX. 4154 02:51:40,157 --> 02:51:44,361 IN THIS PROJECTION, LOCAL COMPUTATION 4155 02:51:44,361 --> 02:51:46,863 HAPPEN WITHIN THE SYNPASE BEFORE 4156 02:51:46,863 --> 02:51:50,634 IT GETS TO THE CELL BODY. THIS CAN'T 4157 02:51:50,634 --> 02:51:51,768 BE IMPLEMENTED WITH A SINGLE SYNAPSE. 4158 02:51:51,768 --> 02:51:56,840 WE NEED MORE CONNECTOMES, ACROSS 4159 02:51:56,840 --> 02:51:59,075 MORE SPECIES AT HIGHER RESOLUTION. 4160 02:51:59,075 --> 02:52:03,747 WE NEED MID-SCALE NEUROMORPHIC 4161 02:52:03,747 --> 02:52:05,482 SYSTEMS TO CATALYZE ROBOTIC 4162 02:52:05,482 --> 02:52:06,583 INTELLIGENCE. I WANT TO THANK ALL 4163 02:52:06,583 --> 02:52:07,918 THE UNIVERSITIES INVOLVED 4164 02:52:07,918 --> 02:52:10,554 AND THE FUNDING OPPORTUNITIES. 4165 02:52:10,554 --> 02:52:14,858 THANK YOU. 4166 02:52:14,858 --> 02:52:16,126 >> THANK YOU. 4167 02:52:16,126 --> 02:52:23,166 THIS TIME I WOULD LIKE TO INVITE 4168 02:52:23,166 --> 02:52:25,101 KAI MILLER FROM THE MAYO 4169 02:52:25,101 --> 02:52:27,904 CLINIC. 4170 02:52:27,904 --> 02:52:29,773 >> IT'S A GREAT PLEASURE TO BE 4171 02:52:29,773 --> 02:52:32,075 HERE IN PARTICULAR, I WOULD LIKE 4172 02:52:32,075 --> 02:52:34,277 TO THANK DR. HWANG THE PROGRAM 4173 02:52:34,277 --> 02:52:36,580 OFFICER FOR INVITING ME. 4174 02:52:36,580 --> 02:52:38,982 I'M A CLINICIAN SO I SHOULD 4175 02:52:38,982 --> 02:52:41,685 START OUT BY SAYING THERE IS 4176 02:52:41,685 --> 02:52:51,828 DOMAIN EXPERTISE -- I HAVE LESS 4177 02:52:51,828 --> 02:52:53,930 EXPERTISE THAN THOSE IN THE 4178 02:52:53,930 --> 02:52:56,399 AUDIENCE BUT I CAN DESCRIBE THE 4179 02:52:56,399 --> 02:52:58,969 END APPLICATION FOR HEALTHCARE. 4180 02:52:58,969 --> 02:53:01,738 THESE ARE MY DISCLOSURES. 4181 02:53:01,738 --> 02:53:04,007 I WOULD LIKE TO THANK THE NIH 4182 02:53:04,007 --> 02:53:07,644 FOR SUPPORTING OUR WORK. 4183 02:53:07,644 --> 02:53:11,915 FUNCTIONAL NEUROSURGEONS ARE 4184 02:53:11,915 --> 02:53:14,184 DOCTORS WHO INVASIVELY INTERACT 4185 02:53:14,184 --> 02:53:16,186 WITH NEUROCIRCUITRY. WE THINK OF 4186 02:53:16,186 --> 02:53:20,991 THE BRAIN, THE SPINE AND PERIPHERAL 4187 02:53:20,991 --> 02:53:21,658 NERVE TO MEASURE AND MODIFY ITS 4188 02:53:21,658 --> 02:53:26,029 COMPUTATIONAL PURPOSE USING 4189 02:53:26,029 --> 02:53:28,331 ELECTRODES AND TISSUE 4190 02:53:28,331 --> 02:53:29,866 DESTRUCTION AND NOW WITH 4191 02:53:29,866 --> 02:53:30,867 EMERGING GENOMIC TOOLS. 4192 02:53:30,867 --> 02:53:33,436 TO BEGIN WITH, WE HAD OPEN 4193 02:53:33,436 --> 02:53:35,705 SURGERIES AND WE USED TO DO THINGS 4194 02:53:35,705 --> 02:53:38,541 LIKE INDUCE FOCAL STROKES. 4195 02:53:38,541 --> 02:53:41,344 BUT OVER TIME WE START 4196 02:53:41,344 --> 02:53:44,748 TO DELIVER HIGH FREQUENCY-POWERED 4197 02:53:44,748 --> 02:53:46,383 ULTRASOUND OR LASER LIGHT 4198 02:53:46,383 --> 02:53:48,952 TO DESTROY STRUCTURES INVOLVED 4199 02:53:48,952 --> 02:53:53,290 WITH CIRCUITRY AND STIMULATE. 4200 02:53:53,290 --> 02:53:55,926 THE AMOUNT OF INTERVENTION WE 4201 02:53:55,926 --> 02:53:59,729 HAVE IS CRUDE COMPARED TO 4202 02:53:59,729 --> 02:54:03,366 TECHNOLOGY ON OUR PHONES AND 4203 02:54:03,366 --> 02:54:05,135 COMPUTING. 4204 02:54:05,135 --> 02:54:06,736 MEDICINE DOES LAG AND THE 4205 02:54:06,736 --> 02:54:07,604 PURPOSE BEHIND THE LAG IS THE 4206 02:54:07,604 --> 02:54:10,874 NEED FOR SAFETY AND THE NEED FOR 4207 02:54:10,874 --> 02:54:12,442 VALIDATION PRIOR TO IMPLEMENTING 4208 02:54:12,442 --> 02:54:13,009 THINGS. 4209 02:54:13,009 --> 02:54:14,644 WHEN WE INTERVENE WITH THE 4210 02:54:14,644 --> 02:54:16,846 NERVOUS SYSTEM, IT'S IMPORTANT 4211 02:54:16,846 --> 02:54:18,782 TO REALIZE THERE IS A DIFFERENCE 4212 02:54:18,782 --> 02:54:21,551 BETWEEN TREATING THE SYMPTOMS OF 4213 02:54:21,551 --> 02:54:24,587 A DISEASE AND CHANGING THE 4214 02:54:24,587 --> 02:54:26,122 CIRCUIT DYSFUNCT DYSFUNCTIONS 4215 02:54:26,122 --> 02:54:26,923 THAT UNDERLIE THE DISEASE. 4216 02:54:26,923 --> 02:54:28,825 THAT MEANS WITH WE PUT DEVICES 4217 02:54:28,825 --> 02:54:31,194 IN THE BRAIN, MOST OFTEN WHAT 4218 02:54:31,194 --> 02:54:33,530 WE'RE DOING IS DELIVERING 4219 02:54:33,530 --> 02:54:36,399 ELECTRICAL CURRENT AT HIGH 4220 02:54:36,399 --> 02:54:37,667 FREQUENCY TO DISRUPT AN AREA. 4221 02:54:37,667 --> 02:54:39,703 MOVING FORWARD, WE WOULD LIKE TO 4222 02:54:39,703 --> 02:54:41,705 TREAT MORE NUANCE DISEASES THAN 4223 02:54:41,705 --> 02:54:43,173 WE DO NOW. MOVING FORWARD, 4224 02:54:43,173 --> 02:54:50,246 WHERE WE TREAT TREMORS, WE WANT 4225 02:54:50,246 --> 02:54:52,349 TO START TREATING PSYCHIATRIC 4226 02:54:52,349 --> 02:54:54,984 DISEASES, WE'RE GOING NO NEED 4227 02:54:54,984 --> 02:54:56,786 APPROACHES THAT MATCH THE 4228 02:54:56,786 --> 02:54:57,954 STATISTICS OF THE BRAIN. 4229 02:54:57,954 --> 02:55:01,124 I THINK OF THIS AS WHAT IT MEANS 4230 02:55:01,124 --> 02:55:05,862 TO BE NEUROMORPHIC AND ADAPT TO 4231 02:55:05,862 --> 02:55:08,598 THE STATE OF THE BRAIN SO WE CAN 4232 02:55:08,598 --> 02:55:12,635 MODIFY THE WAY THE BRAIN 4233 02:55:12,635 --> 02:55:14,070 COMPUTES IN REALTIME. 4234 02:55:14,070 --> 02:55:17,540 THE TYPE OF APPROACH FOR 4235 02:55:17,540 --> 02:55:18,875 NEUROMORPHIC SYSTEMS AND 4236 02:55:18,875 --> 02:55:20,410 EMBODIMENT MAY APPLY. 4237 02:55:20,410 --> 02:55:23,446 IF I SAY WHAT'S THE USE OF A 4238 02:55:23,446 --> 02:55:23,980 CLINICIAN? 4239 02:55:23,980 --> 02:55:27,817 I CAN PROVIDE SOME CONTEXT AND 4240 02:55:27,817 --> 02:55:29,219 DESCRIBE A NEED THAT WE HAVE, 4241 02:55:29,219 --> 02:55:32,555 BUT IT'S IMPORTANT TO REMEMBER 4242 02:55:32,555 --> 02:55:35,558 THAT THE MISSION STATEMENT OF 4243 02:55:35,558 --> 02:55:39,429 THE NIH AND MISSION STATEMENT OF 4244 02:55:39,429 --> 02:55:43,666 OTHER INSTITUTIONS THAT DELIVER 4245 02:55:43,666 --> 02:55:46,770 CARE IS TO MEET A CLINICAL NEED. 4246 02:55:46,770 --> 02:55:49,239 BUT THE MISSION FOR COMPANIES ARE 4247 02:55:49,239 --> 02:55:50,407 BOUND BY THEIR CHARTER TO MAKE A PROFIT. 4248 02:55:50,407 --> 02:55:53,176 WE HAVE TO KEEP IN MIND THAT THE 4249 02:55:53,176 --> 02:55:55,011 WAY WE DESIGN THINGS HAS TO 4250 02:55:55,011 --> 02:55:57,113 INTEGRATE INTO THE ECONOMIC 4251 02:55:57,113 --> 02:55:59,649 PICTURE WHICH PLAYS A ROLE IN 4252 02:55:59,649 --> 02:56:02,952 THE TYPES OF WAYS WE MAKE 4253 02:56:02,952 --> 02:56:04,053 DECISION IN THE CLINICAL CONTEXT. 4254 02:56:04,053 --> 02:56:07,490 THE ECONOMIC INCENTIVES DRIVE 4255 02:56:07,490 --> 02:56:08,892 WHICH TOOLS. 4256 02:56:08,892 --> 02:56:12,362 IT'S NOT NECESSARILY THE BEST 4257 02:56:12,362 --> 02:56:14,330 TOOL BUT THE ONE THAT WE HAVE 4258 02:56:14,330 --> 02:56:17,100 ACCESS TO. 4259 02:56:17,100 --> 02:56:18,868 ECONOMICS DRIVE THAT. 4260 02:56:18,868 --> 02:56:20,503 THAT'S NOT THE SAME THING AS 4261 02:56:20,503 --> 02:56:23,807 WHAT IS SCIENTIFICALLY COMPELLING. 4262 02:56:23,807 --> 02:56:24,340 I'M A SCIENTST AND A SURGEON. 4263 02:56:24,340 --> 02:56:27,544 THE THINGS THAT INTEREST ME 4264 02:56:27,544 --> 02:56:28,678 INTELLECTUALLY ARE DIFFERENT 4265 02:56:28,678 --> 02:56:32,982 THAN THE THINGS THAT INFLUENCE 4266 02:56:32,982 --> 02:56:35,385 THE DECISION I 4267 02:56:35,385 --> 02:56:37,687 MAKE ON HOW TO TREAT A 4268 02:56:37,687 --> 02:56:38,555 PARTICULAR CHILD. 4269 02:56:38,555 --> 02:56:41,925 A LOT OF US HERE THINK ABOUT THE 4270 02:56:41,925 --> 02:56:43,393 MESSAGING TO THE PUBLIC. 4271 02:56:43,393 --> 02:56:45,328 IT'S DIFFERENT THAN WHAT 4272 02:56:45,328 --> 02:56:50,366 PATIENTS NEED AND ECONOMIC 4273 02:56:50,366 --> 02:56:52,469 INCENTIVES. 4274 02:56:52,469 --> 02:56:57,073 IF WE SAY WHAT DOES ARTIFICIAL 4275 02:56:57,073 --> 02:57:00,910 INTELLIGENCE HAVE TO OFFER US 4276 02:57:00,910 --> 02:57:01,644 FUNCTIONAL NEUROSURGEON? I THINK WE 4277 02:57:01,644 --> 02:57:05,982 CAN USE IT FOR DISCOVERING STRUCTURE. 4278 02:57:05,982 --> 02:57:08,852 I DO ELECTROPHYSIOLOGICAL SIGNAL 4279 02:57:08,852 --> 02:57:12,622 DECODING. TRY TO UNDERSTAND HOW 4280 02:57:12,622 --> 02:57:21,931 GENETIC PREDISPOSITION DETERMINE 4281 02:57:21,931 --> 02:57:24,234 THE END PHENOTYPE. STRUCTURAL 4282 02:57:24,234 --> 02:57:25,335 SEGMENTATION, DOCUMENTATION,CHART 4283 02:57:25,335 --> 02:57:28,304 SYNTHESIS ARE THE MOST BORING THING 4284 02:57:28,304 --> 02:57:29,272 THAT WE DO. 4285 02:57:29,272 --> 02:57:31,941 THIS IS THE MOST STRAIGHTFORWARD 4286 02:57:31,941 --> 02:57:34,711 IMPLEMENTATION WE HAVE FOR LLMS 4287 02:57:34,711 --> 02:57:35,745 NOW. BUT FOR DELIVERING MEDICAL 4288 02:57:35,745 --> 02:57:37,847 CARE, THIS IS 50% OF THE PERSON 4289 02:57:37,847 --> 02:57:44,254 TIME IS DEDICATED 4290 02:57:44,254 --> 02:57:50,326 TO DOCUMENTATION AND CHARTING. 4291 02:57:50,326 --> 02:57:52,562 AND A LOT OF IT IS NOT AUTOMATED 4292 02:57:52,562 --> 02:57:54,497 IN INTUITIVE OR WORKABLE WAYS. 4293 02:57:54,497 --> 02:57:55,665 CLINICAL PREDICTION. 4294 02:57:55,665 --> 02:57:56,900 FORECASTING HOW THINGS CHANGE 4295 02:57:56,900 --> 02:57:58,768 OVER THE LIFESPAN. 4296 02:57:58,768 --> 02:58:01,938 I THINK I'VE HEARD A LOT ABOUT 4297 02:58:01,938 --> 02:58:02,872 ARTIFICIAL INTELLIGENCE WHERE 4298 02:58:02,872 --> 02:58:03,806 PEOPLE ARE THINKING ABOUT THIS 4299 02:58:03,806 --> 02:58:06,109 AHEAD OF TIME. 4300 02:58:06,109 --> 02:58:08,177 THIS IS JUST A SNAPSHOT TAKEN 4301 02:58:08,177 --> 02:58:12,515 FROM ONE OF OUR MANUSCRIPTS ABOUT 4302 02:58:12,515 --> 02:58:16,586 HOW WE MAKE A CONENSUS DECISION FOR 4303 02:58:16,586 --> 02:58:18,154 EACH INDIVIDUAL PATIENT WHO MIGHT 4304 02:58:18,154 --> 02:58:19,789 HAVE EPILEPSY. 4305 02:58:19,789 --> 02:58:21,724 WHAT WE TO HAVE IN THE FUTURE IS 4306 02:58:21,724 --> 02:58:25,161 DIGITAL TWINNING. USING A LARGE DATABASE OF PATIENTS, 4307 02:58:25,161 --> 02:58:29,165 WITHOUT RELYING ON CONVERSATIONS 4308 02:58:29,165 --> 02:58:32,635 BETWEEN EXPERTS, WE WANT TO HAVE 4309 02:58:32,635 --> 02:58:35,038 ALGORITHMS THAT HAVE THE PAST 4310 02:58:35,038 --> 02:58:36,773 KNOWLEDGE OF MANY, MANY PATIENTS 4311 02:58:36,773 --> 02:58:42,712 TREATED WITH EITHER STIMULATION 4312 02:58:42,712 --> 02:58:47,483 OR A RESECTION OR LASER ABLATION. 4313 02:58:47,483 --> 02:58:51,020 AMONG THIS PAST SET OF PATIENTS, 4314 02:58:51,020 --> 02:58:56,225 WHICH PATIENT IS THE CURRENT PATIENT 4315 02:58:56,225 --> 02:59:00,930 CLOSE TO? CAN WE MODEL IT BASED ON 4316 02:59:00,930 --> 02:59:03,299 PAST STATISTICS IN AN INTUITIVE WAY? 4317 02:59:03,299 --> 02:59:05,301 DIGITAL TWIN HAS A ROLE IN CLINICAL 4318 02:59:05,301 --> 02:59:06,502 MEDICINE AND DECISION-MAKING FOR PATIENTS. 4319 02:59:06,502 --> 02:59:08,705 FOR THOSE OF YOU WHO MAY NOT BE 4320 02:59:08,705 --> 02:59:11,140 FAMILIAR WITH WHAT IT'S LIKE TO 4321 02:59:11,140 --> 02:59:13,109 BE A PATIENT AND GO THROUGH THE 4322 02:59:13,109 --> 02:59:14,944 WORK FLOW, THIS IS HOW I THINK 4323 02:59:14,944 --> 02:59:18,581 OF ORGANIZING THINGS WITHIN AN 4324 02:59:18,581 --> 02:59:19,148 INSTITUTION. 4325 02:59:19,148 --> 02:59:23,119 THEY'LL SEE A PHYSICIAN WHERE A 4326 02:59:23,119 --> 02:59:24,187 DIAGNOSIS IS MADE. THERE IS A ROLE 4327 02:59:24,187 --> 02:59:27,056 FOR A.I. HERE. WE DO SURGICAL 4328 02:59:27,056 --> 02:59:30,159 PLANNING. WE PUT ELECTRODES IN ONE 4329 02:59:30,159 --> 02:59:32,929 SPOT IF I'M BURNING AN AREA OF 4330 02:59:32,929 --> 02:59:35,832 THE BRAIN OR IF I WANT TO USE 4331 02:59:35,832 --> 02:59:39,702 MOTION CAPTURE TO LOCALIZE AREAS IN 4332 02:59:39,702 --> 02:59:48,011 THE BRAIN, MACHINE LEARNING CAN PLAY 4333 02:59:48,011 --> 02:59:53,416 AN IMPORTANT ROLE. WE USE ROBOTICS 4334 02:59:53,416 --> 02:59:57,353 IN THE OPERATING ROOM. WE HAVE 4335 02:59:57,353 --> 02:59:59,822 ONLINE DIAGNOSTICS BASED ON IMAGING 4336 02:59:59,822 --> 03:00:01,658 ELECTROPHYSIOLOGY. THERE' A ROLE FOR 4337 03:00:01,658 --> 03:00:03,259 A.I. IN PREDICTING RESPONSE TO 4338 03:00:03,259 --> 03:00:05,561 INTERVENTION. ANOTHER APPLICATION 4339 03:00:05,561 --> 03:00:08,498 FOR DIGITAL TWINNING. IN THE NEXT 4340 03:00:08,498 --> 03:00:10,967 STEP NOW THAT WE'VE DONE THIS, CAN WE 4341 03:00:10,967 --> 03:00:11,634 PREDICT INDIVIDUAL CARE AFTERWARDS? 4342 03:00:11,634 --> 03:00:13,536 MY RESEARCH AND THE RESEARCH OF 4343 03:00:13,536 --> 03:00:17,940 MANY OTHER NEUROSURGEONS WHO ARE 4344 03:00:17,940 --> 03:00:19,609 SCIENTISTS IS WORKING ON THE NEXT 4345 03:00:19,609 --> 03:00:23,579 GENERATION OF IMPLANTABLE DEVICES. 4346 03:00:23,579 --> 03:00:25,314 PARAMETER OPTIMIZATION AND 4347 03:00:25,314 --> 03:00:26,282 CLOSED-LOOP CONTROL IS THE OBVIOUS 4348 03:00:26,282 --> 03:00:29,952 CLEAREST MOST COMPELLING ASPECT 4349 03:00:29,952 --> 03:00:33,990 OF A.I. FOR NEUROSURGERY. 4350 03:00:33,990 --> 03:00:41,297 WE NEED TO DISCUSS HOW TO INTEGRATE 4351 03:00:41,297 --> 03:00:48,838 INFORMATION AND MULTIMODAL DATA 4352 03:00:48,838 --> 03:00:50,206 ACROSS INSTITUTIONS FOR NEW THERAPY. 4353 03:00:50,206 --> 03:00:52,608 DOMINIQUE DUNCAN WHO SPOKE YESTERDAY, 4354 03:00:52,608 --> 03:00:55,078 THE TYPES OF ROLES THAT THESE 4355 03:00:55,078 --> 03:00:57,613 SCIENTISTS PLAY ARE IMPORTANT FOR 4356 03:00:57,613 --> 03:00:58,614 NEXT GENERATION THERAPY AND IMPORTANT 4357 03:00:58,614 --> 03:01:03,686 FOR THE NEW ALGORITHMS THAT WILL BE 4358 03:01:03,686 --> 03:01:04,354 DEVELOPED AND IMPLEMENTED. 4359 03:01:04,354 --> 03:01:09,492 I THINK THERE IS A KEY ROLE FOR 4360 03:01:09,492 --> 03:01:13,096 HAVING NEW ALGORITHMS THAT -- 4361 03:01:13,096 --> 03:01:14,997 TAKING NEW ALGORITHMS AND 4362 03:01:14,997 --> 03:01:16,566 ALLOWING ONE TO IMPLEMENT THEM 4363 03:01:16,566 --> 03:01:19,035 FROM DATA THAT IS AGGREGATED 4364 03:01:19,035 --> 03:01:20,803 ACROSS A WIDE VARIETY OF 4365 03:01:20,803 --> 03:01:21,604 INSTITUTIONS. 4366 03:01:21,604 --> 03:01:23,206 THAT IS SOMETHING THAT THE NIH 4367 03:01:23,206 --> 03:01:25,975 IS UNIQUELY EQUIPPED TO DO BECAUSE 4368 03:01:25,975 --> 03:01:30,346 NIH IS WHAT INCENTIVIZES US 4369 03:01:30,346 --> 03:01:33,182 PHYSICIANS AND OTHERS -- WHAT 4370 03:01:33,182 --> 03:01:34,917 INCENTIVIZES US IN TERMS OF 4371 03:01:34,917 --> 03:01:36,619 DECIDING THE KIND OF RESEARCH WE'RE 4372 03:01:36,619 --> 03:01:39,055 GOING TO DO AND DECIDING HOW TO 4373 03:01:39,055 --> 03:01:41,457 DO IT AND GETTING FUNDS AND 4374 03:01:41,457 --> 03:01:41,924 RESOURCES WE NEED. 4375 03:01:41,924 --> 03:01:45,027 I WOULD LIKE TO GIVE A COUPLE OF 4376 03:01:45,027 --> 03:01:46,162 EXAMPLES ABOUT HOW WE DESIGN DEVICES 4377 03:01:46,162 --> 03:01:51,701 THAT INTERACT WITH THE BRAIN TO 4378 03:01:51,701 --> 03:01:56,572 PROVIDE A CONDUIT TO THE OUTSIDE WORLD: BRAIN COMPUTER INTERFACE. 4379 03:01:56,572 --> 03:02:01,077 INDUCING PLASTICITY TO HELP PEOPLE 4380 03:02:01,077 --> 03:02:03,846 WHO HAVE BEEN INJURED BY STROKE OR 4381 03:02:03,846 --> 03:02:07,583 EPILEPSY. TO MANAGE DISEASE SYMPTOMS 4382 03:02:07,583 --> 03:02:09,352 THIS IS OPEN LOOP STIMULATION NOW. 4383 03:02:09,352 --> 03:02:11,587 THINKING PARTICULARLY ABOUT EMBODIED 4384 03:02:11,587 --> 03:02:13,923 AND NEUROMORPHIC APPROACHES. 4385 03:02:13,923 --> 03:02:15,825 IMAGINE YOU HAVE A BRAIN, WE ALL 4386 03:02:15,825 --> 03:02:17,960 DO. 4387 03:02:17,960 --> 03:02:19,762 WE WANT TO SENSE FROM THE BRAIN 4388 03:02:19,762 --> 03:02:21,731 THESE TYPES OF DEVICE. 4389 03:02:21,731 --> 03:02:23,499 THERE IS A TECHNIQUE FOR 4390 03:02:23,499 --> 03:02:24,333 RECORDING INFORMATION IN THE 4391 03:02:24,333 --> 03:02:25,268 BRAIN. AN ELECTRICAL EXAMPLE: 4392 03:02:25,268 --> 03:02:28,571 WE HAVE SOME COMPUTATION AND 4393 03:02:28,571 --> 03:02:29,772 DECODING THAT GOES INTO THIS. 4394 03:02:29,772 --> 03:02:33,443 WE USE THAT FOR SOME SORT OF 4395 03:02:33,443 --> 03:02:34,844 INTERNAL/EXTERNAL EFFECT. THE 4396 03:02:34,844 --> 03:02:37,280 INTERNAL EFFECT WOULD BE BRAIN 4397 03:02:37,280 --> 03:02:40,216 STIMULATION. RETRAINING OR MANAGING 4398 03:02:40,216 --> 03:02:50,626 THE DISEASE WITH ELECTRODES 4399 03:02:50,626 --> 03:02:55,698 IMPLANTED IN THE BRAIN. WE CAN 4400 03:02:55,698 --> 03:02:57,867 INTERFACE WITH THE OUTSIDE WITH BCI. 4401 03:02:57,867 --> 03:03:00,369 ADOPTING AN EMBODIED OR NEUROMORPHIC 4402 03:03:00,369 --> 03:03:02,438 APPROACH MIGHT BE USEFUL. EMBODIED 4403 03:03:02,438 --> 03:03:04,640 MEASUREMENTS MATCHES THE INTERVENTION 4404 03:03:04,640 --> 03:03:06,742 THAT WE'VE PERFORMED TO THE PHYSICAL 4405 03:03:06,742 --> 03:03:09,479 SCALE OF THE SYSTEM. AND THAT COULD 4406 03:03:09,479 --> 03:03:12,949 BE LIMBS IN THE OUTSIDE WORLD, 4407 03:03:12,949 --> 03:03:16,118 THE SCALE OF ELECTRODES COULD 4408 03:03:16,118 --> 03:03:19,922 MATCH THE TYPE OF MEASUREMENTS AND 4409 03:03:19,922 --> 03:03:22,225 THE CIRCUITRY WE'RE INTERVENING 4410 03:03:22,225 --> 03:03:23,059 THE CIRCUITRY OF THE BRAIN IS 4411 03:03:23,059 --> 03:03:25,328 INSTRINSCALLY MULTISCALE. NOT JUST 4412 03:03:25,328 --> 03:03:28,865 A COLUMN WITHIN A NEURON. ITS ACROSS 4413 03:03:28,865 --> 03:03:34,770 SYNAPESES ACROSS REGIONS OF THE 4414 03:03:34,770 --> 03:03:35,972 BRAIN THAT SPREAD CENTIMETERS. 4415 03:03:35,972 --> 03:03:39,275 OUR HARDWARE SHOULD MATCH THE 4416 03:03:39,275 --> 03:03:40,443 PHYSICAL SCALES. 4417 03:03:40,443 --> 03:03:45,948 IN NEUROMORPHIC COMPUTING, THE 4418 03:03:45,948 --> 03:03:48,417 ALGORITHMS COULD BE THOUGHT OF AS 4419 03:03:48,417 --> 03:03:50,052 ALGORITHMS WHERE THE STATISTICS OF 4420 03:03:50,052 --> 03:03:52,755 THE ANALYSIS THAT YOU ARE PROCESSING 4421 03:03:52,755 --> 03:03:55,224 MATCHES THE NEUROLOGICAL STATISTICS. 4422 03:03:55,224 --> 03:03:58,861 THIS IS PRACTICAL -- I'M SORRY, 4423 03:03:58,861 --> 03:04:00,863 LET ME KNOW IF I'M RUNNING TOO 4424 03:04:00,863 --> 03:04:01,664 LONG. 4425 03:04:01,664 --> 03:04:02,999 MATCHING THE STATISTICS OF THE 4426 03:04:02,999 --> 03:04:04,300 BRAIN AND 4427 03:04:04,300 --> 03:04:05,468 MATCHING THE STATISTICS OF THE 4428 03:04:05,468 --> 03:04:07,036 BRAIN IS NOT STATIC. 4429 03:04:07,036 --> 03:04:08,671 THAT'S AN IMPORTANT THING TO 4430 03:04:08,671 --> 03:04:09,405 KNOW. 4431 03:04:09,405 --> 03:04:12,008 THE WAY YOUR BRAIN FUNCTIONS AND 4432 03:04:12,008 --> 03:04:13,509 THE NATURE OF THE CIRCUITS IN 4433 03:04:13,509 --> 03:04:14,810 YOUR BRAIN WHEN YOU ARE AWAKE 4434 03:04:14,810 --> 03:04:16,112 AND ASLEEP ARE DIFFERENT. THE WAY 4435 03:04:16,112 --> 03:04:19,148 IN WHICH YOU NEED TO INTERVENE 4436 03:04:19,148 --> 03:04:21,384 WITH THE BRAIN IN THOSE TWO 4437 03:04:21,384 --> 03:04:23,219 STATES IS VERY DIFFERENT. 4438 03:04:23,219 --> 03:04:30,760 THIS IS AN ARTICLE ABOUT 4439 03:04:30,760 --> 03:04:32,962 TRANSLATION FOR NEUROTECHNOLOGIES. 4440 03:04:32,962 --> 03:04:34,997 IT'S USEFUL. THERE'S A QR CODE. 4441 03:04:34,997 --> 03:04:41,404 THIS IDEA THAT EMBODIED HARDWARE 4442 03:04:41,404 --> 03:04:43,506 WOULD BE HAVING MICROELECTRODES 4443 03:04:43,506 --> 03:04:54,050 WHEN THE PHENOMENA CONCERN ACTION 4444 03:04:54,050 --> 03:04:58,020 POTENTIALS. OR IF YOU WANT TO 4445 03:04:58,020 --> 03:05:00,690 INTERVENE POPULATIONS, USE GRIDS ON THE BRAIN SURFACE. 4446 03:05:00,690 --> 03:05:04,427 THIS IS USEFUL FOR SIGNALS THAT AGGREGATE 10 -100,000 OF NEURONS 4447 03:05:04,427 --> 03:05:06,395 AND OBEY 1/F STATISTICS. YOU CAN LOOK 4448 03:05:06,395 --> 03:05:11,167 AT HOW BRAIN REGIONS INTERACT 4449 03:05:11,167 --> 03:05:12,368 DYNAMICALLY, STIMULATE AND MEASURE 4450 03:05:12,368 --> 03:05:13,035 RESPONSES FROM ONE AREA TO ANOTHER. 4451 03:05:13,035 --> 03:05:16,172 IF YOU'RE LOOKING AT THINGS LIKE 4452 03:05:16,172 --> 03:05:19,008 THE GENERAL HOMEOSTATIC STATE 4453 03:05:19,008 --> 03:05:23,012 OF THE BRAIN, THE WAY WE CONTROL 4454 03:05:23,012 --> 03:05:25,948 LARGE REGIONS FROM THALAMIC OUTPUT 4455 03:05:25,948 --> 03:05:27,616 TO CORTEX. YOU'LL THINK ABOUT 4456 03:05:27,616 --> 03:05:31,120 OSCILLATIONS. THESE ARE THINGS YOU 4457 03:05:31,120 --> 03:05:33,222 CAN PICK UP EVEN OUTSIDE THE HEAD. 4458 03:05:33,222 --> 03:05:36,092 THE SPATIAL SCALE OF ELECTRODES FOR 4459 03:05:36,092 --> 03:05:37,093 THESE CONTEXT ARE DIFFERENT FROM ONE ANOTHER 4460 03:05:37,093 --> 03:05:40,930 BUT THEY'RE TRYING TO INTRINSICALLY 4461 03:05:40,930 --> 03:05:43,366 MATCH THE ANATOMY OF THE BRAIN. 4462 03:05:43,366 --> 03:05:46,769 IT'S DIFFERENT FROM EMBODIMENT 4463 03:05:46,769 --> 03:05:49,038 OF LIMBS. BUT THE WAY YOU DESIGN 4464 03:05:49,038 --> 03:05:50,606 THE HARDWARE SHOULD MATCH IF 4465 03:05:50,606 --> 03:05:51,807 YOU'RE INSIDE THE BRAIN. 4466 03:05:51,807 --> 03:05:57,179 THIS IS AN EXAMPLE OF DECODING 4467 03:05:57,179 --> 03:05:58,681 VISUAL STIMULI. 4468 03:05:58,681 --> 03:06:00,850 YOU'LL SEE WHAT THE PATIENT IS 4469 03:06:00,850 --> 03:06:02,051 SEEING HERE. 4470 03:06:02,051 --> 03:06:03,686 THERE ARE ELECTRODES ON THE 4471 03:06:03,686 --> 03:06:05,955 OTHER SIDE OF THE BRAIN. 4472 03:06:05,955 --> 03:06:11,093 WE'RE MEASURING TWO TYPES OF 4473 03:06:11,093 --> 03:06:13,496 ELECTROPHYSIOLOGY. ONE THAT IS 4474 03:06:13,496 --> 03:06:15,965 FEEDFORWARD OUTPUT IN RESPONSE TO 4475 03:06:15,965 --> 03:06:19,769 VISUAL STIMULUS. THE OTHER IS THIS 4476 03:06:19,769 --> 03:06:20,870 1/F PROCESS REFLECTIVE OF LOCAL CIRCUITRY. 4477 03:06:20,870 --> 03:06:24,340 WE CAN GET REALTIME INFORMATION 4478 03:06:24,340 --> 03:06:26,475 FROM EACH LOCATION AND PLOT THEM 4479 03:06:26,475 --> 03:06:28,544 INTO A LARGER DIMENSIONAL FEATURE SPACE. 4480 03:06:28,544 --> 03:06:30,946 IT'S IMPORTANT IN THE BRAIN TO 4481 03:06:30,946 --> 03:06:33,582 REMEMBER THAT THIS FEATURE SPACE 4482 03:06:33,582 --> 03:06:35,551 IS NOT STATIC. CURRENT SYSTEMS FOR 4483 03:06:35,551 --> 03:06:38,521 DECODING THE BRAIN ARE RELIANT ON 4484 03:06:38,521 --> 03:06:43,726 DAILY CALIBRATION, ROUTINES THAT 4485 03:06:43,726 --> 03:06:47,063 DETERMINE NEW SETS OF PARAMETERS FOR ALGORITHMIC DECODING. 4486 03:06:47,063 --> 03:06:47,596 THIS IS BURDENSOME, NOT PRODUCTIVE. 4487 03:06:47,596 --> 03:06:49,965 MOST OF THE BRAIN COMPUTER 4488 03:06:49,965 --> 03:06:50,866 INTERFACING ONLY WORKS WHEN YOU 4489 03:06:50,866 --> 03:06:53,302 HAVE A SCIENTIST IN THE ROOM 4490 03:06:53,302 --> 03:06:56,605 WITH A PATIENT RECALIBRATING A 4491 03:06:56,605 --> 03:06:58,874 MACHINE EVERY DAY. 4492 03:06:58,874 --> 03:07:01,510 INSTEAD OUR MACHINES NEED TO 4493 03:07:01,510 --> 03:07:03,145 RECALIBRATE THEMSELVES FIRST OF 4494 03:07:03,145 --> 03:07:05,548 ALL BY TRYING TO DECODE THE 4495 03:07:05,548 --> 03:07:08,717 BRAIN IN A MULTISCALE WAY. YOU 4496 03:07:08,717 --> 03:07:12,154 NEED TO ENCODE NOT JUST THE DOMAIN 4497 03:07:12,154 --> 03:07:15,091 SPECIALIST AREA -- HAND AREA -- BUT 4498 03:07:15,091 --> 03:07:17,893 ALSO THE GENERAL STATE OF THE MOTOR CORTEX. 4499 03:07:17,893 --> 03:07:26,836 WE CAN PROBE, REPARAMETIZE DECODING IN REALTIME AND FEED OURALGORITHMS. 4500 03:07:26,836 --> 03:07:37,413 AND THIS IS WHAT WE SHOULD THINK OF 4501 03:07:38,114 --> 03:07:40,049 AS NEUROMORPHIC ALGORITHM WHEN DECODING BRAIN ACTIVITY. 4502 03:07:40,049 --> 03:07:43,919 THIS WOULD BE USEFUL FOR CLOSED-LOOP DEVICES THAT I WANT TO BEIMPLANTING INTO THE BRAIN. 4503 03:07:43,919 --> 03:07:44,854 [APPLAUSE] 4504 03:07:46,756 --> 03:07:48,824 THANK YOU VERY MUCH. 4505 03:07:48,824 --> 03:07:50,092 AT THIS 4506 03:07:51,227 --> 03:07:52,561 >> THANK YOU VERY MUCH. 4507 03:07:52,561 --> 03:07:57,600 AT THIS TIME I WOULD LIKE TO GET 4508 03:07:57,600 --> 03:08:03,572 GIACOMO INDIVERI'S RECORDING. 4509 03:08:03,572 --> 03:08:05,574 >> HELLO, I'M GOING TO PRESENT 4510 03:08:05,574 --> 03:08:09,044 AN OUTLOOK OF NEXT GENERATION 4511 03:08:09,044 --> 03:08:13,916 BRAIN MACHINES INTERFACES. UNLIKE 4512 03:08:13,916 --> 03:08:16,585 MACHINE WORKING WORKLOADS, THEY 4513 03:08:16,585 --> 03:08:21,724 HAVE STRINGENT HARDWARE REQUIREMENTS, 4514 03:08:21,724 --> 03:08:23,926 SMALL POWER BUDGET, LOW LATENCY, CLOSED-LOOP 4515 03:08:23,926 --> 03:08:25,961 MULTI-SCALE DYNAMICS, INPUTS WITH 4516 03:08:25,961 --> 03:08:31,801 A LARGE DYNAMIC AND VARIABLE RANGE. 4517 03:08:31,801 --> 03:08:33,903 FROM THE ALGORITHMIC POINT OF VIEW 4518 03:08:33,903 --> 03:08:37,840 THESE INTERFACES NEED TO BE ALWAYS 4519 03:08:37,840 --> 03:08:41,210 ON. SENSING DATA, DECODING, PROCESSING, 4520 03:08:41,210 --> 03:08:43,012 ENCODING, POSSIBLY STIMULATING 4521 03:08:43,012 --> 03:08:45,381 CONTINUOUSLY. UNLIKE CONVENTIONAL 4522 03:08:45,381 --> 03:08:47,716 WORKLOADS, THE TYPE OF DATA USED OR 4523 03:08:47,716 --> 03:08:51,086 IS AVAILABLE FOR THESE DEVICES IS 4524 03:08:51,086 --> 03:08:51,720 LIMITED. 4525 03:08:51,720 --> 03:08:54,490 THERE ARE NO LARGE DATA SETS 4526 03:08:54,490 --> 03:08:55,824 USED FOR TRAINING. 4527 03:08:55,824 --> 03:08:58,661 IN ADDITION, THESE DEVICES NEED 4528 03:08:58,661 --> 03:09:01,564 TO BE ADAPTED AND HAVE CONTINUAL 4529 03:09:01,564 --> 03:09:05,801 LEARNING, NOT INCREMENTAL LEARNING 4530 03:09:05,801 --> 03:09:07,670 AND CHANGING THEIR PROPERTIES 4531 03:09:07,670 --> 03:09:09,205 CONTINUOUSLY THROUGH TIME AS THE 4532 03:09:09,205 --> 03:09:11,173 DATA COMES THROUGH. 4533 03:09:11,173 --> 03:09:14,376 WITH STREAMING DATA, THE DATA IS 4534 03:09:14,376 --> 03:09:17,213 CALLED BATCH-SIZE ONE WHICH IS 4535 03:09:17,213 --> 03:09:19,081 NOT COMMONLY USED IN DEEP NETWORKS. 4536 03:09:19,081 --> 03:09:21,450 BUT THIS IS A HARD REQUIREMENT FOR 4537 03:09:21,450 --> 03:09:28,057 BRAIN MACHINE INTERFACES. 4538 03:09:28,057 --> 03:09:31,493 IF WE LOOK AT THE FIRST TWO 4539 03:09:31,493 --> 03:09:34,096 REQUIREMENTS, THE EXTREMELY 4540 03:09:34,096 --> 03:09:39,335 SMALL POWER BUDGET AND THE 4541 03:09:39,335 --> 03:09:39,868 ALWAYS ON PROCESSING, IT'S CLEAR 4542 03:09:39,868 --> 03:09:42,938 THAT NEXT GENERATION BMI NEEDS TO BE 4543 03:09:42,938 --> 03:09:45,574 DESIGNED WITH BOTTOM-UP, PHYSICS- 4544 03:09:45,574 --> 03:09:49,111 BASED APPROACH TO SQUEEZE EVERY JOULE 4545 03:09:49,111 --> 03:09:51,413 OUT OF THE SYSTEM. WE NEED TO USE 4546 03:09:51,413 --> 03:09:53,949 DIGITAL AND ANALOG CIRCUITS OPERATING 4547 03:09:53,949 --> 03:09:56,885 ON MULTIPLE TIME SCALES FOR ENDOWING 4548 03:09:56,885 --> 03:10:00,022 THESE SYSTEMS WITH THE ABILITY TO 4549 03:10:00,022 --> 03:10:02,391 PROCESS SIGNALS CONTINUOUSLY WITH 4550 03:10:02,391 --> 03:10:05,294 LARGE DYNAMIC RANGE, WITH STREAMING 4551 03:10:05,294 --> 03:10:08,931 DATA, CONTINUOUS LEARNING, AND SO ON. 4552 03:10:08,931 --> 03:10:12,534 ONE PROMISING APPROACH IS THE 4553 03:10:12,534 --> 03:10:14,637 NEUROMORPHIC COMPUTING ONE. 4554 03:10:14,637 --> 03:10:17,172 WE HAVE BEEN ANALYZING THESE MIXED 4555 03:10:17,172 --> 03:10:18,674 SIGNAL ANALOG / DIGITAL CHIPS FOR 4556 03:10:18,674 --> 03:10:21,777 SOME TIME. WE DESIGN CHIPS USING 4557 03:10:21,777 --> 03:10:23,245 ANALOG SUBTHRESHOLD CIRCUITS TO 4558 03:10:23,245 --> 03:10:25,614 IMPLEMENT SPIKING NEURAL NETWORKS. 4559 03:10:25,614 --> 03:10:32,421 TYPICALLY THE NEURON IS AT THE SIDE 4560 03:10:32,421 --> 03:10:34,857 OF THE ARRAY. A LARGE NUMBER OF SYNAPSES 4561 03:10:34,857 --> 03:10:38,127 RECEIVE INPUTS, ITEGRATE CURRENTS 4562 03:10:38,127 --> 03:10:40,529 USING ANALOG DYNAMICS TO PRODUCE 4563 03:10:40,529 --> 03:10:43,032 SPIKES. THESE CIRCUITS ARE SLOW. THEY 4564 03:10:43,032 --> 03:10:47,569 HAVE THE SAME TIME CONSTANTS AS 4565 03:10:47,569 --> 03:10:50,172 THE BIOLOGICAL SYSTEMS WE WANT THEM 4566 03:10:50,172 --> 03:10:54,109 TO INTERACT WITH: NEURONS, MUSCLES. 4567 03:10:54,109 --> 03:10:57,846 BECAUSE OF THE WAY THEY'RE BUILT BY 4568 03:10:57,846 --> 03:11:03,452 COPY PASTING MANY NEURONS TOGETHER. 4569 03:11:03,452 --> 03:11:05,287 THEY ARE MASSIVELY PARALLE AND 4570 03:11:05,287 --> 03:11:08,290 COMPATIBLE WITH EMERGING MEMORY TECHNOLOGIES. 4571 03:11:08,290 --> 03:11:11,860 HOWEVER, THEY ARE NOISY. ANALOG CIRCUITS 4572 03:11:11,860 --> 03:11:13,562 HAVE DEVICE MISMATCH. EACH NEURON IS 4573 03:11:13,562 --> 03:11:17,700 DIFFERENT. THEY'RE HIGHLY VARIABLE. 4574 03:11:17,700 --> 03:11:21,070 ONCE THESE ANALOG SIGNALS PRODUCE 4575 03:11:21,070 --> 03:11:23,038 AN ACTION POTENTIAL, WE CONVERT THEM 4576 03:11:23,038 --> 03:11:26,041 INTO SPIKES AND WE CAN USE DIGITAL 4577 03:11:26,041 --> 03:11:29,945 NETWORKS TO CONFIGURE THE WAY 4578 03:11:29,945 --> 03:11:31,547 THESE NEURONS ARE CONNECTED TO 4579 03:11:31,547 --> 03:11:36,218 EACH OTHER USING ROUTERS. 4580 03:11:36,218 --> 03:11:37,886 A POINT THAT IS IMPORTANT IS THAT 4581 03:11:37,886 --> 03:11:39,254 THESE NEURONS AND SYNAPSES ARE 4582 03:11:39,254 --> 03:11:42,725 VARIABLE. THERE IS A LARGE AMOUNT OF 4583 03:11:42,725 --> 03:11:48,497 DEVICE MISMATCH. 4584 03:11:48,497 --> 03:11:50,466 FORTUNATELY BIOLOGY CAN HELP US 4585 03:11:50,466 --> 03:11:55,637 BECAUSE IT HAS THE SAME TYPE OF 4586 03:11:55,637 --> 03:11:57,072 CHARACTERISTICS. NEURONS ARE VARIABLE 4587 03:11:57,072 --> 03:11:58,874 IN ANALOG CIRCUITS, HERE WE SEE 4588 03:11:58,874 --> 03:12:06,982 AN EXAMPLE OF CHIP MEASUREMENTS 4589 03:12:06,982 --> 03:12:08,684 COMPRISING 256 NEURONS. WE APPLY THE 4590 03:12:08,684 --> 03:12:10,652 SAME INPUT TO ALL NEURONS. NEURONS 4591 03:12:10,652 --> 03:12:12,654 ARE SLIGHTLY DIFFERENT WHEN THEY'RE 4592 03:12:12,654 --> 03:12:16,024 FABRICATED. THE TIME TO FIRST SPIKE 4593 03:12:16,024 --> 03:12:18,494 IS HIGHLY VARIABILITY -- THE 4594 03:12:18,494 --> 03:12:20,896 COEFFICIENT OF VARIATION IS ABOUT 20%. 4595 03:12:20,896 --> 03:12:23,599 NOW, HOW CAN WE DO A ROBUST 4596 03:12:23,599 --> 03:12:26,135 COMPUTATION USING THIS TYPE OF 4597 03:12:26,135 --> 03:12:27,536 VARIABILITY PRESENT IN OUR 4598 03:12:27,536 --> 03:12:28,771 COMPUTING SUBSTRATE? 4599 03:12:28,771 --> 03:12:31,373 ONE THING THAT COMES TO MIND IS 4600 03:12:31,373 --> 03:12:32,641 TO USE AVERAGING. 4601 03:12:32,641 --> 03:12:35,010 IF WE AVERAGE POPULATIONS OF 4602 03:12:35,010 --> 03:12:36,879 NEURONS, WE CAN REDUCE THE 4603 03:12:36,879 --> 03:12:39,648 VARIABILITY BY THE 4604 03:12:39,648 --> 03:12:42,518 SQUARE ROOT OF THE NUMBER OF 4605 03:12:42,518 --> 03:12:43,285 NEURNS WE AVERAGE WITH. 4606 03:12:43,285 --> 03:12:48,624 IF WE USE NONLINEAR DYNAMICS, 4607 03:12:48,624 --> 03:12:51,693 IT'S USEFUL AND IT'S BEEN SHOWN 4608 03:12:51,693 --> 03:12:56,799 THAT THESE TYPES OF NONLINEARITIES 4609 03:12:56,799 --> 03:12:58,567 CAN REDUCE VARIANCE BY A FACT OR N 4610 03:12:58,567 --> 03:13:00,002 NOT SQUARE ROOT OF N. IT'S POWERFUL. 4611 03:13:00,002 --> 03:13:05,641 THE OTHER MECHANISM FOR ENDOWING THE 4612 03:13:05,641 --> 03:13:06,375 SYSTEM WITH ROBUSTNESS IS LEARNING. 4613 03:13:06,375 --> 03:13:07,910 WE QUANTIFY THIS VARIABILITY IN 4614 03:13:07,910 --> 03:13:12,581 OUR CHIPS AND THOUGHT WHAT IF WE 4615 03:13:12,581 --> 03:13:14,416 AVERAGE TWO, FOUR, EIGHT NEURONS? 4616 03:13:14,416 --> 03:13:17,753 AND WE MEASURE THE COEFFICIENT 4617 03:13:17,753 --> 03:13:21,089 OF VARIATION AS A FUNCTION OF 4618 03:13:21,089 --> 03:13:22,291 POPULATION SIZE AND INTEGRATION TIME. 4619 03:13:22,291 --> 03:13:22,658 HERE THE COEFFICIENT OF VARIATION 4620 03:13:22,658 --> 03:13:27,463 GOES FROM 20% WITH ONE NEURON 4621 03:13:27,463 --> 03:13:30,632 TO 18% DOWN TO LESS THAN 1%. 4622 03:13:30,632 --> 03:13:36,205 ONCE WE HAVE THIS NUMBER, WE CAN 4623 03:13:36,205 --> 03:13:37,840 CONVERT INTO EQUIVALENT BIT RESOLUTION COLOR CODED IN THIS PICTURE. 4624 03:13:37,840 --> 03:13:41,276 IF WE NEED 8 BIT RESOLUTION, WE NEED 4625 03:13:41,276 --> 03:13:44,646 CLUSTERS OF 8 OR 16 NEURONS AND 4626 03:13:44,646 --> 03:13:47,216 AVERAGE OVER 100 MILLISECONDS 4627 03:13:47,216 --> 03:13:48,517 IT'S EASY TO HAVE HIGH PRECISION 4628 03:13:48,517 --> 03:13:50,919 USING HIGHLY VARIABLE COMPUTING 4629 03:13:50,919 --> 03:13:55,190 SUBSTRATES. 