1
00:00:00,000 --> 00:00:03,517
The following is a conversation
with Yann LeCun, his second time on the

2
00:00:03,600 --> 00:00:06,757
podcast. He is the chief AI scientist at

3
00:00:06,840 --> 00:00:10,357
Meta, formerly Facebook, professor at

4
00:00:10,440 --> 00:00:15,637
NYU, Turing Award winner, one
of the seminal figures in the history of

5
00:00:15,720 --> 00:00:19,757
machine learning and artificial
intelligence, and someone who is

6
00:00:19,840 --> 00:00:24,557
brilliant and opinionated in the best
kind of way, and so is always fun to

7
00:00:24,640 --> 00:00:28,197
talk to. And now a quick
few second mention of

8
00:00:28,280 --> 00:00:31,717
each sponsor. Check them out
in the description. It's the best way to

9
00:00:31,800 --> 00:00:34,477
support this podcast. First is Public

10
00:00:34,560 --> 00:00:36,677
Goods, an online shop I use for

11
00:00:36,760 --> 00:00:40,157
household products. Second is Indeed, a

12
00:00:40,240 --> 00:00:43,117
hiring website. Third is Roka, my

13
00:00:43,200 --> 00:00:47,397
favorite sunglasses and prescription
glasses. Fourth is NetSuite, business

14
00:00:47,480 --> 00:00:49,957
software for managing HR, financials,

15
00:00:50,040 --> 00:00:53,117
and other details. And fifth is Magic

16
00:00:53,200 --> 00:00:56,077
Spoon, low-carb, keto-friendly cereal.

17
00:00:56,160 --> 00:00:59,917
So, the choice is business,
health, or style.

18
00:01:00,000 --> 00:01:04,637
Choose wisely, my friends.
And now onto the full ads. As always, no

19
00:01:04,720 --> 00:01:08,037
ads in the middle. I try to make
this interesting, but if you skip them,

20
00:01:08,120 --> 00:01:12,797
please still check out our sponsors.
I enjoy their stuff, maybe you will, too.

21
00:01:12,880 --> 00:01:14,837
This show is brought to you by Public

22
00:01:14,920 --> 00:01:17,997
Goods, the one-stop shop for affordable,

23
00:01:18,080 --> 00:01:21,197
sustainable, healthy household products.

24
00:01:21,280 --> 00:01:26,557
I use their hand soap, toothpaste,
and toothbrush. I think I use a bunch of

25
00:01:26,640 --> 00:01:30,077
other stuff, too, but that's
that's what comes to mind.

26
00:01:30,160 --> 00:01:33,037
The products often have this minimalist

27
00:01:33,120 --> 00:01:37,597
black and white design that I just
absolutely find beautiful. I love it. I

28
00:01:37,680 --> 00:01:41,757
love minimalism in design. It doesn't
go over the top. It doesn't doesn't have

29
00:01:41,840 --> 00:01:43,837
all these extra things and features that

30
00:01:43,920 --> 00:01:46,877
you don't need, just the essentials.

31
00:01:46,960 --> 00:01:51,077
I think it's hard to explain.
But there's something about

32
00:01:51,160 --> 00:01:55,637
the absence of things
that that take up your

33
00:01:55,720 --> 00:01:59,437
attention that allows you
to truly be attentive to

34
00:01:59,520 --> 00:02:03,597
what matters in life.
Anyway, go to publicgoods.com/lex

35
00:02:03,680 --> 00:02:08,357
or use code Lex at checkout to get
$15 off your first order, plus you will

36
00:02:08,440 --> 00:02:11,917
receive your choice of either
free pack of bamboo straws

37
00:02:12,000 --> 00:02:14,557
or reusable food storage wraps.

38
00:02:14,640 --> 00:02:18,997
Visit publicgoods.com/lex
or use code Lex at checkout.

39
00:02:19,080 --> 00:02:20,757
This show is also brought to you by

40
00:02:20,840 --> 00:02:23,957
Indeed, a hiring website. I've used them

41
00:02:24,040 --> 00:02:28,237
as part of many hiring efforts I've done
for the teams I've led in the past. Most

42
00:02:28,320 --> 00:02:32,277
of those for was for engineering,
for research efforts.

43
00:02:32,360 --> 00:02:36,037
They have tools like Indeed instant
match giving you quality candidates

44
00:02:36,120 --> 00:02:40,237
whose resumes on Indeed fit your job
description immediately. For the past

45
00:02:40,320 --> 00:02:44,877
few months I've been going through this
process of um building up a team of uh

46
00:02:44,960 --> 00:02:48,517
folks to help me. I've been
doing quite a bit of hiring.

47
00:02:48,600 --> 00:02:51,637
It's a treacherous

48
00:02:51,720 --> 00:02:56,837
and an exciting process because you get
to meet some friends. So, it's a it's a

49
00:02:56,920 --> 00:03:01,157
beautiful process, but I think it's one
of the most important processes in life.

50
00:03:01,240 --> 00:03:05,277
It's selecting the group of people
with whom you spend your days with.

51
00:03:05,360 --> 00:03:08,877
And so, you got to use the best tools
for the job. Indeed, I think is an

52
00:03:08,960 --> 00:03:13,197
excellent tool. Right now,
you can get a free $75

53
00:03:13,280 --> 00:03:17,477
sponsored job credit to upgrade
your job post at indeed.com/lex.

54
00:03:17,560 --> 00:03:23,040
Terms and conditions apply.
Go to indeed.com/lex.

55
00:03:23,480 --> 00:03:25,037
This show is also brought to you by

56
00:03:25,120 --> 00:03:27,997
Roka, the makers of glasses and

57
00:03:28,080 --> 00:03:32,717
sunglasses that I love wearing
for their design, feel, and

58
00:03:32,800 --> 00:03:36,477
innovation on material, optics,
and grip. Roka was started by two

59
00:03:36,560 --> 00:03:40,837
all-American swimmers from Stanford,
and it was born out of an obsession with

60
00:03:40,920 --> 00:03:43,397
performance. I like the way they feel. I

61
00:03:43,480 --> 00:03:45,357
like the way they look.

62
00:03:45,440 --> 00:03:48,277
Whether I'm uh running like a fast-paced

63
00:03:48,360 --> 00:03:50,637
run, we're talking about 8-minute mile

64
00:03:50,720 --> 00:03:52,757
or faster, or if we're doing a slow pace

65
00:03:52,840 --> 00:03:57,357
run, 9 10-minute mile along the river in

66
00:03:57,440 --> 00:04:00,837
the heat or in the cold, or if I'm just

67
00:04:00,920 --> 00:04:02,757
wearing my suit

68
00:04:02,840 --> 00:04:05,437
out on the town, however that expression

69
00:04:05,520 --> 00:04:10,877
goes, I'm not sure, but they look classy
with a suit, they look badass in running

70
00:04:10,960 --> 00:04:15,397
gear. It's just uh my go-to sunglasses.

71
00:04:15,480 --> 00:04:19,477
Check them out for both prescription
glasses and sunglasses at roka.com and

72
00:04:19,560 --> 00:04:24,837
enter code Lex to save 20% on your first
order. That's roka.com and enter code

73
00:04:24,920 --> 00:04:30,197
Lex. This show is also brought to you
by NetSuite. NetSuite allows you to manage

74
00:04:30,280 --> 00:04:34,997
financials, human resources,
inventory, e-commerce,

75
00:04:35,080 --> 00:04:38,997
and uh many more business-related
details all in one place. I'm not sure

76
00:04:39,080 --> 00:04:42,357
why I was doing upspeak on that sentence.

77
00:04:42,440 --> 00:04:46,597
Maybe because I'm very excited about
NetSuite. Anyway, there's a lot of messy

78
00:04:46,680 --> 00:04:49,237
things you have to get
right when running a company.

79
00:04:49,320 --> 00:04:52,037
If you're a business owner,
if you're an entrepreneur,

80
00:04:52,120 --> 00:04:56,637
if you're a founder of a startup,
this is something

81
00:04:56,720 --> 00:05:00,877
you have to think about. You have to use
the best tools for the job to make sure

82
00:05:00,960 --> 00:05:04,997
all the messy things required to run
a business are taken care of for you so

83
00:05:05,080 --> 00:05:07,117
you can focus on the things that you're

84
00:05:07,200 --> 00:05:09,837
best at, that where your brilliance

85
00:05:09,920 --> 00:05:14,917
shines. If you are starting a business,
I wish you the best of luck. It's a

86
00:05:15,000 --> 00:05:19,277
difficult journey, but it's worth it.

87
00:05:19,360 --> 00:05:23,797
Anyway, right now special financing
is back. Head to netsuite.com/lex

88
00:05:23,880 --> 00:05:25,757
to get their one-of-a-kind financing

89
00:05:25,840 --> 00:05:29,477
program. That's netsuite.com/lex.

90
00:05:29,560 --> 00:05:34,517
netsuite.com/lex This episode
is also brought to you by

91
00:05:34,600 --> 00:05:38,837
Magic Spoon. The OG,
not quite OG, but really

92
00:05:38,920 --> 00:05:42,877
old-school sponsor of this podcast that I
love. It's a low-carb, keto-friendly

93
00:05:42,960 --> 00:05:45,357
cereal. It has 0 g of sugar. It's

94
00:05:45,440 --> 00:05:50,237
delicious. I don't say that enough. It's
really is delicious. Given that it's 0 g

95
00:05:50,320 --> 00:05:51,917
of sugar, it's very surprising how

96
00:05:52,000 --> 00:05:55,117
delicious it is. 13 to 14 g of protein,

97
00:05:55,200 --> 00:05:58,477
only 4 net g of carbs, and 140 calories

98
00:05:58,560 --> 00:06:00,877
in in each serving. You can build your

99
00:06:00,960 --> 00:06:05,797
own box or get a variety pack with
available flavors of cocoa, fruity,

100
00:06:05,880 --> 00:06:09,037
frosted, peanut butter,
blueberry, and cinnamon.

101
00:06:09,120 --> 00:06:12,437
Cocoa is my favorite. It's the flavor
of champions. I don't know why I keep

102
00:06:12,520 --> 00:06:14,477
saying that, but it seems to be true.

103
00:06:14,560 --> 00:06:17,837
Anyway, Magic Spoon has a 100% happiness

104
00:06:17,920 --> 00:06:22,677
guarantee. So, if you don't like it,
they will refund it. Who else

105
00:06:22,760 --> 00:06:25,637
will give you a 100% happiness guarantee?

106
00:06:25,720 --> 00:06:30,477
Go to magicspoon.com/lex and use
code Lex at checkout to save $5

107
00:06:30,560 --> 00:06:33,797
off your order. That's magicspoon.com/lex

108
00:06:33,880 --> 00:06:35,837
and use code Lex.

109
00:06:35,920 --> 00:06:38,477
This is the Lex Fridman podcast, and

110
00:06:38,560 --> 00:06:43,600
here's my conversation with Yann LeCun.

111
00:06:53,960 --> 00:06:55,980
>> [music]

112
00:06:58,160 --> 00:07:00,317
>> You co-wrote
the article Self-Supervised

113
00:07:00,400 --> 00:07:04,117
Learning: The Dark Matter of Intelligence.
Great title, by the way.

114
00:07:04,200 --> 00:07:06,717
With Ishan Misra. So, let me ask, what

115
00:07:06,800 --> 00:07:09,277
is self-supervised learning and why is

116
00:07:09,360 --> 00:07:11,837
it the dark matter of intelligence?

117
00:07:11,920 --> 00:07:14,277
>> I'll start by the dark matter part.

118
00:07:14,360 --> 00:07:17,557
Uh There is obviously a kind of learning

119
00:07:17,640 --> 00:07:20,437
that humans and animals are

120
00:07:20,520 --> 00:07:23,677
uh are doing that we currently are not

121
00:07:23,760 --> 00:07:27,597
reproducing properly with machines
or with AI, right? So, the most popular

122
00:07:27,680 --> 00:07:31,797
approaches to machine learning today
are or paradigms, I should say, are

123
00:07:31,880 --> 00:07:34,637
supervised learning
and reinforcement learning.

124
00:07:34,720 --> 00:07:37,037
And they are extremely efficient.

125
00:07:37,120 --> 00:07:41,677
Supervised learning requires many
samples for learning anything.

126
00:07:41,760 --> 00:07:46,357
And reinforcement learning requires
a ridiculously large number of trial and

127
00:07:46,440 --> 00:07:49,677
errors to for, you know,
a system to learn anything.

128
00:07:49,760 --> 00:07:54,147
Um And that's why we don't
have self-driving cars.

129
00:07:54,230 --> 00:07:56,677
>> [laughter] >> That's a big
leap from one to the other.

130
00:07:56,760 --> 00:07:59,357
Okay, so that to solve difficult

131
00:07:59,440 --> 00:08:02,877
problems, you have to have a lot of uh

132
00:08:02,960 --> 00:08:07,077
human annotation for supervised learning
to work. And to solve those difficult

133
00:08:07,160 --> 00:08:09,917
problems with reinforcement
learning, you have to have

134
00:08:10,000 --> 00:08:13,637
some way to maybe simulate that problem
such that you can do that large-scale

135
00:08:13,720 --> 00:08:16,317
kind of learning that
reinforcement learning requires.

136
00:08:16,400 --> 00:08:19,197
>> Right. So, how is it that, you know,

137
00:08:19,280 --> 00:08:20,837
most teenagers can learn to drive a car

138
00:08:20,920 --> 00:08:24,157
in about 20 hours of practice?

139
00:08:24,240 --> 00:08:26,917
Whereas uh even with millions of hours

140
00:08:27,000 --> 00:08:31,957
of simulated practice, a self-driving
car can't actually learn to drive itself

141
00:08:32,040 --> 00:08:35,117
properly. Um and so,
obviously, we're missing

142
00:08:35,200 --> 00:08:38,597
something, right? And And it's quite
obvious for a lot of people that, you

143
00:08:38,680 --> 00:08:42,517
know, the immediate response you get from
many people is, well, you know, humans

144
00:08:42,600 --> 00:08:47,997
use their background knowledge
to learn faster. And they're right. Now,

145
00:08:48,080 --> 00:08:51,997
how was that background knowledge
acquired? And that's to be questioned.

146
00:08:52,080 --> 00:08:54,477
So, now you have to ask, you know, how

147
00:08:54,560 --> 00:08:59,357
do babies in the first few months
of life learn how the world works? Mostly

148
00:08:59,440 --> 00:09:02,837
by observation because they
can hardly act in the world.

149
00:09:02,920 --> 00:09:05,757
Um and they learn an enormous amount
of background knowledge about the world

150
00:09:05,840 --> 00:09:09,797
that may be the the basis
of what we call common sense.

151
00:09:09,880 --> 00:09:15,117
Uh this type of learning is not learning
a task. It's not being reinforced for

152
00:09:15,200 --> 00:09:20,237
anything. It's just observing the world
and figuring out how how it works.

153
00:09:20,320 --> 00:09:24,877
Building world models. Learning world
models. Um how do we do this? And how do

154
00:09:24,960 --> 00:09:28,197
we reproduce this in in machines?
So, self-supervised learning

155
00:09:28,280 --> 00:09:32,877
is, you know, one instance
or one attempt at trying to

156
00:09:32,960 --> 00:09:34,957
reproduce this kind of learning.

157
00:09:35,040 --> 00:09:39,477
>> Okay, so you're looking at just
observation. so not even the interacting

158
00:09:39,560 --> 00:09:43,677
part of a child. It's just sitting there
watching mom and dad walk around, pick

159
00:09:43,760 --> 00:09:47,357
up stuff, all of that. That's the That's
what you mean about background knowledge.

160
00:09:47,440 --> 00:09:51,877
>> Perhaps not even watching mom and dad,
just, you know, watching the world go by.

161
00:09:51,960 --> 00:09:55,077
>> Just having eyes open or having eyes
closed or the very act of opening and

162
00:09:55,160 --> 00:09:58,917
closing eyes that the world appears
and disappears, all of that basic

163
00:09:59,000 --> 00:10:03,277
information. And you're
saying in in order to learn

164
00:10:03,360 --> 00:10:06,677
to drive like the reason
humans are able to learn

165
00:10:06,760 --> 00:10:10,277
to drive quickly, some faster than
others, is because of the background

166
00:10:10,360 --> 00:10:13,597
knowledge they were able
to watch cars operate in the world

167
00:10:13,680 --> 00:10:17,517
in the many years leading up to it, the
physics of basic objects, all that kind

168
00:10:17,600 --> 00:10:18,837
of stuff. >> That's right.
I mean, the basic physics

169
00:10:18,920 --> 00:10:21,917
of objects, you don't even know you you
don't even need to know, you know, how a

170
00:10:22,000 --> 00:10:24,597
car works, right? Because that you
can learn fairly quickly. I mean, the

171
00:10:24,680 --> 00:10:28,477
example I use very often
is you're driving next to a cliff.

172
00:10:28,560 --> 00:10:33,277
And you know in advance because of your,
you know, understanding of intuitive

173
00:10:33,360 --> 00:10:36,557
physics that if you turn the wheel
to the right, the car will veer to the

174
00:10:36,640 --> 00:10:39,997
right, will run off the cliff,
fall off the cliff, and nothing good

175
00:10:40,080 --> 00:10:44,277
will come out of this, right?
Um but if you are a sort of,

176
00:10:44,360 --> 00:10:47,877
you know, tabular rasa reinforcement
learning system that doesn't have a

177
00:10:47,960 --> 00:10:51,837
model of the world, uh you
have to repeat falling off this

178
00:10:51,920 --> 00:10:55,157
cliff thousands of times before you
figure out it's a bad idea. And then a

179
00:10:55,240 --> 00:10:59,277
few more thousand times before you
figure out how to not do it. And then a

180
00:10:59,360 --> 00:11:02,557
few more million times before you figure
out how to not do it in every situation

181
00:11:02,640 --> 00:11:04,397
you ever encounter.

182
00:11:04,480 --> 00:11:06,797
>> So, self-supervised learning still has

183
00:11:06,880 --> 00:11:11,957
to have some source of truth
being told to it by somebody.

184
00:11:12,040 --> 00:11:14,797
>> And is >> So, you have
to figure out a way without

185
00:11:14,880 --> 00:11:17,317
human assistance or without significant

186
00:11:17,400 --> 00:11:22,277
amount of human assistance to get
that truth from the world. So, the mystery

187
00:11:22,360 --> 00:11:26,517
there is um how much signal
is there, how much truth

188
00:11:26,600 --> 00:11:30,357
is there that the world gives you,
whether it's the human world, like you

189
00:11:30,440 --> 00:11:35,117
watch YouTube or something like that,
or it's the more natural world. So, how

190
00:11:35,200 --> 00:11:39,197
much signal is there? >> So,
here's the trick. There is way more

191
00:11:39,280 --> 00:11:42,517
signal in sort of a
self-supervised setting

192
00:11:42,600 --> 00:11:46,357
than there is in either a supervised
or reinforcement setting.

193
00:11:46,440 --> 00:11:50,477
And this is going to my,
you know, analogy of the cake.

194
00:11:50,560 --> 00:11:53,437
>> Yes. [laughter] >> Uh the,
you know, the cake as someone has

195
00:11:53,520 --> 00:11:56,357
called it, where when you try to figure

196
00:11:56,440 --> 00:12:00,597
out how much information you ask the
machine to predict and how much feedback

197
00:12:00,680 --> 00:12:02,797
you give the machine at every trial.

198
00:12:02,880 --> 00:12:05,557
In reinforcement learning, you give
the machine a single scalar. You tell the

199
00:12:05,640 --> 00:12:09,277
machine you did good, you did bad.
And you and you and you only

200
00:12:09,360 --> 00:12:11,437
tell this to the machine once in a while.

201
00:12:11,520 --> 00:12:15,917
When I say you, it could be the the
universe telling the machine, right? Um

202
00:12:16,000 --> 00:12:19,637
but it's just one scalar. And so,
as a consequence, there is you you cannot

203
00:12:19,720 --> 00:12:22,437
possibly learn something very
complicated without many, many, many

204
00:12:22,520 --> 00:12:25,197
trials where you get many, many feedbacks

205
00:12:25,280 --> 00:12:27,917
uh of this type. Supervised learning,

206
00:12:28,000 --> 00:12:30,437
you you give a few bits to the machine

207
00:12:30,520 --> 00:12:33,037
uh at every every sample.

208
00:12:33,120 --> 00:12:37,557
Let's say you're training a system on,
you know, recognizing images

209
00:12:37,640 --> 00:12:39,517
on ImageNet. There is 1,000 categories.

210
00:12:39,600 --> 00:12:43,957
That's a little less than 10
bits of information per sample.

211
00:12:44,040 --> 00:12:47,437
But, self-supervised learning, here
is the setting. You Ideally, we don't know

212
00:12:47,520 --> 00:12:49,157
how to do this yet, but ideally, you

213
00:12:49,240 --> 00:12:52,517
would show a machine a segment of a

214
00:12:52,600 --> 00:12:56,397
video and then stop the video and ask
ask the machine to predict what's going

215
00:12:56,480 --> 00:12:59,320
to happen next.

