1
00:00:05,862 --> 00:00:07,664
GOOD AFTERNOON, EVERYONE. 

2
00:00:07,664 --> 00:00:09,432
WELCOME TO THE WALS WEDNESDAY

3
00:00:09,432 --> 00:00:12,502
AFTERNOON LECTURE SERIES TODAY. 

4
00:00:12,502 --> 00:00:14,871
I'M GRACE WONG, PROGRAM DIRECTOR

5
00:00:14,871 --> 00:00:16,306
NATIONAL INSTITUTE OF

6
00:00:16,306 --> 00:00:18,174
NEUROLOGICAL DISORDERS AND

7
00:00:18,174 --> 00:00:19,943
STROKE AND I ALSO WORK FOR THE

8
00:00:19,943 --> 00:00:22,378
BRAIN INITIATIVE. I'M DELIGHT TO

9
00:00:22,378 --> 00:00:24,747
BED HERE TODAY. BUT BEFORE WE

10
00:00:24,747 --> 00:00:26,049
START SOME VERY IMPORTANT

11
00:00:26,049 --> 00:00:27,217
INFORMATION FOR THOSE OF YOU WHO

12
00:00:27,217 --> 00:00:29,953
ARE HERE FOR CONTINUING MEDICAL

13
00:00:29,953 --> 00:00:32,288
EDUCATION CREDITS THE CODE IS,

14
00:00:32,288 --> 00:00:36,459
LOOK AT THAT, 57659. IF YOU ARE

15
00:00:36,459 --> 00:00:38,661
ATTENDING VIRTUALLY, AND HAVE

16
00:00:38,661 --> 00:00:40,630
QUESTIONS, PLEASE HIT THE SEND

17
00:00:40,630 --> 00:00:43,166
LIVE FEEDBACK ON THE VIDEOCAST

18
00:00:43,166 --> 00:00:45,501
PAGE AND YOUR QUESTIONS WILL BE

19
00:00:45,501 --> 00:00:48,504
ASKED AT THE END OF THE LECTURE.

20
00:00:48,504 --> 00:00:50,139
SO IT IS MY PLEASURE TO

21
00:00:50,139 --> 00:00:53,443
INTRODUCE OUR SPEAKER TODAY, DR.

22
00:00:53,443 --> 00:00:56,079
BRAD AIMONE, WHO I HAVE ACTUALLY

23
00:00:56,079 --> 00:00:57,680
KNOWN FOR UP TO A DECADE.  WE

24
00:00:57,680 --> 00:00:59,849
JUST DISCOVERED THIS OVER LUNCH.

25
00:00:59,849 --> 00:01:02,318
I REALLY GOT TO KNOW BRAD IN

26
00:01:02,318 --> 00:01:03,820
2020 AT MY PREVIOUS JOB AS A

27
00:01:03,820 --> 00:01:06,022
PROGRAM DIRECTOR AT THE NATIONAL

28
00:01:06,022 --> 00:01:08,691
SCIENCE FOUNDATION. I STARTED A

29
00:01:08,691 --> 00:01:10,894
PROGRAM EMERGING FRONTIERS AN

30
00:01:10,894 --> 00:01:12,128
RESEARCH INNOVATION PROGRAM ON

31
00:01:12,128 --> 00:01:15,632
BRAIN IRSPIRED DYNAMICS FOR

32
00:01:15,632 --> 00:01:16,499
INTERRELATED CIRCUIT AND

33
00:01:16,499 --> 00:01:18,401
ARTIFICIAL INTELLIGENCE AND BRAD

34
00:01:18,401 --> 00:01:20,203
WAS GENEROUS WITH WITH HIS

35
00:01:20,203 --> 00:01:23,039
PERSPECTIVE AND KNOWLEDGE. HE IS

36
00:01:23,039 --> 00:01:26,342
A DISTINGUISHED FELLOW AT SANDIA

37
00:01:26,342 --> 00:01:27,143
NATIONAL LAB, HE LED SEVERAL

38
00:01:27,143 --> 00:01:29,712
RESEARCH PROJECTS THAT LEVERAGES

39
00:01:29,712 --> 00:01:31,147
COMPUTATIONAL NEUROSCIENCE TO

40
00:01:31,147 --> 00:01:32,916
ADVANCE ARTIFICIAL INTELLIGENCE.

41
00:01:32,916 --> 00:01:35,551
AND IN USING NEUROMORPHIC

42
00:01:35,551 --> 00:01:38,554
COMPUTING PLATFORMS FOR FUTURE

43
00:01:38,554 --> 00:01:39,389
ENERGY EFFICIENT SCIENTIFIC

44
00:01:39,389 --> 00:01:44,327
COMPUTING.  HE RECENTLY LED A

45
00:01:44,327 --> 00:01:45,061
MULTI-INSTITUTION DEPARTMENT OF

46
00:01:45,061 --> 00:01:48,731
ENERGY OFFICE OF SCIENCE

47
00:01:48,731 --> 00:01:50,800
MICROELECTRONICS CO-DESIGN

48
00:01:50,800 --> 00:01:53,469
PROJECT ON NOVEL PROBABILISTIC

49
00:01:53,469 --> 00:01:54,804
NEUROMORPHIC COMPUTING CALLED

50
00:01:54,804 --> 00:01:58,408
COIN FLIPSS. IN 2023 BRAD WAS

51
00:01:58,408 --> 00:02:00,310
RECIPIENT OF THE THE PRESTIGIOUS

52
00:02:00,310 --> 00:02:01,444
PRIZE FOR NEUROMORPHIC

53
00:02:01,444 --> 00:02:03,913
ENGINEERING. AND HE WAS ALSO A

54
00:02:03,913 --> 00:02:05,748
NATIONAL ACADEMY OF ENGINEERING

55
00:02:05,748 --> 00:02:09,252
IN THIS FRONTIERS OF ENGINEERING

56
00:02:09,252 --> 00:02:11,287
INVITEE SINCE 2017. BEFORE

57
00:02:11,287 --> 00:02:13,823
JOINING THE DEPARTMENT OF ENERGY

58
00:02:13,823 --> 00:02:16,259
BRAD STUDIED WITH RUSTY GAUGE AT

59
00:02:16,259 --> 00:02:19,162
THE SALK INSTITUTE AS

60
00:02:19,162 --> 00:02:20,697
COMPUTATIONAL NEUROSCIENCE IN

61
00:02:20,697 --> 00:02:22,365
NEUROGENESIS.  YOU WILL HEAR A

62
00:02:22,365 --> 00:02:24,133
LITTLE BIT OF THAT RESEARCH

63
00:02:24,133 --> 00:02:25,668
TODAY AND E HE WILL TELL HOW

64
00:02:25,668 --> 00:02:28,171
NEUROMORPHIC COMPUTING CAN HELP

65
00:02:28,171 --> 00:02:31,708
US UNDERSTAND THE BRAIN. BRAD.

66
00:02:31,708 --> 00:02:32,275
[APPLAUSE]

67
00:02:32,275 --> 00:02:34,344
>> LET'S WELL COP OUR
SPEAKER. 

68
00:02:34,344 --> 00:02:37,146
-- WELCOME OUR SPEAKER.

69
00:02:37,146 --> 00:02:40,316
>> THANK YOU, GRACE. THANK
YOU,

70
00:02:40,316 --> 00:02:42,118
EVERYONE FOR THIS AMAZING

71
00:02:42,118 --> 00:02:43,753
OPPORTUNITY TO COME HERE AND

72
00:02:43,753 --> 00:02:46,856
SPEAK WITH YOU TODAY, I HAVE

73
00:02:46,856 --> 00:02:48,858
BEEN -- IT IS QUITE AN HONOR

74
00:02:48,858 --> 00:02:51,127
LOOKING AT THE PAST PRESENTERS

75
00:02:51,127 --> 00:02:53,029
OF THIS LECTURE SERIES AND I

76
00:02:53,029 --> 00:02:54,397
THINK IT IS -- I THINK IT IS AN

77
00:02:54,397 --> 00:02:56,599
EXCITING TIME TO BE ABLE TO COME

78
00:02:56,599 --> 00:02:57,300
AND SHARE SOME OF THE THINGS

79
00:02:57,300 --> 00:03:00,603
THAT ARE GOING ON IN MY FIELD,

80
00:03:00,603 --> 00:03:02,538
NEUROMORPHIC COMPUTING PEELED,

81
00:03:02,538 --> 00:03:05,274
WHICH IS A NEW FIELD FOR ME, AS

82
00:03:05,274 --> 00:03:07,176
GRACE MENTIONED, MY ORIGINAL

83
00:03:07,176 --> 00:03:10,179
WORK AS A NEUROSCIENTIST WAS

84
00:03:10,179 --> 00:03:13,049
LOOKING AT ADULT NEUROGENESIS, I

85
00:03:13,049 --> 00:03:14,817
WILL TELL ABOUT TO START. OVER

86
00:03:14,817 --> 00:03:18,354
THE LAST DECADE FOCUSED ON

87
00:03:18,354 --> 00:03:21,457
COMPUTING SIDE. SOMEWHAT

88
00:03:21,457 --> 00:03:23,426
SURPRISINGLY NEUROMORPHIC

89
00:03:23,426 --> 00:03:25,461
COMPUTING HASN'T HAD A LOT TO DO

90
00:03:25,461 --> 00:03:26,729
WITH NEUROSCIENCE HISTORICALLY

91
00:03:26,729 --> 00:03:28,831
BUT IT IS CHANGING AND WE HAVE A

92
00:03:28,831 --> 00:03:30,933
REAL OPPORTUNITY IF WE TAKE --

93
00:03:30,933 --> 00:03:32,402
CHOOSE TO TAKE ADVANTAGE OF IT

94
00:03:32,402 --> 00:03:33,736
TO BRING THESE FIELDS CLOSERRING

95
00:03:33,736 --> 00:03:36,539
TO AND HELP US GAIN NEW INSIGHT

96
00:03:36,539 --> 00:03:39,008
HOUSE THE BRAIN WORKS. SO I'M

97
00:03:39,008 --> 00:03:44,013
GOING TO START WITH A -- KIND OF

98
00:03:44,013 --> 00:03:45,348
METAPHOR IF YOU WILL OR

99
00:03:45,348 --> 00:03:46,816
INSPIRATION IF YOU WILL. AND

100
00:03:46,816 --> 00:03:48,484
LOOKING AT HOW MODELING AND

101
00:03:48,484 --> 00:03:50,019
SIMULATION HAS HAD A TREMENDOUS

102
00:03:50,019 --> 00:03:52,055
IMPACT IN OUR SOCIETY AS A

103
00:03:52,055 --> 00:03:55,792
WHOLE. THIS IS HURRICANE MILTON

104
00:03:55,792 --> 00:03:57,693
THAT IF YOU ARE FAMILIAR WITH

105
00:03:57,693 --> 00:03:59,796
THE NEWS MOST PEOPLE ARE, HIT

106
00:03:59,796 --> 00:04:03,099
TAMPA A COUPLE OF WEEKS AGO. AND

107
00:04:03,099 --> 00:04:04,434
WHAT IS AMAZING ABOUT THIS IS,

108
00:04:04,434 --> 00:04:08,371
WE ARE ALL FAMILIAR WITH THESE

109
00:04:08,371 --> 00:04:10,706
CONES OF UNCERTAINTY WHERE THE

110
00:04:10,706 --> 00:04:11,741
STORM TRACKS COME AND THIS IS

111
00:04:11,741 --> 00:04:13,276
CASE THE AVERAGE PERSON IS

112
00:04:13,276 --> 00:04:15,011
SEEING MODELING SIMULATION

113
00:04:15,011 --> 00:04:16,112
IMPACT THEIR LIVES IN REAL TIME.

114
00:04:16,112 --> 00:04:18,681
DAY TO DAY WE GET A NEW FORECAST

115
00:04:18,681 --> 00:04:20,683
WHERE THE STORM MAY GO. THIS IS

116
00:04:20,683 --> 00:04:22,051
AN I BELIEVE USUAL ONE, WENTS

117
00:04:22,051 --> 00:04:23,920
FROM WEST TO EAST, NOT WHAT

118
00:04:23,920 --> 00:04:25,388
HURRICANE DOS IN THAT PART OF

119
00:04:25,388 --> 00:04:29,325
THE COUNTRY. BUT WHAT IS EVEN

120
00:04:29,325 --> 00:04:31,627
MORE AMAZING, THIS STORM DID

121
00:04:31,627 --> 00:04:33,096
ALMOST EXACTLY WHAT THE MODELS

122
00:04:33,096 --> 00:04:34,997
PREDICTED. THAT DOESN'T ALWAYS

123
00:04:34,997 --> 00:04:36,232
HAPPEN BUT IT IS HAPPENING MORE

124
00:04:36,232 --> 00:04:37,467
AND MORE OFTEN BECAUSE OUR

125
00:04:37,467 --> 00:04:39,435
ABILITY TO SIMULATE THESE

126
00:04:39,435 --> 00:04:44,073
SYSTEMS IS COMPLEX WEATHER, INCS

127
00:04:44,073 --> 00:04:45,441
GOTTEN BETTER AND BETTER EVERY

128
00:04:45,441 --> 00:04:47,043
YEAR AND PART OF THE REASON WHY

129
00:04:47,043 --> 00:04:51,581
IS IF YOU LOOK AT THESE SORT OF

130
00:04:51,581 --> 00:04:53,583
SUPERCOMPUTERS THAT NOAA HAS, IN

131
00:04:53,583 --> 00:04:57,053
VIRGINIA OR ARIZONA, THESE LARGE

132
00:04:57,053 --> 00:04:58,087
SCALE GIANT HIGH PERFORMANCE

133
00:04:58,087 --> 00:05:00,490
COMPUTING SYSTEMS, IN EFFECT CAN

134
00:05:00,490 --> 00:05:03,926
BE DEPLOYED SAVING UNTOLD NUMBER

135
00:05:03,926 --> 00:05:07,163
OF LIVES AND TREMENDOUS ECONOMIC

136
00:05:07,163 --> 00:05:08,831
IMPACT. THE QUESTION, WHY ARE WE

137
00:05:08,831 --> 00:05:11,267
NOT ABLE TO DO THIS TODAY FOR

138
00:05:11,267 --> 00:05:12,935
MENTAL HEALTH AND NEUROLOGICAL

139
00:05:12,935 --> 00:05:14,203
DISORDERS? THIS IS THE

140
00:05:14,203 --> 00:05:15,671
CHALLENGE, MOTIVATING THOUGHT.

141
00:05:15,671 --> 00:05:17,039
WOULDN'T IT BE NICE IF WE COULD

142
00:05:17,039 --> 00:05:18,908
MAKE THAT SAME PREDICTION?

143
00:05:18,908 --> 00:05:20,510
OTHERS ARE THINKING ABOUT THIS

144
00:05:20,510 --> 00:05:23,079
AS WELL. BUT I WANT TO COME AT

145
00:05:23,079 --> 00:05:24,647
IT FROM A DIFFERENT PERSPECTIVE.

146
00:05:24,647 --> 00:05:27,717
WHICH REALLY IS LOOKING FROM THE

147
00:05:27,717 --> 00:05:30,887
BOTTOM UP. AS SOMEONE WHO

148
00:05:30,887 --> 00:05:34,190
MODELED BIOLOGICAL PROCESS, KIND

149
00:05:34,190 --> 00:05:36,425
OF HIGH LEVEL GRANULARITY, AT

150
00:05:36,425 --> 00:05:39,095
LEAST TIME, HOW DO WE GO FROM

151
00:05:39,095 --> 00:05:40,563
THIS LEVEL UNDERSTANDING INTO

152
00:05:40,563 --> 00:05:42,165
THIS SORT OF SYSTEM

153
00:05:42,165 --> 00:05:43,933
UNDERSTANDING OF HOW THE BRAIN

154
00:05:43,933 --> 00:05:46,135
WORKS AND HOW DISEASES MAY BE

155
00:05:46,135 --> 00:05:51,340
UNDERSTOOD AND BE TREATED? IF WE

156
00:05:51,340 --> 00:05:52,909
WANT TO START BUILDING THESE

157
00:05:52,909 --> 00:05:58,314
SORT OF HIGH FIDELITY MODELS OF

158
00:05:58,314 --> 00:06:09,091
NEURO-PROCESSES NEURAL MY WORK S

159
00:06:18,234 --> 00:06:21,237
WELL AS BROADER COMMUNITY

160
00:06:21,237 --> 00:06:22,838
STARTED HELP SHAPE A PATH

161
00:06:22,838 --> 00:06:24,507
FORWARD HOW WE ARE GOING TO

162
00:06:24,507 --> 00:06:25,975
POTENTIALLY WHAT WE NEED TO DO

163
00:06:25,975 --> 00:06:29,612
TO GET THERE. THEN I WANT TO

164
00:06:29,612 --> 00:06:31,080
TALK ABOUT THIS FIELD OF

165
00:06:31,080 --> 00:06:32,548
NEUROMORPHIC COMPUTING WHICH IS

166
00:06:32,548 --> 00:06:34,116
A REALLY A NEW THING I IMAGINE

167
00:06:34,116 --> 00:06:36,552
FOR MANY OF YOU PROBABLY A NEW

168
00:06:36,552 --> 00:06:37,687
CONCEPT, IT IS SOMETHING THAT IS

169
00:06:37,687 --> 00:06:39,589
REALLY AN EXCITING AREA THAT'S

170
00:06:39,589 --> 00:06:40,856
REALLY EMERGING ON THE SCENE

171
00:06:40,856 --> 00:06:44,360
RIGHT NOW. AND TALK ABOUT HOW

172
00:06:44,360 --> 00:06:47,029
THAT MAY ACTUALLY HAVE REAL

173
00:06:47,029 --> 00:06:49,298
UNPREDICTED INSIGHTS HOW NEURAL

174
00:06:49,298 --> 00:06:51,767
COMPUTATION MAY BE OCCURRING. SO

175
00:06:51,767 --> 00:06:54,870
TO START, I WANT TO GO BACK TO

176
00:06:54,870 --> 00:06:58,207
ADULT NEUROGENESIS, THIS IS A

177
00:06:58,207 --> 00:07:00,176
PROCESS THAT IS QUITE I STILL

178
00:07:00,176 --> 00:07:01,877
FIND VERY EXCITING THE IDEA THAT

179
00:07:01,877 --> 00:07:03,813
RIGHT NOW ALL OF US IN THIS ROOM

180
00:07:03,813 --> 00:07:07,984
ALL OF US LISTENING ONLINE, ARE

181
00:07:07,984 --> 00:07:09,685
EXPERIENCING THE ADDITION OF NEW

182
00:07:09,685 --> 00:07:11,587
NEURONS INTO OUR HIPPOCAMPUS NOT

183
00:07:11,587 --> 00:07:13,389
A TON OF THEM MIND YOU BUT THEY

184
00:07:13,389 --> 00:07:16,692
ARE HAPPENING. DEVELOPMENT DOES

185
00:07:16,692 --> 00:07:18,894
NOT END WHEN YOU ARE FIVE YEARS

186
00:07:18,894 --> 00:07:21,330
OLD THE WAY WE WERE TAUGHT

187
00:07:21,330 --> 00:07:22,565
GROWING UP. THERE ARE NEW

188
00:07:22,565 --> 00:07:25,368
NEURONS IN THE ADULT BRAIN. THIS

189
00:07:25,368 --> 00:07:28,070
IS A LOCALIZED PROCESS, AS YOU

190
00:07:28,070 --> 00:07:30,506
SEE HERE, KIND OF HIGHLIGHTED IN

191
00:07:30,506 --> 00:07:31,874
THE DENTIGEROUS OF THE

192
00:07:31,874 --> 00:07:33,876
HIPPOCAMPUS.  THE BOTTOM RIGHT

193
00:07:33,876 --> 00:07:35,878
IS AN IMAGE FROM (INDISCERNIBLE)

194
00:07:35,878 --> 00:07:38,014
A POST DOC WHEN I WAS A GRAD

195
00:07:38,014 --> 00:07:40,249
STUDENT WHO TOOK THESE BEAUTIFUL

196
00:07:40,249 --> 00:07:43,219
PICTURES OF THESE GFP POSITIVE

197
00:07:43,219 --> 00:07:44,920
NEURONS, THE GREEN CELLS ARE THE

198
00:07:44,920 --> 00:07:46,489
NEWBORN CELLS, ABOUT A MONTH OLD

199
00:07:46,489 --> 00:07:49,191
IN A MOUSE. AND IT IS

200
00:07:49,191 --> 00:07:51,727
PARTICULARLY A SMALL POPULATION

201
00:07:51,727 --> 00:07:53,296
OF GRANULE CELLS WHICH ARE THE

202
00:07:53,296 --> 00:07:54,664
DOMINANT CELL POPULATION IN THE

203
00:07:54,664 --> 00:07:56,565
DEN DAY WHICH ARE THE PINK CELLS

204
00:07:56,565 --> 00:07:59,035
YOU SEE THERE. SO AS I

205
00:07:59,035 --> 00:07:59,802
MENTIONED, THIS WAS KIND OF ONE

206
00:07:59,802 --> 00:08:01,170
OF THESE PEOPLE LOVE TO USE THIS

207
00:08:01,170 --> 00:08:03,606
AS AN EXAMPLE, ONE ICE DOGMAS

208
00:08:03,606 --> 00:08:06,409
THAT WAS OVERTURNED WE KNOW

209
00:08:06,409 --> 00:08:07,610
NEUROGENESIS DOES EXIST THOUGH

210
00:08:07,610 --> 00:08:11,981
WE WERE TOLD IT DOES NOT. BUT IT

211
00:08:11,981 --> 00:08:14,083
IS NOT A TON OF CELLS AS THEY

212
00:08:14,083 --> 00:08:16,819
SAID. HUMANS MITT IT MAY BE ANY

213
00:08:16,819 --> 00:08:18,788
GIVEN TIME HAVING A FEW THOUSAND

214
00:08:18,788 --> 00:08:21,257
NEURONS IMMATURE CRITICAL STATE.

215
00:08:21,257 --> 00:08:22,491
STILL TO BE DETERMINED.  THIS IS

216
00:08:22,491 --> 00:08:25,328
VERY DIFFICULT TO MEASURE IN

217
00:08:25,328 --> 00:08:27,763
HUMANS FOR OBVIOUS REASONS. WE

218
00:08:27,763 --> 00:08:29,231
DO KNOW THOUGH IN RODENT THIS IS

219
00:08:29,231 --> 00:08:30,933
HAS A REAL IMPACT ON MEMORY

220
00:08:30,933 --> 00:08:33,269
FORMATION. IT SEEMS TO BE

221
00:08:33,269 --> 00:08:37,506
CLEARLY IMPORTANT FOR COGNITION.