4630 03:13:55,190 --> 03:14:01,997 BIOLOGY USE POPULATION CODES. 4631 03:14:01,997 --> 03:14:03,866 WE KNOW FROM BIOLOGY HOW TO USE 4632 03:14:03,866 --> 03:14:06,368 POPULATION TO ENCODE SIGNALS. THIS IS 4633 03:14:06,368 --> 03:14:07,803 AN EXAMPLE MEASURED FROM NON-HUMAN 4634 03:14:07,803 --> 03:14:10,539 PRIMATES, DATA MEASURED FROM CAT 4635 03:14:10,539 --> 03:14:13,709 VISUAL CORTEX AND OUR CHIPS WHERE WE 4636 03:14:13,709 --> 03:14:16,011 HAD FEEDFORWARD OR RECURRENT NETWORKS 4637 03:14:16,011 --> 03:14:22,284 IN WHICH WE CAN AMPLIFY THE SIGNAL 4638 03:14:22,284 --> 03:14:23,252 AND SUPPRESS THE NODES WITH SELECTIVE 4639 03:14:23,252 --> 03:14:25,320 AMPLIFICATION. WE CAN USE A 4640 03:14:25,320 --> 03:14:26,722 POPULATION OF NEURONS, AVERAGE OVER 4641 03:14:26,722 --> 03:14:34,329 THE FIRING RATE TO GET A PRECISE 4642 03:14:34,329 --> 03:14:35,931 ESTIMATE OF THE FEATURE PRESENTED AS 4643 03:14:35,931 --> 03:14:37,599 INPUT. IT COULD BE THE TONE OF THE 4644 03:14:37,599 --> 03:14:43,138 SOUND OR STIMULUS OR THE 4645 03:14:43,138 --> 03:14:44,706 PRESSURE ON THE SKIN AND SO ON. 4646 03:14:44,706 --> 03:14:45,207 NOW THAT WE HAVE POPULATIONS, WE 4647 03:14:45,207 --> 03:14:47,376 CAN CREATE SYSTEMS AND 4648 03:14:47,376 --> 03:14:48,644 APPLICATIONS BUT HOW DO WE DO 4649 03:14:48,644 --> 03:14:51,813 THAT? HOW DO WE GO FROM BRIDGING 4650 03:14:51,813 --> 03:14:52,281 BETWEEN POPULATIONS OF NEURONS 4651 03:14:52,281 --> 03:14:54,650 AND CREATING ALGORITHMS FOR 4652 03:14:54,650 --> 03:14:57,753 PROGRAMMING NEUROMORPHIC SYSTEMS 4653 03:14:57,753 --> 03:15:00,422 TO SOLVE PRACTICAL APPLICATIONS? 4654 03:15:00,422 --> 03:15:04,927 HOW DO WE BUILD A NEUROMORPHIC 4655 03:15:04,927 --> 03:15:06,028 SOLUTION STACK? WE DON'T PROGRAM 4656 03:15:06,028 --> 03:15:08,463 COMPUTERS BY LOOKING AT SINGLE 4657 03:15:08,463 --> 03:15:13,101 TRANSISTORS, WE ABSTRACT WITH LOGIC GATES. 4658 03:15:13,101 --> 03:15:16,505 WITH HIERARCHY OF ABSTRACTIONS, WE GO 4659 03:15:16,505 --> 03:15:19,341 FROM MACHINE LANGUAGE TO COMPILERS TO 4660 03:15:19,341 --> 03:15:27,549 SOFTWARE. IN BIOLOGY, WE WOULD THINK 4661 03:15:27,549 --> 03:15:30,519 AT THE LEVEL OF NETWORKS. WE'VE BEEN 4662 03:15:30,519 --> 03:15:32,120 LOOKING AT DIFFERENT NETWORKS FROM 4663 03:15:32,120 --> 03:15:34,122 COMPUTATIONAL NEUROSCIENCE TO BUILD 4664 03:15:34,122 --> 03:15:35,924 A SOLUTION STACK. LACK OF TIME, WILL 4665 03:15:35,924 --> 03:15:39,828 MENTION WINNER-TAKE-ALL NETWORKS 4666 03:15:39,828 --> 03:15:41,530 AND NEURAL STATE MACHINES USED BY 4667 03:15:41,530 --> 03:15:46,168 THE COMMUNITY. WINNER-TAKE-ALL IS A 4668 03:15:46,168 --> 03:15:49,438 USEFUL COMPUTATIONAL PRIMITIVE 4669 03:15:49,438 --> 03:15:51,473 BECAUSE THEY CAN BE USED TO PROCESS 4670 03:15:51,473 --> 03:15:54,209 SIGNALS IN THE ANALOG DOMAIN LIKE 4671 03:15:54,209 --> 03:15:56,845 FOR SOUND, VISION, AND PRESSURE OR 4672 03:15:56,845 --> 03:15:59,581 NEURAL RECORDINGS. YOU CAN AMPLIFY OR 4673 03:15:59,581 --> 03:16:02,551 FILTER SIGNALS. BY CHANGING THE 4674 03:16:02,551 --> 03:16:09,791 PARAMETERS ON THE SAME NETWORK, YOU 4675 03:16:09,791 --> 03:16:10,721 CAN IMPLEMENT NONLINEAR BEHAVIORS: 4676 03:16:10,721 --> 03:16:11,916 SELECTIVE AMPLIFICATION, SIGNALRESTORATION, 4677 03:16:11,916 --> 03:16:12,361 MULTISTABILITY. 4678 03:16:12,361 --> 03:16:15,998 WE CAN GO FROM PROCESSING WAVES TO 4679 03:16:15,998 --> 03:16:19,901 MANIPULATING SYMBOLS, FROM SIGNAL 4680 03:16:19,901 --> 03:16:24,640 PROCESSING TO COMPUTATION. 4681 03:16:24,640 --> 03:16:27,075 IN NEURAL STATE MACHINES, BY HAVING 4682 03:16:27,075 --> 03:16:31,747 WINNER-TAKE-ALL NETWORKS, WE CAN HAVE 4683 03:16:31,747 --> 03:16:34,783 ELEMENTS WITH PERSISTING ACTIVITIES. 4684 03:16:34,783 --> 03:16:39,721 AND WE CAN SWITCH FROM ONE STATE TO 4685 03:16:39,721 --> 03:16:50,198 ANOTHER BY CHANGING THE INPUT. DATA 4686 03:16:57,906 --> 03:16:59,708 FROM CHIPS PROBE COGNITION IN NON-HUMAN PRIMATES. 4687 03:16:59,708 --> 03:17:02,377 USING POPULATIONS OF NEURONS IS ONE 4688 03:17:02,377 --> 03:17:04,279 WAY TO GET ROBUST COMPUTATION. 4689 03:17:04,279 --> 03:17:08,350 SINGLE NEURONS ARE NOISY. 4690 03:17:08,350 --> 03:17:11,353 LEARNING IS THE OTHER MECHANISM. 4691 03:17:11,353 --> 03:17:12,788 WE'VE IMPLEMENTED LEARNING RULES 4692 03:17:12,788 --> 03:17:16,625 USING POPULATIONS OF NEURONS. 4693 03:17:16,625 --> 03:17:24,833 INHIBITORY NEURONS ARE IMPORTANT IN 4694 03:17:24,833 --> 03:17:27,235 ROBUST LEARNING ALGORITHMS ON CHIP. 4695 03:17:27,235 --> 03:17:32,307 I CAN POINT YOU TO THE PAPERS 4696 03:17:32,307 --> 03:17:34,509 WITH MORE DETAILS THAN THIS. 4697 03:17:34,509 --> 03:17:36,712 BY COMBINING POPULATIONS OF NEURONS 4698 03:17:36,712 --> 03:17:38,246 LEARNING WE CAN GO FROM ALGORITHMS TO 4699 03:17:38,246 --> 03:17:40,649 APPLICATIONS. REMEMBER NEUROMORPHIC 4700 03:17:40,649 --> 03:17:42,117 COMPUTING NEEDS TO USE THE LANGUAGE 4701 03:17:42,117 --> 03:17:44,386 OF NEUROSCIENCE AND DYNAMICAL SYSTEMS 4702 03:17:44,386 --> 03:17:46,121 THE COMPUTATION NEEDS TO BE EXPRESSED 4703 03:17:46,121 --> 03:17:50,992 AS SIGNAL PROCESSING STEPS FROM WAVES 4704 03:17:50,992 --> 03:17:53,695 TO SYMBOLS. ALGORITHMS CAN BE CREATED 4705 03:17:53,695 --> 03:17:54,596 COMPOSING PRIMITIVES AND IMPLEMENTING 4706 03:17:54,596 --> 03:17:56,231 LEARNING. BY FOLLOWING THIS APPROACH, 4707 03:17:56,231 --> 03:17:58,066 WE'VE BEEN SHOWING THAT IT'S 4708 03:17:58,066 --> 03:18:00,268 POSSIBLE TO SOLVE SEVERAL TYPES 4709 03:18:00,268 --> 03:18:03,372 OF PROBLEMS IN THE DOMAIN OF BMI 4710 03:18:03,372 --> 03:18:07,309 GOING FROM HEART RATE VARIABILITY 4711 03:18:07,309 --> 03:18:09,511 TO EMG DECODING TO HFO DETECTION. 4712 03:18:09,511 --> 03:18:12,547 I WOULD JUST LIKING TO GO 4713 03:18:12,547 --> 03:18:14,015 QUICKLY OVER HFO IF 4714 03:18:14,015 --> 03:18:15,717 TIME ALLOWS TO TELL YOU THAT 4715 03:18:15,717 --> 03:18:18,186 IT'S POSSIBLE TO IMPLEMENT 4716 03:18:18,186 --> 03:18:21,790 NETWORKS THAT CAN SOLVE PROBLEMS 4717 03:18:21,790 --> 03:18:23,558 RELATED TO EPILEPTIC SEIZURES. 4718 03:18:23,558 --> 03:18:24,693 HFOS ARE HIGH FREQUENCY OSCILLATIONS 4719 03:18:24,693 --> 03:18:28,230 IN THE HIGH FREQUENCY BAND OF 4720 03:18:28,230 --> 03:18:34,936 EEG DATA AND TRAINED DOCTORS CAN 4721 03:18:34,936 --> 03:18:36,738 DISTINGUISH TRUE FROM FALSE HFOS. 4722 03:18:36,738 --> 03:18:39,040 WE CREATED A NETWORK TO DISTINGUISH 4723 03:18:39,040 --> 03:18:43,645 THIS BY IMPLEMENTING IT ON CHIP WITH 4724 03:18:43,645 --> 03:18:46,915 LOW OISE AMPLIFIERS, FILTERS, SNNS. 4725 03:18:46,915 --> 03:18:52,220 BY COMBINING THESE SIGNALS WITH 4726 03:18:52,220 --> 03:18:57,692 NONLINEAR SYNAPSES AND SPIKING NEURONS 4727 03:18:57,692 --> 03:18:59,628 WE'RE ABLE TO DETECT TRUE HFOS FROM 4728 03:18:59,628 --> 03:19:03,365 FALSE HFOS WITHOUT BACKPROPAGATION 4729 03:19:03,365 --> 03:19:10,806 WE DID HAVE TO MATCH THE TIME 4730 03:19:10,806 --> 03:19:12,007 CONSTANT OF THE SYNAPSES TO THE DATA. 4731 03:19:12,007 --> 03:19:22,017 JUST A SINGLE CHIP - 614 MICROWATTS 4732 03:19:22,017 --> 03:19:24,453 ONLY 78% ACCURACY IS ENOUGH TO GIVE 4733 03:19:24,453 --> 03:19:27,055 THE SURGEON AN EXTRA CUE FOR 4734 03:19:27,055 --> 03:19:32,594 MAKING A DECISION. 4735 03:19:32,594 --> 03:19:33,829 TO CONCLUDE, I WOULD LIKE TO 4736 03:19:33,829 --> 03:19:35,997 APPOINT OUT THAT IT'S POSSIBLE 4737 03:19:35,997 --> 03:19:42,270 TO BUILD INTELLIGENT BMIS AND 4738 03:19:42,270 --> 03:19:44,406 EFFICIENT INTELLIGENT BMIS. 4739 03:19:44,406 --> 03:19:48,543 IF WE FOLLOW THIS PHYSICS-BASED 4740 03:19:48,543 --> 03:19:50,712 APPROACH AND EXPLORE ALL THE 4741 03:19:50,712 --> 03:19:52,147 PROPERTIES OF THE SILICON, AND 4742 03:19:52,147 --> 03:19:55,617 ALSO IF WE LISTEN TO THE BRAIN 4743 03:19:55,617 --> 03:19:57,886 AND USE BIOPHYSICALLY REALISTIC 4744 03:19:57,886 --> 03:19:59,888 COMPUTATIONAL MODELS THAT ARE 4745 03:19:59,888 --> 03:20:03,425 WELL-MATCHED TO THE PHYSICS OF THE 4746 03:20:03,425 --> 03:20:05,961 COMPUTING SUBSTRATE AND TEST THEM. 4747 03:20:05,961 --> 03:20:08,830 THIS PROCESS OF DESIGNING MODELS 4748 03:20:08,830 --> 03:20:11,566 AND TESTING THEM AND REFINING 4749 03:20:11,566 --> 03:20:13,101 THEM IS SOMETHING WE'VE BEEN 4750 03:20:13,101 --> 03:20:15,770 DOING FOR THE PAST 10 YEARS AND 4751 03:20:15,770 --> 03:20:17,506 SOMETHING THAT CAN LEAD TO THE 4752 03:20:17,506 --> 03:20:21,610 NEXT GENERATION OF BRAIN MACHINE 4753 03:20:21,610 --> 03:20:22,143 INTERFACES. 4754 03:20:22,143 --> 03:20:23,979 I THANK YOU FOR YOUR ATTENTION. 4755 03:20:23,979 --> 03:20:33,588 I'M HAPPY TO TAKE QUESTIONS. 4756 03:20:33,588 --> 03:20:38,426 >> RALPH ETIENNE-CUMMINGS 4757 03:20:38,426 --> 03:20:39,728 FROM JOHNS HOPKINS UNIVERSITY. 4758 03:20:39,728 --> 03:20:43,965 >> I COME AT THE END WHICH MEANS 4759 03:20:43,965 --> 03:20:47,202 THAT MY THUNDER IS MOSTLY 4760 03:20:47,202 --> 03:20:47,903 STOLEN. 4761 03:20:47,903 --> 03:20:49,671 NONETHELESS I HOPE I CAN PULL IT 4762 03:20:49,671 --> 03:20:52,941 ALL TOGETHER IN A WAY THAT WILL 4763 03:20:52,941 --> 03:20:56,344 PROVIDE SOME DIRECTIONS ON WHAT 4764 03:20:56,344 --> 03:21:05,020 WE THINK NEUROMORPHIC IN MEDICINE 4765 03:21:05,020 --> 03:21:06,221 WILL LOOK LIKE GOING FORWARD. THIS IS 4766 03:21:06,221 --> 03:21:11,726 A READOUT FROM A NSF-NIH SPONSORED 4767 03:21:11,726 --> 03:21:12,160 WORKSHOP IN BALTIMORE IN OCTOBER. 4768 03:21:12,160 --> 03:21:16,665 WE WERE ESSENTIALLY LOOKING FOR 4769 03:21:16,665 --> 03:21:21,903 NEUROMORPHIC PRINCIPLES 4770 03:21:21,903 --> 03:21:23,571 ESSENTIAL FOR HEALTHCARE. 4771 03:21:23,571 --> 03:21:28,977 THIS WORKSHOP WAS ORGANIZED BY 4772 03:21:28,977 --> 03:21:32,147 THE FOLLOWING FOLKS. 4773 03:21:32,147 --> 03:21:34,683 IN PARTICULAR GRACE WAS ONE 4774 03:21:34,683 --> 03:21:36,885 OF THE ORGANIZERS AS WELL AND 4775 03:21:36,885 --> 03:21:39,721 KEPT US TO TASK TO MAKE SURE 4776 03:21:39,721 --> 03:21:43,258 THAT WE DID THE RIGHT THINGS. 4777 03:21:43,258 --> 03:21:46,428 IT'S A COMBINATION OF 4778 03:21:46,428 --> 03:21:48,129 ENGINEERS AND PHYSICIANS. 4779 03:21:48,129 --> 03:21:52,267 WE HAD THE GOOD FORTUNE OF A 4780 03:21:52,267 --> 03:21:54,669 SCIENCE WRITER 4781 03:21:54,669 --> 03:21:56,271 WHO WILL PUT TOGETHER THE 4782 03:21:56,271 --> 03:22:02,777 REPORT AT THE END, AND ALSO SHE 4783 03:22:02,777 --> 03:22:05,046 RAN THE QUESTIONING AND THAT WAS 4784 03:22:05,046 --> 03:22:06,381 INTERACTIVE WITH A LOT OF 4785 03:22:06,381 --> 03:22:06,982 DEBATES. 4786 03:22:06,982 --> 03:22:08,984 WHAT WAS THE WORKSHOP ABOUT? 4787 03:22:08,984 --> 03:22:11,753 IT WAS ABOUT BRINGING TOGETHER 4788 03:22:11,753 --> 03:22:13,121 COMMUNITIES THAT SPEAK ABOUT 4789 03:22:13,121 --> 03:22:17,892 WAYS THAT WE CAN LEARN FROM 4790 03:22:17,892 --> 03:22:20,362 NEUROSCIENCE AND BIOMEDICAL 4791 03:22:20,362 --> 03:22:22,197 ENGINEERING IN ORDER TO PUT 4792 03:22:22,197 --> 03:22:25,600 TOGETHER IDEAS THAT WOULD BE 4793 03:22:25,600 --> 03:22:27,602 HELPFUL IN SOLVING PROBLEMS IN 4794 03:22:27,602 --> 03:22:27,936 MEDICINE. 4795 03:22:27,936 --> 03:22:30,705 WE WANTED A ROAD MAP TO SAY 4796 03:22:30,705 --> 03:22:35,043 WHAT SHOULD BE THE POTENTIAL PLACES 4797 03:22:35,043 --> 03:22:36,578 FUNDING AGENCIES CAN THINK ABOUT 4798 03:22:36,578 --> 03:22:38,713 FUNDING GOING FORWARD. 4799 03:22:38,713 --> 03:22:40,382 THE ORGANIZATION -- YOU CAN SEE 4800 03:22:40,382 --> 03:22:42,584 HERE, IT WAS ESSENTIALLY BASED 4801 03:22:42,584 --> 03:22:44,853 ON A FLOW THAT STARTED CENTRALLY 4802 03:22:44,853 --> 03:22:45,754 FROM THE CORTEX. 4803 03:22:45,754 --> 03:22:53,194 WE TALKED ABOUT VARIOUS ASPECTS. 4804 03:22:53,194 --> 03:22:57,132 WE STARTED WITH A WONDERFUL 4805 03:22:57,132 --> 03:23:03,972 KEYNOTE SPEECH FROM TIM DENISON. 4806 03:23:03,972 --> 03:23:06,541 WE SAW THE REVIEW PAPER THAT KAI 4807 03:23:06,541 --> 03:23:10,045 TALKED ABOUT THAT THEY PUBLISHED. 4808 03:23:10,045 --> 03:23:11,446 TIM PROVIDED A FUNDAMENTAL REVIEW OF 4809 03:23:11,446 --> 03:23:13,782 WHAT CAN BE DONE AND WHAT HE LOOKS 4810 03:23:13,782 --> 03:23:15,116 FORWARD GOING FORWARD. 4811 03:23:15,116 --> 03:23:17,719 FROM THERE WE WENT INTO 4812 03:23:17,719 --> 03:23:18,620 PROSTHETICS AND ONE OF THE 4813 03:23:18,620 --> 03:23:22,957 HIGHLIGHTS FOR ME WAS THE WORK 4814 03:23:22,957 --> 03:23:28,263 THAT RANU JUNG IS DOING FROM THE 4815 03:23:28,263 --> 03:23:31,566 UNIVERSITY OF ARKANSAS. IT'S SOME OF 4816 03:23:31,566 --> 03:23:35,737 THE MOST ADVANCED REAL WORLD 4817 03:23:35,737 --> 03:23:37,772 NEUROPROSTHETICS. A LOT OF GREAT WORK 4818 03:23:37,772 --> 03:23:40,708 BY OTHER PEOPLE AS WELL. 4819 03:23:40,708 --> 03:23:42,844 AS WE MOVE TO THE NEXT DAY, WE 4820 03:23:42,844 --> 03:23:46,681 WENT INTO MATERIALS AND 4821 03:23:46,681 --> 03:23:47,715 ARCHITECTURES. 4822 03:23:47,715 --> 03:23:50,351 WE CLOSED OUT WITH APPLICATIONS 4823 03:23:50,351 --> 03:23:53,388 AND COMPUTATIONS. 4824 03:23:53,388 --> 03:23:55,890 AND, AGAIN, THERE WAS A LOT OF 4825 03:23:55,890 --> 03:23:56,124 DEBATE. 4826 03:23:56,124 --> 03:23:57,425 THERE WAS A LOT OF DISCUSSIONS 4827 03:23:57,425 --> 03:24:00,128 ABOUT WHAT IT MEANS AND WHAT ARE 4828 03:24:00,128 --> 03:24:01,296 THE PRINCIPLES. 4829 03:24:01,296 --> 03:24:04,766 THE WORKSHOP WAS SUCCESSFULLY 4830 03:24:04,766 --> 03:24:07,969 ATTENDED AND WE'RE PLEASED THAT FOLKS 4831 03:24:07,969 --> 03:24:09,571 ATTENDED IN THE NUMBERS THAT 4832 03:24:09,571 --> 03:24:12,240 THEY DID. 4833 03:24:12,240 --> 03:24:16,444 WE BASICALLY TRIED TO UNDERSTAND 4834 03:24:16,444 --> 03:24:18,747 WHAT NEUROMORPHIC ENGINEERING 4835 03:24:18,747 --> 03:24:19,247 MEANS AND HOW TO EXTRACT PRINCIPLES. 4836 03:24:19,247 --> 03:24:21,983 I USE THIS AS A MOTIVATOR FOR 4837 03:24:21,983 --> 03:24:26,721 INTERPLAY BETWEEN FUNCTION AND FORM. 4838 03:24:26,721 --> 03:24:31,559 HOW DO YOU MODEL THE FORM OF THE 4839 03:24:31,559 --> 03:24:32,560 NERVOUS SYSTEM AND/OR THE 4840 03:24:32,560 --> 03:24:35,196 FUNCTION OF THE NERVOUS SYSTEM 4841 03:24:35,196 --> 03:24:38,166 INTO ENGINEERED SYSTEMS AND TRY TO 4842 03:24:38,166 --> 03:24:43,204 BUILD SYSTEMS THAT APPLY TO A 4843 03:24:43,204 --> 03:24:48,977 PARTICULAR PROBLEM? DA VINCI 4844 03:24:48,977 --> 03:24:54,516 WHAT WILL WE SAW WAS THAT FROM A 4845 03:24:54,516 --> 03:24:57,886 NEURO-- BASICALLY LATE 80S 4846 03:24:57,886 --> 03:25:01,022 EARLY 90S THERE WAS A NOTION THAT 4847 03:25:01,022 --> 03:25:05,226 IF YOU LOOK AT CIRCUITS, YOU CAN 4848 03:25:05,226 --> 03:25:09,497 IMPLEMENT COMPUTATIONAL MODELS OF 4849 03:25:09,497 --> 03:25:11,366 NERVOUS SYSTEM THAT TAKE ADVANTAGE OF THE COMPUTATION IN INTEGRATEDCIRCUITS. 4850 03:25:11,366 --> 03:25:15,270 SOME OF THE FIRST NEUROMORPHIC 4851 03:25:15,270 --> 03:25:17,105 ENGINEERING DONE IN TERMS OF 4852 03:25:17,105 --> 03:25:19,974 BUILDING A NEURON DATES BACK TO 4853 03:25:19,974 --> 03:25:23,545 THE 1960S WHERE AN ELECTROCHEMICAL 4854 03:25:23,545 --> 03:25:34,055 CELL BUILT BY PAUL MUELLER WAS 4855 03:25:34,989 --> 03:25:36,124 A BILIPID LAYER, HE APPLIED ELECTRIC 4856 03:25:36,124 --> 03:25:39,561 SIGNALS TO OPEN AND CLOSE ION 4857 03:25:39,561 --> 03:25:40,195 CHANNELS TO GENERATE SPIKES. 4858 03:25:40,195 --> 03:25:42,830 THAT WAS ONE OF THE FIRST EXAMPLE 4859 03:25:42,830 --> 03:25:49,737 OF A NEUROMORPHIC SYSTEM. AND THEN 4860 03:25:49,737 --> 03:25:51,873 HE PUT TOGETHER A LOT OF 4861 03:25:51,873 --> 03:25:53,641 NETWORKS TO DO SPEECH 4862 03:25:53,641 --> 03:25:55,009 RECOGNITION AND SPEECH 4863 03:25:55,009 --> 03:25:55,443 DECOMPOSITION. 4864 03:25:55,443 --> 03:26:00,782 WHEN WE THINK OF THE PRINCIPLES, 4865 03:26:00,782 --> 03:26:03,585 THIS IS A LITTLE CONTROVERSIAL 4866 03:26:03,585 --> 03:26:05,553 BUT IT'S HOW I SEE IT. 4867 03:26:05,553 --> 03:26:10,258 WE LEARN FROM BIOINSPIRATION. 4868 03:26:10,258 --> 03:26:12,727 WE LEARN THE PRINCIPLES SUCH AS 4869 03:26:12,727 --> 03:26:14,896 COMPUTATION, MODEL, ALGORITHMS, 4870 03:26:14,896 --> 03:26:17,131 PROCESSES AND SO ON, LEARNING. 4871 03:26:17,131 --> 03:26:18,566 THE IMPLEMENTATION TO ME IS THE 4872 03:26:18,566 --> 03:26:20,902 NEXT STEP. 4873 03:26:20,902 --> 03:26:21,903 I'M AGNOSTIC. 4874 03:26:21,903 --> 03:26:24,038 I DON'T MIND WHETHER IT'S 4875 03:26:24,038 --> 03:26:25,406 IMPLEMENTED HARDWARE OR SOFTWARE 4876 03:26:25,406 --> 03:26:28,710 OR SOME OTHER MECHANISM, IT IS 4877 03:26:28,710 --> 03:26:31,279 WHAT IT IS. 4878 03:26:31,279 --> 03:26:32,747 ULTIMATELY WE GET TO THE 4879 03:26:32,747 --> 03:26:35,283 APPLICATIONS WHERE WE TALK ABOUT 4880 03:26:35,283 --> 03:26:37,018 EMBODIMENTS. WHAT WE SAW IN 4881 03:26:37,018 --> 03:26:39,721 THE CLINIC AND TO LEARN SCIENCE, 4882 03:26:39,721 --> 03:26:42,457 TO BASICALLY FEED THAT LOOP BACK AND 4883 03:26:42,457 --> 03:26:45,460 LEARN BACK WHAT WAS GOING ON IN 4884 03:26:45,460 --> 03:26:46,527 THE SCIENTIFIC NOTION TO BEGIN 4885 03:26:46,527 --> 03:26:47,295 WITH. THAT FLOW IS WHAT I SEE AS 4886 03:26:47,295 --> 03:26:50,031 ESSENTIALLY UNDERSTANDING THE 4887 03:26:50,031 --> 03:26:52,233 NEUROMORPHIC PRINCIPLES. 4888 03:26:52,233 --> 03:26:53,868 I LIKE THE SHORT AND SWEET 4889 03:26:53,868 --> 03:26:57,972 I LIKE THE SHORT AND SWEET 4890 03:26:57,972 --> 03:26:59,941 MORNING ABOUT BASICALLY COMING 4891 03:26:59,941 --> 03:27:01,476 UP WITH COMPUTERS, COMPUTATION 4892 03:27:01,476 --> 03:27:05,179 MODELS THAT SCALE IN THE SAME 4893 03:27:05,179 --> 03:27:08,283 WAY THE BRAIN DOES WITH POWER 4894 03:27:08,283 --> 03:27:13,788 EFFICIENCY OF THE BRAIN. 4895 03:27:13,788 --> 03:27:19,761 HOW HAS NEUROMORPHIC EVOLVED 4896 03:27:19,761 --> 03:27:21,929 OVER THE LAST 30 YEARS? 4897 03:27:21,929 --> 03:27:28,436 WE WERE EXTREMELY 4898 03:27:28,436 --> 03:27:29,570 HARDWARE-ORIENTED IN THE 4899 03:27:29,570 --> 03:27:30,171 BEGINNING. WE MOVED AWAY 4900 03:27:30,171 --> 03:27:32,507 INTO SENSORY PROCESSING 4901 03:27:32,507 --> 03:27:34,008 IN THE EARLY YEARS 4902 03:27:34,008 --> 03:27:36,044 THEN GOT INTO MATERIAL SCIENCE 4903 03:27:36,044 --> 03:27:38,646 AND THOUGHT ABOUT HOW TO DO 4904 03:27:38,646 --> 03:27:41,249 ACCELERATIONS. 4905 03:27:41,249 --> 03:27:42,483 NOW WE'RE STARTING TO THINK 4906 03:27:42,483 --> 03:27:45,887 ABOUT HOW TO DEAL WITH THE DATA 4907 03:27:45,887 --> 03:27:47,622 SCIENCE AND A.I. ELEMENT OF IT WHICH 4908 03:27:47,622 --> 03:27:53,494 IS STARTING TO PLAY A ROLE IN THESE 4909 03:27:53,494 --> 03:27:57,665 AREAS AND ULTIMATELY WE HAVE TO CARE 4910 03:27:57,665 --> 03:28:01,836 ABOUT APPLICATION, BIOMORPHIC 4911 03:28:01,836 --> 03:28:12,347 HEALTHCARE. I HAVE A STUDENT 4912 03:28:15,550 --> 03:28:19,954 AKWASI AKWABOAH HE'S GOTTEN THE BUG 4913 03:28:19,954 --> 03:28:22,623 OF OLD NEUROMORPH, SUPER INTO THE 4914 03:28:22,623 --> 03:28:23,524 CIRCUIT ASPECTS OF IT. 4915 03:28:23,524 --> 03:28:26,828 HE THINKS ABOUT IT FROM A 4916 03:28:26,828 --> 03:28:28,162 PERSPECTIVE WHERE HE SAYS I'M 4917 03:28:28,162 --> 03:28:29,831 TRYING TO SOLVE A PROBLEM HERE. 4918 03:28:29,831 --> 03:28:31,466 THE PROBLEM HE'S TRYING TO SOLVE 4919 03:28:31,466 --> 03:28:35,570 IS LOOKING AT THE SIMILARITY 4920 03:28:35,570 --> 03:28:38,773 BETWEEN TWO STREAMS OF SPIKES. 4921 03:28:38,773 --> 03:28:43,511 WE CAN APPLY THIS INTO A DEPTH 4922 03:28:43,511 --> 03:28:43,811 EXTRACTION INTO A STEREOOPSIS 4923 03:28:43,811 --> 03:28:48,883 SCENARIO. WE CAN USE IT FOR 4924 03:28:48,883 --> 03:28:49,851 SOUND LOCALIZATION BUT HE WAS 4925 03:28:49,851 --> 03:28:51,953 INTERESTED IN THE QUESTION OF IF 4926 03:28:51,953 --> 03:28:54,155 YOU HAVE ELECTRODES AND I WANT 4927 03:28:54,155 --> 03:28:56,524 TO MAKE SURE THAT I APPLY THE 4928 03:28:56,524 --> 03:28:58,693 RIGHT SPIKE TRAINS SO THAT I GET 4929 03:28:58,693 --> 03:29:03,931 THE RIGHT RESPONSE IN THE 4930 03:29:03,931 --> 03:29:05,600 TISSUE, HOW DO I DO IT? 4931 03:29:05,600 --> 03:29:07,535 I CAN RECORD LOCALLY FROM WHERE I 4932 03:29:07,535 --> 03:29:10,104 STIMULATE AND GET A PROFILE OF THE 4933 03:29:10,104 --> 03:29:15,042 IMPACT OF THE STIMULATION AND ADAPT 4934 03:29:15,042 --> 03:29:17,211 A DYNAMIC MODEL TO CHANGE THE 4935 03:29:17,211 --> 03:29:18,746 WAY THE SPIKES ARE DELIVERED IN 4936 03:29:18,746 --> 03:29:21,649 ORDER TO MAKE SURE THAT WHAT YOU 4937 03:29:21,649 --> 03:29:23,084 EXPECT TO GET IS WHAT YOU 4938 03:29:23,084 --> 03:29:23,651 ACTUALLY GET. 4939 03:29:23,651 --> 03:29:26,487 TO DO THIS, YOU HAVE TO EMBED 4940 03:29:26,487 --> 03:29:27,789 SOME CIRCUITS. 4941 03:29:27,789 --> 03:29:32,059 I THINK -- EMBED SOME CIRCUITS 4942 03:29:32,059 --> 03:29:33,828 IN THE ELECTRODE ITSELF IN ORDER 4943 03:29:33,828 --> 03:29:36,230 TO BE ABLE TO DO THESE TWO KINDS 4944 03:29:36,230 --> 03:29:39,133 OF COMPUTATIONAL COMPONENTS. 4945 03:29:39,133 --> 03:29:43,004 SO HE STAGES IT AS A LOSS 4946 03:29:43,004 --> 03:29:44,639 FUNCTION THAT HAS TO BE 4947 03:29:44,639 --> 03:29:46,908 MINIMIZED, IF YOU WILL. 4948 03:29:46,908 --> 03:29:50,978 AND THEN ALL THE DIFFERENT 4949 03:29:50,978 --> 03:29:52,613 COMPONENTS CAN BE IMPLEMENTED IN 4950 03:29:52,613 --> 03:29:54,749 DIFFERENT PIECES OF THE CIRCUIT. 4951 03:29:54,749 --> 03:29:56,818 WHAT IS ALSO INTERESTING TO ME 4952 03:29:56,818 --> 03:30:01,022 IS THE FACT THAT CERTAIN PIECES 4953 03:30:01,022 --> 03:30:02,957 COME FROM DIFFERENT PART OF THE 4954 03:30:02,957 --> 03:30:04,559 LITERATURE THAT SUPPORTS 4955 03:30:04,559 --> 03:30:08,062 NEUROMORPHICS OVER THE YEARS. 4956 03:30:08,062 --> 03:30:11,332 THIS IS A BIT FROM CELL PHONE CAMERAS 4957 03:30:11,332 --> 03:30:15,303 AND DOING ADDITION, DIVISION AND 4958 03:30:15,303 --> 03:30:17,538 MULTIPLICATION FROM THE EARLY 4959 03:30:17,538 --> 03:30:21,409 WORK ON COCHLEAR IMPLANTS. 4960 03:30:21,409 --> 03:30:24,812 THIS IS A BIT FROM THE ADAPTATION 4961 03:30:24,812 --> 03:30:25,847 CIRCUIT THAT TOBI DELBRUCK USED 4962 03:30:25,847 --> 03:30:31,118 IN ORDER TO GET A RANGE OF 4963 03:30:31,118 --> 03:30:32,353 VISUAL SENSING. 4964 03:30:32,353 --> 03:30:34,856 SO PUT TOGETHER NOW CAN BE 4965 03:30:34,856 --> 03:30:42,230 APPLIED IN AN ELECTRODE SCENARIO 4966 03:30:42,230 --> 03:30:45,066 WHERE YOU ADAPT ELECTRODE AND 4967 03:30:45,066 --> 03:30:45,633 SIMULATION CHARACTERISTICS BASED ON 4968 03:30:45,633 --> 03:30:48,369 CHANGES IN THE ENVIRONMENT 4969 03:30:48,369 --> 03:30:49,704 AROUND THE ELECTRODE. 4970 03:30:49,704 --> 03:30:58,980 LOVE THIS PAPER BY GERT CAUWENBERGHS 4971 03:30:58,980 --> 03:31:01,883 THAT MAPS ANGSTROMS TO METER 4972 03:31:01,883 --> 03:31:05,720 PARALLEL MAGNITUDES OF WHERE 4973 03:31:05,720 --> 03:31:13,995 NEUROMORPHICS PLAYS A ROLE. 4974 03:31:13,995 --> 03:31:17,598 WHAT IS COMPUTATION? WHAT ARE 4975 03:31:17,598 --> 03:31:24,705 ALGORITHMS? WHAT ARE THE INDIVIDUAL 4976 03:31:24,705 --> 03:31:28,209 COMPONENTS OF THE CIRCUIT AND THE ROLE OF SYNTHESIS? 4977 03:31:28,209 --> 03:31:30,811 THIS IS ONE OF WHERE WE LAY OUT 4978 03:31:30,811 --> 03:31:33,781 ALL THE PLACES THAT WE THINK 4979 03:31:33,781 --> 03:31:35,383 NEUROMORPHICS CAN PLAY A ROLE. 4980 03:31:35,383 --> 03:31:38,386 WHAT ARE THE DIFFERENT AREAS AND 4981 03:31:38,386 --> 03:31:41,989 WE DIVIDE IT INTO DIAGNOSIS, BIO- 4982 03:31:41,989 --> 03:31:52,767 SIGNAL ANALYSIS, NEURAL INTERFACES 4983 03:31:54,936 --> 03:31:58,239 AND NEUROMORPHIC TOOLS. RECOMMENDED. 4984 03:31:58,239 --> 03:32:01,809 COUPLE OF READOUT FROM THE WORKSHOP. 4985 03:32:01,809 --> 03:32:04,045 NUMBER ONE, DEFINITIONAL NOTION 4986 03:32:04,045 --> 03:32:06,280 ARGUEMENTS. WE UNDERSTOOD THAT 4987 03:32:06,280 --> 03:32:10,785 HEY, WE NEED NOT BE PARSIMONIOUSLY 4988 03:32:10,785 --> 03:32:16,891 ATTENDING TO DIRECT NEUROBIOLOGICAL 4989 03:32:16,891 --> 03:32:22,396 REPLICATION TO GET THE JOB DONE. 4990 03:32:22,396 --> 03:32:25,833 SOME DEGREE OF FLOW BETWEEN 4991 03:32:25,833 --> 03:32:29,103 PARSIMONY AND FUNCTIONAL. SO FORM 4992 03:32:29,103 --> 03:32:31,539 AND FUNCTION WHERE IT FITS IS DRIVEN 4993 03:32:31,539 --> 03:32:34,275 BY THE PROBLEM AT HAND. 4994 03:32:34,275 --> 03:32:38,312 WE TALKED ABOUT THE FACT THAT 4995 03:32:38,312 --> 03:32:42,083 THE TERM NEUROMORPHIC ITSELF -- 4996 03:32:42,083 --> 03:32:46,721 THE TERM "NEUROMORPHIC" IS 4997 03:32:46,721 --> 03:32:47,221 PROBLEMATIC. 4998 03:32:47,221 --> 03:32:49,924 IT'S ASSUMED TO BE SOMEWHAT 4999 03:32:49,924 --> 03:32:56,564 EXPERIMENTAL EVEN THOUGH WE LIKE 5000 03:32:56,564 --> 03:32:59,100 A.I., NEUROMORPHIC GETS PEOPLE A 5001 03:32:59,100 --> 03:33:01,035 LITTLE BIT WORRIED FOR SOME 5002 03:33:01,035 --> 03:33:01,736 REASON. 5003 03:33:01,736 --> 03:33:03,738 THAT'S SOMETHING WE NEED TO 5004 03:33:03,738 --> 03:33:04,372 OVERCOME. 5005 03:33:04,372 --> 03:33:05,973 IT'S A HURDLE THAT WE NEED TO 5006 03:33:05,973 --> 03:33:09,577 GET OVER. 5007 03:33:09,577 --> 03:33:13,581 ONE AREA THAT BRAD WAS 5008 03:33:13,581 --> 03:33:17,551 POINTING OUT, WE LIKE 5009 03:33:17,551 --> 03:33:19,887 BENCHMARKING BECAUSE IT GIVES 5010 03:33:19,887 --> 03:33:23,891 US A SET APPROACH ON HOW GOOD AN 5011 03:33:23,891 --> 03:33:28,863 ALGORITHM IS BUT IT CAN BE 5012 03:33:28,863 --> 03:33:33,067 STIFLING BECAUSE WE'RE LOCALIZE 5013 03:33:33,067 --> 03:33:36,437 TO ATTEMPT ONE SET OF PROBLEMS 5014 03:33:36,437 --> 03:33:39,640 AND DON'T LOOK BEYOND IT. 5015 03:33:39,640 --> 03:33:43,210 AN ECONOMIST SAID USELESS 5016 03:33:43,210 --> 03:33:45,646 SCIENCE IS IMPORTANT. 5017 03:33:45,646 --> 03:33:48,282 IF WE'RE BOUNDED BY BENCHMARKS 5018 03:33:48,282 --> 03:33:51,886 ONLY, WE TENDING TO LOCALIZED 5019 03:33:51,886 --> 03:33:53,821 AND THAT CAN BE PROBLEMATIC. 5020 03:33:53,821 --> 03:33:55,356 THE ADVANTAGES THAT WE'VE HEARD 5021 03:33:55,356 --> 03:33:58,192 A LOT ABOUT. 5022 03:33:58,192 --> 03:33:59,860 MATERIALS WAS ALSO ANOTHER KEY 5023 03:33:59,860 --> 03:34:01,362 ELEMENT OF WHAT WE HEARD ABOUT. 5024 03:34:01,362 --> 03:34:09,003 THERE WAS A PRESENTATION BY 5025 03:34:09,003 --> 03:34:10,705 ZHENAN BAO FROM STANFORD ON 5026 03:34:10,705 --> 03:34:14,375 HOW TO APPLY FLEXIBLE, SOFT 5027 03:34:14,375 --> 03:34:16,077 MATERIALS TO DO REPLICATION OF 5028 03:34:16,077 --> 03:34:19,113 SKIN AND GET THE APPROPRIATE 5029 03:34:19,113 --> 03:34:27,888 BEHAVIORS YOU'D EXPECT IN DIFFERENT 5030 03:34:27,888 --> 03:34:28,456 PATHWAYS OF RECEPTORS OF FINGERS. 5031 03:34:28,456 --> 03:34:30,324 THAT WAS AMAZING. 5032 03:34:30,324 --> 03:34:36,430 BIOCOMPATIBILITY AND SELF-HEALING 5033 03:34:36,430 --> 03:34:39,567 MATERIALS IS SOMETHING THAT WAS ALSO 5034 03:34:39,567 --> 03:34:40,901 A VERY IMPORTANT COMPONENT 5035 03:34:40,901 --> 03:34:44,004 IN THE END, WHAT FOLKS WERE 5036 03:34:44,004 --> 03:34:47,274 KEYED ON IS THE FACT THAT WE 5037 03:34:47,274 --> 03:34:50,578 DON'T HAVE A GOOD WAY TO DEVELOP 5038 03:34:50,578 --> 03:34:51,245 THESE TECHNOLOGIES. 5039 03:34:51,245 --> 03:34:54,048 HOW TO DEVELOP MATERIALS THAT'S 5040 03:34:54,048 --> 03:34:56,450 SHAREABLE AND REPRODUCIBLE 5041 03:34:56,450 --> 03:34:58,385 HAT OTHER LABS CAN TAKE 5042 03:34:58,385 --> 03:34:59,553 ADVANTAGE OF AND THIS IS 5043 03:34:59,553 --> 03:35:01,856 SOMETHING THAT FUNDING AGENCIES 5044 03:35:01,856 --> 03:35:03,824 CAN HELP PUT TOGETHER. 5045 03:35:03,824 --> 03:35:06,694 CLINICAL TRANSLATION, WE GOT A 5046 03:35:06,694 --> 03:35:09,663 LONG LIST OF STUFF THAT HAS TO 5047 03:35:09,663 --> 03:35:11,599 BE RELEVANT TO THE CLINICAL 5048 03:35:11,599 --> 03:35:12,500 APPLICATION. 5049 03:35:12,500 --> 03:35:15,302 ESSENTIALLY IN THE END, IT'S HOW 5050 03:35:15,302 --> 03:35:21,242 DO WE GET MULTIMODEL AND CONTINUOUS 5051 03:35:21,242 --> 03:35:29,550 MONITORING? CONSIDER DIFFERENT 5052 03:35:29,550 --> 03:35:31,519 BIOMARKERS? SOMETHING KAI REFERRED 5053 03:35:31,519 --> 03:35:32,987 TO IN HIS TALK: CLINICAL WORK FLOWS. 5054 03:35:32,987 --> 03:35:34,855 MAKING SURE WHATEVER YOU DO, YOU HAVE 5055 03:35:34,855 --> 03:35:38,025 TO BE CONCERNED ON WHAT THE CLINICAL 5056 03:35:38,025 --> 03:35:39,493 WORK FLOW IS AND YOUR SYSTEM HAS TO 5057 03:35:39,493 --> 03:35:41,829 FIT INTO THAT AND REGULATORY 5058 03:35:41,829 --> 03:35:50,337 REQUIREMENTS PLAYS A BIG ROLE 5059 03:35:50,337 --> 03:35:50,538 A PLACE THAT IS RELEVANT IS 5060 03:35:50,538 --> 03:35:56,343 INFRASTRUCTURE FOR INNOVATION, 5061 03:35:56,343 --> 03:35:59,713 FABRICATION ACCESS. ENGINEERS OF A 5062 03:35:59,713 --> 03:36:04,718 CERTAIN GENERATION REMEMBER MOSIS. 5063 03:36:04,718 --> 03:36:08,556 WE COULD FABRICATE CHIPS QUICKLY. WE 5064 03:36:08,556 --> 03:36:18,432 NEED SOMETHING SIMILAR. 5065 03:36:18,432 --> 03:36:26,807 LASTLY, WHAT ARE THE PATH FORWARD? 5066 03:36:26,807 --> 03:36:37,351 LEVERAGING ALGORITHMS TO GET SIZE, 5067 03:36:37,351 --> 03:36:38,252 WEIGHT AND POWER IS OBVIOUS WAY TO GO 5068 03:36:38,252 --> 03:36:44,158 THINK PHYSIOMORPHIC. BEYOND NEURO. 5069 03:36:44,158 --> 03:36:48,062 CLINICAL NEEDS SHOULD DRIVE THE 5070 03:36:48,062 --> 03:36:50,931 APPLICATIONS INSTEAD OF 5071 03:36:50,931 --> 03:36:51,732 TECHNOLOGY. 5072 03:36:51,732 --> 03:36:54,869 GETTING AT SOLUTIONS. 5073 03:36:54,869 --> 03:36:58,873 BECAUSE OF THE TIME, I'LL LET 5074 03:36:58,873 --> 03:37:00,741 YOU GET THIS DISTRIBUTED. 5075 03:37:00,741 --> 03:37:02,209 >> THE VIDEO WILL BE AVAILABLE, 5076 03:37:02,209 --> 03:37:04,511 BUT NOT THE SLIDES. 5077 03:37:04,511 --> 03:37:06,146 >> THE VIDEO WILL BE 5078 03:37:06,146 --> 03:37:10,417 DISTRIBUTED. 5079 03:37:10,417 --> 03:37:12,653 MULTIMODAL DATA SETS FROM REAL 5080 03:37:12,653 --> 03:37:14,788 WORLD SITUATIONS WAS ANOTHER KEY. 5081 03:37:14,788 --> 03:37:17,958 IT WAS DEEMED TO BE REALLY KEY FOR 5082 03:37:17,958 --> 03:37:18,826 DEVELOPMENT. 5083 03:37:18,826 --> 03:37:23,163 THE LAST PART THAT IS CRUCIAL. 5084 03:37:23,163 --> 03:37:24,932 THEORETICAL ANALYSIS FRAMEWORK. 5085 03:37:24,932 --> 03:37:27,301 WITHOUT THAT, WE CAN'T 5086 03:37:27,301 --> 03:37:29,837 UNDERSTAND WHAT WE'RE SEEING AND 5087 03:37:29,837 --> 03:37:31,805 WE CAN'T INNOVATE AS WELL. 5088 03:37:31,805 --> 03:37:34,074 I'LL STOP HERE AND TAKE 5089 03:37:34,074 --> 03:37:35,743 QUESTIONS. 5090 03:37:35,743 --> 03:37:38,579 >> THANK YOU RALPH, THIS IS 5091 03:37:38,579 --> 03:37:41,081 IMPORTANT. I WANT TO INVITE OUR 5092 03:37:41,081 --> 03:37:47,154 DISCUSSANTS TO JOIN US. I KNOW 5093 03:37:47,154 --> 03:37:48,455 YOU RUSHED OVER THE REGULATORY SLIDE 5094 03:37:48,455 --> 03:37:52,359 WE HAVE SOMEONE FROM THE FDA IN THE 5095 03:37:52,359 --> 03:37:53,861 AUDIENCE IF ANYONE HAS QUESTIONS 5096 03:37:53,861 --> 03:37:57,965 ABOUT THAT. 5097 03:37:57,965 --> 03:38:03,370 >> RALPH, COULD I ASK A QUICK 5098 03:38:03,370 --> 03:38:03,804 QUESTION? 5099 03:38:03,804 --> 03:38:05,439 I LIKED YOUR SLIDE WHERE YOU 5100 03:38:05,439 --> 03:38:07,341 TALKED ABOUT THE LAST STEP 5101 03:38:07,341 --> 03:38:09,743 BEFORE YOU GO INTO BIOMEDICAL 5102 03:38:09,743 --> 03:38:12,579 APPLICATIONS IS HARDWARE-SOFTWARE 5103 03:38:12,579 --> 03:38:13,948 INTERFACE. WE TALKED ABOUT DENDRITIC 5104 03:38:13,948 --> 03:38:22,156 COMPUTING. ARE THERE ANY NEUROMORPHIC 5105 03:38:22,156 --> 03:38:24,525 ALGORITHMS YOU WOULD LIKE TO SEE 5106 03:38:24,525 --> 03:38:26,827 INSTANTIATED IN HARDWARE SO THE 5107 03:38:26,827 --> 03:38:29,430 FIELD CAN USE THAT AT THE THAT LAST 5108 03:38:29,430 --> 03:38:35,035 LEVEL OF INTEGRATION? 