216
00:12:59,440 --> 00:13:00,677
And so, you let the machine predict, and

217
00:13:00,760 --> 00:13:04,357
then you let time go by and show the

218
00:13:04,440 --> 00:13:09,277
machine what actually happened, and hope
the machine will, you know, learn to do

219
00:13:09,360 --> 00:13:12,317
a better job at predicting next
time around. There's a huge amount of

220
00:13:12,400 --> 00:13:17,597
information you give the machine because
it's an entire video clip. It's a of uh

221
00:13:17,680 --> 00:13:21,277
you know, the of the the future after
the video clip you fed it um in the

222
00:13:21,360 --> 00:13:24,717
first place. >> So,
both for language and for vision,

223
00:13:24,800 --> 00:13:29,677
there's a subtle, seemingly trivial
construction, but maybe that's

224
00:13:29,760 --> 00:13:33,717
representative of what is required
to create intelligence, which is

225
00:13:33,800 --> 00:13:38,200
filling the gap. So

226
00:13:38,360 --> 00:13:41,637
>> Sounds dumb, but can you it's

227
00:13:41,720 --> 00:13:45,837
it is possible that you can solve all
of intelligence in this way. Just for both

228
00:13:45,920 --> 00:13:48,237
language, just

229
00:13:48,320 --> 00:13:50,997
give a sentence and continue it or give

230
00:13:51,080 --> 00:13:56,077
a sentence and there's a gap in it,
some words blanked out and you fill in

231
00:13:56,160 --> 00:13:59,837
what words go there.
For vision, you give a sequence of

232
00:13:59,920 --> 00:14:04,157
images and predict what's going to happen
next or you fill in what happened

233
00:14:04,240 --> 00:14:07,277
in between.
Do you think it's possible that

234
00:14:07,360 --> 00:14:10,557
formulation alone

235
00:14:10,640 --> 00:14:14,957
as a signal for self-supervised learning
can solve intelligence for vision and

236
00:14:15,040 --> 00:14:16,797
language? >> I think
that's a bit short at the

237
00:14:16,880 --> 00:14:20,357
moment. Um so whether this will

238
00:14:20,440 --> 00:14:23,837
take us all the way to, you know,
human-level intelligence or something or

239
00:14:23,920 --> 00:14:28,357
just cat-level intelligence uh
is not clear, but among all the

240
00:14:28,440 --> 00:14:32,477
possible approaches that people have
proposed, I think it's a bit short. So

241
00:14:32,560 --> 00:14:35,157
I think this idea of

242
00:14:35,240 --> 00:14:37,077
an intelligent system filling in the

243
00:14:37,160 --> 00:14:39,837
blanks, either, you know, predicting the

244
00:14:39,920 --> 00:14:44,477
future, inferring the past,
filling in missing information,

245
00:14:44,560 --> 00:14:48,637
uh you know, I'm currently filling
the blank of what is behind your head and

246
00:14:48,720 --> 00:14:52,757
what you what your head looks like and,
you know, from from the back. Uh because

247
00:14:52,840 --> 00:14:55,677
I have, you know, basic knowledge
about how humans are made.

248
00:14:55,760 --> 00:14:58,237
And I don't know if you're going to, you
know, what where you're going to say, at

249
00:14:58,320 --> 00:14:59,997
which point you're going to speak,
whether you're going to move your head

250
00:15:00,080 --> 00:15:02,917
this way or that way, which way you're
going to look, but I know you're not

251
00:15:03,000 --> 00:15:06,157
going to just dematerialize
and reappear 3 m

252
00:15:06,240 --> 00:15:10,237
down the hall. Uh you know,
because I know what's possible and what's

253
00:15:10,320 --> 00:15:12,957
impossible uh according
to intuitive physics.

254
00:15:13,040 --> 00:15:15,877
>> But you have a model of what's possible
and impossible and then you'd be very

255
00:15:15,960 --> 00:15:19,757
surprised if it happens and then then
you'll have to reconstruct your model.

256
00:15:19,840 --> 00:15:23,277
>> Right. So that that's the model of the
world. It's what tells you you know what

257
00:15:23,360 --> 00:15:26,957
fills in the blanks. So given your
partial information about the state of

258
00:15:27,040 --> 00:15:31,997
the world given by your perception
your your model of the world fills in

259
00:15:32,080 --> 00:15:35,677
the missing information and that
includes predicting the future,

260
00:15:35,760 --> 00:15:38,837
retrodicting the past, you know,
and filling in things you

261
00:15:38,920 --> 00:15:42,597
don't immediately perceive.
>> And that doesn't have to be purely

262
00:15:42,680 --> 00:15:46,717
generic vision or visual information
or generic language. You can go to

263
00:15:46,800 --> 00:15:50,877
specifics like uh predicting

264
00:15:50,960 --> 00:15:53,757
what control decision you make when
you're driving in a lane. You have a

265
00:15:53,840 --> 00:15:57,557
sequence of images from a a vehicle

266
00:15:57,640 --> 00:16:00,477
and then you can you have information if

267
00:16:00,560 --> 00:16:04,917
you record it on video where the car
ended up going so you can go back in

268
00:16:05,000 --> 00:16:08,757
time and predict where the car went
based on the visual information. That's

269
00:16:08,840 --> 00:16:11,317
very specific, domain specific.

270
00:16:11,400 --> 00:16:12,917
>> Right, but the question is whether we

271
00:16:13,000 --> 00:16:15,357
can come up with sort of a generic

272
00:16:15,440 --> 00:16:18,597
uh method for you know, training

273
00:16:18,680 --> 00:16:20,437
machines to do this kind of prediction

274
00:16:20,520 --> 00:16:23,437
or filling in the blanks. So right now

275
00:16:23,520 --> 00:16:28,117
uh this type of approach has been
unbelievably successful in the context

276
00:16:28,200 --> 00:16:30,037
of natural language processing.

277
00:16:30,120 --> 00:16:34,197
Every modern natural language processing
is pre-trained in self-supervised manner

278
00:16:34,280 --> 00:16:37,677
to fill in the blanks. So you you show
it a sequence of words, you remove 10%

279
00:16:37,760 --> 00:16:40,677
of them and then you train some gigantic
neural net to predict the words that are

280
00:16:40,760 --> 00:16:44,157
missing. That and once you pre-train that

281
00:16:44,240 --> 00:16:46,277
network, you can use the

282
00:16:46,360 --> 00:16:48,677
internal representation learned by it as

283
00:16:48,760 --> 00:16:54,077
input to you know, something that you
train supervised or or whatever. Um

284
00:16:54,160 --> 00:16:57,877
that's been incredibly successful,
not so successful in images although it's

285
00:16:57,960 --> 00:17:02,597
making progress.
And uh and it's it's based on uh sort of

286
00:17:02,680 --> 00:17:06,797
manual data augmentation. Uh we can go
into this later, but what has not been

287
00:17:06,880 --> 00:17:09,077
successful yet is training from video.

288
00:17:09,160 --> 00:17:13,437
So getting a machine to learn
to represent the visual world for example

289
00:17:13,520 --> 00:17:16,757
by just watching video.
Nobody has really succeeded in doing this.

290
00:17:16,840 --> 00:17:19,757
>> Okay, well, let's kind of give
a high-level overview. What's the

291
00:17:19,840 --> 00:17:21,997
difference in

292
00:17:22,080 --> 00:17:24,917
kind and in difficulty between vision

293
00:17:25,000 --> 00:17:29,477
and language? So, you said
people haven't been able to

294
00:17:29,560 --> 00:17:33,397
really kind of crack the problem
of vision open in terms of self-supervised

295
00:17:33,480 --> 00:17:36,677
learning, but that may not be
necessarily because it's fundamentally

296
00:17:36,760 --> 00:17:38,957
more difficult. Maybe like when we're

297
00:17:39,040 --> 00:17:42,197
talking about achieving like passing

298
00:17:42,280 --> 00:17:46,037
the Turing test in the full spirit
of the Turing test in language might be

299
00:17:46,120 --> 00:17:49,917
harder than vision. That's That's not
obvious. So, what In your view, what

300
00:17:50,000 --> 00:17:55,117
which is harder or perhaps are
they just the same problem? When uh

301
00:17:55,200 --> 00:17:58,597
the farther we get to solving each, the
more we realize it's all the same thing.

302
00:17:58,680 --> 00:18:02,317
It's all the same cake. >> I think
what I'm looking for are methods that

303
00:18:02,400 --> 00:18:05,877
make make them look essentially
like the same cake, but

304
00:18:05,960 --> 00:18:10,637
currently they're not. And the main issue
with learning world models or learning

305
00:18:10,720 --> 00:18:12,077
predictive models

306
00:18:12,160 --> 00:18:15,917
is that uh the prediction is never a

307
00:18:16,000 --> 00:18:20,957
single thing cuz the world
is not entirely predictable.

308
00:18:21,040 --> 00:18:22,597
>> Yeah. >> It may be
deterministic or stochastic.

309
00:18:22,680 --> 00:18:25,997
We can get into the philosophical
discussion about it, but uh

310
00:18:26,080 --> 00:18:29,357
but even if it's deterministic,
it's not entirely predictable.

311
00:18:29,440 --> 00:18:34,597
And so if I play a short video
clip and then I ask you to

312
00:18:34,680 --> 00:18:37,357
predict what's going to happen next,
there's many, many plausible

313
00:18:37,440 --> 00:18:42,437
continuations for that uh video clip,
and the number of continuation grows

314
00:18:42,520 --> 00:18:45,477
with the interval of time
that you're asking the

315
00:18:45,560 --> 00:18:48,397
system to make a prediction for.

316
00:18:48,480 --> 00:18:52,757
And so, one big question with
self-supervised learning is how you

317
00:18:52,840 --> 00:18:57,077
represent this uncertainty, how you
represent multiple discrete outcomes,

318
00:18:57,160 --> 00:18:59,197
how you represent a sort of continuum of

319
00:18:59,280 --> 00:19:00,877
possible outcomes,

320
00:19:00,960 --> 00:19:04,757
um etc. And you know, if you are a sort

321
00:19:04,840 --> 00:19:08,357
of a classical machine learning person,
you say, "Oh, you just represent a

322
00:19:08,440 --> 00:19:13,477
distribution, right?"
And that we know how to do when we're

323
00:19:13,560 --> 00:19:17,037
predicting words, missing
words in the text because

324
00:19:17,120 --> 00:19:20,517
uh you can have a neural net give
a score for every word in the dictionary.

325
00:19:20,600 --> 00:19:24,357
It's a big, you know, it's a big list
of numbers, you know, maybe 100,000 or so.

326
00:19:24,440 --> 00:19:26,157
You can turn them into a a probability

327
00:19:26,240 --> 00:19:28,117
distribution that gives that tells you

328
00:19:28,200 --> 00:19:33,477
when I say a sentence, you know,
the you know, the cat is chasing the blank

329
00:19:33,560 --> 00:19:35,917
in the kitchen. You know,
there are only a few words

330
00:19:36,000 --> 00:19:39,957
that make sense there. You know,
it could be a mouse, it could be a lizard

331
00:19:40,040 --> 00:19:42,207
spot or you know,
something like that, right?

332
00:19:42,290 --> 00:19:46,157
>> [snorts] >> Uh and and if
I if I say the the blank is

333
00:19:46,240 --> 00:19:49,677
chasing the blank in the savanna,
you also have a bunch of plausible options

334
00:19:49,760 --> 00:19:53,797
for those two words, right?
Um that that that because you you have

335
00:19:53,880 --> 00:19:56,877
kind of a you know, underlying reality
that you can refer to to sort of fill in

336
00:19:56,960 --> 00:20:01,357
those those blanks. Um so

337
00:20:01,440 --> 00:20:05,837
you cannot say for sure in the savanna
if it's a you know, a lion or a cheetah

338
00:20:05,920 --> 00:20:08,237
or whatever. You cannot know if it's a

339
00:20:08,320 --> 00:20:11,437
zebra or gnu or you know, whatever.

340
00:20:11,520 --> 00:20:14,437
Wildebeest, same thing. Um

341
00:20:14,520 --> 00:20:17,237
but uh

342
00:20:17,320 --> 00:20:21,677
but you can represent the uncertainty
by just a a long list of numbers. Now,

343
00:20:21,760 --> 00:20:23,637
if I if I do the same thing with video

344
00:20:23,720 --> 00:20:26,197
and I ask you to predict a video clip,

345
00:20:26,280 --> 00:20:30,717
it's not a discrete set
of potential frames. You have to have

346
00:20:30,800 --> 00:20:34,037
somewhere representing a sort
of infinite number of plausible

347
00:20:34,120 --> 00:20:38,437
continuations of multiple frames in a
you know, high-dimensional continuous

348
00:20:38,520 --> 00:20:42,037
space. And we just have no
idea how to do this properly.

349
00:20:42,120 --> 00:20:44,797
>> Uh finite high-dimensional.

350
00:20:44,880 --> 00:20:47,197
So, like you cuz >> It's
finite high-dimensional, yes.

351
00:20:47,280 --> 00:20:50,597
>> Just like the words.
Like try to get it to uh

352
00:20:50,680 --> 00:20:56,157
down to a a small finite set of like
under a million, something like that.

353
00:20:56,240 --> 00:20:59,957
>> Something like that. >> I mean,
it's kind of ridiculous that we're

354
00:21:00,040 --> 00:21:04,757
doing a distribution over every single
possible word for language and it works.

355
00:21:04,840 --> 00:21:07,757
It It like that's
a really dumb way to do it.

356
00:21:07,840 --> 00:21:10,557
Um like there seems to there there seems

357
00:21:10,640 --> 00:21:14,997
to be like there should be some
more compressed representation of the

358
00:21:15,080 --> 00:21:18,037
distribution over the words.
>> You're right about that.

359
00:21:18,120 --> 00:21:21,117
>> And so >> I agree. >> do you
have any interesting ideas about

360
00:21:21,200 --> 00:21:24,637
how to represent all of reality
in a compressed way such you can form a

361
00:21:24,720 --> 00:21:26,797
distribution over it?
>> That's one of the big questions, you

362
00:21:26,880 --> 00:21:28,237
know, how do how do you do that? Right.

363
00:21:28,320 --> 00:21:32,557
I mean what's what's kind of you know,
another thing that that really is

364
00:21:32,640 --> 00:21:36,917
stupid about um I shouldn't say stupid
but like simplistic about current

365
00:21:37,000 --> 00:21:41,317
approaches to self-supervision
in in NLP in text

366
00:21:41,400 --> 00:21:45,957
is that not only do you represent
a joint distribution over words but for

367
00:21:46,040 --> 00:21:48,837
multiple words that are missing
those distributions are essentially

368
00:21:48,920 --> 00:21:51,397
independent of each other.

369
00:21:51,480 --> 00:21:53,677
Um and you know, you don't pay too much

370
00:21:53,760 --> 00:21:58,597
of a price for this. So you So you can't
So that you know, the system you know,

371
00:21:58,680 --> 00:22:01,517
in the the the sentence
that I gave earlier

372
00:22:01,600 --> 00:22:05,837
uh if it gives a a certain probability
for a lion and and cheetah and then a

373
00:22:05,920 --> 00:22:07,717
certain probability for

374
00:22:07,800 --> 00:22:12,917
uh you know, gazelle uh
wildebeest and and and zebra

375
00:22:13,000 --> 00:22:17,477
uh those two probabilities
are independent of each other.

376
00:22:17,560 --> 00:22:20,637
Uh and it's not the case that those
things are independent. Lions actually

377
00:22:20,720 --> 00:22:25,197
attack like bigger animals than than
cheetahs. So you know, there's a huge

378
00:22:25,280 --> 00:22:29,757
independence hypothesis in in this
process which is not actually true. The

379
00:22:29,840 --> 00:22:33,637
reason for this is that we
don't know how to represent uh

380
00:22:33,720 --> 00:22:38,077
properly distributions over
combinatorial uh sequences of symbols

381
00:22:38,160 --> 00:22:41,037
essentially when they're because
the number grows exponentially with the

382
00:22:41,120 --> 00:22:44,597
length of the of the symbols.
And so we have to use tricks for this

383
00:22:44,680 --> 00:22:49,717
but um those techniques can you know,
get around like don't even deal with it.

384
00:22:49,800 --> 00:22:52,837
So So the big question is like would there

385
00:22:52,920 --> 00:22:57,557
be some sort of abstract
latent representation of text

386
00:22:57,640 --> 00:22:59,837
that would say that you know, when I

387
00:22:59,920 --> 00:23:02,997
when I switch lion for gazelle uh lion

388
00:23:03,080 --> 00:23:07,437
for cheetah, I also have
to switch zebra for gazelle.

389
00:23:07,520 --> 00:23:10,877
>> Yeah, so this independence
assumption let me

390
00:23:10,960 --> 00:23:14,877
throw some criticism at you that I
often hear and see how you respond.

391
00:23:14,960 --> 00:23:18,957
So, this kind of filling in the blanks
is just statistics. You're not learning

392
00:23:19,040 --> 00:23:20,837
anything

393
00:23:20,920 --> 00:23:23,517
like the deep underlying concepts.

394
00:23:23,600 --> 00:23:27,077
You're just mimicking stuff from

395
00:23:27,160 --> 00:23:30,597
the past. You're not learning anything
new such that you can use it to

396
00:23:30,680 --> 00:23:33,357
generalize about the world.

397
00:23:33,440 --> 00:23:35,637
Uh or okay, let me just say the crude

398
00:23:35,720 --> 00:23:38,157
version, which is it's just statistics.

399
00:23:38,240 --> 00:23:42,197
It's not intelligence. Uh what do you
have to say to that? What do you usually

400
00:23:42,280 --> 00:23:44,557
say to that if you kind
of hear this kind of thing?

401
00:23:44,640 --> 00:23:47,197
>> I don't get into those discussions
because they are they're kind of

402
00:23:47,280 --> 00:23:50,637
pointless. Um so, first
of all, it's quite possible

403
00:23:50,720 --> 00:23:52,397
that intelligence is just statistics.

404
00:23:52,480 --> 00:23:54,637
It's just statistics of a particular kind.

405
00:23:54,720 --> 00:23:58,637
>> Yes. >> Uh where >> the philosophical
question. Is Is kind of is

406
00:23:58,720 --> 00:24:02,197
is intel- Is it possible that
intelligence is just statistics?

407
00:24:02,280 --> 00:24:06,797
>> Yeah. But, what kind
of statistics? So, uh if

408
00:24:06,880 --> 00:24:10,917
you are asking the question are
the model of the world the models of

409
00:24:11,000 --> 00:24:14,717
the world that we learn, um do they
have some notion of causality? Yes.

410
00:24:14,800 --> 00:24:17,317
>> Mhm. >> So,
if the criticism comes from people

411
00:24:17,400 --> 00:24:20,637
who say you know, uh current machine
learning system don't care about

412
00:24:20,720 --> 00:24:25,637
causality, which by the way is wrong.
Uh you know, I agree with them. Yeah, you

413
00:24:25,720 --> 00:24:29,877
should, you know, your model of the world
should have your actions as one of your

414
00:24:29,960 --> 00:24:32,557
the inputs. And that will
drive you to learn causal

415
00:24:32,640 --> 00:24:36,717
models of the world where you know what
you know, what uh intervention in the

416
00:24:36,800 --> 00:24:41,237
world will cause what result. Or you can
do this by observation of other agents

417
00:24:41,320 --> 00:24:44,437
uh acting in the world
and and observing the effect.

418
00:24:44,520 --> 00:24:48,197
Uh other humans, for example. So, I think

419
00:24:48,280 --> 00:24:52,797
you know, at some level of description,
uh intelligence is just statistics.

420
00:24:52,880 --> 00:24:57,077
Uh but that doesn't mean you don't you
don't you know, you won't have models

421
00:24:57,160 --> 00:25:02,037
that have, you know, deep mechanistic
explanation for what goes on. Uh, the

422
00:25:02,120 --> 00:25:03,597
question is how do you learn them?

423
00:25:03,680 --> 00:25:05,797
That's that's the question
I'm interested in.

424
00:25:05,880 --> 00:25:09,197
Uh, because you know,
a lot of people who actually

425
00:25:09,280 --> 00:25:13,277
voiced a criticism say that
those mechanistic model has to

426
00:25:13,360 --> 00:25:15,997
have to come from someplace else. They
have to come from human designers, they

427
00:25:16,080 --> 00:25:18,037
have to come from, I don't know what.

428
00:25:18,120 --> 00:25:22,597
And obviously, we learn them.
Uh, or if we don't learn them as an

429
00:25:22,680 --> 00:25:27,557
individual, nature learn them
for us using evolution. So,

430
00:25:27,640 --> 00:25:29,237
regardless of what you think, those

431
00:25:29,320 --> 00:25:32,077
processes have been learned somehow.

432
00:25:32,160 --> 00:25:34,877
>> So, if you look at the the human brain,

433
00:25:34,960 --> 00:25:39,477
just like when we humans introspect
about how the brain works, it seems like

434
00:25:39,560 --> 00:25:43,437
when we think about what
is intelligence, we think about the

435
00:25:43,520 --> 00:25:46,957
high-level stuff like the models we've
constructed, concepts like cognitive

436
00:25:47,040 --> 00:25:51,717
science, like concepts of memory
and reasoning module, almost like these

437
00:25:51,800 --> 00:25:55,317
high-level modules.
Is there's Is this serve as a good

438
00:25:55,400 --> 00:25:57,237
analogy?