222
00:08:37,506 --> 00:08:40,042
AND WE ALSO KNOW IT IS HIGHLY

223
00:08:40,042 --> 00:08:43,279
CORRELATED WITH A LOT OF THINGS

224
00:08:43,279 --> 00:08:45,181
THAT ARE CLINICALLY HAVE A LOT

225
00:08:45,181 --> 00:08:46,982
OF INTEREST, THINGS LIKE

226
00:08:46,982 --> 00:08:49,652
EPILEPSY, DEPRESSION, STRESS,

227
00:08:49,652 --> 00:08:51,754
AGING, ALL OF THOSE SEEM TO HAVE

228
00:08:51,754 --> 00:08:55,291
SOME INTERACTION WITH

229
00:08:55,291 --> 00:08:56,659
NEUROGENESIS AND AT THE SAME

230
00:08:56,659 --> 00:08:58,761
TIME SOME THINGS LIKE SSRIs

231
00:08:58,761 --> 00:09:09,305
APPEAR THE INTERACT AND TO TALK

232
00:09:18,214 --> 00:09:20,616
ABOUT A LITTLE BIT IS THE

233
00:09:20,616 --> 00:09:24,954
COMPUTATIONAL IMPACT OF THIS.

234
00:09:24,954 --> 00:09:25,821
WHAT IS CONVENIENT ABOUT THIS

235
00:09:25,821 --> 00:09:27,656
PROCESS IS THAT THIS -- WE WERE

236
00:09:27,656 --> 00:09:30,326
ADDING NEW -- APPARENTLY THE

237
00:09:30,326 --> 00:09:31,160
ADDITIONAL NEURONS IS OCCURRING

238
00:09:31,160 --> 00:09:32,828
IN A BRAIN REGION RELATIVELY

239
00:09:32,828 --> 00:09:34,063
WELL STUDIED FROM A

240
00:09:34,063 --> 00:09:36,499
COMPUTATIONAL POINT OF VIEW.

241
00:09:36,499 --> 00:09:38,701
DENTIGEROUS IS WHERE LTP WAS

242
00:09:38,701 --> 00:09:41,303
ORIGINALLY SEEN IN VIVO IN

243
00:09:41,303 --> 00:09:43,806
MAMMALS. WHAT WE -- WHAT PEOPLE

244
00:09:43,806 --> 00:09:45,374
WITHIN OBSERVING WHILE I WAS A

245
00:09:45,374 --> 00:09:46,842
GRAD STUDENT A LOT OF THE FIELD

246
00:09:46,842 --> 00:09:48,344
WAS RESONATING AROUND THIS IDEA

247
00:09:48,344 --> 00:09:50,012
YOU SEE ON THE LEFT HERE, THAT

248
00:09:50,012 --> 00:09:51,380
THESE KNEW NEURONS AS THEY COME

249
00:09:51,380 --> 00:09:52,681
IN -- NEW NEURONS AS THEY COME

250
00:09:52,681 --> 00:09:54,383
IN OVER THE COURSE OF A MOUSE

251
00:09:54,383 --> 00:09:55,818
MULTIPLE WEEKS OF DEVELOPMENT,

252
00:09:55,818 --> 00:09:57,987
THAT THEY ARE -- THEY GO THROUGH

253
00:09:57,987 --> 00:09:59,955
THAT'S PERIODS OF INCREASED

254
00:09:59,955 --> 00:10:00,990
EXCITABILITY AND INCREASE

255
00:10:00,990 --> 00:10:03,893
ABILITY TO LEARN. THEY START OFF

256
00:10:03,893 --> 00:10:04,927
QUIET, THEIR STEM CELLS NOT

257
00:10:04,927 --> 00:10:08,030
WIRED INTO ANYTHING, BECOME

258
00:10:08,030 --> 00:10:10,032
HYPERACTIVE, THEY ARE LIKE

259
00:10:10,032 --> 00:10:11,300
HYPERACTIVE LITTLE CHILDREN AS

260
00:10:11,300 --> 00:10:12,334
YOU SEE ON THE VIDEO ON THE

261
00:10:12,334 --> 00:10:14,336
RIGHT.  AND MATURE CELLS

262
00:10:14,336 --> 00:10:16,238
EVENTUALLY MATURE AND GET QUIET

263
00:10:16,238 --> 00:10:17,373
AND ONLY RESPOND TO WHAT THEY

264
00:10:17,373 --> 00:10:20,276
ARE SUPPOSED TO. SO THIS SORT OF

265
00:10:20,276 --> 00:10:23,245
MODEL OF COMPUTATION WAS

266
00:10:23,245 --> 00:10:24,947
APPEALING, IT WAS ATTRACTIVE

267
00:10:24,947 --> 00:10:26,816
BECAUSE WHAT WE HAVE IS A CASE

268
00:10:26,816 --> 00:10:29,051
WHERE THERE IS A CLASSIC CODING

269
00:10:29,051 --> 00:10:32,121
SCHEME IN THE DENTIGEROUS WHERE

270
00:10:32,121 --> 00:10:33,255
EACH NEURON RESPONDS IN EFFECT

271
00:10:33,255 --> 00:10:35,491
LIKE ALMOST LIKE A GRANDMOTHER

272
00:10:35,491 --> 00:10:37,493
CELL, NOT QUITE A GRANDMOTHER

273
00:10:37,493 --> 00:10:39,628
CELL BUT RESPONDS HIGHLY

274
00:10:39,628 --> 00:10:42,598
SELECTIVELY TO MALL SET OF

275
00:10:42,598 --> 00:10:43,833
THINGS BUT YOUNG NEURONS

276
00:10:43,833 --> 00:10:45,868
TRANSIENT PERIOD OF TIME ARE

277
00:10:45,868 --> 00:10:46,902
RECEPTIVE TO NEW THINGS AND IN

278
00:10:46,902 --> 00:10:48,237
THAT PROCESS LEARN TO RESPOND TO

279
00:10:48,237 --> 00:10:51,207
THOSE LATER. BASICALLY SAMPLED

280
00:10:51,207 --> 00:10:53,275
LOTS OF POSSIBILITIES AND BECOME

281
00:10:53,275 --> 00:10:56,345
EXPERT IN ONE OR TWO THINGS.

282
00:10:56,345 --> 00:10:57,847
THIS WAS COMPELLING AND

283
00:10:57,847 --> 00:10:59,982
INTERESTING FROM COMPUTATIONAL

284
00:10:59,982 --> 00:11:00,950
POINT OF VIEW. IT MAKES YOU

285
00:11:00,950 --> 00:11:02,985
START THINKING WHAT DOES IT MEAN

286
00:11:02,985 --> 00:11:06,088
TO BUILD A COMPUTING SYSTEM AN

287
00:11:06,088 --> 00:11:07,423
ALGORITHM IF YOU WILL, THIS IS

288
00:11:07,423 --> 00:11:11,527
BEFORE AI TOOK OFF. SO WHEN WE

289
00:11:11,527 --> 00:11:13,963
STARTED THINKING WHAT THIS

290
00:11:13,963 --> 00:11:15,598
MEANT, I WON'T GO INTO THE

291
00:11:15,598 --> 00:11:17,366
MODELING ITSELF IN DEPTH BUT THE

292
00:11:17,366 --> 00:11:20,469
CONCLUSION WE HAD WAS THAT

293
00:11:20,469 --> 00:11:21,937
DENTIGEROUS HAS BEEN KNOWN AND

294
00:11:21,937 --> 00:11:24,073
BELIEVED FOR MANY YEARS TO BE

295
00:11:24,073 --> 00:11:25,941
IMPORTANT PATTERN SEPARATION THE

296
00:11:25,941 --> 00:11:28,177
ABILITY TO DISCRIMINATE BETWEEN

297
00:11:28,177 --> 00:11:29,078
RELATIVELY SIMILAR THINGS, YOU

298
00:11:29,078 --> 00:11:31,113
CAN SEE THIS FROM ANIMAL

299
00:11:31,113 --> 00:11:33,382
BEHAVIOR SIDE WHERE ANIMAL

300
00:11:33,382 --> 00:11:35,684
CONTEXT FEAR CONDITIONING

301
00:11:35,684 --> 00:11:37,753
DISCRIMINATE BETWEEN TWO SIMILAR

302
00:11:37,753 --> 00:11:39,021
ENVIRONMENTS.  DIFFERENT

303
00:11:39,021 --> 00:11:40,723
POPULATIONS OF NEURONS RESPONDS

304
00:11:40,723 --> 00:11:46,996
TO SIMILAR INPUTS, SO FORTH.  

305
00:11:46,996 --> 00:11:49,365
BUT WHAT WAS COMPLICATED STILL,

306
00:11:49,365 --> 00:11:51,467
IF THE DENTY GYRUS IS GOING TO

307
00:11:51,467 --> 00:11:52,268
PATTERN SEPARATION BY

308
00:11:52,268 --> 00:11:53,569
CONSTRUCTION AND ANATOMICAL

309
00:11:53,569 --> 00:11:54,703
REASONS TO THINK THAT BE THE

310
00:11:54,703 --> 00:11:57,606
CASE, WHY WOULD YOU NEED TO HAVE

311
00:11:57,606 --> 00:11:59,074
THIS COMPLICATED PROCESS OF

312
00:11:59,074 --> 00:12:01,844
NEUROGENESIS TO DO THAT? YOU

313
00:12:01,844 --> 00:12:03,512
DON'T NEED NEUROGENESIS TO

314
00:12:03,512 --> 00:12:06,482
PATTERN SEPARA SEPARATE. THEN

315
00:12:06,482 --> 00:12:07,650
ESPECIALLY GIVEN THE REST OF THE

316
00:12:07,650 --> 00:12:09,451
BRAIN SHUT OFF NEUROGENESIS. THE

317
00:12:09,451 --> 00:12:11,287
VAST MAJORITY OF THE BRAIN DOES

318
00:12:11,287 --> 00:12:13,088
NOT HAVE NEURONS COMING IN REAL

319
00:12:13,088 --> 00:12:15,090
LIFE SO CLEARLY SOMETHING MORE

320
00:12:15,090 --> 00:12:16,792
HAS TO BE HAPPENING. BUT WE JUST

321
00:12:16,792 --> 00:12:20,029
DIDN'T HAVE THE TOOLS TO

322
00:12:20,029 --> 00:12:23,766
UNDERSTAND IT. SO REALLY, THE

323
00:12:23,766 --> 00:12:26,635
CHALLENGE THAT I HAD AT THE

324
00:12:26,635 --> 00:12:28,304
TIME, THIS WAS I WAS FINISHING

325
00:12:28,304 --> 00:12:30,205
MY Ph.D., SOMETIMES POST DOC

326
00:12:30,205 --> 00:12:32,474
AND TRYING TO UNDERSTAND IF I

327
00:12:32,474 --> 00:12:34,476
HAD THIS CONCEPTUAL MODEL OF

328
00:12:34,476 --> 00:12:35,744
WHAT NEW NEURONS MAYBE DOING, I

329
00:12:35,744 --> 00:12:37,446
WANT TO GET TO THE -- I WANT TO

330
00:12:37,446 --> 00:12:41,050
-- WE ARE IN IT TO TREAT

331
00:12:41,050 --> 00:12:42,184
DISEASES AND IMPROVE COGNITION

332
00:12:42,184 --> 00:12:45,421
WITH AGING AND OVERCOME EFFECTS

333
00:12:45,421 --> 00:12:46,455
OF MENTAL HEALTH ON COGNITION

334
00:12:46,455 --> 00:12:48,557
AND SO FORTH. WE KNOW

335
00:12:48,557 --> 00:12:49,925
NEUROGENESIS IS LIKELY INVOLVED

336
00:12:49,925 --> 00:12:51,126
BUT HOW DO WE CONNECT THOSE

337
00:12:51,126 --> 00:12:54,330
DOTS? WE NEED TO EXPAND IN SOME

338
00:12:54,330 --> 00:12:55,998
DIRECTION. AND THE TWO OBVIOUS

339
00:12:55,998 --> 00:12:58,100
WAYS LOOKING AT IT FROM MY

340
00:12:58,100 --> 00:13:00,469
PERSPECTIVE AS A COMPUTATIONAL

341
00:13:00,469 --> 00:13:01,470
PERSON, WAS MAYBE IT IS A

342
00:13:01,470 --> 00:13:03,839
CHALLENGE WITH MODELING, THAT WE

343
00:13:03,839 --> 00:13:05,207
SIMPLY DON'T HAVE A BIG ENOUGH

344
00:13:05,207 --> 00:13:08,143
MODEL.  WE DON'T HAVE A

345
00:13:08,143 --> 00:13:10,312
REALISTIC SIZE MODEL, AT THIS

346
00:13:10,312 --> 00:13:12,881
TIME THIS IS BEFORE AI BEFORE

347
00:13:12,881 --> 00:13:16,018
GPUs THAT RUN BILLIONS OF

348
00:13:16,018 --> 00:13:17,620
NEURONS MODELS. MAYBE I NEED A

349
00:13:17,620 --> 00:13:20,222
BIGGER MODEL.  BUT EVEN IF YOU

350
00:13:20,222 --> 00:13:21,924
LOOK AT THIS, IT IS A LOT OF

351
00:13:21,924 --> 00:13:23,392
NEURONS IN OUR BRAINS, EVEN

352
00:13:23,392 --> 00:13:24,893
COMPARED TO TODAY'S COMPUTING

353
00:13:24,893 --> 00:13:26,228
CAPABILITIES, WE HAVE A LOOT OF

354
00:13:26,228 --> 00:13:28,697
SYNAPSES, A LOT OF -- LOT OF

355
00:13:28,697 --> 00:13:30,065
SYNAPSES, LOT OF TIMES MILLION

356
00:13:30,065 --> 00:13:31,967
SECONDS IN A MONTH. THEN THERE'S

357
00:13:31,967 --> 00:13:33,736
THE OTHER -- MILLISECONDS. THE

358
00:13:33,736 --> 00:13:36,271
OTHER ISSUE ON THE RIGHT T

359
00:13:36,271 --> 00:13:38,007
BIOLOGICAL COMPLEXITY IS PREMEN

360
00:13:38,007 --> 00:13:39,241
DOUSE. SO MUCH WE DON'T KNOW.

361
00:13:39,241 --> 00:13:41,143
ONE EXAMPLE OF WHAT WE DIDN'T

362
00:13:41,143 --> 00:13:42,077
KNOW THAT WILL BE IMPORTANT IF

363
00:13:42,077 --> 00:13:44,413
YOU ARE TALKING ABOUT

364
00:13:44,413 --> 00:13:46,649
NEUROLOGICAL CONDITIONS, MENTAL

365
00:13:46,649 --> 00:13:50,419
HEALTH IS THAT IT IS NOT JUST

366
00:13:50,419 --> 00:13:51,887
GABA AND GLUTAMATE AND DONE,

367
00:13:51,887 --> 00:13:54,423
THERE IS A LOT OF IMMENSE

368
00:13:54,423 --> 00:13:55,791
COMPLEXITY AND NEUROMODULATORY

369
00:13:55,791 --> 00:13:56,492
SYSTEMS THAT IS REALLY

370
00:13:56,492 --> 00:13:58,594
IMPORTANT. SO I -- WHEN I WAS

371
00:13:58,594 --> 00:13:59,528
FIRST THINKING ABOUT WHERE I

372
00:13:59,528 --> 00:14:05,401
WANT TO GO AFTER MY TIME AT ECSD

373
00:14:05,401 --> 00:14:07,269
AND SALK I WAS MOTIVATED AND

374
00:14:07,269 --> 00:14:08,971
INSPIRED TO THINK WHAT CAN --

375
00:14:08,971 --> 00:14:13,742
CAN WE START PUTTING THE

376
00:14:13,742 --> 00:14:17,913
COMPLEXITY OF DOPAMINE

377
00:14:17,913 --> 00:14:19,048
NOREPINEPHRINE, THESE MODULATORY

378
00:14:19,048 --> 00:14:20,315
SYSTEMS CRITICAL IN COGNITION

379
00:14:20,315 --> 00:14:22,418
AND AFFECT AND MOOD, THAT ARE

380
00:14:22,418 --> 00:14:27,222
OFTEN IMPLICATED IN DISORDERS.  