5109 03:38:35,035 --> 03:38:37,338 >> SO IF I UNDERSTAND YOUR 5110 03:38:37,338 --> 03:38:39,540 QUESTION CORRECTLY, IN MY PAST 5111 03:38:39,540 --> 03:38:43,210 LIFE WHERE NOT SO LONG AGO I WAS 5112 03:38:43,210 --> 03:38:45,813 INTERESTED IN SPINAL CORD 5113 03:38:45,813 --> 03:38:46,747 STIMULATION TO RESTORE LOCOMOTION. 5114 03:38:46,747 --> 03:38:50,484 WHAT WE DID THERE IS WE BUILT 5115 03:38:50,484 --> 03:38:58,759 HARDWARE MODELS OF THE SPINAL CORD 5116 03:38:58,759 --> 03:39:01,395 AS AN EFFERENT COPY. AND USED THAT 5117 03:39:01,395 --> 03:39:09,636 TO STIMULATE THE SPINAL CORD TO 5118 03:39:09,636 --> 03:39:11,438 RECOVER LOCOMOTION ACTIVITIES. 5119 03:39:11,438 --> 03:39:14,141 I COULD TAKE THE CENTRAL PATTERN 5120 03:39:14,141 --> 03:39:16,176 GENERATOR MODEL AND USE IT 5121 03:39:16,176 --> 03:39:17,144 AS A MECHANISM TO STIMULATE. 5122 03:39:17,144 --> 03:39:23,517 MORE RECENTLY -- NOW WE CAN IMAGE 5123 03:39:23,517 --> 03:39:27,821 AN ENTIRE BEHAVING ZEBRA FISH DOING 5124 03:39:27,821 --> 03:39:32,059 PREY CAPTURE. YOU HAVE THE ENTIRE 5125 03:39:32,059 --> 03:39:36,096 NEURONAL CONNECTOME ALMOST 5126 03:39:36,096 --> 03:39:36,897 INSTANTANEOUSLY. 5127 03:39:36,897 --> 03:39:38,399 I WOULD LIKE TO SEE THAT 5128 03:39:38,399 --> 03:39:39,900 IMPLEMENTED INTO HARDWARE WHERE 5129 03:39:39,900 --> 03:39:42,069 YOU CAN TURN OFF DIFFERENT PIECES 5130 03:39:42,069 --> 03:39:45,572 AND SEE IF YOU DID THE SAME ABLATION 5131 03:39:45,572 --> 03:39:48,008 EXPERIMENT ON THE ANIMAL WHERE 5132 03:39:48,008 --> 03:39:51,111 THE ANIMAL WOULD CHANGE ITS 5133 03:39:51,111 --> 03:39:57,184 BEHAVIOR AS THE ANIMAL IN THE 5134 03:39:57,184 --> 03:39:57,384 MODEL. 5135 03:39:57,384 --> 03:40:01,588 THAT LINK BETWEEN A REALTIME 5136 03:40:01,588 --> 03:40:03,824 CONTINUOUS MEASUREMENT OF THE 5137 03:40:03,824 --> 03:40:07,094 NERVOUS SYSTEM AND THE TWIN AND 5138 03:40:07,094 --> 03:40:09,997 THE TWIN INTERACTING WITH THE TWO. 5139 03:40:09,997 --> 03:40:15,235 >> I WOULD LIKE TO FOCUS THIS 5140 03:40:15,235 --> 03:40:20,274 DISCUSSION TO REMIND US OF OUR GOALS 5141 03:40:20,274 --> 03:40:23,210 TO DISCUSS NEUROAI PRINCIPLES AND 5142 03:40:23,210 --> 03:40:25,045 TECHNOLOGIES ACROSS AGENCIES TO 5143 03:40:25,045 --> 03:40:29,716 TRANSFORM NEUROSCIENCE AND BRAIN HEALTH. 5144 03:40:29,716 --> 03:40:32,286 UNDERSTAND HOW OTHER AGENCIES COULD 5145 03:40:32,286 --> 03:40:36,690 BENEFIT FROM BRAIN DATA, KNOWLEDGE- 5146 03:40:36,690 --> 03:40:38,926 BASES. WITH THAT REFRAMING, I WOULD 5147 03:40:38,926 --> 03:40:43,430 LIKE TO INVITE JENNIFER TO ASK 5148 03:40:43,430 --> 03:40:44,731 HER PROVOCATIVE OR LINKING 5149 03:40:44,731 --> 03:40:47,267 QUESTION. 5150 03:40:47,267 --> 03:40:51,038 >> I GUESS I GET TO START. 5151 03:40:51,038 --> 03:40:53,474 SO, FIRST THING IS A LOT OF GOOD 5152 03:40:53,474 --> 03:40:53,841 TALKS. 5153 03:40:53,841 --> 03:40:56,343 IT WAS WONDERFUL TO SEE 5154 03:40:56,343 --> 03:40:58,946 EVERYTHING THERE AND THE WHOLE 5155 03:40:58,946 --> 03:41:01,215 CONVERSATIONS AROUND BETWEEN 5156 03:41:01,215 --> 03:41:04,518 BOTH THE EMBODIED COMPUTATION, 5157 03:41:04,518 --> 03:41:06,386 TALKING ABOUT THE MEDICAL 5158 03:41:06,386 --> 03:41:08,489 ASPECTS SIMULTANEOUSLY WITH THE 5159 03:41:08,489 --> 03:41:08,956 ROBOTICS. 5160 03:41:08,956 --> 03:41:10,057 ALL OF THESE THINGS ARE 5161 03:41:10,057 --> 03:41:11,458 INTERESTING. 5162 03:41:11,458 --> 03:41:14,661 ALSO THE DISCUSSIONS THAT RALPH 5163 03:41:14,661 --> 03:41:18,332 IN TERMS OF SHOWING THE 5164 03:41:18,332 --> 03:41:21,001 PREVIOUS OUTBRIEF, DISCUSSIONS 5165 03:41:21,001 --> 03:41:24,671 BETWEEN BIOHEALTH AND NEUROMORPHIC 5166 03:41:24,671 --> 03:41:27,407 ITEMS AND DIRECTIONS. IT WAS COOL. 5167 03:41:27,407 --> 03:41:29,143 AS I THINK ABOUT THIS FURTHER IS 5168 03:41:29,143 --> 03:41:33,147 GIVEN ALL OF THIS, YOU WERE 5169 03:41:33,147 --> 03:41:36,850 TALKING ABOUT THE NEUROMORPHIC 5170 03:41:36,850 --> 03:41:37,718 CONCEPTS. 5171 03:41:37,718 --> 03:41:40,020 HOW DO WE SCALE THIS UP TO BE 5172 03:41:40,020 --> 03:41:43,457 SOMETHING WE COULD USE IN SOME 5173 03:41:43,457 --> 03:41:45,726 OF THESE HEALTH AND MEDICAL 5174 03:41:45,726 --> 03:41:46,193 APPLICATIONS. 5175 03:41:46,193 --> 03:41:48,162 IN PARTICULAR, WHAT WOULD THAT 5176 03:41:48,162 --> 03:41:50,030 MEAN IN TERMS OF SCALING. 5177 03:41:50,030 --> 03:41:53,767 CERTAINLY, COMPUTING IS THERE AND 5178 03:41:53,767 --> 03:41:55,969 THERE ARE DISCUSSIONS THAT COME 5179 03:41:55,969 --> 03:41:58,572 FROM THE PREVIOUS SESSION BUT 5180 03:41:58,572 --> 03:42:00,507 ALSO IN THE SENSORS, 5181 03:42:00,507 --> 03:42:10,651 ACTUATION, AND 5182 03:42:11,852 --> 03:42:16,924 THE WHOLE PICTURE. 5183 03:42:16,924 --> 03:42:18,892 >> SO THE SCALING PART I SUPPOSE 5184 03:42:18,892 --> 03:42:21,828 IS A QUESTION OF WHAT 5185 03:42:21,828 --> 03:42:22,529 SPECIFICALLY YOU ARE REFERRING 5186 03:42:22,529 --> 03:42:22,729 TO. 5187 03:42:22,729 --> 03:42:26,600 IF YOU'RE REFERRING TO WHAT IS 5188 03:42:26,600 --> 03:42:31,071 INSIDE -- IMPLANTED OR ON THE 5189 03:42:31,071 --> 03:42:33,373 BODY, THE WEARABLES. 5190 03:42:33,373 --> 03:42:35,242 I WOULD ARGUE THAT THE SCALE IS 5191 03:42:35,242 --> 03:42:36,310 NOT AS IMPORTANT. 5192 03:42:36,310 --> 03:42:40,113 IT'S MORE OF A LOCAL, SMALLER 5193 03:42:40,113 --> 03:42:46,320 SYSTEM. 5194 03:42:46,320 --> 03:42:46,887 BUT REFERRING BACK TO 5195 03:42:46,887 --> 03:42:48,855 KWABENA'S TALK WHERE HE WAS 5196 03:42:48,855 --> 03:42:51,024 ESSENTIALLY SAYING EVEN IF WE 5197 03:42:51,024 --> 03:42:54,428 WERE TO ASSUME THAT WE CAN GO TO THE 5198 03:42:54,428 --> 03:42:57,364 CLOUD, GET ALL THE COMPUTATION THERE 5199 03:42:57,364 --> 03:42:59,633 THE TYPE OF COMPUTATION WE NEED TO DO 5200 03:42:59,633 --> 03:43:04,238 WOULD BE SO MUCH THAT THE IMPACT ON 5201 03:43:04,238 --> 03:43:06,106 THE ENVIRONMENT WOULD BE MASSIVE. 5202 03:43:06,106 --> 03:43:08,075 THAT'S WHERE THE SCALING IS 5203 03:43:08,075 --> 03:43:10,010 SUPER IMPORTANT TO MAKE THOSE SYSTEMS 5204 03:43:10,010 --> 03:43:13,513 MUCH MORE EFFICIENT AND AVAILABLE. 5205 03:43:13,513 --> 03:43:17,150 SO THAT WE CAN USE IT KIND OF 5206 03:43:17,150 --> 03:43:22,055 OFF-LOADING OF THE MEDICAL, THE 5207 03:43:22,055 --> 03:43:23,423 DETECTION OF CANCER IN THOSE 5208 03:43:23,423 --> 03:43:24,258 IMAGES AND SO FORTH. 5209 03:43:24,258 --> 03:43:27,628 EVEN IF YOU DO STUFF OFF THE 5210 03:43:27,628 --> 03:43:30,664 BODY, YOU OFF-LOAD IT, YOU STILL 5211 03:43:30,664 --> 03:43:32,566 NEED TO WORRY ABOUT SCALING AT 5212 03:43:32,566 --> 03:43:33,200 THAT LEVEL. 5213 03:43:33,200 --> 03:43:42,542 THAT'S WHERE I WOULD SEE IT. 5214 03:43:42,542 --> 03:43:46,747 >> GO AHEAD, GINA. 5215 03:43:46,747 --> 03:43:50,550 >> IN TERMS OF SCALING FROM COMPUTING 5216 03:43:50,550 --> 03:43:52,653 IN CLOSED LOOP PERSPECTIVE, I THINK 5217 03:43:52,653 --> 03:43:56,823 IT CONNECTS ALSO TO THE NECESSITY 5218 03:43:56,823 --> 03:44:01,161 TO HAVE CAPABILITIES TO INTEGRATE 5219 03:44:01,161 --> 03:44:10,270 WITH THE SENSORS, THE SIMULATION, 5220 03:44:10,270 --> 03:44:12,005 THE ACTUATION. THAT IS INFRASTRUCTURE 5221 03:44:12,005 --> 03:44:15,275 NEEDED TO BE ABLE TO DEMONSTRATE 5222 03:44:15,275 --> 03:44:19,046 THAT CLOSED-LOOP BEHAVIOR BETWEEN 5223 03:44:19,046 --> 03:44:22,616 SENSING, COMPUTING AND SIMULATION. 5224 03:44:22,616 --> 03:44:24,251 >> I WOULD LIKE TO JUMP IN ON 5225 03:44:24,251 --> 03:44:25,218 THAT. 5226 03:44:25,218 --> 03:44:33,226 WE'RE TALKING ABOUT SENSORS AND 5227 03:44:33,226 --> 03:44:36,496 ACTUATION. WE JUST CAN'T HAVE COMPUTING 5228 03:44:36,496 --> 03:44:38,231 ON THE CHIP. THE INFORMATION HAS TO 5229 03:44:38,231 --> 03:44:40,067 COME INTO THE CHIP AND COME OUT. 5230 03:44:40,067 --> 03:44:42,469 ONE OF THE BIG THINGS IS HOW DO 5231 03:44:42,469 --> 03:44:44,104 WE TRANSFORM THAT SENSORY 5232 03:44:44,104 --> 03:44:46,506 INFORMATION INTO A SYSTEM THAT 5233 03:44:46,506 --> 03:44:53,113 CAN BE HANDLED NEUROMORPHICALLY. 5234 03:44:53,113 --> 03:44:54,214 A LOT OF COMMERCIAL SENSORS ARE 5235 03:44:54,214 --> 03:44:56,316 NOT DESIGNED TO INTERFACE WITH THE 5236 03:44:56,316 --> 03:44:58,919 NEUROMORPHIC SYSTEM. NOT THAT THEY 5237 03:44:58,919 --> 03:45:00,020 CAN'T BUT THEY JUST AREN'T. 5238 03:45:00,020 --> 03:45:02,222 AN EXAMPLE, MOST MOTOR CONTROL AT 5239 03:45:02,222 --> 03:45:03,657 THE LOWEST LEVEL IS DONE WITH PULSES 5240 03:45:03,657 --> 03:45:04,991 AND TIMING. 5241 03:45:04,991 --> 03:45:06,660 THERE IS A MIDDLE LAYER IN 5242 03:45:06,660 --> 03:45:09,229 BETWEEN THAT TAKES DIGITAL SIGNALS 5243 03:45:09,229 --> 03:45:11,732 AND CONVERTS THEM INTO PULSES. 5244 03:45:11,732 --> 03:45:12,199 NO REASON IT CAN'T BE DONE 5245 03:45:12,199 --> 03:45:14,401 NEUROMORPHICALLY. THAT'S NOT HOW THE 5246 03:45:14,401 --> 03:45:15,068 SYSTEM IS CONSTRUCTED. 5247 03:45:15,068 --> 03:45:17,838 IN TERMS OF INTERFACING WITH ALL 5248 03:45:17,838 --> 03:45:23,009 THESE DIFFERENT ACTUATORS, WE HAVE 5249 03:45:23,009 --> 03:45:26,580 TO RECKON WITH THE MIDDLE LAYER 5250 03:45:26,580 --> 03:45:28,081 AND THE COMMUNICATION ASPECT. 5251 03:45:28,081 --> 03:45:33,253 >> ANOTHER CONSIDERATION FOR 5252 03:45:33,253 --> 03:45:41,728 SCALING, DR. HAYS SAID IF WE DESIGN 5253 03:45:41,728 --> 03:45:46,833 THE MOST SOPHISTICATED THING BUT 5254 03:45:46,833 --> 03:45:49,403 THAT A SAILOR CAN'T IMPLEMENT, 5255 03:45:49,403 --> 03:45:52,105 IT'S NOT GOING TO BE USED. 5256 03:45:52,105 --> 03:45:54,541 PHYSICIAL SURGEONS ARE NOT THE 5257 03:45:54,541 --> 03:45:55,809 END-USER. IT'S MAKING SURE THAT 5258 03:45:55,809 --> 03:45:58,111 THINGS ARE PACKAGED 5259 03:45:58,111 --> 03:46:00,914 APPROPRIATELY. WHEN WE DEVELOP 5260 03:46:00,914 --> 03:46:04,317 DEVICES, THE PACKAGING AND CONNECTORS 5261 03:46:04,317 --> 03:46:14,461 TAKE LONGER TO SOLVE --IT'S THE 5262 03:46:14,461 --> 03:46:21,468 ABILITY TO GET PACKAGING THAT WORKS 5263 03:46:21,468 --> 03:46:23,770 IN THE END USER AND COMMERCIAL SPACE. 5264 03:46:23,770 --> 03:46:25,839 THEN THE OTHER IS MAKING SURE 5265 03:46:25,839 --> 03:46:26,940 THAT PEOPLE THAT ARE GOING TO BE 5266 03:46:26,940 --> 03:46:29,009 USING IT, THAT IT CAN SCALE TO 5267 03:46:29,009 --> 03:46:34,281 THE RELATIVE MOST COMMON 5268 03:46:34,281 --> 03:46:35,849 DENOMINATOR NEEDS THAT DIFFERENT 5269 03:46:35,849 --> 03:46:38,351 PEOPLE HAVE. 5270 03:46:38,351 --> 03:46:39,119 >> GREAT QUESTIONS. 5271 03:46:39,119 --> 03:46:43,190 LET AS MOVE ON TO MYRIAM. 5272 03:46:43,190 --> 03:46:44,157 >> THANK YOU VERY MUCH ALL OF 5273 03:46:44,157 --> 03:46:44,691 YOU. 5274 03:46:44,691 --> 03:46:47,661 I LEARNED A LOT FROM THESE 5275 03:46:47,661 --> 03:46:48,428 INTERESTING TALKS. 5276 03:46:48,428 --> 03:46:51,531 WE LEARNED A LOT IN THIS WHOLE 5277 03:46:51,531 --> 03:46:55,869 TWO WEEKS ON HOW NEUROSCIENCE 5278 03:46:55,869 --> 03:46:57,971 WILL HELP NEUROMORPHIC. 5279 03:46:57,971 --> 03:47:02,909 WE'RE INTERESTED IN NEUROAI. 5280 03:47:02,909 --> 03:47:05,378 ONE THING THAT IS MISSING IS A.I. 5281 03:47:05,378 --> 03:47:11,451 IF YOU DON'T CARE ABOUT NEURO, 5282 03:47:11,451 --> 03:47:14,287 HEARING ABOUT THE PROBLEMS WE 5283 03:47:14,287 --> 03:47:16,756 CAN SOLVE, IF I'M SITTING IN 5284 03:47:16,756 --> 03:47:18,758 CONFERENCES THAT ARE COMPLETELY 5285 03:47:18,758 --> 03:47:22,095 HARDWARE OR A.I. ORIENTED, 5286 03:47:22,095 --> 03:47:25,832 THEY ARE TACKLING SIMILAR PROBLEMS 5287 03:47:25,832 --> 03:47:29,302 COMPLETELY WITHOUT 5288 03:47:29,302 --> 03:47:30,170 NEUROSCIENCE INSPIRATION. 5289 03:47:30,170 --> 03:47:36,443 HOW WOULD I COMPARE A NEUROINSPIRED 5290 03:47:36,443 --> 03:47:39,145 SOLUTION, FOR EXAMPLE AN EFFICIENCY 5291 03:47:39,145 --> 03:47:40,647 PROBLEM, WITH SOMEBODY WHO TACKLES 5292 03:47:40,647 --> 03:47:42,282 THE SAME PROBLEM WITHOUT NEURO 5293 03:47:42,282 --> 03:47:44,784 INSPIRATION. THAT'S ONE QUESTION. 5294 03:47:44,784 --> 03:47:47,921 I'M AMONG THE TEACHERS THAT HAVE 5295 03:47:47,921 --> 03:47:48,989 ONE QUESTION WITH TEN PARTS. 5296 03:47:48,989 --> 03:47:51,057 ALONG THIS LINE, I'M THINKING 5297 03:47:51,057 --> 03:47:57,697 OF, FOR EXAMPLE, GENERALIZABILITY 5298 03:47:57,697 --> 03:47:58,999 IS ISSUE IN ROBOTICS RIGHT NOW. 5299 03:47:58,999 --> 03:48:01,268 WHY DO I NEED TO HAVE A ROBOT 5300 03:48:01,268 --> 03:48:05,539 THAT ALWAYS RUNS ON A NEUROMORPHIC 5301 03:48:05,539 --> 03:48:05,739 CHIP BECAUSE OF GENERALIZABILITY? 5302 03:48:05,739 --> 03:48:07,974 WHY DON'T I THINK ABOUT HAVING 5303 03:48:07,974 --> 03:48:13,313 A.I. TAKING CARE OF ALL THE 5304 03:48:13,313 --> 03:48:13,880 PROBLEMS. 5305 03:48:13,880 --> 03:48:17,317 THE MOMENT THAT I NEED THAT 5306 03:48:17,317 --> 03:48:18,285 GENERALIZABILITY PIECE, I TURN ON 5307 03:48:18,285 --> 03:48:20,854 MY NEUROMORPHIC CHIP THAT HELPS 5308 03:48:20,854 --> 03:48:22,188 WITH THE ENVIRONMENTAL CHANGE 5309 03:48:22,188 --> 03:48:24,424 AND IT GOES BACK TO A.I. 5310 03:48:24,424 --> 03:48:28,728 MY QUESTION IS HOW CAN I HAVE 5311 03:48:28,728 --> 03:48:39,172 NEUROAI PLUS A.I.? 5312 03:48:41,107 --> 03:48:43,410 >> I UNDERSTOOD YOUR QUESTION. 5313 03:48:43,410 --> 03:48:46,613 AT LEAST IN MY SITUATION, I 5314 03:48:46,613 --> 03:48:49,716 CAN'T RUN A TRANSFORMER ON MY 5315 03:48:49,716 --> 03:48:50,750 HARDWARE. 5316 03:48:50,750 --> 03:48:54,154 IT TAKES WAY TOO MUCH POWER TO 5317 03:48:54,154 --> 03:48:56,556 PERFORM A SINGLE INFERENCE WITH 5318 03:48:56,556 --> 03:48:57,324 A TRANSFORMER. THEY'RE HUGE. 5319 03:48:57,324 --> 03:48:59,593 AND IT SUCKS DOWN LOTS OF POWER 5320 03:48:59,593 --> 03:49:02,495 JUST TO PERFORM INFERENCE. 5321 03:49:02,495 --> 03:49:03,930 I DON'T HAVE THAT POWER. 5322 03:49:03,930 --> 03:49:07,167 >> YOUR SITUATION IS SPECIAL. 5323 03:49:07,167 --> 03:49:09,202 >> I AGREE. 5324 03:49:09,202 --> 03:49:12,172 WE HAVE ROBOTIC SITUATION THAT 5325 03:49:12,172 --> 03:49:14,507 ARE NOT THAT LIMITED TO POWER 5326 03:49:14,507 --> 03:49:17,177 AND THEY CAN AFFORD IT. 5327 03:49:17,177 --> 03:49:18,078 >> FAIR ENOUGH. 5328 03:49:18,078 --> 03:49:20,313 IT'S STILL A LOT OF POWER. 5329 03:49:20,313 --> 03:49:26,186 I MEAN, THESE THINGS ARE SUCKING 5330 03:49:26,186 --> 03:49:28,655 DOWN 400 WATTS. 5331 03:49:28,655 --> 03:49:30,223 THAT'S A LOT OF POWER. 5332 03:49:30,223 --> 03:49:32,459 SOME ARE GETTING DOWN LOWER, BUT 5333 03:49:32,459 --> 03:49:35,428 I DON'T KNOW. 5334 03:49:35,428 --> 03:49:39,065 YOU KNOW, I KNOW MY SITUATION IS 5335 03:49:39,065 --> 03:49:40,700 DIFFERENT, BUT UNTIL THEY CAN 5336 03:49:40,700 --> 03:49:42,402 START GETTING SMALLER 5337 03:49:42,402 --> 03:49:44,004 TRANSFORMERS THAT ARE NOT SO 5338 03:49:44,004 --> 03:49:51,444 LARGE THAT REQUIRE SUCH LARGE 5339 03:49:51,444 --> 03:49:51,945 GPUS, I THINK IT IS A PROBLEM. 5340 03:49:51,945 --> 03:49:54,147 A LOT OF PEOPLE JUST SAY CONNECT 5341 03:49:54,147 --> 03:49:55,715 TO THE INTERNET. AND IT'S LIKE 5342 03:49:55,715 --> 03:49:57,917 WELL A LOT OF APPLICATIONS CAN'T 5343 03:49:57,917 --> 03:50:02,856 CONNECT TO THE CLOUD. 5344 03:50:02,856 --> 03:50:06,026 >> I COULD JUMP ON THAT TOO. 5345 03:50:06,026 --> 03:50:07,060 FUNDAMENTALLY I DON'T SEE A 5346 03:50:07,060 --> 03:50:09,429 REASON WHY WE CAN'T HAVE THE 5347 03:50:09,429 --> 03:50:12,399 HYBRID SYSTEMS OF NEUROAI AND 5348 03:50:12,399 --> 03:50:12,799 A.I. 5349 03:50:12,799 --> 03:50:16,369 RIGHT NOW WE'RE IN THE 5350 03:50:16,369 --> 03:50:18,505 EXPERIMENTATION PHASE. 5351 03:50:18,505 --> 03:50:20,407 IN MOST ENGINEERING CONTEXT WE 5352 03:50:20,407 --> 03:50:21,941 TEND IT USE THE TOOL BEST FOR 5353 03:50:21,941 --> 03:50:22,509 THE JOB. 5354 03:50:22,509 --> 03:50:25,278 IF ALL THE ROBOT HAS TO DO IS 5355 03:50:25,278 --> 03:50:27,380 CLASSIFY THERE IS A GRAPE OR 5356 03:50:27,380 --> 03:50:30,283 ORANGE IN FRONT OF ME, THERE IS 5357 03:50:30,283 --> 03:50:32,752 NO REASON FOR BUILDING A NEUROMORPHIC 5358 03:50:32,752 --> 03:50:34,888 SYSTEM FOR IMAGE CLASSIFICATION. 5359 03:50:34,888 --> 03:50:37,357 THERE ARE SOME ASPECTS THAT ARE 5360 03:50:37,357 --> 03:50:40,126 BETTER HANDLED IN NEUROAI AND 5361 03:50:40,126 --> 03:50:41,828 OTHERS IN NORMAL A.I. 5362 03:50:41,828 --> 03:50:45,365 I THINK SOME OF THESE EXAMPLES WOULD 5363 03:50:45,365 --> 03:50:49,869 BE MULTIMODAL, INTEGRATING OVER 5364 03:50:49,869 --> 03:50:51,771 MULTIPLE SENSORY SYSTEMS IN LONG TIME 5365 03:50:51,771 --> 03:50:53,740 HORIZON. THERE IS NO REASON WHY 5366 03:50:53,740 --> 03:51:00,280 WE CAN'T HAVE THE NEUROMORPHIC 5367 03:51:00,280 --> 03:51:04,150 COPROCESSOR THAT HANDLES THE 5368 03:51:04,150 --> 03:51:05,218 NEUROMORPHIC APPLICATIONS AND THE 5369 03:51:05,218 --> 03:51:07,721 A.I. PART. SYNERGY IS GOING TO ENABLE 5370 03:51:07,721 --> 03:51:08,855 POWERFUL SYSTEMS. 5371 03:51:08,855 --> 03:51:10,423 >> THAT'S A GREAT POINT. 5372 03:51:10,423 --> 03:51:13,226 I DO WANT TO POINT OUT A LOT OF 5373 03:51:13,226 --> 03:51:15,528 THE PRESENTATIONS DISCUSSED ABOUT 5374 03:51:15,528 --> 03:51:19,666 HAVING UNIFIED DATA EMBEDDINGS 5375 03:51:19,666 --> 03:51:23,636 BETWEEN SENSING, COMPUTING AND 5376 03:51:23,636 --> 03:51:34,180 ACTUATION. IT IS DIFFICULT TO BRING 5377 03:51:35,582 --> 03:51:36,449 THESE DIFFERENT CHIPS IN THE DIGITAL 5378 03:51:36,449 --> 03:51:39,519 DOMAIN. IT MIGHT BE A TRADEOFF THAT 5379 03:51:39,519 --> 03:51:46,392 WILL DEPEND ON THE APPLICATION. 5380 03:51:46,392 --> 03:51:50,130 >> I AGREE THAT SOMETHING 5381 03:51:50,130 --> 03:51:53,967 DOESN'T MAKE SENSE TO BE DONE THROUGH 5382 03:51:53,967 --> 03:51:57,604 NEUROMORPHIC, LIKE IMAGE CLASSIFICATION. 5383 03:51:57,604 --> 03:52:01,307 ON THE OTHER SIDE, AT NEURIPS, I 5384 03:52:01,307 --> 03:52:04,077 HEARD AN AMAZING PRESENTATION BY 5385 03:52:04,077 --> 03:52:04,577 I THINK LINDA SMITH. 5386 03:52:04,577 --> 03:52:12,218 SHE WAS LOOKING FROM NEWBORNS 5387 03:52:12,218 --> 03:52:16,422 TO INFANTS AT HOW THEY DEVELOP. 5388 03:52:16,422 --> 03:52:20,593 SHE STARTED SAYING NEWBORNS ARE 5389 03:52:20,593 --> 03:52:27,700 ATTRACTED BY STRONG EDGES, 5390 03:52:27,700 --> 03:52:30,103 FURNITURE, AND THEN THEY BECOME 5391 03:52:30,103 --> 03:52:33,173 ATTRACTED TO COLORFUL AND BRIGHT 5392 03:52:33,173 --> 03:52:35,508 THINGS AND THEN MORE DETAILS AND HOW 5393 03:52:35,508 --> 03:52:37,443 THEY INTERACT WITH OBJECTS WHEN 5394 03:52:37,443 --> 03:52:41,948 THEY PLAY. IT'S NOT A RANDOM 5395 03:52:41,948 --> 03:52:45,084 PASSIVE PRESENTATION OF STIMULI, 5396 03:52:45,084 --> 03:52:47,453 BUT IT HAS A STRUCTURE. 5397 03:52:47,453 --> 03:52:50,456 THEY REPEAT AND LOOK AT THE SAME 5398 03:52:50,456 --> 03:52:51,891 THING MULTIPLE TIMES, THEN SWITCH. 5399 03:52:51,891 --> 03:52:54,794 WHY CAN'T WE LEARN FROM THAT TO 5400 03:52:54,794 --> 03:52:56,196 DEVELOP SYSTEMS THAT ARE MORE 5401 03:52:56,196 --> 03:52:57,630 EFFICIENT THAT OPTIMIZE THE FACT 5402 03:52:57,630 --> 03:53:02,202 THAT WE NEED A LOT OF DATA. 5403 03:53:02,202 --> 03:53:03,670 THAT DON'T GENERALIZE. THEY'RE 5404 03:53:03,670 --> 03:53:04,404 BRITTLE AND NOT EXPLAINABLE. 5405 03:53:04,404 --> 03:53:07,907 MAYBE THAT'S NOT NEUROMORPHIC, 5406 03:53:07,907 --> 03:53:10,677 BUT TO ME IT'S AT APPROACH THAT 5407 03:53:10,677 --> 03:53:13,346 LOOKS INTO HOW CAN THE SYSTEM BE 5408 03:53:13,346 --> 03:53:16,716 BUILT IN AWAY THAT IS GENERALLY 5409 03:53:16,716 --> 03:53:21,054 TRYING NOT TO BURN DOWN THE 5410 03:53:21,054 --> 03:53:21,287 PLANET. 5411 03:53:21,287 --> 03:53:23,423 CLIMATE CHANGE, THEY IGNORE 5412 03:53:23,423 --> 03:53:24,924 THAT. 5413 03:53:24,924 --> 03:53:28,061 LET'S BUILD FIVE NUCLEAR PLANTS FOR 5414 03:53:28,061 --> 03:53:30,930 A.I. I THINK IT'S USEFUL, BUT MOST OF 5415 03:53:30,930 --> 03:53:34,934 THE TIMES IT'S MISUSED. 5416 03:53:34,934 --> 03:53:37,103 >> GIACOMO WANTS TO JUMP IN. 5417 03:53:37,103 --> 03:53:40,240 >> I HOPE EVERYONE CAN HEAR ME. 5418 03:53:40,240 --> 03:53:44,110 >> I'M GOING TO BE A LITTLE BIT 5419 03:53:44,110 --> 03:53:47,447 CONTROVERSIAL HERE AGAIN. 5420 03:53:47,447 --> 03:53:56,489 I THINK ALL A.I. IS 5421 03:53:56,489 --> 03:53:57,323 IS NEUROAI. 5422 03:53:57,323 --> 03:54:00,560 EVEN IN THE TRANSFORMER CASE, 5423 03:54:00,560 --> 03:54:02,262 THAT ATTENTION MECHANISM FOR PICKING 5424 03:54:02,262 --> 03:54:09,836 CAN BE CONSTRUCTED IN A NEUROMORPHIC 5425 03:54:09,836 --> 03:54:11,638 OR BIO-INSPIRED WAY. 5426 03:54:11,638 --> 03:54:14,340 I THINK IT'S NOT A DICHOTOMY 5427 03:54:14,340 --> 03:54:18,311 THAT I WOULD DRAW IS ONE THAT 5428 03:54:18,311 --> 03:54:21,514 WOULD SAY HEY, APPLY THE RIGHT 5429 03:54:21,514 --> 03:54:22,548 APPROACH THAT SOLVES THE PROBLEM. 5430 03:54:22,548 --> 03:54:25,518 THAT'S THE HYBRID SYSTEM. 5431 03:54:25,518 --> 03:54:27,687 >> GIACOMO, IF YOU CAN UNMUTE. 5432 03:54:27,687 --> 03:54:28,288 >> YES. 5433 03:54:28,288 --> 03:54:31,491 I AM. 5434 03:54:31,491 --> 03:54:33,459 I HOPE EVERYONE CAN HEAR ME. 5435 03:54:33,459 --> 03:54:35,695 I'M SORRY I COULDN'T BE THERE IN 5436 03:54:35,695 --> 03:54:37,297 PERSON. 5437 03:54:37,297 --> 03:54:40,400 AT LEAST I COULD SEE AND I AGREE 5438 03:54:40,400 --> 03:54:42,468 COMPLETELY WITH EVERYTHING SAID. 5439 03:54:42,468 --> 03:54:46,539 I WANT TO ADD AS IT HAS BEEN 5440 03:54:46,539 --> 03:54:48,775 MENTIONED, NEUROMORPHIC DOES 5441 03:54:48,775 --> 03:54:53,146 NOT HAVE TO REPLACE A.I. BUT IT 5442 03:54:53,146 --> 03:54:54,447 COULD COMPLEMENT A.I. 5443 03:54:54,447 --> 03:54:59,252 YOU HAVE SOMETHING OF THE ORDER 5444 03:54:59,252 --> 03:55:02,588 OF IT ALWAYS ON ACTING AS A 5445 03:55:02,588 --> 03:55:03,222 WATCHDOG. 5446 03:55:03,222 --> 03:55:05,358 IF ANYTHING REQUIRES MORE 5447 03:55:05,358 --> 03:55:07,627 COMPUTATIONAL POWER, THAT'S WHEN 5448 03:55:07,627 --> 03:55:10,930 EITHER MICRO PROCESSOR OR THE 5449 03:55:10,930 --> 03:55:12,598 WHOLE PROCESSOR OR GPU COULD BE 5450 03:55:12,598 --> 03:55:13,099 TURNED ON. 5451 03:55:13,099 --> 03:55:15,468 IT'S POSSIBLE TO USE BOTH. 5452 03:55:15,468 --> 03:55:18,871 IT MAKES SENSE IN AUTOMOTIVE 5453 03:55:18,871 --> 03:55:27,347 DRIVING AND OTHER INDUSTRIAL 5454 03:55:27,347 --> 03:55:28,514 APPLICATIONS TO SEE HOW TO USE BOTH. 5455 03:55:28,514 --> 03:55:31,684 TO WE FIND SOLUTIONS USING THE 5456 03:55:31,684 --> 03:55:33,353 NEUROMORPHIC APPROACH, WE HAVE TO 5457 03:55:33,353 --> 03:55:36,990 FIND THE SAME ROUTE THAT NATURE AND 5458 03:55:36,990 --> 03:55:38,758 EVOLUTION BECAUSE WE HAVE SAME 5459 03:55:38,758 --> 03:55:42,362 CONSTRAINTS. THEN WE MIGHT BE LED TO 5460 03:55:42,362 --> 03:55:45,665 SOLUTIONS THAT ARE MORE EFFICIENT 5461 03:55:45,665 --> 03:55:47,567 MORE VOLUME THAT OCCUPY LESS AREA 5462 03:55:47,567 --> 03:55:48,601 AND VOLUME. 5463 03:55:48,601 --> 03:55:51,337 I THINK THIS ALLOWS US TO FIND 5464 03:55:51,337 --> 03:55:53,606 OTHER SOLUTIONS, BUT THEN WE 5465 03:55:53,606 --> 03:55:56,709 SHOULDN'T THINK OF REPLACING A.I., 5466 03:55:56,709 --> 03:55:59,445 WE SHOULD USE IT WHENEVER THAT 5467 03:55:59,445 --> 03:56:00,947 IS BEST THING TO DO. 5468 03:56:00,947 --> 03:56:03,449 WHENEVER THERE IS DIGITAL DATA, 5469 03:56:03,449 --> 03:56:06,452 IT DOESN'T MAKE SENSE TO USE 5470 03:56:06,452 --> 03:56:08,588 NEUROMORPHIC. 5471 03:56:08,588 --> 03:56:11,090 I MEAN, TRANSFERRING THE DATA IS 5472 03:56:11,090 --> 03:56:13,893 GOING TO BURN ALL THE POWER THAT 5473 03:56:13,893 --> 03:56:15,561 YOU WOULD SAVE. 5474 03:56:15,561 --> 03:56:18,965 LIKE RALPH LIKES TO SAY BURNING 5475 03:56:18,965 --> 03:56:26,005 1 AMP PLUS .1 MILLIAMP OR 1 AMP 5476 03:56:26,005 --> 03:56:27,707 PLUS .1 MICROAMP MAKES NO DIFFERENCE. 5477 03:56:27,707 --> 03:56:30,410 WE NEED TO GET THE HIGHEST 5478 03:56:30,410 --> 03:56:32,378 ACCURACY WITHOUT SACRIFICING 5479 03:56:32,378 --> 03:56:33,079 POWER. 5480 03:56:33,079 --> 03:56:35,381 >> WE HAVE THREE MORE 5481 03:56:35,381 --> 03:56:40,653 DISCUSSIONS IN SEVEN MINUTES. 5482 03:56:40,653 --> 03:56:49,495 OUR NEXT DISCUSSION -- 5483 03:56:49,495 --> 03:56:51,264 >> HELLO, ARE YOU ABLE TO HEAR 5484 03:56:51,264 --> 03:56:51,597 ME? 5485 03:56:51,597 --> 03:56:53,499 >> WE CAN HEAR YOU. 5486 03:56:53,499 --> 03:56:56,135 >> SO, I JUST WANTED TO ASK A 5487 03:56:56,135 --> 03:56:59,472 BRIEF QUESTION SINCE WE'RE SHORT 5488 03:56:59,472 --> 03:57:06,813 ON TIME. I KEEP THINKING 5489 03:57:06,813 --> 03:57:09,615 ABOUT JET AIRPLANES. 5490 03:57:09,615 --> 03:57:13,052 THERE ARE THINGS THAT ENGINEERS 5491 03:57:13,052 --> 03:57:16,789 DO BETTER THAN WE DO. 5492 03:57:16,789 --> 03:57:19,325 WHAT THE MOST LOW HANGING FRUIT 5493 03:57:19,325 --> 03:57:20,660 OPPORTUNITIES FOR DIRECT 5494 03:57:20,660 --> 03:57:24,163 INTERACTIONS BETWEEN ANIMALS AND 5495 03:57:24,163 --> 03:57:26,199 HUMANS AND MORE ENGINEERED 5496 03:57:26,199 --> 03:57:29,202 SYSTEMS AND DO WE NEED MORE 5497 03:57:29,202 --> 03:57:34,340 NEUROMORPHIC TO MAKE THAT INTERFACE? 5498 03:57:34,340 --> 03:57:39,312 >> YOU CAN SCAN THE QR CODE OR 5499 03:57:39,312 --> 03:57:44,417 TYPE IT INTO THE ZOOM CHAT AND 5500 03:57:44,417 --> 03:57:46,185 WE'LL HAVE COME BACK TO YOUR QUESTION 5501 03:57:46,185 --> 03:57:46,853 WHEN WE HAVE IT IN WRITING. 5502 03:57:46,853 --> 03:57:48,221 CHRIS, YOU'RE NEXT. 5503 03:57:48,221 --> 03:57:50,456 >> I'LL TRY TO BRING LITTLE 5504 03:57:50,456 --> 03:57:50,957 SKEPTICISM 5505 03:57:50,957 --> 03:57:55,061 I ENJOYED ALL THE TALKS AND 5506 03:57:55,061 --> 03:57:57,497 EVERYONE IS FREE TO JOIN IN BUT 5507 03:57:57,497 --> 03:57:59,198 I WANTED TO PICK A FIGHT WITH 5508 03:57:59,198 --> 03:58:00,433 DR. MILLER. 5509 03:58:00,433 --> 03:58:02,702 I'M GLAD YOU BROUGHT UP THE 5510 03:58:02,702 --> 03:58:03,936 TENSION BETWEEN ECONOMIC 5511 03:58:03,936 --> 03:58:04,804 INTERESTS. 5512 03:58:04,804 --> 03:58:06,706 SCIENTIFIC INTERESTS, YOU SAID 5513 03:58:06,706 --> 03:58:08,274 PUBLIC INTERESTS AND I'LL THROW 5514 03:58:08,274 --> 03:58:11,110 THE INTEREST OF PEOPLE LIVING 5515 03:58:11,110 --> 03:58:17,550 WITH NEW NEUROLOGIC DISORDERS AS 5516 03:58:17,550 --> 03:58:17,884 WELL. 5517 03:58:17,884 --> 03:58:22,755 WE HAVE OVERLAP IN THE 5518 03:58:22,755 --> 03:58:23,456 NEUROTECHNOLOGY WORLD WHERE 5519 03:58:23,456 --> 03:58:25,992 THERE IS A CULTURE OF NEEDING TO BE 5520 03:58:25,992 --> 03:58:27,827 TECHNOLOGY-DRIVEN TO FEEL LIKE 5521 03:58:27,827 --> 03:58:29,862 THEY'RE EXCITE BUT IT PULLS AT 5522 03:58:29,862 --> 03:58:32,031 THE COMPETING INTEREST OF MAKING 5523 03:58:32,031 --> 03:58:33,933 SOMETHING SIMPLE AND SCALABLE 5524 03:58:33,933 --> 03:58:35,868 AND AFFECT IMPACT ON PATIENTS. 5525 03:58:35,868 --> 03:58:38,871 I THINK WE CAN AGREE WITH 5526 03:58:38,871 --> 03:58:40,173 LOWER-POWERED THINGS ARE BETTER 5527 03:58:40,173 --> 03:58:42,708 AND A.I. CAN HELP US WITH CERTAIN 5528 03:58:42,708 --> 03:58:44,677 TASKS BUT THERE ARE SOME THINGS 5529 03:58:44,677 --> 03:58:48,147 YOU SAID, THERE IS AN IMPLICIT 5530 03:58:48,147 --> 03:58:49,882 ASSUMPTION IF WE MAKE TECHNOLOGY 5531 03:58:49,882 --> 03:58:57,290 THAT IS MORE BRAIN-LIKE, NEUROMORPHIC 5532 03:58:57,290 --> 03:59:00,760 IT'S ADAPTIVE, IT'S CLOSED-LOOP AND 5533 03:59:00,760 --> 03:59:01,727 WILL BE BETTER FOR IMPLANTABLE NEUROTECH. 5534 03:59:01,727 --> 03:59:03,262 I'M TRYING TO THINK OF EVIDENCE 5535 03:59:03,262 --> 03:59:04,564 OF THAT. IT'S HARD TO THINK OF SOME. 5536 03:59:04,564 --> 03:59:10,169 WE NOW HAVE ADAPTIVE DBS FOR 5537 03:59:10,169 --> 03:59:10,603 PARKINSONS. 5538 03:59:10,603 --> 03:59:13,839 THERE ARE GAIS BUT THEY'RE 5539 03:59:13,839 --> 03:59:14,540 MODEST COMPARED TO REGULAR DBS. 5540 03:59:14,540 --> 03:59:18,711 IF WE WANT TO INDUCE PLASTICITY, 5541 03:59:18,711 --> 03:59:21,614 THERE MAY NOT BE A 5542 03:59:21,614 --> 03:59:22,915 NEUROMORPHIC SIGNAL FOR THAT. 5543 03:59:22,915 --> 03:59:26,085 THERE ARE EXAMPLES OF TRYING TO 5544 03:59:26,085 --> 03:59:32,058 STOP SOMETHING LIKE 5545 03:59:32,058 --> 03:59:32,825 DEFIBRILLATION IS NOT CARDIOMORPHIC. 5546 03:59:32,825 --> 03:59:35,795 IF YOU'RE NOT TALKING ABOUT 5547 03:59:35,795 --> 03:59:40,032 PROSTHETICS, MANY OF THE NEURO- 5548 03:59:40,032 --> 03:59:41,901 TECHNIQUES IS TO STOP SOMETHING 5549 03:59:41,901 --> 03:59:45,605 NEURODEGENERATION, EPILEPSY. 5550 03:59:45,605 --> 03:59:48,641 I WANT TO BE A BELIEVER, BUT I 5551 03:59:48,641 --> 03:59:50,676 ACTUALLY DON'T KNOW WHERE THIS 5552 03:59:50,676 --> 03:59:51,944 ASSUMPTION COMES FROM. 5553 03:59:51,944 --> 03:59:53,279 WE PROBABLY CAN'T ANSWER THIS 5554 03:59:53,279 --> 03:59:55,081 RIGHT HERE, BUT WHAT IT BRINGS 5555 03:59:55,081 --> 03:59:59,051 UP TO ME IS THERE IS A NEED FOR 5556 03:59:59,051 --> 04:00:01,354 A SIGNIFICANT SCIENTIFIC PROGRAM 5557 04:00:01,354 --> 04:00:05,191 LIKE WHAT DO WE DO WITH ALL THIS 5558 04:00:05,191 --> 04:00:06,959 TECHNOLOGY ONCE WE HAVE IT. SOLUTIONS 5559 04:00:06,959 --> 04:00:10,997 ARE NOT TECHNOLOGY LED, BUT LED BY NEEDS. 5560 04:00:10,997 --> 04:00:13,432 I WANTED TO ASK, WHAT DO YOU 5561 04:00:13,432 --> 04:00:15,901 THINK ABOUT THE NEED FOR A MAJOR 5562 04:00:15,901 --> 04:00:17,103 SCIENTIFIC PROGRAM TO GO 5563 04:00:17,103 --> 04:00:18,471 ALONGSIDE THIS TO FIGURE OUT 5564 04:00:18,471 --> 04:00:20,906 WHAT TO DO WITH ALL THE 5565 04:00:20,906 --> 04:00:22,608 TECHNOLOGY AND WHAT WOULD SUCH A 5566 04:00:22,608 --> 04:00:25,011 PROGRAM HAVE TO LOOK LIKE? 5567 04:00:25,011 --> 04:00:27,113 WHAT WOULD BE ITS FEATURES? 5568 04:00:27,113 --> 04:00:31,717 >> WHAT I CAN SAY IS I GUESS THE 5569 04:00:31,717 --> 04:00:36,255 CLOSEST THING WE HAVE TO A TRUE 5570 04:00:36,255 --> 04:00:42,261 CLOSED LOOP SYSTEM IS RESPONSIVE 5571 04:00:42,261 --> 04:00:44,797 NEUROSTIMULATION FOR EPILEPSY. 5572 04:00:44,797 --> 04:00:46,866 IT'S AN IDEA THAT YOU 5573 04:00:46,866 --> 04:00:49,268 CONTINUOUSLY RECORD A STREAM OF 5574 04:00:49,268 --> 04:00:51,103 INFORMATION FROM THE AREA IN THE 5575 04:00:51,103 --> 04:00:52,505 BRAIN WHERE YOU THINK THE SEIZURE 5576 04:00:52,505 --> 04:00:53,439 IS STARTING FROM. THE RUDIMENTARY 5577 04:00:53,439 --> 04:00:56,242 FILTERING PROCESS ON DEVICE AND IT 5578 04:00:56,242 --> 04:00:59,512 TRIGGERS A SINGLE OR MULTIPLE 5579 04:00:59,512 --> 04:01:01,480 PULSES OF ELECTRICAL STIMULATION 5580 04:01:01,480 --> 04:01:02,715 IN RESPONSE TO A DETECTED EVENT. 5581 04:01:02,715 --> 04:01:06,852 WE DON'T KNOW IF THE CLOSED LOOP 5582 04:01:06,852 --> 04:01:07,920 ASPECT HELPS AT ALL. 5583 04:01:07,920 --> 04:01:11,657 BUT WE KNOW, FOR EXAMPLE, THAT 5584 04:01:11,657 --> 04:01:13,025 BECAUSE WE WERE RECORDING AND 5585 04:01:13,025 --> 04:01:15,461 TRACKING OVER TIME, WE NOW 5586 04:01:15,461 --> 04:01:17,430 UNDERSTAND THAT THERE ARE STRONG 5587 04:01:17,430 --> 04:01:20,733 DIURNAL PATTERNS TO SEIZURES AND THE 5588 04:01:20,733 --> 04:01:23,502 LIKELIHOOD THAT PATIENTS WILL 5589 04:01:23,502 --> 04:01:25,805 HAVE A SEIZURE. 5590 04:01:25,805 --> 04:01:26,906 AND THAT WAS SOMETHING THAT WE 5591 04:01:26,906 --> 04:01:28,808 DIDN'T KNOW UNTIL WE HAD THIS 5592 04:01:28,808 --> 04:01:29,241 SYSTEM. 5593 04:01:29,241 --> 04:01:32,178 BUT THE SENSITIVITY OF 5594 04:01:32,178 --> 04:01:34,480 STIMULATION IS PROBABLY GETTING 5595 04:01:34,480 --> 04:01:36,882 BETTER AND IT'S HARD TO -- YOU 5596 04:01:36,882 --> 04:01:39,518 NEED TO AGGREGATE A LOT OF DATA 5597 04:01:39,518 --> 04:01:41,587 BEFORE YOU CAN SAY CONCLUSIVELY 5598 04:01:41,587 --> 04:01:43,756 THIS IS MAKING PATIENT 5599 04:01:43,756 --> 04:01:44,790 SEIZURES BETTER. 