439
00:25:57,320 --> 00:26:01,677
Like, are we ignoring the uh

440
00:26:01,760 --> 00:26:06,477
the dark matter, the the basic low-level
mechanisms, just like we ignore the way

441
00:26:06,560 --> 00:26:09,357
the operating system works,
we're just using the

442
00:26:09,440 --> 00:26:11,637
uh the the high-level software, we're

443
00:26:11,720 --> 00:26:16,637
ignoring that at the low level, the
neural network might be doing something

444
00:26:16,720 --> 00:26:20,957
like statistics. Like,
I mean, sorry to use this word

445
00:26:21,040 --> 00:26:24,557
probably incorrectly and crudely, but
doing this kind of filling the gap kind

446
00:26:24,640 --> 00:26:28,357
of learning and just kind of updating
the model constantly in order to be able

447
00:26:28,440 --> 00:26:32,637
to support the raw sensory information
mation to predict it and then adjust to

448
00:26:32,720 --> 00:26:36,197
the prediction when it's wrong.
But like, high level when we look at our

449
00:26:36,280 --> 00:26:40,237
brain at the high level, it feels like
we're doing like we're playing chess,

450
00:26:40,320 --> 00:26:44,837
like we're we're like playing with
high-level concepts and we're stitching

451
00:26:44,920 --> 00:26:48,797
them together and we're putting them
into long-term memory, but really what's

452
00:26:48,880 --> 00:26:52,837
going underneath is something we're not
able to introspect, which is this kind

453
00:26:52,920 --> 00:26:57,957
of uh uh simple large neural network
that's just filling in the gaps.

454
00:26:58,040 --> 00:27:01,757
>> Right. Well, okay, so there's a lot
of questions and a lot of answers there.

455
00:27:01,840 --> 00:27:04,597
Okay, so first of all, there's a whole
school of thought in neuroscience,

456
00:27:04,680 --> 00:27:06,837
computational neuroscience in particular,

457
00:27:06,920 --> 00:27:11,717
um that likes the idea of predictive
coding, which is really related to the

458
00:27:11,800 --> 00:27:14,437
idea I was talking about in
self-supervised learning. So, everything

459
00:27:14,520 --> 00:27:17,917
is about prediction. The essence
of intelligence is the ability to predict.

460
00:27:18,000 --> 00:27:21,037
>> Yeah. >> And everything
the brain does is trying to predict

461
00:27:21,120 --> 00:27:24,037
uh predict everything
from everything else.

462
00:27:24,120 --> 00:27:28,077
Okay, and that that's really sort of the
underlying principle, if you want, that

463
00:27:28,160 --> 00:27:32,077
uh self-supervised learning is trying
to kind of reproduce this idea of

464
00:27:32,160 --> 00:27:34,997
prediction as kind
of an essential mechanism

465
00:27:35,080 --> 00:27:38,237
of uh task-independent
learning, if you want.

466
00:27:38,320 --> 00:27:41,437
The next step is what kind
of intelligence are you

467
00:27:41,520 --> 00:27:45,437
interested in reproducing? And of course,
you know, we all think about,

468
00:27:45,520 --> 00:27:49,197
you know, trying to reproduce sort of,
you know, high-level cognitive processes

469
00:27:49,280 --> 00:27:53,477
in humans. But like with machines,
we're not even at the level of

470
00:27:53,560 --> 00:27:58,237
even reproducing the learning
processes in a in a cat brain.

471
00:27:58,320 --> 00:28:01,757
Um you know, the most intelligent of our
intelligent systems don't don't have as

472
00:28:01,840 --> 00:28:04,597
much common sense as as a house cat.

473
00:28:04,680 --> 00:28:07,437
>> [snorts] >> So, um how
is it that cats learn? And

474
00:28:07,520 --> 00:28:10,757
you know, cats don't do a whole lot of uh
reasoning. They certainly have causal

475
00:28:10,840 --> 00:28:14,917
models. They certainly have uh
because, you know, many cats can

476
00:28:15,000 --> 00:28:16,317
figure out like how they can act on the

477
00:28:16,400 --> 00:28:18,637
world to get what they want. Um they

478
00:28:18,720 --> 00:28:21,997
certainly have uh a fantastic model of

479
00:28:22,080 --> 00:28:25,757
intuitive physics, uh certainly
of the the the dynamics of

480
00:28:25,840 --> 00:28:29,757
their own bodies, but but also of prey
and things like that, right? So,

481
00:28:29,840 --> 00:28:31,797
um they they're they're pretty smart.

482
00:28:31,880 --> 00:28:35,037
They They only do this with
about 800 million neurons.

483
00:28:35,120 --> 00:28:38,117
Um we are not anywhere close to

484
00:28:38,200 --> 00:28:40,117
reproducing this kind of uh thing. So,

485
00:28:40,200 --> 00:28:43,237
to some extent, I I could I could say

486
00:28:43,320 --> 00:28:47,557
let's not even worry about
like the high-level cognition

487
00:28:47,640 --> 00:28:50,357
uh and kind of, you know, long-term
planning and reasoning that humans can

488
00:28:50,440 --> 00:28:53,357
do until we figure out like, you know,
can we even reproduce what Cassio is

489
00:28:53,440 --> 00:28:54,357
doing?

490
00:28:54,440 --> 00:28:57,917
Now, that said, this ability to learn

491
00:28:58,000 --> 00:29:01,237
world models, I think is the key to the

492
00:29:01,320 --> 00:29:05,677
possibility of learning machines that
can also reason. So, whenever I give a

493
00:29:05,760 --> 00:29:08,597
talk, I say there are there are three
challenges in the three main challenges

494
00:29:08,680 --> 00:29:11,917
in machine learning. The first one is uh
you know, getting machines to learn to

495
00:29:12,000 --> 00:29:15,997
represent the world um and I'm
proposing self-supervised learning.

496
00:29:16,080 --> 00:29:20,277
>> [snorts] >> Uh the second is uh
getting machines to reason in ways

497
00:29:20,360 --> 00:29:23,517
that are compatible with essentially
gradient-based learning

498
00:29:23,600 --> 00:29:26,437
because this is what deep
learning is all about, really.

499
00:29:26,520 --> 00:29:28,717
Um [snorts] and the third
one is something we have

500
00:29:28,800 --> 00:29:33,197
no idea how to solve, at least I
have no idea how to solve, is uh

501
00:29:33,280 --> 00:29:37,117
can we get machines to learn
hierarchical representations of action

502
00:29:37,200 --> 00:29:41,517
plans? You know, like, you know, we know
how to train them to learn hierarchical

503
00:29:41,600 --> 00:29:43,507
representations of perception.

504
00:29:43,590 --> 00:29:45,237
>> [snorts] >> You know,
we accomplish on that sensing

505
00:29:45,320 --> 00:29:49,357
like that and so on. But, what about
action plans? Can we get them to

506
00:29:49,440 --> 00:29:52,397
spontaneously learn good hierarchical
representations of actions?

507
00:29:52,480 --> 00:29:56,317
>> Also gradient-based. >> Yeah,
all of that has, you know, needs

508
00:29:56,400 --> 00:29:59,517
to be somewhat differentiable so
that you can apply sort of gradient-based

509
00:29:59,600 --> 00:30:03,147
learning, uh which is really
what deep learning is about.

510
00:30:03,230 --> 00:30:03,877
>> [snorts]

511
00:30:03,960 --> 00:30:07,837
>> So, it's background knowledge, ability

512
00:30:07,920 --> 00:30:11,317
to reason in a way uh

513
00:30:11,400 --> 00:30:15,717
that's differentiable that is somehow
connected deeply integrated

514
00:30:15,800 --> 00:30:19,637
with our background knowledge or builds
on top of that background knowledge and

515
00:30:19,720 --> 00:30:23,157
then given that background knowledge,
be able to make hierarchical plans.

516
00:30:23,240 --> 00:30:26,877
>> Right. >> So, if if
you take classical optimal

517
00:30:26,960 --> 00:30:30,637
control, there's something
in classical optimal control called uh

518
00:30:30,720 --> 00:30:34,277
model predictive control and it's,
you know, it's been around since the the

519
00:30:34,360 --> 00:30:37,677
early '60s. NASA uses
that to compute trajectories

520
00:30:37,760 --> 00:30:42,517
of rockets. And the basic idea is that you
You a pre- a predictive model >> [snorts]

521
00:30:42,600 --> 00:30:47,037
>> of the rocket, let's say, or whatever
system you are you intend to control,

522
00:30:47,120 --> 00:30:49,637
>> [snorts] >> which,
given the state of the system at

523
00:30:49,720 --> 00:30:54,237
time t and given an action that
you're taking on the system, so for

524
00:30:54,320 --> 00:30:58,157
a rocket it would be thrust and,
you know, all the controls you can have.

525
00:30:58,240 --> 00:31:01,637
Uh it gives you the state of the system
at time t plus delta t, right? So

526
00:31:01,720 --> 00:31:04,067
basically differential equation,
something like that.

527
00:31:04,150 --> 00:31:08,157
>> [snorts] >> Um and if you
have this model and you have

528
00:31:08,240 --> 00:31:12,997
this model in the form of some sort
of neural net or some sort of uh set of

529
00:31:13,080 --> 00:31:16,037
formula that you can back propagate
gradient through, you can do what's

530
00:31:16,120 --> 00:31:19,277
called model predictive
control or gradient based uh

531
00:31:19,360 --> 00:31:22,077
model predictive control. So, you have

532
00:31:22,160 --> 00:31:25,077
uh you can unroll that uh

533
00:31:25,160 --> 00:31:27,917
that model in time, you you

534
00:31:28,000 --> 00:31:30,157
you you feed it a

535
00:31:30,240 --> 00:31:33,037
hypothesized sequence of actions,

536
00:31:33,120 --> 00:31:37,197
and then you have some objective
function that measures how well at the

537
00:31:37,280 --> 00:31:40,397
end of the trajectory the system has uh
succeeded or matched what you wanted to

538
00:31:40,480 --> 00:31:44,517
do. Um you know, is the robot harmed?
Have you grasped the object you want to

539
00:31:44,600 --> 00:31:47,037
grasp? If it's a rocket, you know, are

540
00:31:47,120 --> 00:31:51,757
you at the right place near the space
station? Things like that. Um and by

541
00:31:51,840 --> 00:31:54,957
back propagation through time, and again
this was invented in the 1960s by

542
00:31:55,040 --> 00:31:58,837
optimal control theories,
uh you can figure out

543
00:31:58,920 --> 00:32:02,357
uh what is the optimal
sequence of actions that will

544
00:32:02,440 --> 00:32:07,077
uh you know, get my system
to the the best final state.

545
00:32:07,160 --> 00:32:11,277
So, that's a form of reasoning.
It's basically planning, and a lot of

546
00:32:11,360 --> 00:32:14,637
planning uh systems in robotics
are actually based on this.

547
00:32:14,720 --> 00:32:17,597
Uh and uh uh and you can
think of this as a form

548
00:32:17,680 --> 00:32:21,957
of reasoning. So, you know, to take
the example of the teenager driving a car

549
00:32:22,040 --> 00:32:25,397
again, you have a pretty good dynamical
model of the car. It doesn't need to be

550
00:32:25,480 --> 00:32:29,357
very accurate, but you know, again, that
if you turn the wheel to the right and

551
00:32:29,440 --> 00:32:31,757
there is a cliff, you're going to run
off the cliff, right? You don't need to

552
00:32:31,840 --> 00:32:34,077
have a very accurate
model to predict that.

553
00:32:34,160 --> 00:32:38,037
And you can run this in your mind
and decide not to do it for that reason.

554
00:32:38,120 --> 00:32:40,757
Because you can predict in advance that
the result is going to be bad. So, you

555
00:32:40,840 --> 00:32:45,637
can sort of imagine different scenarios,
and uh and then, you know, employ

556
00:32:45,720 --> 00:32:48,437
uh or take the first step in the
scenario that is most favorable, and

557
00:32:48,520 --> 00:32:49,757
then repeat the process of planning.

558
00:32:49,840 --> 00:32:53,037
That's called receding horizon model
predictive control. So, you know, you

559
00:32:53,120 --> 00:32:57,757
know, all those things have names, you
know, uh going back, you know, decades. Um

560
00:32:57,840 --> 00:33:02,317
and so, if you're not not the you know,
in classical optimal control, the model of

561
00:33:02,400 --> 00:33:04,717
the world is not generally learned.

562
00:33:04,800 --> 00:33:07,557
Uh this, you know, sometimes a few
parameters you have to identify. That's

563
00:33:07,640 --> 00:33:11,477
called systems identification.
But, uh but generally,

564
00:33:11,560 --> 00:33:14,877
the model is mostly
deterministic and mostly built by

565
00:33:14,960 --> 00:33:19,237
hand. So, the big question of AI,
uh I think the big challenge of AI for

566
00:33:19,320 --> 00:33:23,037
the next decade is how do we get
machines to learn predictive models of

567
00:33:23,120 --> 00:33:27,077
the world that deal with uncertainty,
and deal with the real world in all its

568
00:33:27,160 --> 00:33:31,317
complexity. So, it's not just trajectory
of a rocket, which you can reduce to

569
00:33:31,400 --> 00:33:35,117
first principles. It's not It's not even
just a trajectory of a robot arm, which

570
00:33:35,200 --> 00:33:38,517
again, you can model by, you know,
careful mathematics. But, it's

571
00:33:38,600 --> 00:33:42,797
everything else. Everything we observe
in the world, you know, people behavior,

572
00:33:42,880 --> 00:33:45,997
um you know, physical systems that involve

573
00:33:46,080 --> 00:33:50,957
collective phenomena,
like water, or or, you know,

574
00:33:51,040 --> 00:33:56,277
trees and you know, branches in a in a
tree or something, or or uh like complex

575
00:33:56,360 --> 00:33:58,637
things that, you know,
humans have no trouble

576
00:33:58,720 --> 00:34:02,237
developing abstract representations and
predictive model for, but we still don't

577
00:34:02,320 --> 00:34:05,757
know how to do with machines.
>> Where do you put in in these three,

578
00:34:05,840 --> 00:34:09,117
maybe in the in the planning stages,

579
00:34:09,200 --> 00:34:11,997
the the game theoretic nature of this

580
00:34:12,080 --> 00:34:16,437
world, where your actions not only
respond to the dynamic nature of the

581
00:34:16,520 --> 00:34:19,477
world, the environment,
but also affect it?

582
00:34:19,560 --> 00:34:24,477
So, if there's other humans involved, is
this is this point number four, or is it

583
00:34:24,560 --> 00:34:28,557
somehow integrated into the hierarchical
representation of action, in your view?

584
00:34:28,640 --> 00:34:31,997
>> I think it's integrated. It's just
um it's just that now your model of the

585
00:34:32,080 --> 00:34:35,037
world has to deal with, you know,
it just makes it more complicated, right?

586
00:34:35,120 --> 00:34:38,037
The fact that uh humans are
complicated and not easily

587
00:34:38,120 --> 00:34:41,997
predictable, that makes your model
of the world much more complicated. That

588
00:34:42,080 --> 00:34:45,437
much more complicated. >> Well,
there's a che- I mean, I I suppose

589
00:34:45,520 --> 00:34:50,077
chess is an analogy. So,
Monte Carlo tree search.

590
00:34:50,160 --> 00:34:53,437
>> Right. >> It There
is a I go, you go, I go, you

591
00:34:53,520 --> 00:34:56,117
go. Like um Andrej Karpathy recently

592
00:34:56,200 --> 00:34:59,877
gave a talk at MIT about car doors.

593
00:34:59,960 --> 00:35:03,797
I think there's some machine learning too,
but mostly car doors. And there's a

594
00:35:03,880 --> 00:35:07,677
dynamic nature to the car like
the person opening the door checking

595
00:35:07,760 --> 00:35:10,317
I mean, he wasn't talking about that. He
was talking about the perception problem

596
00:35:10,400 --> 00:35:13,517
of what the the anthology of what
defines a car door, this big

597
00:35:13,600 --> 00:35:17,757
philosophical question. But to me it
was interesting cuz like it's obvious that

598
00:35:17,840 --> 00:35:21,517
the person opening the car doors they're
trying to get out like here in New York

599
00:35:21,600 --> 00:35:25,117
trying to get out of the car.
You slowing down is going to signal

600
00:35:25,200 --> 00:35:28,397
something. You speeding up is going
to signal something. And that's a dance.

601
00:35:28,480 --> 00:35:33,277
It's a asynchronous chess game.
I don't know.

602
00:35:33,360 --> 00:35:37,677
>> [laughter] >> So, I It feels like um

603
00:35:37,760 --> 00:35:41,397
it's not just I mean, I guess you can
integrate all of that into one giant

604
00:35:41,480 --> 00:35:46,477
model. Like the entirety of the the
the these little interactions. Cuz it's

605
00:35:46,560 --> 00:35:49,677
not as complicated as chess. It's just
like a little dance. We do like a little

606
00:35:49,760 --> 00:35:51,917
dance together. And then we figure it out.

607
00:35:52,000 --> 00:35:55,757
>> Well, in some ways it's way
more complicated than chess because

608
00:35:55,840 --> 00:35:59,597
uh because it's continuous,
it's uncertain in a continuous manner.

609
00:35:59,680 --> 00:36:01,797
Uh >> It doesn't feel more complicated.

610
00:36:01,880 --> 00:36:04,997
>> But it doesn't feel more complicated
because that's what we are we've evolved

611
00:36:05,080 --> 00:36:07,757
to solve. This is the kind of problem
we've evolved to solve. And so we're

612
00:36:07,840 --> 00:36:12,437
good at it because, you know,
nature has made us good at it.

613
00:36:12,520 --> 00:36:16,037
Nature has not made us good at chess.
We completely suck at chess.

614
00:36:16,120 --> 00:36:19,437
>> Yeah. >> Um in fact,
that's why we designed it as a

615
00:36:19,520 --> 00:36:20,837
as a game is to be challenging.

616
00:36:20,920 --> 00:36:23,477
>> Mhm. >> [clears throat] >> Uh
but and if there is something that,

617
00:36:23,560 --> 00:36:28,477
you know, recent progress in the chess and
go has made us realize is that humans are

618
00:36:28,560 --> 00:36:32,197
really terrible at those things. Like
really bad, you know? There was a story,

619
00:36:32,280 --> 00:36:37,037
right, before AlphaGo that uh uh
you know, the best Go player thought

620
00:36:37,120 --> 00:36:40,477
they were maybe two or three stones
behind, you know, an ideal player that

621
00:36:40,560 --> 00:36:44,757
they would call God. Uh in fact,
no, they are like nine or 10

622
00:36:44,840 --> 00:36:47,277
stones behind. I mean, we're just bad.

623
00:36:47,360 --> 00:36:52,237
So, we're not good at and it's because
we have limited working memory. We know

624
00:36:52,320 --> 00:36:55,717
we're not very good at like doing
this tree exploration. You know, computers

625
00:36:55,800 --> 00:36:59,397
are much better um at doing
than we are. But we are much

626
00:36:59,480 --> 00:37:02,477
better at learning differentiable
models of the world.

627
00:37:02,560 --> 00:37:06,917
I mean, I say differentiable
in a kind of you know, I should say

628
00:37:07,000 --> 00:37:09,317
not differentiable in the sense that,
you know, we run backprop through it,

629
00:37:09,400 --> 00:37:14,077
but in the sense that uh our brain has
some mechanism for estimating gradients

630
00:37:14,160 --> 00:37:15,957
of of some kind. >> Yeah.

631
00:37:16,040 --> 00:37:17,917
>> And that's what, you know, makes us

632
00:37:18,000 --> 00:37:21,117
efficient. So, if you have an agent that

633
00:37:21,200 --> 00:37:25,557
consists of a a model
of the world, which, you know,

634
00:37:25,640 --> 00:37:28,557
in the human brain is basically
the entire front half of your brain.

635
00:37:28,640 --> 00:37:33,357
>> [snorts] >> Um
an objective function which

636
00:37:33,440 --> 00:37:37,477
uh in human in in humans is a combination
of two things. There is your

637
00:37:37,560 --> 00:37:40,877
sort of intrinsic motivation module,
which is in the basal ganglia, you know,

638
00:37:40,960 --> 00:37:44,517
at the base of your brain. That's the
thing that measures pain and hunger and

639
00:37:44,600 --> 00:37:47,357
things like that, like immediate um

640
00:37:47,440 --> 00:37:51,917
uh feelings and emotions.
Um and then there is

641
00:37:52,000 --> 00:37:54,437
you know, the equivalent of what people
in reinforcement learning call a critic,

642
00:37:54,520 --> 00:37:57,517
which is a sort of module that predicts

643
00:37:57,600 --> 00:38:01,557
ahead what the outcome of a

644
00:38:01,640 --> 00:38:05,677
of a situation will be. And so,
it's it's not a cost function,

645
00:38:05,760 --> 00:38:09,557
but it's sort of not an objective
function, but it's sort of a

646
00:38:09,640 --> 00:38:13,117
you know, trained predictor of the
ultimate objective function. And that

647
00:38:13,200 --> 00:38:16,917
also is differentiable. And so, if all
of this is differentiable, your cost

648
00:38:17,000 --> 00:38:19,517
function, your your critic, your

649
00:38:19,600 --> 00:38:22,317
uh Uh, you know, your your world model.