381
00:14:27,222 --> 00:14:28,957
WE KNOW FROM ELECTROPHYSIOLOGY

382
00:14:28,957 --> 00:14:30,359
WORK THESE ARE IMPORTANT WHEN IT

383
00:14:30,359 --> 00:14:33,996
COMES TO THINGS LIKE BEHAVIOR OF

384
00:14:33,996 --> 00:14:35,731
GRANULE CELLS AND INNER NEURON

385
00:14:35,731 --> 00:14:40,436
POPULATIONS. ALONG WITH POST DOC

386
00:14:40,436 --> 00:14:41,904
(INAUDIBLE) IN RUSTY'S LAB AT

387
00:14:41,904 --> 00:14:45,374
TIME WHO LEFT AND WENT TO START

388
00:14:45,374 --> 00:14:46,942
POSITION ON CHINA, I DID SOME

389
00:14:46,942 --> 00:14:51,680
MODELING WORK WHERE WE PUT

390
00:14:51,680 --> 00:14:54,950
DOPAMINE, VERY ABSTRACT MODEL SO

391
00:14:54,950 --> 00:14:56,418
DETAILS ARE BASICALLY WHAT YOU

392
00:14:56,418 --> 00:14:58,520
SEE AT THE TOP OF THE CHART,

393
00:14:58,520 --> 00:15:00,422
SIMPLIFIED MODEL OF DOPAMINE

394
00:15:00,422 --> 00:15:02,057
BASED ON PHYSIOLOGY RESULT SHE

395
00:15:02,057 --> 00:15:03,826
WAS SEEING WHERE WE OBSERVE WHEN

396
00:15:03,826 --> 00:15:05,794
YOU PUT THIN TO THE PATTERN

397
00:15:05,794 --> 00:15:06,962
SEPARATION MODEL -- PUT THIS

398
00:15:06,962 --> 00:15:08,397
INTO THE PATTERN SEPARATION

399
00:15:08,397 --> 00:15:11,300
MODEL WE HAVE YOU SEE A TEMPORAL

400
00:15:11,300 --> 00:15:13,869
ASYNCHRONY HAPPENING. IF I HAVE

401
00:15:13,869 --> 00:15:15,304
A BEHAVIORAL MODEL, ADDING

402
00:15:15,304 --> 00:15:18,140
NEURONS IN THERE, AND THEN I HAD

403
00:15:18,140 --> 00:15:19,775
A DOPAMINE EVENT, THINGS THAT

404
00:15:19,775 --> 00:15:21,677
HAPPEN BEFORE THE DOPAMINE EVENT

405
00:15:21,677 --> 00:15:22,945
WERE ENCODED DIFFERENTLY THAN

406
00:15:22,945 --> 00:15:24,379
THE THINGS THAT HAPPENED AFTER

407
00:15:24,379 --> 00:15:26,949
THE DOPAMINE EVENT. IF YOU ARE

408
00:15:26,949 --> 00:15:29,952
FAMILIAR WITH AI AND SOME OF THE

409
00:15:29,952 --> 00:15:32,454
AI TECHNIQUES HAVE GONE IN

410
00:15:32,454 --> 00:15:33,589
RECENT YEARS THIS IS NOT

411
00:15:33,589 --> 00:15:35,190
DIFFERENT FROM REINFORCEMENT

412
00:15:35,190 --> 00:15:37,960
LEARNING WHICH IS DOPAMINE

413
00:15:37,960 --> 00:15:39,628
IMPLICATED IN, AND DIFFERENCE

414
00:15:39,628 --> 00:15:41,230
LEARNING. THESE ARE FAIRLY

415
00:15:41,230 --> 00:15:42,598
FUNDAMENTAL BUILDING BLOCKS OF

416
00:15:42,598 --> 00:15:48,403
AI. SO WE FURTHER WENT AND THIS

417
00:15:48,403 --> 00:15:52,141
WAS WORK YANG LINK'S LAB DID

418
00:15:52,141 --> 00:15:54,476
BEHAVIORAL EXPERIMENTS AN

419
00:15:54,476 --> 00:15:55,110
PHYSIOLOGY EXPERIMENT WHERE IS

420
00:15:55,110 --> 00:15:57,713
YOU GIVE A MOUSE CONDENSED MILK

421
00:15:57,713 --> 00:15:59,782
AND YOU CAN OBSERVE IN AFFECT

422
00:15:59,782 --> 00:16:01,917
AFTER DOPAMINE EVENT OR

423
00:16:01,917 --> 00:16:04,019
PRESUMABLY CONDENSED MILK

424
00:16:04,019 --> 00:16:05,587
EQUIVALENT OF DOPAMINE EVENT,

425
00:16:05,587 --> 00:16:07,322
INDUCES A REWARD RESPONSE,

426
00:16:07,322 --> 00:16:09,758
LEARNING IS SUPPRESSED. NOT ALL

427
00:16:09,758 --> 00:16:11,994
THAT DIFFERENT THAN WHAT MODEL

428
00:16:11,994 --> 00:16:14,830
SUGGEST. THEN THIS WENT IN

429
00:16:14,830 --> 00:16:15,764
CONTEXT FEAR CONDITIONING AND

430
00:16:15,764 --> 00:16:17,966
OTHER THINGS, NOT FOR SAKE OF

431
00:16:17,966 --> 00:16:20,135
TIME NOT SHOW THOSE RESULTS. I

432
00:16:20,135 --> 00:16:20,836
THINK THIS WAS EXCITING BECAUSE

433
00:16:20,836 --> 00:16:22,404
IT SHOWED THAT YES, THERE IS A

434
00:16:22,404 --> 00:16:23,972
REAL REASON TO INCLUDE THE

435
00:16:23,972 --> 00:16:25,774
COMPLEXITY AND MODULATORY

436
00:16:25,774 --> 00:16:28,377
SYSTEMS IN THESE MODELS. BUT THE

437
00:16:28,377 --> 00:16:32,581
PROBLEM IS, THIS IS ONE

438
00:16:32,581 --> 00:16:33,682
MODULATOR WITH A LOT OF WORK

439
00:16:33,682 --> 00:16:35,984
THAT WAS DONE BY NUMBER OF

440
00:16:35,984 --> 00:16:37,920
EXPERIMENTALISTS ON THIS PAPER,

441
00:16:37,920 --> 00:16:42,157
PNAS AT THE TIME. AND WE WERE

442
00:16:42,157 --> 00:16:45,594
ABLE TO NAIL DOWN ONE

443
00:16:45,594 --> 00:16:47,362
EXPERIMENT, ONE LITTLE MODEL

444
00:16:47,362 --> 00:16:48,831
RESULT AND IT IS LIKE THERE IS

445
00:16:48,831 --> 00:16:51,266
SO MUCH MORE. IT IS JUST -- THIS

446
00:16:51,266 --> 00:16:55,037
IS CHIPPING OFF A TINY BIT AND

447
00:16:55,037 --> 00:16:56,104
HOPING THAT WE HAVE TO DO THIS

448
00:16:56,104 --> 00:16:58,640
FOREVER. SO THAT WAS A LITTLE

449
00:16:58,640 --> 00:17:00,776
UNDERWHELMING. IT WAS EXCITING,

450
00:17:00,776 --> 00:17:03,412
I LOVE THE RESULT BUT IS THIS A

451
00:17:03,412 --> 00:17:05,314
PATH FORWARD TO TRANSMITTER BY

452
00:17:05,314 --> 00:17:06,648
TRANSMITTER CELL TYPE BY CELL

453
00:17:06,648 --> 00:17:08,083
TYPE CHIP AWAY AT WHAT IS GOING

454
00:17:08,083 --> 00:17:13,088
ON. AS WAY OF FRAMING THIS, MANY

455
00:17:13,088 --> 00:17:16,024
OF YOU ARE FAMILIAR WITH PAT

456
00:17:16,024 --> 00:17:17,459
CHURCHHILL AND TERRY

457
00:17:17,459 --> 00:17:19,161
(INDISCERNIBLE) HAD A BOOK IN

458
00:17:19,161 --> 00:17:20,996
1988, TEN YEARS AGO HAD A

459
00:17:20,996 --> 00:17:22,164
FOLLOW-UP PAPER LOOKING AT TIME

460
00:17:22,164 --> 00:17:24,800
SCALES OF EXPERIMENTAL

461
00:17:24,800 --> 00:17:25,834
MODALITIES SPATIAL SCALES AND

462
00:17:25,834 --> 00:17:27,603
THIS DIFFERS TREMENDOUSLY

463
00:17:27,603 --> 00:17:30,172
WHETHER DOING SINGLE SPIKES AND

464
00:17:30,172 --> 00:17:33,609
PHYSIOLOGY OR FMRI IMAGING. THIS

465
00:17:33,609 --> 00:17:35,844
SAME THING HAPPENS WITH

466
00:17:35,844 --> 00:17:38,146
PLASTICITY. NEUROGENESIS IS

467
00:17:38,146 --> 00:17:40,215
OCCURRING OF COURSE AS I

468
00:17:40,215 --> 00:17:43,318
MENTIONED WEEKS AND MONTHS

469
00:17:43,318 --> 00:17:44,353
SCALES, CIRCUIT SCALES WITH HOD

470
00:17:44,353 --> 00:17:45,621
LAY TORRS ARE DIFFERENT TIME

471
00:17:45,621 --> 00:17:47,923
SCALES. SO IN THE SENSE YOU

472
00:17:47,923 --> 00:17:50,259
START LOOKING AT THIS, THE TIME

473
00:17:50,259 --> 00:17:53,061
SCALES THE SPATIAL SCALES, THE

474
00:17:53,061 --> 00:17:54,363
COMPLEXITY SCALES, BECOMES A

475
00:17:54,363 --> 00:17:57,199
LITTLE TAUNTING. -- DAUNTING BUT

476
00:17:57,199 --> 00:17:58,567
WE HAVE GIANT COMPUTERS.

477
00:17:58,567 --> 00:18:02,337
SHOULDN'T THAT BE ENOUGH? SO FOR

478
00:18:02,337 --> 00:18:03,572
THIS REASON, LOOK AT THAT

479
00:18:03,572 --> 00:18:04,907
COMPLEXITY THINKING I'M NOT

480
00:18:04,907 --> 00:18:07,409
GOING TO SOLVE THE PROBLEM ONE

481
00:18:07,409 --> 00:18:09,611
SYSTEM AT A TIME, ESPECIALLY AS

482
00:18:09,611 --> 00:18:12,714
A MODELER. I LOVE WORKING WITH

483
00:18:12,714 --> 00:18:14,983
EXPERIMENTALISTS BUT THAT IS NOT

484
00:18:14,983 --> 00:18:15,851
NECESSARILY THE BEST WAY TO

485
00:18:15,851 --> 00:18:18,820
START A CAREER IS WHERE YOU ARE

486
00:18:18,820 --> 00:18:19,855
DEPENDENT UPON BUNCH OF PEOPLE

487
00:18:19,855 --> 00:18:22,357
TO GIVE YOU DATA. SO I WEPT TO

488
00:18:22,357 --> 00:18:24,259
SANDIA NATIONAL LABS.  IF YOU

489
00:18:24,259 --> 00:18:26,094
ARE NOT FAMILIAR WITH DOE

490
00:18:26,094 --> 00:18:26,995
LABORATORIES THEY GENERALLY

491
00:18:26,995 --> 00:18:28,230
DON'T HAVE A LOT OF

492
00:18:28,230 --> 00:18:28,864
NEUROSCIENCE, THAT'S BEGINNING

493
00:18:28,864 --> 00:18:31,166
TO CHANGE.  AT THE TIME THEY

494
00:18:31,166 --> 00:18:33,101
DIDN'T HAVE MUCH AT ALL. BUT

495
00:18:33,101 --> 00:18:35,771
THEY DO HAVE REALLY BIG

496
00:18:35,771 --> 00:18:37,673
COMPUTERS. IN FACT THE BIGGEST

497
00:18:37,673 --> 00:18:40,042
COMPUTERS IN THE WORLD RIGHT NOW

498
00:18:40,042 --> 00:18:42,711
ARE OAKRIDGE NATIONAL LAB AND

499
00:18:42,711 --> 00:18:44,746
ARGON NATIONAL LAB. WE ARE

500
00:18:44,746 --> 00:18:46,315
GETTING TO THE 18 BLOTCHES A T A

501
00:18:46,315 --> 00:18:49,785
FILE, SANDIA, THIS IS 13, 40

502
00:18:49,785 --> 00:18:52,321
YEARS AGO SANDIA HAD ONE OF THE

503
00:18:52,321 --> 00:18:53,855
BEGGEST SUPERCOMPUTERS AVAILABLE

504
00:18:53,855 --> 00:18:55,958
TO SANDIA STAFF TO RUN

505
00:18:55,958 --> 00:18:57,960
EXPERIMENTS ON. SO I WENT AND

506
00:18:57,960 --> 00:19:00,162
SAID LET'S GOES TO REALISTIC

507
00:19:00,162 --> 00:19:06,468
SCALE DE DENTIGEROUS SIMULATION,

508
00:19:06,468 --> 00:19:08,136
MAYBE TO SOLVE THE PROBLEM. I

509
00:19:08,136 --> 00:19:08,670
DIDN'T IT SEEMS TO ME THE

510
00:19:08,670 --> 00:19:10,105
ORIGINAL MODEL BUT GRANULE CELLS

511
00:19:10,105 --> 00:19:11,106
WITH INTERNAL CORTEX INPUTS, I

512
00:19:11,106 --> 00:19:13,141
SAID LET'S LOOK AT ALL THE

513
00:19:13,141 --> 00:19:13,809
DIFFERENT INTERNEURON

514
00:19:13,809 --> 00:19:15,410
POPULATIONS.  THERE'S MANY MORE

515
00:19:15,410 --> 00:19:16,778
THAN THIS BUT THIS IS MORE THAN

516
00:19:16,778 --> 00:19:20,749
YOU SEE IN MOST MODELS. SIX

517
00:19:20,749 --> 00:19:22,684
DIFFERENT INNER NEURON

518
00:19:22,684 --> 00:19:24,886
POPULATIONS DIFFERENT TYPES OF

519
00:19:24,886 --> 00:19:26,822
CELLS, GRID CELLS AND NON-GRID

520
00:19:26,822 --> 00:19:28,757
CELL INPUTS AND LOOK AT

521
00:19:28,757 --> 00:19:31,493
SOPHISTICATED FUNCTION, PATTERN

522
00:19:31,493 --> 00:19:32,094
SEPARATION FUNCTION BEFORE AND

523
00:19:32,094 --> 00:19:34,896
ABILITY TO ENCODE INFORMATION. 

524
00:19:34,896 --> 00:19:36,798
WHAT WE OBSERVED, I BELIEVE THE

525
00:19:36,798 --> 00:19:38,934
FIRST TIME WE SAW THIS HAS BEEN

526
00:19:38,934 --> 00:19:41,403
SHOWN INTUITIVELY MAKES SENSE,

527
00:19:41,403 --> 00:19:43,739
IS THAT THERE IS A REAL VALUE OF

528
00:19:43,739 --> 00:19:46,675
GOING TO LARGE SCALE. THIS IS 15

529
00:19:46,675 --> 00:19:49,311
YEARS AGO, 25,000 NEURON MODEL

530
00:19:49,311 --> 00:19:51,013
WAS STILL AT THE TIME ONE OF THE

531
00:19:51,013 --> 00:19:52,914
BIGGEST MODELS AROUND. BUT THAT

532
00:19:52,914 --> 00:19:55,884
IS ONLY 10% SIZE OF MOUSE

533
00:19:55,884 --> 00:19:57,786
DENTIGEROUS, FOR CONTEXT RAT --

534
00:19:57,786 --> 00:19:59,688
MOUSE IS 300,000 CELLS WHICH YOU

535
00:19:59,688 --> 00:20:05,193
SEE TOP RIGHT, AND HUMAN IS

536
00:20:05,193 --> 00:20:07,662
ABOUT 10 MILLION CELLS. IF WE

537
00:20:07,662 --> 00:20:09,664
SCALE UP ON SUPERCOMPUTER AS WE

538
00:20:09,664 --> 00:20:12,134
INCREASE THE SIZE OF THE

539
00:20:12,134 --> 00:20:13,468
NETWORK, WE SEE ON THE BOTTOM

540
00:20:13,468 --> 00:20:15,203
RIGHT BASICALLY THE ACTIVITY OF

541
00:20:15,203 --> 00:20:18,707
THE OVERALL NETWORK DECREASES.

542
00:20:18,707 --> 00:20:20,008
THIS IS THERE ARE REAL REASONS

543
00:20:20,008 --> 00:20:22,244
FOR THIS, LOOK AT -- THERE IS A

544
00:20:22,244 --> 00:20:23,912
DIFFERENCE BETWEEN NEURON THAT

545
00:20:23,912 --> 00:20:26,181
HAS 10,000 SYNAPSES IN A NEURON

546
00:20:26,181 --> 00:20:28,283
THAT HAS 100 SYNAPSES OR A

547
00:20:28,283 --> 00:20:31,920
THOUSAND SYNAPSES. SCALING IS

548
00:20:31,920 --> 00:20:33,922
NOT LINEAR WHEN TALKING THINGS

549
00:20:33,922 --> 00:20:37,125
THAT ARE HIGHLY CONNECTED IN A

550
00:20:37,125 --> 00:20:38,693
COMPLEX GRAPH. A LOT OF MODELS

551
00:20:38,693 --> 00:20:41,063
THAT WE HAD, THIS KIND OF IS

552
00:20:41,063 --> 00:20:42,631
LIKE A WARNING SIGN, A LOT OF

553
00:20:42,631 --> 00:20:44,332
MODELS THAT LOOK AT BRAIN

554
00:20:44,332 --> 00:20:45,600
REGIONS AND SAY I HAVE A

555
00:20:45,600 --> 00:20:48,103
THOUSAND NEURON MODEL WHATEVER

556
00:20:48,103 --> 00:20:49,838
REGION, UNLESS YOUR REGION IS A

557
00:20:49,838 --> 00:20:55,177
THOUSAND NEURONS, LIKE A LOCUST

558
00:20:55,177 --> 00:20:57,345
AURELEUS MODEL, YOU ARE PROBABLY

559
00:20:57,345 --> 00:20:58,847
GETTING BIASED RESULTS BECAUSE

560
00:20:58,847 --> 00:20:59,915
THERE'S COMPLICATIONS IN

561
00:20:59,915 --> 00:21:01,450
SCALING. WHAT WAS INTERESTING

562
00:21:01,450 --> 00:21:04,553
THOUGH WAS THAT WHEN WE ADD NEW

563
00:21:04,553 --> 00:21:06,822
NEURONS TO THIS, WE SAW THAT THE

564
00:21:06,822 --> 00:21:11,226
OVERALL SCALING EFFECT DIDN'T

565
00:21:11,226 --> 00:21:13,795
CHANGE MATURE CELLS BUT YOUNG

566
00:21:13,795 --> 00:21:16,565
CELLS DIDN'T CARE, THE YOUNG

567
00:21:16,565 --> 00:21:18,700
NEURONS WEIRDLY, THEY ONLY --

568
00:21:18,700 --> 00:21:19,835
ALREADY HAVE FEWER CONNECTIONS

569
00:21:19,835 --> 00:21:22,537
THAN MATURE CELLS DO, THEY ARE

570
00:21:22,537 --> 00:21:23,538
SMALL NETWORK NEURONS TO BEGIN

571
00:21:23,538 --> 00:21:25,240
WITH SO I THAT DIDN'T CARE THE

572
00:21:25,240 --> 00:21:26,508
SIZE OF THE MODEL. BUT THE

573
00:21:26,508 --> 00:21:29,711
MATURE CELLS DID.  WE LOOKED AT

574
00:21:29,711 --> 00:21:30,412
PATTERN SEPARATION AND

575
00:21:30,412 --> 00:21:32,013
INFORMATION CODING, WE SAW THAT

576
00:21:32,013 --> 00:21:34,082
IT WAS IMPROVED WITH NEW

577
00:21:34,082 --> 00:21:35,684
NEURONS. SO THIS WAS EXCITING

578
00:21:35,684 --> 00:21:37,919
FOR ONE REASON, WHICH I DIDN'T

579
00:21:37,919 --> 00:21:39,955
REALLY GO INTO WHICH IS THAT AS

580
00:21:39,955 --> 00:21:41,423
I MENTION HUMAN NEUROGENESIS IS

581
00:21:41,423 --> 00:21:43,825
THOUGHT A LOW LEVEL. MUCH LOWER

582
00:21:43,825 --> 00:21:45,694
THAN MICE RELATIVELY BUT THIS

583
00:21:45,694 --> 00:21:46,761
WOULD SUGGEST AS YOU GO TO

584
00:21:46,761 --> 00:21:49,431
LARGER SCALES YOU DON'T NEED AS

585
00:21:49,431 --> 00:21:50,532
MANY NEW NEURONS. THAT WAS

586
00:21:50,532 --> 00:21:53,835
QUITING. BUT WHAT WASN'T

587
00:21:53,835 --> 00:21:55,937
EXCITING IS AFTER DOING ALL THAT

588
00:21:55,937 --> 00:21:57,339
WORK, AND THAT TOOK A WHILE TO

589
00:21:57,339 --> 00:22:00,876
BUILD THE MODEL AND RUN THESE

590
00:22:00,876 --> 00:22:03,011
THINGS, I REALIZE WE DIDN'T

591
00:22:03,011 --> 00:22:05,747
LEARN MUCH. THIS WAS A BIT

592
00:22:05,747 --> 00:22:07,082
UNDERWHELMING AT THE TIME

593
00:22:07,082 --> 00:22:08,550
BECAUSE WE WERE LOOKING -- I

594
00:22:08,550 --> 00:22:09,518
REALIZED IT WAS BECAUSE WE WERE

595
00:22:09,518 --> 00:22:11,520
LOOKING FOR PATTERN SEPARATION

596
00:22:11,520 --> 00:22:13,588
AND WE FOUND PATTERN SEPARATION 

597
00:22:13,588 --> 00:22:17,159
THIS IS A CHALLENGE THAT OFTEN

598
00:22:17,159 --> 00:22:18,760
HAPPENS IN NEUROSCIENCE, A BIG

599
00:22:18,760 --> 00:22:20,862
CHALLENGE IN SYSTEMS AND

600
00:22:20,862 --> 00:22:21,463
COMPUTATIONAL NEUROSCIENCE IN

601
00:22:21,463 --> 00:22:25,934
PARTICULAR, WHICH IS THAT THE

602
00:22:25,934 --> 00:22:27,836
SYSTEM, THINGS ARE SO

603
00:22:27,836 --> 00:22:28,770
COMPLICATED WE CONSTRUCT OUR

604
00:22:28,770 --> 00:22:30,906
EXPERIMENTS FOR HYPOTHESIS BUT

605
00:22:30,906 --> 00:22:32,040
SINCE SYSTEMS ARE COMPLICATED IT

606
00:22:32,040 --> 00:22:34,676
IS EASY TO FINDS VALIDATION. 

607
00:22:34,676 --> 00:22:36,111
CONFIRMATION BIAS VALE, IT IS

608
00:22:36,111 --> 00:22:37,112
CERTAINLY IN THE CASE OF

609
00:22:37,112 --> 00:22:38,780
SIMULATIONS.  THINGS LIKE

610
00:22:38,780 --> 00:22:39,648
FINDING THE CORRELATIONS IN DATA

611
00:22:39,648 --> 00:22:45,453
IS NOT HARD.  SO I WAS A LITTLE

612
00:22:45,453 --> 00:22:46,721
-- THIS WAS A LITTLE

613
00:22:46,721 --> 00:22:50,325
DISCOURAGING. AND IN A REAL WAY

614
00:22:50,325 --> 00:22:53,361
LED TO ME KIND OF DIALING BACK A

615
00:22:53,361 --> 00:22:56,698
LITTLE ON THE THE BOTTOMUP

616
00:22:56,698 --> 00:22:59,034
SIMULATION EFFORT. PART OF THE

617
00:22:59,034 --> 00:23:01,369
REASON WAS THAT WHEN YOU START

618
00:23:01,369 --> 00:23:03,171
SEEING, ASK WHAT YOU REALLY

619
00:23:03,171 --> 00:23:04,506
NEED, YOU NEED LOTS OF DATA; WE

620
00:23:04,506 --> 00:23:05,907
DIDN'T HAVE THAT MUCH AT THE

621
00:23:05,907 --> 00:23:07,809
TIME, GETTING IT BY DEN DAY WAS

622
00:23:07,809 --> 00:23:09,778
NOT JUST THE DEN DAY, EVERYTHING

623
00:23:09,778 --> 00:23:11,379
ELSE MATTERS TOO. THE TIME

624
00:23:11,379 --> 00:23:15,150
SCALES ARE A BIG DEAL. AND WE

625
00:23:15,150 --> 00:23:16,651
HAVE THIS BIG GROWING ISSUE THAT

626
00:23:16,651 --> 00:23:17,819
I DIDN'T KNOW WHAT TO LOOK FOR

627
00:23:17,819 --> 00:23:20,021
WHEN IT CAME TO COMPUTATIONAL

628
00:23:20,021 --> 00:23:21,590
MODELS. LIKE WHAT IS A TASK I

629
00:23:21,590 --> 00:23:24,025
WANT TO DO. CONTEXT FEAR

630
00:23:24,025 --> 00:23:25,293
CONDITIONING WE SAW EFFECTS ON

631
00:23:25,293 --> 00:23:28,063
IN MOUSE, IT INVOLVES A LOT OF

632
00:23:28,063 --> 00:23:32,934
SYSTEMS LOT MORE IN DENTATE SO

633
00:23:32,934 --> 00:23:34,636
HOW DO WE MATCH THESE? WE HAD

634
00:23:34,636 --> 00:23:36,638
THESE SORT OFFER HARD LESSONS --

635
00:23:36,638 --> 00:23:38,607
OF HARD LESSONS YOU CAN MODEL

636
00:23:38,607 --> 00:23:41,076
SOMETHING COMPLICATED OR SMALL

637
00:23:41,076 --> 00:23:43,612
UNCOMPLICATED OR BIG AND SIMPLE,

638
00:23:43,612 --> 00:23:47,749
WE NEED BETTER BEHAVIORAL TASK. 

639
00:23:47,749 --> 00:23:49,684
THIS WAS ME ABOUT TEN YEARS AGO.

640
00:23:49,684 --> 00:23:52,754
WHAT HAPPENED ABOUT TEN YEARS

641
00:23:52,754 --> 00:23:55,523
AGO? I THINK EVERYONE HERE IS

642
00:23:55,523 --> 00:23:56,591
FAMILIAR WITH ONE THING THAT

643
00:23:56,591 --> 00:23:58,159
HAPPENED, THERE WAS A MAJOR

644
00:23:58,159 --> 00:23:59,694
INVESTMENT IN THE BRAIN

645
00:23:59,694 --> 00:24:02,163
INITIATIVE THAT WANTED -- I

646
00:24:02,163 --> 00:24:03,898
WOULD SAY WONDERFULLY SAID LET'S

647
00:24:03,898 --> 00:24:06,668
GET TO THIS BIOLOGICAL

648
00:24:06,668 --> 00:24:08,236
COMPLEXITY CHALLENGE. WE DON'T

649
00:24:08,236 --> 00:24:11,973
KNOW ENOUGH. I SHOWED YOU WHY

650
00:24:11,973 --> 00:24:13,475
FROM ONE EXAMPLE WE DON'T KNOW

651
00:24:13,475 --> 00:24:14,843
ENOUGH ABOUT ASPECT OF THE BRAIN

652
00:24:14,843 --> 00:24:17,412
TO MAKE DEEPER INSIGHT. AND THE

653
00:24:17,412 --> 00:24:19,214
BRAIN INITIATIVE FACES HEAD ON.