5600 04:01:44,790 --> 04:01:46,292 THE LIKELIHOOD THAT SOMEONE WILL 5601 04:01:46,292 --> 04:01:49,929 HAVE A SEIZURE IS DEPENDENT UPON 5602 04:01:49,929 --> 04:01:51,430 THEIR PARTICULAR FORM OF 5603 04:01:51,430 --> 04:01:52,031 EPILEPSY. 5604 04:01:52,031 --> 04:01:55,201 WE KNOW THAT BECAUSE OF THESE 5605 04:01:55,201 --> 04:01:55,801 CLOSED-LOOP DEVICES. 5606 04:01:55,801 --> 04:01:56,936 THAT DOESN'T MEAN THAT'S GOING 5607 04:01:56,936 --> 04:01:59,372 TO CHANGE THE CIRCUIT FOR 5608 04:01:59,372 --> 04:02:01,874 EPILEPSY, BUT IT DOES MEAN BY 5609 04:02:01,874 --> 04:02:04,810 HAVING A CLOSED-LOOP SYSTEM FOR 5610 04:02:04,810 --> 04:02:06,245 AN INDIVIDUAL PERSON KNOW WHAT 5611 04:02:06,245 --> 04:02:08,848 THEIR SEIZURE PATTERNS ARE AND 5612 04:02:08,848 --> 04:02:11,016 ADJUST THE PLANS FOR STIMULATION 5613 04:02:11,016 --> 04:02:13,085 AND RESPONSE. THAT'S RUDIMENTARY 5614 04:02:13,085 --> 04:02:16,021 IN TERMS OF MATCHING THE STATISTICS 5615 04:02:16,021 --> 04:02:17,556 OF SEIZURES, BUT IT'S DIFFERENT 5616 04:02:17,556 --> 04:02:18,858 FROM PERSON-TO-PERSON. 5617 04:02:18,858 --> 04:02:23,529 THAT WOULD BE ONE EXAMPLE. 5618 04:02:23,529 --> 04:02:24,730 THE MEDTRONIC DEVICE I THINK 5619 04:02:24,730 --> 04:02:27,900 THAT THEY'RE TRYING TO PUSH A 5620 04:02:27,900 --> 04:02:29,668 PRODUCT BEFORE THEY FIGURE OUT 5621 04:02:29,668 --> 04:02:31,470 THE APPROPRIATE MECHANISM, THERE 5622 04:02:31,470 --> 04:02:36,242 ARE BETTER APPLICATIONS THAN 5623 04:02:36,242 --> 04:02:38,511 PARKINGSON'S FOR CLOSED-LOOP DEVICES. 5624 04:02:38,511 --> 04:02:40,813 FOR EXAMPLE THE WORK COMING OUT 5625 04:02:40,813 --> 04:02:43,582 OF THE UNIVERSITY OF FLORIDA THAT 5626 04:02:43,582 --> 04:02:47,853 AYSE GUNDUZ SHOWED THAT YOU CAN 5627 04:02:47,853 --> 04:02:51,090 DECODE THE ONSET OF TICK AND 5628 04:02:51,090 --> 04:02:53,259 STIMULATE IN RESPONSE FOR PATIENTS 5629 04:02:53,259 --> 04:03:01,200 THAT SUFFER FROM TOURETTE SYNDROME. 5630 04:03:01,200 --> 04:03:03,202 JERKING OF THE BODY IS COMMON. 5631 04:03:03,202 --> 04:03:07,473 WE CAN DETECT THE IMPENDING NEED 5632 04:03:07,473 --> 04:03:09,108 TO EXERCISE A TICK. 5633 04:03:09,108 --> 04:03:14,280 I GUESS THAT DOES CLOSE THE LOOP 5634 04:03:14,280 --> 04:03:16,782 BUT THAT'S AN APPLICATION WHERE 5635 04:03:16,782 --> 04:03:18,284 THE DEVICE IS NEEDED. 5636 04:03:18,284 --> 04:03:20,052 THE POPULATION AND SIZE FOR 5637 04:03:20,052 --> 04:03:24,356 PEOPLE IMPLANTED IS SMALL 5638 04:03:24,356 --> 04:03:25,825 ENOUGH WHERE THAT INS'T THE MAIN 5639 04:03:25,825 --> 04:03:26,992 APPLICATION THAT'S BEING CHASED. 5640 04:03:26,992 --> 04:03:29,528 >> MY QUESTION IS MORE LIKE, IN 5641 04:03:29,528 --> 04:03:32,932 BOTH OF THOSE EXAMPLES, THE TYPE 5642 04:03:32,932 --> 04:03:36,669 OF STIMULATION DELIVERED IS VERY 5643 04:03:36,669 --> 04:03:37,970 NON-BIOLOGICAL. 5644 04:03:37,970 --> 04:03:39,672 IT'S A DISRUPTIVE STIMULATION. 5645 04:03:39,672 --> 04:03:42,475 IF THERE IS AN ASSERTION THAT IF 5646 04:03:42,475 --> 04:03:45,845 WE MAKE THEM MORE BIOLOGICAL AND 5647 04:03:45,845 --> 04:03:49,415 MAKE THEM MORE LIKE NEUROMORPHIC 5648 04:03:49,415 --> 04:03:51,484 IN SOME WAY THAT IT CAN BE 5649 04:03:51,484 --> 04:03:54,186 BETTER, HOW DO WE TRY AND GET 5650 04:03:54,186 --> 04:03:56,856 ANSWERS TO QUESTIONS LIKE THAT 5651 04:03:56,856 --> 04:03:58,891 BEFORE A COMPLETE HARDWARE 5652 04:03:58,891 --> 04:04:00,759 PACKAGE THAT IS ABLE TO DO IT 5653 04:04:00,759 --> 04:04:03,696 HAS BEEN BUILT AND ALL THAT TIME 5654 04:04:03,696 --> 04:04:05,965 AND MONEY HAS BEEN INVESTED SO 5655 04:04:05,965 --> 04:04:07,733 WE CAN GUIDE THE DEVELOPMENT OF 5656 04:04:07,733 --> 04:04:09,468 THINGS LIKE THAT WITHOUT HAVING 5657 04:04:09,468 --> 04:04:11,270 TO REACT TO TECHNOLOGY HANDED TO 5658 04:04:11,270 --> 04:04:13,439 US. 5659 04:04:13,439 --> 04:04:16,408 >> SO SOME THINGS WE DO 5660 04:04:16,408 --> 04:04:17,610 CURRENTLY -- AND PEOPLE MAY OR 5661 04:04:17,610 --> 04:04:19,879 MAY NOT BE FAMILIAR WITH THIS, 5662 04:04:19,879 --> 04:04:24,250 BUT FOR EPILEPSY, WE IMPLANT 5663 04:04:24,250 --> 04:04:25,417 STEREO-EEG WITH DEPTH ELECTRODES 5664 04:04:25,417 --> 04:04:27,953 THROUGHOUT THE BRAIN AND TRY TO 5665 04:04:27,953 --> 04:04:29,622 CHARACTERIZE WHERE A SEIZURE IS 5666 04:04:29,622 --> 04:04:32,291 COMING FROM. THEN WE DO TEST THERAPY 5667 04:04:32,291 --> 04:04:34,393 FOR A COUPLE OF DAYS TO SEE IF IT 5668 04:04:34,393 --> 04:04:39,431 HELPS. THE SAME APPROACH IS ADOPTED 5669 04:04:39,431 --> 04:04:40,533 FOR DEPRESSION AND OCD ACROSS A 5670 04:04:40,533 --> 04:04:47,273 NUMBER OF INSTITUTIONS. WE'VE DONE 5671 04:04:47,273 --> 04:04:47,640 THIS POST-ANOXIC BRAIN INJURY. 5672 04:04:47,640 --> 04:04:50,776 WE HAVE A GOOD TEST ENVIRONMENT. 5673 04:04:50,776 --> 04:04:54,914 I THINK THAT REPRESENTS EXTRA 5674 04:04:54,914 --> 04:04:59,618 SURGERY AND ONE TO TWO WEEK 5675 04:04:59,618 --> 04:05:01,620 INPATIENT STAY. 5676 04:05:01,620 --> 04:05:04,757 THE IDEA WE SHOULD DELIVER THE 5677 04:05:04,757 --> 04:05:07,626 MOST COMMON STIMULATION PATTERNS 5678 04:05:07,626 --> 04:05:09,762 AT A FIXED FREQUENCY. 5679 04:05:09,762 --> 04:05:11,030 A LOT OF THESE HAVE PATHOLOGICAL 5680 04:05:11,030 --> 04:05:14,466 BIOMARKERS ARE OSCILLATORY WITH A 5681 04:05:14,466 --> 04:05:20,139 FIXED TIME SCALE. IT DOESN'T MAKE 5682 04:05:20,139 --> 04:05:21,340 SENSE TO TREAT SOMETHING THAT HAS 5683 04:05:21,340 --> 04:05:23,976 A FIXED TIME SCALE AND TREAT IT 5684 04:05:23,976 --> 04:05:26,679 WITH SOMETHING THAT DOESN'T. 5685 04:05:26,679 --> 04:05:30,316 THERE ARE ACTUALLY TOOLS OUT THERE 5686 04:05:30,316 --> 04:05:32,084 AVAILABLE ON A RESEARCH BASIS 5687 04:05:32,084 --> 04:05:34,386 FOR COMPANIES. 5688 04:05:34,386 --> 04:05:37,323 WE AND OTHERS, I THINK YOU GUYS 5689 04:05:37,323 --> 04:05:40,426 AT LEAST WERE PUTTING IN A PROPOSAL 5690 04:05:40,426 --> 04:05:44,396 ON THIS POTENTIALLY WITH US. 5691 04:05:44,396 --> 04:05:47,733 THE IDEA THAT YOU DELIVER 5692 04:05:47,733 --> 04:05:50,603 STIMULATION IN NEW WAYS, I THINK 5693 04:05:50,603 --> 04:05:51,570 THERE ARE TOOLS FOR THAT. 5694 04:05:51,570 --> 04:05:53,439 THOSE ARE NOT COMMERCIAL GRADE 5695 04:05:53,439 --> 04:05:55,708 TOOLS BUT THEY COULD QUICKLY 5696 04:05:55,708 --> 04:05:56,609 BECOME THAT. 5697 04:05:56,609 --> 04:06:00,112 SORRY FOR TAKING SO LONG. 5698 04:06:00,112 --> 04:06:02,615 >> AND I WANT TO ADDRESS ONE 5699 04:06:02,615 --> 04:06:02,881 THING. 5700 04:06:02,881 --> 04:06:05,951 YOU WERE TALKING ABOUT CENTRAL 5701 04:06:05,951 --> 04:06:06,852 STIMULATION. 5702 04:06:06,852 --> 04:06:09,622 BUT THERE IS ANOTHER ASPECT 5703 04:06:09,622 --> 04:06:12,524 WHICH IS CENTRAL PERIPHERAL 5704 04:06:12,524 --> 04:06:14,827 WHICH I THINK IS MORE RELEVANT 5705 04:06:14,827 --> 04:06:16,362 IN SOME SENSE. 5706 04:06:16,362 --> 04:06:19,765 THINKING ABOUT BLOOD PRESSURE'S 5707 04:06:19,765 --> 04:06:23,335 LINK TO FOLKS WITH SPINAL-CORD 5708 04:06:23,335 --> 04:06:23,902 INJURIES. 5709 04:06:23,902 --> 04:06:27,306 THERE IS A MASSIVE LINK THERE. 5710 04:06:27,306 --> 04:06:29,508 IF YOU COULD MONITOR BLOOD 5711 04:06:29,508 --> 04:06:30,876 PRESSURE, IT DOESN'T EXIST 5712 04:06:30,876 --> 04:06:31,443 TODAY AT LEAST NOT WELL. 5713 04:06:31,443 --> 04:06:33,979 AND USE THAT TO DETERMINE WHEN 5714 04:06:33,979 --> 04:06:36,048 TO STIMULATE THE NERVOUS SYSTEM 5715 04:06:36,048 --> 04:06:40,319 IN ORDER TO BASICALLY REACH 5716 04:06:40,319 --> 04:06:43,789 APPROPRIATE EQUILIBRIUM, THAT'S A 5717 04:06:43,789 --> 04:06:47,760 PLACE NEUROMORPHICS COULD BE A BIG DEAL. 5718 04:06:47,760 --> 04:06:50,295 >> VAGAL NERVE STIMULATION HAVE HAD 5719 04:06:50,295 --> 04:06:52,498 THAT FOR THE LAST SEVEN YEARS 5720 04:06:52,498 --> 04:06:56,135 THE CURRENT GENERATION DETECT 5721 04:06:56,135 --> 04:07:00,105 RAPID INCREASES IN HEART RATE AND 5722 04:07:00,105 --> 04:07:04,143 WILL DELIVER STIMULATION IN RESPONSE 5723 04:07:04,143 --> 04:07:04,910 IT'S A PERIPHERAL SENSOR. 5724 04:07:04,910 --> 04:07:08,514 >> THAT'S HEART RATE. 5725 04:07:08,514 --> 04:07:10,249 >> RIGHT. 5726 04:07:10,249 --> 04:07:11,884 >> I'M DOING POORLY ON TIME 5727 04:07:11,884 --> 04:07:12,718 MANAGEMENT. 5728 04:07:12,718 --> 04:07:15,421 WE'RE FOUR MINUTE OVER THE 5729 04:07:15,421 --> 04:07:18,924 SCHEDULED BREAK TIME AND WE 5730 04:07:18,924 --> 04:07:21,760 STILL HAVE KAREN WHO HASN'T HAD 5731 04:07:21,760 --> 04:07:24,163 A CHANCE TO ASK HER QUESTION. 5732 04:07:24,163 --> 04:07:26,498 DO WE WANT TO TAKE TIME FROM THE 5733 04:07:26,498 --> 04:07:28,500 4:00 WRAP-UP TO CONTINUE SOME OF 5734 04:07:28,500 --> 04:07:30,903 THESE QUESTIONS? 5735 04:07:30,903 --> 04:07:33,672 >> YES, WE CAN DO THAT. 5736 04:07:33,672 --> 04:07:43,849 FOR SURE. 5737 04:07:46,418 --> 04:07:48,554 >> WE HAVE A QUESTION COME IN 5738 04:07:48,554 --> 04:07:50,589 AND WE WILL HAVE TO TAKE THAT 5739 04:07:50,589 --> 04:07:51,490 ONE AT 4:00. 5740 04:07:51,490 --> 04:07:53,525 I'M SORRY FOR THE PEOPLE STANDING 5741 04:07:53,525 --> 04:07:54,593 PATIENTLY. 5742 04:07:54,593 --> 04:07:56,061 WE CAN TAKE YOUR QUESTION AT 5743 04:07:56,061 --> 04:07:56,261 4:00. 5744 04:07:56,261 --> 04:07:59,264 THERE WILL BEING A RAPID FIRE 5745 04:07:59,264 --> 04:08:00,699 WRAP-UP BECAUSE THIS DISCUSSION 5746 04:08:00,699 --> 04:08:01,333 IS SO INTERESTING. 5747 04:08:01,333 --> 04:08:03,502 THANK YOU, EVERYBODY. 5748 04:08:03,502 --> 04:08:05,871 WE'LL RESUME AT 4:00 PM. 5749 04:08:05,871 --> 04:08:07,906 WELCOME BACK PANELISTS 5750 04:08:07,906 --> 04:08:09,374 AND DISCUSSANTS. 5751 04:08:09,374 --> 04:08:11,643 WE'RE GOING TO CONTINUE WITH 5752 04:08:11,643 --> 04:08:13,712 SESSION FOUR. 5753 04:08:13,712 --> 04:08:19,251 FOR ANYONE WHO HAS QUESTIONS 5754 04:08:19,251 --> 04:08:20,886 THAT -- WE'LL PROBABLY NOT HAVE 5755 04:08:20,886 --> 04:08:23,255 TIME TO ASK ALL YOUR QUESTIONS, 5756 04:08:23,255 --> 04:08:26,692 SO FEEL FREE TO SCAN THE QR CODE 5757 04:08:26,692 --> 04:08:28,494 AND ASK YOUR QUESTION THERE. 5758 04:08:28,494 --> 04:08:31,196 WE'LL BE ABLE TO INCLUDE THAT IN 5759 04:08:31,196 --> 04:08:33,365 OUR POST WORKSHOP REPORT. 5760 04:08:33,365 --> 04:08:40,405 IF WE'RE REALLY AMBITIOUS, WE'LL 5761 04:08:40,405 --> 04:08:42,274 FIND SOMEONE TO ANSWER THOSE 5762 04:08:42,274 --> 04:08:43,208 QUESTIONS. 5763 04:08:43,208 --> 04:08:45,177 LET ME GET STARTED WITH WHERE WE 5764 04:08:45,177 --> 04:08:47,713 LEFT OFF. 5765 04:08:47,713 --> 04:08:55,654 BING HAD A QUESTION. 5766 04:08:55,654 --> 04:08:56,655 MACHINES HAVE CAPABILITIES WE 5767 04:08:56,655 --> 04:08:59,958 DON'T HAVE, RADARS, WHEELS AND 5768 04:08:59,958 --> 04:09:01,326 JET ENGINES. 5769 04:09:01,326 --> 04:09:06,298 WHAT ARE LOW-HANGING FRUIT FOR 5770 04:09:06,298 --> 04:09:09,168 INTERFACING HUMAN COMMUNITIES WITH 5771 04:09:09,168 --> 04:09:14,573 ENGINEERED SYSTEMS. AND DO WE NEED 5772 04:09:14,573 --> 04:09:19,878 NEUROMORPHIC COMPUTING TO DO THIS? 5773 04:09:19,878 --> 04:09:21,313 >> I DON'T THINK THE ARGUMENT IS 5774 04:09:21,313 --> 04:09:25,317 THAT WE NEED ALL ASPECTS OF 5775 04:09:25,317 --> 04:09:27,186 BIOLOGICAL COMPUTATION TO DO 5776 04:09:27,186 --> 04:09:29,121 THESE TASKS. 5777 04:09:29,121 --> 04:09:32,491 IT IS THE CLASSIC CASE OF BIRDS 5778 04:09:32,491 --> 04:09:35,027 FLAP AND PLANES USE ENGINES. 5779 04:09:35,027 --> 04:09:37,529 THE KEY THING TO KNOW IS THERE 5780 04:09:37,529 --> 04:09:40,699 ARE ASPECTS OF BIOLOGICAL 5781 04:09:40,699 --> 04:09:41,400 INTELLIGENCE THAT WE CANNOT 5782 04:09:41,400 --> 04:09:43,502 CURRENTLY DO WITH THE STRATEGIES 5783 04:09:43,502 --> 04:09:44,636 WE'RE CURRENTLY USING. 5784 04:09:44,636 --> 04:09:47,272 WE'RE GOING TO HAVE TO FIGURE 5785 04:09:47,272 --> 04:09:49,942 OUT WHAT THOSE PRINCIPLES ARE IN 5786 04:09:49,942 --> 04:09:50,542 NEURAL SYSTEMS. 5787 04:09:50,542 --> 04:09:53,011 FOR ME PERSONALLY, I'M NOT 5788 04:09:53,011 --> 04:09:54,479 CONVINCED THAT THE SPIKE IS 5789 04:09:54,479 --> 04:09:56,648 IMPORTANT AND NECESSARY FOR 5790 04:09:56,648 --> 04:09:59,284 THESE KINDS OF 5791 04:09:59,284 --> 04:10:00,219 NEURALL COMPUTATIONS. 5792 04:10:00,219 --> 04:10:06,124 BUT NEUROMORPHIC AND NEUROAI 5793 04:10:06,124 --> 04:10:09,294 SYSTEMS -- WE HAVE TO FIGURE OUT 5794 04:10:09,294 --> 04:10:12,064 THE CORE PRINCIPLES BEFORE WE 5795 04:10:12,064 --> 04:10:16,969 CAN DISTILL THOSE CORE 5796 04:10:16,969 --> 04:10:19,471 PRINCIPLES DOWN. 5797 04:10:19,471 --> 04:10:22,541 >> I THINK THIS DOVETAILS INTO 5798 04:10:22,541 --> 04:10:24,243 THE QUESTION I ASKED BEFORE WE 5799 04:10:24,243 --> 04:10:26,378 ENDED THE SESSION, WE'RE TALKING 5800 04:10:26,378 --> 04:10:29,281 A LOT ABOUT TECHNOLOGY BUT THERE 5801 04:10:29,281 --> 04:10:30,549 ARE FOUNDATIONAL SCIENCE 5802 04:10:30,549 --> 04:10:32,417 QUESTIONS THAT WE HAVE TO ASK 5803 04:10:32,417 --> 04:10:34,219 ESPECIALLY IN THE DOMAIN OF 5804 04:10:34,219 --> 04:10:36,755 HUMAN INTERFACE EXAMINING 5805 04:10:36,755 --> 04:10:37,990 NEUROTECHNOLOGY ABOUT WHAT THE 5806 04:10:37,990 --> 04:10:40,559 VALUE OF THE SYSTEMS IS AND HOW 5807 04:10:40,559 --> 04:10:41,960 TO USE THEM BEFORE THEY ARE 5808 04:10:41,960 --> 04:10:44,129 BUILT BASED ON THE PERSPECTIVES 5809 04:10:44,129 --> 04:10:45,898 OF CLINICIANS AND LIVED 5810 04:10:45,898 --> 04:10:47,032 EXPERIENCE FOLKS AND WHAT THE 5811 04:10:47,032 --> 04:10:49,101 NEEDS ARE SO WE'RE NOT JUST 5812 04:10:49,101 --> 04:10:51,069 DESIGNING A TECHNOLOGY AND THEN 5813 04:10:51,069 --> 04:10:52,571 TRYING TO FIGURE OUT WHAT NEEDS 5814 04:10:52,571 --> 04:10:56,575 IT MIGHT BE ABLE TO MEET, BUT 5815 04:10:56,575 --> 04:10:57,409 WE'RE DESIGNING TO THE NEED 5816 04:10:57,409 --> 04:10:59,044 RATHER THAN LETTING THE 5817 04:10:59,044 --> 04:10:59,444 TECHNOLOGY LEAD. 5818 04:10:59,444 --> 04:11:02,214 I THINK THERE IS A NEED FOR A 5819 04:11:02,214 --> 04:11:03,348 FOUNDATIONAL SCIENCE PROGRAM TO 5820 04:11:03,348 --> 04:11:05,017 GO ALONG WITH ALL THE THINGS 5821 04:11:05,017 --> 04:11:06,585 THAT WE'RE DISCUSSING SO WE'RE 5822 04:11:06,585 --> 04:11:07,452 BUILDING THINGS THAT ARE GOING 5823 04:11:07,452 --> 04:11:09,721 TO BE USEFUL AND NOT JUST 5824 04:11:09,721 --> 04:11:11,657 ASSUMING THEY'LL BE USEFUL ONCE 5825 04:11:11,657 --> 04:11:12,925 THEY ARE BUILT. 5826 04:11:12,925 --> 04:11:16,595 >> ALL RIGHT. MAKES A LOT 5827 04:11:16,595 --> 04:11:19,531 OF SENSE ALONG THE LINES OF 5828 04:11:19,531 --> 04:11:20,666 PARTICIPATORY DESIGN. 5829 04:11:20,666 --> 04:11:24,636 I THINK THE NEXT QUESTION THAT 5830 04:11:24,636 --> 04:11:30,309 WE HAD WAS REALLY A COMMENT FROM 5831 04:11:30,309 --> 04:11:31,543 GIACOMO. 5832 04:11:31,543 --> 04:11:38,984 GIACOMO, ARE YOU STILL ON LINE? 5833 04:11:38,984 --> 04:11:41,053 >> YES, YES, I AM. 5834 04:11:41,053 --> 04:11:43,989 >> HE'S IN EUROPE. 5835 04:11:43,989 --> 04:11:49,761 SO THE NEUROMORPHIC COMMUNITY IS 5836 04:11:49,761 --> 04:11:50,028 PRAGMATIC. SO IF THERE'S A BENEFIT 5837 04:11:50,028 --> 04:11:51,830 IN USING NEUROMORPHIC PROCESSING ON 5838 04:11:51,830 --> 04:11:53,332 NON BIOLOGICAL DATA SUCH AS RADAR, WE'LL DO IT. 5839 04:11:53,332 --> 04:11:56,768 BUT IF THERE IS A NEED TO USE 5840 04:11:56,768 --> 04:11:58,203 CONVENTIONAL COMPUTING, WE'LL DO 5841 04:11:58,203 --> 04:11:59,137 THAT AS WELL. 5842 04:11:59,137 --> 04:12:00,772 ONE THING DOES NOT RULE THE 5843 04:12:00,772 --> 04:12:02,107 OTHER OUT. 5844 04:12:02,107 --> 04:12:04,409 THE ONE PLACE WHERE THE 5845 04:12:04,409 --> 04:12:06,011 NEUROMORPHIC APPROACH WILL HELP 5846 04:12:06,011 --> 04:12:08,313 HERE IS THE NEED TO DO THINGS AT 5847 04:12:08,313 --> 04:12:10,082 LOW POWER ONLINE ESPECIALLY 5848 04:12:10,082 --> 04:12:13,418 WHEN THE SENSORY DATA IS SPARSE. 5849 04:12:13,418 --> 04:12:17,689 OH, THERE IS GIACOMO. 5850 04:12:17,689 --> 04:12:19,491 NOTHING ELSE TO ADD? 5851 04:12:19,491 --> 04:12:20,492 >> NO. 5852 04:12:20,492 --> 04:12:22,094 >> THEN I'LL TURN IT OVER TO 5853 04:12:22,094 --> 04:12:24,162 KAREN WHO HAS BEEN PATIENTLY 5854 04:12:24,162 --> 04:12:26,798 WAITING TO ASK HER PROVOCATIVE 5855 04:12:26,798 --> 04:12:28,266 OR LINKING QUESTION. 5856 04:12:28,266 --> 04:12:29,368 >> THANKS TO EVERYBODY FOR 5857 04:12:29,368 --> 04:12:32,671 COMING BACK AND FOR SPEAKERS TO 5858 04:12:32,671 --> 04:12:34,206 STAY ONLINE ESPECIALLY GIACOMO. 5859 04:12:34,206 --> 04:12:37,876 I KNOW IT'S LATE. 5860 04:12:37,876 --> 04:12:38,510 FIRST I WANTED TO COMMENT ON 5861 04:12:38,510 --> 04:12:40,112 WHAT CHRIS SAID BECAUSE IT'S SO 5862 04:12:40,112 --> 04:12:41,713 IMPORTANT THAT THE TYPES OF 5863 04:12:41,713 --> 04:12:46,518 QUESTIONS WE DECIDE TO ASK WITH 5864 04:12:46,518 --> 04:12:48,920 NEUROAI THEY'RE NOT 5865 04:12:48,920 --> 04:12:50,288 MORALLY NEUTRAL. A DECISION ON 5866 04:12:50,288 --> 04:12:53,692 WHAT WE'RE MODELING, WHAT WE CHOOSE 5867 04:12:53,692 --> 04:12:55,761 TO USE THE SCIENCE FOR, AND WHAT WE 5868 04:12:55,761 --> 04:12:59,398 CLAIM TO BE SAYING, THESE ALL COME 5869 04:12:59,398 --> 04:13:03,769 WITH EMBEDDED VALUES AND CONFLICTS 5870 04:13:03,769 --> 04:13:06,405 AND NEED TO BE DETERMINED IN ANY 5871 04:13:06,405 --> 04:13:08,440 NEURO AI EFFORTS WITH THE COMMUNITIES 5872 04:13:08,440 --> 04:13:11,843 FOR WHOM THIS MIGHT MATTER MOST 5873 04:13:11,843 --> 04:13:13,045 INCLUDING THE BUILDERS AND 5874 04:13:13,045 --> 04:13:14,579 FUTURE USERS. 5875 04:13:14,579 --> 04:13:17,916 I'M THINKING ABOUT GIACOMO. 5876 04:13:17,916 --> 04:13:21,386 I'M GLAD YOU'RE STILL HERE AND 5877 04:13:21,386 --> 04:13:22,387 RALPH I'M GLAD YOU'RE STILL HERE. 5878 04:13:22,387 --> 04:13:24,756 I WANT TO TALK ABOUT BCIS. 5879 04:13:24,756 --> 04:13:30,295 IN THE NEUROETHICS LITERATURE, 5880 04:13:30,295 --> 04:13:34,166 WE'VE HAD DISCUSSION ON BCIS 5881 04:13:34,166 --> 04:13:36,568 IN GENERAL AND HOW THEY 5882 04:13:36,568 --> 04:13:38,937 INTERVENE WITH THE BRAIN AND 5883 04:13:38,937 --> 04:13:43,442 WHAT MEANS FOR INDIVIDUALS, THEIR 5884 04:13:43,442 --> 04:13:45,410 IDENTITY, ATYPICAL SAFETY CONCERNS. 5885 04:13:45,410 --> 04:13:47,179 NOT JUST IMPLANTING A DEVICE BUT 5886 04:13:47,179 --> 04:13:52,117 WHAT IT DOES TO INFLUENCE THEIR 5887 04:13:52,117 --> 04:13:52,951 PERCEPTION OF THE WORLD ET 5888 04:13:52,951 --> 04:13:54,386 CETERA. 5889 04:13:54,386 --> 04:13:57,656 SCIENTISTS WHO WORKS ON BCI'S 5890 04:13:57,656 --> 04:13:59,624 REPORT SHARING THOSE CONCERNS AND 5891 04:13:59,624 --> 04:14:01,126 THEY'RE TRYING TO FIGURE OUT 5892 04:14:01,126 --> 04:14:02,160 WHAT TO DO WITH THAT. 5893 04:14:02,160 --> 04:14:05,330 A LOT OF THE SOLUTIONS ARE 5894 04:14:05,330 --> 04:14:08,834 AROUND IN ETHICAL LEGAL CATEGORY, 5895 04:14:08,834 --> 04:14:09,434 CONSENT. 5896 04:14:09,434 --> 04:14:12,537 THAT'S A GOOD START BUT PEOPLE 5897 04:14:12,537 --> 04:14:15,173 OFTEN MISTAKE THAT CONTENT IS A 5898 04:14:15,173 --> 04:14:15,874 LIABILITY DEVICE. 5899 04:14:15,874 --> 04:14:19,678 I HAVE YET IN ALL MY YEARS 5900 04:14:19,678 --> 04:14:22,814 WORKING SEEN ONE WRITTEN AT AN 5901 04:14:22,814 --> 04:14:26,084 APPROPRIATE HUMAN READABLE LEVEL. 5902 04:14:26,084 --> 04:14:32,357 WHAT I ADVOCATE FOR IS A 5903 04:14:32,357 --> 04:14:33,125 HUMAN-IN-THE-LOOP ETHICS BY DESIGN 5904 04:14:33,125 --> 04:14:36,261 APPROACH. BRINGS ME TO AN OPERATIONAL 5905 04:14:36,261 --> 04:14:38,063 QUESTION ON HOW TECHNOLOGY 5906 04:14:38,063 --> 04:14:41,199 WORKS AND HOW IT MIGHT BE USED 5907 04:14:41,199 --> 04:14:41,500 CLINICALLY. 5908 04:14:41,500 --> 04:14:44,503 THE FIRST IS A BASIC ONE. 5909 04:14:44,503 --> 04:14:47,873 WE SAY OPERATIONALLY, THERE ARE 5910 04:14:47,873 --> 04:14:50,642 GREAT THINGS ABOUT BCIS. 5911 04:14:50,642 --> 04:14:52,911 THERE IS POTENTIAL ENERGY 5912 04:14:52,911 --> 04:14:54,646 EFFICIENCY AND MAYBE MORE 5913 04:14:54,646 --> 04:14:56,181 REAL-TIME CAPABILITIES. 5914 04:14:56,181 --> 04:14:56,915 IT'S SMALLER. 5915 04:14:56,915 --> 04:14:58,250 THERE IS EVEN THE POSSIBILITY 5916 04:14:58,250 --> 04:15:00,252 THAT THERE ARE MORE NATIVE 5917 04:15:00,252 --> 04:15:01,586 CONNECTIONS YOU CAN MAKE WHEN 5918 04:15:01,586 --> 04:15:03,221 INTERFACING WITH THE BRAIN. 5919 04:15:03,221 --> 04:15:05,157 EVEN POTENTIAL FOR EDGE 5920 04:15:05,157 --> 04:15:07,259 COMPUTING OR A SECURITY FEATURE. 5921 04:15:07,259 --> 04:15:08,393 THIS IS SOMETHING WE TALKED 5922 04:15:08,393 --> 04:15:09,861 ABOUT THIS MORNING. 5923 04:15:09,861 --> 04:15:12,497 MANY BCIS OFTEN CONNECT 5924 04:15:12,497 --> 04:15:18,703 BETWEEN UNSECURE INTERNET 5925 04:15:18,703 --> 04:15:19,371 CONNECTIONS OR BLUETOOTH. 5926 04:15:19,371 --> 04:15:21,039 I WONDER ABOUT -- THERE ARE 5927 04:15:21,039 --> 04:15:22,440 BENEFITS TO AUTONOMY. 5928 04:15:22,440 --> 04:15:26,044 YOU CAN RESTORE SOMEONE'S 5929 04:15:26,044 --> 04:15:27,913 AUTONOMY AND AGENCY TO OPERATE 5930 04:15:27,913 --> 04:15:31,216 IN THE WORLD WITH THESE DEVICES. 5931 04:15:31,216 --> 04:15:33,919 BUT DOES NEUROMORPHIC COMPUTING 5932 04:15:33,919 --> 04:15:41,326 IT'S POTENTIAL TO NATIVELY INTERACT 5933 04:15:41,326 --> 04:15:43,795 EXACERBATE TENSIONS ABOUT SAFETY FEATURES? 5934 04:15:43,795 --> 04:15:45,764 IT MIGHT BE INTERACTING WITH THE 5935 04:15:45,764 --> 04:15:48,233 BRAIN AND OTHER WAYS WHERE IT 5936 04:15:48,233 --> 04:15:49,901 MITIGATES THOSE TYPES OF 5937 04:15:49,901 --> 04:15:50,335 CONCERNS? 5938 04:15:50,335 --> 04:15:54,472 I MENTIONED SAFETY ISSUE AND I'M 5939 04:15:54,472 --> 04:15:57,242 CURIOUS IF THOSE QUESTIONS CAME 5940 04:15:57,242 --> 04:16:00,345 UP AT THE WORKSHOP YOU WERE 5941 04:16:00,345 --> 04:16:03,281 TALKING ABOUT, RALPH. 5942 04:16:03,281 --> 04:16:04,716 >> THE CYBERSECURITY ASPECT CAME 5943 04:16:04,716 --> 04:16:06,418 UP. 5944 04:16:06,418 --> 04:16:09,921 AND IT WAS SOMETHING THAT WAS 5945 04:16:09,921 --> 04:16:13,625 PERCEIVED AS BEING AN AREA OF 5946 04:16:13,625 --> 04:16:16,127 INVESTMENT HASN'T BEEN ENOUGH 5947 04:16:16,127 --> 04:16:16,962 INVESTMENT INTO STUDY. 5948 04:16:16,962 --> 04:16:19,097 ALSO FROM THE PERSPECTIVE OF 5949 04:16:19,097 --> 04:16:19,531 STANDARDS. 5950 04:16:19,531 --> 04:16:25,203 THAT WAS ANOTHER COMPONENT. 5951 04:16:25,203 --> 04:16:26,738 I THINK IEEE IS DEVELOPING 5952 04:16:26,738 --> 04:16:33,111 STANDARDS ON A.I. AND SPECIALLY 5953 04:16:33,111 --> 04:16:36,181 ON NEUROAI WITH CONNECTION 5954 04:16:36,181 --> 04:16:38,583 BETWEEN NERVOUS SYSTEM AND A.I. 5955 04:16:38,583 --> 04:16:43,855 THERE IS NO SOLUTION JUST YET. 5956 04:16:43,855 --> 04:16:47,425 SO MOVING AWAY KIND OF THAT'S 5957 04:16:47,425 --> 04:16:51,329 THE HIGH LEVEL, THE METHOD SIDE 5958 04:16:51,329 --> 04:16:54,232 OF WHAT WE CAN DO TO PROTECT THE 5959 04:16:54,232 --> 04:16:56,401 USERS OF THE DEVICES. 5960 04:16:56,401 --> 04:17:00,071 MAYBE THIS IS WHERE THE OLD 5961 04:17:00,071 --> 04:17:01,273 SCHOOL NEUROMORPHS CAN BE HELPFUL. 5962 04:17:01,273 --> 04:17:04,175 THE HARDWARE IS SO SPECIFICALLY 5963 04:17:04,175 --> 04:17:06,945 DESIGNED IN A WAY TO INTERFACE 5964 04:17:06,945 --> 04:17:08,446 WITH THE NERVOUS SYSTEM AS WELL AS 5965 04:17:08,446 --> 04:17:12,817 TO MIMIC IT IN SUCH A WAY THAT IT 5966 04:17:12,817 --> 04:17:15,887 TRIES TO USE THE PHYSICS OF 5967 04:17:15,887 --> 04:17:18,423 COMPUTATION IN ORDER TO SOLVE THE 5968 04:17:18,423 --> 04:17:20,058 PROBLEM AT HAND. 5969 04:17:20,058 --> 04:17:22,160 THERE IS TYPICALLY NOT A LOT OF 5970 04:17:22,160 --> 04:17:23,361 EXTERNAL CONNECTIONS. 5971 04:17:23,361 --> 04:17:27,599 THERE IS NO USB OR BLUETOOTH 5972 04:17:27,599 --> 04:17:30,235 ELEMENT TO IT. 5973 04:17:30,235 --> 04:17:33,038 MAYBE IN THAT SENSE WE CAN 5974 04:17:33,038 --> 04:17:37,309 MITIGATE SOME OF THE DANGER BY 5975 04:17:37,309 --> 04:17:39,644 HAVING NOT THE TRADITIONAL 5976 04:17:39,644 --> 04:17:41,846 INTERFACES, BUT MORE SPECIFIC 5977 04:17:41,846 --> 04:17:43,715 INTERFACE THAT JUST CLOSES THE 5978 04:17:43,715 --> 04:17:44,416 LOOP. 5979 04:17:44,416 --> 04:17:46,117 THE EXAMPLE THAT I GAVE WHERE 5980 04:17:46,117 --> 04:17:54,125 WE'RE WATCHING AN ELECTRODE AND 5981 04:17:54,125 --> 04:17:55,894 ADAPTING LOCALLY, THAT HAS NO 5982 04:17:55,894 --> 04:17:58,363 NEED FOR EXTERNAL 5983 04:17:58,363 --> 04:17:58,863 COMMUNICATIONS. 5984 04:17:58,863 --> 04:18:01,366 SO THE SAFETY ASPECT ON HACKING 5985 04:18:01,366 --> 04:18:03,735 INTO IT AND SO ON CAN BE 5986 04:18:03,735 --> 04:18:04,069 AVOIDED. 5987 04:18:04,069 --> 04:18:06,838 BUT THE OTHER SIDE IS WHAT IF IT 5988 04:18:06,838 --> 04:18:07,872 RUNS OFF THE RAILS? 5989 04:18:07,872 --> 04:18:09,741 WHAT IF THE CONTROL LOOP 5990 04:18:09,741 --> 04:18:11,276 IS OUT OF TRACK? 5991 04:18:11,276 --> 04:18:17,082 HOW DO YOU SOLVE FOR THAT? 5992 04:18:17,082 --> 04:18:19,718 >> THAT'S A POINT I HOPE GIACOMO 5993 04:18:19,718 --> 04:18:20,852 WILL ADDRESS. 5994 04:18:20,852 --> 04:18:22,053 >> EXACTLY. 5995 04:18:22,053 --> 04:18:25,156 I AGREE COMPLETELY THE 5996 04:18:25,156 --> 04:18:26,324 NEUROMORPHIC TYPE OF SYSTEMS 5997 04:18:26,324 --> 04:18:29,728 HAVE LESS OF A PROBLEM. 5998 04:18:29,728 --> 04:18:34,032 IT'S LESS OF THIS INTERPRETABILITY. 5999 04:18:34,032 --> 04:18:39,771 THEY ARE MORE INTERPRETABLE THAN THE 6000 04:18:39,771 --> 04:18:40,271 BLACK BOX APPROACH OF STANDARD A.I. 6001 04:18:40,271 --> 04:18:42,640 THERE IS STILL THE PROBLEM OF 6002 04:18:42,640 --> 04:18:44,709 HAVING A SYSTEM, ESPECIALLY IF 6003 04:18:44,709 --> 04:18:48,012 IT HAS ADAPTATION PROPERTIES OR 6004 04:18:48,012 --> 04:18:49,314 ON-CHIP LEARNING PROPERTIES OF HAVING 6005 04:18:49,314 --> 04:18:51,716 A SYSTEM THAT IS AUTONOMOUS AND 6006 04:18:51,716 --> 04:18:53,985 THAT INTERACTS WITH ANOTHER 6007 04:18:53,985 --> 04:18:55,553 AUTONOMOUS SYSTEM, THE REAL BRAIN 6008 04:18:55,553 --> 04:18:57,889 OR REAL NERVOUS OR PERIPERAL SYSTEMS. 6009 04:18:57,889 --> 04:18:59,791 IT'S A PROBLEM THAT WE ARE 6010 04:18:59,791 --> 04:19:01,426 STARTING TO FACE ESPECIALLY ALSO 6011 04:19:01,426 --> 04:19:04,396 BECAUSE IN THE NEUROMORPHIC 6012 04:19:04,396 --> 04:19:08,400 COMMUNITY, SOME OF US ARE TRYING 6013 04:19:08,400 --> 04:19:12,737 TO LINK THE ARTIFICIAL NEURONS 6014 04:19:12,737 --> 04:19:13,838 WITH REAL NEURONS EVEN ORGANOIDS. 6015 04:19:13,838 --> 04:19:15,273 SO YOU CAN IMAGINE THERE ARE 6016 04:19:15,273 --> 04:19:17,442 MANY ETHICAL QUESTIONS THAT 6017 04:19:17,442 --> 04:19:18,843 ARISE IN THIS DOMAIN. 6018 04:19:18,843 --> 04:19:23,748 I CAN TELL YOU WE'VE BEEN 6019 04:19:23,748 --> 04:19:25,450 INVOLVING PEOPLE THAT ARE 6020 04:19:25,450 --> 04:19:27,685 EXPERTS IN ETHICS IN OUR 6021 04:19:27,685 --> 04:19:28,987 WORKSHOPS AND DISCUSSIONS TO TRY 6022 04:19:28,987 --> 04:19:31,322 TO START TO WORRY ABOUT THESE 6023 04:19:31,322 --> 04:19:33,057 THINGS EVEN BEFORE WE START TO 6024 04:19:33,057 --> 04:19:34,793 DESIGN THESE CIRCUITS AND CHIPS 6025 04:19:34,793 --> 04:19:38,430 THAT WILL EVENTUALLY MIGHT END 6026 04:19:38,430 --> 04:19:41,466 UP IN BRAIN MACHINE INTERFACE. 6027 04:19:41,466 --> 04:19:43,101 I AGREE, IT'S AN IMPORTANT PROBLEM. 6028 04:19:43,101 --> 04:19:43,701 THE COMMUNITY IS AWARE OF THE 6029 04:19:43,701 --> 04:19:48,206 PROBLEM. 6030 04:19:48,206 --> 04:19:51,209 CHIARA LED A WORKSHOP IN THE 6031 04:19:51,209 --> 04:19:52,844 NEUROTECH AND A.I. INITIATIVE SHE 6032 04:19:52,844 --> 04:19:54,312 LED WITH HER COLLEAGUES IN 6033 04:19:54,312 --> 04:19:57,982 EUROPE SOME YEARS AGO. 6034 04:19:57,982 --> 04:19:59,984 EVEN IF THE PAST RECENT YEARS 6035 04:19:59,984 --> 04:20:03,988 WE'VE BEEN INVOLVING COLLEAGUES 6036 04:20:03,988 --> 04:20:05,990 I DON'T HAVE A SOLUTION TO THE 6037 04:20:05,990 --> 04:20:06,224 PROBLEM. 6038 04:20:06,224 --> 04:20:07,959 >> I THINK IT'S GOOD TO HEAR 6039 04:20:07,959 --> 04:20:09,494 YOU'RE THINKING ABOUT IT AND I 6040 04:20:09,494 --> 04:20:12,030 WANTED TO POINT OUT THIS IS AN 6041 04:20:12,030 --> 04:20:13,198 OPPORTUNITY BECAUSE IT'S NOT A TOPIC. 6042 04:20:13,198 --> 04:20:16,501 HBP HAS BEEN INVOLVED THAT'S HOW 6043 04:20:16,501 --> 04:20:18,703 WE'VE INTERACTED. 6044 04:20:18,703 --> 04:20:24,042 WE'VE LOOKED AT ETHICAL ISSUES. 6045 04:20:24,042 --> 04:20:26,744 BUT IT'S NOT A PART OF THE 6046 04:20:26,744 --> 04:20:31,983 CONVERSATION OF IMPLANTABLE BCI, 6047 04:20:31,983 --> 04:20:36,120 FDA COMMUNITY COLLABORATORY. 6048 04:20:36,120 --> 04:20:38,523 THIS IS SOMETHING TO DISCUSS TO 6049 04:20:38,523 --> 04:20:40,992 UPGRADE OUR PROCESSES AS WE 6050 04:20:40,992 --> 04:20:42,060 EXPLORE THESE THINGS. 6051 04:20:42,060 --> 04:20:45,530 >> MAY I ASK A QUICK QUESTION. 6052 04:20:45,530 --> 04:20:47,799 FOR PACE MAKERS AND SO ON, THAT 6053 04:20:47,799 --> 04:20:50,034 IS ALSO HAS THE SAME KIND OF 6054 04:20:50,034 --> 04:20:53,204 RISKS, IF YOU WILL AS THE BCI 6055 04:20:53,204 --> 04:20:53,738 SCENARIO. 6056 04:20:53,738 --> 04:20:58,443 WHAT DO THEY DO THERE? 6057 04:20:58,443 --> 04:21:02,413 >> I SAY THAT IF YOU HAVE A 6058 04:21:02,413 --> 04:21:05,250 HEART TRANSPLANT, YOU'RE 6059 04:21:05,250 --> 04:21:06,885 PROBABLY STILL YOU. 6060 04:21:06,885 --> 04:21:11,155 IF YOU GET A BRAIN TRANSPLANT 6061 04:21:11,155 --> 04:21:12,957 YOU ARE NOT YOU. 6062 04:21:12,957 --> 04:21:15,927 NOT TO DISPARAGE THE HEART, BUT 6063 04:21:15,927 --> 04:21:18,029 I THINK THE TYPES MUCH CHANGES 6064 04:21:18,029 --> 04:21:19,898 ALSO MIGHT BE DIFFERENT. THEY 6065 04:21:19,898 --> 04:21:23,034 MIGHT BE SLOWER OR FASTER. 6066 04:21:23,034 --> 04:21:25,970 THERE ARE THINGS WE DON'T UNDERSTAND 6067 04:21:25,970 --> 04:21:30,275 EVEN WITH REGULAR BRAIN COMPUTER 6068 04:21:30,275 --> 04:21:31,543 INTERFACES, THE EFFECTS OF TIME. 6069 04:21:31,543 --> 04:21:32,777 THERE ARE CHALLENGES WITH THE 6070 04:21:32,777 --> 04:21:35,647 BRAIN BECAUSE IT'S -- WELL THERE 6071 04:21:35,647 --> 04:21:38,716 ARE A LOT OF THINGS SPECIAL 6072 04:21:38,716 --> 04:21:39,984 ABOUT THE BRAIN. 6073 04:21:39,984 --> 04:21:42,053 OTHER THINGS TO THINK ABOUT WITH 6074 04:21:42,053 --> 04:21:47,892 THE THESE MORE SOPHISTICATED 6075 04:21:47,892 --> 04:21:49,694 TECHNIQUES AS THEY RAPIDLY EVOLVE. 6076 04:21:49,694 --> 04:21:51,963 WE DO NOT HAVE ANYWHERE IN THE WORLD 6077 04:21:51,963 --> 04:21:57,669 A GOOD STANDARD FOR ENSURING 6078 04:21:57,669 --> 04:21:59,337 BACKWARDS COMPATIBILITY AND UPGRADES. 6079 04:21:59,337 --> 04:22:00,204 IT WOULD BE IMPORTANT TO 6080 04:22:00,204 --> 04:22:02,540 CONSIDER THAT AS WELL. 6081 04:22:02,540 --> 04:22:06,744 I RECALL CHRIS AND I HAVE A 6082 04:22:06,744 --> 04:22:10,148 FRIEND, BRANDY ELLIS, WHO HAS 6083 04:22:10,148 --> 04:22:10,982 A DBS IMPLANT. 6084 04:22:10,982 --> 04:22:12,584 SHE SAID PUBLICLY BEFORE SHE 6085 04:22:12,584 --> 04:22:15,019 MISSED THE OPPORTUNITY TO GET AN 6086 04:22:15,019 --> 04:22:15,753 UPGRADE. IT WAS RIGHT BEFORE 6087 04:22:15,753 --> 04:22:18,523 GETTING IMPLANTED AND NOW SHE HAS 6088 04:22:18,523 --> 04:22:21,092 A LESS SECURE DEVICE. THERE AREN'T 6089 04:22:21,092 --> 04:22:22,360 WAYS TO CHANGE THAT FOR HER. 6090 04:22:22,360 --> 04:22:24,963 HER DEVICE WORKS BEAUTIFULLY FOR 6091 04:22:24,963 --> 04:22:25,563 HER AT THE SAME TIME. 6092 04:22:25,563 --> 04:22:27,465 WE DON'T HAVE MECHANISMS IN 6093 04:22:27,465 --> 04:22:30,101 PLACE FOR THAT. 6094 04:22:30,101 --> 04:22:32,570 MANY KIND OF BCIS WILL HAVE 6095 04:22:32,570 --> 04:22:34,038 SIMILAR PROBLEMS. 6096 04:22:34,038 --> 04:22:36,641 >> AND I GUESS WITH PERIPHERAL 6097 04:22:36,641 --> 04:22:39,143 SYSTEMS LIKE RETINAL IMPLANTS. 