650
00:38:22,400 --> 00:38:26,877
Then you can use gradient-based type
methods to do planning, to do reasoning,

651
00:38:26,960 --> 00:38:32,677
to do learning, you know, to do all the
things that we'd like an intelligent agent

652
00:38:32,760 --> 00:38:36,117
uh, to do.
>> And the gradient-based learning

653
00:38:36,200 --> 00:38:39,677
like what's your intuition? That's
probably at the core of what can solve

654
00:38:39,760 --> 00:38:41,997
intelligence. So, you don't need

655
00:38:42,080 --> 00:38:47,517
like um logic-based
reasoning uh, in your view.

656
00:38:47,600 --> 00:38:50,717
>> I don't know how to make
logic-based reasoning compatible with

657
00:38:50,800 --> 00:38:52,077
uh, efficient learning. >> Yeah.

658
00:38:52,160 --> 00:38:55,197
>> And Okay. I mean, there is a big
question, perhaps a philosophical

659
00:38:55,280 --> 00:38:58,117
question. I mean,
it's not that philosophical, but

660
00:38:58,200 --> 00:39:01,997
uh, that we can ask is is that,
you know, all the learning algorithms we

661
00:39:02,080 --> 00:39:07,037
know from engineering and computer science
proceed by optimizing some objective

662
00:39:07,120 --> 00:39:12,837
function. >> Yeah. >> Right? Uh,
so one question we may ask is uh,

663
00:39:12,920 --> 00:39:16,637
is does learning in the brain
minimize an objective function?

664
00:39:16,720 --> 00:39:19,637
I mean, it could be a,
you know, a composite of multiple

665
00:39:19,720 --> 00:39:22,877
objective functions, but it's
still an objective function.

666
00:39:22,960 --> 00:39:25,437
Uh, second, if it does optimize an

667
00:39:25,520 --> 00:39:30,157
objective function, does it do does
it do it by uh, some sort of gradient

668
00:39:30,240 --> 00:39:33,837
estimation? You know, it doesn't need
to be backprop, but, you know, some way of

669
00:39:33,920 --> 00:39:36,757
estimating the gradient
in efficient manner.

670
00:39:36,840 --> 00:39:41,117
Whose complexity is on the same order
of magnitude as you know, actually running

671
00:39:41,200 --> 00:39:44,717
the inference. >> [snorts] >> Uh,

672
00:39:44,800 --> 00:39:48,197
cuz you can't afford to do things like,
you know, perturbing a weight in your

673
00:39:48,280 --> 00:39:50,037
brain to figure out what the effect is.

674
00:39:50,120 --> 00:39:52,637
And then sort of uh, you know,
you you can't do sort of

675
00:39:52,720 --> 00:39:54,597
estimating gradient by perturbation.

676
00:39:54,680 --> 00:39:57,317
It's To me, it seems very implausible

677
00:39:57,400 --> 00:40:00,797
that the brain uses some sort of uh,

678
00:40:00,880 --> 00:40:04,917
you know, zeroth-order black
box gradient-free optimization

679
00:40:05,000 --> 00:40:08,717
because it's so much less efficient
than gradient optimization. So, it has to

680
00:40:08,800 --> 00:40:11,157
have a way of estimating gradients.

681
00:40:11,240 --> 00:40:13,077
>> Is it possible that some kind of

682
00:40:13,160 --> 00:40:16,397
logic-based reasoning emerges in pockets

683
00:40:16,480 --> 00:40:20,477
as a useful, like you said, if the brain
has an objective function, maybe it's a

684
00:40:20,560 --> 00:40:23,277
mechanism for creating
objective functions.

685
00:40:23,360 --> 00:40:25,917
It's It's It's a mechanism for um

686
00:40:26,000 --> 00:40:30,717
creating knowledge bases, for example,
that can then be queried. Like maybe

687
00:40:30,800 --> 00:40:33,517
it's like an efficient representation
of knowledge that's learned in a

688
00:40:33,600 --> 00:40:35,717
gradient-based way or something like that.

689
00:40:35,800 --> 00:40:39,277
>> Well, so I think there is a lot
of different types of uh intelligence. So,

690
00:40:39,360 --> 00:40:43,157
first of all, I think the type of logical
reasoning that we think about um

691
00:40:43,240 --> 00:40:46,997
that we are >> [snorts] >> you know,
maybe stemming from you know,

692
00:40:47,080 --> 00:40:49,997
sort of classical AI
of the 1970s and '80s.

693
00:40:50,080 --> 00:40:54,917
Um I think humans use
that relatively rarely

694
00:40:55,000 --> 00:40:56,637
and are not particularly good at it.

695
00:40:56,720 --> 00:40:58,877
>> But we judge each other based on our

696
00:40:58,960 --> 00:41:02,437
ability to uh solve those rare problems.

697
00:41:02,520 --> 00:41:07,117
It's called an IQ test. >> so.
Like I'm I'm not very good at chess.

698
00:41:07,200 --> 00:41:10,437
>> Yes, I'm I'm judging you
this whole time because of it.

699
00:41:10,520 --> 00:41:13,317
Well, we we actually >> With
With your with your uh you know,

700
00:41:13,400 --> 00:41:15,187
heritage, I'm sure you're good at chess.

701
00:41:15,270 --> 00:41:18,117
>> [laughter] >> No, stereotypes.
Not all stereotypes are

702
00:41:18,200 --> 00:41:19,917
true.

703
00:41:20,000 --> 00:41:22,317
>> Well, I'm terrible at chess. So, um

704
00:41:22,400 --> 00:41:28,077
you know, but I think perhaps uh another
type of intelligence that I have is this

705
00:41:28,160 --> 00:41:32,597
uh uh you know, ability of sort
of building models of the world from

706
00:41:32,680 --> 00:41:37,557
uh you know, reasoning obvious
obviously, but also also data.

707
00:41:37,640 --> 00:41:41,117
Uh and those those models generally are
more kind of analogical, right? So, it's

708
00:41:41,200 --> 00:41:44,237
it's uh it's reasoning by simulation

709
00:41:44,320 --> 00:41:48,957
and but by analogy, where you use one
model to apply to a new situation. Even

710
00:41:49,040 --> 00:41:52,997
though you've never seen that situation,
you can sort of um connect it to a

711
00:41:53,080 --> 00:41:57,957
situation you've encountered before.
Uh and and your reasoning is more,

712
00:41:58,040 --> 00:42:01,877
you know, akin to some sort of internal
simulation. So, you you are kind of

713
00:42:01,960 --> 00:42:04,677
simulating what's happening when you're
building, I don't know, a box out of

714
00:42:04,760 --> 00:42:08,317
wood or something, right?
You kind of imagine in in advance

715
00:42:08,400 --> 00:42:10,437
like what would be the result of,
you know, cutting the wood in this

716
00:42:10,520 --> 00:42:14,037
particular way. Are you going to use,
you know, screws or nails or whatever.

717
00:42:14,120 --> 00:42:17,117
Um when you are interacting with someone,
you also have a model of that

718
00:42:17,200 --> 00:42:21,237
person and and sort
of interact with that person,

719
00:42:21,320 --> 00:42:24,917
you know, having this model
in mind uh to kind of uh

720
00:42:25,000 --> 00:42:29,277
tell the person what what you think
is useful to them. So, I think this

721
00:42:29,360 --> 00:42:34,157
this ability to construct models
of the world is basically the essence the

722
00:42:34,240 --> 00:42:35,797
essence of intelligence.

723
00:42:35,880 --> 00:42:37,917
And the ability to use it then to

724
00:42:38,000 --> 00:42:41,357
uh plan uh actions that will

725
00:42:41,440 --> 00:42:44,957
uh fulfill a particular uh criterion um

726
00:42:45,040 --> 00:42:47,357
of course is is is necessary as well.

727
00:42:47,440 --> 00:42:51,037
>> So, I'm going to ask you a series
of impossible questions as we keep ask as

728
00:42:51,120 --> 00:42:55,397
I've been doing. So, so if that's
the fundamental sort of dark matter of

729
00:42:55,480 --> 00:42:59,517
intelligence, this ability to form
a background model, what's your intuition

730
00:42:59,600 --> 00:43:03,397
about how much knowledge is required?

731
00:43:03,480 --> 00:43:07,197
You you know, I think dark
matter you you put a percentage of

732
00:43:07,280 --> 00:43:12,077
uh on it of of uh the composition
of the universe and

733
00:43:12,160 --> 00:43:13,637
how much of it is dark matter, how much

734
00:43:13,720 --> 00:43:17,117
of it is dark energy. How much

735
00:43:17,200 --> 00:43:21,877
uh information do you think
is required to to to be a house cat?

736
00:43:21,960 --> 00:43:25,397
So, you have to be able to uh when
you see a box, go in it. When you see a

737
00:43:25,480 --> 00:43:29,957
human, compute the most evil action. If
there's a thing that's near an edge, you

738
00:43:30,040 --> 00:43:34,277
knock it off.
All of that plus the extra stuff you

739
00:43:34,360 --> 00:43:38,317
mentioned, which is a uh great
self-awareness of the physics of your of

740
00:43:38,400 --> 00:43:40,677
your own body and the and the world.

741
00:43:40,760 --> 00:43:43,997
How much knowledge is required
do you think to solve it?

742
00:43:44,080 --> 00:43:47,597
>> Um >> I don't even know how
to measure an answer to that question.

743
00:43:47,680 --> 00:43:51,077
>> I'm not sure how to measure it,
but whatever it is, it fits in about

744
00:43:51,160 --> 00:43:54,317
about 800,000 neurons. Uh

745
00:43:54,400 --> 00:43:58,157
800 million neurons, sorry.
>> The representation does.

746
00:43:58,240 --> 00:44:00,917
>> Everything.
All knowledge, everything, right?

747
00:44:01,000 --> 00:44:05,597
Um you know, it's less than a billion.
A dog is 2 billion, but a cat is less than

748
00:44:05,680 --> 00:44:10,277
1 billion. >> Right.
And so multiply that by 1,000 and you get

749
00:44:10,360 --> 00:44:14,237
the number of synapses.
And I think almost all of it is is

750
00:44:14,320 --> 00:44:18,597
learned through this you know a sort
of self-supervised learning although you

751
00:44:18,680 --> 00:44:21,877
know I I think a tiny sliver is learned
through reinforcement learning and

752
00:44:21,960 --> 00:44:25,157
certainly very little through you
know classical supervised learning

753
00:44:25,240 --> 00:44:28,557
although it's not even clear how
supervised learning actually works in

754
00:44:28,640 --> 00:44:33,277
in biological world.
So I think almost all of it is

755
00:44:33,360 --> 00:44:37,557
is self-supervised learning.
>> It's driven by

756
00:44:37,640 --> 00:44:42,477
the the sort of ingrained objective
functions that a cat or human have at

757
00:44:42,560 --> 00:44:45,357
the base of their brain
which kind of drives their

758
00:44:45,440 --> 00:44:49,677
their behavior.
So you know nature tells us

759
00:44:49,760 --> 00:44:53,117
you're hungry. It doesn't
tell us how to feed a feed

760
00:44:53,200 --> 00:44:56,237
ourselves. That's that's something that
the rest of our brain has to figure out

761
00:44:56,320 --> 00:44:58,837
right? >> Well it's
interesting cuz there might be

762
00:44:58,920 --> 00:45:01,517
more like deeper objective functions

763
00:45:01,600 --> 00:45:06,117
that are lying the whole thing.
So hun- hunger may be some kind of

764
00:45:06,200 --> 00:45:10,277
Now you go to like neurobiology
might be just the brain

765
00:45:10,360 --> 00:45:14,397
um trying to maintain homeostasis.

766
00:45:14,480 --> 00:45:18,877
So hunger is just one
of the human perceivable

767
00:45:18,960 --> 00:45:23,437
symptoms of the brain being unhappy
with the way things are currently. So it

768
00:45:23,520 --> 00:45:26,877
could be just like one really dumb
objective function at the core.

769
00:45:26,960 --> 00:45:30,357
>> But that's how that's how
a behavior is is is driven.

770
00:45:30,440 --> 00:45:33,957
The fact that you know the orbital ganglia

771
00:45:34,040 --> 00:45:38,397
drive us to do things that are that are
different from say an orangutan or

772
00:45:38,480 --> 00:45:41,917
certainly a cat is what
makes you know human nature

773
00:45:42,000 --> 00:45:44,877
versus orangutan nature versus cat nature.

774
00:45:44,960 --> 00:45:46,757
So for example

775
00:45:46,840 --> 00:45:50,437
you know our basal ganglia drives us to

776
00:45:50,520 --> 00:45:54,157
seek the company of other humans.

777
00:45:54,240 --> 00:45:58,037
And that's because nature has figured
out that we need to be social animals

778
00:45:58,120 --> 00:46:03,277
for our species to survive
and it's true of of uh uh

779
00:46:03,360 --> 00:46:07,557
It's not true of orangutans. Orangutans
are solitary animals. Um they don't seek

780
00:46:07,640 --> 00:46:10,677
the company of others.
In fact, they avoid them.

781
00:46:10,760 --> 00:46:13,317
Uh in fact, they scream at them when
they come too close because they're

782
00:46:13,400 --> 00:46:17,477
territorial.
Uh cuz for for their survival, you know,

783
00:46:17,560 --> 00:46:20,197
uh evolution has figured
out that's the best thing.

784
00:46:20,280 --> 00:46:24,157
I mean, they're occasionally social,
of course, for, you know, um

785
00:46:24,240 --> 00:46:28,757
reproduction and stuff like that. But um
but but they're mostly solitary. So,

786
00:46:28,840 --> 00:46:31,837
uh so all of those behaviors are not
part of intelligence. You know, people

787
00:46:31,920 --> 00:46:34,197
say, "Oh, you're never going to have
intelligent machines because, you know,

788
00:46:34,280 --> 00:46:37,757
human intelligence is social." But then
you look at orangutans, you look at

789
00:46:37,840 --> 00:46:40,717
octopus. Octopus never know their parents.

790
00:46:40,800 --> 00:46:44,517
They barely interact with any other. And
they And they get to be really smart in

791
00:46:44,600 --> 00:46:47,357
less than a less than a year,
in like half a year.

792
00:46:47,440 --> 00:46:49,517
Um you know, in a year or they're adults.

793
00:46:49,600 --> 00:46:52,717
In 2 years, they're dead. So, um

794
00:46:52,800 --> 00:46:56,677
there are things that we think
as humans are intimately linked with

795
00:46:56,760 --> 00:46:59,677
intelligence, like social interaction,

796
00:46:59,760 --> 00:47:01,677
like language.

797
00:47:01,760 --> 00:47:06,757
We think I think we give way too much
importance to language as a substrate of

798
00:47:06,840 --> 00:47:10,357
intelligence as humans.
Because we think our reasoning is so

799
00:47:10,440 --> 00:47:14,197
linked with language. >> So,
for for to solve the house cat

800
00:47:14,280 --> 00:47:18,317
intelligence problem, you think you could
do it on a desert island. You could have

801
00:47:18,400 --> 00:47:22,677
>> Pretty much. >> You could
just have a cat sitting there um

802
00:47:22,760 --> 00:47:27,637
looking at the waves at the ocean
waves and figure a lot of it out.

803
00:47:27,720 --> 00:47:33,157
>> It needs to have sort of, you know,
the right set of drives uh to kind of,

804
00:47:33,240 --> 00:47:37,637
you know, get it to do the thing and learn
the appropriate things, right? But uh

805
00:47:37,720 --> 00:47:42,557
um like for example, you know,
uh baby humans are are driven to

806
00:47:42,640 --> 00:47:46,517
learn to stand up and walk. Okay,
you know, it's not that's kind of

807
00:47:46,600 --> 00:47:50,037
this desire is hardwired. How to do
it precisely is not. That's learned.

808
00:47:50,120 --> 00:47:52,357
>> Mhm. >> But the desire to >> To walk?

809
00:47:52,440 --> 00:47:56,837
>> move around and stand up.
Um that's sort of

810
00:47:56,920 --> 00:47:58,957
hardwired. It's very simple to hardwire

811
00:47:59,040 --> 00:48:03,877
this kind of stuff. >> What Oh,
like the desire to Well,

812
00:48:03,960 --> 00:48:07,557
that's interesting.
You're hardwired to want to walk.

813
00:48:07,640 --> 00:48:12,397
That's not a There's got to be
a deeper need for walking.

814
00:48:12,480 --> 00:48:15,557
I I think it was probably socially
imposed by society that you need to

815
00:48:15,640 --> 00:48:19,797
walk. All the other bipedal >> There
are a lot of simple animals that,

816
00:48:19,880 --> 00:48:24,117
you know, would probably
walk without ever watching any

817
00:48:24,200 --> 00:48:27,877
other members of the species.
>> It seems like a scary thing to have to

818
00:48:27,960 --> 00:48:30,397
do cuz you suck at bipedal walking at

819
00:48:30,480 --> 00:48:33,877
first. It seems crawling is much safer.

820
00:48:33,960 --> 00:48:37,317
Much more like why are you in a hurry?

821
00:48:37,400 --> 00:48:39,197
>> [laughter] >> Well,
because because you have this

822
00:48:39,280 --> 00:48:42,277
thing that drives you
to do it, you know? Um

823
00:48:42,360 --> 00:48:46,997
which is sort of part of uh
the sort of human development.

824
00:48:47,080 --> 00:48:49,717
>> Is that understood actually
what >> Not entirely. No.

825
00:48:49,800 --> 00:48:53,037
>> What is the What's the reason to get
on two feet? It's really hard. Like most

826
00:48:53,120 --> 00:48:54,717
animals don't get on two feet.

827
00:48:54,800 --> 00:48:57,717
>> They get on four feet. You know,
many mammals get on four feet.

828
00:48:57,800 --> 00:49:00,477
>> Yeah, they feel >> Very quickly.
Some of them extremely quickly.

829
00:49:00,560 --> 00:49:04,517
>> But I don't you know, like from the
last time I've interacted with a table,

830
00:49:04,600 --> 00:49:08,397
that's much more stable than a thing on
two legs. It's just a really hard problem.

831
00:49:08,480 --> 00:49:11,317
>> Yeah, I mean birds have
figured it out with two feet.

832
00:49:11,400 --> 00:49:15,477
>> It Well, technically we can go into
anthology. They have four. I guess they

833
00:49:15,560 --> 00:49:17,077
have two feet. >> They have two feet.

834
00:49:17,160 --> 00:49:20,757
>> Chickens. >> You know,
dinosaurs had two feet. Many of them.

835
00:49:20,840 --> 00:49:25,237
>> Allegedly. >> [laughter] >> I'm
just now learning that T-Rex was

836
00:49:25,320 --> 00:49:29,077
eating grass, not other animals.
T-Rex might have been a friendly

837
00:49:29,160 --> 00:49:32,157
friendly pet. What do you think about uh

838
00:49:32,240 --> 00:49:36,677
I don't know if you looked at the
test for general intelligence that

839
00:49:36,760 --> 00:49:40,117
François Chollet put together. I don't
know if you got a chance to look at that

840
00:49:40,200 --> 00:49:43,717
kind of thing. Like what What's
your intuition about how to solve

841
00:49:43,800 --> 00:49:47,237
like an IQ type of test? >> I don't know.
I think it's so outside of

842
00:49:47,320 --> 00:49:50,157
my radar screen that it's not really

843
00:49:50,240 --> 00:49:52,517
relevant, I think, in the short term.

844
00:49:52,600 --> 00:49:55,837
>> Well, I guess one
way to ask another way

845
00:49:55,920 --> 00:49:58,517
perhaps more closer to what

846
00:49:58,600 --> 00:50:00,637
to your work is like how do you solve

847
00:50:00,720 --> 00:50:04,597
MNIST uh with very little example data?

848
00:50:04,680 --> 00:50:07,637
>> That's right. And that's the answer
to this probably is self-supervised

849
00:50:07,720 --> 00:50:11,717
learning. Just learn to represent
images and then learning uh you know, to

850
00:50:11,800 --> 00:50:15,477
recognize handwritten digits on top
of this will only require a few samples.

851
00:50:15,560 --> 00:50:17,397
And we observe this in in humans, right?