654
00:24:19,214 --> 00:24:21,116
THE HUMAN BRAIN PROJECT DID THE

655
00:24:21,116 --> 00:24:22,417
SAME THING AT A DIFFERENT SCALE

656
00:24:22,417 --> 00:24:24,352
AND HAD A DIFFERENT EMPHASIS OF

657
00:24:24,352 --> 00:24:25,253
COURSE BUT IT WAS HAPPENING

658
00:24:25,253 --> 00:24:27,589
AROUND THE SAME TIME. AND

659
00:24:27,589 --> 00:24:28,990
ANOTHER THING HAPPENED AROUND

660
00:24:28,990 --> 00:24:30,358
THE THIS TIME WHICH IS

661
00:24:30,358 --> 00:24:31,726
ARTIFICIAL INTELLIGENCE

662
00:24:31,726 --> 00:24:36,131
EXPLODED. AI IS NOT BUY LOGICAL

663
00:24:36,131 --> 00:24:38,066
INTELLIGENCE, THAT IS WHY IT IS

664
00:24:38,066 --> 00:24:39,134
ARTIFICIAL, NOT BIOLOGICAL, WHAT

665
00:24:39,134 --> 00:24:41,136
I THINK IF YOU LOOK AT WHAT

666
00:24:41,136 --> 00:24:43,438
HAPPENED IN SOME WAYS IT DID THE

667
00:24:43,438 --> 00:24:45,206
OPPOSITE, IT STRIPPED WHATEVER

668
00:24:45,206 --> 00:24:46,441
BIOLOGICAL COMPLEXITY REMAINED

669
00:24:46,441 --> 00:24:47,842
IN THINGS LIKE NEURAL NETWORKS

670
00:24:47,842 --> 00:24:50,912
AND FOCUSED ON LET'S MAKING THIS

671
00:24:50,912 --> 00:24:52,881
BIG AS POSSIBLE AS SIMPLE AS

672
00:24:52,881 --> 00:24:53,848
POSSIBLE. BUT AT THE SAME TIME

673
00:24:53,848 --> 00:24:57,552
WE ARE DOING BOTH THE AB OHIO

674
00:24:57,552 --> 00:24:58,386
MAYBE THESE THINGS THINGS CAN

675
00:24:58,386 --> 00:24:59,354
COMBINE AND ONE WAY THEY COMBINE

676
00:24:59,354 --> 00:25:01,723
IS NEUROMORPHIC COMPUTING. THAT

677
00:25:01,723 --> 00:25:03,291
CHUTELY I BELIEVE IS A PATH BY

678
00:25:03,291 --> 00:25:06,861
WHICH WE CAN GET COMPLEXITY AND

679
00:25:06,861 --> 00:25:08,430
SCALE TOGETHER AND THIS

680
00:25:08,430 --> 00:25:11,199
SOMETHING THAT IS NOT IN LIEU OF

681
00:25:11,199 --> 00:25:14,069
THESE OTHER THINGS. WE

682
00:25:14,069 --> 00:25:15,470
ABSOLUTELY HAVE TO HAVE THIS

683
00:25:15,470 --> 00:25:20,241
FOCUS ON BIOLOGICAL COMPLEXITY.

684
00:25:20,241 --> 00:25:21,309
THE EYE FIELD IS NOT GOING

685
00:25:21,309 --> 00:25:22,644
ANYWHERE. ARTIFICIAL NETWORKS

686
00:25:22,644 --> 00:25:24,579
ARE THE WORLD WE LIVE IN NOW.

687
00:25:24,579 --> 00:25:25,680
HOW DO WE TAKE ADVANTAGE OF

688
00:25:25,680 --> 00:25:27,349
THIS? WE CAN START PUSHING TO

689
00:25:27,349 --> 00:25:33,254
THE TOP RIGHT OF THIS. SO WHAT

690
00:25:33,254 --> 00:25:34,489
IS NEUROMORPHIC COMPUTING? I

691
00:25:34,489 --> 00:25:35,957
HAVE THROWN ET AROUND A FEW

692
00:25:35,957 --> 00:25:37,092
TIMES. NEUROMORPHIC COMPUTEING

693
00:25:37,092 --> 00:25:41,262
IS A FIELD EXISTED -- PARDON ME,

694
00:25:41,262 --> 00:25:43,331
EXISTED SINCE THE 19 # 0s,

695
00:25:43,331 --> 00:25:47,802
CARBON MEAD IS -- CARVER MEAD,

696
00:25:47,802 --> 00:25:49,104
ELECTRICAL ENGINEER CRITICAL AT

697
00:25:49,104 --> 00:25:51,239
STAGES OF BUILDING INTEGRATED

698
00:25:51,239 --> 00:25:52,407
CIRCUITS. TOWARDS THE ENDS OF

699
00:25:52,407 --> 00:25:56,978
HIS RESEARCH CAREER, HE AND GRAD

700
00:25:56,978 --> 00:25:59,147
STUDENT (INDISCERNIBLE) DECIDED

701
00:25:59,147 --> 00:26:01,816
TO BEGIN TO EMULATE PARTS OF

702
00:26:01,816 --> 00:26:02,851
NEURAL SYSTEMS AS UNDERSTOOD IN

703
00:26:02,851 --> 00:26:07,622
THE '80s IN SILICON. THAT SORT

704
00:26:07,622 --> 00:26:11,092
OF IS MEAD'S VISION ORIGINALLY,

705
00:26:11,092 --> 00:26:12,994
LET'S PUT WHAT WE KNOW ABOUT

706
00:26:12,994 --> 00:26:16,131
DIFFERENT SENSORY SYSTEMS AND

707
00:26:16,131 --> 00:26:18,166
BUILD SILICON CHIPS OUT OF IT TO

708
00:26:18,166 --> 00:26:20,101
CAPTURE THAT. THAT FIELD TOOK

709
00:26:20,101 --> 00:26:22,270
OFF. THERE IS A COMMUNITY OF

710
00:26:22,270 --> 00:26:24,973
RESEARCHERS THEY FOCUS A LOT ON

711
00:26:24,973 --> 00:26:26,241
THINGS LIKE EVENT BASE SENSORS

712
00:26:26,241 --> 00:26:28,410
THAT LED TO ARTIFICIAL RETINAS

713
00:26:28,410 --> 00:26:31,679
AND COCHLEAS AND SO FORTH, BASED

714
00:26:31,679 --> 00:26:33,448
CAMERAS. AND ABOUT TEN YEARS AGO

715
00:26:33,448 --> 00:26:35,683
THAT FIELD WAS GOING ALONG LOOK

716
00:26:35,683 --> 00:26:39,120
AT ANALOG, NON-DIGITAL SILICON

717
00:26:39,120 --> 00:26:42,524
ELECTRONICS, ABOUT TEN YEARS AGO

718
00:26:42,524 --> 00:26:45,026
THERE WAS A RENEWAL OF INTEREST

719
00:26:45,026 --> 00:26:47,228
FROM THE MOORE COMPUTER SCIENCE

720
00:26:47,228 --> 00:26:48,263
COMMUNITY, PARTLY DRIVEN BY

721
00:26:48,263 --> 00:26:50,665
STUFF GOING ON IN AI LOOKING

722
00:26:50,665 --> 00:26:53,802
TOWARDS LARGE SCALE SYSTEMS.  

723
00:26:53,802 --> 00:26:54,836
BASICALLY BECAUSE THE BRAIN IS

724
00:26:54,836 --> 00:26:56,738
REALLY EFFICIENT AT COMPUTING

725
00:26:56,738 --> 00:26:58,106
AND THE COMPUTERS TODAY ARE NOT

726
00:26:58,106 --> 00:27:02,010
EFFICIENT AT COMPUTING. IN TERMS

727
00:27:02,010 --> 00:27:03,912
OF ENERGY. SO THERE ARE TWO

728
00:27:03,912 --> 00:27:05,280
BROAD CLASSES OF NEUROMORPHIC

729
00:27:05,280 --> 00:27:08,650
COMPUTING. I WON'T GO INTO GORY

730
00:27:08,650 --> 00:27:12,387
DETAILDETAILS.  WE FOCUS AT SANN

731
00:27:12,387 --> 00:27:14,756
LARGE DIGITAL CMOSS TECHNOLOGY.

732
00:27:14,756 --> 00:27:16,658
THESE ARE THE UNDERLYING

733
00:27:16,658 --> 00:27:19,727
FABRICATION TECHNOLOGY OF A

734
00:27:19,727 --> 00:27:22,964
INTEL CHIP YOU ARE SEEING BOXED

735
00:27:22,964 --> 00:27:24,732
HERE BOTTOM RIGHT AT SANDIA.

736
00:27:24,732 --> 00:27:27,368
THOSE CHIPS ARE THE SAME

737
00:27:27,368 --> 00:27:28,436
MATERIALS AND FABRICATION IN

738
00:27:28,436 --> 00:27:30,939
YOUR CELL PHONE, LAPTOP,

739
00:27:30,939 --> 00:27:34,576
WHATEVER. BUT THE ARCHITECTURE

740
00:27:34,576 --> 00:27:36,411
IS NOT PROGRAMMABLE THE WAY YOU

741
00:27:36,411 --> 00:27:38,179
PROGRAM A NORMAL COMPUTER, IT

742
00:27:38,179 --> 00:27:42,116
DOESN'T NATURALLY REPYTHON, IT

743
00:27:42,116 --> 00:27:43,551
IS AN ARCHITECTURE RATHER THAT

744
00:27:43,551 --> 00:27:45,553
LOOKS MORE LIKE THE BRAIN,

745
00:27:45,553 --> 00:27:48,356
PROGRAM CIRCUITS OF NEURONS. THE

746
00:27:48,356 --> 00:27:49,824
OTHER ALTERNATIVE IS ANALOG

747
00:27:49,824 --> 00:27:51,359
SYSTEMS, I WON'T GO INTO DETAIL,

748
00:27:51,359 --> 00:27:55,463
WE DO THIS WORK AT SANDIA. IT IS

749
00:27:55,463 --> 00:27:57,866
FURTHER OUT, MORE CAPTURING THE

750
00:27:57,866 --> 00:28:00,134
BIOPHYSIC, HAS LONGER PATH

751
00:28:00,134 --> 00:28:03,872
TOWARDS SCALING, PROBABLY THE

752
00:28:03,872 --> 00:28:06,641
FUTURE BUT I'M GOING TO FOCUS

753
00:28:06,641 --> 00:28:10,578
MORE ON DIGITAL CMOS TECHNOLOGY.

754
00:28:10,578 --> 00:28:13,147
THIS IS MORE COMPUTER ARCHITECT

755
00:28:13,147 --> 00:28:13,781
OAR CENTED SLIDE, THERE'S LIST

756
00:28:13,781 --> 00:28:16,718
OF THINGS FROM THE BRAIN

757
00:28:16,718 --> 00:28:17,252
TRANSLATED TO ELECTRICAL

758
00:28:17,252 --> 00:28:18,019
ENGINEERING OR COMPUTER

759
00:28:18,019 --> 00:28:21,089
ENGINEERING CONCEPTS LIKE

760
00:28:21,089 --> 00:28:22,257
SPIKING COMPLEX NETWORKS,

761
00:28:22,257 --> 00:28:23,858
ABILITY OF SYNAPSES AND NEURONS

762
00:28:23,858 --> 00:28:25,927
TO EXIST IN THE SAME PHYSICAL

763
00:28:25,927 --> 00:28:27,295
LOCATION, LEARNING. ALL THESE

764
00:28:27,295 --> 00:28:28,997
THINGS ARE WHAT THESE HARDWARE

765
00:28:28,997 --> 00:28:32,033
PLATFORMS TRY TO CAPTURE. IT IS

766
00:28:32,033 --> 00:28:33,701
RAPIDSLY EVOLVING FIELD.  I HI

767
00:28:33,701 --> 00:28:37,005
THE FULL DETAILS OF IT THEY WILL

768
00:28:37,005 --> 00:28:38,206
CONTINUES TO CHANGE OVER THE

769
00:28:38,206 --> 00:28:40,408
COURSE OF THE COMING YEARS AND

770
00:28:40,408 --> 00:28:42,544
DECADES BUT RIGHT NOW WE CAN GET

771
00:28:42,544 --> 00:28:44,112
A BILLION NEURONS ON TO EACH OF

772
00:28:44,112 --> 00:28:45,813
THESE LITTLE CHIPS YOU SEE ON

773
00:28:45,813 --> 00:28:47,615
THE TOP RIGHT, MILLION NEURONS

774
00:28:47,615 --> 00:28:50,251
EACH OF THESE CHIPS. WE CAN GET

775
00:28:50,251 --> 00:28:51,486
BILLIONS OF NEURONS IN A

776
00:28:51,486 --> 00:28:52,787
PLATFORM WHICH IS ABOUT THE SIZE

777
00:28:52,787 --> 00:28:55,056
OF THIS LARGE MICROWAIVE HEAR,

778
00:28:55,056 --> 00:28:56,958
WE HAVE ONE AT SANDIA, THE

779
00:28:56,958 --> 00:28:58,326
LARGEST NEUROMORPHIC SYSTEM THAT

780
00:28:58,326 --> 00:29:00,895
IS IN OUR LAB THERE. EACH OF HE

781
00:29:00,895 --> 00:29:05,033
IS CHIPS IS ONE WATT OF POWER.

782
00:29:05,033 --> 00:29:07,735
THAT IS ABOUT 100TH OF CPU THAT

783
00:29:07,735 --> 00:29:09,737
MIGHT BE IN YOUR LAPTOP AND

784
00:29:09,737 --> 00:29:12,373
MAYBE ABOUT A THOUSANDTH OF

785
00:29:12,373 --> 00:29:17,245
PRODUCTION LEVEL GPU. GIVE OR

786
00:29:17,245 --> 00:29:19,814
TAKE. I WILL TALK NEUROMORPHIC,

787
00:29:19,814 --> 00:29:21,416
GIVE THREE VIGNETTES THIS LAST

788
00:29:21,416 --> 00:29:23,785
PART OF THE TALK. WHERE I WANT

789
00:29:23,785 --> 00:29:26,821
TO SHOW WHAT NEUROMORPHIC VIEW

790
00:29:26,821 --> 00:29:28,590
OF COMPUTATION CAN TELL US ABOUT

791
00:29:28,590 --> 00:29:33,494
BRAIN. GIVEN THIS SORT OF

792
00:29:33,494 --> 00:29:35,063
NEUROGENESIS INSPIRED VISION

793
00:29:35,063 --> 00:29:36,230
BEFORE WHERE THOSE CHALLENGES

794
00:29:36,230 --> 00:29:38,766
ARE. THE OBVIOUS ONE IS THAT WE

795
00:29:38,766 --> 00:29:40,602
CAN GET TO THINGS AT SCALE. THAT

796
00:29:40,602 --> 00:29:41,769
IS WHAT I WILL START WITH. THE

797
00:29:41,769 --> 00:29:44,539
SECOND ONE IS THAT I BELIEVE IT

798
00:29:44,539 --> 00:29:45,940
PROVIDES INTERESTING CONSTRAINTS

799
00:29:45,940 --> 00:29:47,742
HOW TO ANALYZE DATA, THAT COMES

800
00:29:47,742 --> 00:29:49,978
OUT OF LARGER NEURAL SYSTEMS

801
00:29:49,978 --> 00:29:52,180
BECAUSE IN SOME LEVEL IF WE HAVE

802
00:29:52,180 --> 00:29:53,748
MILLIONS OF NEURONS IN CHIP WE

803
00:29:53,748 --> 00:29:55,083
ARE GOING TO ANALYZE THAT DATA

804
00:29:55,083 --> 00:29:57,418
IF WE SIMULATE IT. THE FINAL

805
00:29:57,418 --> 00:29:59,787
THING WHICH IS A LITTLE BIT MORE

806
00:29:59,787 --> 00:30:02,590
OUT THERE, I WILL ADMIT, IS THAT

807
00:30:02,590 --> 00:30:04,525
WE MIGHT BE ABLE TO ACTUALLY

808
00:30:04,525 --> 00:30:05,226
UNCOVER SOME NEW DIRECTIONS TO

809
00:30:05,226 --> 00:30:06,961
THINK ABOUT HOW THE BRAIN IS

810
00:30:06,961 --> 00:30:08,630
WORKING, BY LOOKING WHAT THESE

811
00:30:08,630 --> 00:30:11,032
CHIPS ARE GOOD AT.  SO START OFF

812
00:30:11,032 --> 00:30:12,166
BY TALKING A LITTLE BIT ABOUT

813
00:30:12,166 --> 00:30:13,701
OUR ABILITY TO SIMULATE NEURAL

814
00:30:13,701 --> 00:30:17,972
SYSTEMS AT SCALE. SO OF COURSE

815
00:30:17,972 --> 00:30:22,477
THE DREAM IS IF I HAD A

816
00:30:22,477 --> 00:30:24,078
CONNECTOME OF A BRAIN,

817
00:30:24,078 --> 00:30:27,515
EVENTUALLY HUMAN BRAIN BUT

818
00:30:27,515 --> 00:30:28,616
CONNECTOME OF WHATEVER BRAIN YOU

819
00:30:28,616 --> 00:30:30,551
CARE ABOUT CAN I MODEL THAT

820
00:30:30,551 --> 00:30:33,354
CONNECTOME RELATIVELY EASILY ON

821
00:30:33,354 --> 00:30:36,024
A NEUROMORPHIC CHIP? CONNECTOMES

822
00:30:36,024 --> 00:30:37,291
ARE LOTS AND LOTS OF NEURONS,

823
00:30:37,291 --> 00:30:40,528
LOTS OF SYNAPSES. WE GENERALLY

824
00:30:40,528 --> 00:30:43,398
DON'T HAVE THEM. WELL AS MANY OF

825
00:30:43,398 --> 00:30:45,566
YOU KNOW, ABOUT A MONTH AGO LESS

826
00:30:45,566 --> 00:30:50,938
THAN A MONTH AGO THE FLY WIRE

827
00:30:50,938 --> 00:30:52,907
TEAM (INDISCERNIBLE) AND

828
00:30:52,907 --> 00:30:54,642
COLLEAGUES LARGELY PRINCETON THE

829
00:30:54,642 --> 00:30:56,644
BRAIN INITIATIVE HELPED FUND A

830
00:30:56,644 --> 00:30:59,981
LARGE PART OF, THE TEAM

831
00:30:59,981 --> 00:31:02,183
PUBLISHED A SERIES OF FIVE

832
00:31:02,183 --> 00:31:05,720
PAPERS, THESE ARE NEW, I MED MY

833
00:31:05,720 --> 00:31:06,954
WAY THROUGH A COUPLE OF THEM,

834
00:31:06,954 --> 00:31:07,889
NOT THROUGH EVERY DETAIL YET BUT

835
00:31:07,889 --> 00:31:09,657
PART OF WHAT THEY HAD IN THIS

836
00:31:09,657 --> 00:31:14,295
PAPER IS A FULLOME OF A

837
00:31:14,295 --> 00:31:15,096
DROSOPHILA BRAIN.  THIS IS

838
00:31:15,096 --> 00:31:17,565
EXCITING BECAUSE WE HAD A

839
00:31:17,565 --> 00:31:20,301
CONNECTOME FOR A LONG TIME OF C

840
00:31:20,301 --> 00:31:22,070
ELEGANS BUT 307 OR SO NEURONS

841
00:31:22,070 --> 00:31:24,172
THAT ARE MOSTLY ANALOG IS

842
00:31:24,172 --> 00:31:26,107
DIFFERENT FROM A MODEL SYSTEM

843
00:31:26,107 --> 00:31:27,809
THAT IS IN THE HUNDREDS OF

844
00:31:27,809 --> 00:31:30,478
THOUSANDS OF NEURONS. AND CAN DO

845
00:31:30,478 --> 00:31:32,447
SOME VERY SOPHISTICATED

846
00:31:32,447 --> 00:31:34,182
BEHAVIORS. SO ONE OF THE PAPERS

847
00:31:34,182 --> 00:31:38,186
PUBLISHED HERE IS A BOTTLE OF

848
00:31:38,186 --> 00:31:41,289
SENSORY MOTOR PROCESSING

849
00:31:41,289 --> 00:31:42,390
OLFACTORY RESPONSE, ON -- THAT

850
00:31:42,390 --> 00:31:45,927
IS A MODEL OF THE FULL BRAIN. SO

851
00:31:45,927 --> 00:31:48,162
ADD THE CONNECTOME AND IMPLEMENT

852
00:31:48,162 --> 00:31:50,465
AD MODEL OF IT.  SO WE ASKED THE

853
00:31:50,465 --> 00:31:53,334
QUESTION, I DIDN'T DO IT,

854
00:31:53,334 --> 00:31:55,069
BECAUSE I HAVE BEEN TRAVELING

855
00:31:55,069 --> 00:31:57,371
BUT COUPLE OF MY COLLEAGUES HERE

856
00:31:57,371 --> 00:32:02,910
AT SANDIA, (INDISCERNIBLE), THEY

857
00:32:02,910 --> 00:32:04,045
DECIDED JUST KIND OF LIKE WHAT

858
00:32:04,045 --> 00:32:06,114
WE HAD THESE CHIPS, WE HAVE THIS

859
00:32:06,114 --> 00:32:07,582
CONNECTOME E THEY PUBLISH

860
00:32:07,582 --> 00:32:08,683
EVERYTHING WITH IT.  WE NEVER

861
00:32:08,683 --> 00:32:10,685
TALKED TO THE FIVE WIRE PEOPLE

862
00:32:10,685 --> 00:32:12,954
THEY DON'T KNOW WE DID THIS. BUT

863
00:32:12,954 --> 00:32:14,555
THEY PUBLISHED IT, IT IS THERE,

864
00:32:14,555 --> 00:32:16,891
AND WE SAID LET'S TAKE YOUR

865
00:32:16,891 --> 00:32:19,093
MODEL AND OR TAKE THAT MODEL AND

866
00:32:19,093 --> 00:32:22,163
STICK IT THROUGH A COUPLE OF OUR

867
00:32:22,163 --> 00:32:23,531
TOOLS, (INDISCERNIBLE) A SPIKING

868
00:32:23,531 --> 00:32:24,365
PROGRAMMING STOOL AND STACKS

869
00:32:24,365 --> 00:32:27,502
WHICH IS A LARGE SCALE

870
00:32:27,502 --> 00:32:29,670
SIMULATION TOOL AND MAP ON TO

871
00:32:29,670 --> 00:32:32,306
LOIHI. THEY DID, I WILL GIVE

872
00:32:32,306 --> 00:32:34,008
THEM THE CREDIT HERE BUT WE

873
00:32:34,008 --> 00:32:35,476
ACTUALLY WERE ABLE TO TAKE THIS

874
00:32:35,476 --> 00:32:37,278
MODEL OFF THE SHELF, LITS RALLY

875
00:32:37,278 --> 00:32:41,749
OFF THE SHELF, DOWNLOAD FROM GET

876
00:32:41,749 --> 00:32:44,519
LAB AND JUST STICK ON TO LOIHI

877
00:32:44,519 --> 00:32:49,590
AND I TOOK WORK, IT WASN'T

878
00:32:49,590 --> 00:32:50,958
INSTANTANEOUS BUT WE SEE

879
00:32:50,958 --> 00:32:52,960
DYNAMICS COMING OFF NEURONS IN

880
00:32:52,960 --> 00:32:55,096
THE CHIPS THAT ARE LARGELY

881
00:32:55,096 --> 00:32:58,666
ANALOGOUS TO WHAT WE SHOULD SEE.