6098 04:22:39,143 --> 04:22:43,781 WE HAD THE ARGUS SYSTEM BEING 6099 04:22:43,781 --> 04:22:45,316 DISCONTINUED NOW PEOPLE WANT IT 6100 04:22:45,316 --> 04:22:46,951 AND CAN'T GET TO IT. 6101 04:22:46,951 --> 04:22:48,586 THERE ARE TWO SIDES OF THE 6102 04:22:48,586 --> 04:22:49,721 PROBLEM. 6103 04:22:49,721 --> 04:22:51,856 >> WITH THE RETINAL IMPLANT, 6104 04:22:51,856 --> 04:22:54,859 PEOPLE WHEN THE COMPANY WAS SOLD 6105 04:22:54,859 --> 04:22:57,595 WERE NOT ABLE TO GET UPGRADES OR 6106 04:22:57,595 --> 04:22:59,631 REPLACEMENTS FOR THEIR 6107 04:22:59,631 --> 04:23:01,432 TECHNOLOGY ANYMORE AND THEY 6108 04:23:01,432 --> 04:23:03,067 BECAME BLIND AGAIN. 6109 04:23:03,067 --> 04:23:05,370 I HEARD SOMEBODY TELL ME THE 6110 04:23:05,370 --> 04:23:08,439 SAME THING WITH A SPINAL CORD 6111 04:23:08,439 --> 04:23:12,076 STIMULATOR THAT RESTORED THEIR MOVEMENT 6112 04:23:12,076 --> 04:23:15,513 WHEN THEIR EQUIPMENT STOPPED 6113 04:23:15,513 --> 04:23:22,020 WORKING THEY BECAME PARALYZED 6114 04:23:22,020 --> 04:23:22,220 AGAIN. IT'S DEVASTATING. 6115 04:23:22,220 --> 04:23:23,888 >> I WOULD LIKE TO ASK A 6116 04:23:23,888 --> 04:23:25,790 QUESTION ON ZOOM. 6117 04:23:25,790 --> 04:23:27,725 I WOULD LIKE TO DISAGREE WITH 6118 04:23:27,725 --> 04:23:29,527 THE IDEA THAT WE FIRST NEED TO 6119 04:23:29,527 --> 04:23:31,462 UNDERSTAND THE PRINCIPLES BEFORE 6120 04:23:31,462 --> 04:23:33,598 WE UNDERSTAND THE TECHNOLOGY. 6121 04:23:33,598 --> 04:23:35,600 UNDERSTANDING BY BUILDING IS 6122 04:23:35,600 --> 04:23:38,970 STILL A VALID APPROACH. 6123 04:23:38,970 --> 04:23:40,605 UNDERSTANDING COMES FROM BITS 6124 04:23:40,605 --> 04:23:42,373 AND PIECES AND FINDING LIMITATIONS 6125 04:23:42,373 --> 04:23:43,675 AND UNDERSTANDING WHAT IS CORE. 6126 04:23:43,675 --> 04:23:46,411 THIS IS SPICY. 6127 04:23:46,411 --> 04:23:48,546 I WANTED TO GET THE PANEL'S TAKE 6128 04:23:48,546 --> 04:23:52,016 ON THAT. 6129 04:23:52,016 --> 04:23:54,185 >> SOUNDS LIKE THAT MAY BE 6130 04:23:54,185 --> 04:23:55,753 ADDRESSED TO ME. 6131 04:23:55,753 --> 04:23:59,624 I DON'T DISAGREE WITH THAT. 6132 04:23:59,624 --> 04:24:00,458 YET THERE IS A LOT OF VALUE IN 6133 04:24:00,458 --> 04:24:02,560 THE ITERATIVE BACK AND FORTH OF 6134 04:24:02,560 --> 04:24:04,362 THIS ALL. 6135 04:24:04,362 --> 04:24:07,398 AND TO BE CLEAR, I SEE A LOT OF 6136 04:24:07,398 --> 04:24:11,335 VALUE FOR NEUROAI AND 6137 04:24:11,335 --> 04:24:12,036 NEUROMORPHICS IN A LOT OF AREAS. 6138 04:24:12,036 --> 04:24:15,006 MOST OF MY COMMENTS ARE DIRECTED 6139 04:24:15,006 --> 04:24:17,408 TO NEUROTECHNOLOGY AND BRAIN IMPLANTS 6140 04:24:17,408 --> 04:24:18,009 THE HONEST ANSWER 6141 04:24:18,009 --> 04:24:19,777 IN THE WORK WE DO RIGHT NOW, IF 6142 04:24:19,777 --> 04:24:22,747 YOU GAVE ME A NEUROMORPHIC 6143 04:24:22,747 --> 04:24:24,382 DEVICE, THERE WOULD BE A LOT OF 6144 04:24:24,382 --> 04:24:26,217 TIME, ENERGY AND MONEY THAT 6145 04:24:26,217 --> 04:24:26,918 WOULD HAVE GONE INTO THE 6146 04:24:26,918 --> 04:24:29,587 DEVELOPMENT OF IT AND WE DON'T 6147 04:24:29,587 --> 04:24:31,656 KNOW WHAT TO DO WITH IT RIGHT NOW 6148 04:24:31,656 --> 04:24:34,225 TO MAKE THE LIVES OF OUR PATIENTS 6149 04:24:34,225 --> 04:24:35,093 BETTER. 6150 04:24:35,093 --> 04:24:37,595 IT'S POSSIBLE THAT THAT DEVICE 6151 04:24:37,595 --> 04:24:38,529 WOULD HAVE BEEN DESIGNED AND 6152 04:24:38,529 --> 04:24:40,732 BUILT FOR A SPECIFIC PURPOSE 6153 04:24:40,732 --> 04:24:43,167 THAT IS NOT MOST NEEDED IN THAT 6154 04:24:43,167 --> 04:24:43,701 APPLICATION. 6155 04:24:43,701 --> 04:24:45,103 I LOVE THE TECHNOLOGY 6156 04:24:45,103 --> 04:24:45,870 DEVELOPMENT. 6157 04:24:45,870 --> 04:24:48,806 I AGREE IT'S AN ITERATIVE 6158 04:24:48,806 --> 04:24:50,274 PROCESS BUT I'M TRYING TO MAKE 6159 04:24:50,274 --> 04:24:51,709 THE STATEMENT THERE IS A LOT OF 6160 04:24:51,709 --> 04:24:55,747 VALUE IN THE PARALLEL LINE OF 6161 04:24:55,747 --> 04:24:57,281 SCIENTIFIC INQUIRY TO UNDERSTAND 6162 04:24:57,281 --> 04:24:58,549 WHAT WE CAN DO WITH THE 6163 04:24:58,549 --> 04:25:00,384 TECHNOLOGY, NOT IN PLACE OF IT, 6164 04:25:00,384 --> 04:25:01,919 BUT NOT JUST TECHNOLOGY 6165 04:25:01,919 --> 04:25:04,255 DEVELOPMENT IN PLACE OF DOING 6166 04:25:04,255 --> 04:25:05,890 THE SCIENCE EITHER IS WHAT I'M 6167 04:25:05,890 --> 04:25:07,658 TRYING TO ADVOCATE HERE. 6168 04:25:07,658 --> 04:25:10,094 WE CAN END UP WITH A PLACE WHERE 6169 04:25:10,094 --> 04:25:11,529 THE TECHNOLOGY IS DESIGNED TO A 6170 04:25:11,529 --> 04:25:13,631 NEED RATHER THAN NEEDS TRYING TO 6171 04:25:13,631 --> 04:25:15,266 FIT INTO A TECHNOLOGY DESIGN 6172 04:25:15,266 --> 04:25:17,835 THAT'S BEENS SEPARATED FROM IT. 6173 04:25:17,835 --> 04:25:20,204 I DON'T WHOLESALE DISAGREE, BUT 6174 04:25:20,204 --> 04:25:23,508 I'M TRYING TO ADVOCATE THERE IS 6175 04:25:23,508 --> 04:25:24,842 SOMETHING ELSE THAT NEEDS TO BE 6176 04:25:24,842 --> 04:25:25,877 PART OF THE DISCUSSION. 6177 04:25:25,877 --> 04:25:31,849 >> I HAD A QUESTION WHICH IS 6178 04:25:31,849 --> 04:25:32,984 THE INTERESTING THING ABOUT THIS SET 6179 04:25:32,984 --> 04:25:35,653 OF TALKS IN PARTICULAR THE 6180 04:25:35,653 --> 04:25:39,991 NEUROTECH TALKS AS WELL AS THE 6181 04:25:39,991 --> 04:25:41,859 EMBODIED ROBOTICS TALKS. 6182 04:25:41,859 --> 04:25:43,961 IT BRINGS TO THE FRONT THAT WE 6183 04:25:43,961 --> 04:25:46,531 CAN DEVELOP TECHNOLOGIES IN A 6184 04:25:46,531 --> 04:25:48,833 GENERIC SENSE UP TO A POINT. 6185 04:25:48,833 --> 04:25:51,102 AT WHICH POINT YOU NEED TO 6186 04:25:51,102 --> 04:25:53,437 BUCKLE DOWN AND INVEST ON ADVANCING 6187 04:25:53,437 --> 04:25:55,173 THE TECHNOLOGY READINESS LEVEL (TRL). 6188 04:25:55,173 --> 04:25:56,908 FOR A SPECIFIC TECHNOLOGY, 6189 04:25:56,908 --> 04:25:58,576 REDUCE ALL THE RISK ASSOCIATED 6190 04:25:58,576 --> 04:26:00,678 WITH ITS USE IN A GIVEN 6191 04:26:00,678 --> 04:26:01,779 APPLICATION. 6192 04:26:01,779 --> 04:26:03,314 AT WHICH POINT IT STOPS BEING 6193 04:26:03,314 --> 04:26:05,383 USEFUL FOR OTHER THINGS OR IT MAY 6194 04:26:05,383 --> 04:26:09,220 BECOME USEFUL FOR OTHER THINGS BY 6195 04:26:09,220 --> 04:26:11,689 LUCK. AT WHAT POINT DO WE SWITCH 6196 04:26:11,689 --> 04:26:13,157 FROM DEVELOPING CONCEPTS? THIS 6197 04:26:13,157 --> 04:26:19,263 APPLIES TO NEUROAI BROADLY AND 6198 04:26:19,263 --> 04:26:20,231 NEUROMORPHICS SPECIFICALLY. 6199 04:26:20,231 --> 04:26:21,899 WE CAN DO SOMETHING GENERAL 6200 04:26:21,899 --> 04:26:25,002 WHERE WE'RE ALL LIKE I PREFER 6201 04:26:25,002 --> 04:26:27,238 THIS OR THAT BUT WE'RE ALL 6202 04:26:27,238 --> 04:26:29,340 CHASING THE SAME GOAL IN THE 6203 04:26:29,340 --> 04:26:30,208 DISTANT FUTURE. 6204 04:26:30,208 --> 04:26:33,277 AT SOME POINT YOU HAVE TO SWITCH 6205 04:26:33,277 --> 04:26:35,213 AND COMMIT TO A TECHNOLOGY AND 6206 04:26:35,213 --> 04:26:39,016 INVEST TIME AND RESOURCES TO 6207 04:26:39,016 --> 04:26:41,285 VALIDATE THAT AND IF YOU'RE NOT 6208 04:26:41,285 --> 04:26:42,720 CAREFUL, YOU BUILD A SOLUTION 6209 04:26:42,720 --> 04:26:46,224 WHICH IS CUSTOM TO THAT 6210 04:26:46,224 --> 04:26:47,358 SIX-LEGGED ROBOT AND IT'S NOT 6211 04:26:47,358 --> 04:26:48,960 GOOD FOR ANYTHING ELSE. 6212 04:26:48,960 --> 04:26:51,262 WHICH MAY NOT BE WHAT WE WANT. 6213 04:26:51,262 --> 04:26:53,231 IF YOU WAIT LONG ENOUGH, YOU 6214 04:26:53,231 --> 04:26:56,100 BUILD SOMETHING WHERE THAT IS 6215 04:26:56,100 --> 04:26:57,969 INCREDIBLY INFORMATIVE AND 6216 04:26:57,969 --> 04:26:59,370 TRANSFERS ACROSS. 6217 04:26:59,370 --> 04:27:01,672 HOW DO WE IDENTIFY THIS HAND-OFF 6218 04:27:01,672 --> 04:27:05,076 POINT OR IS IT EVEN A HAND-OFF 6219 04:27:05,076 --> 04:27:10,114 POINT, AND BASICALLY DO THIS IN 6220 04:27:10,114 --> 04:27:11,916 A -- I THINK WHAT YOU'RE 6221 04:27:11,916 --> 04:27:13,551 ADVOCATING IS EYES OPEN THAT THIS IS 6222 04:27:13,551 --> 04:27:14,685 GOING TO HAPPEN 6223 04:27:14,685 --> 04:27:17,488 AS WE MATURE THIS TECHNOLOGY AND 6224 04:27:17,488 --> 04:27:19,490 BECOME INTENTIONAL ABOUT IT AS 6225 04:27:19,490 --> 04:27:21,392 OPPOSED TO CRASHING INTO A WALL 6226 04:27:21,392 --> 04:27:23,594 AND SAYING MAYBE I SHOULDN'T 6227 04:27:23,594 --> 04:27:28,165 HAVE DONE THAT. 6228 04:27:28,165 --> 04:27:29,934 >> IF I CAN ADD TO THAT. 6229 04:27:29,934 --> 04:27:32,770 I THINK IT'S IMPORTANT TO 6230 04:27:32,770 --> 04:27:35,573 DISTINGUISH IN TERMS OF 6231 04:27:35,573 --> 04:27:36,941 NEUROMORPHIC PLATFORMS. 6232 04:27:36,941 --> 04:27:39,010 THERE IS THE NEUROMORPHIC APPROACH 6233 04:27:39,010 --> 04:27:41,178 WHICH I THINK IS WHAT KWABENA IS 6234 04:27:41,178 --> 04:27:43,414 PROPOSING FOR NEUROMORPHIC SUPREMACY 6235 04:27:43,414 --> 04:27:45,583 WHERE WE AIM FOR LARGE-SCALE GENERAL- 6236 04:27:45,583 --> 04:27:48,653 PURPOSE TYPE EVER BRAIN INSPIRED 6237 04:27:48,653 --> 04:27:49,320 COMPUTING SYSTEMS AND THEN THERE'S 6238 04:27:49,320 --> 04:27:52,790 THE NEUROMORPHIC APPROACH OF THE 6239 04:27:52,790 --> 04:27:55,493 EDGE COMPUTING, ULTRA LOW POWER, 6240 04:27:55,493 --> 04:27:56,894 VERY SPECIALIZED SENSORY PROCESSING 6241 04:27:56,894 --> 04:27:58,829 SYSTEMS. FOR THE SENSORY PROCESSING 6242 04:27:58,829 --> 04:28:01,232 SYSTEM SIDE OF THINGS, THOSE 6243 04:28:01,232 --> 04:28:03,868 WOULD BE REALLY DEVELOPED FOR A 6244 04:28:03,868 --> 04:28:04,568 SPECIFIC APPLICATION. 6245 04:28:04,568 --> 04:28:07,305 THERE IS NOTHING YET, ALL THE 6246 04:28:07,305 --> 04:28:08,973 THINGS WE HAVE AVAILABLE NOW ARE 6247 04:28:08,973 --> 04:28:09,974 RESEARCH PLATFORMS. 6248 04:28:09,974 --> 04:28:17,448 THIS IS WHAT MIKE DAVIS SAYS FOR 6249 04:28:17,448 --> 04:28:20,584 LOIHI, SAME FOR SPINNAKER 2, AND 6250 04:28:20,584 --> 04:28:22,553 CHIPS FROM HOPKINS, SAN DIEGO, ZURICH. 6251 04:28:22,553 --> 04:28:22,887 THEY ARE ALL RESEARCH PLATFORMS 6252 04:28:22,887 --> 04:28:26,991 AS SOON AS A CUSTOMER COMES WITH 6253 04:28:26,991 --> 04:28:29,060 SOME SPECIFICATIONS, PROBABLY 6254 04:28:29,060 --> 04:28:30,928 WE'D STRIP OUT MOST OF THE 6255 04:28:30,928 --> 04:28:33,698 THINGS ON THESE PLATFORM AND 6256 04:28:33,698 --> 04:28:34,999 MAKE IT SPECIFIC TO THAT 6257 04:28:34,999 --> 04:28:37,068 PARTICULAR APPLICATION, BUT NOT 6258 04:28:37,068 --> 04:28:38,235 FOR ANYTHING ELSE. IF WE WANT 6259 04:28:38,235 --> 04:28:41,472 TO REDUCE POWER CONSUMPTION, WE 6260 04:28:41,472 --> 04:28:45,710 CANNOT HAVE GENERAL-PURPOSE 6261 04:28:45,710 --> 04:28:47,144 PROGRAMMABILITY. IT'S A TRADE OFF. 6262 04:28:47,144 --> 04:28:51,849 BUT THIS APPLIES ONY TO EMBEDDED 6263 04:28:51,849 --> 04:28:53,517 SENSORY SYSTEMS, FOR EXAMPLE, BRAIN MACHINE INTERFACES 6264 04:28:53,517 --> 04:28:57,288 AND NOT TO THE MORE LARGE-SCALE 6265 04:28:57,288 --> 04:29:02,093 NEURAL PROCESSING SYSTEMS THAT 6266 04:29:02,093 --> 04:29:04,228 KWABENA ADVOCATES. 6267 04:29:04,228 --> 04:29:07,331 >> TO THAT POINT, I WAS GOING TO 6268 04:29:07,331 --> 04:29:10,101 POINT OUT THAT THIS ISN'T NEW TO 6269 04:29:10,101 --> 04:29:11,502 ANY TECHNOLOGY. 6270 04:29:11,502 --> 04:29:13,971 THERE HAS ALWAYS BEEN A VALLEY 6271 04:29:13,971 --> 04:29:18,109 OF DEATH BETWEEN -- ON THE 6272 04:29:18,109 --> 04:29:20,344 INNOVATION LIFE CYCLE. 6273 04:29:20,344 --> 04:29:26,117 THERE IS A VALLEY OF DEATH FOR 6274 04:29:26,117 --> 04:29:26,884 PRETTY MUCH ANY TECHNOLOGY. 6275 04:29:26,884 --> 04:29:28,252 ONE COULD ARGUE THAT THE 6276 04:29:28,252 --> 04:29:32,523 RESEARCH COMMUNITY AND ARGUABLY 6277 04:29:32,523 --> 04:29:33,190 GOVERNMENT FOSTERS THE 6278 04:29:33,190 --> 04:29:37,661 GENERATION OF THE IDEAS BUT IT'S 6279 04:29:37,661 --> 04:29:41,165 INDUSTRY THAT DOES THAT 6280 04:29:41,165 --> 04:29:42,800 REFINEMENT AS TO WHICH SOLUTION 6281 04:29:42,800 --> 04:29:46,637 IS GOING TO BUBBLE UP OR NOT. 6282 04:29:46,637 --> 04:29:50,207 AND THAT REFINEMENT PROCESS IS 6283 04:29:50,207 --> 04:29:51,609 REALLY -- IF YOU LOOK AT 6284 04:29:51,609 --> 04:29:53,377 HISTORY, IT'S BEEN THE SURVIVAL 6285 04:29:53,377 --> 04:29:55,046 OF THE FITTEST OF THE INDUSTRY 6286 04:29:55,046 --> 04:29:55,913 ORGANIZATIONS. 6287 04:29:55,913 --> 04:30:00,217 THOSE THAT SURVIVE AND THOSE 6288 04:30:00,217 --> 04:30:01,185 THAT DON'T. 6289 04:30:01,185 --> 04:30:03,487 PROBABLY NOT THE BEST ANSWER, 6290 04:30:03,487 --> 04:30:06,090 BUT IT IS WHAT WE SEE EVERY DAY 6291 04:30:06,090 --> 04:30:10,728 WITH DIFFERENT TECHNOLOGIES. 6292 04:30:10,728 --> 04:30:13,664 >> IN SOME WAYS AND MAYBE I'M 6293 04:30:13,664 --> 04:30:15,766 BEING OVERLY BROAD. 6294 04:30:15,766 --> 04:30:18,936 A.I. IS A NEW FIELD. 6295 04:30:18,936 --> 04:30:23,207 NEUROAI IS A NEWER FIELD AND 6296 04:30:23,207 --> 04:30:24,975 NEUROMORPHIC, THERE'S NOT LIKE AN 6297 04:30:24,975 --> 04:30:27,845 EXISTING INDUSTRY OUT THERE DOING IT. 6298 04:30:27,845 --> 04:30:28,946 I HEAR WHAT YOU'RE SAYING, IF I COME 6299 04:30:28,946 --> 04:30:31,182 UP WITH A NEW WIDGET FOR THE AUTO- 6300 04:30:31,182 --> 04:30:34,285 MOTIVES INDUSTRY, THERE'S AN EXISTING 6301 04:30:34,285 --> 04:30:41,959 STRUCTURE FOR IT. BUT THIS IS NOT OUR 6302 04:30:41,959 --> 04:30:43,694 CASE. WE HOPE THINGS START THINGS UP. 6303 04:30:43,694 --> 04:30:50,367 IS THERE A WAY TO BE MANIPULATE THIS 6304 04:30:50,367 --> 04:30:53,437 VALLEY OF DEATH FROM WHERE WE'RE 6305 04:30:53,437 --> 04:30:56,707 SITTING NOW GIVEN THAT 6306 04:30:56,707 --> 04:30:59,310 INDUSTRY HAS TO FORM 6307 04:30:59,310 --> 04:30:59,977 ORGANICALLY AS WE LOOK FORWAR? 6308 04:30:59,977 --> 04:31:01,178 >> CAN I ADD SOMETHING FURTHER 6309 04:31:01,178 --> 04:31:01,879 ON THAT. 6310 04:31:01,879 --> 04:31:03,614 ONE OF THE THINGS BROUGHT UP WAS 6311 04:31:03,614 --> 04:31:06,350 THAT WHEN WE START TO GET REAL 6312 04:31:06,350 --> 04:31:08,119 APPLICATIONS, IT HAS TO BE 6313 04:31:08,119 --> 04:31:12,056 EXTREMELY SPECIFIC IN ITS 6314 04:31:12,056 --> 04:31:12,356 APPLICATION. 6315 04:31:12,356 --> 04:31:15,493 YET, IF I WERE TO LOOK AT A 6316 04:31:15,493 --> 04:31:16,861 DESIGN PERSPECTIVE, IT'S NOT 6317 04:31:16,861 --> 04:31:18,729 WHAT WE SEE IN PRACTICE. 6318 04:31:18,729 --> 04:31:22,500 IT IS QUITE EXPENSIVE TO DO FULL 6319 04:31:22,500 --> 04:31:24,368 NEW IC DESIGN STARTS. 6320 04:31:24,368 --> 04:31:26,670 IN PRACTICE WHAT YOU SEE IS EVEN 6321 04:31:26,670 --> 04:31:28,939 ON THE DIGITAL SIDE, THERE ARE 6322 04:31:28,939 --> 04:31:30,508 MANY DIFFERENT COMPONENTS THAT 6323 04:31:30,508 --> 04:31:32,543 ARE CONFIGUREABLE OF SOME FORM AND 6324 04:31:32,543 --> 04:31:33,777 STILL EFFICIENT. 6325 04:31:33,777 --> 04:31:36,714 THERE IS A REASON WE HAVE FPGAS 6326 04:31:36,714 --> 04:31:38,682 AND NOT CUSTOM CHIPS IN THOSE SPACES. 6327 04:31:38,682 --> 04:31:41,519 THERE IS A REASON THAT WE HAVE 6328 04:31:41,519 --> 04:31:44,889 MICRO PROCESSORS EVERWHERE BECAUSE 6329 04:31:44,889 --> 04:31:48,025 EVERYONE CAN CUSTOMIZE IT WITH 6330 04:31:48,025 --> 04:31:48,859 GOOD EFFICIENCY. 6331 04:31:48,859 --> 04:31:50,027 I THINK THAT'S GOING TO BE TRUE 6332 04:31:50,027 --> 04:31:53,864 WITH A LOT OF ANALOGUE 6333 04:31:53,864 --> 04:31:54,131 COMPUTING. 6334 04:31:54,131 --> 04:31:55,833 WE'LL NEED THAT TO MAKE IT AT ALL 6335 04:31:55,833 --> 04:31:57,434 AFFORDABLE. 6336 04:31:57,434 --> 04:32:01,739 AND FURTHER -- THERE ARE ALWAYS 6337 04:32:01,739 --> 04:32:03,607 BE ONE OR TWO APPLICATIONS THAT 6338 04:32:03,607 --> 04:32:04,842 WILL BE FINE. 6339 04:32:04,842 --> 04:32:06,677 BUT YOU CAN USUALLY COUNT THOSE 6340 04:32:06,677 --> 04:32:08,479 ON A FEW FINGERS. 6341 04:32:08,479 --> 04:32:10,214 >> THAT'S WHAT I WAS ABOUT TO 6342 04:32:10,214 --> 04:32:11,182 SAY. THE MOST SUCCESSFUL 6343 04:32:11,182 --> 04:32:18,989 NEUROMORPHIC TECHNIQUE IS THE 6344 04:32:18,989 --> 04:32:19,990 LOGITECH MOUSE. IT'S VERY SPECIFIC. 6345 04:32:19,990 --> 04:32:24,328 YOU MOVE A MOUSE ON A SCREEN. 6346 04:32:24,328 --> 04:32:26,597 >> IT WAS DONE MORE THAN 25 YEARS AGO 6347 04:33:12,543 --> 04:33:15,746 SO. SO TODAY WE'VE HAD 6348 04:33:18,015 --> 04:33:20,651 A BIT INTENTIONALLY THAN YESTERDAY. 6349 04:33:20,651 --> 04:33:25,089 I THINK WE WANTED TO FOCUS ON TODAY 6350 04:33:25,422 --> 04:33:28,759 WAS THIS SORT OF, QUESTION ABOUT HOW, 6351 04:33:29,426 --> 04:33:32,029 NEUROMORPHIC TECHNOLOGY, EMBODIMENT, 6352 04:33:32,029 --> 04:33:35,032 SOME OF THE SORT OF NEUROTECH THINGS WE JUST DISCUSSED, 6353 04:33:35,532 --> 04:33:37,901 MIGHT PLAY INTO 6354 04:33:37,901 --> 04:33:40,671 THE SORT OF BIGGER NEUROAI SORT OF QUESTIONS. 6355 04:33:40,671 --> 04:33:40,871 RIGHT. 6356 04:33:40,871 --> 04:33:43,874 AND HOPEFULLY WE GOT 6357 04:33:44,408 --> 04:33:47,611 A SENSE OF BOTH THE WHERE, FOR PEOPLE 6358 04:33:47,611 --> 04:33:48,579 WHO AREN'T AS FAMILIAR WITH 6359 04:33:48,579 --> 04:33:50,014 THESE FIELDS, YOU GET A BETTER, 6360 04:33:50,014 --> 04:33:52,449 SENSE OF WHAT THESE FIELDS ARE ABOUT, 6361 04:33:52,449 --> 04:33:54,118 WHAT THE QUESTIONS ARE. 6362 04:33:54,118 --> 04:33:56,420 BUT I THINK ALSO A SENSE OF THE OPPORTUNITY. 6363 04:33:56,420 --> 04:33:59,890 I THINK RIGHT NOW WHAT WE'RE SEEING IS WE'RE CLEARLY GAINING, 6364 04:34:00,391 --> 04:34:03,394 THROUGH THE BRAIN INITIATIVE AND OTHER EFFORTS. 6365 04:34:03,627 --> 04:34:08,132 NEUROSCIENCE HAS REACHED A POINT WHERE IT IS ABLE TO OFFER 6366 04:34:09,066 --> 04:34:10,801 SYSTEM LEVEL UNDERSTANDING OF 6367 04:34:10,801 --> 04:34:13,237 WHAT'S GOING ON IN THE BRAIN. 6368 04:34:13,237 --> 04:34:14,238 BUT AT THE SAME TIME, 6369 04:34:15,806 --> 04:34:17,975 THE NEUROMORPHIC FIELD IN PARTICULAR 6370 04:34:17,975 --> 04:34:20,611 IS REACHING SCALES, WHICH IS ABLE TO ABSORB 6371 04:34:20,611 --> 04:34:22,946 THOSE SORT OF DETAILS AND COMPLEXITIES 6372 04:34:22,946 --> 04:34:24,248 OUT OF NEUROSCIENCE. 6373 04:34:24,248 --> 04:34:27,251 AND SO I THINK THIS IS PRETTY EXCITING TIME. 6374 04:34:27,251 --> 04:34:30,654 I MEAN, I THINK EVEN TWO YEARS AGO, 6375 04:34:31,221 --> 04:34:33,290 DARE I SAY, I DON'T THINK THIS WORKSHOP, 6376 04:34:33,290 --> 04:34:36,293 OR AT LEAST TODAY, WOULD HAVE BEEN AS POSSIBLE. 6377 04:34:37,094 --> 04:34:39,229 THIS IS A VERY EXCITING TIME, 6378 04:34:39,229 --> 04:34:41,332 OBVIOUSLY IN NEUROAI, BUT ALSO FOR 6379 04:34:41,332 --> 04:34:44,335 THE ROLE OF NEUROMORPHIC COMING BACK 6380 04:34:46,170 --> 04:34:46,837 INTO NEUROSCIENCE. 6381 04:34:46,837 --> 04:34:49,073 SO I'M 6382 04:34:49,073 --> 04:34:50,641 SO, YOU KNOW, I THINK, WELL, 6383 04:34:50,641 --> 04:34:52,076 YOU KNOW, THE SLIDES AREN'T SUPER IMPORTANT, 6384 04:34:52,076 --> 04:34:54,244 BUT I THINK WHAT WE REALLY WANT TO KIND OF POINT OUT 6385 04:34:54,244 --> 04:34:56,914 IS THAT THERE IS THIS SORT OF VIRTUOUS CYCLE. 6386 04:34:56,914 --> 04:34:58,015 AND I THINK IT'S, YOU KNOW, 6387 04:34:58,015 --> 04:34:59,516 WE'VE BEEN TRYING TO POPULATE IT. RIGHT. 6388 04:34:59,516 --> 04:35:02,519 I THINK WE'VE SEEN EXAMPLES OVER THE LAST, YOU KNOW, 6389 04:35:03,253 --> 04:35:06,156 STAGE WORTH OF TALKS, A COUPLE OF SESSIONS WHERE, 6390 04:35:06,156 --> 04:35:07,658 YOU KNOW, WE'RE SEEING THINGS LIKE, 6391 04:35:08,792 --> 04:35:12,096 NOT ONLY CAN WE PUT NEURAL MODELS IN THE CONNECTOME 6392 04:35:12,196 --> 04:35:15,933 ON NEUROMORPHIC HARDWARE TODAY AND THEN USE THAT IN SIMULATION, 6393 04:35:15,933 --> 04:35:19,269 BUT WE CAN TAKE SOME OF THESE ACTIVE THEORETICAL FRAMEWORKS 6394 04:35:19,269 --> 04:35:22,706 LIKE THE MANIFOLD, YOU KNOW, AND, AND THINGS LIKE THAT. 6395 04:35:22,706 --> 04:35:25,843 AND WE ACTUALLY CAN INTERROGATE A.I. MODELS, CAN, 6396 04:35:26,210 --> 04:35:28,512 YOU KNOW, INTERROGATE NEUROMORPHIC SYSTEMS 6397 04:35:28,512 --> 04:35:31,515 AND USE THESE NEUROSCIENCE TOOLS TO UNDERSTAND 6398 04:35:31,682 --> 04:35:34,318 WHAT'S GOING ON IN THE NEUROAI FIELD. 6399 04:35:34,318 --> 04:35:37,321 BUT WE CAN ALSO UNDERSTAND. 6400 04:35:38,355 --> 04:35:40,924 AND I THINK THE OTHER THE OTHER THING IS, IS THAT, 6401 04:35:40,924 --> 04:35:44,428 THE NEUROMORPHIC FIELD IS POSITIONED 6402 04:35:44,428 --> 04:35:48,298 TO GREATLY IMPROVE FROM THESE NEUROSCIENCE IDEAS. 6403 04:35:48,298 --> 04:35:51,268 RIGHT? I THINK, YOU KNOW, 6404 04:35:51,468 --> 04:35:53,470 THOSE OF US WHO'VE BEEN DOING NEUROMORPHIC FOR A WHILE 6405 04:35:53,470 --> 04:35:56,473 AND SOME FOR, FOR A VERY LONG WHILE, 6406 04:35:56,874 --> 04:35:58,208 YOU KNOW, WE'VE HAD THIS SORT OF BELIEF 6407 04:35:58,208 --> 04:35:59,943 THAT WE WERE GOING TO HIT THIS POINT 6408 04:35:59,943 --> 04:36:02,413 THAT THE ENERGY COSTS OF THE WORLD ARE GOING TO 6409 04:36:03,947 --> 04:36:06,950 OVERWHELM MUCH OF, OF, 6410 04:36:07,117 --> 04:36:09,753 WHAT'S WHAT'S GOING ON. 6411 04:36:09,753 --> 04:36:11,054 AND I THINK NEUROMORPHIC NOW 6412 04:36:11,054 --> 04:36:14,057 REACHED THE POINT WHERE WE'RE ABLE TO, TO IMPACT THAT. 6413 04:36:14,558 --> 04:36:17,161 BUT I THINK IT'S GOING TO REQUIRE THE SORT OF BRAIN, 6414 04:36:17,161 --> 04:36:19,797 YOU KNOW, REAL BRAIN INSPIRATION TO COME BACK INTO THE PROCESS 6415 04:36:19,797 --> 04:36:20,230 AND DO IT. 6416 04:36:20,230 --> 04:36:21,398 AND I THINK THIS IS, 6417 04:36:21,398 --> 04:36:24,735 WHAT I WOULD SAY IS WE WANT AS MUCH 6418 04:36:25,636 --> 04:36:29,139 INPUT FROM THE NEUROSCIENCE FIELD AS POSSIBLE. 6419 04:36:29,139 --> 04:36:32,609 AND I DO BELIEVE THE NEUROAI COMMUNITY IS BEST POSITIONED 6420 04:36:32,609 --> 04:36:35,612 TO BE THAT LIAISON. 6421 04:36:39,817 --> 04:36:42,286 OH, PERFECT. 6422 04:36:42,286 --> 04:36:43,187 THANK YOU. 6423 04:36:43,187 --> 04:36:45,956 SO ANYWAY, WE HAVE A LOOP EVERYONE LIKES THE LOOP. 6424 04:36:45,956 --> 04:36:48,258 ACTUALLY NEVER USED A LOOP BEFORE IN A SLIDE 6425 04:36:48,258 --> 04:36:50,027 BEFORE THIS MEETING, I'M GOING TO START USING IT. 6426 04:36:50,027 --> 04:36:52,162 I THINK. 6427 04:36:52,162 --> 04:36:55,766 SO ANYWAY, THE FIRST SESSION, JUST AS A BRIEF RECAP, RIGHT? 6428 04:36:55,766 --> 04:36:56,200 YOU KNOW, 6429 04:36:57,201 --> 04:36:58,836 I THINK I SAID MOST OF THIS, RIGHT. 6430 04:36:58,836 --> 04:37:02,339 I DO THINK THAT THIS IDEA OF, OF EMBODIMENT, 6431 04:37:03,106 --> 04:37:04,341 AND I THINK, YOU KNOW, WE CAN DEBATE 6432 04:37:04,341 --> 04:37:05,542 WHEN EMBODIMENT REALLY MEANS. 6433 04:37:05,542 --> 04:37:07,644 BUT I THINK THIS IDEA THAT 6434 04:37:07,644 --> 04:37:10,647 THE BRAIN DOES NOT EXIST IN ISOLATION, EVEN IN, 6435 04:37:11,548 --> 04:37:12,983 YOU KNOW, 6436 04:37:12,983 --> 04:37:16,019 NOT NECESSARILY LOOKING AT, INTERVENTIONS, 6437 04:37:16,019 --> 04:37:17,287 JUST UNDERSTANDING HOW THE BRAIN WORKS. 6438 04:37:17,287 --> 04:37:18,722 WE HAVE TO UNDERSTAND HOW IT INTERACT, 6439 04:37:18,722 --> 04:37:21,725 HOW THE BRAIN IS INTERACTING WITH THE WORLD AROUND IT. 6440 04:37:22,159 --> 04:37:24,561 AND THIS IS A WAY OF BUILDING MODEL SYSTEMS 6441 04:37:24,561 --> 04:37:25,863 IN ORDER TO UNDERSTAND THINGS. 6442 04:37:25,863 --> 04:37:29,066 AND I THINK AT THE SAME TIME, WE CAN LOOK AT HOW A.I. 6443 04:37:29,633 --> 04:37:31,969 YOU KNOW, JUST EVEN MAINSTREAM DEEP NETWORKS. 6444 04:37:31,969 --> 04:37:34,972 AND SO FORTH ARE INTERACTING WITH THE WORLD 6445 04:37:35,472 --> 04:37:37,908 AND SIMILARLY INTERROGATE THOSE USING 6446 04:37:37,908 --> 04:37:39,643 THE NEUROSCIENCE TOOLS THAT WE HAVE DEVELOPED TODAY. 6447 04:37:39,643 --> 04:37:40,744 AND I THINK THAT, 6448 04:37:40,744 --> 04:37:43,747 SOME OF THAT IMPACT WAS VERY WAS VERY IMPACTFUL. 6449 04:37:44,148 --> 04:37:44,915 IMPORTANT. 6450 04:37:44,915 --> 04:37:48,552 SO, YOU KNOW, I DO THINK THAT THE, THE, THE IMPACT, YOU KNOW, 6451 04:37:49,319 --> 04:37:51,255 IS FEEDING BACK INTO NEUROSCIENCE, 6452 04:37:51,255 --> 04:37:53,557 BUT ALSO GOING OUT AND MAKING THE WORLD A BETTER PLACE. 6453 04:37:53,557 --> 04:37:54,091 BUT ALSO. 6454 04:37:54,091 --> 04:37:56,727 GETTING INTO NEURO HEALTH AND NEUROTECHNOLOGY. 6455 04:37:56,727 --> 04:37:58,896 AND THAT'S WHERE I THINK WE'LL HAND OFF TO GINA. 6456 04:38:00,831 --> 04:38:02,199 YES. THANK YOU VERY MUCH, BRAD. 6457 04:38:02,199 --> 04:38:06,270 ACTUALLY, MY, FOLLOW UP SLIDE, IS, 6458 04:38:07,271 --> 04:38:10,107 USING I DON'T KNOW IF THIS IS WORKING AS. 6459 04:38:10,107 --> 04:38:11,708 A THE SPACE. FOR. 6460 04:38:11,708 --> 04:38:14,678 YES. FOR SESSION FOUR, 6461 04:38:14,678 --> 04:38:19,283 WE TOOK THAT, CONCEPT OF IMPACT 6462 04:38:19,583 --> 04:38:24,621 AND HOW NEUROAI CAN SUPPORT ADVANCES IN DIFFERENT 6463 04:38:25,622 --> 04:38:28,625 MEDICAL AND HEALTHCARE RELEVANT APPLICATIONS. 6464 04:38:29,192 --> 04:38:32,863 AND, ONE CONCEPT THAT WAS DISCUSSED TIME 6465 04:38:32,863 --> 04:38:36,466 AND TIME AGAIN WAS THIS IDEA OF EMBODIMENT 6466 04:38:36,466 --> 04:38:40,137 AND HOW THAT HAS TO TRANSLATE INTO END-TO-END SYSTEMS, 6467 04:38:40,537 --> 04:38:43,540 WHERE NEUROMORPHIC IS A PART OF IT. 6468 04:38:43,740 --> 04:38:46,777 IT HAS TO BE INTEGRATED WITH MULTIMODAL SENSORS. 6469 04:38:47,210 --> 04:38:51,548 IT HAS TO HAVE REALLY HIGH FAN-IN 6470 04:38:51,548 --> 04:38:54,885 AND FAN-OUT COMPUTING, BIOLOGICALLY REALISTIC 6471 04:38:56,153 --> 04:39:00,591 ACTIONS AND DECISIONS, SIMULATION IN A CLOSED LOOP. 6472 04:39:00,591 --> 04:39:03,594 SO THIS IDEA OF HAVING THE NEUROMORPHIC 6473 04:39:03,627 --> 04:39:06,296 COMPUTING AS PART OF A CLOSED LOOP, 6474 04:39:06,296 --> 04:39:09,232 IT CAME FROM, FROM THE EMBODIMENT PIECE. 6475 04:39:09,232 --> 04:39:12,669 AND ALSO TO, TO THAT EXTENT, 6476 04:39:14,104 --> 04:39:17,107 BEING ABLE TO HAVE TWINS BY, 6477 04:39:17,741 --> 04:39:22,346 OF THE BIOLOGICAL REGIONS, DIGITAL TWINS, HARDWARE TWINS 6478 04:39:22,579 --> 04:39:24,581 TO SUPPORT 6479 04:39:24,581 --> 04:39:27,284 THE, THE RESEARCH AND DEVELOPMENT IN THESE 6480 04:39:27,284 --> 04:39:30,253 END-TO-END SYSTEMS. 6481 04:39:31,321 --> 04:39:35,125 ANOTHER IMPORTANT PIECE WAS WHAT ARE THE APPLICATIONS 6482 04:39:35,192 --> 04:39:39,963 THAT, ARE, SUPPORTIVE, AND WHERE WE SEE 6483 04:39:39,963 --> 04:39:44,334 THE NEUROMORPHIC COMPUTING ADVANTAGE, WHERE IS NEUROAI, 6484 04:39:45,702 --> 04:39:50,474 BRINGING IN ITS CONTRIBUTION FOR NIH RELEVANT APPLICATIONS. 6485 04:39:50,474 --> 04:39:54,177 AND WE ENDED UP OUR DISCUSSIONS FOR SESSION FOUR WITH THAT. 6486 04:39:55,045 --> 04:39:57,948 THERE WERE A LOT OF DISCUSSION ABOUT, 6487 04:39:57,948 --> 04:40:00,951 DIAGNOSTICS AND NEURAL INTERFACES, 6488 04:40:01,051 --> 04:40:05,188 THE DIFFERENT BIO SIGNAL ANALYSIS, ROBOTICS, ETC., 6489 04:40:05,188 --> 04:40:09,660 BUT THAT COMES WITH DATA SETS AND SUPPORTING 6490 04:40:10,661 --> 04:40:14,998 DATA THAT IS ABLE TO, TO PROVIDE 6491 04:40:14,998 --> 04:40:18,035 RESEARCHERS AND ENGINEERS WITH 6492 04:40:19,569 --> 04:40:23,707 METRICS AND WAYS OF ACTUALLY DEVELOPING THESE TECHNOLOGIES. 6493 04:40:24,207 --> 04:40:28,412 AND, IN TERMS OF METRICS, A LOT WAS DISCUSSED, OF COURSE, 6494 04:40:29,079 --> 04:40:31,581 ENERGY EFFICIENCY IS A BIG PIECE, 6495 04:40:31,581 --> 04:40:35,585 BUT ALSO THE IDEA OF PERFORMANCE IN TERMS OF ACCURACIES, 6496 04:40:36,153 --> 04:40:39,523 REPROGRAMMABILITY, VALIDATION, SAFETY, 6497 04:40:39,523 --> 04:40:42,859 CLINICAL BENEFITS, USABILITY, RELIABILITY. 6498 04:40:42,859 --> 04:40:45,162 WHAT'S THE TECHNOLOGICAL MATURITY? 6499 04:40:46,096 --> 04:40:50,767 AND THE INTERPRETABILITY AND THE EXPLAINABILITY OF THE MODELS? 6500 04:40:51,668 --> 04:40:53,904 AND, WE HAD A LIVELY DISCUSSION 6501 04:40:53,904 --> 04:40:57,341 ABOUT THE SAFETY AND ETHICS OF, OF THIS, 6502 04:40:58,942 --> 04:41:00,877 AND, THE, 6503 04:41:00,877 --> 04:41:05,549 THE LAST PIECE OF IT WAS, WHAT WOULD BE NEXT? 6504 04:41:05,549 --> 04:41:07,584 WHAT SHOULD COME NEXT? 6505 04:41:07,584 --> 04:41:10,587 AND, THE DISCUSSION 6506 04:41:10,787 --> 04:41:14,725 ALTERNATED AND PROPOSES TWO COMPLEMENTARY 6507 04:41:15,359 --> 04:41:17,027 RESEARCH PROGRAMS. 6508 04:41:17,027 --> 04:41:18,862 ONE MORE ON THE FOUNDATIONAL PIECE 6509 04:41:18,862 --> 04:41:22,766 OF IDENTIFYING WHAT WOULD BE THE POTENTIAL KEY 6510 04:41:23,633 --> 04:41:26,203 APPLICATIONS FOR NEUROAI. 6511 04:41:26,203 --> 04:41:29,706 AND, THE OTHER ONE WAS ON TECHNOLOGY DEVELOPMENT 6512 04:41:29,706 --> 04:41:34,077 ON SUPPORTING, THESE DEVELOPMENT OF NEUROAI SOLUTIONS. 6513 04:41:34,745 --> 04:41:38,615 SO, WITH THIS, I'M LOOKING FORWARD TO THE WRAP UP 6514 04:41:38,615 --> 04:41:41,718 FOR THE ENTIRE WORKSHOP AND TO SEE THE BIG VISION 6515 04:41:41,718 --> 04:41:44,721 FOR THE FUTURE. 6516 04:41:51,928 --> 04:41:55,799 I'M GOING TO HAND IT OFF TO PAUL MIDDLEBROOKS. 6517 04:41:56,800 --> 04:41:58,435 ALL RIGHT, SO, 6518 04:41:58,435 --> 04:42:02,639 WE'RE GOING TO HAVE A DISCUSSION TO WRAP UP THIS SESSION BEFORE 6519 04:42:02,639 --> 04:42:05,642 JOHN COMES OUT AND REALLY WRAPS IT UP. 6520 04:42:07,477 --> 04:42:09,212 MAYBE ONE WAY I'LL OPEN 6521 04:42:09,212 --> 04:42:12,215 THIS DISCUSSION UP IS BY NOTING THAT, 6522 04:42:14,718 --> 04:42:17,721 THERE WASN'T A TON OF THEORY, 6523 04:42:18,288 --> 04:42:20,557 IN ALL OF THE TALKS. 6524 04:42:20,557 --> 04:42:23,627 AND SO THEY'RE SORT OF TWO QUESTIONS. 6525 04:42:23,627 --> 04:42:25,996 ONE IS, 6526 04:42:25,996 --> 04:42:28,965 SPEAKING TO WHAT DORIS YESTERDAY NOTED 6527 04:42:29,599 --> 04:42:33,503 WHEN SHE SAID, DON'T DISCOUNT OUR INTUITIONS AS BIOLOGISTS 6528 04:42:33,503 --> 04:42:36,506 AS WE ARE DOING RESEARCH. 6529 04:42:39,009 --> 04:42:41,278 WITH NEUROAI AND USING ARTIFICIAL 6530 04:42:41,278 --> 04:42:45,449 INTELLIGENCE MODELS, THERE'S A SLIGHT RISK. 6531 04:42:45,849 --> 04:42:47,517 YOU KNOW, THESE DATA DRIVEN APPROACHES, 6532 04:42:47,517 --> 04:42:48,351 THERE'S A SLIGHT RISK 6533 04:42:48,351 --> 04:42:51,655 THAT WE LOSE SOME OF THAT CREATIVITY AND INTUITION, 6534 04:42:51,922 --> 04:42:53,356 AND WE'RE JUST LISTENING TO THE MODELS. 6535 04:42:53,356 --> 04:42:57,027 SO ONE QUESTION IS HOW DO WE MAINTAIN THAT INTUITION 6536 04:42:57,594 --> 04:42:59,796 AS WE PUSH FORWARD IN NEUROAI? 6537 04:43:12,309 --> 04:43:14,010 I WOULD BE HAPPY TO RESPOND 6538 04:43:14,010 --> 04:43:17,013 TO THAT BY SORT OF EXTENDING THE QUESTION, PERHAPS. 6539 04:43:17,781 --> 04:43:20,784 SO YESTERDAY, BLAKE 6540 04:43:21,351 --> 04:43:23,987 POSED 6541 04:43:23,987 --> 04:43:26,890 A SORT OF DICHOTOMY OF APPROACHES, ONE OF WHICH 6542 04:43:26,890 --> 04:43:31,294 WAS APPLYING THEORETICAL IDEAS, AS YOU'RE SAYING, 6543 04:43:31,895 --> 04:43:34,898 AND THE OTHER ONE WAS TAKING SORT OF A, 6544 04:43:35,799 --> 04:43:39,970 BOTTOM UP APPLYING DATA, TRYING TO MAKE DO 6545 04:43:39,970 --> 04:43:42,973 WITH WHATEVER WE HAVE JUST TO BUILD THINGS THAT WORK. 6546 04:43:43,073 --> 04:43:44,975 AND I THINK TALKING ABOUT, 6547 04:43:44,975 --> 04:43:47,744 AND MAYBE I COULD JUST SUGGEST TALKING ABOUT THAT TENSION 6548 04:43:47,744 --> 04:43:48,879 AND HOW IT INTERSECTS 6549 04:43:48,879 --> 04:43:51,882 THE THINGS WE'VE BEEN SEEING OVER THE PAST FEW DAYS. 6550 04:43:56,353 --> 04:44:00,891 I THINK THIS YOU'RE POSING THIS DICHOTOMY IN BETWEEN, 6551 04:44:00,891 --> 04:44:03,894 YOU KNOW, BIOLOGY AND THEORY. 6552 04:44:05,662 --> 04:44:09,032 WHAT WE DO, FOR EXAMPLE, IN, IN THE, YOU KNOW, 6553 04:44:09,032 --> 04:44:12,068 NEUROMORPHIC WORKSHOPS, IN TELLURIDE, 6554 04:44:12,068 --> 04:44:16,106 IN CAPO CACCIA AND OTHERS, A NEW ONE IN, IN BANGALORE. 