852
00:50:17,480 --> 00:50:22,157
You you show a a young child a a picture
book with a couple picture of an

853
00:50:22,240 --> 00:50:26,037
elephant and that's it. The The child
knows what an elephant is. And

854
00:50:26,120 --> 00:50:29,977
we we we see this today with practical
systems that we you know, we train

855
00:50:30,060 --> 00:50:32,397
[snorts] image recognition systems with

856
00:50:32,480 --> 00:50:36,157
uh enormous amounts of of images either

857
00:50:36,240 --> 00:50:40,237
either completely self-supervised or very
weakly supervised. For example, you

858
00:50:40,320 --> 00:50:44,957
can uh train a neural net to predict
uh whatever hashtag people type on

859
00:50:45,040 --> 00:50:47,917
Instagram, right? And you can do
this with billions of images cuz there's

860
00:50:48,000 --> 00:50:52,357
billions per day that are showing up.
Um so, the the amount of training data

861
00:50:52,440 --> 00:50:55,517
there is essentially unlimited.
And [snorts] then you take the output

862
00:50:55,600 --> 00:50:59,997
representation, you know, a couple
layers down from the output uh of what

863
00:51:00,080 --> 00:51:04,997
the system learned and feed this as input
to a classifier for any object in

864
00:51:05,080 --> 00:51:07,637
the world that you want and it works
pretty well. So, that's transfer

865
00:51:07,720 --> 00:51:09,637
learning, okay? [clears throat] Uh or

866
00:51:09,720 --> 00:51:12,197
weakly supervised uh transfer learning.

867
00:51:12,280 --> 00:51:17,037
Uh people are making very very fast
progress using self-supervised learning

868
00:51:17,120 --> 00:51:19,757
uh for for the with this
kind of scenario as well.

869
00:51:19,840 --> 00:51:22,557
>> [snorts] >> Um and you know,
my guess is that that's

870
00:51:22,640 --> 00:51:24,317
that's going to be the the future.

871
00:51:24,400 --> 00:51:26,757
>> For self-supervised learning, how much

872
00:51:26,840 --> 00:51:29,597
cleaning do you think is needed for

873
00:51:29,680 --> 00:51:34,637
filtering um uh malicious
signal or what's a better

874
00:51:34,720 --> 00:51:38,557
term? But like a lot of people
use hashtags on Instagram

875
00:51:38,640 --> 00:51:41,957
to uh get like good SEO >> Mhm.

876
00:51:42,040 --> 00:51:44,917
>> that doesn't fully represent
the contents of the image.

877
00:51:45,000 --> 00:51:49,437
Like they'll put a picture of a cat
and hashtag it with like science, awesome,

878
00:51:49,520 --> 00:51:54,357
fun, I don't know. All kind of Why would
you put science? That's not very good SEO.

879
00:51:54,440 --> 00:51:58,157
>> The way the way my my colleagues
who worked on this project at

880
00:51:58,240 --> 00:52:02,477
at Facebook now Meta
Meta AI a few years ago

881
00:52:02,560 --> 00:52:05,757
dealt with this is that they only
selected something like 17,000 tags that

882
00:52:05,840 --> 00:52:08,637
correspond to kind of physical things or

883
00:52:08,720 --> 00:52:12,917
or situations. Like you know,
that has some visual content.

884
00:52:13,000 --> 00:52:17,797
Um so you know you wouldn't
have like hash TBT or

885
00:52:17,880 --> 00:52:21,757
anything like that. >> Oh,
so they they keep a very select set

886
00:52:21,840 --> 00:52:23,197
of hashtags is what you're saying.

887
00:52:23,280 --> 00:52:26,557
>> Yeah. >> Okay. >> But it's
still it's still on the order of

888
00:52:26,640 --> 00:52:29,757
you know, 10 to 20,000.
So it's fairly large.

889
00:52:29,840 --> 00:52:33,197
>> Okay. Can you tell me
about data augmentation?

890
00:52:33,280 --> 00:52:37,717
What the heck is data augmentation
and how is it used maybe contrastive

891
00:52:37,800 --> 00:52:42,717
learning for for video?
What are some cool ideas here?

892
00:52:42,800 --> 00:52:45,917
>> Right. So data augmentation, I mean
first data augmentation, you know, is

893
00:52:46,000 --> 00:52:49,637
the idea of artificially increasing the
size of your training set by distorting

894
00:52:49,720 --> 00:52:52,677
the images that you
have in ways that don't

895
00:52:52,760 --> 00:52:54,237
change the nature of the image, right?

896
00:52:54,320 --> 00:52:57,797
So you take you do MNIST, you can do data
augmentation on MNIST. And people have

897
00:52:57,880 --> 00:53:00,917
done this since 1990s, right? You take a

898
00:53:01,000 --> 00:53:05,437
MNIST digit and you shift it
a little bit or you change the size or

899
00:53:05,520 --> 00:53:08,877
rotate it, skew it you know, etc.

900
00:53:08,960 --> 00:53:12,717
>> Add noise. >> Add noise,
etc. And it it works better.

901
00:53:12,800 --> 00:53:15,837
If you train a supervised classifier
with augmented data, you're going to get

902
00:53:15,920 --> 00:53:17,437
better results.

903
00:53:17,520 --> 00:53:20,717
Now, it's become really interesting over

904
00:53:20,800 --> 00:53:23,717
the last couple years because

905
00:53:23,800 --> 00:53:25,597
a lot of self-supervised learning

906
00:53:25,680 --> 00:53:28,037
techniques to pre-train vision systems

907
00:53:28,120 --> 00:53:31,517
are based on data augmentation. And the

908
00:53:31,600 --> 00:53:35,237
the basic techniques
is originally inspired by

909
00:53:35,320 --> 00:53:38,917
techniques that I worked on in the early
'90s and Jeff Hinton worked on also in

910
00:53:39,000 --> 00:53:42,757
the early '90s. They were sort
of parallel work. I used to call this

911
00:53:42,840 --> 00:53:47,237
Siamese network. So, basically you
take uh two identical copies of the same

912
00:53:47,320 --> 00:53:49,717
network. They share the same weights.

913
00:53:49,800 --> 00:53:54,517
And you show two uh different views
of the same object. Either those two

914
00:53:54,600 --> 00:53:57,917
different views may have been obtained
by data augmentation, or maybe it's two

915
00:53:58,000 --> 00:54:01,837
different views of the same scene from
a camera that you moved, or at different

916
00:54:01,920 --> 00:54:05,397
times, or something like that, right?
Or two pictures of the same person, things

917
00:54:05,480 --> 00:54:09,997
like that. And then you train this neural
net, those two identical copies of this

918
00:54:10,080 --> 00:54:13,957
neural net, to produce an output
representation, a vector,

919
00:54:14,040 --> 00:54:17,357
uh in such a way that the
representation for those two

920
00:54:17,440 --> 00:54:21,997
images are as close to each other
as possible, as identical to each other as

921
00:54:22,080 --> 00:54:26,597
possible, right? Because you want the
system to be basically learn a function

922
00:54:26,680 --> 00:54:29,837
that will that will be invariant, that
will not change, whose output will not

923
00:54:29,920 --> 00:54:34,397
change when you transform those inputs
uh in in those in those particular ways,

924
00:54:34,480 --> 00:54:37,637
right? >> [snorts] >> Um so,
that's easy to do.

925
00:54:37,720 --> 00:54:40,917
What's complicated is how do you make
sure that when you show two images that

926
00:54:41,000 --> 00:54:42,917
are different, the system
will produce different things?

927
00:54:43,000 --> 00:54:47,917
>> Mhm. >> Because if you don't
have a specific provision for this,

928
00:54:48,000 --> 00:54:51,437
uh the system will just ignore the inputs
uh when you train it. It will end

929
00:54:51,520 --> 00:54:54,477
up ignoring the input and just produce
a constant vector that is the same for

930
00:54:54,560 --> 00:54:56,597
every input, right? >> Yes.
>> That's called a collapse.

931
00:54:56,680 --> 00:55:00,357
>> Uh-huh. >> Now, how do you avoid
collapse? So, there's two ideas.

932
00:55:00,440 --> 00:55:04,877
Uh one idea that I proposed in the early
'90s uh with my colleagues at Bell Labs,

933
00:55:04,960 --> 00:55:09,837
uh Jan Bromley and a couple other people,
uh which we now call contrastive

934
00:55:09,920 --> 00:55:13,397
learning, which is to have negative
examples, right? So, you have pairs

935
00:55:13,480 --> 00:55:16,397
of images that you know are different.

936
00:55:16,480 --> 00:55:20,597
And you show them to the network, and uh
those two copies, and then you you push

937
00:55:20,680 --> 00:55:22,517
the two output vectors
away from each other.

938
00:55:22,600 --> 00:55:24,957
>> Mhm. >> And that would
eventually guarantee that

939
00:55:25,040 --> 00:55:28,557
things that are semantically similar
produce similar representations, and

940
00:55:28,640 --> 00:55:30,677
things that are different
produce different representations.

941
00:55:30,760 --> 00:55:33,357
>> Mhm. >> Uh we actually
came up with this idea

942
00:55:33,440 --> 00:55:35,717
for a project of doing uh signature

943
00:55:35,800 --> 00:55:40,797
verification. So, we would collect
signature signatures from like multiple

944
00:55:40,880 --> 00:55:43,957
signatures on the same person and then
train a neural net to produce the same

945
00:55:44,040 --> 00:55:47,677
representation. And then you know

946
00:55:47,760 --> 00:55:52,437
force the system to produce different
representation for different signatures.

947
00:55:52,520 --> 00:55:56,677
This was actually the the problem
was proposed by people from a what was a

948
00:55:56,760 --> 00:56:00,237
subsidiary of AT&T at the time called NCR.

949
00:56:00,320 --> 00:56:02,317
And they were interested in storing

950
00:56:02,400 --> 00:56:07,477
representation of the signature
on the 80 bytes of the magnetic strip of a

951
00:56:07,560 --> 00:56:10,037
credit card. >> [snorts] >> So,
we came up with this idea of having

952
00:56:10,120 --> 00:56:13,397
a neural net with 80 outputs.
You know, that we would quantize on

953
00:56:13,480 --> 00:56:16,957
bytes to so that we could encode the
the >> And then encoding was then used to

954
00:56:17,040 --> 00:56:19,037
compare whether signature matches or not.

955
00:56:19,120 --> 00:56:22,157
>> That's right. So, then you would
you know sign you would run through the

956
00:56:22,240 --> 00:56:25,037
neural net and then you would compare
the output vector to whatever is stored

957
00:56:25,120 --> 00:56:26,637
on your card. >> Did it actually work?

958
00:56:26,720 --> 00:56:28,927
>> It worked but they
ended up not using it.

959
00:56:29,010 --> 00:56:33,397
>> [laughter] >> Um because
nobody cares actually. I mean the

960
00:56:33,480 --> 00:56:38,997
American you know finish payment system
is incredibly lax in that respect compared

961
00:56:39,080 --> 00:56:41,597
to Europe for example.
>> the signatures. What's the purpose of

962
00:56:41,680 --> 00:56:45,037
signatures anyway? This is very >> Nobody
looks at them. Nobody cares, you know.

963
00:56:45,120 --> 00:56:49,517
>> And and it's uh yeah. >> Yeah, no.
So, uh so that that's contrastive running,

964
00:56:49,600 --> 00:56:53,517
right? So, you need positive and negative
pairs and the problem with that is that

965
00:56:53,600 --> 00:56:56,717
you know, even though I had
the original paper on this

966
00:56:56,800 --> 00:56:59,997
I'm actually not very positive about
it because it doesn't work in high

967
00:57:00,080 --> 00:57:02,997
dimension.
If your representation is high dimensional

968
00:57:03,080 --> 00:57:06,237
there's just too many ways
for two things to be different.

969
00:57:06,320 --> 00:57:09,957
And and so you would need lots
and lots and lots of negative pairs.

970
00:57:10,040 --> 00:57:12,957
Uh so, there is a a particular
implementation of this which is

971
00:57:13,040 --> 00:57:17,197
relatively recent from actually
the Google Toronto group

972
00:57:17,280 --> 00:57:21,077
uh where you know, Jeff Hinton is the
the senior member there. It's called

973
00:57:21,160 --> 00:57:23,917
SimCLR, S I M C L C L R. >> Mhm.

974
00:57:24,000 --> 00:57:27,037
>> Uh and it you know, basically
a particular way of implementing this idea

975
00:57:27,120 --> 00:57:30,557
of contrastive learning.
The particular objective function.

976
00:57:30,640 --> 00:57:34,397
Now what I'm much more enthusiastic about

977
00:57:34,480 --> 00:57:36,557
these days is non-contrasting methods.

978
00:57:36,640 --> 00:57:39,077
So, other ways to guarantee
that >> [snorts]

979
00:57:39,160 --> 00:57:42,197
>> uh uh the representations will be

980
00:57:42,280 --> 00:57:46,197
different for different
uh different inputs.

981
00:57:46,280 --> 00:57:49,757
And it's actually based on an idea that um

982
00:57:49,840 --> 00:57:53,877
Jeff Hinton proposed in the early '90s
with his student at the time, Subbarao.

983
00:57:53,960 --> 00:57:56,517
And it's based on the idea of maximizing
the mutual information between the

984
00:57:56,600 --> 00:58:00,197
outputs of the two systems. You only
show positive pairs. You only show pairs

985
00:58:00,280 --> 00:58:03,247
of images that you know
are somewhat similar.

986
00:58:03,330 --> 00:58:06,997
>> [snorts] >> And you train
the two networks to be informative,

987
00:58:07,080 --> 00:58:11,757
but also to be as informative
uh of each other as possible. So,

988
00:58:11,840 --> 00:58:16,717
basically, one representation has to be
predictable from the other, essentially.

989
00:58:16,800 --> 00:58:20,717
Uh and, you know, he proposed
that idea had, you know,

990
00:58:20,800 --> 00:58:24,077
a couple of papers in the early '90s,
and then nothing was done about it for

991
00:58:24,160 --> 00:58:29,397
decades. And I kind of revived this idea
together with my postdocs at at FAIR.

992
00:58:29,480 --> 00:58:32,717
Uh particularly a postdoc
called Stéphane Deny, who's now a

993
00:58:32,800 --> 00:58:35,997
junior professor in in Finland
at University of Alto.

994
00:58:36,080 --> 00:58:41,197
Um we came [snorts] up with something
called uh that we called Barlow Twins.

995
00:58:41,280 --> 00:58:42,837
And it's a a particular way of

996
00:58:42,920 --> 00:58:46,357
maximizing the information content of uh

997
00:58:46,440 --> 00:58:50,837
of a vector, you know,
using some uh hypotheses.

998
00:58:50,920 --> 00:58:54,557
And we have uh kind of a another
version of it that's more recent

999
00:58:54,640 --> 00:58:58,957
now called VICReg, VICREG. That that
means variance invariance covariance

1000
00:58:59,040 --> 00:59:02,477
regularization. it's the thing I'm
the most excited about in in machine

1001
00:59:02,560 --> 00:59:05,517
learning in the last 15 years. I mean,
I'm I'm not I'm really really excited

1002
00:59:05,600 --> 00:59:09,557
about this. >> What uh kind
of data augmentation is useful for

1003
00:59:09,640 --> 00:59:12,117
that non-contrastive learning method?

1004
00:59:12,200 --> 00:59:15,437
Are we talking about Does that not
matter that much? Or it seems like a

1005
00:59:15,520 --> 00:59:17,837
very important part of the step.

1006
00:59:17,920 --> 00:59:19,597
>> Yeah. >> How you
generate the images that are

1007
00:59:19,680 --> 00:59:21,517
similar but sufficiently different?

1008
00:59:21,600 --> 00:59:24,317
>> Yeah, that's right. It's a important
step, and it's also an annoying step

1009
00:59:24,400 --> 00:59:28,197
because you need to have that knowledge
of what the augmentation uh

1010
00:59:28,280 --> 00:59:32,117
you can do that do not change
the nature of the of the object.

1011
00:59:32,200 --> 00:59:35,277
>> Yeah. >> And so um
the standard scenario which

1012
00:59:35,360 --> 00:59:39,557
you know a lot of people working
in this area are using is you use uh

1013
00:59:39,640 --> 00:59:43,477
the the type of distortion So
so basically you do uh geometric

1014
00:59:43,560 --> 00:59:46,397
distortion. So one basically just shift
the image a little bit, it's called

1015
00:59:46,480 --> 00:59:50,317
cropping. Another one kind of changes
the scale a little bit. Another one kind

1016
00:59:50,400 --> 00:59:53,797
of rotates it. Another one changes the
colors, you know, you kind of do a shift

1017
00:59:53,880 --> 00:59:56,197
in color balance or something like that.

1018
00:59:56,280 --> 01:00:00,757
Um saturation another one sort of blurs
it. Another one adds noise. So you have

1019
01:00:00,840 --> 01:00:04,277
like a a catalog of kind of standard
things and people try to use the same

1020
01:00:04,360 --> 01:00:07,437
ones for different algorithms
so that they can compare.

1021
01:00:07,520 --> 01:00:10,717
Um but some algorithms uh some
self-supervised algorithm actually can

1022
01:00:10,800 --> 01:00:14,397
deal with much bigger like more
aggressive data augmentation

1023
01:00:14,480 --> 01:00:18,957
and some can't don't. So that kind of
makes the whole thing difficult. But but

1024
01:00:19,040 --> 01:00:21,797
that's the kind of distortions
we're talking about and

1025
01:00:21,880 --> 01:00:24,237
and so you you you you train with those

1026
01:00:24,320 --> 01:00:26,717
distortions. And then uh

1027
01:00:26,800 --> 01:00:29,237
you you chop off the last layer or

1028
01:00:29,320 --> 01:00:33,077
couple layers of the of the the network.

1029
01:00:33,160 --> 01:00:36,717
And you use the representation as input
to a classifier. You train the classifier.

1030
01:00:36,800 --> 01:00:40,637
>> Mhm. >> [clears throat] >> Um
on ImageNet let's say or whatever.

1031
01:00:40,720 --> 01:00:45,717
And measure the performance.
And interestingly enough the methods that

1032
01:00:45,800 --> 01:00:49,037
are really good at eliminating the
information that is irrelevant which is

1033
01:00:49,120 --> 01:00:53,547
the distortions between those images
do a good job at eliminating it.

1034
01:00:53,630 --> 01:00:56,837
>> [laughter] >> And uh
as a consequence you cannot use

1035
01:00:56,920 --> 01:01:00,957
those the representations in those
systems for things like object detection

1036
01:01:01,040 --> 01:01:04,027
and localization because
that information is gone.

1037
01:01:04,110 --> 01:01:06,117
>> [snorts] >> So the type
of data augmentation you

1038
01:01:06,200 --> 01:01:09,157
need to do depends on the
task you want eventually

1039
01:01:09,240 --> 01:01:13,037
the system to to solve and the type
of data augmentation standard data

1040
01:01:13,120 --> 01:01:16,717
augmentation that we use today are only
appropriate for object recognition or

1041
01:01:16,800 --> 01:01:19,557
image classification.
They're not appropriate for things like

1042
01:01:19,640 --> 01:01:22,997
>> Okay, help me out and what
uh localization is So, you're

1043
01:01:23,080 --> 01:01:27,037
saying it's just not good at the negative
uh like at cla- at classifying the

1044
01:01:27,120 --> 01:01:29,877
negative, so that's why it can't
be used for the localization.

1045
01:01:29,960 --> 01:01:32,517
>> No, it's just that
you train the system,

1046
01:01:32,600 --> 01:01:35,517
you know, you you you give it an image
and then you give it the same image

1047
01:01:35,600 --> 01:01:38,677
shifted and scaled.
You tell it that's the same image.

1048
01:01:38,760 --> 01:01:41,117
>> Yeah. >> So, the system
basically is trying to

1049
01:01:41,200 --> 01:01:44,037
eliminate the information
about position and size.

1050
01:01:44,120 --> 01:01:47,957
So, now And now you want to use
that >> Oh, like literally

1051
01:01:48,040 --> 01:01:49,997
>> out where an object
is and what size it is.

1052
01:01:50,080 --> 01:01:51,957
>> bounding box.
It did be able to actually

1053
01:01:52,040 --> 01:01:56,397
Okay. It can still find It can still
find the object in the image. It's just

1054
01:01:56,480 --> 01:02:00,917
not very good at finding the exact
boundaries of that object. Interesting.

1055
01:02:01,000 --> 01:02:03,957
Interesting, which you know,
there that's an interesting

1056
01:02:04,040 --> 01:02:07,677
sort of philosophical question.
How important How important is object

1057
01:02:07,760 --> 01:02:11,837
localization anyway? We're We're
like obsessed by measuring like image

1058
01:02:11,920 --> 01:02:16,077
segmentation. Obsessed by measuring
perfectly knowing the boundaries of

1059
01:02:16,160 --> 01:02:17,677
objects when

1060
01:02:17,760 --> 01:02:19,917
arguably

1061
01:02:20,000 --> 01:02:24,717
that's not that essential to an
understanding what are the contents of

1062
01:02:24,800 --> 01:02:29,357
the scene. >> On the other hand, I think
evolutionarily the first vision systems

1063
01:02:29,440 --> 01:02:32,797
in animals were basically all about
localization, very little about

1064
01:02:32,880 --> 01:02:36,317
recognition.
And in the human brain, you have two

1065
01:02:36,400 --> 01:02:41,357
separate pathways for um
recognizing the nature of of a scene

1066
01:02:41,440 --> 01:02:45,597
or an object and localizing
objects. So, you use the

1067
01:02:45,680 --> 01:02:49,317
first pathway called the ventral pathway
for, you know, telling what you're

1068
01:02:49,400 --> 01:02:51,957
looking at. >> [snorts] >> Uh
the other pathway, the dorsal

1069
01:02:52,040 --> 01:02:56,637
pathway, is used for navigation, for
grasping, for everything else. And, you

1070
01:02:56,720 --> 01:03:00,237
know, basically a lot of the things
you need for survival are >> [snorts]

1071
01:03:00,320 --> 01:03:03,600
>> localization and detection.