882
00:32:58,666 --> 00:33:00,568
THIS IS -- I MEAN QUITE

883
00:33:00,568 --> 00:33:02,003
LITERALLY HOT OFF THE PRESSES.

884
00:33:02,003 --> 00:33:03,538
WE HAVEN'T GONE THROUGH

885
00:33:03,538 --> 00:33:04,739
VALIDATED EVERY SINGLE THING

886
00:33:04,739 --> 00:33:06,307
YET, THERE'S SOME DESIGN

887
00:33:06,307 --> 00:33:07,341
CONSIDERATION WE HAD TO DO BUT

888
00:33:07,341 --> 00:33:10,278
THIS SAYS THESE CHIPS COMING OFF

889
00:33:10,278 --> 00:33:14,148
THINGS LIKE INTEL'S LOIHI AND

890
00:33:14,148 --> 00:33:16,918
SPINNAKER, SPIN CLOUD SPINNAKER

891
00:33:16,918 --> 00:33:18,152
TWO CHIPS CAN RUN BIOLOGICAL

892
00:33:18,152 --> 00:33:21,689
MODELS AND IT IS A QUESTION OF

893
00:33:21,689 --> 00:33:23,658
ENGINEERING NOW HOW WE FIT THE

894
00:33:23,658 --> 00:33:25,393
BIOLOGY DATA WE ARE PRODUCING OR

895
00:33:25,393 --> 00:33:27,028
Y'ALL ARE PRODUCING TO THE CHIPS

896
00:33:27,028 --> 00:33:28,996
BEING PRODUCED.  SO H IS PRETTY

897
00:33:28,996 --> 00:33:32,500
EXCITING. I THINK THIS WAS QUITE

898
00:33:32,500 --> 00:33:33,868
EXCITING AND NEUROMORPHIC

899
00:33:33,868 --> 00:33:34,869
COMPUTING HAS A REAL VALUE. 

900
00:33:34,869 --> 00:33:36,904
THIS IS NOT A HUGE NETWORK

901
00:33:36,904 --> 00:33:38,673
NOWADAYS FOR PUT ON GPU. I DON'T

902
00:33:38,673 --> 00:33:41,809
THINK THERE IS A REAL -- WE

903
00:33:41,809 --> 00:33:43,411
DON'T HAVE TO HAVE NEUROMORPHIC

904
00:33:43,411 --> 00:33:46,447
TO STUDY FRUIT FLY BRAINS WE MAY

905
00:33:46,447 --> 00:33:48,149
FIND VALUE BUT AS WE SCALE

906
00:33:48,149 --> 00:33:50,051
LARGER SYSTEMS WE WILL. SO THIS

907
00:33:50,051 --> 00:33:53,955
IS -- THERE IS A PATH. SO THE

908
00:33:53,955 --> 00:33:58,693
SECOND THING. SO IF YOU CAN SEE

909
00:33:58,693 --> 00:34:00,394
FROM THIS LAST PLOT THERE ARE A

910
00:34:00,394 --> 00:34:03,598
LOT OF NEURONS.  THIS IS ONLY

911
00:34:03,598 --> 00:34:06,334
PLOTTING HUNDRED THOUSAND

912
00:34:06,334 --> 00:34:09,237
NEURONS. SYSTEM WHEN WE SIMULATE

913
00:34:09,237 --> 00:34:11,539
ACTUAL CORTICAL NEURONS, WE DID

914
00:34:11,539 --> 00:34:14,075
ONE WITH SUPERCOMPUTERS OVER 26

915
00:34:14,075 --> 00:34:17,111
MILLION NEURONS USING THE STACKS

916
00:34:17,111 --> 00:34:17,879
TOOL, 20 MILLION NEURONS COMING

917
00:34:17,879 --> 00:34:20,481
OUT MODELING LOSE COURTSCAL

918
00:34:20,481 --> 00:34:22,049
ARCHITECT COURT CAT CONNECTIONS

919
00:34:22,049 --> 00:34:24,986
WE CALL THIS A MASTER PLOT IN

920
00:34:24,986 --> 00:34:26,587
SYSTEMS NEURO WHERE YOU HAVE

921
00:34:26,587 --> 00:34:28,289
OVER TIME LOTS OF NEURONS,

922
00:34:28,289 --> 00:34:29,624
FIRING AND NOT FIRING. THIS IS

923
00:34:29,624 --> 00:34:33,027
JUST LOOKS LIKE NOISE. SO WE

924
00:34:33,027 --> 00:34:35,863
EFFECTIVELY CREATE AD NEW BIG

925
00:34:35,863 --> 00:34:37,665
DATA PROBLEM. -- CREATED A NEW

926
00:34:37,665 --> 00:34:38,799
BIG DATA PROBLEM. ONE THING THAT

927
00:34:38,799 --> 00:34:39,967
WAS AN INTERESTING THOUGHT WHEN

928
00:34:39,967 --> 00:34:41,369
WE PUT THESE BIG NEURAL

929
00:34:41,369 --> 00:34:42,970
SIMULATIONS ON TO THESE

930
00:34:42,970 --> 00:34:45,840
SUPERCOMPUTERS, THIS WAS ON THE

931
00:34:45,840 --> 00:34:47,742
OAKRIDGE SUMMIT MACHINE, WE SAW

932
00:34:47,742 --> 00:34:50,444
-- WE HAVE THIS TEMPTATION WE

933
00:34:50,444 --> 00:34:53,748
ALL HAVE, I'M SILL LATED EVERY

934
00:34:53,748 --> 00:34:55,516
NEURON AND SHOULD RECORD EVERY

935
00:34:55,516 --> 00:34:57,184
NEURON WHICH CREATE AS DATA

936
00:34:57,184 --> 00:34:59,687
PROBLEM BUT WHEN YOU DO THE

937
00:34:59,687 --> 00:35:02,590
NEUROMORPHIC SYSTEM IF YOU

938
00:35:02,590 --> 00:35:05,126
RECORD FROM EVERY NEURON FROM

939
00:35:05,126 --> 00:35:05,960
EVERY SYSTEM RATION YOU LOSE THE

940
00:35:05,960 --> 00:35:08,062
ADVANTAGE OF BEING NEUROMORPHIC

941
00:35:08,062 --> 00:35:09,363
IF YOU HAVE TO GIGABYTES OF

942
00:35:09,363 --> 00:35:14,035
DATA. SO INPUT OUTPUT IS

943
00:35:14,035 --> 00:35:15,603
EXPENSIVE SO WE WANTED TO

944
00:35:15,603 --> 00:35:17,104
OBSERVE A SUBSET OF ACTIVITY

945
00:35:17,104 --> 00:35:18,272
WHICH IS WHAT WE DEAL WITH WITH

946
00:35:18,272 --> 00:35:19,640
THE BRAIN. I CAN'T IMAGINE

947
00:35:19,640 --> 00:35:21,075
FUTURE WHERE WE ARE GOING TO

948
00:35:21,075 --> 00:35:23,744
RECORD THE SPIKING ACTIVITY OF

949
00:35:23,744 --> 00:35:24,979
PATIENTS OF EVERY NEURON IN

950
00:35:24,979 --> 00:35:25,913
PATIENT BRAIN. WE ARE ALWAYS

951
00:35:25,913 --> 00:35:26,914
GOING TO BE LOOKING AT A

952
00:35:26,914 --> 00:35:29,016
SUBSETS.  SO WHAT DO WE DO IF

953
00:35:29,016 --> 00:35:30,851
ONLY LOOK AT SUBSET? HOW DO WE

954
00:35:30,851 --> 00:35:34,288
GO FROM SPIKES TO SOMETHING A

955
00:35:34,288 --> 00:35:36,824
LITTLE MORE ABSTRACT? THIS IS

956
00:35:36,824 --> 00:35:38,526
WORK BY MY POST DOC WHO CAME

957
00:35:38,526 --> 00:35:40,828
FROM THE BIRD SONG COMMUNITY IN

958
00:35:40,828 --> 00:35:42,863
GET NEAR'S LAB AT UCSD. HE HAD

959
00:35:42,863 --> 00:35:44,465
THIS REALIZATION THAT AS WE ARE

960
00:35:44,465 --> 00:35:46,801
LOOK AT THIS LARGE SCALE DATA WE

961
00:35:46,801 --> 00:35:49,603
TENDS TO DO WHAT PEOPLE TRAINED

962
00:35:49,603 --> 00:35:52,873
IN SERIAL COMPUTING PROGRAM DO,

963
00:35:52,873 --> 00:35:55,876
IS THINK ABOUT CHUNK OF TIME,

964
00:35:55,876 --> 00:35:57,345
WHAT SPIKING EVENTS HAPPEN

965
00:35:57,345 --> 00:35:58,779
DIFFERENT PERIOD OF TIME AND

966
00:35:58,779 --> 00:36:01,549
THAT SORT OF PATTERN OF TIME

967
00:36:01,549 --> 00:36:03,417
PROGRESSES FORWARD, TRAJECTORY

968
00:36:03,417 --> 00:36:05,953
THROUGH TIME AND SPACE.  AND

969
00:36:05,953 --> 00:36:08,622
WHAT BRAD'S REALIZATION WAS

970
00:36:08,622 --> 00:36:10,091
THERE IS NO REASON THE BRAIN

971
00:36:10,091 --> 00:36:12,560
NEEDS TO BE -- THE BRAIN IS NOT

972
00:36:12,560 --> 00:36:14,562
CHUNKING THINGS UP IN TIME, TIME

973
00:36:14,562 --> 00:36:15,596
IS AN ARTIFICIAL CONSTRUCT

974
00:36:15,596 --> 00:36:16,764
BECAUSE THAT'S HOW WE MEASURE

975
00:36:16,764 --> 00:36:18,566
AND SYSTEM LATE THINGS BUT THESE

976
00:36:18,566 --> 00:36:20,368
NEURONS ARE -- THEY ARE --

977
00:36:20,368 --> 00:36:23,604
SIMULATE THINGS, THEY ARE FIRING

978
00:36:23,604 --> 00:36:25,606
ASYNCHRONOUSLY BASED ON THEIR

979
00:36:25,606 --> 00:36:27,308
INPUTS, MAYBE A LITTLE LOOSE

980
00:36:27,308 --> 00:36:28,042
INFORMATION.

981
00:36:28,042 --> 00:36:29,677
THAT'S WORK LOOK AT CAUSAL

982
00:36:29,677 --> 00:36:30,811
RELATIONSHIPS BETWEEN NEURONS IN

983
00:36:30,811 --> 00:36:32,880
THESE MASTER PLOTS AND WHAT BRAD

984
00:36:32,880 --> 00:36:33,848
REALIZED IF I'M DOING A

985
00:36:33,848 --> 00:36:35,616
SIMULATION I HAVE A CONNECTOME.

986
00:36:35,616 --> 00:36:37,084
I SIMULATED IT.  I KNOW WHAT

987
00:36:37,084 --> 00:36:38,986
CONNECTS TO WHAT. IF YOU COMBINE

988
00:36:38,986 --> 00:36:40,354
THAT ASSORTMENT OF FUNCTIONAL

989
00:36:40,354 --> 00:36:41,889
CAUSALITY MEASURES, THAT PEOPLE

990
00:36:41,889 --> 00:36:44,258
HAVE BEEN DOING FOR A FEW YEARS,

991
00:36:44,258 --> 00:36:45,593
WITH STRUCTURAL CONSTRAINTS, YOU

992
00:36:45,593 --> 00:36:49,163
GET SOMETHING THAT IS ANALOGOUS

993
00:36:49,163 --> 00:36:51,499
TO A SPIKE DEPENDENT PLASTICITY

994
00:36:51,499 --> 00:36:53,434
ROLE, WON'T GO INTO DETAIL BUT

995
00:36:53,434 --> 00:36:55,436
COMING OFF THAT WE DID SOMETHING

996
00:36:55,436 --> 00:36:58,839
WE CALL GRAPHICAL NEURAL

997
00:36:58,839 --> 00:37:00,307
ACTIVITY THREADS. WE LIKE

998
00:37:00,307 --> 00:37:02,309
ACRONYMS.  WE ARE A NATIONAL

999
00:37:02,309 --> 00:37:06,380
LAB. GOT TO LOVE THE ACRONYMS SO

1000
00:37:06,380 --> 00:37:09,750
OUR GNATS DECOMPOSE THESE INTO

1001
00:37:09,750 --> 00:37:12,086
SERIES OF OVERLAPPING

1002
00:37:12,086 --> 00:37:14,188
POPULATIONS OF CAUSALLY LINKED

1003
00:37:14,188 --> 00:37:16,323
NEURON PATTERNS. WHAT'S REALLY

1004
00:37:16,323 --> 00:37:18,092
GREAT, THIS IS A PAPER WE HAVE

1005
00:37:18,092 --> 00:37:20,227
ON ARCHIVE WRITING UP FOR MORE

1006
00:37:20,227 --> 00:37:25,299
FORMAL WAY RIGHT NOW, IS THAT

1007
00:37:25,299 --> 00:37:27,134
YOU SEE THESE SAME SORT OF

1008
00:37:27,134 --> 00:37:29,136
CORRELATED PATTERNS CAN EXIST

1009
00:37:29,136 --> 00:37:29,970
SEVERAL TIMES OF THE SYSTEM, IT

1010
00:37:29,970 --> 00:37:32,740
IS NOT JUST A ONE OFF THING,

1011
00:37:32,740 --> 00:37:34,475
SEEING LIKE ENSEMBLES THAT ARE

1012
00:37:34,475 --> 00:37:35,509
SPATIAL TEMPORAL THAT WILL

1013
00:37:35,509 --> 00:37:37,545
REPEAT THEMSELVES OVER TIME.

1014
00:37:37,545 --> 00:37:39,547
MOREOVER, IF I WITH GIVE CERTAIN

1015
00:37:39,547 --> 00:37:42,049
SET OF INPUTS I SEE CERTAIN

1016
00:37:42,049 --> 00:37:44,285
PATTERNS RELIABLY REACTIVATE

1017
00:37:44,285 --> 00:37:46,053
THEMSELVES MAYBE NOT EVERY TIME,

1018
00:37:46,053 --> 00:37:47,288
THERE'S STOW CHASTITY IN THE

1019
00:37:47,288 --> 00:37:48,422
MODEL AND RANDOM WHEN THEY POP

1020
00:37:48,422 --> 00:37:53,527
UP BUT THIS IS A FAIRLY SCALABLE

1021
00:37:53,527 --> 00:37:59,100
APPROACH TO ABSTRACT AWAY NEURON

1022
00:37:59,100 --> 00:38:02,436
ONE FIRE AT TIME 1, NEURON 2

1023
00:38:02,436 --> 00:38:04,405
FIRED AT TIME 5, THIS IS ZOOMING

1024
00:38:04,405 --> 00:38:05,606
OUT MAYBE MAKE SOMETHING A

1025
00:38:05,606 --> 00:38:06,774
LITTLE MORE USEFUL FROM A

1026
00:38:06,774 --> 00:38:07,875
COMPUTING POINT OF VIEW,

1027
00:38:07,875 --> 00:38:09,076
POTENTIALLY FROM A MONITORING

1028
00:38:09,076 --> 00:38:12,947
POINT OF VIEW. SO THIS IS EARLY

1029
00:38:12,947 --> 00:38:15,816
DAYS WE JUST GOT CRC PROJECT

1030
00:38:15,816 --> 00:38:18,185
THROUGH DOE THAT'S WORKING WITH

1031
00:38:18,185 --> 00:38:24,125
TIM GETNER AT UCSD AND JOE

1032
00:38:24,125 --> 00:38:26,460
ROSEFELD AT USF TO LOOK AT MORE

1033
00:38:26,460 --> 00:38:28,028
DETAIL WITH REAL DATA BUT WE ARE

1034
00:38:28,028 --> 00:38:29,897
PRETTY EXCITED.  GRAB SOME

1035
00:38:29,897 --> 00:38:37,171
WATER. THE FINAL STEP I WOULD

1036
00:38:37,171 --> 00:38:40,508
LIKE TO TALK ABOUT BRIEFLY, IS

1037
00:38:40,508 --> 00:38:42,243
THIS IDEA THAT WE MIGHT BE ABLE

1038
00:38:42,243 --> 00:38:44,011
TO GET A DEEPER UNDERSTANDING

1039
00:38:44,011 --> 00:38:45,646
ABOUT HOW THE BRAIN'S

1040
00:38:45,646 --> 00:38:47,214
FUNCTIONING FROM LOOKING AT WHAT

1041
00:38:47,214 --> 00:38:48,449
THE NEUROMORPHIC SYSTEMS ARE

1042
00:38:48,449 --> 00:38:53,954
GOOD AT.  SO HYPOTHESIS HERE, I

1043
00:38:53,954 --> 00:38:56,690
WILL SAY I THINK THIS IS

1044
00:38:56,690 --> 00:38:59,293
SOMEWHAT A GRATIFIED HYPOTHESIS 

1045
00:38:59,293 --> 00:39:00,761
IT IS CERTAINLY IN THE

1046
00:39:00,761 --> 00:39:01,862
NEUROMORPHIC FIELD. 

1047
00:39:01,862 --> 00:39:02,863
NEUROMORPHIC COMPUTING

1048
00:39:02,863 --> 00:39:05,199
POTENTIALLY HELP US UNDERSTAND

1049
00:39:05,199 --> 00:39:06,967
THINGS THE BRAIN IS ACTUALLY

1050
00:39:06,967 --> 00:39:09,637
GOOD AT BY VIRTUE OF HAVING BEEN

1051
00:39:09,637 --> 00:39:11,238
CONSTRUCTED AT THE SAME

1052
00:39:11,238 --> 00:39:12,439
ARCHITECTURES AND SAME DESIGN

1053
00:39:12,439 --> 00:39:13,574
PRINCIPLES THAT WE BELIEVE THE

1054
00:39:13,574 --> 00:39:18,279
BRAIN HAS.  CAN IT IN A SENSE

1055
00:39:18,279 --> 00:39:20,514
NEUROMORPHIC COMPUTING HELP US

1056
00:39:20,514 --> 00:39:21,749
GET SOMETHING BETTER THAN

1057
00:39:21,749 --> 00:39:23,017
PATTERN SEPARATION? SOMETHING A

1058
00:39:23,017 --> 00:39:25,419
LITTLE MORE COMPUTATIONALLY RICH

1059
00:39:25,419 --> 00:39:28,189
AND USEFUL. I WANT TO CONTRAST

1060
00:39:28,189 --> 00:39:34,128
THIS BY SAYING NEUROMORPHIC

1061
00:39:34,128 --> 00:39:35,796
EXCUSING IS NECESSARY USE AI TO

1062
00:39:35,796 --> 00:39:38,032
DO THIS. PART OF THE REASON IS

1063
00:39:38,032 --> 00:39:39,400
THAT ARTIFICIAL INTELLIGENCE

1064
00:39:39,400 --> 00:39:42,369
WHILE POWERFUL AS IT IS, IS NOT

1065
00:39:42,369 --> 00:39:44,738
CONSTRAINED BY PHYSICAL SPACE OR

1066
00:39:44,738 --> 00:39:46,307
BY ENERGY CONSTRAINTS, WE KNOW

1067
00:39:46,307 --> 00:39:47,741
IN BECAUSE WE CAN LOOK AT THE

1068
00:39:47,741 --> 00:39:50,311
ENERGY COST OF TODAY'S AI

1069
00:39:50,311 --> 00:39:52,079
ALGORITHMS IN PRODUCTION. IT IS

1070
00:39:52,079 --> 00:39:54,215
ALWAYS A POSSIBILITY IF I NEED

1071
00:39:54,215 --> 00:39:55,783
NEURAL NETWORK TO DO SOMETHING I

1072
00:39:55,783 --> 00:39:57,685
THROW MORE DATA AND MORE LAYERS

1073
00:39:57,685 --> 00:40:00,321
AND GET IT TO WORK. THAT DOESN'T

1074
00:40:00,321 --> 00:40:02,756
TELL US ANYTHING ABOUT HOW BRAIN

1075
00:40:02,756 --> 00:40:04,858
WORKS BECAUSE I THE BRAIN CAN'T

1076
00:40:04,858 --> 00:40:06,860
SAY I NEED TO LEARN GERMAN,

1077
00:40:06,860 --> 00:40:08,395
GOING THERE NEXT YEAR, ADD A FEW

1078
00:40:08,395 --> 00:40:09,863
MORE LAYERS TO CORTEX. THAT IS

1079
00:40:09,863 --> 00:40:13,033
NOT AN OPTION FOR US. OUR BRAINS

1080
00:40:13,033 --> 00:40:16,570
EXISTED WITH CONSTRAINTS. SO I

1081
00:40:16,570 --> 00:40:19,039
DON'T THINK THE FACT THAT NEURAL

1082
00:40:19,039 --> 00:40:22,076
NETWORKS ARE SUCCESSFUL AT DOING

1083
00:40:22,076 --> 00:40:23,877
COGNITIVE THINGS IS AS USEFUL AS

1084
00:40:23,877 --> 00:40:26,046
WE MIGHT WISH THEY WERE. I THINK

1085
00:40:26,046 --> 00:40:27,548
WE HAVE TO INCLUDE THE

1086
00:40:27,548 --> 00:40:29,350
CONSTRAINT OF EFFICIENCY. THE

1087
00:40:29,350 --> 00:40:33,754
REASON IS, IS BECAUSE IT IS NOT

1088
00:40:33,754 --> 00:40:36,323
AMOUNT OF WORK THAT AN ALGORITHM

1089
00:40:36,323 --> 00:40:37,891
REQUIRES BUT THE AM OF WORK

1090
00:40:37,891 --> 00:40:39,793
UNDER CONSTRAINT, OUR BRAIN HAVE

1091
00:40:39,793 --> 00:40:42,129
TO DEAL WITH THAT, THESE

1092
00:40:42,129 --> 00:40:43,130
NEUROMORPHIC CHIPS HAVE TO DEAL

1093
00:40:43,130 --> 00:40:44,999
WITH THAT. I HAVE A CONCRETE LOW

1094
00:40:44,999 --> 00:40:46,533
POWER DESIGNED LIKE THE BRAIN

1095
00:40:46,533 --> 00:40:47,801
BUT I CAN ONLY FIT SO MUCH ON

1096
00:40:47,801 --> 00:40:50,771
THEM. SO WHAT CAN I DO WITH

1097
00:40:50,771 --> 00:40:53,674
THAT? IF YOU BELIEVE THAT THE

1098
00:40:53,674 --> 00:40:55,409
NEUROMORPHIC CHIPS THAT WE ARE

1099
00:40:55,409 --> 00:40:57,311
DEVELOPING HAVE AT LEAST

1100
00:40:57,311 --> 00:40:58,545
OVERLAPPING SET OF CONSTRAINTS

1101
00:40:58,545 --> 00:41:00,080
THE BRAIN HAS, WE BELIEVE WE

1102
00:41:00,080 --> 00:41:02,149
MIGHT BE ABLE TO ACTUALLY MAKE

1103
00:41:02,149 --> 00:41:04,051
INSIGHTS ABOUT WHAT WE CAN DO.