6555 04:44:17,340 --> 04:44:19,776 WE KEEP INVITING BIOLOGISTS, 6556 04:44:19,776 --> 04:44:23,380 THAT DESCRIBE WHAT HAPPENS, WHAT THEY OBSERVE, 6557 04:44:24,714 --> 04:44:27,517 WHAT I THINK IS 6558 04:44:27,517 --> 04:44:30,020 NOT MISSING, BUT 6559 04:44:30,020 --> 04:44:33,023 NEEDED IS TO HAVE 6560 04:44:34,558 --> 04:44:37,427 A SMOOTHER TRANSITION BETWEEN WHAT 6561 04:44:37,427 --> 04:44:40,463 A BIOLOGIST OR A NEUROSCIENTIST OBSERVES 6562 04:44:41,097 --> 04:44:44,100 AND DESCRIBES, 6563 04:44:45,135 --> 04:44:46,770 WITH THEORIES, 6564 04:44:46,770 --> 04:44:49,739 MATHEMATICS, COMPUTATIONAL NEUROSCIENCE 6565 04:44:49,739 --> 04:44:52,742 THAT TRANSLATE THAT INTO SOMETHING 6566 04:44:52,742 --> 04:44:55,745 THAT WE CAN IMPLEMENT IN NEUROMORPHIC TECHNOLOGY. 6567 04:44:56,579 --> 04:44:58,581 WE ARE TRYING TO DO THIS EXERCISE. 6568 04:44:58,581 --> 04:45:01,384 SOMETIMES WE DON'T HAVE THE RIGHT BACKGROUND TO DO SO. 6569 04:45:01,384 --> 04:45:02,652 SO IT'S THAT'S WHY 6570 04:45:02,652 --> 04:45:06,489 IT'S A REALLY MULTIDISCIPLINARY TYPE OF COMMUNITY. 6571 04:45:06,790 --> 04:45:08,758 SO, SO IF YOU HAD I THINK WE CAN COME BACK 6572 04:45:08,758 --> 04:45:10,894 TO THE SAME TOPIC, BUT JUST JUMPING OFF OF THAT. 6573 04:45:10,894 --> 04:45:12,929 IF YOU HAD TEN YEARS 6574 04:45:12,929 --> 04:45:15,999 HOW WOULD YOU DESIGN A PROGRAM TO GET WHAT 6575 04:45:15,999 --> 04:45:19,002 YOU NEED? 6576 04:45:24,441 --> 04:45:27,444 IT'S. 6577 04:45:30,747 --> 04:45:33,750 YOU MEAN DESIGN A PROGRAM IN TERMS OF. 6578 04:45:34,150 --> 04:45:39,489 IMPLEMENTING, YOU KNOW, INFRASTRUCTURE, PEOPLE TRAINING? 6579 04:45:40,657 --> 04:45:42,759 YOU WERE TALKING ABOUT WHAT YOU NEED. 6580 04:45:42,759 --> 04:45:46,363 SO HOW I MEAN, WHAT WE, YOU KNOW, HERE AT THE NIH, RIGHT? 6581 04:45:46,529 --> 04:45:48,798 WE NEED TO TALK ABOUT IMPLEMENTATION. 6582 04:45:48,798 --> 04:45:51,801 SO HOW WOULD THIS BE IMPLEMENTED? 6583 04:45:52,802 --> 04:45:53,636 ONE FIRST STEP 6584 04:45:53,636 --> 04:45:57,941 IS THESE WORKSHOPS WHERE WE HAVE PEOPLE COMING TOGETHER. 6585 04:45:57,941 --> 04:46:00,877 AND I THINK YESTERDAY IT WAS MENTIONED, 6586 04:46:01,378 --> 04:46:04,381 WE DON'T NEED TO HAVE PEOPLE TRAINED IN EVERYTHING, 6587 04:46:04,381 --> 04:46:06,182 BUT WE NEED TO HAVE PEOPLE 6588 04:46:06,182 --> 04:46:09,452 THAT UNDERSTAND EACH OTHER THAT USE THE SAME VOCABULARY. 6589 04:46:09,452 --> 04:46:12,055 AND THAT'S NOT THE CASE YET. 6590 04:46:12,055 --> 04:46:14,491 WHEN DIFFERENT PERSONS 6591 04:46:14,491 --> 04:46:17,494 FROM DIFFERENT COMMUNITIES TALK ABOUT, 6592 04:46:18,328 --> 04:46:21,698 USING CERTAIN KEY WORDS, THAT KEYWORD HAS DIFFERENT MEANINGS 6593 04:46:21,731 --> 04:46:23,199 IN DIFFERENT COMMUNITIES. 6594 04:46:23,199 --> 04:46:26,870 AND, AND SO THAT'S WHY IT'S IMPORTANT. 6595 04:46:27,837 --> 04:46:31,041 TO CREATE AN INFRASTRUCTURE WHERE WE CAN EXCHANGE IDEAS 6596 04:46:31,041 --> 04:46:34,044 AND COMPETENCIES. AND THAT, 6597 04:46:34,310 --> 04:46:37,881 KIND OF FACILITATES THE SYNERGIES. 6598 04:46:40,550 --> 04:46:41,217 I'LL AMPLIFY 6599 04:46:41,217 --> 04:46:42,686 THAT A LITTLE BIT, AND THEN I'LL GO BACK 6600 04:46:42,686 --> 04:46:45,155 TO THE ORIGINAL QUESTION FOR A MOMENT. 6601 04:46:45,155 --> 04:46:46,356 I THINK IT 6602 04:46:46,356 --> 04:46:48,992 I WOULD ACTUALLY SAY, YES, WE DEFINITELY NEED TO HAVE WAYS 6603 04:46:48,992 --> 04:46:52,328 TO GET WITHIN A PROGRAM HOW PEOPLE COMMUNICATE. 6604 04:46:52,929 --> 04:46:55,598 AND IF IT WERE ME, I WOULD DEFINITELY TRY TO HAVE A PROGRAM 6605 04:46:55,598 --> 04:46:57,934 WHERE YOU HAD AT LEAST A COUPLE CORE TEAMS 6606 04:46:57,934 --> 04:47:01,004 THAT WERE FAIRLY BROAD THAT WOULD THEN ALL BE 6607 04:47:01,004 --> 04:47:04,374 COMMUNICATING TOGETHER, PROBABLY ALL INTERFACING 6608 04:47:04,374 --> 04:47:07,510 WITHIN THE SAME SORT OF HIGH- LEVEL FRAMEWORK AND TOOLS. 6609 04:47:08,878 --> 04:47:11,181 I THINK THAT A COUPLE TIMES THIS HAS BEEN DONE IN THE PAST 6610 04:47:11,181 --> 04:47:13,583 HAS BEEN INCREDIBLY EFFECTIVE. 6611 04:47:13,583 --> 04:47:16,386 BECAUSE IT'S NOT JUST EDUCATION ACROSS THE BOUNDARIES, 6612 04:47:16,386 --> 04:47:19,522 BUT EVERYBODY USING THAT CAPABILITY. 6613 04:47:19,522 --> 04:47:21,124 ALL THE WAY ACROSS. 6614 04:47:21,124 --> 04:47:23,193 AND SO YOU'D WANT TO HAVE A COUPLE BROAD TEAMS TO 6615 04:47:23,193 --> 04:47:24,060 BE ABLE TO DO THAT 6616 04:47:25,762 --> 04:47:26,930 SO I COULD IMMEDIATELY SEE 6617 04:47:26,930 --> 04:47:28,465 THAT WHEN WE BUILD A TEN YEAR PROGRAM 6618 04:47:28,465 --> 04:47:30,767 TO DO SOMETHING ALONG THOSE LINES. 6619 04:47:30,767 --> 04:47:33,770 THE OTHER COMMENT YOU MENTIONED, WHICH WAS, 6620 04:47:34,571 --> 04:47:35,905 YOU KNOW, THE QUESTION OF THEORY. 6621 04:47:35,905 --> 04:47:36,806 I MEAN, I THINK 6622 04:47:36,806 --> 04:47:39,809 CERTAINLY THEORY MATTERS, BUT I THINK WHERE YOU'RE 6623 04:47:39,876 --> 04:47:42,545 GOING WITH THAT IS, YOU KNOW, IT WAS THE DATA DRIVEN, 6624 04:47:42,545 --> 04:47:46,382 DRIVEN APPROACH VERSUS WHAT CAN WE FIGURE OUT DIRECTLY. 6625 04:47:46,382 --> 04:47:49,385 AND I WOULD SAY ANYTIME WE TRY TO BUILD SOMETHING. 6626 04:47:49,753 --> 04:47:51,821 AT LEAST I, FOR MY SENSE OF IT, 6627 04:47:51,821 --> 04:47:53,823 ANYTHING THAT I CAN DEFINE EXPLICITLY 6628 04:47:53,823 --> 04:47:56,759 RIGHT FROM THE BEGINNING. I WOULD USE THAT. 6629 04:47:56,759 --> 04:47:59,395 YOU KNOW, I DON'T WANT TO MAKE IT A COMPLETE BLACK BOX IF I 6630 04:47:59,395 --> 04:48:00,563 DON'T NEED TO DO THAT. 6631 04:48:00,563 --> 04:48:04,134 I ONLY PUT MORE BLACK BOXES WHEN I NEED TO. 6632 04:48:04,434 --> 04:48:07,003 GOOD EXAMPLE WOULD BE IF I'M DOING NEUROMORPHIC 6633 04:48:07,003 --> 04:48:07,470 STRUCTURES 6634 04:48:07,470 --> 04:48:09,339 AROUND, SAY, SPEECH RECOGNITION, 6635 04:48:09,339 --> 04:48:11,174 I ALREADY KNOW THE FRONT END VERY WELL. 6636 04:48:11,174 --> 04:48:14,144 THE MODEL FROM IS RELATED TO THE HUMAN COCHLEA. 6637 04:48:14,477 --> 04:48:15,745 I ALREADY KNOW WHAT THAT LOOKS LIKE. 6638 04:48:15,745 --> 04:48:18,348 I ALREADY KNOW WHAT THE FIRST COUPLE LAYERS LOOK LIKE. 6639 04:48:18,348 --> 04:48:21,751 I GO BUILD THIS AND THEN I GO ON TOP OF THAT. 6640 04:48:22,085 --> 04:48:24,521 AND THAT'S USEFUL BECAUSE THE MORE THIS 6641 04:48:24,521 --> 04:48:26,790 IF I'M BUILDING A TRUE EMBEDDED STRUCTURE, I 6642 04:48:26,790 --> 04:48:26,990 WANT TO 6643 04:48:26,990 --> 04:48:29,793 SIMPLIFY THE BLACK BOX AS MUCH AS I CAN 6644 04:48:29,793 --> 04:48:31,027 OR ALL SORTS OF OTHER REASONS. 6645 04:48:34,597 --> 04:48:37,600 WHICH I. 6646 04:48:39,802 --> 04:48:42,772 IMAGINE, PATRICK, YOU HAVE SO YOU HAVE INDUSTRY EXPERIENCE. 6647 04:48:42,772 --> 04:48:44,307 AND I THINK OF THESE THINGS IN TERMS 6648 04:48:44,307 --> 04:48:47,944 OF, OF FOR EXAMPLE, I ASKED MY PI 6649 04:48:48,044 --> 04:48:49,179 YESTERDAY 6650 04:48:49,179 --> 04:48:52,015 VIA TEXT WHAT TOOL WOULD YOU NEED? 6651 04:48:52,015 --> 04:48:53,716 AND HE SAID A SOFTWARE ENGINEER 6652 04:48:53,716 --> 04:48:55,185 AND I COMPLETELY AGREE WITH HIM. 6653 04:48:55,185 --> 04:48:58,188 AND I IMAGINE YOU MIGHT HAVE THOUGHTS ABOUT THAT. 6654 04:48:58,555 --> 04:49:01,558 YES. I AM A REFORMED SOFTWARE ENGINEER. 6655 04:49:01,658 --> 04:49:04,894 I DID DO THAT FOR A LITTLE BIT THOSE, INTERESTING. 6656 04:49:04,894 --> 04:49:06,996 SO RIGHT AFTER MY POSTDOC, 6657 04:49:06,996 --> 04:49:09,832 I GOT A JOB AS A SOFTWARE ENGINEER AT GOOGLE, 6658 04:49:09,832 --> 04:49:12,035 AND DID THAT FOR A LITTLE BIT UNTIL THEY REALIZED 6659 04:49:12,035 --> 04:49:14,804 THAT THEY SHOULD REALLY, YOU KNOW, TAKE ME OFF 6660 04:49:14,804 --> 04:49:17,807 OF THE REAL MACHINES THAT DO THE REAL THINGS. 6661 04:49:18,474 --> 04:49:21,678 BUT IN ALL SERIOUSNESS, IT WAS I THINK I REALLY, 6662 04:49:23,213 --> 04:49:25,748 ENLIGHTENING EXPERIENCE 6663 04:49:25,748 --> 04:49:29,085 AND SEEING WHAT WAS THE SHIFT 6664 04:49:29,085 --> 04:49:31,154 IN PARADIGMS FOR DOING ENGINEERING, 6665 04:49:31,154 --> 04:49:33,489 ESPECIALLY FOR SOFTWARE ENGINEERING. 6666 04:49:33,489 --> 04:49:36,860 AND SO I THINK AS WE'RE CONTEMPLATING THIS WORLD 6667 04:49:36,860 --> 04:49:39,229 IN WHICH MAYBE WE HAVE LARGER TEAMS, 6668 04:49:39,229 --> 04:49:41,064 WE HAVE CENTRALIZED ENGINEERING, 6669 04:49:41,064 --> 04:49:43,733 MAYBE WE HAVE LIKE A HUB- AND-SPOKES MODEL IN WHICH WE DO, 6670 04:49:43,733 --> 04:49:45,235 YOU KNOW, CENTRALIZED RESEARCH. 6671 04:49:45,235 --> 04:49:48,271 AND MANY PEOPLE ARE BROUGHT IN TO DO THIS. 6672 04:49:48,271 --> 04:49:50,707 AND WE'LL NEED TO EMBRACE, 6673 04:49:50,707 --> 04:49:54,844 THAT KIND OF FRAMEWORK FOR RESEARCH, SOFTWARE ENGINEERING 6674 04:49:55,411 --> 04:49:56,346 THAT IS MORE ROBUST 6675 04:49:56,346 --> 04:49:59,349 THAN THE KINDS OF APPROACHES THAT WE USE NOW, BECAUSE 6676 04:50:00,516 --> 04:50:04,454 IF WE WANT TO FACILITATE BOTH THE DISCOVERY SCIENCE, 6677 04:50:04,721 --> 04:50:06,889 AS WELL AS THE ENGINEERING THAT'S NECESSARY, 6678 04:50:06,889 --> 04:50:09,092 WE'LL NEED TO BRING THESE DISCIPLINES TOGETHER. 6679 04:50:09,092 --> 04:50:11,127 BUT RIGHT NOW, IT IS HARD. 6680 04:50:11,127 --> 04:50:13,630 IT IS CHALLENGING FOR PEOPLE IN LABS TO HIRE 6681 04:50:13,630 --> 04:50:16,466 GOOD RESEARCH SOFTWARE ENGINEERS FOR INSTANCE. 6682 04:50:18,401 --> 04:50:19,736 WHY IS IT ALWAYS A CHALLENGE? 6683 04:50:19,736 --> 04:50:22,739 BECAUSE THEY CAN MAKE MORE MONEY ELSEWHERE? 6684 04:50:22,739 --> 04:50:23,606 THAT'S A BIG REASON. 6685 04:50:23,606 --> 04:50:25,541 YEAH, ABSOLUTELY. 6686 04:50:25,541 --> 04:50:28,077 IT'S ALSO, 6687 04:50:28,077 --> 04:50:30,880 IT IS ALSO LIKE A CAREER TRAP. 6688 04:50:30,880 --> 04:50:34,284 IF SOMEONE HAS THE POSSIBILITY OF GOING INTO INDUSTRY RIGHT NOW 6689 04:50:34,784 --> 04:50:39,055 AND CLIMBING UP THE LADDER AS SOFTWARE ENGINEER, YOU KNOW, 6690 04:50:39,055 --> 04:50:42,058 THEY CAN, THEY HAVE LOTS OF POSSIBILITIES FOR GROWTH. 6691 04:50:42,458 --> 04:50:43,793 IF YOU ARE 6692 04:50:43,793 --> 04:50:46,896 INSIDE OF A RESEARCH LAB AND YOU'RE A SOFTWARE ENGINEER, 6693 04:50:47,630 --> 04:50:50,733 MUCH HARDER TO FIGURE OUT WAYS 6694 04:50:50,733 --> 04:50:53,936 TO GET, PEER REVIEW, 6695 04:50:55,271 --> 04:50:58,274 TO HAVE PEER MENTORSHIP. 6696 04:50:58,308 --> 04:51:00,143 YOU DON'T HAVE NECESSARILY POSSIBILITIES 6697 04:51:00,143 --> 04:51:01,678 TO GROW WITHIN THE ORGANIZATION. 6698 04:51:01,678 --> 04:51:03,446 OFTENTIMES THERE WILL BE 1 OR 2 RESEARCH 6699 04:51:03,446 --> 04:51:07,517 SOFTWARE ENGINEERS WITHIN A SILO, WHICH IS THE PI 6700 04:51:07,517 --> 04:51:08,151 IS THE SILO. 6701 04:51:08,151 --> 04:51:10,820 AND, IN THIS CASE, SO WHAT ARE YOUR OPTIONS 6702 04:51:10,820 --> 04:51:11,921 FOR ADVANCEMENTS? 6703 04:51:11,921 --> 04:51:13,289 YOU'LL MAKE ABOUT A QUARTER OF THE SALARY 6704 04:51:13,289 --> 04:51:14,924 THAT YOU WOULD MAKE AN INDUSTRY. 6705 04:51:14,924 --> 04:51:19,996 AND SO YEAH, I MEAN, IT'S, IT IS OFTEN AND I SHOULD SAY THAT, 6706 04:51:21,831 --> 04:51:23,466 IT OFTEN TIMES IT'S NOT EVEN 6707 04:51:23,466 --> 04:51:26,469 POSSIBLE TO FUND THESE POSITIONS. 6708 04:51:27,670 --> 04:51:30,173 SO, SO YEAH, THERE'S A FEW PLACES 6709 04:51:30,173 --> 04:51:33,876 THAT HAVE REALLY REALIZED THE ROLE OF CENTRALIZED ENGINEERING. 6710 04:51:34,777 --> 04:51:37,647 IT'S JUST TO NAME A FEW 6711 04:51:37,647 --> 04:51:40,650 I THINK HHMI JANELIA IS ONE PLACE THAT HAVE EMBRACED THIS 6712 04:51:41,050 --> 04:51:43,119 AT THE ALLEN INSTITUTE, 6713 04:51:43,119 --> 04:51:44,287 SOME OF THE NEWER INSTITUTES 6714 04:51:44,287 --> 04:51:45,955 IN CHARLOTTESVILLE COMING ONLINE THAT ARE TRYING 6715 04:51:45,955 --> 04:51:48,157 TO DO A NEW SCIENCE AT THE ART INSTITUTE 6716 04:51:48,157 --> 04:51:51,160 ARE REALLY TRYING TO EMBRACE THIS, BUT OTHERWISE IT'S, 6717 04:51:52,161 --> 04:51:55,398 WE'RE LIMITED BY, YOU KNOW, THIS BOTTLENECK WHERE, 6718 04:51:56,165 --> 04:51:58,634 ESSENTIALLY IT'S GRAD STUDENTS THAT ARE DOING THAT WORK 6719 04:51:58,634 --> 04:51:59,802 OF RESEARCH, SOFTWARE ENGINEERING. 6720 04:51:59,802 --> 04:52:01,571 AND, YOU KNOW, MAYBE SOME OF THAT WORK 6721 04:52:01,571 --> 04:52:04,574 IS NOT BEST DONE BY GRAD STUDENTS. 6722 04:52:07,176 --> 04:52:09,212 YEAH, I AGREE WITH THAT. 6723 04:52:09,212 --> 04:52:13,182 BUT SO IS THE SOLUTION JUST TO CENTRALIZE EVERYTHING 6724 04:52:13,182 --> 04:52:14,417 IN INSTITUTIONS THEN? 6725 04:52:18,855 --> 04:52:21,357 NO. 6726 04:52:21,357 --> 04:52:24,427 IS THE SOLUTION TO CENTRALIZE EVERYTHING 6727 04:52:24,427 --> 04:52:25,328 IN INSTITUTIONS, 6728 04:52:25,328 --> 04:52:27,930 LIKE THE EXAMPLES THAT PATRICK WAS GIVING 6729 04:52:27,930 --> 04:52:30,032 ABOUT HOW ESSENTIALLY IT'S 6730 04:52:30,032 --> 04:52:33,036 A MANAGEMENT PROBLEM, RIGHT? 6731 04:52:36,939 --> 04:52:41,544 I. YEAH, I MEAN, I THINK THERE'S REALLY HARD 6732 04:52:41,544 --> 04:52:46,048 INSTITUTIONAL DESIGN QUESTIONS THERE ABOUT WHO IS THE SOFTWARE 6733 04:52:46,048 --> 04:52:49,452 ENGINEERING GROUP AND WHO ARE THEY REPORTING TO, 6734 04:52:49,452 --> 04:52:52,555 AND HOW CLOSELY ARE THEY INTEGRATED WITH THE SCIENTISTS. 6735 04:52:53,456 --> 04:52:55,792 YOU KNOW, THAT ALSO INTERSECT WHAT PATRICK IS SAYING 6736 04:52:55,792 --> 04:52:57,927 ABOUT CAREER PATH AND MENTORSHIP. 6737 04:52:57,927 --> 04:53:01,431 AND SO I DON'T HAVE ANY REALLY GOOD ANSWERS, 6738 04:53:01,431 --> 04:53:04,434 BUT I THINK THE TECH FIRMS HAVE THOUGHT ABOUT THIS, 6739 04:53:05,268 --> 04:53:07,770 ABOUT HOW, YOU KNOW, WHO'S A RESEARCH SOFTWARE 6740 04:53:07,770 --> 04:53:11,474 ENGINEER, WHO'S A DEVOPS PERSON, WHO IS A, YOU KNOW, 6741 04:53:12,241 --> 04:53:15,211 A TESTING PERSON. 6742 04:53:15,411 --> 04:53:16,612 HOW DO YOU SCALE UP? 6743 04:53:16,612 --> 04:53:18,581 WHO HOW DO THEY PARTNER WITH, 6744 04:53:18,581 --> 04:53:21,584 YOU KNOW, DATA SCIENTIST DOING THE, PROTOTYPING? 6745 04:53:21,584 --> 04:53:23,486 AND I THINK IT'S, YOU KNOW, I DON'T HAVE ANY ANSWERS, 6746 04:53:23,486 --> 04:53:25,488 BUT I THINK IT'S WORTH THINKING ABOUT THOSE ISSUES. 6747 04:53:25,488 --> 04:53:26,422 THEY'RE REALLY IMPORTANT. 6748 04:53:29,158 --> 04:53:32,095 WE CONTINUE. 6749 04:53:32,095 --> 04:53:34,630 I'M STILL STUCK ON YOUR INTUITION QUESTION, 6750 04:53:34,630 --> 04:53:37,967 BUT I THINK WE'VE KIND OF GONE IN ANOTHER -- AN INTERESTING PATH. 6751 04:53:37,967 --> 04:53:40,002 BUT I HEARD IT 6752 04:53:40,002 --> 04:53:42,472 MAYBE FROM A DIFFERENT PERSPECTIVE. 6753 04:53:42,472 --> 04:53:45,608 WHAT DORIS WAS SAYING, I MEAN, IT WAS KIND OF TO ME, 6754 04:53:45,908 --> 04:53:48,077 IF SHE'S HERE, SHE CAN TELL US WHAT SHE MOST MEANT. 6755 04:53:48,077 --> 04:53:51,080 BUT I WAS THINKING THAT 6756 04:53:51,147 --> 04:53:54,150 IT'S ACTUALLY KIND OF, MAINTAINING A CERTAIN 6757 04:53:54,951 --> 04:53:58,287 HUMILITY AND CURIOSITY ABOUT YOUR KNOWLEDGE SET. 6758 04:53:58,287 --> 04:54:00,256 AND I THINK YOU HAD ASKED THIS QUESTION IN THE BEGINNING. 6759 04:54:00,256 --> 04:54:01,090 YOU HAD SAID, 6760 04:54:01,090 --> 04:54:02,158 WHAT DO OUR TRAINEES NEED? 6761 04:54:02,158 --> 04:54:03,526 AND SOMEBODY SAID, CURIOSITY. 6762 04:54:03,526 --> 04:54:06,295 AND YOU'RE LIKE, WELL, YOU NEED CURIOSITY FOR EVERYTHING. 6763 04:54:06,295 --> 04:54:07,597 AND I WAS LIKE, YOU KNOW, I'M 6764 04:54:07,597 --> 04:54:10,633 THINKING, ACTUALLY, I'VE BEEN HEARING AS AN OUTSIDER 6765 04:54:10,833 --> 04:54:14,537 AND I HAVE FELT SO LUCKY TO THOSE OF YOU WHO'VE SHARED 6766 04:54:14,537 --> 04:54:16,138 WITH ME YOUR EXPERIENCES, 6767 04:54:17,440 --> 04:54:18,040 YOU KNOW, I'VE HEARD A 6768 04:54:18,040 --> 04:54:21,043 LOT OF CULTURAL CONFLICT ACROSS DIFFERENT DISCIPLINES. 6769 04:54:21,043 --> 04:54:23,946 I LOVED TALKING TO THE STUDENTS AT THE POSTERS. 6770 04:54:23,946 --> 04:54:26,949 IF ANY OF YOU ARE HERE, THAT WAS REALLY MY HIGHLIGHT. 6771 04:54:26,949 --> 04:54:29,485 BECAUSE I ASKED THEM ABOUT THEIR EXPERIENCES 6772 04:54:29,485 --> 04:54:33,189 ON WHAT IT WAS TO BE IN THIS INTERSECTION OF A SPACE. 6773 04:54:33,689 --> 04:54:35,124 HOW LONELY THAT IS. 6774 04:54:35,124 --> 04:54:37,527 ALSO, AND HOW, 6775 04:54:37,527 --> 04:54:40,163 YOU KNOW, THEY'RE LUCKING OUT WITH SOMEBODY WHO'S OPEN MINDED 6776 04:54:40,163 --> 04:54:43,399 TO LET THEM TRY SOMETHING ELSE AND WHAT THEY HAVE MAYBE, 6777 04:54:44,033 --> 04:54:45,701 YOUNG UNDERGRADS OR WHATEVER 6778 04:54:45,701 --> 04:54:47,236 AREN'T THE BEST SOFTWARE ENGINEERS 6779 04:54:47,236 --> 04:54:50,206 FOR OTHER REASONS, I DON'T KNOW, BUT WHAT I THINK THAT, 6780 04:54:51,240 --> 04:54:53,342 MY EXPERIENCE FROM MY OLD PROFESSORING DAYS 6781 04:54:53,342 --> 04:54:54,911 WAS THAT ACTUALLY 6782 04:54:54,911 --> 04:54:57,580 STUDENTS ARE REALLY GREAT ABOUT BEING CURIOUS, 6783 04:54:57,580 --> 04:54:59,515 AND THEY DON'T HAVE THOSE INGRAINED ASSUMPTIONS 6784 04:54:59,515 --> 04:55:02,018 THAT GET HAMMERED INTO YOU BY THE TIME YOU GET YOUR PHD, 6785 04:55:02,018 --> 04:55:03,586 WHERE YOU'RE REALLY INDOCTRINATED 6786 04:55:03,586 --> 04:55:06,055 IN A MEDIEVAL SYSTEM OF BELIEFS OF YOUR MENTOR, 6787 04:55:06,055 --> 04:55:07,356 YOU KNOW, YOU JUST KIND OF 6788 04:55:07,356 --> 04:55:08,925 AND IT'S REALLY HARD TO BREAK OUT OF THAT. 6789 04:55:10,159 --> 04:55:11,727 AND TO THE POINT, 6790 04:55:11,727 --> 04:55:15,898 I THINK SO, LIKE MIKE WAS MAKING ABOUT MARK. 6791 04:55:15,898 --> 04:55:16,532 SORRY, MARK, 6792 04:55:16,532 --> 04:55:17,900 I APOLOGIZE THAT YOU WERE TALKING ABOUT 6793 04:55:17,900 --> 04:55:19,902 ONCE YOU GET THERE'S THIS OTHER, 6794 04:55:19,902 --> 04:55:22,271 THIS RELATES TO WHAT YOU WERE SAYING TO PATRICK. 6795 04:55:22,271 --> 04:55:23,739 WHAT I ALSO HEARD FROM THE STUDENTS, 6796 04:55:23,739 --> 04:55:25,908 WHICH MANY OF YOU BEEN SAYING HERE IS THAT, YOU KNOW, 6797 04:55:25,908 --> 04:55:28,945 WE ARE NOT PREPARING THEM FOR CAREERS IN NEUROAI. 6798 04:55:28,945 --> 04:55:31,347 AND IT'S NOT THAT WE NEED TO HAVE A PHD IN NEUROAI. 6799 04:55:31,347 --> 04:55:32,548 IT'S JUST THAT 6800 04:55:32,548 --> 04:55:35,885 MAYBE THEY NEED INSTITUTIONS THAT ARE BUILT 6801 04:55:35,885 --> 04:55:38,754 MORE AROUND HAVING RESEARCH AND EDUCATION TRACKS. 6802 04:55:38,754 --> 04:55:39,088 YOU KNOW, 6803 04:55:39,088 --> 04:55:40,890 MAYBE THERE NEED TO BE INCENTIVE STRUCTURES 6804 04:55:40,890 --> 04:55:42,291 THAT SUPPORT INNOVATION 6805 04:55:42,291 --> 04:55:44,794 AND EVEN CULTIVATE TOWARDS UNDERSTANDING WHAT IT WOULD MEAN 6806 04:55:44,794 --> 04:55:47,363 TO BE A DEVOPS PERSON OR WHATEVER. 6807 04:55:47,363 --> 04:55:48,698 YOU KNOW, THEY DON'T REALLY LEARN A 6808 04:55:48,698 --> 04:55:50,399 LOT OF THAT UNTIL THEY LEAVE. 6809 04:55:51,434 --> 04:55:52,201 AND I DON'T KNOW THAT 6810 04:55:52,201 --> 04:55:55,271 WE'RE SETTING UP A GOOD BRIDGE FOR THEM TO LEAVE ACADEMIA. 6811 04:55:55,271 --> 04:55:56,706 IF WE'RE SETTING UP A GOOD BRIDGE FOR THEM, 6812 04:55:56,706 --> 04:55:58,874 FOR THOSE TYPES OF CAREERS WHERE THEY'RE NEEDED, 6813 04:55:58,874 --> 04:56:01,877 OR EVEN TO DO MORE INNOVATIVE WORK IN THE UNIVERSITY. 6814 04:56:05,748 --> 04:56:08,951 SO BLAKE STARTED THE CONFERENCE BY BRINGING UP 6815 04:56:08,951 --> 04:56:11,954 ALAN TURING AND JOHN VON NEUMANN AND HOW THEY 6816 04:56:12,755 --> 04:56:15,758 DIED IN THEIR EARLY 50S. 6817 04:56:16,125 --> 04:56:18,794 SO I KNOW THIS IS A TERRIBLE QUESTION. 6818 04:56:18,794 --> 04:56:21,797 BUT, YOU KNOW, SO TEN YEARS ON WITH THE BRAIN INITIATIVE, 6819 04:56:21,998 --> 04:56:23,799 WE CAN THINK IN TERMS OF THE NEXT TEN YEARS, 6820 04:56:23,799 --> 04:56:26,669 OR WE COULD THINK IN TERMS OF, LET'S SAY, 50 YEARS. 6821 04:56:26,669 --> 04:56:29,672 DO YOU HAVE ANY SENSE OF THE VISION 6822 04:56:30,339 --> 04:56:32,742 WHERE NEUROAI WILL BE AND HOW TO GET THERE 6823 04:56:32,742 --> 04:56:34,844 IN 50 YEARS, BARRING THE SINGULARITY? 6824 04:56:39,482 --> 04:56:40,583 ASSUMING THE SINGULARITY 6825 04:56:40,583 --> 04:56:43,586 DOES NOT HAPPEN. OH. 6826 04:56:45,655 --> 04:56:49,091 OH, I THINK IT WILL BE IN A POST AGI WORLD IN 50 YEARS. 6827 04:56:49,158 --> 04:56:51,494 I DON'T THINK THAT THAT'S SUPER 6828 04:56:51,494 --> 04:56:54,497 THAT IT'S THAT CONTROVERSIAL. 6829 04:56:54,497 --> 04:56:58,801 SO I THINK OUR WORLD IS GOING TO BE VERY DIFFERENT. 6830 04:56:59,168 --> 04:57:00,836 MAYBE NOT IN TEN YEARS, BUT 6831 04:57:00,836 --> 04:57:01,203 WELL, YEAH. 6832 04:57:01,203 --> 04:57:02,505 I MEAN, REALLY I'M ASKING MORE 6833 04:57:02,505 --> 04:57:03,773 AND MORE ABOUT A TEN YEAR VISION. 6834 04:57:03,773 --> 04:57:06,776 AND 50 IS A RIDICULOUS NUMBER, I REALIZE. BUT, 6835 04:57:07,476 --> 04:57:11,514 I MEAN, DO YOU HAVE A SENSE OF WOULD YOU KNOW WHAT GOAL 6836 04:57:11,514 --> 04:57:14,483 YOU WOULD WANT TO ACHIEVE BY TEN YEARS FROM NOW? 6837 04:57:18,320 --> 04:57:21,624 I GUESS I'D BE I'D BE HAPPY TO TAKE A TAKE A SHOT. 6838 04:57:21,957 --> 04:57:24,860 SO, ONE OF, 6839 04:57:24,860 --> 04:57:25,161 YOU KNOW, 6840 04:57:25,161 --> 04:57:26,962 I THINK I'D START BY TALKING ABOUT, LIKE, 6841 04:57:26,962 --> 04:57:29,732 THE ARC OF BASIC RESEARCH ON THE WHOLE. RIGHT. 6842 04:57:29,732 --> 04:57:35,337 SO IN THE EARLY 70S, RICHARD NIXON LAUNCHED A WAR ON CANCER, 6843 04:57:35,504 --> 04:57:37,673 AND THAT LED TO A BUNCH OF MONEY GOING TO NIH. 6844 04:57:37,673 --> 04:57:40,876 AND, NIH USED A LOT OF THAT MONEY FOR 6845 04:57:42,044 --> 04:57:43,412 BASIC SCIENCE RESEARCH INTO 6846 04:57:43,412 --> 04:57:46,515 CELL BIOLOGY TO UNDERSTAND HOW CELLS WORKED BETTER. 6847 04:57:46,882 --> 04:57:51,353 AND DECADES LATER, THAT LED TO NEW KINDS OF CURES FOR CANCER. 6848 04:57:52,655 --> 04:57:55,725 YOU KNOW, IMMUNOTHERAPY DRUGS, YOU, YOU IMMUNOTHERAPIES 6849 04:57:55,725 --> 04:57:58,928 YOU CAN SAY, CAME RIGHT OUT OF A LOT OF THAT FUNDING. 6850 04:57:59,395 --> 04:58:00,162 AND SO I THINK THAT'S 6851 04:58:00,162 --> 04:58:03,165 THE NATURE OF BASIC SCIENCE THAT IT'S YOU INVEST IN 6852 04:58:03,532 --> 04:58:06,669 UNDERSTANDING SYSTEMS WITH THE POINT OF BEING A, 6853 04:58:06,669 --> 04:58:09,872 YOU KNOW, SORT OF MEDIUM TO LONG TERM DEVELOPING 6854 04:58:10,840 --> 04:58:13,175 IMPROVEMENTS IN HUMAN HEALTH AND HUMAN SOCIETY. 6855 04:58:13,175 --> 04:58:15,077 AND WHAT WOULD THOSE BE? 6856 04:58:15,077 --> 04:58:16,712 AND SO I THINK THAT'S A GOOD PLACE 6857 04:58:16,712 --> 04:58:18,180 TO START ABOUT THINKING ABOUT, LIKE, 6858 04:58:18,180 --> 04:58:19,081 WHAT ARE WE THINKING ABOUT FOR NEUROAI? 6859 04:58:19,081 --> 04:58:22,585 I SEE A LOT OF THAT, 6860 04:58:23,953 --> 04:58:26,222 THE BASIC SCIENCE THAT WE'RE TALKING ABOUT HERE 6861 04:58:26,222 --> 04:58:28,424 AND THE AMAZING WORK 6862 04:58:28,424 --> 04:58:30,793 THAT MANY PEOPLE ARE HERE DOING TO ADVANCE 6863 04:58:30,793 --> 04:58:32,795 KNOWLEDGE OF NETWORKS IN THE BRAIN. 6864 04:58:32,795 --> 04:58:35,264 IT'S REALLY POINTING TOWARDS 6865 04:58:35,264 --> 04:58:38,434 FUTURE UNDERSTANDING OF BRAIN DISEASES. 6866 04:58:38,434 --> 04:58:41,437 AND WE KNOW THAT A LOT OF I MEAN, I THINK IT'S, 6867 04:58:41,437 --> 04:58:44,607 I THINK MANY OF US THINK THAT MANY BRAIN PATHOLOGIES 6868 04:58:44,607 --> 04:58:46,208 ARISE FROM NETWORK PROPERTIES. 6869 04:58:47,476 --> 04:58:49,044 AND UNDERSTANDING DISEASE, 6870 04:58:49,044 --> 04:58:52,248 , A VARIETY OF DISEASES FROM THE EFFECTS 6871 04:58:52,248 --> 04:58:55,451 OF ALZHEIMER'S TO PARKINSON'S TO SCHIZOPHRENIA, ETC., ETC. 6872 04:58:55,451 --> 04:58:57,953 ARE GOING TO ARISE FROM UNDERSTANDING OF NETWORKS. 6873 04:58:57,953 --> 04:59:00,322 AND THE WORK THAT WE HAVE TALKED ABOUT HERE OVER THE PAST 6874 04:59:00,322 --> 04:59:03,826 TWO DAYS IS VERY MUCH UNDERSTANDING NETWORKS, 6875 04:59:04,426 --> 04:59:07,596 AND I THINK MY SENSE OF THE VISION FOR NEUROAI OVER 6876 04:59:07,596 --> 04:59:08,130 THE NEXT, 6877 04:59:08,130 --> 04:59:10,099 LET'S SAY, 10 TO 20 YEARS, IS REALLY 6878 04:59:10,099 --> 04:59:12,668 THAT WE'RE GOING TO START TO UNDERSTAND BRAIN DISEASES 6879 04:59:12,668 --> 04:59:15,237 BETTER AS WE UNDERSTAND BETTER HOW BRAIN NETWORKS WORK. 6880 04:59:15,237 --> 04:59:18,641 AND A.I. YOU KNOW, PRINCIPLES ARE GOING TO INFORM THAT. 6881 04:59:19,942 --> 04:59:21,577 WHAT DO YOU MEAN BY NETWORK PROPERTIES? 6882 04:59:21,577 --> 04:59:23,145 COULD YOU JUST EXPAND ON THAT? 6883 04:59:23,145 --> 04:59:24,647 YEAH. 6884 04:59:24,647 --> 04:59:24,980 YEAH. 6885 04:59:24,980 --> 04:59:26,148 SO I HAVE 6886 04:59:26,148 --> 04:59:27,516 YOU KNOW, I STARTED MY CAREER STUDYING 6887 04:59:27,516 --> 04:59:31,453 SINGLE NEURONS, AS I THINK PAUL DID, AND MANY OF US DID IN, 6888 04:59:31,453 --> 04:59:33,255 YOU KNOW, IN ANIMALS THAT ARE BEHAVING 6889 04:59:33,255 --> 04:59:36,425 AND CORRELATING THE PROPERTIES OF SINGLE NEURONS WITH BEHAVIOR. 6890 04:59:36,425 --> 04:59:39,128 AND WE HAVE LEARNED ENORMOUS AMOUNTS FROM THAT. 6891 04:59:39,128 --> 04:59:40,462 PEOPLE LIKE, YOU KNOW, 6892 04:59:41,764 --> 04:59:42,131 HUBEL AND 6893 04:59:42,131 --> 04:59:45,334 WIESEL TO BILL NEWSOME AND ETC., ETC. 6894 04:59:45,334 --> 04:59:47,703 HAVE ADVANCED A LOT OF OUR KNOWLEDGE OF LIKE 6895 04:59:47,703 --> 04:59:51,307 THE NEURAL BASIS OF BEHAVIOR. 6896 04:59:51,841 --> 04:59:55,511 AND OVER THE PAST TEN YEARS, THANKS IN PART TO TOOLS 6897 04:59:55,511 --> 04:59:56,378 DEVELOPED BY THE BRAIN INITIATIVE, 6898 04:59:56,378 --> 04:59:58,214 WE HAVE RECORDED FROM MANY, MANY, MANY NEURONS. 6899 04:59:58,214 --> 05:00:01,116 AND WE'RE STARTING TO UNDERSTAND POPULATION LEVEL PHENOMENA 6900 05:00:01,116 --> 05:00:04,153 AND HOW GROUPS OF NEURONS WITHIN AREAS WORK TOGETHER 6901 05:00:04,820 --> 05:00:06,722 AND HOW CIRCUITS CREATE THAT FUNCTION. 6902 05:00:06,722 --> 05:00:10,626 YOU KNOW, WHETHER IT'S SYNAPSES OR DIFFERENT CELL TYPES 6903 05:00:11,026 --> 05:00:15,130 AND ALSO HOW NETWORKS, EXTEND ACROSS BRAIN REGIONS, YOU KNOW, 6904 05:00:16,198 --> 05:00:18,467 AND PEOPLE HERE LIKE KANAKA HAVE WORKED, 6905 05:00:18,467 --> 05:00:20,669 AMONGST MANY OTHERS, HAVE WORKED ON THESE THINGS. 6906 05:00:20,669 --> 05:00:21,871 AND SO WHEN I SAY NETWORK PROPERTIES, I'M 6907 05:00:21,871 --> 05:00:22,638 TALKING ABOUT THAT. 6908 05:00:22,638 --> 05:00:26,242 LINKING CIRCUITS TO THE ACTIVITY OF POPULATIONS OF NEURONS 6909 05:00:26,242 --> 05:00:28,477 THAT GIVE RISE TO COMPUTATIONS AND ULTIMATELY BEHAVIOR. 6910 05:00:31,347 --> 05:00:33,816 SO I WOULD. 6911 05:00:33,816 --> 05:00:34,183 FRAME THIS. 6912 05:00:34,183 --> 05:00:34,884 IN A COUPLE OF WAYS, 6913 05:00:34,884 --> 05:00:38,420 BUT I THINK I WOULD SAY I WOULD LIKE TO SEE A REAL 6914 05:00:38,420 --> 05:00:41,423 I'D LIKE TO SEE IN TEN YEARS, 6915 05:00:42,525 --> 05:00:45,527 A SILICON VERSION OF MULTIPLE LAYERS OF 6916 05:00:45,694 --> 05:00:49,531 OF CORTICAL LAYERS, LIKE TRULY WITH THE FULL PYRAMIDAL. 6917 05:00:49,531 --> 05:00:53,702 SO, I MEAN, I MEAN REALLY BIOPHYSICAL CLOSE TO 6918 05:00:53,702 --> 05:00:55,104 THIS SUCH THAT 6919 05:00:55,104 --> 05:00:58,073 THEN WE CAN USE THAT FOR ENGINEERING APPLICATIONS. 6920 05:00:58,073 --> 05:00:58,941 I THINK THAT IT DOES A 6921 05:00:58,941 --> 05:00:59,775 NUMBER OF THINGS. 6922 05:00:59,775 --> 05:01:02,511 ONE, AND BY THE WAY, WE COULD ALSO DO THIS FOR A COUPLE 6923 05:01:02,511 --> 05:01:04,747 OTHER DIFFERENT STRUCTURES IF WE WANTED TO. 6924 05:01:04,747 --> 05:01:06,081 I'M JUST TO USE THIS TO START WITH. 6925 05:01:06,081 --> 05:01:08,117 SOMEBODY SAYS, I WANT TO TAKE THE FLIERS. 6926 05:01:08,117 --> 05:01:10,552 WE'LL TALK ABOUT THAT TOO. IT'S COOL. 6927 05:01:10,552 --> 05:01:11,854 BUT I THINK IT IS A COUPLE THINGS. 6928 05:01:11,854 --> 05:01:15,190 ONE IS IT CERTAINLY OBVIOUSLY GETS ALL THE 6929 05:01:15,190 --> 05:01:17,259 CIRCUITS UP THAT UP AT A CERTAIN LEVEL. 6930 05:01:17,259 --> 05:01:18,894 I THINK IT ALSO GETS YOU TO A POINT WHERE 6931 05:01:18,894 --> 05:01:20,229 WE CAN ACTUALLY USE THIS 6932 05:01:20,229 --> 05:01:22,898 FOR THIS APPLICATION, BECAUSE THE TECHNOLOGY BE READY. 6933 05:01:22,898 --> 05:01:24,967 BUT I THINK MORE IMPORTANTLY, IT FORCES. 6934 05:01:24,967 --> 05:01:25,734 US TO ASK A LOT OF 6935 05:01:25,734 --> 05:01:26,869 QUESTIONS, 6936 05:01:26,869 --> 05:01:29,872 TO INTEGRATE OUR UNDERSTANDING OF WHAT DO WE MEAN BY 6937 05:01:30,472 --> 05:01:31,373 WHAT HAPPENS WHEN I 6938 05:01:31,373 --> 05:01:34,376 IN TERMS OF THE ENGINEERING, IN TERMS OF THE COMPUTATION 6939 05:01:34,643 --> 05:01:37,479 OF, SAY, MULTIPLE PYRAMIDAL CELLS, 6940 05:01:37,479 --> 05:01:39,315 CORTICAL COLUMN AND SO FORTH? 6941 05:01:39,315 --> 05:01:40,549 AND I THINK THAT WILL 6942 05:01:41,483 --> 05:01:42,217 CREATE A 6943 05:01:42,217 --> 05:01:46,155 HUGE ENTHUSIASM IN TERMS OF A WHOLE SET OF THINGS 6944 05:01:46,155 --> 05:01:48,257 IN COMPUTATIONAL NEUROSCIENCE. 6945 05:01:48,257 --> 05:01:51,260 THAT PROBABLY WILL CHANGE THE WAY WE TALK ABOUT 6946 05:01:51,427 --> 05:01:52,261 THE MODELS. 6947 05:01:52,261 --> 05:01:54,396 THE WAY WE TALK ABOUT, YOU KNOW, 6948 05:01:54,396 --> 05:01:55,931 NEURAL ENCODING 6949 05:01:55,931 --> 05:01:59,068 ALL THE WAY UP AND DOWN THE CHAIN. THAT WILL PROBABLY 6950 05:01:59,101 --> 05:02:02,271 -- IT WILL FINALLY HELP TO INTEGRATE THINGS AS WELL AS 6951 05:02:03,105 --> 05:02:05,407 BRING A LOT OF COMMUNITIES TOGETHER. 6952 05:02:05,407 --> 05:02:08,477 SO IF I HAD A TEN YEAR PLAN, THAT'S WHAT I WOULD GO FOR. 6953 05:02:16,485 --> 05:02:16,986 OH, SORRY. 6954 05:02:16,986 --> 05:02:20,322 SO, IF ANYONE HAS QUESTIONS, PLEASE COME GO TO THE MIC, 6955 05:02:20,789 --> 05:02:22,591 AND ASK QUESTIONS. SO THIS IS IT'S AN OPEN 6956 05:02:22,591 --> 05:02:24,627 DISCUSSION, NOT A CLOSED DISCUSSION. 6957 05:02:24,627 --> 05:02:27,630 ANY ANSWERS TO THE QUESTIONS PAUL RAISED? 6958 05:02:27,696 --> 05:02:30,366 IT'S A TWO WAY STREET. 6959 05:02:30,366 --> 05:02:33,369 CAN I CONTINUE ALONG THIS LINE? 6960 05:02:33,669 --> 05:02:34,003 OKAY. 6961 05:02:34,003 --> 05:02:36,372 SO I THINK WE HAVE LIKE, 6962 05:02:36,372 --> 05:02:40,142 PRETTY LIKE ALREADY WE'RE MAKING A LOT OF PROGRESS ON THE FLY. 6963 05:02:40,142 --> 05:02:43,312 AND SO I WOULD NOT BE SURPRISED IF WE COULD REALLY SAY 6964 05:02:43,312 --> 05:02:45,714 THAT IN TEN YEARS THE FLY WOULD BE SOLVED. 6965 05:02:45,714 --> 05:02:49,118 WE'RE ALREADY ABLE TO MEASURE FROM ABOUT 10% OF THE NEURONS 6966 05:02:49,118 --> 05:02:50,853 IN THE MOUSE. 6967 05:02:52,688 --> 05:02:53,889 PEOPLE HAVE SAID THAT 6968 05:02:53,889 --> 05:02:55,391 BUILDING THE CONNECTOME FOR THE MOUSE 6969 05:02:55,391 --> 05:02:58,327 WOULD TAKE ABOUT 15 YEARS, BUT THOSE ESTIMATES HAVE 6970 05:02:58,327 --> 05:03:02,665 BECOME MUCH SHORTER AS EXTENSION MICROSCOPY IS, HAS IMPROVED. 6971 05:03:03,065 --> 05:03:05,734 SO I THINK THE VIRTUAL MOUSE, 6972 05:03:05,734 --> 05:03:07,002 YOU KNOW, THE KIND OF GRANULARITY 6973 05:03:07,002 --> 05:03:09,705 THAT WE'RE SEEING IN FLIES, I THINK 6974 05:03:09,705 --> 05:03:11,874 MIGHT BE FEASIBLE 6975 05:03:11,874 --> 05:03:14,877 JUST SO WITHIN THE NEXT TEN YEARS. 6976 05:03:15,477 --> 05:03:20,249 FOR NON-HUMAN PRIMATES AND FOR HUMANS, 6977 05:03:20,249 --> 05:03:21,884 I THINK THAT WE WOULD NEED TO VASTLY 6978 05:03:21,884 --> 05:03:24,887 INCREASE OUR CAPACITIES TO RECORD FROM NEURONS. 6979 05:03:25,087 --> 05:03:26,321 SO IF YOU LOOK AT, YOU KNOW, 6980 05:03:26,321 --> 05:03:30,225 THE SO-CALLED MOORE'S LAW FOR, FOR NEURAL RECORDINGS, IT 6981 05:03:30,225 --> 05:03:33,228 DOUBLES ABOUT EVERY FIVE YEARS WHEN IT COMES TO ELECTRODES. 