1072
01:03:03,880 --> 01:03:08,077
>> Is similarity learning or contrastive
learning or these non-contrastive

1073
01:03:08,160 --> 01:03:10,837
methods the same
as understanding something?

1074
01:03:10,920 --> 01:03:14,917
Just because you know a distorted cat
is the same as a non-distorted cat, does

1075
01:03:15,000 --> 01:03:18,677
Does mean you understand
what it means to be a cat?

1076
01:03:18,760 --> 01:03:22,037
>> To some extent. I mean, it's
a superficial understanding, obviously.

1077
01:03:22,120 --> 01:03:25,797
>> But like what is the ceiling of this
method, do you think? Is this just one

1078
01:03:25,880 --> 01:03:30,357
trick on the path to doing
self-supervised learning? Can we go

1079
01:03:30,440 --> 01:03:33,877
really, really far? >> I think
we can go really far. So, if we

1080
01:03:33,960 --> 01:03:38,557
figure out how to uh use
techniques of that type, perhaps

1081
01:03:38,640 --> 01:03:41,397
very different, but you
know, of the same nature,

1082
01:03:41,480 --> 01:03:46,637
to uh train a system from from video
to do video prediction, essentially.

1083
01:03:46,720 --> 01:03:51,957
Um I think we'll have a path,
um you know, towards

1084
01:03:52,040 --> 01:03:56,317
I you know, I wouldn't say unlimited,
but but a path towards some level of

1085
01:03:56,400 --> 01:04:02,187
uh you know, physical common sense
in in machines. And I also think that

1086
01:04:02,270 --> 01:04:04,757
>> [snorts] >> um that

1087
01:04:04,840 --> 01:04:09,597
ability to learn how the world works
from a sort of high-throughput channel

1088
01:04:09,680 --> 01:04:11,757
like like vision >> Mhm.

1089
01:04:11,840 --> 01:04:15,437
>> is a necessary step towards uh

1090
01:04:15,520 --> 01:04:19,357
sort of real artificial intelligence.
In other words, I believe in grounded

1091
01:04:19,440 --> 01:04:22,877
intelligence. I don't think we can train
a machine to be intelligent purely from

1092
01:04:22,960 --> 01:04:25,757
text. Because I think the amount of

1093
01:04:25,840 --> 01:04:26,917
information about the world that's

1094
01:04:27,000 --> 01:04:32,877
contained in text is tiny
compared to what we need to know.

1095
01:04:32,960 --> 01:04:35,437
>> Mhm. >> So, for example,
let let's uh and and

1096
01:04:35,520 --> 01:04:38,677
you know, people have attempted to do
this for for 30 years, right? The The

1097
01:04:38,760 --> 01:04:40,317
Cyc project and things like that, right?

1098
01:04:40,400 --> 01:04:43,917
Of basically kind of writing down all
the facts that are known and >> Mhm.

1099
01:04:44,000 --> 01:04:47,757
>> uh hoping that some some sort of common
sense will emerge. Um I think it's

1100
01:04:47,840 --> 01:04:51,917
basically hopeless. Uh but let me take
an example. You take an object, I I

1101
01:04:52,000 --> 01:04:55,637
describe a situation to you. I take
an object, I put it on the table, and I

1102
01:04:55,720 --> 01:04:59,677
push the table. It's completely obvious
to you that the object will be pushed

1103
01:04:59,760 --> 01:05:03,317
with the table, right?
Because it's sitting on it.

1104
01:05:03,400 --> 01:05:06,997
Um there's no text in the world,
I believe, that explains this.

1105
01:05:07,080 --> 01:05:11,197
And so, if you train a machine
as powerful as it could be, you know, your

1106
01:05:11,280 --> 01:05:13,717
GPT 5000

1107
01:05:13,800 --> 01:05:16,517
um whatever it is, it's never going to

1108
01:05:16,600 --> 01:05:21,197
learn about this.
Um that information is just is not

1109
01:05:21,280 --> 01:05:24,637
present in any text. >> Well,
the the question like with the Cyc

1110
01:05:24,720 --> 01:05:26,317
project, the dream I think is to have

1111
01:05:26,400 --> 01:05:28,477
like

1112
01:05:28,560 --> 01:05:32,357
like 10 million say facts like that

1113
01:05:32,440 --> 01:05:37,397
that give you a head start.
Like a parent guiding you.

1114
01:05:37,480 --> 01:05:41,117
Now, we humans don't need a parent
to tell us that the table will move uh

1115
01:05:41,200 --> 01:05:44,797
sorry, the smartphone will
move with the table. But

1116
01:05:44,880 --> 01:05:47,797
we get a lot of guidance in other ways.

1117
01:05:47,880 --> 01:05:50,357
So, it's possible that we
can give it a quick shortcut.

1118
01:05:50,440 --> 01:05:52,957
>> What about cat? A cat knows that.

1119
01:05:53,040 --> 01:05:57,277
>> No, but they evolved.
So >> No, they learn like us.

1120
01:05:57,360 --> 01:05:59,277
>> The the sorry, the physics of stuff?

1121
01:05:59,360 --> 01:06:04,037
>> Yeah. >> [sighs and gasps] >>
Well, yeah, so you're saying it's

1122
01:06:04,120 --> 01:06:08,997
uh so you're putting a lot of intelligence
onto the nurture side, not the nature.

1123
01:06:09,080 --> 01:06:13,837
>> Yes. >> Cuz we've seen to have
you know, there's a very inefficient

1124
01:06:13,920 --> 01:06:15,837
arguably process of evolution that got

1125
01:06:15,920 --> 01:06:19,717
us from bacteria to who we are today.

1126
01:06:19,800 --> 01:06:21,717
Started at the bottom, now we're here.

1127
01:06:21,800 --> 01:06:26,077
So, uh the question is how Okay, so

1128
01:06:26,160 --> 01:06:30,237
the question is how fundamental is that
that the nature of the whole hardware

1129
01:06:30,320 --> 01:06:33,797
uh and then is there any
way to shortcut it if it's

1130
01:06:33,880 --> 01:06:36,917
fundamental? If it's not, if it's most
of intelligence, most of the cool stuff

1131
01:06:37,000 --> 01:06:38,877
we've been talking about is mostly

1132
01:06:38,960 --> 01:06:41,157
nurture mostly trained, we figured out

1133
01:06:41,240 --> 01:06:43,997
by observing the world. We can form that

1134
01:06:44,080 --> 01:06:48,277
uh big beautiful sexy background
model that you're talking about just by

1135
01:06:48,360 --> 01:06:50,837
sitting there. Um

1136
01:06:50,920 --> 01:06:53,717
then okay, then you need to then like

1137
01:06:53,800 --> 01:06:58,597
maybe uh it is all
supervised learning all the

1138
01:06:58,680 --> 01:07:00,917
way down. It's all
supervised learning, say.

1139
01:07:01,000 --> 01:07:03,997
>> Whatever it is that makes
uh you know, human intelligence

1140
01:07:04,080 --> 01:07:08,237
different from other animals, which you
know, a lot of people think is language

1141
01:07:08,320 --> 01:07:10,597
and logical reasoning
and this kind of stuff.

1142
01:07:10,680 --> 01:07:14,717
It cannot be that complicated because it
only popped up in the last million years.

1143
01:07:14,800 --> 01:07:19,037
>> Yeah. >> Uh and >> Yeah, it it >>
you know, it it and it only involves,

1144
01:07:19,120 --> 01:07:21,717
you know, less than 1%
of our genome, right? Which

1145
01:07:21,800 --> 01:07:25,277
is the difference between human
genome and chimps or whatever.

1146
01:07:25,360 --> 01:07:28,757
So, uh it can't be
that complicated, you know,

1147
01:07:28,840 --> 01:07:32,757
it can't be that fundamental. I mean,
the most of the so complicated stuff

1148
01:07:32,840 --> 01:07:37,037
already exists in cats and dogs and you
know, certainly primates, non-human

1149
01:07:37,120 --> 01:07:40,517
primates. >> Yeah,
that little thing with humans

1150
01:07:40,600 --> 01:07:43,597
might be just a something about social

1151
01:07:43,680 --> 01:07:48,717
interaction and ability to maintain
ideas across like a collective of

1152
01:07:48,800 --> 01:07:54,117
people. It it's it sounds very dramatic
and very impressive, but it probably isn't

1153
01:07:54,200 --> 01:07:56,877
mechanistically speaking. >> It is,
but we're not there yet. Like, we

1154
01:07:56,960 --> 01:08:01,397
you know, we we have I mean,
this is number 634,

1155
01:08:01,480 --> 01:08:05,317
you know, in the list
of problems we have to solve.

1156
01:08:05,400 --> 01:08:10,277
>> So, basic physics of the world
is is is is number one. What do you um

1157
01:08:10,360 --> 01:08:14,237
Just a quick tangent on data
augmentation. So, a lot

1158
01:08:14,320 --> 01:08:15,837
of it is

1159
01:08:15,920 --> 01:08:19,717
uh hard-coded versus learned.

1160
01:08:19,800 --> 01:08:22,837
Do you have any intuition that maybe

1161
01:08:22,920 --> 01:08:27,197
there could be some weird data
augmentation like generative type of

1162
01:08:27,280 --> 01:08:31,197
data augmentation, like doing
something weird to images which then

1163
01:08:31,280 --> 01:08:34,197
improves the the similarity learning

1164
01:08:34,280 --> 01:08:36,917
process? So, not just kind of dumb

1165
01:08:37,000 --> 01:08:41,557
simple distortions, but by you shaking
your head just saying that even simple

1166
01:08:41,640 --> 01:08:44,917
distortions are enough. >> I think no.
I think data augmentation is

1167
01:08:45,000 --> 01:08:48,037
a temporary necessary evil. >> Huh.

1168
01:08:48,120 --> 01:08:51,997
>> Um so, what people are working
on now is is two things. One is

1169
01:08:52,080 --> 01:08:56,637
uh the type of self-supervised learning
like trying to to translate the type of

1170
01:08:56,720 --> 01:09:00,557
self-supervised learning people use
in language, translating this to images,

1171
01:09:00,640 --> 01:09:02,077
which is basically a denoising

1172
01:09:02,160 --> 01:09:04,277
autoencoder method. Right? So, you you

1173
01:09:04,360 --> 01:09:09,237
take an image, you you block
you mask some parts of it.

1174
01:09:09,320 --> 01:09:12,677
Uh and then you you train some giant
neural net to reconstruct the parts that

1175
01:09:12,760 --> 01:09:18,157
you that are that are missing.
And until very recently

1176
01:09:18,240 --> 01:09:20,997
there was no there was no
working methods for that.

1177
01:09:21,080 --> 01:09:24,557
All the auto encoder type methods
for images weren't producing very good

1178
01:09:24,640 --> 01:09:27,877
representation. But there's
a paper now coming out of

1179
01:09:27,960 --> 01:09:32,677
the fair group at Menlo Park
that actually works very well. So,

1180
01:09:32,760 --> 01:09:35,037
that doesn't require the data
augmentation, that requires only

1181
01:09:35,120 --> 01:09:40,077
masking. >> Okay. Only masking for images

1182
01:09:40,160 --> 01:09:42,157
Okay. [snorts] >> Right.
So, you mask part of the image

1183
01:09:42,240 --> 01:09:46,917
and you train a system which you know
in this case is a transformer because

1184
01:09:47,000 --> 01:09:51,237
you can you can the transformer
represents the image as

1185
01:09:51,320 --> 01:09:55,237
uh non-overlapping patches, so it's easy
to mask patches and things like that.

1186
01:09:55,320 --> 01:09:59,077
>> Okay, then my question transfers to
that problem then masking. Like why should

1187
01:09:59,160 --> 01:10:01,797
the mask be a square or rectangle?

1188
01:10:01,880 --> 01:10:04,997
>> Yeah, so it doesn't matter. Like
you know, I think we're going to come up

1189
01:10:05,080 --> 01:10:09,877
probably in the future with sort of uh
you know, ways to mask that are it's you

1190
01:10:09,960 --> 01:10:14,797
know, kind of random essentially.
I mean, they are random already, but

1191
01:10:14,880 --> 01:10:18,717
>> No, no, but like
something that's challenging.

1192
01:10:18,800 --> 01:10:22,437
Like optimally challenging.
So, like I mean,

1193
01:10:22,520 --> 01:10:23,917
maybe it's a metaphor that doesn't

1194
01:10:24,000 --> 01:10:26,557
apply, but you're it seems like there's

1195
01:10:26,640 --> 01:10:32,077
an data augmentation or masking
is an interactive element with it. Like

1196
01:10:32,160 --> 01:10:33,997
you're almost like playing with an image.

1197
01:10:34,080 --> 01:10:37,597
>> Yeah. >> And like it's like the way
we play with an image in our minds.

1198
01:10:37,680 --> 01:10:41,597
>> No, but it's like dropout.
It's like Boltzmann machine training. You

1199
01:10:41,680 --> 01:10:46,317
you know, every every every
time you see a percept you also

1200
01:10:46,400 --> 01:10:49,957
you you can you can perturb
it in some way and then

1201
01:10:50,040 --> 01:10:53,437
the the principle
of the training procedure

1202
01:10:53,520 --> 01:10:57,917
is to minimize the difference of the
output of the representation between the

1203
01:10:58,000 --> 01:11:02,797
the clean version and the corrupted
version essentially, right? And you you

1204
01:11:02,880 --> 01:11:05,717
you can do this in real time, right?
So, you know, Boltzmann machine works like

1205
01:11:05,800 --> 01:11:09,997
this, right? You you you show a percept,
uh you tell the machine that's a good

1206
01:11:10,080 --> 01:11:12,157
combination of activities of your input

1207
01:11:12,240 --> 01:11:17,397
neurons. Uh and then you either uh

1208
01:11:17,480 --> 01:11:22,517
let them go their merry way without
clamping them to values or you only do

1209
01:11:22,600 --> 01:11:26,157
this with a subset. >> Yeah. >> [snorts]
>> And what you do is you're training the

1210
01:11:26,240 --> 01:11:30,397
system so that the the stable state
of the entire network is the same

1211
01:11:30,480 --> 01:11:35,077
regardless of whether it's his entire
input or whether it's his only part of it.

1212
01:11:35,160 --> 01:11:38,397
Um you know, the noise in autoencoder
method is basically the same thing,

1213
01:11:38,480 --> 01:11:41,557
right? You you're you're training
a system to reproduce the input the

1214
01:11:41,640 --> 01:11:45,637
complete input and filling the blanks
regardless of which which parts are are

1215
01:11:45,720 --> 01:11:49,437
missing and that's really the underlying
principle. And you could imagine sort of

1216
01:11:49,520 --> 01:11:53,917
even in the brain some sort of neural
principle where, you know, neurons can uh

1217
01:11:54,000 --> 01:11:56,717
oscillate, right? So, they they
take their activity and then

1218
01:11:56,800 --> 01:12:00,877
temporarily they kind of shut off
to um you know, force the rest of the

1219
01:12:00,960 --> 01:12:02,717
system to basically

1220
01:12:02,800 --> 01:12:05,237
reconstruct the input without their

1221
01:12:05,320 --> 01:12:10,437
help, you know? And and and I mean,
you you could imagine, you know,

1222
01:12:10,520 --> 01:12:13,117
you know, more or less
biologically possible processes.

1223
01:12:13,200 --> 01:12:14,317
>> Something like that. And I guess with

1224
01:12:14,400 --> 01:12:18,317
this denoising autoencoder and masking

1225
01:12:18,400 --> 01:12:21,637
and data augmentation,
you don't have to worry about being

1226
01:12:21,720 --> 01:12:25,517
super efficient. You can just
do do as much as you want.

1227
01:12:25,600 --> 01:12:28,077
>> Yeah. >> And and get better over time.

1228
01:12:28,160 --> 01:12:31,637
Cuz I was thinking like you might want
to be clever about the way you do all of

1229
01:12:31,720 --> 01:12:36,077
these procedures, you know,
but that's only if it's

1230
01:12:36,160 --> 01:12:39,877
somehow costly to do every
iteration, but it's not really.

1231
01:12:39,960 --> 01:12:43,557
>> Not really. Maybe.
And then there is, you know, data

1232
01:12:43,640 --> 01:12:46,117
augmentation without
explicit data augmentation.

1233
01:12:46,200 --> 01:12:49,237
This data augmentation by waiting,
which is, you know, the the sort of video

1234
01:12:49,320 --> 01:12:55,157
prediction. >> Right. >> Um you're
observing a video clip observing the

1235
01:12:55,240 --> 01:12:58,317
you know the continuation
of that video clip and try

1236
01:12:58,400 --> 01:13:02,197
you try to learn a representation
using the joint embedding architectures

1237
01:13:02,280 --> 01:13:06,597
in such a way that the representation
of the future clip is easily predictable

1238
01:13:06,680 --> 01:13:08,517
from the representation of the of the

1239
01:13:08,600 --> 01:13:10,397
observed clip.

1240
01:13:10,480 --> 01:13:14,637
>> Do you think YouTube
has enough raw data

1241
01:13:14,720 --> 01:13:18,237
from which to learn how to be a cat?

1242
01:13:18,320 --> 01:13:23,277
>> I think so. >> So it the amount
of data is not the constraint. I

1243
01:13:23,360 --> 01:13:25,957
mean >> it would require
some selection I think.

1244
01:13:26,040 --> 01:13:29,797
>> Some actually >> Some some
selection of maybe the right

1245
01:13:29,880 --> 01:13:33,077
type of data you need some >>
the rabbit hole of just cat videos that

1246
01:13:33,160 --> 01:13:36,397
might you might need to watch
some lectures or something.

1247
01:13:36,480 --> 01:13:39,397
>> No, you >> How matter would that be

1248
01:13:39,480 --> 01:13:42,397
if it like watches lectures about

1249
01:13:42,480 --> 01:13:46,457
intelligence and then learns watches
your lectures and NYU and learns

1250
01:13:46,540 --> 01:13:48,157
[clears throat] from that
how to be intelligent.

1251
01:13:48,240 --> 01:13:49,737
>> I don't think it would be enough.

1252
01:13:49,820 --> 01:13:54,357
>> [laughter] [gasps] >> What's
your do you find multimodal

1253
01:13:54,440 --> 01:13:57,717
learning interesting we've been talking
about visual language like if combining

1254
01:13:57,800 --> 01:14:00,117
those together maybe audio
all those kinds of things.

1255
01:14:00,200 --> 01:14:05,117
>> lot of things that I find interesting
in the short term but are not addressing

1256
01:14:05,200 --> 01:14:08,877
the important problem that I think are
really kind of the big challenges. So I

1257
01:14:08,960 --> 01:14:10,397
think you know things like multitask

1258
01:14:10,480 --> 01:14:15,237
learning continual learning you know

1259
01:14:15,320 --> 01:14:19,077
adversarial issues. I mean those have
you know great practical interests in

1260
01:14:19,160 --> 01:14:22,797
the relative short term.
Possibly but I don't think they're

1261
01:14:22,880 --> 01:14:25,677
fundamental you know active learning
even to some extent reinforcement

1262
01:14:25,760 --> 01:14:28,677
learning. I think those things will

1263
01:14:28,760 --> 01:14:32,157
become either obsolete or or useless or

1264
01:14:32,240 --> 01:14:36,117
easy once we figured out how to do

1265
01:14:36,200 --> 01:14:37,837
self-supervised representation learning

1266
01:14:37,920 --> 01:14:41,157
or or learning predictive world models.

1267
01:14:41,240 --> 01:14:46,317
And so I think that's what you know
the entire community should be focusing on

1268
01:14:46,400 --> 01:14:48,997
at least people who are interested
in sort of fundamental questions or you

1269
01:14:49,080 --> 01:14:52,637
know, really kind of pushing the envelope
of AI towards the next the next

1270
01:14:52,720 --> 01:14:55,477
stage. But of course there is like
a huge amount of you know, very

1271
01:14:55,560 --> 01:14:58,637
interesting work to do in sort
of practical questions that that have you

1272
01:14:58,720 --> 01:15:01,317
know, short-term impact.
>> Well, you know,

1273
01:15:01,400 --> 01:15:05,557
it's it's difficult to talk about that
the temporal scale because all of human

1274
01:15:05,640 --> 01:15:09,837
civilization will eventually be
destroyed because the the the the sun

1275
01:15:09,920 --> 01:15:14,077
will die out and even if Elon
Musk is successful multi-planetary

1276
01:15:14,160 --> 01:15:16,437
colonization across the galaxy,

1277
01:15:16,520 --> 01:15:18,797
eventually the entirety of it will just

1278
01:15:18,880 --> 01:15:22,317
become giant black holes and um

1279
01:15:22,400 --> 01:15:24,077
>> That's going to take a while though.