1104
00:41:04,051 --> 00:41:06,053
VERY QUICKLY, THIS IS WHAT A

1105
00:41:06,053 --> 00:41:08,922
NEUROMORPHIC SYSTEM LOOKS LIKE. 

1106
00:41:08,922 --> 00:41:12,559
IF YOU ARE FAMILIAR WITH FIRE

1107
00:41:12,559 --> 00:41:14,695
NEURON IT IS A SIMPLE MODEL OF

1108
00:41:14,695 --> 00:41:18,732
SPIKE NEURON, MOST NEUROMORPHIC

1109
00:41:18,732 --> 00:41:20,668
CHIPS HAVE HUNDREDS OF THESE

1110
00:41:20,668 --> 00:41:22,303
FIRE NEURONS ON INDIVIDUAL

1111
00:41:22,303 --> 00:41:24,171
COMPUTE CORE AND A CHIP HAS FEW

1112
00:41:24,171 --> 00:41:25,339
HUNDRED CORES AND THEN I CAN

1113
00:41:25,339 --> 00:41:27,908
WIRE UP A BUNCH OF CHIPS

1114
00:41:27,908 --> 00:41:31,378
TOGETHTOGETHER. A NICE HIERARCHT

1115
00:41:31,378 --> 00:41:34,782
I'M PROGRAMMING NEURONS TALKING

1116
00:41:34,782 --> 00:41:35,949
TO OTHER NEURONS SYNAPSES.

1117
00:41:35,949 --> 00:41:37,584
SAME AS OUR BRAIN. THE TWO

1118
00:41:37,584 --> 00:41:38,952
FUNCTIONS IN THE REMAINING TEN

1119
00:41:38,952 --> 00:41:40,087
MINUTES I WANT TO SHOW, THAT WE

1120
00:41:40,087 --> 00:41:42,056
HAVE DISCOVERED NEUROMORPHIC

1121
00:41:42,056 --> 00:41:44,992
EXUDING IS HIGHLY EFFICIENT AT. 

1122
00:41:44,992 --> 00:41:47,127
THINGS THAT WE DID NOT EXPECT TO

1123
00:41:47,127 --> 00:41:49,596
BE SUCCESSFUL AT. MY DAY JOB AT

1124
00:41:49,596 --> 00:41:51,065
SANDIA IS TO LOOK AT

1125
00:41:51,065 --> 00:41:52,599
NEUROMORPHIC HARDWARE AND SAY

1126
00:41:52,599 --> 00:41:53,934
WHAT CAN THIS DO FOR COMPUTING

1127
00:41:53,934 --> 00:41:57,171
IN GENERAL. SANDIA AS A WHOLE IS

1128
00:41:57,171 --> 00:42:00,641
WORRIED -- DOE IS VERY WORRIED

1129
00:42:00,641 --> 00:42:01,809
ABOUT ENERGY COST ASSOCIATED

1130
00:42:01,809 --> 00:42:03,110
WITH COMPUTING TODAY.  SO WHAT

1131
00:42:03,110 --> 00:42:05,612
CAN NEUROMORPHIC COMPUTING

1132
00:42:05,612 --> 00:42:06,213
COEFFICIENTLY? WE HAVE

1133
00:42:06,213 --> 00:42:08,115
DISCOVERED TWO THINGS. ONE IS

1134
00:42:08,115 --> 00:42:09,350
THAT PROBLEMS THAT REQUIRE

1135
00:42:09,350 --> 00:42:11,485
WIDESPREAD SAMPLING YOU MIGHT --

1136
00:42:11,485 --> 00:42:14,455
OUR BRAINS DO THE SYNAPTIC

1137
00:42:14,455 --> 00:42:15,689
SAMPLING AND PROBLEMS THAT ARE

1138
00:42:15,689 --> 00:42:17,691
SPATIALLY DISTRIBUTED SO

1139
00:42:17,691 --> 00:42:19,059
PHYSICAL SOLUTIONS TO PROBLEMS

1140
00:42:19,059 --> 00:42:20,728
THAT HAVE SPATIAL CONSTRAINTS.

1141
00:42:20,728 --> 00:42:23,464
THE FIRST THING WAS DONE

1142
00:42:23,464 --> 00:42:25,165
MYOPRIMARILY WILLIAM WHO DID THE

1143
00:42:25,165 --> 00:42:27,334
FLY WIRE MAPPING AND DARBY SMITH

1144
00:42:27,334 --> 00:42:30,537
WHO WAS POST DOC NOW STAFF, BOTH

1145
00:42:30,537 --> 00:42:33,073
USED TO BE POST DOCS NOW STAFF

1146
00:42:33,073 --> 00:42:36,276
AT SANDIA MATHEMATICIANS, THEY

1147
00:42:36,276 --> 00:42:37,845
ARE FANTASTIC TO WORK WITH, I

1148
00:42:37,845 --> 00:42:40,848
HIGHLY RECOMMEND IT.  I MENTION

1149
00:42:40,848 --> 00:42:42,383
EARLIER THE HURRICANE

1150
00:42:42,383 --> 00:42:43,183
PREDICTION, THIS IS A DIFFERENT

1151
00:42:43,183 --> 00:42:46,787
HURRICANE. BUT OF COURSE THERE

1152
00:42:46,787 --> 00:42:47,855
ARE A LOT OF PROBLEMS IN THE

1153
00:42:47,855 --> 00:42:50,958
WORLD WE CARE ABOUT THAT RELY ON

1154
00:42:50,958 --> 00:42:52,092
UNCERTAINTY TO UNDERSTAND. OUR

1155
00:42:52,092 --> 00:42:53,227
COMPUTERS ARE NOT GOOD DEALING

1156
00:42:53,227 --> 00:42:55,963
WITH UNCERTAINTY, FLIPING A

1157
00:42:55,963 --> 00:42:58,432
TRANSISTER IS A BINARY DISCRETE

1158
00:42:58,432 --> 00:43:00,868
DETERMINISTIC EVENT SO WHAT CAN

1159
00:43:00,868 --> 00:43:03,904
WE MODEL THINGS PROBABILISTIC

1160
00:43:03,904 --> 00:43:06,140
PROBLEMS ON NEUROMORPHIC

1161
00:43:06,140 --> 00:43:07,174
HARDWARE. THIS IS SOMETHING

1162
00:43:07,174 --> 00:43:10,077
PEOPLE WOULD NEVER HAVE GUESSED

1163
00:43:10,077 --> 00:43:11,979
IS POSSIBLE. BUT ABOUT FIVE

1164
00:43:11,979 --> 00:43:14,214
YEARS AGO WE REALIZED THE PAPER

1165
00:43:14,214 --> 00:43:16,717
CAME OUT IN 2022, WE SAW ACTUAL

1166
00:43:16,717 --> 00:43:17,951
NEUROMORPHIC SCALING ADVANTAGE

1167
00:43:17,951 --> 00:43:19,286
FOR MODELING RANDOM WALKS.  IF

1168
00:43:19,286 --> 00:43:20,521
YOU DON'T KNOW WHAT A RANDOM

1169
00:43:20,521 --> 00:43:21,989
WALK IS, IMAGINE STOCK PRICE

1170
00:43:21,989 --> 00:43:23,657
JITTERING UP AND DOWN, YOU MAY

1171
00:43:23,657 --> 00:43:25,092
GO UP OVER TIME, YOU KNOW IT

1172
00:43:25,092 --> 00:43:26,527
WILL GO UP EVERY TIME BUT THE

1173
00:43:26,527 --> 00:43:31,398
PATH IT TAKES TO GETS THERE IS

1174
00:43:31,398 --> 00:43:35,402
STOCHASTIC, NOT PRE-DETERMINED. 

1175
00:43:35,402 --> 00:43:37,371
SO WONDERING HOW THE MODEL WORKS

1176
00:43:37,371 --> 00:43:38,939
OTHER THAN TO SAY WE WERE ABLE

1177
00:43:38,939 --> 00:43:40,073
TO WIRE UP POPULATIONS OF

1178
00:43:40,073 --> 00:43:43,277
NEURONS IN A WAY THAT REPRESENT

1179
00:43:43,277 --> 00:43:44,545
THE STATE SPACE OVER WHICH

1180
00:43:44,545 --> 00:43:46,513
RANDOM WALK HAS O OCCUR. A MONTE

1181
00:43:46,513 --> 00:43:48,315
CARLO PROCESS EVOLVING OVER TIME

1182
00:43:48,315 --> 00:43:50,417
BASICALLY BE AKIN TO A SPIKE

1183
00:43:50,417 --> 00:43:51,852
MOVING OVER THIS MESH OF

1184
00:43:51,852 --> 00:43:53,587
NEURONS. EVERY RANDOM WALKER IN

1185
00:43:53,587 --> 00:43:55,289
MY MODEL IS A DIFFERENT SPIKE

1186
00:43:55,289 --> 00:43:58,826
MOVING OVER MY CHIP. THAT IS A

1187
00:43:58,826 --> 00:44:00,294
RADICALLY DIFFERENT WAY THINKING

1188
00:44:00,294 --> 00:44:02,062
ABOUT MONTE CARLO SIMULATIONS AN

1189
00:44:02,062 --> 00:44:04,097
COMPUTING AND IT WORKS REALLY

1190
00:44:04,097 --> 00:44:06,099
WELL UNDER AMORPHIC HARDWARE.

1191
00:44:06,099 --> 00:44:09,903
WHAT ESSENTIALLY HAPPENS, MY

1192
00:44:09,903 --> 00:44:12,005
11-YEAR-OLD DID THIS IN A BLEND,

1193
00:44:12,005 --> 00:44:15,008
AMAZES ME, I CAN'T DO THIS IN A

1194
00:44:15,008 --> 00:44:15,943
BLENDER B. EVERY NEURON IN THE

1195
00:44:15,943 --> 00:44:19,580
SYSTEM IS BEHAVING INDEPENDENTLY

1196
00:44:19,580 --> 00:44:21,582
OF EVERY OTHER NEURON, MESH

1197
00:44:21,582 --> 00:44:24,451
POINT BEHAVING INDEPENDENTLY OF

1198
00:44:24,451 --> 00:44:26,687
EVERY OTHER MESH POINT SO THE

1199
00:44:26,687 --> 00:44:28,889
RANDOM WALKS GO IN FULL PERFECT

1200
00:44:28,889 --> 00:44:30,557
PARALLELISM, AND THIS TURNS OUT

1201
00:44:30,557 --> 00:44:32,493
A HUGE ENERGY SAVINGS.  WE SAW

1202
00:44:32,493 --> 00:44:35,596
THIS IN OUR BENCHMARKING AT THE

1203
00:44:35,596 --> 00:44:38,365
TIME, YOU SEE THE NUMBER OF

1204
00:44:38,365 --> 00:44:39,600
WALKER PER JEWELS MEASURING

1205
00:44:39,600 --> 00:44:40,734
ENERGY NOT POWER SUBPOENA ORDERS

1206
00:44:40,734 --> 00:44:42,269
OF MAGNITUDE BETTER THAN YOU SEE

1207
00:44:42,269 --> 00:44:45,606
ON CPU AND GPU UNDER AMORPHIC

1208
00:44:45,606 --> 00:44:47,274
HARDWARE SO WENT A LITTLE SLOWER

1209
00:44:47,274 --> 00:44:48,809
ON THE ORIGINAL NEUROMORPHIC

1210
00:44:48,809 --> 00:44:50,611
SYSTEMS THE NEW SYSTEMS ARE ABLE

1211
00:44:50,611 --> 00:44:51,879
TO GO FASTER BUT THE POWER

1212
00:44:51,879 --> 00:44:53,547
ADVANTAGE IS STILL THERE. SO

1213
00:44:53,547 --> 00:44:56,283
THIS IS SAYING THINGS LIKE

1214
00:44:56,283 --> 00:44:59,620
PROBABILISTIC SAMPLING. MONTE

1215
00:44:59,620 --> 00:45:01,321
CARLO BASE YEN INTEREST, REAL

1216
00:45:01,321 --> 00:45:03,156
PROBLEMS THE WORLD NEEDS TO DEAL

1217
00:45:03,156 --> 00:45:05,425
WITH ARE VERY NICE NEUROMORPHIC

1218
00:45:05,425 --> 00:45:07,094
HARD WARE. OUR BRAIN SIMILARLY

1219
00:45:07,094 --> 00:45:10,898
HAVE TO DEAL WITH UNSORETY. SO

1220
00:45:10,898 --> 00:45:12,900
-- UNCERTAINTY. THE SECOND AREA

1221
00:45:12,900 --> 00:45:15,102
IS ALSO WORK WITH MY CURRENT

1222
00:45:15,102 --> 00:45:18,038
POST DOC BRADS THEILMAN WHO DID

1223
00:45:18,038 --> 00:45:20,674
THE GNATS STUFF, WE LOOK AT

1224
00:45:20,674 --> 00:45:22,276
SPATIALLY DISTRIBUTED PHYSICS

1225
00:45:22,276 --> 00:45:25,679
PROBLEM. SO IF YOU DO

1226
00:45:25,679 --> 00:45:27,514
COMPUTATIONAL PHYSICS, MANY

1227
00:45:27,514 --> 00:45:30,317
PROBLEMS CAN BE HOE DESCRIBED AS

1228
00:45:30,317 --> 00:45:31,451
PARTIAL EQUATION THAT LOOKS LIKE

1229
00:45:31,451 --> 00:45:33,787
THIS. THE WAY YOU DO THIS

1230
00:45:33,787 --> 00:45:37,357
CLASSICALLY IS USING ELEMENT

1231
00:45:37,357 --> 00:45:39,826
METHODS. IF YOU ARE NOT

1232
00:45:39,826 --> 00:45:41,161
FAMILIAR, IT IS BASICALLY TRYING

1233
00:45:41,161 --> 00:45:42,195
TO UNDERSTAND SOME PHYSICAL

1234
00:45:42,195 --> 00:45:43,664
DYNAMIC LIKE HEAT OR WHATEVER

1235
00:45:43,664 --> 00:45:47,768
OVER SOME SURFACE, EVERY POINT

1236
00:45:47,768 --> 00:45:49,870
ON THAT MESH, MY AREA MODELING

1237
00:45:49,870 --> 00:45:51,772
OVER MESH OR SOMETHING, AT EVERY

1238
00:45:51,772 --> 00:45:53,674
POINT I BASICALLY COMPUTE THE

1239
00:45:53,674 --> 00:45:55,242
PHYSICS AT THAT POINT BASED ON

1240
00:45:55,242 --> 00:45:58,345
THE NEIGHBORING POINTS. THIS IS

1241
00:45:58,345 --> 00:46:00,347
A GOLD STANDARD HOW YOU DO

1242
00:46:00,347 --> 00:46:03,383
THINGS LIKE TURBULENCE MODELING

1243
00:46:03,383 --> 00:46:05,185
OR HEAT ACTIVE DEUCE MODELING SO

1244
00:46:05,185 --> 00:46:06,587
FORTH -- DIFFUSION MODELING.

1245
00:46:06,587 --> 00:46:08,355
THAT'S A LOT OF WORK IN

1246
00:46:08,355 --> 00:46:10,223
SCIENTIFIC LEARNING TO

1247
00:46:10,223 --> 00:46:11,592
APPROXIMATE THIS BUT CAN YOU MAP

1248
00:46:11,592 --> 00:46:14,628
THIS ON TO NEUROMORPHIC CHIPS.

1249
00:46:14,628 --> 00:46:16,196
AGAIN, NO ONE ANYWHERE IN THE

1250
00:46:16,196 --> 00:46:17,698
NEUROMORPHIC FIELD THINKS THIS

1251
00:46:17,698 --> 00:46:20,467
IS POSSIBLE. BUT WE WANTED TO --

1252
00:46:20,467 --> 00:46:22,636
BASED ON THAT PROBABILISTIC

1253
00:46:22,636 --> 00:46:23,737
RESULT BEFORE CAN YOU USE SOME

1254
00:46:23,737 --> 00:46:25,405
SORT OF SAMPLING APPROACH TO DO

1255
00:46:25,405 --> 00:46:28,842
THIS? AND WHAT BRAD HAD A

1256
00:46:28,842 --> 00:46:30,077
REALIZATION FROM HIS

1257
00:46:30,077 --> 00:46:31,411
NEUROSCIENCE BACKGROUND REALIZED

1258
00:46:31,411 --> 00:46:33,981
IS THAT YOU CAN ACTUALLY -- IF

1259
00:46:33,981 --> 00:46:36,416
YOU HAVE THESE CORTICAL MODELS

1260
00:46:36,416 --> 00:46:38,852
WHERE YOU HAVE CLUSTERS OF

1261
00:46:38,852 --> 00:46:40,654
NEURONS EXCITATORY AND INHIBIT

1262
00:46:40,654 --> 00:46:42,322
POOTORY LOCALIZE OPTSMIZING

1263
00:46:42,322 --> 00:46:44,024
THEMSELVES TO REPRESENTS

1264
00:46:44,024 --> 00:46:45,058
SOMETHING YOU CAN WIRE THOSE UP

1265
00:46:45,058 --> 00:46:48,662
IN A MESH NOT UNLIKE WHAT A FINE

1266
00:46:48,662 --> 00:46:51,431
ELEMENT MESH SHOW AND

1267
00:46:51,431 --> 00:46:52,232
MATHEMATICALLY YOU CAN MAKE

1268
00:46:52,232 --> 00:46:53,433
THOSE RELATIONSHIPS EXPLICIT

1269
00:46:53,433 --> 00:46:55,202
SUCH THAT I CAN TAKE THE PHYSICS

1270
00:46:55,202 --> 00:46:58,305
MODEL THAT COMPUTATIONAL

1271
00:46:58,305 --> 00:47:01,608
PHYSICIST CARES ABOUT, I CAN

1272
00:47:01,608 --> 00:47:04,745
TAKE THAT A MATRIX AND I CAN

1273
00:47:04,745 --> 00:47:06,513
REPRESENT IT AS THE CONNECTIONS

1274
00:47:06,513 --> 00:47:09,583
BETWEEN THESE MESH POINTS YOU

1275
00:47:09,583 --> 00:47:12,319
CAN SEE HERE. THE DETAILS YOU

1276
00:47:12,319 --> 00:47:14,855
CAN GET -- IT CAN GET

1277
00:47:14,855 --> 00:47:16,223
COMPLICATED BUT IT IS NOT

1278
00:47:16,223 --> 00:47:18,258
REQUIRING ANY DYNAMICS OR MATH

1279
00:47:18,258 --> 00:47:19,626
THAT IS NOT SOMETHING

1280
00:47:19,626 --> 00:47:20,427
NEUROSCIENTISTS HAVE BEEN

1281
00:47:20,427 --> 00:47:22,796
LOOKING AT FOR YEARS, IN FACT,

1282
00:47:22,796 --> 00:47:25,932
IT IS RATHER NICELY PARALLEL

1283
00:47:25,932 --> 00:47:28,535
WITH THINGS IN LIKE MORE CONTROL

1284
00:47:28,535 --> 00:47:29,970
PERSPECTIVES HOW CORTICAL

1285
00:47:29,970 --> 00:47:31,872
CIRCUITS MAY WORK. IT SCALES

1286
00:47:31,872 --> 00:47:34,841
VERY NICELY ON THINGS LIKE LOIHI

1287
00:47:34,841 --> 00:47:38,679
2 WHICH WE HAVE BEEN DOING. IF I

1288
00:47:38,679 --> 00:47:40,013
HID THE LABELS IN THE BOTTOM

1289
00:47:40,013 --> 00:47:41,715
THERE YOU WOULDN'T BE ABLE TO

1290
00:47:41,715 --> 00:47:42,983
TELL WHICH WAS NEUROMORPHIC

1291
00:47:42,983 --> 00:47:44,651
SOLUTION AND WHICH THE GROUNDS

1292
00:47:44,651 --> 00:47:46,186
TRUTH SOLUTION. THIS IS

1293
00:47:46,186 --> 00:47:47,320
SOMETHING THAT AGAIN EVERYONE

1294
00:47:47,320 --> 00:47:50,023
THINKS AN AI SOLUTION OR

1295
00:47:50,023 --> 00:47:52,025
AMORPHIC SOLUTION HAS TO BE SOME

1296
00:47:52,025 --> 00:47:53,193
COURSE APPROXIMATION OF WHAT WE

1297
00:47:53,193 --> 00:47:55,696
CAN DO WITH FANCY CALCULATORS

1298
00:47:55,696 --> 00:47:57,831
AND DETERMINISTIC COMPUTING.  IT

1299
00:47:57,831 --> 00:47:59,399
DOESN'T HAVE TO BE. TURNS OUT

1300
00:47:59,399 --> 00:48:01,168
THESE NEUROMORPHIC CHIPS USING

1301
00:48:01,168 --> 00:48:04,171
BRAIN INSPIRED ALGORITHMS, USING

1302
00:48:04,171 --> 00:48:06,106
NEURONS AS THE UNDERLYING

1303
00:48:06,106 --> 00:48:07,674
COMPUTE UNITS CAN SOLVE THESE

1304
00:48:07,674 --> 00:48:09,209
PROBLEMS AT LEVELS OF ACCURACY

1305
00:48:09,209 --> 00:48:12,212
AND PRECISION THAT THE

1306
00:48:12,212 --> 00:48:13,380
COMPUTATIONAL PHYSICIST MAY CARE

1307
00:48:13,380 --> 00:48:15,348
ABOUT. IF YOU ARE LIKE CORTICAL

1308
00:48:15,348 --> 00:48:16,550
SYSTEMS NEUROSCIENTIST AND LOOK

1309
00:48:16,550 --> 00:48:18,118
AT THE MASTER PLOT HERE COMING

1310
00:48:18,118 --> 00:48:21,354
OFF THE MODEL, THIS CAN BE DATA

1311
00:48:21,354 --> 00:48:25,659
COMING OFF OF SOME SORT OF

1312
00:48:25,659 --> 00:48:26,560
MULTI-ELECTRODE ARRAY OR

1313
00:48:26,560 --> 00:48:28,361
SOMETHING LIKE THAT. SO THIS WAS

1314
00:48:28,361 --> 00:48:31,398
PRETTY EXCITING BECAUSE IT TELLS

1315
00:48:31,398 --> 00:48:33,934
US THESE ARE PHYSICS PROBLEMS,

1316
00:48:33,934 --> 00:48:37,070
THESE AREN'T BIOLOGY PROBLEMS.