6982 05:03:34,096 --> 05:03:36,765 IF WE WANT TO GET, YOU KNOW, PRETTY REASONABLE COVERAGE 6983 05:03:36,765 --> 05:03:38,400 OF THE ENTIRE CORTEX, 6984 05:03:38,400 --> 05:03:40,703 IT WOULD NEED TO INCREASE AT A MUCH FASTER RATE THAN THAT. 6985 05:03:40,703 --> 05:03:43,072 BUT THAT'S, YOU KNOW, WITHIN THE REALM OF POSSIBILITY, 6986 05:03:43,072 --> 05:03:46,075 IF WE PUT THE ENGINEERING EFFORT, INTO IT 6987 05:03:46,275 --> 05:03:49,078 AND I THINK THAT, WE'LL GET 6988 05:03:49,078 --> 05:03:53,115 WE'LL PROBABLY BE ABLE TO GET AT THAT POINT RECORDINGS FROM 6989 05:03:53,582 --> 05:03:57,119 YOU KNOW, TENS OF PEOPLE AND TENS OF ANIMALS 6990 05:03:57,553 --> 05:03:59,788 INTERACTING AND SOCIAL INTERACTIONS, 6991 05:04:01,690 --> 05:04:02,925 WIRELESSLY RECORDING. 6992 05:04:02,925 --> 05:04:04,960 I COULD DEFINITELY SEE THAT HAPPEN. 6993 05:04:04,960 --> 05:04:05,794 WILL WE NEED THOSE 6994 05:04:05,794 --> 05:04:08,797 RECORDINGS, THOUGH, IF WE HAVE FOUNDATION MODELS? 6995 05:04:09,098 --> 05:04:10,532 YEAH, OF COURSE WE NEED THOSE RECORDINGS. 6996 05:04:10,532 --> 05:04:12,534 AND JUST TO BE CLEAR, YOU KNOW, WHEN I WAS MENTIONING 6997 05:04:12,534 --> 05:04:15,871 100,000 HOURS, YESTERDAY THAT WE ALREADY 6998 05:04:15,871 --> 05:04:18,874 HAVE, AS, 6999 05:04:19,274 --> 05:04:21,877 ANDREAS CORRECTLY POINTED OUT, YOU KNOW, A LOT OF 7000 05:04:21,877 --> 05:04:24,947 THIS IS DATA IN LOW ENTROPY SITUATIONS. 7001 05:04:25,747 --> 05:04:28,584 A LOT OF THIS IS DATA WHEN PATIENTS ARE MONITORED, 7002 05:04:28,584 --> 05:04:29,985 FOR INSTANCE, WHERE THEY'RE NOT DOING 7003 05:04:29,985 --> 05:04:32,988 ANY PARTICULAR TASK AND WE DON'T HAVE THE LABELS FOR THEM. 7004 05:04:33,055 --> 05:04:35,958 SO, SO IT'S IMPORTANT TO CAVEAT THIS. 7005 05:04:35,958 --> 05:04:38,760 WHAT I'M SAYING IS THAT WE HAVE A GOOD BASIS LIKE, 7006 05:04:38,760 --> 05:04:40,295 IT FEELS TO ME LIKE THE FIRST TEN YEARS 7007 05:04:40,295 --> 05:04:42,164 OF THE BRAIN INITIATIVE WAS REALLY LIKE ABOUT 7008 05:04:42,164 --> 05:04:45,434 BUILDING A STRONG FOUNDATION, HAVING ALL THESE NEW 7009 05:04:45,434 --> 05:04:48,437 NEUROTECHNOLOGIES, HAVING ALL OF THESE DATA ARCHIVES. 7010 05:04:48,437 --> 05:04:50,606 BUILDING THE INFRASTRUCTURE TO BE ABLE TO GET, 7011 05:04:53,976 --> 05:04:55,611 TO BE ABLE TO, 7012 05:04:55,611 --> 05:04:59,414 PUT METADATA ON DATA THAT ALREADY EXISTS, SO THAT WE 7013 05:04:59,414 --> 05:05:01,150 CAN ACTUALLY LEVERAGE THIS DATA. 7014 05:05:01,150 --> 05:05:02,918 AND NOW I THINK IT'S THE RIGHT 7015 05:05:02,918 --> 05:05:05,921 TIME TO KIND OF PUT IT INTO HIGH GEAR. 7016 05:05:07,456 --> 05:05:10,459 IF ANYONE WANT TO COMMENT ON THAT BEFORE WE TAKE A QUESTION. 7017 05:05:11,760 --> 05:05:12,628 OKAY. 7018 05:05:12,628 --> 05:05:14,329 QUESTION. 7019 05:05:14,329 --> 05:05:15,931 HI. THIS IS A GREAT DISCUSSION. 7020 05:05:15,931 --> 05:05:18,200 ONE QUESTION IS IN THE NEXT TEN YEARS, 7021 05:05:18,200 --> 05:05:20,669 WOULD IT ALSO BE POSSIBLE TO INTEGRATE NON- 7022 05:05:20,669 --> 05:05:23,438 ELECTRICAL SIGNALS AND NON-NEURONAL CELLS 7023 05:05:23,438 --> 05:05:25,541 AT THE SYSTEMS LEVEL INTO THESE MODELS? 7024 05:05:25,541 --> 05:05:26,675 WOULD THAT BE HELPFUL? 7025 05:05:26,675 --> 05:05:28,277 JENNIFER I WAS GOING TO ASK YOU THAT AS WELL. 7026 05:05:28,277 --> 05:05:28,777 I MEAN DO 7027 05:05:28,777 --> 05:05:31,647 WE JUST NEED THE PYRAMIDAL NEURONS OR WHAT LEVEL OF? 7028 05:05:31,647 --> 05:05:34,016 I WOULD THINK YOU I MEAN CERTAINLY. 7029 05:05:34,016 --> 05:05:35,617 THE ELECTRICAL SIGNALS YOU'D LIKE THERE. 7030 05:05:35,617 --> 05:05:38,954 I THINK I WOULD LOVE TO SEE, YOU KNOW ALSO WHERE 7031 05:05:38,954 --> 05:05:41,957 IT MAKES SENSE, THE CHEMICAL SIGNALS 7032 05:05:42,357 --> 05:05:44,026 AND OTHER RELATED THINGS. 7033 05:05:44,026 --> 05:05:47,563 CERTAINLY THE DISCUSSIONS AROUND SOME OF THE NON NEURAL BLOCKS, 7034 05:05:47,563 --> 05:05:50,566 WHICH ARE ACTUALLY DOING MODULATIONS AT, 7035 05:05:50,632 --> 05:05:53,235 YOU KNOW, MINUTES TO LONGER TIMESCALES 7036 05:05:53,235 --> 05:05:54,069 I THINK IS ABSOLUTELY 7037 05:05:54,069 --> 05:05:56,205 PART OF THAT CONVERSATION. 7038 05:05:56,205 --> 05:05:58,040 I THINK IT HAS TO BE PART OF THAT CONVERSATION 7039 05:05:58,040 --> 05:06:00,309 IF WE'RE GOING TO BE REAL ABOUT THIS. 7040 05:06:00,309 --> 05:06:02,578 SO EXACTLY WHAT THAT ROADMAP 7041 05:06:02,578 --> 05:06:02,978 IS, I THINK 7042 05:06:02,978 --> 05:06:05,414 IT'S STILL A LITTLE LESS THAN OBVIOUS, BUT 7043 05:06:05,414 --> 05:06:06,815 I'D LIKE US TO GO DOWN THAT PATH. 7044 05:06:11,720 --> 05:06:12,254 ANDRES. 7045 05:06:12,254 --> 05:06:12,554 YEAH. 7046 05:06:12,554 --> 05:06:16,725 SO I WHAT I THINK A BIG, BOLD VISION. 7047 05:06:16,725 --> 05:06:18,861 I THINK WE NEED A BOLD NEW 7048 05:06:18,861 --> 05:06:22,364 VISION FOR THE NEXT TEN YEARS AND NOT CONTINUE 7049 05:06:22,731 --> 05:06:25,400 IN ADDITION TO CONTINUE DOING WHAT WE'RE CURRENTLY DOING. 7050 05:06:25,400 --> 05:06:28,737 I THINK WE NEED SOMETHING NEW, BOTH AT THE STRUCTURAL LEVEL 7051 05:06:28,737 --> 05:06:31,740 AND IN TERMS OF THE SCALE OF THE PROJECT. 7052 05:06:32,107 --> 05:06:34,042 AND IT DOESN'T HAVE TO BE ONE. IT COULD BE MULTIPLE. 7053 05:06:34,042 --> 05:06:37,946 BUT HERE IS THE ONE IDEA THAT I WOULD ADVOCATE FOR. 7054 05:06:38,347 --> 05:06:41,049 IF YOU LOOK AT THE LAST TEN YEARS, WHAT WE'VE ACHIEVED, 7055 05:06:41,049 --> 05:06:44,052 AS IT WAS SAID MANY TIMES, IS REALLY NEUROTECHNOLOGY 7056 05:06:44,286 --> 05:06:47,623 TO BE ABLE TO DO EXPERIMENTS AT SCALE AND ALSO IS 7057 05:06:47,856 --> 05:06:48,924 DEEP LEARNING, RIGHT? 7058 05:06:48,924 --> 05:06:51,593 I MEAN, A.I. HAVE SOLVED PROTEIN FOLDING. 7059 05:06:51,593 --> 05:06:53,729 IT HAS DONE AMAZING STUFF. 7060 05:06:53,729 --> 05:06:56,131 AND I THINK IN THE NEXT TEN YEARS 7061 05:06:56,131 --> 05:06:58,767 WE NEED TO BRING THESE TWO THINGS TOGETHER. 7062 05:06:58,767 --> 05:07:04,106 BUT TO DO THESE, WE NEED TO DO A LOT OF LARGE SCALE DATA COLLECTION. 7063 05:07:04,539 --> 05:07:07,609 AND I THINK TO DO THAT WELL, IT NEEDS TO BE DONE 7064 05:07:08,477 --> 05:07:11,780 MOSTLY IN A CENTRALIZED WAY, WHERE WE'RE GOING TO BE ABLE 7065 05:07:11,780 --> 05:07:12,915 TO DESIGN THE RIGHT 7066 05:07:12,915 --> 05:07:15,951 EXPERIMENTS WHENN THE ANIMALS ARE GOING TO BE ENGAGED 7067 05:07:15,951 --> 05:07:17,552 IN COMPLEX TASKS. 7068 05:07:17,552 --> 05:07:21,123 SO THEN WE CAN GATHER THIS DATA AND THEN BUILD, AS IT WAS 7069 05:07:21,123 --> 05:07:22,891 SAID, THE FOUNDATION MODEL. 7070 05:07:22,891 --> 05:07:24,259 AND I THINK THAT FOUNDATION MODEL 7071 05:07:24,259 --> 05:07:25,961 IS GOING TO BE ABLE TO BE USED BY EVERYBODY, 7072 05:07:25,961 --> 05:07:27,796 BECAUSE YOU CAN THINK OF IT LIKE AN API 7073 05:07:27,796 --> 05:07:29,731 WHERE YOU'LL BE ABLE TO GO IN 7074 05:07:29,731 --> 05:07:30,565 RUN IN SILICO 7075 05:07:30,565 --> 05:07:33,568 EXPERIMENTS, DESIGN EXPERIMENTS, AND, AND TEST THEM BACK 7076 05:07:33,602 --> 05:07:35,170 IN YOUR OWN LAB. 7077 05:07:35,170 --> 05:07:38,940 SO I DO THINK WE NEED LIKE, AND I THINK 7078 05:07:38,940 --> 05:07:40,142 THIS IS THE RIGHT TIME TO DO IT. 7079 05:07:40,142 --> 05:07:43,278 WE NEED LIKE A BIG INVESTMENT IN NEW TYPE OF INVESTMENT 7080 05:07:43,278 --> 05:07:46,281 FROM THE FEDERAL GOVERNMENT, LIKE THE GENOME PROJECT. 7081 05:07:46,481 --> 05:07:47,516 THAT'S GOING TO SAY, OKAY, 7082 05:07:47,516 --> 05:07:50,686 IN THE LAST TEN YEARS WE'VE HAD NEUROTECHNOLOGY. 7083 05:07:50,686 --> 05:07:53,155 NOW WE HAVE DEEP LEARNING. LET'S BRING IT TOGETHER. 7084 05:07:53,155 --> 05:07:56,158 AND I SEE THIS AS HEDGING OUR BETS IN THE FOLLOWING WAY. 7085 05:07:56,325 --> 05:08:00,295 EVEN IF A.I. IN THE NEXT FIVE YEARS LIKE I SAW, YOU KNOW, 7086 05:08:00,295 --> 05:08:04,466 OPENAI, ANTHROPIC AI'S CEO HAD A PODCAST YESTERDAY 7087 05:08:04,466 --> 05:08:08,203 AND HE CLAIMED THAT IN 2 OR 3 YEARS WE'LL HAVE AGI. 7088 05:08:08,203 --> 05:08:09,705 SO YOU CAN IMAGINE THAT THE 7089 05:08:09,705 --> 05:08:12,708 IF WE CREATE AGENTS THAT BECOME AUTONOMOUS, 7090 05:08:12,841 --> 05:08:14,109 THEY'RE GOING TO BE IN SILICO 7091 05:08:14,109 --> 05:08:16,445 FOR THE NEXT FEW YEARS, LIKE WE'LL BE SOME VERSION OF 7092 05:08:16,445 --> 05:08:17,946 LARGE LANGUAGE MODELS. 7093 05:08:17,946 --> 05:08:19,514 SO IF WE BUILD THESE FOUNDATION 7094 05:08:19,514 --> 05:08:21,450 MODELS, NOT ONLY SCIENTISTS WILL BE 7095 05:08:21,450 --> 05:08:23,385 ABLE TO MINE AND DO EXPERIMENTS IN SILICO, 7096 05:08:23,385 --> 05:08:26,888 BUT THESE A.I. AGENTS CAN THEN GENERATE HYPOTHESES. 7097 05:08:27,322 --> 05:08:30,959 IF, ON THE OTHER HAND, THESE PATHS TOWARDS AGI 7098 05:08:30,992 --> 05:08:33,695 GET STUCK AND THE SCALING LAWS DON'T WORK, 7099 05:08:33,695 --> 05:08:36,565 THEN THE TYPE OF EXPERIMENTS THAT WE NEUROSCIENCE CAN DO, 7100 05:08:36,565 --> 05:08:39,568 WE'RE GOING TO PROVIDE A NEW DIMENSION OF NEURAL DATA 7101 05:08:39,735 --> 05:08:40,969 TO REALLY HELP A.I. 7102 05:08:40,969 --> 05:08:43,972 AND THERE'S THE WHOLE ISSUE OF BUILDING FOUNDATION 7103 05:08:43,972 --> 05:08:45,674 MODELS FOR NEUROPSYCHIATRIC DISEASES. 7104 05:08:45,674 --> 05:08:49,644 BUT I DO THINK WE NEED SOMETHING MORE OF A BOLD VISION, 7105 05:08:50,312 --> 05:08:50,979 SOMETHING THAT'S 7106 05:08:50,979 --> 05:08:54,383 GOING TO BRING NEW TYPES OF FUNDING TO BE ABLE TO DO THIS. 7107 05:08:54,683 --> 05:08:55,417 AND IT WILL INVOLVE 7108 05:08:55,417 --> 05:08:59,221 MANY PARTS OF THE GOVERNMENT LIKE NIH, OF COURSE. 7109 05:08:59,588 --> 05:09:02,657 DOE, DARPA, DOD 7110 05:09:03,425 --> 05:09:04,192 ANYWAY, THAT'S. 7111 05:09:08,497 --> 05:09:11,500 COMMENTS? 7112 05:09:12,367 --> 05:09:13,702 SO I, 7113 05:09:13,702 --> 05:09:14,369 I GUESS I, 7114 05:09:14,369 --> 05:09:17,372 I THINK ABOUT THIS VISION, I APPRECIATED WHAT YOU ACTUALLY 7115 05:09:17,372 --> 05:09:20,575 I WAS VERY MOVED ANDREAS WHEN YOU WERE TALKING EARLIER 7116 05:09:20,575 --> 05:09:24,946 ABOUT HOW USEFUL OR MAYBE NOT USEFUL THE DATA THAT WE HAVE IS. 7117 05:09:25,680 --> 05:09:28,083 I THINK I WAS STRUCK WITH HOW LIMITED YOU SAID IT MIGHT BE. 7118 05:09:28,083 --> 05:09:29,451 I THINK YOU'VE REPEATED THIS TOO. 7119 05:09:29,451 --> 05:09:31,119 LIKE WE NEED THE LABELS ON IT. 7120 05:09:31,119 --> 05:09:34,122 WE EITHER WE LACK METADATA, ETC. 7121 05:09:34,389 --> 05:09:38,360 SO I AGREE WITH A BOLD VISION 7122 05:09:38,593 --> 05:09:41,029 AND I WOULD SUPPORT A VISION ALSO THAT 7123 05:09:41,029 --> 05:09:43,732 TAKES CARE OF THE PROCESSES IN WHICH IT DOES THAT. 7124 05:09:43,732 --> 05:09:46,268 AND ONE OF THAT THING THAT CAN BE DONE 7125 05:09:46,268 --> 05:09:49,271 SOONER IS THIS IS SOMETHING I WAS ASKING, 7126 05:09:50,472 --> 05:09:52,040 IS SHE HERE? DOMINIQUE? 7127 05:09:52,040 --> 05:09:52,874 ANYWAY, 7128 05:09:52,874 --> 05:09:55,877 SOMETHING WE WERE TALKING ABOUT WAS WITH THE EXISTING DATA 7129 05:09:56,278 --> 05:09:57,546 OF ALL THOSE EIGHT 7130 05:09:57,546 --> 05:10:01,082 REPOSITORIES THAT WERE DISCUSSED AND INTRODUCED, OF THOSE, 7131 05:10:01,550 --> 05:10:04,553 HOW REPRESENTATIVE DO WE KNOW THOSE ARE? 7132 05:10:04,719 --> 05:10:06,221 AND SO SOME OF THEM ON PURPOSE, 7133 05:10:06,221 --> 05:10:07,823 CERTAIN KINDS OF DEMOGRAPHIC INFORMATION 7134 05:10:07,823 --> 05:10:10,559 IS NOT PUT ON THEM FOR AVOIDING RE-IDENTIFICATION 7135 05:10:10,559 --> 05:10:12,427 BECAUSE THEY'RE FROM SO FEW PEOPLE. 7136 05:10:12,427 --> 05:10:16,298 BUT OTHERS, HAVE SEX DATA ON THEM, FOR EXAMPLE. 7137 05:10:16,565 --> 05:10:20,168 AND WE ALREADY KNOW OF THE HUMAN DATA, AT LEAST THAT MAYBE, 7138 05:10:20,502 --> 05:10:25,240 YOU KNOW, IT'S MAYBE 2% OF THAT DATA ACTUALLY IS, 7139 05:10:25,907 --> 05:10:27,008 FEATURING PEOPLE WHO ARE 7140 05:10:27,008 --> 05:10:29,678 TYPICALLY UNDERREPRESENTED IN SCIENTIFIC STUDIES. 7141 05:10:29,678 --> 05:10:31,313 SO I THINK THAT 7142 05:10:31,313 --> 05:10:34,282 AND I'M NOT SAYING THAT MEANS WE CAN'T USE THE DATA AT ALL, 7143 05:10:34,416 --> 05:10:35,317 BUT I THINK THAT MEANS 7144 05:10:35,317 --> 05:10:37,819 THAT WE NEED TO BE VERY CLEAR ON WHAT WE'RE USING 7145 05:10:37,819 --> 05:10:40,155 IF WE START TRYING TO BUILD MODELS ON THAT, 7146 05:10:40,155 --> 05:10:42,190 IF WE START TRYING TO MAKE CLAIMS 7147 05:10:42,190 --> 05:10:44,025 ABOUT HOW THE BRAIN WORKS. 7148 05:10:44,025 --> 05:10:45,994 THE OTHER PIECE THAT I WOULD SAY 7149 05:10:45,994 --> 05:10:48,997 IN THIS BIG CONVERSATION, WE ALREADY MENTIONED AGI. 7150 05:10:49,197 --> 05:10:51,633 WELL, WE DIDN'T TALK ABOUT WHAT I MENTIONED EARLIER IN MY TALK 7151 05:10:51,633 --> 05:10:53,802 WAS RESPONSIBLE CONCEPTUALIZATION. 7152 05:10:53,802 --> 05:10:56,805 A LOT OF THE TERMS INHERENT IN A.I. ARE BORROWED. 7153 05:10:57,372 --> 05:10:57,839 YOU KNOW, 7154 05:10:57,839 --> 05:10:59,541 THEY ARE BORROWED FROM BRAIN 7155 05:10:59,541 --> 05:11:02,110 OR THEY'RE BORROWED TO REFERENCES TO HUMANS. 7156 05:11:02,110 --> 05:11:03,145 SOME OF THAT'S BY DESIGN, 7157 05:11:03,145 --> 05:11:03,778 BUT SOME OF THAT'S 7158 05:11:03,778 --> 05:11:07,282 KIND OF MORE CONVENIENT, LIKE INTELLIGENCE, AUTONOMY, 7159 05:11:07,649 --> 05:11:11,019 MOTIVATION. AND ALTHOUGH SCIENTISTS, 7160 05:11:11,019 --> 05:11:12,621 SOME SCIENTISTS MIGHT CLAIM THAT THEY HAVE 7161 05:11:12,621 --> 05:11:14,389 THE SAME DEFINITION AS OUR PEERS, 7162 05:11:14,389 --> 05:11:16,658 WE HAVE SEEN HOW THEY ACTUALLY DON'T. 7163 05:11:16,658 --> 05:11:18,326 AND IF YOU'RE NOT INTENTIONALLY 7164 05:11:18,326 --> 05:11:21,530 USING THOSE TO CALL UPON HUMAN FEATURES, 7165 05:11:22,030 --> 05:11:25,033 THEN WE MIGHT THINK TWICE ABOUT HOW WE'RE SPEAKING ABOUT THEM. 7166 05:11:25,033 --> 05:11:27,802 BECAUSE OUR PEERS, AND ESPECIALLY THE PUBLIC, 7167 05:11:27,802 --> 05:11:29,738 WHEN THEY HEAR US TALK ABOUT WHAT WE'RE BUILDING 7168 05:11:29,738 --> 05:11:33,141 AND THE TYPES OF TECHNOLOGY WE ARE TRYING TO CREATE, 7169 05:11:33,508 --> 05:11:35,744 WE'LL HEAR SOMETHING THAT'S ALONG THE LINES 7170 05:11:35,744 --> 05:11:38,647 OF THE POPULAR IMAGINATION, WHICH ACTUALLY SCIENTISTS 7171 05:11:38,647 --> 05:11:42,684 ARE JUST AS VULNERABLE TO FOLK PSYCHOLOGY AND CULTURAL 7172 05:11:43,318 --> 05:11:45,387 NOTIONS, TOO. SO I THINK THAT WOULD BE 7173 05:11:45,387 --> 05:11:48,156 AN IMPORTANT PART OF THAT VISION. 7174 05:11:48,156 --> 05:11:49,591 AND THE LAST PIECE, I HOPE, 7175 05:11:49,591 --> 05:11:51,426 IS THAT THERE'S AN ETHICS- BY-DESIGN APPROACH 7176 05:11:51,426 --> 05:11:52,727 WHERE WE'RE ASKING MORE QUESTIONS, 7177 05:11:52,727 --> 05:11:55,397 NOT JUST ABOUT RESPONSIBLE CONCEPTUALIZATION AND DATA, 7178 05:11:55,397 --> 05:11:58,066 BUT HAVING HUMANS IN THE LOOP AS WE THINK ABOUT APPLICATIONS 7179 05:11:58,066 --> 05:11:58,767 FOR THEM IN HEALTH. 7180 05:12:02,237 --> 05:12:02,604 YEAH. 7181 05:12:02,604 --> 05:12:05,840 I SO, WHAT KAREN IS SAYING ABOUT, 7182 05:12:06,841 --> 05:12:09,411 YOU KNOW, BEING CAREFUL ABOUT DATA AND PUBLIC OUTREACH 7183 05:12:09,411 --> 05:12:12,047 AND THINKING ABOUT HOW WE SPEAK TO THE PUBLIC AND COMMUNICATE 7184 05:12:12,047 --> 05:12:15,050 WHAT WE'RE DOING AND THE BROAD SCOPE OF OUR VISION AND, 7185 05:12:16,351 --> 05:12:18,420 MAKING, YOU KNOW, THINKING ABOUT CAREFULLY 7186 05:12:18,420 --> 05:12:21,189 ABOUT HOW TO COMMUNICATE, I THINK IS SUPER IMPORTANT. 7187 05:12:21,189 --> 05:12:23,224 I ALSO WANTED TO JUST RESPOND TO ANDREAS 7188 05:12:23,224 --> 05:12:25,927 AND APPARENTLY SAY SOMETHING THAT MIGHT BE SOMEWHAT 7189 05:12:25,927 --> 05:12:28,930 UNPOPULAR IN THE ROOM BASED ON THE REACTION TO HIS COMMENTS. 7190 05:12:29,264 --> 05:12:30,131 I 7191 05:12:30,131 --> 05:12:30,966 THINK WE 7192 05:12:30,966 --> 05:12:32,601 CERTAINLY NEED A BROAD VISION 7193 05:12:32,601 --> 05:12:34,769 AND WE NEED TO BE ABLE TO COMMUNICATE IT. 7194 05:12:34,769 --> 05:12:36,538 THAT IS INCREDIBLY ESSENTIAL. 7195 05:12:36,538 --> 05:12:40,208 I ALSO THINK WE NEED MULTI-SCALE PROJECTS. 7196 05:12:40,208 --> 05:12:44,379 I THINK WE SHOULD BE VERY CAREFUL OF BIG INVESTMENTS IN 7197 05:12:45,380 --> 05:12:46,748 SINGLE PROJECTS. 7198 05:12:46,748 --> 05:12:49,751 AND I, I THINK THE BRAIN INITIATIVE, 7199 05:12:50,252 --> 05:12:53,555 AS ORIGINALLY ENVISIONED HERE IN THE US, 7200 05:12:54,089 --> 05:12:57,626 DID AMAZING WORK FUNDING SOME SMALL TEAMS 7201 05:12:57,626 --> 05:13:00,895 OF SMALL NUMBERS OF PIS AND SOME BIGGER TEAMS 7202 05:13:00,895 --> 05:13:03,832 THAT LASTED LONGER AND SOME REALLY BIG PROJECTS. 7203 05:13:03,832 --> 05:13:06,768 AND FROM MY PERSPECTIVE, A LOT OF THE, 7204 05:13:06,768 --> 05:13:07,435 YOU KNOW, WE'RE TALKING ABOUT 7205 05:13:07,435 --> 05:13:09,304 INTEGRATING PEOPLE HERE, ABOUT INTEGRATING THEORY 7206 05:13:09,304 --> 05:13:11,439 AND BRINGING THEORY AND EXPERIMENTS TOGETHER. 7207 05:13:11,439 --> 05:13:13,742 THE BRAIN INITIATIVE WAS INCREDIBLY SUCCESSFUL 7208 05:13:13,742 --> 05:13:16,111 AT BRINGING TOGETHER SMALL GROUPS OF PIS 7209 05:13:16,111 --> 05:13:19,080 TO REALLY DO DEEP WORK ON THEORY AND EXPERIMENT TOGETHER, 7210 05:13:19,948 --> 05:13:22,283 AND FUNDING LOTS OF DIFFERENT TEAMS LIKE THAT. 7211 05:13:22,283 --> 05:13:26,021 AND I THINK THAT HAS BEEN A SUPER SUCCESSFUL MODEL. 7212 05:13:26,021 --> 05:13:28,290 AND AND SO, YES, BOLD VISION. 7213 05:13:28,290 --> 05:13:30,425 YES, A BIG PROJECT. YES. SOME BIG PROJECTS. 7214 05:13:30,425 --> 05:13:32,994 BUT ALSO LET'S THINK ABOUT MULTI-SCALE TEAMS 7215 05:13:32,994 --> 05:13:35,163 AND BRING PEOPLE TOGETHER TO THINK ABOUT DEEP STUFF 7216 05:13:35,163 --> 05:13:37,332 THAT IS GOING TO HAVE REALLY BIG BASIC RETURNS. 7217 05:13:38,600 --> 05:13:41,603 CAN I JUST ADD SOMETHING TO THAT? 7218 05:13:41,636 --> 05:13:45,774 I'M ALSO VERY INSPIRED BY ANDREAS'S VISION, 7219 05:13:46,141 --> 05:13:49,110 WHICH I THINK BUILDS ON THE SUCCESSES 7220 05:13:49,310 --> 05:13:53,181 OF THE BRAIN PROJECTS OVER THE LAST 7221 05:13:53,214 --> 05:13:56,451 TEN YEARS, BRAIN INITIATIVE OVER THE LAST TEN YEARS. 7222 05:13:56,818 --> 05:13:59,821 BUT IN SORT OF SUPPORT OF WHAT YOU WERE SAYING, 7223 05:14:01,189 --> 05:14:03,558 OVER I MEAN, THE WAY THE BRAI 7224 05:14:03,558 --> 05:14:07,395 INITIATIVE STARTED ESPECIALLY WAS WITH A BUNCH OF SMALL TEAMS 7225 05:14:07,729 --> 05:14:12,967 THAT WERE PUT TOGETHER OFTEN IN RESPONSE TO SORT OF, CALLS 7226 05:14:12,967 --> 05:14:17,405 FOR THE DEVELOPMENT OF NEURO- TECHNOLOGIES OF VARIOUS SORTS. 7227 05:14:17,739 --> 05:14:20,675 AND SO I ACTUALLY THINK THAT, 7228 05:14:20,675 --> 05:14:24,346 THE, THE IDEAL SITUATION WOULD BE IF WE HAD BOTH. 7229 05:14:24,546 --> 05:14:27,816 WE SORT OF MOVE FORWARD WITH THE EXCITING RESULTS. 7230 05:14:27,816 --> 05:14:30,819 WE BUILD ON THE EXCITING RESULTS THAT WE ALREADY HAVE, 7231 05:14:30,852 --> 05:14:33,254 AND THEN WE SORT OF TARGET 7232 05:14:33,254 --> 05:14:36,524 THE AREAS WHICH ARE NOT READY FOR BIG SCIENCE YET. 7233 05:14:36,524 --> 05:14:37,425 FOR EXAMPLE, 7234 05:14:38,393 --> 05:14:39,027 THERE WERE A LOT 7235 05:14:39,027 --> 05:14:42,030 OF REALLY INTERESTING TALKS ON 7236 05:14:42,030 --> 05:14:45,233 HOW DO WE BUILD ENERGY EFFICIENT COMPUTATION. 7237 05:14:45,233 --> 05:14:46,401 BUT I DON'T THINK WE'RE THERE YET. 7238 05:14:46,401 --> 05:14:46,901 I MEAN, 7239 05:14:46,901 --> 05:14:47,368 I THINK THERE WERE 7240 05:14:47,368 --> 05:14:49,704 A LOT OF INTERESTING TALKS, BUT I DON'T THINK 7241 05:14:49,704 --> 05:14:52,440 THAT THERE'S ANYTHING YET THAT WE'RE READY TO SCALE UP. 7242 05:14:52,440 --> 05:14:56,678 AND I THINK ONE OF THE MAJOR CONSENSUSES FROM SORT OF, 7243 05:14:57,278 --> 05:14:58,279 YOU KNOW, THE, 7244 05:14:58,279 --> 05:15:01,282 THE VARIOUS SESSIONS WAS THAT THAT IS ONE, YOU KNOW, ENERGY, 7245 05:15:02,117 --> 05:15:05,754 THE INEFFICIENCY OF -- SORT OF THE CURRENT PARADIGM 7246 05:15:06,020 --> 05:15:08,289 IS GOING TO BE ONE OF THE BIG PROBLEMS THAT WE FACE. 7247 05:15:08,289 --> 05:15:09,357 AND I DON'T THINK 7248 05:15:09,357 --> 05:15:11,893 THE SOLUTION TO THAT IS PUT TOGETHER ONE BIG TEAM. 7249 05:15:11,893 --> 05:15:15,597 THE SOLUTION, I THINK, IS TO PUT OUT CALLS FOR MANY, 7250 05:15:15,597 --> 05:15:18,700 MANY TEAMS TO USE AS MANY APPROACHES 7251 05:15:18,700 --> 05:15:21,803 AS WE CAN TO FIGURE OUT WHAT THE SOLUTION IS. 7252 05:15:21,803 --> 05:15:24,539 AND I THINK, YOU KNOW, WE CAN IDENTIFY 7253 05:15:24,539 --> 05:15:25,473 A BUNCH OF SORT OF 7254 05:15:25,473 --> 05:15:26,708 SIMILAR PROBLEMS 7255 05:15:26,708 --> 05:15:30,445 THAT WERE TOUCHED ON AT THIS MEETING THAT WOULD BENEFIT FROM, 7256 05:15:30,445 --> 05:15:33,448 I THINK, THE SORT OF SMALLER HOPE, 7257 05:15:34,149 --> 05:15:38,153 NOT EVEN HYPOTHESIS DRIVEN, BUT BUT SORT OF ARTISANAL SCIENCE, 7258 05:15:38,153 --> 05:15:40,388 YOU KNOW, LET A THOUSAND FLOWERS BLOOM 7259 05:15:40,388 --> 05:15:42,157 APPROACH THAT I THINK YOU'RE ADVOCATING. 7260 05:15:42,157 --> 05:15:45,293 SO I THINK IT DEPENDS ON WHAT THE, 7261 05:15:45,727 --> 05:15:48,897 WHAT STAGE WE'RE AT RIGHT NOW FOR A GIVEN PROJECT, 7262 05:15:48,897 --> 05:15:52,233 OR A GIVEN SORT OF SET OF QUESTIONS, THAT WOULD BE MY 7263 05:15:52,967 --> 05:15:53,568 TAKE ON IT 7264 05:15:55,103 --> 05:15:55,436 AND GO, 7265 05:15:55,436 --> 05:15:58,439 YOU CAN GO AHEAD 7266 05:15:59,073 --> 05:16:01,176 SO YEAH. 7267 05:16:01,176 --> 05:16:04,846 SO I JUST LIKE TO POINT OUT ONE THING, TONY, WHICH, 7268 05:16:05,814 --> 05:16:06,848 OF COURSE, 7269 05:16:06,848 --> 05:16:08,983 I DON'T THINK THAT ANDREAS WAS EITHER SAYING, LIKE, 7270 05:16:08,983 --> 05:16:10,685 WE SHOULD SUCK ALL THE AIR OUT OF THE ROOM 7271 05:16:10,685 --> 05:16:12,921 AND JUST GIVING MONEY TO HIM. 7272 05:16:12,921 --> 05:16:14,189 I CAN ASSURE YOU. 7273 05:16:14,189 --> 05:16:15,623 HE MIGHT BE. HE MIGHT BE SAYING THAT. 7274 05:16:15,623 --> 05:16:16,591 YOU MIGHT BE SAYING. 7275 05:16:16,591 --> 05:16:18,193 ALL RIGHT, 7276 05:16:18,193 --> 05:16:22,130 BUT, YOU KNOW, ONE EXAMPLE I SHOWED IN, IN MY SLIDES, 7277 05:16:22,463 --> 05:16:26,434 YESTERDAY WAS, THIS LARGE LANGUAGE MODEL LLAMA. 7278 05:16:26,601 --> 05:16:26,868 RIGHT. 7279 05:16:26,868 --> 05:16:28,770 AND THAT'S AN ARTIFACT THAT WAS CREATED 7280 05:16:28,770 --> 05:16:31,639 BY A VERY COORDINATED RESEARCH TEAM. 7281 05:16:31,639 --> 05:16:35,310 THERE'S 500 PEOPLE ON THAT PAPER, AND IT'S BEEN CITED, 7282 05:16:35,543 --> 05:16:37,312 YOU KNOW, 10,000 TIMES IN A YEAR. 7283 05:16:37,312 --> 05:16:40,982 IT'S A BIG ARTIFACT THAT IT'S USED BY A LOT OF PEOPLE. AND, 7284 05:16:41,950 --> 05:16:45,119 IT IS LIKE, WELL, BEYOND THE CAPACITIES 7285 05:16:45,153 --> 05:16:48,156 OF A SINGLE LAB THAT'S SUPPORTED BY A FEW ARE R01S. 7286 05:16:48,323 --> 05:16:48,623 RIGHT. 7287 05:16:48,623 --> 05:16:52,126 SO I THINK IN THAT SENSE, IF WE LOOK AT 7288 05:16:52,126 --> 05:16:54,128 WHETHER IT'S THE LARGE HADRON COLLIDER, 7289 05:16:54,128 --> 05:16:57,298 WHETHER IT'S, THE, THE HUMAN GENOME PROJECT 7290 05:16:57,665 --> 05:17:01,169 OR WHETHER IT'S THESE NEWFANGLED MODELS THAT ARE CREATED 7291 05:17:01,169 --> 05:17:05,340 BY THESE, FOUNDATIONAL TEAMS INSIDE OF BIG TECH, 7292 05:17:05,573 --> 05:17:08,209 IT REQUIRES A LOT OF ENGINEERING AND A COORDINATED EFFORT, 7293 05:17:08,209 --> 05:17:11,212 WHICH DOESN'T MEAN THAT YOU HAVE TO SUCK THE AIR 7294 05:17:11,212 --> 05:17:14,515 OUT OF THE ROOM AND 7295 05:17:14,916 --> 05:17:18,152 AND STOP SUPPORTING FUNDAMENTAL AND BASIC SCIENCE. 7296 05:17:18,887 --> 05:17:21,890 IT CAN BE BOTH SPIKE AND SLAB, AS I LIKE TO CALL IT. 7297 05:17:24,325 --> 05:17:26,327 PANAYIOTA. 7298 05:17:26,327 --> 05:17:27,228 HI, SO ACTUALLY MY COMMENT 7299 05:17:27,228 --> 05:17:30,231 WAS FOR THE PREVIOUS QUESTION, BUT MAYBE I'LL BRING IT BACK. 7300 05:17:30,331 --> 05:17:34,235 HAVING TO DO WITH, WHAT JENNIFER SAID ABOUT, YOU KNOW, BUILDING 7301 05:17:34,235 --> 05:17:39,040 THE CORTEX IN A HIGH LEVEL DETAIL ONTO, HARDWARE AND, 7302 05:17:39,507 --> 05:17:42,844 WONDERING WHETHER THERE IS A REALLY BIG RISK THERE. 7303 05:17:44,212 --> 05:17:47,548 SIMILAR TO WHAT THE HUMAN BRAIN PROJECT WAS TRYING TO DO 7304 05:17:47,548 --> 05:17:50,718 BY SIMULATING, LET'S SAY, A VERY BIG CHUNK OF THE CORTEX 7305 05:17:51,085 --> 05:17:53,388 AND IN MY VIEW, AND THE VIEW OF MANY OTHERS 7306 05:17:53,388 --> 05:17:55,023 ACTUALLY FAILED TO DELIVER, 7307 05:17:55,023 --> 05:17:58,226 BECAUSE IN THE END, YOU CAME UP WITH A MODEL 7308 05:17:58,426 --> 05:18:00,895 THAT WAS PRETTY MUCH AS COMPLEX AS THE REAL THING. 7309 05:18:00,895 --> 05:18:03,264 AND IT WAS VERY HARD TO UNDERSTAND 7310 05:18:03,264 --> 05:18:06,567 WHAT THE SIMULATED COLUMN IS ACTUALLY COMPUTING OR DOING. 7311 05:18:06,968 --> 05:18:09,804 AND I THINK IT IS IMPORTANT TO KEEP IN MIND 7312 05:18:09,804 --> 05:18:11,639 THAT THERE IS ALWAYS A TRADE OFF 7313 05:18:11,639 --> 05:18:14,642 BETWEEN A HIGH LEVEL OF COMPLEXITY 7314 05:18:14,642 --> 05:18:18,546 AND AN EASY INTERPRETABILITY OR READOUT OF THE FUNCTION, 7315 05:18:18,813 --> 05:18:20,281 AND ESPECIALLY WHEN WE WANT 7316 05:18:20,281 --> 05:18:23,718 TO BRING THE NEUROAI, LET'S SAY, INTO THE GAME. 7317 05:18:24,218 --> 05:18:27,488 I THINK WHAT IS REALLY MISSING, AND IT IS IMPORTANT HERE, 7318 05:18:27,755 --> 05:18:33,328 IS TO HAVE A COORDINATED EFFORT THAT AIMS AT EXTRACTING THE KEY 7319 05:18:33,494 --> 05:18:38,599 COMPONENTS OF BIOLOGY THAT ARE IMPORTANT FOR BUILDING, 7320 05:18:38,800 --> 05:18:42,270 LET'S SAY, MORE EFFICIENT SYSTEMS AS, AS TONY PUT IT OUT. 7321 05:18:42,637 --> 05:18:45,106 SO I THINK WE, WE REALLY MISS, 7322 05:18:46,074 --> 05:18:48,343 LET'S SAY A THEORETICAL COMPONENT HERE 7323 05:18:48,343 --> 05:18:49,811 THAT TALKS TO THE BIOLOGIST, 7324 05:18:49,811 --> 05:18:53,281 TALKS TO THE HARDWARE EXPERTS, AND TRIES TO FIGURE OUT 7325 05:18:53,581 --> 05:18:57,318 WHICH ARE THE IMPORTANT ELEMENTS THAT WE WANT TO KEEP 7326 05:18:57,618 --> 05:18:59,520 TO CREATE SUCH EFFICIENT SYSTEMS. 7327 05:18:59,520 --> 05:19:02,523 AND TO MY KNOWLEDGE, THERE IS NO SUCH COORDINATED, 7328 05:19:02,657 --> 05:19:05,693 APPROACH ANYWHERE IN EUROPE OR IN THE US. 7329 05:19:05,693 --> 05:19:07,962 AND, YOU KNOW, DEFINING THE MISSING ELEMENTS 7330 05:19:07,962 --> 05:19:10,965 WOULD BE REALLY, YOU KNOW, USEFUL AND IMPORTANT. 7331 05:19:13,634 --> 05:19:16,104 YES. SO THANK YOU, BY THE WAY, 7332 05:19:16,104 --> 05:19:17,305 THANK YOU FOR BRINGING ALL THAT UP. 7333 05:19:17,305 --> 05:19:18,206 THAT WAS VERY HELPFUL. 7334 05:19:18,206 --> 05:19:20,208 AND ALSO THE HISTORY, 7335 05:19:20,208 --> 05:19:23,811 I THINK WHEN I WAS TRYING TO GIVE MY SHORT ANSWER, PART OF 7336 05:19:24,512 --> 05:19:26,047 I ACTUALLY REALLY APPRECIATE THAT 7337 05:19:26,047 --> 05:19:27,281 BECAUSE I THINK THE THEORETICAL 7338 05:19:27,281 --> 05:19:29,650 FRAMEWORK OF THAT AND THE GROUNDING OF THE CORE 7339 05:19:29,650 --> 05:19:31,719 CONCEPTS IS ESSENTIAL. 7340 05:19:31,719 --> 05:19:35,056 AND I THINK PART OF THE REASON I DROVE 7341 05:19:35,056 --> 05:19:36,457 THAT IS SAYING, 7342 05:19:36,457 --> 05:19:39,527 LET'S SEE, LET'S ACTUALLY MAKE THIS BE USEFUL FOR 7343 05:19:39,527 --> 05:19:40,294 SOMETHING THAT'S. 7344 05:19:40,294 --> 05:19:41,496 ENGINEERING APPLICATION 7345 05:19:41,496 --> 05:19:44,632 THAT WOULD FORCE YOU TO DO EXACTLY THE THINGS YOU'RE SAYING 7346 05:19:45,133 --> 05:19:46,234 SO IMPLICITLY. 7347 05:19:46,234 --> 05:19:51,406 I STRONGLY AGREE, AND I'M GLAD YOU BROUGHT IT UP AGAIN. 7348 05:19:51,406 --> 05:19:52,373 I THINK THAT WOULD BE 7349 05:19:52,373 --> 05:19:54,809 A FUNDAMENTAL PART OF THAT KIND OF EFFORT, 7350 05:19:55,877 --> 05:19:58,312 WHETHER DONE AS A LARGE EFFORT OR A SMALL EFFORT. 7351 05:19:58,312 --> 05:20:01,249 BUT THANK YOU. 7352 05:20:01,249 --> 05:20:03,184 YEAH. 7353 05:20:03,184 --> 05:20:05,386 ZACH, SORRY ABOUT THE WAY 7354 05:20:05,386 --> 05:20:06,988 THAT WORKS. 7355 05:20:06,988 --> 05:20:09,991 I, I AGREE PANAYIOTA WITH MANY OF THE, THAT 7356 05:20:09,991 --> 05:20:12,727 INVESTING IN SOME PARTICULAR HARDWARE IS PREMATURE 7357 05:20:12,727 --> 05:20:15,863 BEFORE YOU HAVE SOME CHEAPER SIMULATIONS OF THINGS 7358 05:20:15,863 --> 05:20:17,098 THAT YOU CAN VALIDATE. 7359 05:20:17,098 --> 05:20:18,166 BUT THE THING THAT I THINK 7360 05:20:18,166 --> 05:20:21,169 WE REALLY, REALLY NEED IS TO UNDERSTAND LEARNING. 7361 05:20:21,202 --> 05:20:23,571 WE DON'T UNDERSTAND HOW WE LEARN. 7362 05:20:23,571 --> 05:20:25,673 WE DON'T HAVE THE TECHNOLOGICAL TOOLS 7363 05:20:25,673 --> 05:20:28,876 TO MEASURE THE MECHANISMS OF LEARNING IN THE BRAIN. 7364 05:20:28,876 --> 05:20:31,779 AND WHEN WE LOOK AT WHAT DEEP LEARNING IS DOING, 7365 05:20:31,779 --> 05:20:34,782 IT'S REALLY SUCCESSFUL, BUT IT'S NOT REALLY EFFICIENT. 7366 05:20:35,683 --> 05:20:38,786 IF YOU LOOK AT THE USUAL GRADIENT DESCENT RULE, 7367 05:20:39,620 --> 05:20:42,190 WE KNOW THAT IT JUST BOBS ALONG 7368 05:20:42,190 --> 05:20:45,126 ON THESE LONG VALLEYS WHEN IT COULD USE A BETTER RULE. 7369 05:20:45,693 --> 05:20:47,061 AND THERE ARE SOME ALTERNATIVES 7370 05:20:47,061 --> 05:20:48,896 TO STRAIGHTFORWARD GRADIENT, RIGHT. 7371 05:20:48,896 --> 05:20:50,731 NATURAL GRADIENT AND THINGS LIKE THIS. 7372 05:20:50,731 --> 05:20:53,735 AND MAYBE SOME OF THOSE TRICKS ARE THINGS THAT THE BRAIN DOES. 7373 05:20:53,735 --> 05:20:57,872 IT'S ABLE TO MAKE FASTER LEAPS THAN FIDDLING ALONG 7374 05:20:57,872 --> 05:21:01,275 WITH LITTLE INCREMENTAL CHANGES AND GIVING US EUREKA MOMENTS. 7375 05:21:01,275 --> 05:21:03,711 SO I FEEL LIKE THAT IS A GOOD TEN YEAR PROJECT, 7376 05:21:03,711 --> 05:21:06,414 BRINGING THE TECHNOLOGY TO MEASURE SYNAPTIC CHANGES 7377 05:21:06,414 --> 05:21:09,083 AND NEUROMODULATION ON A LARGE SCALE 7378 05:21:09,083 --> 05:21:11,252 THAT WOULD ALLOW US TO UNDERSTAND EUREKA MOMENTS. 7379 05:21:11,252 --> 05:21:12,487 AND OF COURSE, THE CHALLENGES 7380 05:21:12,487 --> 05:21:15,490 THAT YOU ONLY GET ONE CHANCE TO MAKE A FIRST IMPRESSION. 7381 05:21:15,623 --> 05:21:17,325 SO EVERY TIME YOU'RE STUDYING 7382 05:21:17,325 --> 05:21:20,294 A EUREKA MOMENT, YOU CAN ONLY STUDY IT ONCE. 7383 05:21:20,328 --> 05:21:22,230 SO WE NEED TO FIND SOME PARADIGMS 7384 05:21:22,230 --> 05:21:25,766 WHICH ALLOW US TO GET ANIMALS TO DO LOTS OF EUREKA MOMENTS. 7385 05:21:25,967 --> 05:21:28,636 AND THAT'S HARD, BUT I THINK IT'S FEASIBLE. 7386 05:21:28,636 --> 05:21:30,338 SO YOU'RE ECHOING WHAT BLAKE RICHARD SAID YESTERDAY. 7387 05:21:30,338 --> 05:21:32,807 YES, EXACTLY WITH BLAKE, TOTALLY ON THE SAME PAGE. 7388 05:21:32,807 --> 05:21:34,175 DO YOU THINK MEASURING 7389 05:21:34,175 --> 05:21:38,146 SYNAPTIC STRENGTH ACROSS ALL CONNECTIONS IS THE SOLUTION? 7390 05:21:38,613 --> 05:21:38,846 WELL, 7391 05:21:38,846 --> 05:21:40,314 I THINK YOU ALSO NEED THE MODULATORS, 7392 05:21:40,314 --> 05:21:41,415 BECAUSE I DON'T THINK THAT 7393 05:21:41,415 --> 05:21:43,818 THE ONLY EFFECT IS JUST PURELY LOCAL. 7394 05:21:43,818 --> 05:21:46,821 I THINK THAT THAT WOULD MISS A WHOLE LOT OF STRUCTURE. 7395 05:21:47,188 --> 05:21:49,423 IS THAT DOABLE IN TEN YEARS, YOU THINK? 7396 05:21:49,423 --> 05:21:51,159 DEFINITELY. MAYBE NINE. 7397 05:21:55,096 --> 05:21:56,931 OKAY, SO ACROSS 10 MILLION NEURONS. 7398 05:21:56,931 --> 05:21:58,666 YEAH. 7399 05:21:58,666 --> 05:22:00,201 YEAH, 10,000 SYNAPSES. 