1280
01:15:24,160 --> 01:15:27,797
>> So but but what I'm saying is then
that logic can be used to say it's all

1281
01:15:27,880 --> 01:15:31,477
meaningless.
I'm saying all that to say that

1282
01:15:31,560 --> 01:15:33,797
multitask learning

1283
01:15:33,880 --> 01:15:36,797
might be you're you're calling it

1284
01:15:36,880 --> 01:15:39,437
practical or pragmatic or whatever. That

1285
01:15:39,520 --> 01:15:40,837
might be the thing that achieves

1286
01:15:40,920 --> 01:15:44,517
something very akin to intelligence

1287
01:15:44,600 --> 01:15:47,357
while we're trying to solve the more

1288
01:15:47,440 --> 01:15:51,757
uh general problem of self-supervised
learning of background knowledge. So the

1289
01:15:51,840 --> 01:15:55,477
reason I bring that up, maybe one way
to ask that question. I've been very

1290
01:15:55,560 --> 01:15:59,037
impressed by what Tesla autopilot team
is doing. I don't know if you got any

1291
01:15:59,120 --> 01:16:02,557
chance to glance at this
particular one example of

1292
01:16:02,640 --> 01:16:06,917
multitask learning where they're
literally taking the problem

1293
01:16:07,000 --> 01:16:09,317
like I don't know, Charles Darwin start

1294
01:16:09,400 --> 01:16:14,157
studying animals. They're studying
the problem of driving and asking, okay,

1295
01:16:14,240 --> 01:16:16,917
what are all the things
you have to perceive?

1296
01:16:17,000 --> 01:16:21,157
And the way they're solving it is one,
there's an ontology where you're

1297
01:16:21,240 --> 01:16:24,117
bringing that to the table. So you're
formulating a bunch of different tasks.

1298
01:16:24,200 --> 01:16:26,957
It's like over 100 tasks or something
like that that they're involved in

1299
01:16:27,040 --> 01:16:31,197
driving. And then they're deploying it
and then getting data back from people

1300
01:16:31,280 --> 01:16:35,637
that run into trouble and then trying to
figure out do we add task? Do we like we

1301
01:16:35,720 --> 01:16:39,837
focus on each individual task separately?
In fact, half So the I would

1302
01:16:39,920 --> 01:16:44,757
say I'll classify Andrej Karpathy's talk
into it. So, one was about doors and the

1303
01:16:44,840 --> 01:16:46,597
other one about how much ImageNet sucks.

1304
01:16:46,680 --> 01:16:50,557
He kept going back and forth
on those two topics,

1305
01:16:50,640 --> 01:16:53,197
which ImageNet sucks meaning you can't

1306
01:16:53,280 --> 01:16:55,717
just use a single benchmark. There's so

1307
01:16:55,800 --> 01:17:00,557
like you you have to have like a a giant
suite of benchmarks to understand how

1308
01:17:00,640 --> 01:17:02,117
well your your system actually works.

1309
01:17:02,200 --> 01:17:06,997
>> with him. I mean, he's he's
he's very sensible guy. Um Now,

1310
01:17:07,080 --> 01:17:11,157
okay, it's it's very clear that if
you're faced with a an engineering

1311
01:17:11,240 --> 01:17:14,797
problem that you need to solve in a
relatively short time, particularly if

1312
01:17:14,880 --> 01:17:17,837
you have Elon Musk
breathing down your neck,

1313
01:17:17,920 --> 01:17:21,637
you're going to have to take shortcuts,
right? You you you might think about the

1314
01:17:21,720 --> 01:17:26,037
the fact that the the the right thing
to do and the long-term solution involves,

1315
01:17:26,120 --> 01:17:29,997
you know, some fancy
self-supervisioning, but you have

1316
01:17:30,080 --> 01:17:34,477
you know, Elon Musk breathing down
your neck, uh and, you know, this involves

1317
01:17:34,560 --> 01:17:38,397
uh you know, human lives,
and so you you have to

1318
01:17:38,480 --> 01:17:43,077
basically just do the systematic
engineering and, you know,

1319
01:17:43,160 --> 01:17:47,997
uh fine-tuning and refinements and trial
and error and and and all that stuff.

1320
01:17:48,080 --> 01:17:49,317
Um there's nothing wrong with that.

1321
01:17:49,400 --> 01:17:52,637
That's that's called engineering.
That's called, you know, uh

1322
01:17:52,720 --> 01:17:56,077
putting technology out uh in the

1323
01:17:56,160 --> 01:17:58,877
yeah, in the world, um and and you have

1324
01:17:58,960 --> 01:18:03,517
to kind of iron clad it before
before you do this, you know, um

1325
01:18:03,600 --> 01:18:08,837
uh so much for, you know,
grand grand ideas and principles.

1326
01:18:08,920 --> 01:18:11,917
Um But, you know what? I'm placing myself

1327
01:18:12,000 --> 01:18:16,357
sort of, you know, some you know,
upstream of this, you know,

1328
01:18:16,440 --> 01:18:19,357
quite a bit upstream of this.
>> You're Plato, thinking about Platonic

1329
01:18:19,440 --> 01:18:21,197
forms. You're You're

1330
01:18:21,280 --> 01:18:23,757
>> No, Platonic because eventually I want

1331
01:18:23,840 --> 01:18:28,957
us stuff to get used, but uh it's
okay if it takes 5 or 10 years for the

1332
01:18:29,040 --> 01:18:33,117
community to realize this is the right
thing to do. I've I've done this before.

1333
01:18:33,200 --> 01:18:36,997
It's been the case before that, you You
I've made that that case. I mean, if you

1334
01:18:37,080 --> 01:18:40,917
look back in the mid-2000s, for example,
and you ask yourself the question,

1335
01:18:41,000 --> 01:18:44,487
"Okay, I want to recognize
cars or faces or whatever."

1336
01:18:44,570 --> 01:18:47,477
>> [snorts] >> Um you know,
I can use convolutional nets,

1337
01:18:47,560 --> 01:18:51,637
or I can use some more conventional uh
kind of computer vision techniques, you

1338
01:18:51,720 --> 01:18:56,157
know, using interest point detectors
or SIFT dense features, and you know,

1339
01:18:56,240 --> 01:18:58,157
sticking an SVM on top. At that time,

1340
01:18:58,240 --> 01:19:00,717
the data sets were so small that

1341
01:19:00,800 --> 01:19:05,717
those methods that used more engineering
worked better than ConvNets. There was

1342
01:19:05,800 --> 01:19:09,317
just not enough data for ConvNets.
And ConvNets were were a little a little

1343
01:19:09,400 --> 01:19:12,757
slow with the kind of hardware
that was available at the time.

1344
01:19:12,840 --> 01:19:17,517
And there was a sea change when
uh basically when, you know,

1345
01:19:17,600 --> 01:19:21,877
data sets became bigger and and GPUs
became available. That that's what, you

1346
01:19:21,960 --> 01:19:25,757
know, the two of the main
factors that basically

1347
01:19:25,840 --> 01:19:28,557
made people change
their change their mind.

1348
01:19:28,640 --> 01:19:32,317
Um >> [snorts] >> and you can
you can look at the history

1349
01:19:32,400 --> 01:19:36,477
of like all sub-branches of AI or pattern

1350
01:19:36,560 --> 01:19:40,557
recognition.
And there's a similar trajectory

1351
01:19:40,640 --> 01:19:44,877
followed by techniques where people
start by, you know, engineering the hell

1352
01:19:44,960 --> 01:19:49,277
out of it. Um you know, be it

1353
01:19:49,360 --> 01:19:52,957
optical character recognition,
speech recognition,

1354
01:19:53,040 --> 01:19:56,037
computer vision, like image
recognition in general,

1355
01:19:56,120 --> 01:19:59,677
uh natural language understanding, like,
you know, translation, things like that,

1356
01:19:59,760 --> 01:20:02,397
right? You start to engineer
the hell out of it.

1357
01:20:02,480 --> 01:20:05,557
Um you start to acquire all the knowledge
that prior knowledge you know

1358
01:20:05,640 --> 01:20:09,277
about image formation, about, you know,
the shape of characters, about, you

1359
01:20:09,360 --> 01:20:13,037
know, morphological operations,
about like feature extraction, Fourier

1360
01:20:13,120 --> 01:20:17,037
transforms, you know, Zernike moments,
you know, whatever, right? People have

1361
01:20:17,120 --> 01:20:20,197
come up with thousands of ways
of representing images so that you they

1362
01:20:20,280 --> 01:20:24,077
could be easily classified
afterwards. Same for speech

1363
01:20:24,160 --> 01:20:26,957
recognition, right? There is,
you know, two decades for people to

1364
01:20:27,040 --> 01:20:30,837
figure out a good front
end to preprocess uh,

1365
01:20:30,920 --> 01:20:34,157
uh, speech signals so that, you know,
the information about what is being said

1366
01:20:34,240 --> 01:20:39,437
is preserved, but most of the information
about the identity of the speaker is gone.

1367
01:20:39,520 --> 01:20:41,277
Um, you know, [snorts] catch all

1368
01:20:41,360 --> 01:20:43,837
coefficients or whatever, right? Um,

1369
01:20:43,920 --> 01:20:47,157
and same for for text, right? Uh, you do

1370
01:20:47,240 --> 01:20:49,397
named entity recognition and you parse

1371
01:20:49,480 --> 01:20:53,517
and you you you do uh, tagging of uh, of

1372
01:20:53,600 --> 01:20:57,517
of the parts of speech and, you know,
you do this sort of tree representation

1373
01:20:57,600 --> 01:21:02,157
of uh, clauses and all that stuff,
right? Before you can do anything.

1374
01:21:02,240 --> 01:21:07,197
Um, so, that's how it starts, right? Just

1375
01:21:07,280 --> 01:21:10,997
engineer the hell out of it.
And then, you start having data

1376
01:21:11,080 --> 01:21:13,917
and maybe you have more powerful
computers, maybe you know something

1377
01:21:14,000 --> 01:21:16,877
about statistical learning. So,
you start using machine learning and it's

1378
01:21:16,960 --> 01:21:20,157
usually a small sliver on top of your
kind of hand crafted system where, you

1379
01:21:20,240 --> 01:21:22,517
know, you extract features by hand.

1380
01:21:22,600 --> 01:21:26,157
Okay, and now, you know, nowadays,
the standard way of doing this is that you

1381
01:21:26,240 --> 01:21:28,957
train the entire thing end to end with
a deep learning system and it learns its

1382
01:21:29,040 --> 01:21:33,717
own features and and, you know,
speech recognition systems nowadays,

1383
01:21:33,800 --> 01:21:37,157
uh, or OCR systems are completely end
to end. It's the uh, you know, it's some

1384
01:21:37,240 --> 01:21:40,837
giant neural net that takes raw waveforms

1385
01:21:40,920 --> 01:21:44,317
and produces a sequence of characters
coming out. And it's just a huge neural

1386
01:21:44,400 --> 01:21:47,317
net, right? There's no you know,
Markov model, there's no

1387
01:21:47,400 --> 01:21:51,037
language model that is explicit
other than, you know, something that's

1388
01:21:51,120 --> 01:21:54,197
ingrained in the in the sort of neural
language model, if you want. Same for

1389
01:21:54,280 --> 01:21:59,317
translation, same for all kinds of stuff.
So, you see this continuous evolution

1390
01:21:59,400 --> 01:22:03,277
from you know, less and less hand crafting

1391
01:22:03,360 --> 01:22:07,477
and more and more learning. Um, and uh,

1392
01:22:07,560 --> 01:22:09,877
I I think uh,

1393
01:22:09,960 --> 01:22:12,597
I mean, it's true in biology as well.

1394
01:22:12,680 --> 01:22:17,197
>> So, I mean, we might disagree about
this, maybe not uh, in the this one

1395
01:22:17,280 --> 01:22:21,437
little piece at the end.
You mentioned active learning.

1396
01:22:21,520 --> 01:22:25,357
It feels like active learning, which
is the selection of data and also the

1397
01:22:25,440 --> 01:22:29,597
interactivity, needs to be part of this
giant neural network. You cannot just be

1398
01:22:29,680 --> 01:22:33,757
an observer to do self-supervised
learning. You have to Well,

1399
01:22:33,840 --> 01:22:38,077
I don't self-supervised learning is just
a word, but I would whatever this giant

1400
01:22:38,160 --> 01:22:43,037
stack of a neural network that's
automatically learning, it feels My

1401
01:22:43,120 --> 01:22:46,077
intuition is that

1402
01:22:46,160 --> 01:22:48,837
you have to have a system, whether it's

1403
01:22:48,920 --> 01:22:52,117
a physical robot or a digital robot

1404
01:22:52,200 --> 01:22:57,117
that's interacting with the world
and doing so in a flawed way and improving

1405
01:22:57,200 --> 01:23:02,317
over time. In order
to to form the self-supervised

1406
01:23:02,400 --> 01:23:06,917
learning, well, you can't just
give it a giant sea of data.

1407
01:23:07,000 --> 01:23:08,317
>> Okay, I agree and I disagree.

1408
01:23:08,400 --> 01:23:12,237
>> Okay. >> I agree
in the sense that I think uh

1409
01:23:12,320 --> 01:23:17,037
I agree I agree in two ways. The first The
first way I agree is that if you want uh

1410
01:23:17,120 --> 01:23:20,397
and you certainly need a causal model of
the world that allows you to predict the

1411
01:23:20,480 --> 01:23:23,677
consequences of your actions,
to train that model, you need to take

1412
01:23:23,760 --> 01:23:28,077
actions, right? You need to be able
to act in a world and see the effect

1413
01:23:28,160 --> 01:23:30,357
for you to be to learn
causal models of the world.

1414
01:23:30,440 --> 01:23:33,477
>> Well, so that's not That's not
obvious because you can observe others.

1415
01:23:33,560 --> 01:23:35,997
>> You can observe others.
>> And you can infer that they're similar

1416
01:23:36,080 --> 01:23:37,997
to you, and then you can learn from that.

1417
01:23:38,080 --> 01:23:41,037
>> Yeah, but then you have to kind of
hardwire that part, right? And you know,

1418
01:23:41,120 --> 01:23:44,077
mirror neurons and all
that stuff, right? So, um

1419
01:23:44,160 --> 01:23:47,237
and it's not clear to me how you
would do this in a machine. So,

1420
01:23:47,320 --> 01:23:51,157
um um so, I think
the the action part would

1421
01:23:51,240 --> 01:23:53,837
be necessary for having causal models of

1422
01:23:53,920 --> 01:23:54,957
of the world.

1423
01:23:55,040 --> 01:23:58,517
Uh the second reason it may be necessary

1424
01:23:58,600 --> 01:24:03,597
or at least more efficient is that uh
active learning basically, you know,

1425
01:24:03,680 --> 01:24:07,397
goes for the jugular of what you're what
you don't know, right? Mhm. It's It's

1426
01:24:07,480 --> 01:24:10,397
you know, obvious areas of uncertainty

1427
01:24:10,480 --> 01:24:12,837
uh about your your world and about the

1428
01:24:12,920 --> 01:24:14,877
how the world behaves.

1429
01:24:14,960 --> 01:24:19,117
And you can resolve this uncertainty by

1430
01:24:19,200 --> 01:24:22,877
systematic exploration of that part
that you don't you don't know. And if you

1431
01:24:22,960 --> 01:24:25,477
know that you don't know, then, you know,
it makes you curious. You kind of

1432
01:24:25,560 --> 01:24:28,437
look into situations that And

1433
01:24:28,520 --> 01:24:31,237
you know, across the animal world,

1434
01:24:31,320 --> 01:24:35,677
different species have different
levels of curiosity, right? [laughter]

1435
01:24:35,760 --> 01:24:36,997
Depending on how they're built, right?

1436
01:24:37,080 --> 01:24:41,597
So, you know, [snorts] cats and rats are
incredibly curious. Dogs not so much. I

1437
01:24:41,680 --> 01:24:45,837
mean, less. >> Yeah. So, it could be
useful to have that kind of curiosity.

1438
01:24:45,920 --> 01:24:48,997
>> So, it would be useful, but curiosity
just makes the process faster. It

1439
01:24:49,080 --> 01:24:51,437
doesn't make the process exist.

1440
01:24:51,520 --> 01:24:55,877
The So, what process what learning process

1441
01:24:55,960 --> 01:25:00,557
is it that active learning
makes more efficient?

1442
01:25:00,640 --> 01:25:05,117
And I'm asking that first
question, you know, >> [snorts]

1443
01:25:05,200 --> 01:25:07,717
>> you know, we haven't answered that
question yet. So, you know, I'll worry

1444
01:25:07,800 --> 01:25:09,997
about active learning
once this question is uh

1445
01:25:10,080 --> 01:25:11,597
>> So, that's the more
fundamental question

1446
01:25:11,680 --> 01:25:15,237
to ask. And if active learning or

1447
01:25:15,320 --> 01:25:17,957
interaction increases the efficiency of

1448
01:25:18,040 --> 01:25:20,437
the learning, see, sometimes

1449
01:25:20,520 --> 01:25:25,597
it becomes very different if the increase
is several orders of magnitude,

1450
01:25:25,680 --> 01:25:29,637
right? Like >> That's true.
>> But fundamentally it's still the same

1451
01:25:29,720 --> 01:25:33,117
thing and building up
the intuition about how to

1452
01:25:33,200 --> 01:25:36,357
in a self-supervised way to construct
background models, efficient or

1453
01:25:36,440 --> 01:25:40,557
inefficient, is is the core problem.

1454
01:25:40,640 --> 01:25:44,797
What do you think about Yoshua
Bengio's talking about consciousness

1455
01:25:44,880 --> 01:25:46,437
and all of these kinds of concepts?

1456
01:25:46,520 --> 01:25:50,317
>> Okay. I don't know
what consciousness is.

1457
01:25:50,400 --> 01:25:53,877
But uh >> That's a good opener.

1458
01:25:53,960 --> 01:25:57,597
>> And to some extent, a lot of the things
that are said about consciousness remind

1459
01:25:57,680 --> 01:26:00,117
me of the questions people were asking

1460
01:26:00,200 --> 01:26:02,797
themselves in the 18th century or 17th

1461
01:26:02,880 --> 01:26:07,197
century when they discovered that uh
you know, how the eye works and the

1462
01:26:07,280 --> 01:26:10,277
fact that the image
at the back of the eye was

1463
01:26:10,360 --> 01:26:14,397
upside down, right? Because you have
a lens and And so, on your retina, the

1464
01:26:14,480 --> 01:26:18,357
image that forms is an the of the world,
but it's upside-down. How is it that you

1465
01:26:18,440 --> 01:26:21,797
see right-side up? And you know,
with what we know today in

1466
01:26:21,880 --> 01:26:25,757
science, you know, we realize
this question doesn't make any sense.

1467
01:26:25,840 --> 01:26:29,317
Or or is kind of ridiculous in some way,
right? So, I think a lot of what is said

1468
01:26:29,400 --> 01:26:30,917
about consciousness is of that nature.

1469
01:26:31,000 --> 01:26:34,077
Now, that said, there's a lot
of really smart people that

1470
01:26:34,160 --> 01:26:37,197
for for whom I have a lot of respect
who are talking about this topic, people

1471
01:26:37,280 --> 01:26:40,397
like David Chalmers,
who is a colleague of mine at NYU.

1472
01:26:40,480 --> 01:26:43,717
Um I have [snorts] kind of a

1473
01:26:43,800 --> 01:26:47,677
unorthodox folk uh

1474
01:26:47,760 --> 01:26:51,197
speculative hypothesis
about consciousness. So,

1475
01:26:51,280 --> 01:26:53,037
we're talking about
this idea of world model.

1476
01:26:53,120 --> 01:26:56,477
>> Mhm. >> [clears throat] >>
And uh I think, you know, our entire

1477
01:26:56,560 --> 01:26:58,877
prefrontal cortex basically is

1478
01:26:58,960 --> 01:27:02,437
uh the engine for our world model.

1479
01:27:02,520 --> 01:27:06,517
Uh but when we are attending
at a particular situation,

1480
01:27:06,600 --> 01:27:09,637
we're focused on that situation.
We basically cannot attend to anything

1481
01:27:09,720 --> 01:27:10,517
else.

1482
01:27:10,600 --> 01:27:13,997
And that seems to suggest

1483
01:27:14,080 --> 01:27:17,597
that we basically have only one

1484
01:27:17,680 --> 01:27:21,397
world model engine
in our prefrontal cortex.

1485
01:27:21,480 --> 01:27:25,397
Uh that engine is configurable to the
situation at hand. So, we're building a

1486
01:27:25,480 --> 01:27:29,597
box out of wood, or we are,
you know, driving uh

1487
01:27:29,680 --> 01:27:34,477
down the highway playing chess, we we we
basically have uh a single model of the

1488
01:27:34,560 --> 01:27:37,277
world that we configure
into the situation at hand.

1489
01:27:37,360 --> 01:27:41,117
Which is why we can only
attend to one task at a time.