1317
00:48:37,070 --> 00:48:38,839
BUT OUR NEUROMORPHIC HARDWARE

1318
00:48:38,839 --> 00:48:41,675
SEEMS REALLY GOOD AT DOING THIS.

1319
00:48:41,675 --> 00:48:43,643
AND IF YOU STOP AND THINK FOR A

1320
00:48:43,643 --> 00:48:44,845
SECOND, WHAT ARE THESE PROBLEMS

1321
00:48:44,845 --> 00:48:46,146
REALLY DOING? TRYING TO

1322
00:48:46,146 --> 00:48:47,581
UNDERSTAND WHAT THE PHYSICAL

1323
00:48:47,581 --> 00:48:50,717
WORLD IS, BASED ON SOME PERHAPS

1324
00:48:50,717 --> 00:48:52,486
OBSERVATIONS BUT DEAF ANY

1325
00:48:52,486 --> 00:48:54,187
FACTUALLY PRIOR MODELS, AND WANT

1326
00:48:54,187 --> 00:48:55,322
TO MAKE PREDICTIONS OF WHAT IS

1327
00:48:55,322 --> 00:48:57,124
GOING TO HAPPEN? WE HAVE TO DO

1328
00:48:57,124 --> 00:48:58,258
THAT OUR BRAIN VERSUS TO DO THAT

1329
00:48:58,258 --> 00:48:59,559
AND THAT IS A REAL PROBLEM OUR

1330
00:48:59,559 --> 00:49:01,361
BRAINS HAVE TO SOLVE AND HAVE

1331
00:49:01,361 --> 00:49:02,729
TROUBLE SOLVING WHEN THINGS

1332
00:49:02,729 --> 00:49:04,831
START TO GO WRONG. LIKEWISE, THE

1333
00:49:04,831 --> 00:49:07,567
ABILITY TO MAKE DECISIONS, OFTEN

1334
00:49:07,567 --> 00:49:09,069
REQUIRES SAMPLING BUNCH OF

1335
00:49:09,069 --> 00:49:10,737
POTENTIAL OUTCOMES WHICH IS

1336
00:49:10,737 --> 00:49:12,639
SOMETHING THAT MEANS OUR BRAINS

1337
00:49:12,639 --> 00:49:16,209
REALLY NEED TO DO SOME SORT OF

1338
00:49:16,209 --> 00:49:19,980
SAMPLING PROCESS, PROBABILISTIC

1339
00:49:19,980 --> 00:49:21,214
REASONING.  BASE YEN REASONING

1340
00:49:21,214 --> 00:49:22,482
IS SOMETHING WE LONG THOUGHT THE

1341
00:49:22,482 --> 00:49:27,687
BRAIN WAS DOING BUT -- P BECAUSE

1342
00:49:27,687 --> 00:49:34,227
YEN. . BECAUSE YEN. BASE YEN. SO

1343
00:49:34,227 --> 00:49:35,395
WHAT I WOULD SAY, MAYBE

1344
00:49:35,395 --> 00:49:36,696
CHALLENGE THE NEUROSCIENCE

1345
00:49:36,696 --> 00:49:38,365
COMMUNITY HERE, IS THAT INSTEAD

1346
00:49:38,365 --> 00:49:40,634
OF THINKING OH I THINK THE BRAIN

1347
00:49:40,634 --> 00:49:42,335
DOES THIS BECAUSE I STUDY

1348
00:49:42,335 --> 00:49:43,136
COGNITION, THIS SEEMS IMPORTANT

1349
00:49:43,136 --> 00:49:44,638
WHAT IS FORMULA FOR THIS, HOW

1350
00:49:44,638 --> 00:49:46,873
DOES THAT MAP TO THE BRAIN, MAKE

1351
00:49:46,873 --> 00:49:48,542
WE CAN GO THE OTHER WAY -- MAYBE

1352
00:49:48,542 --> 00:49:51,745
WE CAN GO THE OTHER WAY AROUND.

1353
00:49:51,745 --> 00:49:53,213
CIRCUITS ARE GOOD AT DOING

1354
00:49:53,213 --> 00:49:55,115
CERTAIN THINGS. IF WE CAN

1355
00:49:55,115 --> 00:49:56,383
UNDERSTAND WHAT THOSE ARE LOOK

1356
00:49:56,383 --> 00:49:57,818
INTO THE BRAIN TO SEE IF IT'S

1357
00:49:57,818 --> 00:49:58,852
SOING UP SOMEWHERE AND THERE'S A

1358
00:49:58,852 --> 00:50:01,288
CASE THERE WAS A BEAUTIFUL PAPER

1359
00:50:01,288 --> 00:50:03,290
OUT OF EUROPE A FEW YEARS AGO

1360
00:50:03,290 --> 00:50:07,060
LOOKING AT RANDOM TRAJECTORIES

1361
00:50:07,060 --> 00:50:09,963
DURING HIPPOEXAMPLE REPLAY SO

1362
00:50:09,963 --> 00:50:11,798
THE REPLAY EVENTS MIGHT BE A

1363
00:50:11,798 --> 00:50:14,434
BROWNIAN SAMPLING. AND THERE'S

1364
00:50:14,434 --> 00:50:15,802
CONTROL MODEL QUITE SIMILAR O

1365
00:50:15,802 --> 00:50:18,371
THE MODELS HERE. SO TO WRAP UP,

1366
00:50:18,371 --> 00:50:20,473
WE HAVE THIS CHALLENGE WHERE WE

1367
00:50:20,473 --> 00:50:22,142
ARE LOOKING AT THE COMPLEXITY OF

1368
00:50:22,142 --> 00:50:23,577
THE BRAIN, THE TIME SCALES AND

1369
00:50:23,577 --> 00:50:25,712
THE SIZE OF THE BRAIN AND THIS

1370
00:50:25,712 --> 00:50:26,980
CHALLENGE OF COMPUTATION THAT WE

1371
00:50:26,980 --> 00:50:28,515
ARE FACED WITH.  AND WE DON'T

1372
00:50:28,515 --> 00:50:30,851
NECESSARILY KNOW WHAT TO DO. 

1373
00:50:30,851 --> 00:50:32,185
WHAT I BELIEVE IS THAT

1374
00:50:32,185 --> 00:50:33,753
NEUROMORPHIC FIELD IS NOT THE

1375
00:50:33,753 --> 00:50:35,422
ANSWER HERE. WE CHUTE SLEUTHLY

1376
00:50:35,422 --> 00:50:38,091
HAVE TO GET MORE DATA AND WE

1377
00:50:38,091 --> 00:50:39,826
HAVE TO UNDERSTAND MORE

1378
00:50:39,826 --> 00:50:41,328
CONNECTOME WE NEEDS MORE THAN

1379
00:50:41,328 --> 00:50:44,331
CONNECTOMES.  BUT I ALSO BELIEVE

1380
00:50:44,331 --> 00:50:48,034
HAVING THIS MORE BRAIN INSPIRED

1381
00:50:48,034 --> 00:50:49,169
COMPUTING CONSTRAINT GETTING

1382
00:50:49,169 --> 00:50:50,971
FROM THIS IDEA WE CAN DO

1383
00:50:50,971 --> 00:50:52,439
EVERYTHING BY JUST HAVING ENOUGH

1384
00:50:52,439 --> 00:50:54,207
LAYERS OF NEURONS ON IT BUT

1385
00:50:54,207 --> 00:50:56,009
RATHER CONSTRAINED TO THE

1386
00:50:56,009 --> 00:50:58,578
PHYSICS OF THIS EMERGING

1387
00:50:58,578 --> 00:51:02,015
HARDWARE, MAY GIVE US A PATH TO

1388
00:51:02,015 --> 00:51:09,723
DO THIS. SUPPOSE TO SAY REALISM

1389
00:51:09,723 --> 00:51:11,658
BUT I BELIEVE WE CAN THEN AS WE

1390
00:51:11,658 --> 00:51:15,528
LEARN MORE FROM BIOLOGY SIDE

1391
00:51:15,528 --> 00:51:16,596
COMPUTE TRAJECTORY AND WE HAD A

1392
00:51:16,596 --> 00:51:19,065
QUESTION AT LUNCH ABOUT HOW

1393
00:51:19,065 --> 00:51:20,433
NEUROMODULATORS END UP UNDER

1394
00:51:20,433 --> 00:51:21,468
AMORPHIC COMPUTING.  CERTAINLY

1395
00:51:21,468 --> 00:51:22,802
NO ONE IS THINKING ABOUT THAT IN

1396
00:51:22,802 --> 00:51:25,472
THE NEUROMORPHIC HARDWARE SIDE

1397
00:51:25,472 --> 00:51:26,840
AT DEEP LEVEL BECAUSE WE DON'T

1398
00:51:26,840 --> 00:51:28,541
KNOW ENOUGH ABOUT

1399
00:51:28,541 --> 00:51:29,876
NEUROMODULATORS LIKE SEROTONIN

1400
00:51:29,876 --> 00:51:31,645
TO TELL THEM WHAT THEY NEED. BUT

1401
00:51:31,645 --> 00:51:34,347
I THINK IT IS NOT UNREASONABLE

1402
00:51:34,347 --> 00:51:35,782
THE FIELD IS STILL YOUNG ENOUGH

1403
00:51:35,782 --> 00:51:40,053
WE CAN INFLUENCE IT. WITH THAT,

1404
00:51:40,053 --> 00:51:41,888
I WOULD BE HAPPY TO TAKE ANY

1405
00:51:41,888 --> 00:51:43,456
QUESTIONS, I WOULD LIKE TO FIRST

1406
00:51:43,456 --> 00:51:45,025
THANK -- THIS IS LARGE GROUP OF

1407
00:51:45,025 --> 00:51:47,227
PEOPLE I WORKED WITH OVER THE

1408
00:51:47,227 --> 00:51:48,895
YEARS, FANTASTIC COLLABORATORS,

1409
00:51:48,895 --> 00:51:51,998
EXPERIMENTAL AND COMPUTATIONAL

1410
00:51:51,998 --> 00:51:53,033
MATHEMATICIANS, COMPUTER

1411
00:51:53,033 --> 00:51:56,136
SCIENTISTS, ALL SORTS OF PEOPLE 

1412
00:51:56,136 --> 00:51:56,636
THANK YOU VERY MUCH.

1413
00:51:56,636 --> 00:52:02,042
[APPLAUSE]

1414
00:52:02,042 --> 00:52:03,710
>> CERTAINLY ENCOURAGE
ANYONE

1415
00:52:03,710 --> 00:52:05,612
HERE TO USE THE MIKES.  WE HAVE

1416
00:52:05,612 --> 00:52:07,080
ONE -- MICS.  WE HAVE A QUESTION

1417
00:52:07,080 --> 00:52:08,648
ONLINE TO GET STARTED. THE REST

1418
00:52:08,648 --> 00:52:11,318
OF YOU CAN USE YOUR BRAIN AND/OR

1419
00:52:11,318 --> 00:52:12,686
COMPUTER TO THINK OF QUESTIONS.

1420
00:52:12,686 --> 00:52:17,424
SO THIS IS A GOOD ONE. MISTAKES

1421
00:52:17,424 --> 00:52:19,859
ARE INHERENT PROPERTY OF NEURAL

1422
00:52:19,859 --> 00:52:22,329
NETWORKS BASED BIOLOGICAL AND

1423
00:52:22,329 --> 00:52:24,731
ARTIFICIAL SYSTEMS. CAN

1424
00:52:24,731 --> 00:52:27,300
NEURO-SYMBOLIC BASED AI HELP TO

1425
00:52:27,300 --> 00:52:28,969
DECREASE NUMBER OF MISTAKES MADE

1426
00:52:28,969 --> 00:52:32,973
BY AI?

1427
00:52:32,973 --> 00:52:36,076
>> I THINK IT IS POSSIBLE.
SO

1428
00:52:36,076 --> 00:52:38,178
NEURO-SYMBOLIC PROCESSES, IT IS

1429
00:52:38,178 --> 00:52:38,878
A LITTLE DIFFERENT THAN HOW I

1430
00:52:38,878 --> 00:52:41,081
WAS DESCRIBING THINGS BUT THE

1431
00:52:41,081 --> 00:52:42,849
IDEA IS THAT WE KNOW THERE ARE

1432
00:52:42,849 --> 00:52:45,418
TWO TYPES OF AI WHERE YOU HAVE

1433
00:52:45,418 --> 00:52:48,521
CLASSIC SYMBOLIC MODELS EXPERTS

1434
00:52:48,521 --> 00:52:49,422
CONSTRUCT AND NEURAL NETWORKS

1435
00:52:49,422 --> 00:52:51,224
THAT ARE POWERFUL AND COMBINE

1436
00:52:51,224 --> 00:52:53,159
THE TWO. AND THE HOPE IS IS THAT

1437
00:52:53,159 --> 00:52:55,495
THIS MIGHT PROVIDE

1438
00:52:55,495 --> 00:52:56,363
EXPLAINABILITY AND MORE ACCURATE

1439
00:52:56,363 --> 00:53:03,570
MODELING. I MIGHT TWIST THE

1440
00:53:03,570 --> 00:53:06,039
QUESTION AND SAY OUR BRAINS OUR

1441
00:53:06,039 --> 00:53:09,142
NEURAL CIRCUITS MAKE MISTAKES

1442
00:53:09,142 --> 00:53:10,810
FOR A REASON MANY TIMES, THAT'S

1443
00:53:10,810 --> 00:53:15,548
HOW WE LEARN.  OUR BRAINS ARE

1444
00:53:15,548 --> 00:53:17,417
LIFE LONG COMPUTERS, NOT

1445
00:53:17,417 --> 00:53:19,652
INSTANTANEOUS COMPUTERS. WE

1446
00:53:19,652 --> 00:53:21,421
DEVELOP OUR NEURAL NETWORKS AS

1447
00:53:21,421 --> 00:53:23,890
WE HAD A TRAINING PERIODS, THEN

1448
00:53:23,890 --> 00:53:25,592
WE VALIDATED THEM, DEPLOY THEM

1449
00:53:25,592 --> 00:53:26,559
AND IDEAL I WILL THEY ARE NOT

1450
00:53:26,559 --> 00:53:28,395
CHANGING MUCH, OUR BRAINS HAVE

1451
00:53:28,395 --> 00:53:30,430
TO CHANGE CONTINUOUSLY, ADAPT TO

1452
00:53:30,430 --> 00:53:32,232
BOTH CHANGES INTERNAL AS WELL AS

1453
00:53:32,232 --> 00:53:33,800
EXTERNAL. PART OF WAY THEY ADAPT

1454
00:53:33,800 --> 00:53:36,336
IS THROUGH THESE MISTAKES. SO I

1455
00:53:36,336 --> 00:53:37,904
WOULD SAY IT IS NOT ABOUT

1456
00:53:37,904 --> 00:53:39,472
REDUCING ERRORS, I KNOW THAT'S

1457
00:53:39,472 --> 00:53:43,143
HOW WE QUANTIFY AI PERFORMANCE

1458
00:53:43,143 --> 00:53:46,379
BUT RATHER MAKING SURE THAT WHEN

1459
00:53:46,379 --> 00:53:49,883
THE SYSTEM EXPLORES POSSIBLE

1460
00:53:49,883 --> 00:53:52,452
ALTERNATIVES SCENARIOS. SAMPLING

1461
00:53:52,452 --> 00:53:54,020
IF YOU WILL, THE WAY I DESCRIBE

1462
00:53:54,020 --> 00:53:56,823
IT, THAT ACTUALLY DOESN'T THROW

1463
00:53:56,823 --> 00:54:00,727
THINGS COMPLETELY AWRY, IT

1464
00:54:00,727 --> 00:54:01,661
DOESN'T CRASH THE CAR IF YOU

1465
00:54:01,661 --> 00:54:04,030
WILL IF YOU ARE LIKE

1466
00:54:04,030 --> 00:54:05,265
SELF-DRIVING CAR, YOU DON'T WANT

1467
00:54:05,265 --> 00:54:06,733
THAT, BUT YOU WANT TO EXPLORE --

1468
00:54:06,733 --> 00:54:08,768
YOU HAVE TO MAKE MISTAKES, THAT

1469
00:54:08,768 --> 00:54:11,838
IS WHAT COGNITION IS. SO I THINK

1470
00:54:11,838 --> 00:54:16,209
A LITTLE BIT OF EXPLORATION AND

1471
00:54:16,209 --> 00:54:17,744
MISTAKES IS GOOD FOR A SYSTEM. I

1472
00:54:17,744 --> 00:54:25,018
HOPE THAT ANSWERS THE QUESTION. 

1473
00:54:25,018 --> 00:54:25,418
>>

1474
00:54:25,418 --> 00:54:27,954
>> THANKS, REALLY NICE
TALK. SO

1475
00:54:27,954 --> 00:54:31,257
YOU SHOWED AT SOME POINT THE

1476
00:54:31,257 --> 00:54:32,392
SIMULATION YOU IMPLEMENTED FROM

1477
00:54:32,392 --> 00:54:34,027
THE FLY WIRE DATA. AND ONE THING

1478
00:54:34,027 --> 00:54:35,528
THAT WAS IMPRESSIVE IS IT WAS --

1479
00:54:35,528 --> 00:54:37,097
LOOKED LIKE REALLY LARGE NUMBER

1480
00:54:37,097 --> 00:54:38,932
OF NEURONS. THE FLY, THE FLY

1481
00:54:38,932 --> 00:54:40,800
DATA IS LARGE NUMBER. SO I'M

1482
00:54:40,800 --> 00:54:43,036
INTERESTED IN TRYING TO --

1483
00:54:43,036 --> 00:54:44,904
WONDERING WHAT ARCHITECTURE THAT

1484
00:54:44,904 --> 00:54:46,506
USES.  ARE YOU IMPLEMENTING THE

1485
00:54:46,506 --> 00:54:48,274
SORT OF ANALOG INTEGRATION

1486
00:54:48,274 --> 00:54:51,478
DIRECTLY ON CHIP? AND THEN I

1487
00:54:51,478 --> 00:54:52,378
GUESS MORE BROADLY I'M

1488
00:54:52,378 --> 00:54:56,549
INTERESTED IN WHERE YOU SEE THE

1489
00:54:56,549 --> 00:54:57,684
LARGE-SCALE OF NEUROMORPHIC

1490
00:54:57,684 --> 00:54:59,786
COMPUTING GOING. IS ONE OF THE

1491
00:54:59,786 --> 00:55:01,654
-- I DON'T KNOW PEOPLE SAY THAT

1492
00:55:01,654 --> 00:55:06,192
OUR BRAIN WORKS BETTER THAN CHAT

1493
00:55:06,192 --> 00:55:09,696
GTP ON LIKE HAMBURGER EVERY FEW

1494
00:55:09,696 --> 00:55:12,132
HOURS AND LESS POWER THAN AI USE

1495
00:55:12,132 --> 00:55:13,099
BUS ONE OF THE THINGS WE ARE

1496
00:55:13,099 --> 00:55:15,068
THINKING IN THE FUTURE, IS BEING

1497
00:55:15,068 --> 00:55:17,137
ABLE TO DO REALLY LARGE SCALE

1498
00:55:17,137 --> 00:55:19,906
COMPUTATION IN THIS ANALOG

1499
00:55:19,906 --> 00:55:21,174
FORMAT USING RELATIVELY LOW

1500
00:55:21,174 --> 00:55:25,845
POWER. SO WONDERING WHERE YOU

1501
00:55:25,845 --> 00:55:27,280
THINK SCALE IS GOING AND BEING

1502
00:55:27,280 --> 00:55:28,848
ABLE TO SCALE TO LARGE ENOUGH

1503
00:55:28,848 --> 00:55:29,983
UNITS AND HOW WILL THAT HAPPEN?

1504
00:55:29,983 --> 00:55:32,552
>> YEAH. SO IF I FORGET
THE

1505
00:55:32,552 --> 00:55:34,554
SECOND QUESTION REMIND ME.

1506
00:55:34,554 --> 00:55:36,422
QUICKLY ON THE FLY WIRE WORKS.

1507
00:55:36,422 --> 00:55:39,792
SO THIS IS HOT OFF THE PRESS. WE

1508
00:55:39,792 --> 00:55:40,894
HADN'T FULLY TESTED EVERYTHING

1509
00:55:40,894 --> 00:55:43,429
ON IT. THE PAPER JUST CAME OUT A

1510
00:55:43,429 --> 00:55:45,865
FEW WEEKS AGO. BUT WHAT WAS NICE

1511
00:55:45,865 --> 00:55:47,667
IS THE PAPER THAT WE ARE BASED

1512
00:55:47,667 --> 00:55:49,269
ON IS NOT JUST THE CONNECTOME,

1513
00:55:49,269 --> 00:55:51,304
IT IS ACTUALLY A PAPER DESCRIBED

1514
00:55:51,304 --> 00:55:52,972
THE COMPUTATIONAL MODEL FROM THE

1515
00:55:52,972 --> 00:55:57,410
CONNECTOME. THAT MODEL WAS

1516
00:55:57,410 --> 00:55:59,312
WRITTEN IN BRIAN WHICH IS A

1517
00:55:59,312 --> 00:56:00,980
SPIKING NEURAL NETWORK MODELING

1518
00:56:00,980 --> 00:56:04,384
TOOL WHICH EFFECTIVELY IS DOING

1519
00:56:04,384 --> 00:56:05,952
LEAKY INTEGRATED FIRE NEURONS.