7400 05:22:00,201 --> 05:22:02,503 SO IF YOU WANT TO, LIKE GET ALL OF THEM, 7401 05:22:02,503 --> 05:22:06,641 WE JUST NEED TO SAMPLE FROM WHAT'S LIKE A TRILLION SYNAPSES. 7402 05:22:07,208 --> 05:22:09,143 SURE. 7403 05:22:09,143 --> 05:22:11,646 THAT SOUNDS -- I JUST 7404 05:22:11,646 --> 05:22:14,649 THAT, IT SOUNDS SUPER EXCITING, 7405 05:22:14,849 --> 05:22:17,885 TO BE ABLE TO TRACK LEARNING, BUT AGAIN, IT 7406 05:22:17,885 --> 05:22:18,519 DEPENDS ON 7407 05:22:18,519 --> 05:22:22,190 TECHNOLOGICAL DEVELOPMENTS, THE ABILITY TO MEASURE THINGS 7408 05:22:22,190 --> 05:22:22,857 AT THE, 7409 05:22:22,857 --> 05:22:25,860 SUBCELLULAR LEVEL, TO MEASURE THE SYNAPTIC STRENGTH, 7410 05:22:27,128 --> 05:22:30,831 TO BE ABLE TO TRACK ANIMALS OVER THAT EXTENDED AMOUNT OF TIME. 7411 05:22:31,132 --> 05:22:34,135 MEASURING SYNAPTIC STRENGTH IS NOT TRACKING LEARNING, THOUGH, 7412 05:22:34,735 --> 05:22:35,002 RIGHT? 7413 05:22:35,002 --> 05:22:36,971 THAT'S WHERE THE THEORY COMES IN. 7414 05:22:36,971 --> 05:22:39,207 IF YOU JUST MEASURE SOMETHING, IT DOESN'T, 7415 05:22:39,207 --> 05:22:41,709 WITHOUT A THEORETICAL FRAMEWORK, IT DOESN'T NECESSARILY TELL 7416 05:22:41,709 --> 05:22:44,712 YOU ANYTHING ABOUT THE PROCESS THAT'S HAPPENING. 7417 05:22:45,046 --> 05:22:45,580 RIGHT. 7418 05:22:45,580 --> 05:22:46,981 BUT I MEAN, THE FIRST ORDER 7419 05:22:46,981 --> 05:22:50,351 YOU WOULD EXPECT TO SEE THE LEARNING IN THE SYNAPSES. 7420 05:22:50,351 --> 05:22:52,687 AND THAT'S SAYING THAT THERE'S NOT OTHER STUFF GOING ON THAT, 7421 05:22:55,089 --> 05:22:57,058 BY THE WAY, I SUPPORT THAT IDEA. 7422 05:22:57,058 --> 05:22:59,060 I'M JUST PLAYING DEVIL'S ADVOCATE. 7423 05:22:59,060 --> 05:22:59,393 YEAH. 7424 05:22:59,393 --> 05:23:02,897 BUT THAT BRINGS UP THE IDEA OF, OF DOING CHRONIC RECORDINGS. 7425 05:23:02,897 --> 05:23:03,264 RIGHT. 7426 05:23:03,264 --> 05:23:05,199 AND, AND BRINGING UP NEW TECHNOLOGIES 7427 05:23:05,199 --> 05:23:07,068 IN ORDER TO CREATE CHRONIC RECORDINGS. 7428 05:23:07,068 --> 05:23:10,071 AND DOING THAT REQUIRES 7429 05:23:10,204 --> 05:23:13,207 VERY SIGNIFICANT ENGINEERING. 7430 05:23:14,709 --> 05:23:16,677 GO AHEAD. 7431 05:23:16,677 --> 05:23:19,580 SO, AS WE'RE TALKING ABOUT 7432 05:23:19,580 --> 05:23:23,284 GATHERING MORE AND MORE DATA, YOU KNOW, YOU'RE DOING THE FLY 7433 05:23:23,284 --> 05:23:25,253 CONNECTOME, AND THEN THE MOUSE CONNECTOME 7434 05:23:25,253 --> 05:23:26,420 SOON, 7435 05:23:26,420 --> 05:23:30,057 THESE HUGE INITIATIVES, ALL OF ALL OF WHICH I FULLY SUPPORT. 7436 05:23:31,025 --> 05:23:35,396 BUT I'M CURIOUS HOW PEOPLE THINK ABOUT C ELEGANS, 7437 05:23:36,063 --> 05:23:38,733 BECAUSE WE KNOW ALL 302 OF THOSE NEURONS 7438 05:23:38,733 --> 05:23:41,736 AND WE CAN'T BUILD IT YET. 7439 05:23:45,906 --> 05:23:48,309 THERE ARE SIMULATED ONES. 7440 05:23:48,309 --> 05:23:49,277 YEAH, CERTAINLY. 7441 05:23:49,277 --> 05:23:50,978 THERE'S OPEN WORM. 7442 05:23:50,978 --> 05:23:51,913 YEAH. 7443 05:23:51,913 --> 05:23:54,615 THAT'S IN -- THAT'S ON A COMPUTER. 7444 05:23:54,615 --> 05:23:56,317 WE HAVEN'T BUILT IT. 7445 05:23:56,317 --> 05:23:58,452 RIGHT. WE HAVEN'T EVEN BUILT THE SCALED MODEL OF IT. 7446 05:24:12,066 --> 05:24:14,935 WIRELESS MEANING? 7447 05:24:14,935 --> 05:24:16,971 SO TO REPEAT FOR THE AUDIENCE WHO MAY NOT BE ABLE TO 7448 05:24:16,971 --> 05:24:19,974 HEAR, HE'S SAYING WE DON'T KNOW THE WIRELESS DIAGRAM 7449 05:24:20,241 --> 05:24:20,641 WHICH MEANS 7450 05:24:20,641 --> 05:24:23,644 BUT THERE ARE. 7451 05:24:24,478 --> 05:24:27,481 NEUROTRANSMITTERS THAT ARE NOT GOING THROUGH SYNAPSES. 7452 05:24:27,682 --> 05:24:28,916 YEAH THAT'S TRUE. 7453 05:24:28,916 --> 05:24:31,085 SO HOW DO WE THINK ABOUT THAT. 7454 05:24:31,085 --> 05:24:32,453 WHY DO WE 7455 05:24:32,453 --> 05:24:35,656 WE WANT TO THINK ABOUT TRYING TO SOLVE THAT? 7456 05:24:35,656 --> 05:24:38,659 DO WE THINK IF THAT'S A PROBLEM FOR 302 NEURONS, 7457 05:24:39,160 --> 05:24:40,294 DO YOU THINK THAT'LL BE ONLY 7458 05:24:40,294 --> 05:24:43,297 AN ESCALATED PROBLEM FOR LARGER NUMBERS? 7459 05:24:46,767 --> 05:24:49,036 RIGHT. 7460 05:24:49,036 --> 05:24:50,771 SO DO YOU THINK THAT SPIKES REPLACE 7461 05:24:50,771 --> 05:24:53,708 THAT COMPLETELY? 7462 05:24:55,276 --> 05:24:56,477 THEY USE GRADED TRANSMISSION. 7463 05:24:56,477 --> 05:24:58,979 BUT I BELIEVE THAT'S FOUND IN MAMMALS SOMETIMES. 7464 05:24:58,979 --> 05:25:00,614 IT'S NOT A DOMINANT WAY. 7465 05:25:00,614 --> 05:25:02,116 IT'S NOT THE DOMINANT WAY. 7466 05:25:02,116 --> 05:25:03,851 IT'S TRUE. 7467 05:25:03,851 --> 05:25:04,852 I'VE JUST. 7468 05:25:04,852 --> 05:25:07,488 HEARD IT THAT. 7469 05:25:07,488 --> 05:25:09,323 PEOPLE WHO WORK ON A DAILY RECORD FOR THEIR. 7470 05:25:14,128 --> 05:25:16,530 DOES ANYONE HAVE C. ELEGANS THOUGHTS? 7471 05:25:16,530 --> 05:25:20,134 I MEAN, THERE ARE PEOPLE WORKING ON SIMULATING C ELEGANS, 7472 05:25:21,302 --> 05:25:24,005 BUT I TAKE YOUR POINT ABOUT THE BODY AS WELL. 7473 05:25:24,005 --> 05:25:24,472 I MEAN, 7474 05:25:24,472 --> 05:25:25,673 I THINK THAT'S QUITE IMPORTANT 7475 05:25:25,673 --> 05:25:29,977 BECAUSE I WAS IMPRESSED WITH HOW MUCH EMBODIMENT WAS DISCUSSED. 7476 05:25:30,177 --> 05:25:32,079 I'M REALLY NOT JUST TRYING TO BEAT THE EMBODIMENT DRUM. 7477 05:25:32,079 --> 05:25:33,314 I'M TRYING TO SAY THAT 7478 05:25:33,314 --> 05:25:34,048 THERE PROBABLY 7479 05:25:34,048 --> 05:25:37,084 IS THE WIRELESS SYSTEM OR, YOU KNOW, THESE NEUROMODULATORS 7480 05:25:37,084 --> 05:25:40,087 THAT ARE AT PLAY IN INVERTEBRATE SYSTEMS ARE THAT 7481 05:25:40,087 --> 05:25:42,857 SEEM ALSO LIKELY TO BE IN PLAY IN MAMMALIAN SYSTEMS. 7482 05:25:42,857 --> 05:25:46,327 THAT MAY COMPLICATE MATTERS TO A DEGREE THAT WE DON'T YET 7483 05:25:46,327 --> 05:25:48,396 AND AREN'T YET ABLE TO ANTICIPATE. 7484 05:25:48,396 --> 05:25:50,264 SO I'LL JUST CAVEAT. 7485 05:25:50,264 --> 05:25:52,600 CAN I ADDRESS THAT? THAT'S A GREAT POINT. 7486 05:25:52,600 --> 05:25:54,535 AND INTERESTINGLY, MAYBE IRONICALLY, 7487 05:25:54,535 --> 05:25:57,538 THERE WAS A STUDY PUBLISHED LAST FEW MONTHS THAT ACTUALLY USED 7488 05:25:58,539 --> 05:26:02,276 A.I. ANALYSIS OF GENE EXPRESSION STUDIES TO FIGURE OUT THE LIKELY 7489 05:26:02,276 --> 05:26:05,279 TARGETS OF THESE NEUROMODULATORS THAT ARE ACTING AT A DISTANCE. 7490 05:26:05,913 --> 05:26:09,250 SO I THINK THE THE ROLE OF NEUROMODULATION 7491 05:26:09,250 --> 05:26:10,217 IS GOING TO BE CRITICAL. 7492 05:26:10,217 --> 05:26:13,120 IT'S NOT ALL GOING TO BE IN THE STATIC 7493 05:26:13,120 --> 05:26:16,023 CHEMICAL SYNAPSE WIRING DIAGRAM. 7494 05:26:16,023 --> 05:26:17,758 BUT THAT'S NOT TO DISCOUNT THE 7495 05:26:17,758 --> 05:26:20,261 THE POTENTIAL POWER OF THESE SYSTEMS. 7496 05:26:20,261 --> 05:26:21,962 IF YOU, AS DORIS SAID, 7497 05:26:21,962 --> 05:26:23,731 USE YOUR INTUITION TO GET CREATIVE ABOUT IT. 7498 05:26:23,731 --> 05:26:25,766 I THINK WE CAN ACTUALLY GET PRETTY FAR WITH IT. 7499 05:26:25,766 --> 05:26:26,700 SO IT'S A GREAT POINT 7500 05:26:26,700 --> 05:26:29,703 AND ONE THAT I THINK WE DISCUSSED OFFLINE A BIT. 7501 05:26:29,703 --> 05:26:32,039 BUT THERE ARE APPROACHES TO KIND OF GET AT THIS PROBLEM. 7502 05:26:32,039 --> 05:26:32,373 SO, THEY'RE NOT 7503 05:26:32,373 --> 05:26:34,742 INSOLUBLE. AND THAT COULD BE QUITE 7504 05:26:34,742 --> 05:26:37,745 INSTRUCTIVE IF WE PUT OUR MINDS TO IT. 7505 05:26:40,981 --> 05:26:42,817 TO KIND OF PIGGYBACK ON THAT, 7506 05:26:42,817 --> 05:26:44,718 BUT KIND OF GO TO THE OTHER EXTREME. 7507 05:26:44,718 --> 05:26:44,952 RIGHT. 7508 05:26:44,952 --> 05:26:49,123 SO WE'RE TALKING ABOUT DOING THESE KIND OF VERY LARGE SCALE 7509 05:26:49,123 --> 05:26:52,259 NEURAL ANALYSIS, LARGER SCALE CONNECTOME 7510 05:26:52,259 --> 05:26:55,329 OVER MULTIPLE INDIVIDUALS, RECORDING 7511 05:26:55,329 --> 05:26:58,332 FROM LOTS OF NEURONS OVER LONG PERIODS OF TIME. 7512 05:26:58,833 --> 05:27:01,168 WE ALREADY HAVE A BIG DATA PROBLEM. 7513 05:27:01,168 --> 05:27:04,171 AND THIS IS AN EVEN BIGGER DATA PROBLEM. 7514 05:27:04,238 --> 05:27:04,505 RIGHT? 7515 05:27:04,505 --> 05:27:06,240 I REMEMBER AT EITHER 7516 05:27:06,240 --> 05:27:07,741 THE LAST BRAIN INITIATIVE MEETING 7517 05:27:07,741 --> 05:27:09,977 OR THE ONE BEFORE THAT, SOMEONE 7518 05:27:09,977 --> 05:27:12,446 I THINK IT WAS DIRECTOR NGAI SAID THAT, YOU KNOW, 7519 05:27:12,446 --> 05:27:13,814 IF YOU LOOK AT 7520 05:27:13,814 --> 05:27:16,617 TRYING TO MAKE A CONNECTOME OF SOME OF THESE BIGGER 7521 05:27:16,617 --> 05:27:19,820 MAMMALIAN SYSTEMS, RIGHT, LIKE MOUSE, RAT GOING UP, PROBABLY 7522 05:27:19,820 --> 05:27:20,654 I THINK IT WAS THE RAT. 7523 05:27:21,689 --> 05:27:22,122 IT IS GOING TO 7524 05:27:22,122 --> 05:27:25,159 GENERATE MORE DATA THAN HAS BEEN GENERATED 7525 05:27:25,159 --> 05:27:28,329 BY HUMANITY SINCE THE STONE AGE, RIGHT? 7526 05:27:28,329 --> 05:27:31,532 THIS IS IT'S AN ORDER OF INCREASE IN DATA COMPLEXITY. 7527 05:27:31,532 --> 05:27:34,535 SO I FEEL LIKE A BIG PORTION OF HOW WE MOVE 7528 05:27:34,535 --> 05:27:36,170 INTO THIS NEXT PHASE OF NEUROSCIENCE 7529 05:27:36,170 --> 05:27:38,305 IS EVEN DEVELOPING EVEN MORE TOOLS 7530 05:27:38,305 --> 05:27:41,475 AND METHODOLOGIES AND THEORIES ABOUT HOW YOU ANALYZE 7531 05:27:41,475 --> 05:27:45,212 THESE BIG, EVEN BIGGER ASPECTS OF DATA, I GUESS. 7532 05:27:45,212 --> 05:27:46,514 DOES ANYBODY HAVE THOUGHTS 7533 05:27:46,514 --> 05:27:49,517 ABOUT KIND OF THE LARGER EXPLOSION OF DATA? 7534 05:27:51,285 --> 05:27:53,020 YEAH, I MEAN, I THINK THE DATA THING IS 7535 05:27:53,020 --> 05:27:55,189 SOMETHING THAT IS GOING TO GROW AND WE'RE GOING TO HAVE TO 7536 05:27:55,189 --> 05:27:57,725 DEDICATE ENGINEERING RESOURCES TOO, NO QUESTION. 7537 05:27:57,725 --> 05:28:00,461 AND ALONG THE LINES OF SOFTWARE ENGINEERING, 7538 05:28:00,461 --> 05:28:01,462 THE SOFTWARE ENGINEERING DISCUSSION 7539 05:28:01,462 --> 05:28:04,498 WE ALREADY HAD, I, I JUST WANT TO MAKE ONE COMMENT 7540 05:28:04,498 --> 05:28:07,501 ABOUT THE C ELEGANS, POINT, WHICH I THINK IS GREAT. 7541 05:28:07,501 --> 05:28:07,701 RIGHT? 7542 05:28:07,701 --> 05:28:10,804 WE KNOW THINGS ABOUT THE C ELEGANS WIRING 7543 05:28:10,804 --> 05:28:14,208 AND WHAT YOU KNOW, IT'S A FAIR QUESTION TO ASK WHAT WE HAVE, 7544 05:28:14,808 --> 05:28:17,044 WHAT KIND OF PROPERTIES WE HAVE LEARNED ABOUT THE, 7545 05:28:17,044 --> 05:28:19,246 YOU KNOW, THE HUMAN BRAIN, LET'S SAY, FROM THAT. 7546 05:28:20,514 --> 05:28:23,083 AND I THINK IT JUST REALLY EMPHASIZES 7547 05:28:23,083 --> 05:28:26,820 THE TASK THAT THIS NEUROAI WORKSHOP AND CONCEPT 7548 05:28:26,820 --> 05:28:28,022 IS REALLY PUTTING FORWARD. RIGHT? 7549 05:28:28,022 --> 05:28:31,225 I MEAN, AS WE GO, LOOK AT WHAT'S HAPPENED WITH A.I., 7550 05:28:33,060 --> 05:28:35,529 WITH MACHINE LEARNING SYSTEMS OF ALL SORTS, 7551 05:28:35,529 --> 05:28:38,399 AS YOU GO TO MORE AND MORE AND MORE AND MORE PARAMETERS, 7552 05:28:38,399 --> 05:28:42,770 YOU GET EMERGENT PROPERTIES THAT LEAD TO THINGS LIKE, 7553 05:28:42,770 --> 05:28:44,071 YOU KNOW, MODERN TRANSFORMERS 7554 05:28:44,071 --> 05:28:46,006 AND REALLY LARGE FOUNDATION MODELS. 7555 05:28:46,006 --> 05:28:47,875 AND WE KNOW THAT AS YOU SCALE THEM UP, 7556 05:28:47,875 --> 05:28:50,878 THE CAPABILITIES INCREASE. 7557 05:28:50,878 --> 05:28:54,548 AND I THINK THAT GOING FROM THE STUDY OF THE C ELEGANS 7558 05:28:55,049 --> 05:28:56,517 NERVOUS SYSTEM 7559 05:28:56,517 --> 05:29:00,421 TO THE FLY WIRING DIAGRAM THAT WE'VE SEEN TO, YOU KNOW, 7560 05:29:00,421 --> 05:29:03,424 SMALL MAMMALS AND LARGER MAMMALS AND, 7561 05:29:03,624 --> 05:29:05,292 SCALING UP. I THINK THERE'S A VERY 7562 05:29:05,292 --> 05:29:07,227 EXCITING VISION OVER THE NEXT, 7563 05:29:07,227 --> 05:29:09,863 YOU KNOW, 10 OR 20 YEARS THAT WE ARE GOING TO TAKE 7564 05:29:09,863 --> 05:29:12,800 THESE DATA SETS AND UNDERSTAND EMERGENT PROPERTIES THAT ARISE 7565 05:29:12,800 --> 05:29:14,635 WHEN YOU START WIRING UP LARGE NUMBERS OF NEURONS 7566 05:29:14,635 --> 05:29:17,238 INTO NETWORKS, INTO SYSTEMS LIKE OUR BRAINS. 7567 05:29:17,238 --> 05:29:20,040 AND TO ME, THAT'S A REASON WHY THIS WHAT'S GOING ON 7568 05:29:20,040 --> 05:29:21,241 HERE IS SO EXCITING. REALLY? 7569 05:29:23,877 --> 05:29:25,412 MAYBE ONE LAST QUESTION. 7570 05:29:25,412 --> 05:29:28,248 WE HAVE ABOUT A MINUTE AND A HALF. 7571 05:29:28,248 --> 05:29:29,717 OKAY. 7572 05:29:29,717 --> 05:29:32,086 I'M THE LAST QUESTION. 7573 05:29:32,086 --> 05:29:32,419 OKAY. 7574 05:29:32,419 --> 05:29:35,222 SO BESIDES THE LEARNING, I THINK ANOTHER THING 7575 05:29:35,222 --> 05:29:38,459 THAT WE MAY NEED OR PAYING ATTENTION IS MEMORY. 7576 05:29:38,959 --> 05:29:44,098 SO, I ATTENDED, SFN, AND, IN THE NEUROSCIENCE FIELD, 7577 05:29:44,598 --> 05:29:50,070 IT'S NOT JUST LEARNING HOW BRAIN ENCODE THE INFORMATION 7578 05:29:50,070 --> 05:29:50,804 AND STORES THE 7579 05:29:50,804 --> 05:29:54,942 INFORMATION IN OUR NEURAL SYSTEM. IT'S STILL VERY IMPORTANT 7580 05:29:54,942 --> 05:29:58,846 QUESTIONS AND WE HAVE NOT ANSWERED THAT YET 7581 05:29:59,079 --> 05:30:02,282 IN THE NEUROSCIENCE PART. AND IN THE DEEP LEARNING 7582 05:30:02,282 --> 05:30:02,616 OKAY 7583 05:30:02,616 --> 05:30:03,751 SO BASICALLY PEOPLE 7584 05:30:03,751 --> 05:30:05,819 WORKING ON THE LEARNING, LEARNING AND LEARNING, 7585 05:30:05,819 --> 05:30:10,024 BUT HOW CAN USE THIS AS KIND OF THE TECHNIQUE 7586 05:30:10,024 --> 05:30:11,025 ON NEUROMORPHIC TECHNIQUE TOOLS TO 7587 05:30:11,025 --> 05:30:14,995 ANSWER HOW BRAIN STORES THE MEMORY. 7588 05:30:15,195 --> 05:30:17,498 AND ALSO I THINK THAT THESE ARE VERY IMPORTANT 7589 05:30:17,498 --> 05:30:20,834 BECAUSE MEMORY, IN MY OPINION, 7590 05:30:21,035 --> 05:30:21,935 MEMORY IS RELATED TO 7591 05:30:22,936 --> 05:30:23,771 LEARNING. 7592 05:30:23,771 --> 05:30:27,875 WHEN I TRY TO REMEMBER HOW I LEARN SOMETHING, BECAUSE I 7593 05:30:28,175 --> 05:30:32,379 LEARN AND I MEMORIZE SOMETHING, OKAY, LIKE A 1 PLUS 1 EQUALS 2. 7594 05:30:32,379 --> 05:30:34,682 AND, WHEN I 7595 05:30:34,682 --> 05:30:37,685 FACED A SIMILAR QUESTIONS THAT THEN I RECALL SOME MEMORY. 7596 05:30:38,018 --> 05:30:38,352 OKAY. 7597 05:30:38,352 --> 05:30:41,355 SO BASICALLY I THINK THAT'S HIGHLY RELATED TO EACH OTHER. 7598 05:30:41,555 --> 05:30:44,191 AND ONOTHER POINTS, THAT IS, 7599 05:30:44,191 --> 05:30:47,194 I THINK, FOR THE WHOLE DIGITAL SYSTEM, 7600 05:30:47,327 --> 05:30:50,230 WE ALREADY IN THIS KIND OF DIGITAL WORLD, 7601 05:30:50,230 --> 05:30:53,200 USING DIGITAL COMPUTERS FOR SO MANY YEARS. 7602 05:30:53,200 --> 05:30:55,302 AND THE PEOPLE USE THIS MEMORY 7603 05:30:55,302 --> 05:30:58,305 WHEN YOU TALK MEMORY WITH THE CIRCUIT DESIGN 7604 05:30:58,305 --> 05:31:01,542 GUYS AND THE NEUROSCIENCE GOT FUNDAMENTALLY DIFFERENT 7605 05:31:02,343 --> 05:31:04,144 CONCEPT. 7606 05:31:04,144 --> 05:31:07,047 SO WHEN WE TALK ABOUT THE MEMORY IN THE CIRCUIT 7607 05:31:07,047 --> 05:31:07,881 IS KIND OF 7608 05:31:07,881 --> 05:31:11,685 HOW DO WE STORE THE INFORMATION? 010101 USING DEVICE? 7609 05:31:12,052 --> 05:31:12,252 OKAY. 7610 05:31:12,252 --> 05:31:13,320 THAT'S HOW THE MEMORY 7611 05:31:13,320 --> 05:31:16,423 AND ARCHITECTURES KNOW HOW TO ACCESS THAT. 7612 05:31:17,191 --> 05:31:20,527 BUT IS THAT SAME CONCEPT IN THE NEUROSCIENCE, 7613 05:31:21,061 --> 05:31:24,732 IN THE NEUROMORPHIC COMPUTING? WE USE THE SAME CONCEPT? 7614 05:31:25,065 --> 05:31:26,567 IF YOU USE THE SAME CONCEPT, 7615 05:31:26,567 --> 05:31:29,169 ACTUALLY, THAT'S THE FUNDAMENTAL DIFFERENT. I DO NOT THINK, 7616 05:31:30,337 --> 05:31:30,671 THERE IS 7617 05:31:30,671 --> 05:31:33,607 SOME 0101 STORED IN THE NEURAL SYSTEM. 7618 05:31:33,774 --> 05:31:36,577 SO WHEN WE TALK ABOUT MEMORY, 7619 05:31:36,577 --> 05:31:40,781 MAYBE WE NEED TO THINK OUT OF THE BOX AT A LEAST 7620 05:31:40,781 --> 05:31:42,716 AND NOT, IT'S NOT 7621 05:31:42,716 --> 05:31:44,952 EQUIVALENT TO THE 0101 OR ANY 7622 05:31:44,952 --> 05:31:48,055 MEMORY DEVICES LIKE SRAM IN 7623 05:31:48,055 --> 05:31:50,290 THE DIGITAL SYSTEM. 7624 05:31:50,290 --> 05:31:51,725 SO THAT'S SOMETHING I WANT 7625 05:31:51,725 --> 05:31:52,426 TO PICK UP. 7626 05:31:52,426 --> 05:31:55,929 AND, ANOTHER THING IS I THINK ABOUT MEMORY IS ALSO 7627 05:31:56,930 --> 05:31:59,199 SOME KIND OF CONCEPT THAT WE NEED TO TAKE A PAY ATTENTION. 7628 05:31:59,199 --> 05:32:01,468 THANK YOU VERY MUCH. 7629 05:32:01,468 --> 05:32:03,971 ANYBODY WANT TO COMMENT ON THAT? 7630 05:32:03,971 --> 05:32:04,538 KAREN? I'M SORRY. 7631 05:32:04,538 --> 05:32:06,206 I WAS GOING TO 7632 05:32:06,206 --> 05:32:08,308 STEER US TOWARD ETHICAL CONSIDERATIONS AS WELL, 7633 05:32:08,308 --> 05:32:11,111 BUT I THINK WE NEED TO YIELD THE FLOOR TO, JOHN. 7634 05:32:11,111 --> 05:32:12,413 SO THANK YOU GUYS FOR 7635 05:32:12,413 --> 05:32:14,114 JOINING ME UP HERE IN A THOUGHTFUL DISCUSSION. 7636 05:32:21,622 --> 05:32:23,123 HI EVERYONE, SORRY FOR THE DELAY. 7637 05:32:23,123 --> 05:32:24,324 SO I WANT TO ANNOUNCE 7638 05:32:24,324 --> 05:32:27,327 THE WINNERS OF THE POSTER PRESENTATION AWARD. 7639 05:32:28,195 --> 05:32:32,132 FOR FIRST PLACE, WE HAVE 7640 05:32:32,866 --> 05:32:35,869 ADITYA NAIR, FROM CALTECH. 7641 05:32:37,204 --> 05:32:39,673 MACHINE LEARNING 7642 05:32:39,673 --> 05:32:42,643 GUIDED DISCOVERY OF AN 7643 05:32:51,151 --> 05:32:51,452 INTRINSIC 7644 05:32:51,452 --> 05:32:54,354 LINE 7645 05:32:54,354 --> 05:32:55,622 ATTRACTOR 7646 05:32:55,622 --> 05:32:58,625 DON'T FALL OFF THE STAGE 7647 05:33:03,030 --> 05:33:04,665 OH. REPEAT THE TITLE. 7648 05:33:04,665 --> 05:33:06,467 YEAH, SURE. 7649 05:33:06,467 --> 05:33:09,970 MACHINE LEARNING GUIDED DISCOVERY OF INTRINSIC LINE 7650 05:33:09,970 --> 05:33:11,638 ATTRACTOR 7651 05:33:16,476 --> 05:33:18,312 THAT'S ALL RIGHT. 7652 05:33:18,312 --> 05:33:20,013 NEXT WE HAVE WE HAVE. 7653 05:33:20,013 --> 05:33:23,016 THERE'S TWO RUNNER UPS. 7654 05:33:23,417 --> 05:33:26,186 TO XINHE ZHANG HARVARD UNIVERSITY, 7655 05:33:26,186 --> 05:33:29,723 DECODING BRAIN INTRINSIC DYNAMICS FOR NEUROAI. 7656 05:33:42,035 --> 05:33:45,038 YEAH. YOU. 7657 05:33:47,941 --> 05:33:48,275 RIGHT. 7658 05:33:48,275 --> 05:33:51,245 THAT'S NOT THE RIGHT ONE. 7659 05:33:52,112 --> 05:33:55,115 SO THERE'S A FEW. 7660 05:34:08,896 --> 05:34:11,131 AND THE SECOND RUNNER UP IS HARRISON ESPINO 7661 05:34:11,131 --> 05:34:14,268 FROM THE UNIVERSITY OF CALIFORNIA, IRVINE. 7662 05:34:15,102 --> 05:34:19,139 AND HIS POSTER WAS A RAPID ADAPTING AND CONTINUAL LEARNING, 7663 05:34:19,306 --> 05:34:22,943 SPIKING NEURAL NETWORK PATH PLANNING 7664 05:34:22,943 --> 05:34:25,946 ALGORITHM FOR MOBILE ROBOTS. 7665 05:34:39,293 --> 05:34:42,296 THANK YOU EVERYONE. 7666 05:34:44,698 --> 05:34:45,699 HERE WE GO. 7667 05:34:45,699 --> 05:34:46,033 THANK YOU. 7668 05:34:46,033 --> 05:34:48,201 THEY'RE GREAT. THANKS, EVERYBODY, FOR HANGING IN. 7669 05:34:48,201 --> 05:34:51,705 IT'S BEEN AN INTENSE, TWO DAYS, AND I THINK, 7670 05:34:53,006 --> 05:34:53,941 I HOPE YOU'LL ALL AGREE WITH ME. 7671 05:34:53,941 --> 05:34:56,944 THERE'S A HUGE AMOUNT OF EXCITEMENT IN THIS AREA, 7672 05:34:57,044 --> 05:35:00,180 AND I TRUST YOU BELIEVE THAT BRAIN INITIATIVE CAN 7673 05:35:00,180 --> 05:35:02,015 AND WILL BE PLAYING A ROLE HERE. 7674 05:35:02,015 --> 05:35:04,518 SO THIS IS KIND OF WHAT WE'RE QUITE INTERESTED IN. 7675 05:35:04,518 --> 05:35:07,654 I JUST WANTED TO LEAVE YOU WITH A FEW POINTS TO CONSIDER. 7676 05:35:07,654 --> 05:35:09,690 LOOKING AHEAD. 7677 05:35:09,690 --> 05:35:11,892 YOU KNOW, MANY, MANY QUESTIONS CAME UP OVER THE COURSE 7678 05:35:11,892 --> 05:35:14,861 OF THE LAST TWO DAYS. WE HAD SOME GREAT WRAP UP. 7679 05:35:14,861 --> 05:35:16,697 SUMMARIES JUST NOW. 7680 05:35:16,697 --> 05:35:19,499 BUT ONE THING TO THINK ABOUT IS HOW CAN NEUROAI BRIDGE 7681 05:35:19,499 --> 05:35:21,902 ALL THESE MULTIPLE SCALES, BOTH FROM THE EVOLUTIONARY 7682 05:35:21,902 --> 05:35:24,571 TIME SCALE TO REAL TIME TO THE MILLISECOND TIMESCALE, 7683 05:35:26,139 --> 05:35:27,307 CONSIDERING THE COMPUTATIONAL 7684 05:35:27,307 --> 05:35:28,875 ELEMENTS FROM DENDRITES TO NEURONS. 7685 05:35:28,875 --> 05:35:30,477 WE JUST HAD AN INTERESTING DISCUSSION 7686 05:35:30,477 --> 05:35:33,146 ABOUT NEUROMODULATION, WHICH ALSO CAME UP 7687 05:35:33,146 --> 05:35:34,648 CONSIDERING ALL THE CELL TYPES 7688 05:35:34,648 --> 05:35:35,916 IN THE CIRCUITS, NOT JUST THE NEURONS. 7689 05:35:35,916 --> 05:35:37,150 ASTROCYTES, BELIEVE IT OR NOT, 7690 05:35:37,150 --> 05:35:39,353 ARE THERE TO PLAY A VERY IMPORTANT ROLE. 7691 05:35:39,353 --> 05:35:42,222 AND USING ALL THIS TOGETHER TO UNDERSTAND THE BASIS 7692 05:35:42,222 --> 05:35:45,025 FOR BIOLOGICAL INTELLIGENCE, AND IN TURN, THAT CAN BE USED 7693 05:35:45,025 --> 05:35:48,028 TO DESIGN BETTER 7694 05:35:48,562 --> 05:35:50,931 ARTIFICIAL INTELLIGENCE. 7695 05:35:50,931 --> 05:35:53,567 WE NEED TO CONSIDER HOW CAN THE CROSS-MODAL DATA 7696 05:35:53,567 --> 05:35:55,502 INTEGRATION, STANDARDIZATION AND METRICS. 7697 05:35:55,502 --> 05:35:57,871 VERY IMPORTANT THAT YOU NEED TO, THAT WE UNDERSTAND 7698 05:35:57,871 --> 05:36:00,073 WHAT WE'RE TRYING TO MEASURE AS WE GO ALONG. 7699 05:36:00,073 --> 05:36:02,709 HOW CAN USE THAT TO LEVERAGE LARGE SCALE BRAIN DATA 7700 05:36:02,709 --> 05:36:06,013 AND OTHER DATA TO UNDERSTAND, SENSORIMOTOR INTERACTIONS 7701 05:36:06,013 --> 05:36:09,583 CRUCIAL FOR THEORETICAL ADVANCES AND PRACTICAL APPLICATIONS. 7702 05:36:09,583 --> 05:36:10,350 WE ARE INTERESTED 7703 05:36:10,350 --> 05:36:13,620 IN APPLICATIONS IN ADDITION TO THE THEORY AND THEORY 7704 05:36:13,620 --> 05:36:16,123 IS, OF COURSE, AT THE CORE OF WHAT WE'RE GOING TO BE DOING. 7705 05:36:17,391 --> 05:36:17,791 AND VERY 7706 05:36:17,791 --> 05:36:20,794 IMPORTANT THAT WE NEED TO BE VERY MINDFUL ABOUT 7707 05:36:21,595 --> 05:36:24,464 INCLUDING EVERYTHING WE DO, ETHICAL FRAMEWORKS, 7708 05:36:24,464 --> 05:36:26,600 NOT AS AN ADD ON, BUT FROM THE VERY BEGINNING 7709 05:36:26,600 --> 05:36:29,870 AS WE GET INTO THESE AREAS THAT ARE GOING TO BRING US INTO 7710 05:36:30,270 --> 05:36:33,273 DOMAINS THAT I DON'T THINK WE CAN NECESSARILY IMAGINE 7711 05:36:33,640 --> 05:36:35,108 WHERE WE'LL BE IN TEN YEARS. 7712 05:36:35,108 --> 05:36:38,278 SO IT'S VERY, VERY IMPORTANT FOR US TO BE, MINDFUL 7713 05:36:39,279 --> 05:36:40,380 ABOUT AND INTENTIONAL 7714 05:36:40,380 --> 05:36:43,784 ABOUT HOW WE THINK ABOUT THESE ISSUES AS THEY AFFECT HUMANS 7715 05:36:43,784 --> 05:36:46,787 AND QUITE FRANKLY, THE PLANET, NOT TO BE OVERLY DRAMATIC. 7716 05:36:47,287 --> 05:36:49,156 AND THEN FINALLY, 7717 05:36:49,156 --> 05:36:52,793 I CAN'T OVERSTATE THE IMPORTANCE OF OUR APPRECIATION 7718 05:36:52,793 --> 05:36:57,931 THAT THIS IS A VAST RESEARCH ECOSYSTEM. NIH BRAIN, 7719 05:36:57,931 --> 05:36:59,166 WE'RE JUST A PART OF THIS. 7720 05:36:59,166 --> 05:37:01,635 WE HOPE TO BE IN A LEADERSHIP POSITION 7721 05:37:01,635 --> 05:37:02,169 BECAUSE OF 7722 05:37:02,169 --> 05:37:04,171 THE INTELLECTUAL RESOURCES THAT WE 7723 05:37:04,171 --> 05:37:06,940 HAVE MANAGED TO PULL TOGETHER. 7724 05:37:06,940 --> 05:37:08,809 PUSHING THIS FIELD FORWARD IN A WAY 7725 05:37:08,809 --> 05:37:11,111 THAT I THINK CAN FULFILL ITS TRUE POTENTIAL 7726 05:37:11,111 --> 05:37:14,481 WILL REQUIRE THE COLLABORATION ACROSS AGENCIES, 7727 05:37:14,848 --> 05:37:17,784 NOT JUST FEDERAL AGENCIES, BUT ALSO, NONPROFITS, 7728 05:37:17,784 --> 05:37:20,787 OTHER FUNDERS, PHILANTHROPY, 7729 05:37:20,821 --> 05:37:23,323 AND THAT WE WILL HAVE TO BUILD AN ECOSYSTEM 7730 05:37:23,323 --> 05:37:26,326 THAT IS BUILT ON OPEN SHARING OF RESOURCES 7731 05:37:26,393 --> 05:37:29,196 FOCUSED ON TRAINING THE NEXT GENERATION OF 7732 05:37:29,196 --> 05:37:30,163 NEUROAI RESEARCHERS. 7733 05:37:30,163 --> 05:37:31,164 IT'S INTERESTING 7734 05:37:31,164 --> 05:37:33,033 WHEN WE TALK ABOUT MENTORING THE NEXT GENERATION, 7735 05:37:33,033 --> 05:37:34,668 I MEAN, WHAT THE HECK CAN I TEACH 7736 05:37:35,769 --> 05:37:37,404 SOME OF MY TRAINEES ABOUT THIS FIELD? 7737 05:37:37,404 --> 05:37:38,371 THEY KNOW MORE THAN I DO, 7738 05:37:38,371 --> 05:37:41,208 BUT I THINK IT'S A MATTER OF GIVING THEM THE SPACE 7739 05:37:41,208 --> 05:37:43,610 AND THE FREEDOM TO EXPLORE THESE, 7740 05:37:43,610 --> 05:37:46,613 NEW DOMAINS TO TO BE THE NEXT SET OF CREATORS. 7741 05:37:46,713 --> 05:37:48,448 AND OF COURSE, TRANSLATING THESE INSIGHTS 7742 05:37:48,448 --> 05:37:50,117 FROM BASIC TO BROADER APPLICATIONS. 7743 05:37:50,117 --> 05:37:51,785 SO THESE ARE 7744 05:37:51,785 --> 05:37:54,921 SOME SOMEWHAT GRANULAR THINGS I'D LIKE YOU TO THINK ABOUT. 7745 05:37:54,921 --> 05:37:57,190 BUT JUST OVER THE LAST YEAR, IN LISTENING 7746 05:37:57,190 --> 05:38:00,193 TO THE LAST OF THE DISCUSSIONS, I JUST HAD A FEW OTHER THOUGHTS 7747 05:38:00,260 --> 05:38:01,128 THAT IF YOU'LL ALLOW ME TO 7748 05:38:01,128 --> 05:38:04,097 JUST GO A LITTLE BIT MORE META. 7749 05:38:04,798 --> 05:38:07,534 SO PERMEATING A LOT OF THESE DISCUSSIONS 7750 05:38:07,534 --> 05:38:10,403 WAS THE ISSUE OF SCALING AND SCALABILITY. 7751 05:38:10,403 --> 05:38:12,706 AND I'VE COME TO APPRECIATE 7752 05:38:12,706 --> 05:38:14,274 OVER THE LAST FEW YEARS THAT ABOVE A CERTAIN 7753 05:38:14,274 --> 05:38:17,244 THRESHOLD, SCALING DOESN'T JUST ALLOW YOU TO DO MORE, 7754 05:38:17,244 --> 05:38:19,045 IT ALLOWS YOU TO DO DIFFERENT. 7755 05:38:19,045 --> 05:38:19,813 RIGHT? 7756 05:38:19,813 --> 05:38:22,716 SO IT OPENS UP A WHOLE NEW SET OF PARADIGMS 7757 05:38:22,716 --> 05:38:25,418 FOR UNDERSTANDING COMPLEX PROCESSES 7758 05:38:25,418 --> 05:38:27,654 AND SCALING UP THE ENTERPRISE. 7759 05:38:27,654 --> 05:38:29,456 AND THIS APPLIES HERE 7760 05:38:29,456 --> 05:38:30,891 IN PARTICULAR. 7761 05:38:30,891 --> 05:38:31,791 LET'S NOT DISCOUNT 7762 05:38:31,791 --> 05:38:33,760 THE ROLE OF TECHNOLOGY DISRUPTIONS 7763 05:38:33,760 --> 05:38:36,763 AND DRIVING THE FIELD FORWARD IN A NONLINEAR WAY. 7764 05:38:36,830 --> 05:38:38,498 SO, THERE WAS A MENTION 7765 05:38:38,498 --> 05:38:40,767 JUST IN THE LAST SESSION ABOUT MOORE'S LAW. 7766 05:38:40,767 --> 05:38:43,303 YOU KNOW, CAN WE YOU KNOW, WHAT DOES MOORE'S LAW LOOK LIKE 7767 05:38:43,303 --> 05:38:43,970 FOR ELECTRONICS? 7768 05:38:43,970 --> 05:38:46,139 WHAT DOES IT LOOK LIKE IN TERMS OF WHAT WE'VE LEARNED ABOUT 7769 05:38:47,741 --> 05:38:50,610 THE BRAIN IN TERMS OF CELL TYPES? 7770 05:38:50,610 --> 05:38:52,412 AND I THINK WE'VE EXCEEDED MOORE'S LAW. 7771 05:38:52,412 --> 05:38:54,281 BUT REALLY IT'S NOT A LINEAR. 7772 05:38:54,281 --> 05:38:57,651 OR EVEN A WELL-BEHAVED PROCESS 7773 05:38:57,651 --> 05:39:00,120 BECAUSE IT DEPENDS ON TECHNOLOGY DISRUPTIONS. 7774 05:39:00,120 --> 05:39:04,357 IF YOU LOOK AT SOME ADVANCES, FOR EXAMPLE, IN OUR ABILITY 7775 05:39:04,357 --> 05:39:05,692 TO DO, 7776 05:39:05,692 --> 05:39:08,828 DNA SEQUENCING FROM SLAB GELS, WHICH I RAN BACK 7777 05:39:08,828 --> 05:39:11,431 WHEN I WAS A STUDENT, RIGHT, TO CAPILLARY SEQUENCING, 7778 05:39:11,431 --> 05:39:13,099 THERE WAS A HUGE DISRUPTION 7779 05:39:13,099 --> 05:39:14,267 IN HIGH THROUGHPUT SEQUENCING 7780 05:39:14,267 --> 05:39:15,702 THAT ALLOWED US TO ASK QUESTIONS 7781 05:39:15,702 --> 05:39:18,104 THAT PEOPLE COULD NOT HAVE IMAGINED 7782 05:39:18,104 --> 05:39:21,208 WHEN WE WERE DEBATING WHETHER TO SEQUENCED GENOME TO BEGIN WITH. 7783 05:39:21,875 --> 05:39:23,510 AND SO, 7784 05:39:23,510 --> 05:39:24,211 NEUROPIXELS 7785 05:39:24,211 --> 05:39:25,178 AND OTHER HIGH DENSITY 7786 05:39:25,178 --> 05:39:27,013 RECORDING TECHNIQUES, IF THEY'RE DONE 7787 05:39:27,013 --> 05:39:29,049 IN AN INTELLIGENT AND A SMART WAY, 7788 05:39:29,049 --> 05:39:31,518 CAN REALLY OPEN UP NEW PARADIGMS FOR STUDYING 7789 05:39:31,518 --> 05:39:33,153 HOW NEURAL DYNAMICS 7790 05:39:33,153 --> 05:39:35,822 THAT ARE GOING TO BE SO IMPORTANT FOR GENERATING, 7791 05:39:35,822 --> 05:39:38,191 NEURO-INSPIRED A.I. ALGORITHMS 7792 05:39:38,191 --> 05:39:41,228 NEURO-INSPIRED PLATFORMS AND SO ON. 7793 05:39:41,294 --> 05:39:42,495 I JUST WANTED TO SAY AGAIN, 7794 05:39:43,630 --> 05:39:45,298 IN TERMS OF 7795 05:39:45,298 --> 05:39:48,034 THE IMPORTANCE OF DISRUPTIONS AND DRIVING THE FIELD 7796 05:39:48,034 --> 05:39:51,037 FORWARD, IT'S NOT GOING TO BE A LINEAR OR REGULAR PROCESS. 7797 05:39:51,204 --> 05:39:51,738 THE QUESTION, 7798 05:39:51,738 --> 05:39:53,707 I THINK FOR ALL OF US TO CONSIDER 7799 05:39:53,707 --> 05:39:55,208 IS HOW TO BUILD A RESEARCH 7800 05:39:55,208 --> 05:39:57,143 ECOSYSTEM THAT'S FRIENDLY FOR DISRUPTION. 7801 05:39:57,143 --> 05:39:58,812 YOU CAN'T TELL WHEN THE NEXT DISRUPTION 7802 05:39:58,812 --> 05:40:00,180 IS GOING TO COME, RIGHT. 7803 05:40:00,180 --> 05:40:03,183 BUT WE CAN CREATE AN ECOSYSTEM THAT'S FAVORABLE FOR IT. 7804 05:40:03,383 --> 05:40:05,018 AND AGAIN, THAT INCLUDES 7805 05:40:05,018 --> 05:40:07,187 TRAINING THE NEXT GENERATION OF LEADERS 7806 05:40:07,187 --> 05:40:08,922 AND GIVING THEM THE FREEDOM TO BE CREATIVE 7807 05:40:08,922 --> 05:40:11,491 AND TO TRUST THEIR INTUITION TO BRING OURSELVES 7808 05:40:11,491 --> 05:40:12,892 INTO NEW DOMAINS. 7809 05:40:12,892 --> 05:40:14,761 AND THEN, OF COURSE, THE QUESTION IS, WHAT ROLE 7810 05:40:14,761 --> 05:40:16,730 WILL ARTIFICIAL INTELLIGENCE AND NEURO-INSPIRED 7811 05:40:16,730 --> 05:40:20,500 ARTIFICIAL INTELLIGENCE PLAY IN SHAPING SUCH AN ECOSYSTEM? 7812 05:40:21,101 --> 05:40:24,104 AND THEN MY NEXT TO LAST POINT IS 7813 05:40:24,104 --> 05:40:27,107 DON'T FORGET THE VALUE OF GROUND TRUTH INFORMATION. 7814 05:40:27,107 --> 05:40:27,340 RIGHT? 7815 05:40:27,340 --> 05:40:28,708 WE'RE GOING TO BE BUILDING ALL THESE MODELS. 7816 05:40:28,708 --> 05:40:31,711 WE'RE GOING TO BE FINDING THESE LATENT PATTERNS OF 7817 05:40:31,711 --> 05:40:32,879 AND NEURAL DYNAMICS. 7818 05:40:32,879 --> 05:40:33,914 BUT AT THE END OF THE DAY, 7819 05:40:33,914 --> 05:40:37,083 THEY HAVE TO BRING US BACK TOWARDS SOME REALITY. 7820 05:40:37,784 --> 05:40:39,185 AND THEN FINALLY, I WANT TO BRING UP, 7821 05:40:40,253 --> 05:40:42,389 WHAT'S BECOME A BIT OF A MANTRA IN THE BRAIN INITIATIVE. 7822 05:40:42,389 --> 05:40:45,392 AND IF YOU LOOK AT SOME OF THESE BIG PROJECTS THAT WE STOOD UP, 7823 05:40:45,625 --> 05:40:48,228 INCLUDING THE CELL ATLASING PROJECTS, 7824 05:40:48,228 --> 05:40:50,463 AND I HOPE GOING FORWARD INTO THE CONNECTIVITY 7825 05:40:50,463 --> 05:40:54,401 PROJECTS AND OTHERS AND HOPEFULLY INTO THE A.I. FIELD. 7826 05:40:54,901 --> 05:40:57,971 OUR MANTRA IS THINK BIG, START SMALL AND SCALE FAST. 7827 05:40:58,538 --> 05:41:00,707 AND WITH THAT, I'LL LEAVE YOU. THANK YOU AGAIN FOR BEING HERE. 7828 05:41:00,707 --> 05:41:03,209 THANKS TO THE ORGANIZERS, 7829 05:41:03,209 --> 05:41:05,812 AND OUR CONTRACTORS FOR MAKING ALL THIS HAPPEN, 7830 05:41:05,812 --> 05:41:08,214 ESPECIALLY, IN TERMS OF THE ORGANIZERS, 7831 05:41:08,214 --> 05:41:09,416 ESPECIALLY JOE AND GRACE, 7832 05:41:09,416 --> 05:41:11,584 FOR PUTTING THIS TOGETHER, IT WAS A HUGE LIFT. 7833 05:41:11,584 --> 05:41:14,454 BUT I THINK THIS REALLY GETS US THINKING IN A GOOD WAY 7834 05:41:14,454 --> 05:41:17,123 ABOUT HOW TO SHAPE THE FUTURE IN THIS FIELD. 7835 05:41:17,123 --> 05:41:18,491 AND TO CREATE A NEW FIELD 7836 05:41:19,592 --> 05:41:21,061 FROM WHAT WE'VE BUILT SO FAR. 7837 05:41:21,061 --> 05:41:23,129 SO THANKS AGAIN, EVERYBODY. HAVE A SAFE TRIP BACK.