1520
00:56:05,952 --> 00:56:07,554
THEY ARE BIT MORE COMPLEX BUT

1521
00:56:07,554 --> 00:56:10,156
NOT TREMENDOUSLY SO. SO WE HAVE

1522
00:56:10,156 --> 00:56:13,092
-- WHAT WAS NICE IS THAT MODEL

1523
00:56:13,092 --> 00:56:14,827
BASICALLY WE CAN TAKE DIRECTLY

1524
00:56:14,827 --> 00:56:15,862
FROM THE CODE AND TRANSLATE IT

1525
00:56:15,862 --> 00:56:20,934
ON TO WHAT THE CORE NEURAL

1526
00:56:20,934 --> 00:56:22,168
DIGITAL NEURAL MODELS ARE THAT

1527
00:56:22,168 --> 00:56:25,071
ARE IMPLEMENTED BY THE LOIHI

1528
00:56:25,071 --> 00:56:26,839
TOOL CHIP. SO THEY ARE NOT

1529
00:56:26,839 --> 00:56:29,842
ANALOG, IT IS FULLY DIGITAL, IT

1530
00:56:29,842 --> 00:56:31,511
IS BASICALLY THE WAY THAT CHIP

1531
00:56:31,511 --> 00:56:34,414
WORKS IS THAT IT HAS 100 SOME

1532
00:56:34,414 --> 00:56:35,648
ODD CORES THAT ARE EACH CAPABLE

1533
00:56:35,648 --> 00:56:37,016
OF MODELING ABOUT A THOUSAND

1534
00:56:37,016 --> 00:56:40,520
NEURONS EACH. IN PRACTICE YOU

1535
00:56:40,520 --> 00:56:44,290
GET A FEW HUNDRED NEURONS EACH

1536
00:56:44,290 --> 00:56:47,093
CORE. SO IT WAS MORE OR LESS JIG

1537
00:56:47,093 --> 00:56:48,861
SAW PUZZLE. WE WANT TO TAKE THAT

1538
00:56:48,861 --> 00:56:50,830
MODEL THAT HAD 125,000 NEURONS,

1539
00:56:50,830 --> 00:56:53,766
SPLIT IT UP INTO PIECES AN

1540
00:56:53,766 --> 00:56:55,268
DISTRIBUTE OVER THE LOIHI SYSTEM

1541
00:56:55,268 --> 00:56:57,604
THAT WE HAVE. SO THAT WASN'T --

1542
00:56:57,604 --> 00:56:59,806
WHAT IS NICE IS, WE DIDN'T DO

1543
00:56:59,806 --> 00:57:01,774
THE WORK. I DON'T WANT TO TAKE

1544
00:57:01,774 --> 00:57:03,142
CREDIT FOR THAT. THAT WAS THE

1545
00:57:03,142 --> 00:57:05,111
WORK OF THE FLY WIRE TEAM MAKING

1546
00:57:05,111 --> 00:57:07,747
THAT MODEL WE SHOW YOU CAN PUT

1547
00:57:07,747 --> 00:57:13,620
THAT TO NEUROMODULATOR. THE

1548
00:57:13,620 --> 00:57:15,088
SECOND IS WHERE IS THIS GOING TO

1549
00:57:15,088 --> 00:57:16,689
SCALE AND WHERE IS IT GOING TO

1550
00:57:16,689 --> 00:57:18,391
GO? DEPENDS ON IF YOU WANT TO DO

1551
00:57:18,391 --> 00:57:20,159
MORE BRAIN SIMULATION STUFF. OR

1552
00:57:20,159 --> 00:57:25,031
IF YOU WANT TO GO THIS AI ROUTE 

1553
00:57:25,031 --> 00:57:28,468
I DON'T -- I THINK THE BRAIN

1554
00:57:28,468 --> 00:57:30,169
SIMULATION SIDE WE WILL HAVE A

1555
00:57:30,169 --> 00:57:32,472
PACKET.  THERE'S ENGINEERING

1556
00:57:32,472 --> 00:57:33,873
CONSTRAINTS, NUMBER OF SYNAPSES

1557
00:57:33,873 --> 00:57:35,742
PER NEURON IS BOUND BY LOCAL

1558
00:57:35,742 --> 00:57:39,279
MEMORY AVAILABLE PER CORE, THINK

1559
00:57:39,279 --> 00:57:44,951
ABOUT SOMETHING LIKE LOCUST

1560
00:57:44,951 --> 00:57:46,653
ARCURELIUS NOREPINEPHRINE

1561
00:57:46,653 --> 00:57:47,854
NEURON, THAT IS A BILLION

1562
00:57:47,854 --> 00:57:48,855
NEURONS IN THE BRAIN. IN THAT

1563
00:57:48,855 --> 00:57:50,089
STRUCTURE WE HAVE TO FIGURE HOW

1564
00:57:50,089 --> 00:57:51,958
TO MAP THAT TO IF PHYSICAL

1565
00:57:51,958 --> 00:57:53,426
HARDWARE BUT THAT IS GOING TO

1566
00:57:53,426 --> 00:57:55,528
TAKE WORK, IT IS POSSIBLE AND I

1567
00:57:55,528 --> 00:57:59,832
THINK WE CAN PROBABLY GET TO

1568
00:57:59,832 --> 00:58:01,534
REALISTIC SIZE, WE HAVE A

1569
00:58:01,534 --> 00:58:02,535
BILLION NEURON SYSTEM TODAY THAT

1570
00:58:02,535 --> 00:58:04,203
WE ARE TRYING TO FIGURE HOW TO

1571
00:58:04,203 --> 00:58:08,408
BUILD THINGS UP ON SO

1572
00:58:08,408 --> 00:58:10,943
LIMITATION, THERE ARE

1573
00:58:10,943 --> 00:58:12,045
LIMITATIONS BUT NOTHING

1574
00:58:12,045 --> 00:58:13,813
FUNDAMENTAL THERE. WHEN IT COMES

1575
00:58:13,813 --> 00:58:15,915
TO LARGE LANGUAGE MODELS I WOULD

1576
00:58:15,915 --> 00:58:18,785
SAY THE UNDERLYING STRUCTURE OF

1577
00:58:18,785 --> 00:58:20,853
THOSE MODELS IS NOT VERY

1578
00:58:20,853 --> 00:58:23,089
BIOLOGICBIOLOGICAL. IT MIGHT INE

1579
00:58:23,089 --> 00:58:26,259
THINGS THAT ARE NEURONS AND

1580
00:58:26,259 --> 00:58:28,361
WEIGHTS BUT THOSE MODELS IN MANY

1581
00:58:28,361 --> 00:58:30,396
WAYS ARE NOT GOOD FITS FOR

1582
00:58:30,396 --> 00:58:31,764
NEUROMORPHIC HARD WARE. I WOULD

1583
00:58:31,764 --> 00:58:34,300
SAY -- PEOPLE ARE TRYING TO DO

1584
00:58:34,300 --> 00:58:37,236
THIS, I'M CHEERING THEM ON, I'M

1585
00:58:37,236 --> 00:58:38,604
NOT INCREDIBLY OPTIMISTIC THAT

1586
00:58:38,604 --> 00:58:42,775
WE ARE GOING TO PUTS CHAT GPT ON

1587
00:58:42,775 --> 00:58:46,579
TO LOIHI 2 ANY TIME SOON. WHAT I

1588
00:58:46,579 --> 00:58:49,048
THINK WE MIGHT BE ABLE TO DO IS

1589
00:58:49,048 --> 00:58:51,017
THIS SEE WHAT THESE CHIPS ARE

1590
00:58:51,017 --> 00:58:53,219
REALLY GOOD AT, MIGHT HELP

1591
00:58:53,219 --> 00:58:55,722
CONSTRAIN US TO THINK ABOUT WHAT

1592
00:58:55,722 --> 00:58:59,759
A DIFFERENT FORMULATION OF A

1593
00:58:59,759 --> 00:59:01,394
LARGE LANGUAGE MODEL SHOULD BE,

1594
00:59:01,394 --> 00:59:02,662
DIFFERENT FORMULATION OF LARGE

1595
00:59:02,662 --> 00:59:03,963
LANGUAGE MODEL SHOULD BE AND

1596
00:59:03,963 --> 00:59:06,132
THAT MIGHT HELP US REDESCRIBE IN

1597
00:59:06,132 --> 00:59:07,800
A WAY THAT IS ENERGY EFFICIENT

1598
00:59:07,800 --> 00:59:09,569
WHICH IS A VERY INTERESTING

1599
00:59:09,569 --> 00:59:11,504
SCIENTIFIC QUESTION BUT IT IS --

1600
00:59:11,504 --> 00:59:13,039
I DON'T HAVE AN IMMEDIATE

1601
00:59:13,039 --> 00:59:16,876
ANSWER. THANKS.

1602
00:59:16,876 --> 00:59:23,549
>> WE HAVE ONE MORE
QUESTION.

1603
00:59:23,549 --> 00:59:25,985
>> SO AS YOU SAID WE LEARN
FROM

1604
00:59:25,985 --> 00:59:29,522
OUR ERRORS.  WHEN DO YOU THINK

1605
00:59:29,522 --> 00:59:31,557
YOU WILL HAVE NEUROMORPHIC

1606
00:59:31,557 --> 00:59:34,093
HARDWARE THAT WILL LOOK

1607
00:59:34,093 --> 00:59:37,263
SOMETHING LIKE CHAT GPT AND

1608
00:59:37,263 --> 00:59:38,297
REPROGRAM ITSELF?

1609
00:59:38,297 --> 00:59:41,634
>> IT IS A FUN QUESTION. I
THINK

1610
00:59:41,634 --> 00:59:45,138
IF IT IS FULLY PROGRAMMING

1611
00:59:45,138 --> 00:59:47,306
ITSELF I DON'T KNOW IF WE ARE

1612
00:59:47,306 --> 00:59:48,674
ANYWHERE CLOSE TO THAT. I DO

1613
00:59:48,674 --> 00:59:50,443
THINK WE WANT TO HAVE LEARNING.

1614
00:59:50,443 --> 00:59:51,778
IN FACT THESE CHIPS DO HAVE THE

1615
00:59:51,778 --> 00:59:53,179
ABILITY TO LEARN. WE ARE -- NONE

1616
00:59:53,179 --> 00:59:56,215
OF THE ALGORITHMS REALLY USE

1617
00:59:56,215 --> 00:59:58,785
THAT. WE MIGHT HAVE CHIPS OR

1618
00:59:58,785 --> 01:00:03,189
MIGHT HAVE ALGORITHMS, NOT TOO

1619
01:00:03,189 --> 01:00:05,391
DISTANCE FUTURE THAT ARE

1620
01:00:05,391 --> 01:00:08,628
HOPEFULLY VERY BRAIN INSPIRED

1621
01:00:08,628 --> 01:00:09,629
THAT ARE INTRINSICALLY LEARNING

1622
01:00:09,629 --> 01:00:11,164
SO THE MODELS PUT ON TO THE

1623
01:00:11,164 --> 01:00:14,534
CHIPS WILL ADAM DEPARTMENT

1624
01:00:14,534 --> 01:00:15,501
THEMSELVES OVER TIME. SO NOT

1625
01:00:15,501 --> 01:00:20,072
LEARNING THE SAME AS FPGA

1626
01:00:20,072 --> 01:00:20,973
RECONFIGURES THEMSELVES.  THEY

1627
01:00:20,973 --> 01:00:23,509
ARE SIMILAR TO FPGAs IN TERMS

1628
01:00:23,509 --> 01:00:24,744
HOW WE PROGRAM AND THINK ABOUT

1629
01:00:24,744 --> 01:00:25,645
THEM SO WHAT YOU ARE SUGGESTING

1630
01:00:25,645 --> 01:00:28,047
IS NOT IMPOSSIBLE BUT JUST NOW

1631
01:00:28,047 --> 01:00:30,750
HOW WE ARE REALLY THINKING ABOUT

1632
01:00:30,750 --> 01:00:30,917
IT.

1633
01:00:30,917 --> 01:00:32,385
>> SO NEUROMORPHIC PART,
YOU

1634
01:00:32,385 --> 01:00:33,152
HAVE SPIKING, RIGHT?

1635
01:00:33,152 --> 01:00:33,953
>> YES.

1636
01:00:33,953 --> 01:00:36,589
>> BUT JUST THIS MONDAY I
HEARD

1637
01:00:36,589 --> 01:00:38,591
A TALK BY SOMEONE WHO USED THE

1638
01:00:38,591 --> 01:00:43,162
CONNECTOME FROM THE FLY WIRE AND

1639
01:00:43,162 --> 01:00:46,466
IN THAT SHE SHOWED THAT THERE'S

1640
01:00:46,466 --> 01:00:51,804
A SINGLE SPIKE THAT IS INDUCED

1641
01:00:51,804 --> 01:00:54,941
THAT LEADS TO A SPECIFIC

1642
01:00:54,941 --> 01:00:57,743
BEHAVIOR IN THE FLY. SO THAT IS

1643
01:00:57,743 --> 01:01:04,250
VERY FAR FROM PROBABLEI PROBABIO

1644
01:01:04,250 --> 01:01:06,452
CURIOUS YEAH IT CAN SOLVE RANDOM

1645
01:01:06,452 --> 01:01:08,354
WALKS, A LOT OF CENTRAL LIMITS

1646
01:01:08,354 --> 01:01:12,558
HERE IS REALLY HELPFUL BUT FOR

1647
01:01:12,558 --> 01:01:18,998
SUCH SINGLE SPIKE EVENTS, THAT'S

1648
01:01:18,998 --> 01:01:19,499
HARD TO GET.

1649
01:01:19,499 --> 01:01:24,604
>> I WILL SAY THESE CHIPS
ARE

1650
01:01:24,604 --> 01:01:26,405
RATHER CONVENTIONAL UNDER THE

1651
01:01:26,405 --> 01:01:29,509
HOOD. THESE ARE DIGITAL CMOS

1652
01:01:29,509 --> 01:01:31,511
ELECTRONICS. IF I WANT TO TURN

1653
01:01:31,511 --> 01:01:33,112
OFF NOISE I CAN TURN OFF NOISE

1654
01:01:33,112 --> 01:01:37,016
ON THE CURRENT HARDWARE. I DO

1655
01:01:37,016 --> 01:01:39,352
THINK THAT THE FLY BRAIN AND

1656
01:01:39,352 --> 01:01:41,254
MAMMALIAN BRAINS ARE DIFFERENT

1657
01:01:41,254 --> 01:01:43,723
IN TERMS OF -- I'M NOT AWARE IN

1658
01:01:43,723 --> 01:01:45,591
THE MAMMALIAN BRAINS WHERE YOU

1659
01:01:45,591 --> 01:01:46,959
HAVE SINGLE NEURONS THAT HAVE

1660
01:01:46,959 --> 01:01:49,428
FULLY DETERMINISTIC BEHAVIOR THE

1661
01:01:49,428 --> 01:01:52,131
WAY THAT PERFECTLY WELL DEFINED

1662
01:01:52,131 --> 01:01:55,034
CONNECTOMES LIKE THE FLY

1663
01:01:55,034 --> 01:01:58,237
CONNECTOME IS VERY STRUCTURED. 

1664
01:01:58,237 --> 01:01:59,739
WHEREAS OUR BRAINS ARE MUCH MORE

1665
01:01:59,739 --> 01:02:01,974
RANDOMLY CONNECTED. I DO THINK

1666
01:02:01,974 --> 01:02:03,309
THERE IS FUNDAMENTAL QUESTION

1667
01:02:03,309 --> 01:02:05,444
BUT THOSE ARE MORE COMPUTATIONAL

1668
01:02:05,444 --> 01:02:07,113
NEUROSCIENCE QUESTIONS. WE CAN

1669
01:02:07,113 --> 01:02:08,114
PROBABLY IMPLEMENT EITHER

1670
01:02:08,114 --> 01:02:11,450
VERSION IN NEUROMORPHIC. BUT

1671
01:02:11,450 --> 01:02:14,020
THAT SOFTER NEURON IS THE H 1 --

1672
01:02:14,020 --> 01:02:15,488
I CAN REMEMBER, IT'S BEEN

1673
01:02:15,488 --> 01:02:18,491
STUDIED A WHILE IN FLIES. IT

1674
01:02:18,491 --> 01:02:20,593
WILL BE INTERESTING TO LOOK IN

1675
01:02:20,593 --> 01:02:21,827
-- WE HAVEN'T PLAY WITH THE

1676
01:02:21,827 --> 01:02:23,296
MODEL TOO MUCH TO BE HONEST BUT

1677
01:02:23,296 --> 01:02:23,996
TO SEE HOW THAT WORKS.

1678
01:02:23,996 --> 01:02:27,500
>> THANK YOU. 

1679
01:02:27,500 --> 01:02:29,702
>> JUST QUICKLY, GREAT
TALK,

1680
01:02:29,702 --> 01:02:31,404
GREAT TO SEE THE FLY WIRE

1681
01:02:31,404 --> 01:02:34,407
IMPLEMENT SOD QUICKLY. THE BRAIN

1682
01:02:34,407 --> 01:02:36,509
INITIATIVE IS SELECTING

1683
01:02:36,509 --> 01:02:37,977
COLLECTING OTHER TYPES OF DATA

1684
01:02:37,977 --> 01:02:41,180
INCLUDING CELL TYPE DATA DATA. E

1685
01:02:41,180 --> 01:02:42,315
RELEASED A WHOLE MOUSE BRAIN

1686
01:02:42,315 --> 01:02:43,983
CELL TYPE ATLAS, MULTIPLE

1687
01:02:43,983 --> 01:02:45,885
THOUSANDS OF DIFFERENT TYPES OF

1688
01:02:45,885 --> 01:02:49,221
CELLS. SO THE CONNECTOME IS NOT

1689
01:02:49,221 --> 01:02:51,657
JUST HOMOGENOUS CONNECTION

1690
01:02:51,657 --> 01:02:52,892
GRAPH, IT IS CONNECTIONS BETWEEN

1691
01:02:52,892 --> 01:02:53,993
DIFFERENT TYPES OF CELLS WITH

1692
01:02:53,993 --> 01:02:55,561
DIFFERENT RESPONSES, DIFFERENT

1693
01:02:55,561 --> 01:02:57,029
RESPONSES TO NEUROMODULATORS,

1694
01:02:57,029 --> 01:02:58,831
ALL THAT COMPLEXITY YOU ARE

1695
01:02:58,831 --> 01:03:01,100
TALKING ABOUT. WHAT DO YOU SEE

1696
01:03:01,100 --> 01:03:02,768
AS FAR AS WHETHER -- WELL, HOW

1697
01:03:02,768 --> 01:03:04,870
CAN YOU INTEGRATE THAT KIND OF

1698
01:03:04,870 --> 01:03:06,872
DATA ACROSS DIFFERENT SCALES,

1699
01:03:06,872 --> 01:03:08,341
CELL TYPE CONNECTIVITY AND HOW

1700
01:03:08,341 --> 01:03:11,744
DOES THAT RELATE TO THE MAJOR

1701
01:03:11,744 --> 01:03:13,879
OBSTACLES YOU SEE MAJOR ERROR UP

1702
01:03:13,879 --> 01:03:16,349
TO THE RIGHT OF THAT COMPLEXITY

1703
01:03:16,349 --> 01:03:20,119
VERSUS SIZE PLOT. AND HOW IT

1704
01:03:20,119 --> 01:03:20,620
RELATES?

1705
01:03:20,620 --> 01:03:24,290
>> I LOVE THAT QUESTION
BECAUSE

1706
01:03:24,290 --> 01:03:28,361
I -- PUT THIS IN MY ORIGINAL

1707
01:03:28,361 --> 01:03:30,663
DENTE MODELING STUFF,

1708
01:03:30,663 --> 01:03:31,998
HETEROGENEITY IS FUNDAMENTAL HOW

1709
01:03:31,998 --> 01:03:33,599
THE BRAIN FUNCTIONS AND CRITICAL

1710
01:03:33,599 --> 01:03:35,434
HOW DISEASE INTERACTS WITH

1711
01:03:35,434 --> 01:03:37,436
BRAIN. I IT IS NOT WHAT

1712
01:03:37,436 --> 01:03:40,640
NEUROMORPHIC COMPUTING IS -- WE

1713
01:03:40,640 --> 01:03:42,108
HAVE DONE A LOT OF NEUROMORPHIC

1714
01:03:42,108 --> 01:03:44,110
BUT A LOT OF NEUROMORPHIC

1715
01:03:44,110 --> 01:03:46,112
SYSTEMS CANNED HANDLE THAT, BUT

1716
01:03:46,112 --> 01:03:47,913
JUST HAVEN'T DONE IT YET.

1717
01:03:47,913 --> 01:03:50,683
SOMETHING GPUs ARE TERRIBLE

1718
01:03:50,683 --> 01:03:52,685
HETEROGENEITY THAT'S PLACE

1719
01:03:52,685 --> 01:03:54,887
NEUROMORPHIC HAS ADVANTAGE BUT

1720
01:03:54,887 --> 01:03:56,022
THERE'S REAL WORK THAT WILL BE

1721
01:03:56,022 --> 01:03:57,990
NEEDED TO DIVE INTO SOME OF THIS

1722
01:03:57,990 --> 01:04:00,893
DATA AND START TO REALLY KICK

1723
01:04:00,893 --> 01:04:03,229
THE TIRES ON WHAT WE CAN DO SO A

1724
01:04:03,229 --> 01:04:05,798
FANTASTIC DIRECTION TO GO.

1725
01:04:05,798 --> 01:04:08,467
>> THANK YOU SO MUCH,
BRAD, FOR

1726
01:04:08,467 --> 01:04:10,469
TELLING US ABOUT HOW

1727
01:04:10,469 --> 01:04:11,270
NEUROMORPHIC CAN HELP US

1728
01:04:11,270 --> 01:04:13,539
UNDERSTAND NEUROSCIENCE AND

1729
01:04:13,539 --> 01:04:14,907
THERE IS PA WALS RECEPTION RIGHT

1730
01:04:14,907 --> 01:04:17,576
NOW. SO LET'S GIVE A HAND TO OUR

1731
01:04:17,576 --> 01:04:27,576
SPARE.