1490
01:27:41,200 --> 01:27:44,757
Now, if there is a task
that we do repeatedly, um

1491
01:27:44,840 --> 01:27:48,797
it it goes from the sort of deliberate
reasoning using model of the world and

1492
01:27:48,880 --> 01:27:51,517
prediction and perhaps something like
model predictive control, which I was

1493
01:27:51,600 --> 01:27:55,197
talking about earlier,
to something that is more subconscious,

1494
01:27:55,280 --> 01:27:58,917
that becomes automatic. So, I don't know
if you've ever played against a uh chess

1495
01:27:59,000 --> 01:28:03,837
grandmaster, uh you know,
I get wiped out in, you know, 10

1496
01:28:03,920 --> 01:28:07,557
10 plays, right? Uh and,
you know, I have to think about my

1497
01:28:07,640 --> 01:28:11,197
move for, you know, like 15 minutes.

1498
01:28:11,280 --> 01:28:13,397
Uh and the person in front of me, the

1499
01:28:13,480 --> 01:28:18,317
grandmaster, you know, would just
like react within seconds, right? Uh,

1500
01:28:18,400 --> 01:28:21,357
you know, he doesn't need to think
about it. That's become part of his

1501
01:28:21,440 --> 01:28:23,837
subconscious because,
you know, it's basically just

1502
01:28:23,920 --> 01:28:25,477
pattern recognition at this point.

1503
01:28:25,560 --> 01:28:28,517
>> [snorts] >> Um, same,
you know, you the first few

1504
01:28:28,600 --> 01:28:31,517
hours you drive a car, you are really
attentive, you can't do anything else,

1505
01:28:31,600 --> 01:28:35,157
and then after 20,
30 hours of practice, 50 hours,

1506
01:28:35,240 --> 01:28:37,437
you know, it's subconscious, you can
talk to the person next to you, you

1507
01:28:37,520 --> 01:28:38,997
know, things like that, right?

1508
01:28:39,080 --> 01:28:41,757
Unless the situation becomes
unpredictable, and then you have to stop

1509
01:28:41,840 --> 01:28:46,637
talking. Um, so that suggests you
only have one model in your head.

1510
01:28:46,720 --> 01:28:50,837
And it might suggest the idea that
consciousness basically is the module

1511
01:28:50,920 --> 01:28:54,597
that configures this world model of yours.
You know, you need to have some

1512
01:28:54,680 --> 01:28:59,277
sort of uh, executive
kind of overseer that

1513
01:28:59,360 --> 01:29:02,477
configures your world model
for the situation at hand.

1514
01:29:02,560 --> 01:29:06,477
And that that leads to the kind
of the really curious concept that uh,

1515
01:29:06,560 --> 01:29:10,037
consciousness is not a consequence
of the power of our minds, but of the

1516
01:29:10,120 --> 01:29:13,637
limitation of our brains.
That because we have only one world

1517
01:29:13,720 --> 01:29:18,397
model, we have to be conscious.
If we had as many world models as

1518
01:29:18,480 --> 01:29:21,957
there are as situations we encounter,
then we could do all of them

1519
01:29:22,040 --> 01:29:24,997
simultaneously, and we wouldn't need
this sort of executive control that we

1520
01:29:25,080 --> 01:29:28,117
call consciousness. >> Yeah,
interesting. And somehow maybe

1521
01:29:28,200 --> 01:29:32,037
that executive controller,
I mean, the the hard problem of

1522
01:29:32,120 --> 01:29:35,597
consciousness, there's some kind
of chemicals and biology that's creating a

1523
01:29:35,680 --> 01:29:38,877
feeling like it feels
to experience some of

1524
01:29:38,960 --> 01:29:44,397
these things. That that's kind
of like the hard question is

1525
01:29:44,480 --> 01:29:47,237
what the heck [clears throat] is that,
and why is that useful? Maybe the more

1526
01:29:47,320 --> 01:29:51,837
pragmatic question, why is it useful
to feel like this is really you

1527
01:29:51,920 --> 01:29:55,317
experiencing this versus
just like information

1528
01:29:55,400 --> 01:29:59,837
being processed. Um, yeah,
it could be just a very nice

1529
01:29:59,920 --> 01:30:02,277
side effect of um,

1530
01:30:02,360 --> 01:30:04,797
of the way we evolved that's just very

1531
01:30:04,880 --> 01:30:06,957
um, useful to

1532
01:30:07,040 --> 01:30:11,957
to uh, feel a sense of ownership
to the decisions you make, to

1533
01:30:12,040 --> 01:30:15,597
the perceptions you make, to the model
you're trying to maintain. Like you own

1534
01:30:15,680 --> 01:30:18,477
this thing and it's the only
one you got and if you

1535
01:30:18,560 --> 01:30:20,477
lose it, it's going to really suck. And

1536
01:30:20,560 --> 01:30:23,117
so you should really send the brain some

1537
01:30:23,200 --> 01:30:25,757
signals about it.

1538
01:30:25,840 --> 01:30:28,717
What ideas do you believe might be true

1539
01:30:28,800 --> 01:30:33,197
that most or at least many
people disagree with you with?

1540
01:30:33,280 --> 01:30:35,717
Let's say in the space
of machine learning.

1541
01:30:35,800 --> 01:30:38,957
>> Well, depends who you
talk about, but I think

1542
01:30:39,040 --> 01:30:42,797
So certainly there is uh a bunch
of people who are nativist,

1543
01:30:42,880 --> 01:30:45,517
right? Who think that a lot of the basic
things about the world are kind of

1544
01:30:45,600 --> 01:30:47,677
hardwired in our in our minds.

1545
01:30:47,760 --> 01:30:51,037
Um things like you know, the world
is three-dimensional for example. Is that

1546
01:30:51,120 --> 01:30:55,197
hardwired? Things like uh you know,
object permanence. Is it something that we

1547
01:30:55,280 --> 01:30:58,677
learn uh you know,
before the age of 3 months or

1548
01:30:58,760 --> 01:31:03,437
so? Or or are we born with it?
And there are you know, very you know,

1549
01:31:03,520 --> 01:31:08,517
wide disagreement among the you know,
the cognitive scientists uh for this.

1550
01:31:08,600 --> 01:31:11,237
I think those things are
actually very simple to learn.

1551
01:31:11,320 --> 01:31:14,237
Um you know, is it the case that the

1552
01:31:14,320 --> 01:31:18,317
oriented edge detectors in V1 are
learned or are they hardwired? I think

1553
01:31:18,400 --> 01:31:21,437
they are learned. They might be learned
before birth because it's really easy to

1554
01:31:21,520 --> 01:31:24,877
generate signals from the retina that are
actually will train edge detectors.

1555
01:31:24,960 --> 01:31:28,877
So Um and again, those are
things that can be learned within

1556
01:31:28,960 --> 01:31:32,997
minutes of opening your eyes,
right? I mean uh >> [snorts]

1557
01:31:33,080 --> 01:31:37,117
>> you know, since the 1990s we have uh
algorithms that can learn oriented edge

1558
01:31:37,200 --> 01:31:40,717
detectors completely unsupervised with
the equivalent of a few minutes of real

1559
01:31:40,800 --> 01:31:43,877
time. So uh so those
things have to be learned.

1560
01:31:43,960 --> 01:31:47,757
Um and there's also those you know,
MIT experiments where you kind of

1561
01:31:47,840 --> 01:31:50,957
plug the optical nerve
on the auditory cortex of

1562
01:31:51,040 --> 01:31:54,117
a baby ferret, right? And that auditory
cortex become a visual cortex

1563
01:31:54,200 --> 01:31:58,117
essentially. So you know,
clearly um there's learning

1564
01:31:58,200 --> 01:32:01,717
taking place there.
Um so you know, I think a lot of what

1565
01:32:01,800 --> 01:32:05,117
people think are so basic
that they need to be hardwired,

1566
01:32:05,200 --> 01:32:07,677
I think a lot of those things are
learned because they're easy to learn.

1567
01:32:07,760 --> 01:32:12,077
>> Do you So, you put a lot of value
in the power of learning. What

1568
01:32:12,160 --> 01:32:15,277
kind of things do you
suspect might not be learned?

1569
01:32:15,360 --> 01:32:17,957
Is there something
that could not be learned?

1570
01:32:18,040 --> 01:32:22,477
>> So, your intrinsic drives are not
learned. They They They are the things

1571
01:32:22,560 --> 01:32:24,597
that, you know, make humans

1572
01:32:24,680 --> 01:32:29,517
uh human or make, you know,
cats different from dogs, right? It's

1573
01:32:29,600 --> 01:32:34,957
the the basic drives that are kind
of hardwired in our basal ganglia. Um

1574
01:32:35,040 --> 01:32:37,197
I mean, there are people who are working
on on this kind of stuff. It's called

1575
01:32:37,280 --> 01:32:41,037
intrinsic motivation in the context of
reinforcement learning. Um so, these are

1576
01:32:41,120 --> 01:32:44,957
objective functions where the reward
doesn't come from the external world.

1577
01:32:45,040 --> 01:32:48,357
It's computed by your own brain.
Your own brain computes

1578
01:32:48,440 --> 01:32:52,797
whether you're happy or not, right?
It measures your degree of uh

1579
01:32:52,880 --> 01:32:57,317
comfort or discomfort.
Uh and and because it's your brain

1580
01:32:57,400 --> 01:33:00,437
computing this, presumably it knows also
how to estimate gradient and so that's

1581
01:33:00,520 --> 01:33:02,597
right. So, um

1582
01:33:02,680 --> 01:33:05,397
uh so, it's easier to to learn when your

1583
01:33:05,480 --> 01:33:08,237
objective is is intrinsic.

1584
01:33:08,320 --> 01:33:11,517
Um so, that has to be hardwired.

1585
01:33:11,600 --> 01:33:15,677
Uh the critic that makes long-term
prediction of the outcome, which is the

1586
01:33:15,760 --> 01:33:19,317
eventual result of this, that's learned.

1587
01:33:19,400 --> 01:33:23,277
Uh and perception is learned and your
model of the world is learned. But, let

1588
01:33:23,360 --> 01:33:26,917
me take take an example of, you know,
why the critic uh I mean, example of why

1589
01:33:27,000 --> 01:33:31,757
the critic may be learned, right?
If I uh if I come to you um

1590
01:33:31,840 --> 01:33:35,477
you know, I reach across the table
and I I pinch your arm, right? Complete

1591
01:33:35,560 --> 01:33:38,197
surprise for you. You would
not have expected this from

1592
01:33:38,280 --> 01:33:40,877
>> expecting that the whole time, but yes,
right. [laughter] Let's say for the the

1593
01:33:40,960 --> 01:33:42,997
sake of the story, yes.

1594
01:33:43,080 --> 01:33:49,597
>> Um okay, your basal ganglia is going
to light up cuz it's going to hurt, right?

1595
01:33:49,680 --> 01:33:51,997
Uh >> [snorts] >> and now
your model of the world includes

1596
01:33:52,080 --> 01:33:55,917
the fact that I may pinch
you if I approach my uh

1597
01:33:56,000 --> 01:33:59,357
my uh >> Don't trust humans.
Right. My hand to

1598
01:33:59,440 --> 01:34:04,037
your arm. So, if I try again, you're
going to recoil. And that's your critic,

1599
01:34:04,120 --> 01:34:09,557
uh your predictive you know,
your predictor of your uh ultimate pain

1600
01:34:09,640 --> 01:34:13,717
uh uh system that predicts
that something bad is

1601
01:34:13,800 --> 01:34:15,757
going to happen and you
recoil to avoid it.

1602
01:34:15,840 --> 01:34:17,077
>> So, even that can be learned.

1603
01:34:17,160 --> 01:34:20,797
>> That is learned, definitely.
This is what allows you also to

1604
01:34:20,880 --> 01:34:22,517
uh you know, define some goals, right?

1605
01:34:22,600 --> 01:34:25,637
So, um the fact that you know, you're a

1606
01:34:25,720 --> 01:34:29,877
school child, you wake up in the morning
and you go to school and you know,

1607
01:34:29,960 --> 01:34:34,957
um it's not because you necessarily like
waking up early and going to school, but

1608
01:34:35,040 --> 01:34:37,797
you know that there is a long-term
objective you're trying to optimize.

1609
01:34:37,880 --> 01:34:40,517
>> So, Ernest Becker, I'm not sure
if you're familiar with him. He's the

1610
01:34:40,600 --> 01:34:44,077
philosopher who wrote the book
The Denial of Death. And his idea is that

1611
01:34:44,160 --> 01:34:47,797
one of the core motivations of human
beings is our terror of death, our fear

1612
01:34:47,880 --> 01:34:50,397
of death. Uh that's
what makes us unique from

1613
01:34:50,480 --> 01:34:52,717
cats. Cats are just surviving. They do

1614
01:34:52,800 --> 01:34:54,317
not have a

1615
01:34:54,400 --> 01:34:58,637
deep under- like um cognizance,

1616
01:34:58,720 --> 01:35:03,757
introspection that over
the horizon is the end.

1617
01:35:03,840 --> 01:35:06,837
And he says that I mean, there's
a terror management theory that just all

1618
01:35:06,920 --> 01:35:11,117
these psychological experiments
that show the basically

1619
01:35:11,200 --> 01:35:13,437
this idea that

1620
01:35:13,520 --> 01:35:15,637
all of human civilization, everything we

1621
01:35:15,720 --> 01:35:21,157
create is kind of uh trying to forget if
even for a brief moment that we're going

1622
01:35:21,240 --> 01:35:25,677
to die. When When do you
think humans understand

1623
01:35:25,760 --> 01:35:28,757
that they're going to die?
Is it learned early on

1624
01:35:28,840 --> 01:35:33,397
also, like >> I don't know
to at what point I mean,

1625
01:35:33,480 --> 01:35:36,797
it's a it's a question like, you know,
at what point do you realize that, you

1626
01:35:36,880 --> 01:35:40,197
know, what death really is. And I think
most people don't actually realize what

1627
01:35:40,280 --> 01:35:42,877
death is, right? I mean, most people
believe that you go to heaven or

1628
01:35:42,960 --> 01:35:46,437
something, right? >> Well,
so the the pushback on that what

1629
01:35:46,520 --> 01:35:51,237
Ernest Becker says and um Sheldon
Solomon, all of those folks,

1630
01:35:51,320 --> 01:35:54,877
and I find those ideas a little bit
compelling is that there is moments in

1631
01:35:54,960 --> 01:35:58,517
life early in life.
A lot of this fun happens early in life

1632
01:35:58,600 --> 01:36:00,837
when you are

1633
01:36:00,920 --> 01:36:05,957
uh when you do deeply experience the
terror of this realization. And all the

1634
01:36:06,040 --> 01:36:09,517
things you think about about religion,
all those kinds of things that we kind

1635
01:36:09,600 --> 01:36:14,037
of think about more like teenage years and
later, we're talking about way earlier.

1636
01:36:14,120 --> 01:36:15,957
>> No, it's like 7 or 8 years,
something like that. Yeah.

1637
01:36:16,040 --> 01:36:20,717
>> You realize holy crap, this is

1638
01:36:20,800 --> 01:36:25,077
uh like the mystery, the terror. Like
it's almost like you're a little prey, a

1639
01:36:25,160 --> 01:36:29,397
little baby deer sitting in the darkness
of the jungle, of the woods, looking all

1640
01:36:29,480 --> 01:36:33,117
around you. There's darkness full
of terror. I mean, that's that realization

1641
01:36:33,200 --> 01:36:37,437
says, "Okay, I'm going to go go back
in the comfort of my mind where there is a

1642
01:36:37,520 --> 01:36:42,397
where there is a deep meaning, where there
is a maybe like pretend I'm immortal in

1643
01:36:42,480 --> 01:36:47,117
however way however kind
of idea I can construct to

1644
01:36:47,200 --> 01:36:49,037
help me understand that I'm immortal."

1645
01:36:49,120 --> 01:36:52,677
Religion helps with that. You can you
can delude yourself in all kinds of

1646
01:36:52,760 --> 01:36:57,317
ways. Like lose yourself in the busyness
of each day, have little goals in mind,

1647
01:36:57,400 --> 01:37:00,397
all those kinds of things to think
that it's going to go on forever. And you

1648
01:37:00,480 --> 01:37:04,317
kind of know you're going to die, yeah,
and it's going to be sad, but you don't

1649
01:37:04,400 --> 01:37:08,317
really understand that you're going
to die. And so that's that's their idea.

1650
01:37:08,400 --> 01:37:10,517
And if I find that compelling

1651
01:37:10,600 --> 01:37:13,437
because it does seem to be a core unique

1652
01:37:13,520 --> 01:37:17,717
aspect of human nature that we we're
able to think that we're go we're able

1653
01:37:17,800 --> 01:37:22,597
to really understand that this life
is finite. I find that important.

1654
01:37:22,680 --> 01:37:23,997
>> This There's a bunch
of different things

1655
01:37:24,080 --> 01:37:26,557
there. So, first of all, I don't
think there is a qualitative difference

1656
01:37:26,640 --> 01:37:29,397
between between us and cats in the term.

1657
01:37:29,480 --> 01:37:33,477
I think the difference is that we
just have a better long-term

1658
01:37:33,560 --> 01:37:37,957
ability to predict, you know,
in the long term. And so, we have better

1659
01:37:38,040 --> 01:37:39,237
understanding of how the world works.

1660
01:37:39,320 --> 01:37:42,517
So, we have better understanding of, you
know, finiteness of life and things like

1661
01:37:42,600 --> 01:37:45,437
that. >> So, we have a better
planning engine than cats?

1662
01:37:45,520 --> 01:37:49,877
>> Yeah. >> Okay. >> Um you know,
perhaps >> what's the motivation for

1663
01:37:49,960 --> 01:37:52,077
>> for planning that far >> Well,
I think it's just a side effect of

1664
01:37:52,160 --> 01:37:55,917
the fact that we have just a better
planning engine because it makes us

1665
01:37:56,000 --> 01:37:59,277
as I said, you know, the essence
of intelligence is the ability to predict

1666
01:37:59,360 --> 01:38:03,637
and so the because we're smarter
as a side effect we also have this

1667
01:38:03,720 --> 01:38:09,437
ability to kind of make predictions about
our own future existence or lack thereof.

1668
01:38:09,520 --> 01:38:14,477
>> Okay. >> Um you say religion helps with
that. I think religion hurts actually.

1669
01:38:14,560 --> 01:38:17,917
Um It makes people worry about like, you
know, what's going to happen after their

1670
01:38:18,000 --> 01:38:22,877
death, etc. If you believe that, you know,
you just don't exist after death. Like,

1671
01:38:22,960 --> 01:38:24,837
you know, it solves completely
the problem. At least

1672
01:38:24,920 --> 01:38:28,317
>> you're saying if you don't believe
in God, you don't worry about what happens

1673
01:38:28,400 --> 01:38:31,077
after death? >> Yeah. >> I don't know.

1674
01:38:31,160 --> 01:38:36,157
>> worry about the about you know, this
life because that's the only one you have.

1675
01:38:36,240 --> 01:38:39,117
>> I think it's well, I don't know I
don't know. If I were to say what Ernest

1676
01:38:39,200 --> 01:38:42,837
Becker says then I'll
say I agree with him more

1677
01:38:42,920 --> 01:38:47,837
than not is um you do deeply worry.

1678
01:38:47,920 --> 01:38:52,357
Uh if you if you believe there's no God,
there's still a deep worry like of the

1679
01:38:52,440 --> 01:38:54,357
mystery of it all. Like, how does that

1680
01:38:54,440 --> 01:38:57,637
make any sense that it just ends? I

1681
01:38:57,720 --> 01:39:00,837
don't think we can truly understand that

1682
01:39:00,920 --> 01:39:02,997
this right I mean, so much of our life,

1683
01:39:03,080 --> 01:39:06,317
the consciousness, the ego is invested

1684
01:39:06,400 --> 01:39:09,157
in this in this being and then

1685
01:39:09,240 --> 01:39:13,477
>> Science keeps bringing
humanity down from its pedestal.

1686
01:39:13,560 --> 01:39:16,717
And that's that's just
another another example of it.

1687
01:39:16,800 --> 01:39:20,957
>> That's wonderful, but for us individual
humans, we don't like to be brought down

1688
01:39:21,040 --> 01:39:25,437
from a pedestal. You're saying like
what but see you're fine with it because

1689
01:39:25,520 --> 01:39:28,957
well, so what Ernest Becker would say
is you're fine with it because that's just

1690
01:39:29,040 --> 01:39:32,277
a more peaceful existence for you,
but you're not really fine. You're hiding

1691
01:39:32,360 --> 01:39:34,037
from In fact, some of the people that

1692
01:39:34,120 --> 01:39:37,237
experienced the deepest trauma that it

1693
01:39:37,320 --> 01:39:39,917
earlier in life, they often before they

1694
01:39:40,000 --> 01:39:44,317
to extensive therapy will say that I'm
fine. It's like when you talk to people

1695
01:39:44,400 --> 01:39:46,037
who are truly angry, how are you doing?

1696
01:39:46,120 --> 01:39:49,757
I'm fine. The The question is,
what's going on?

1697
01:39:49,840 --> 01:39:53,717
>> Now, I had a near-death
experience. I had a very bad

1698
01:39:53,800 --> 01:39:56,517
motorbike accident when I was 17, so.

1699
01:39:56,600 --> 01:40:01,197
>> And >> Uh but that didn't
have any impact on my

1700
01:40:01,280 --> 01:40:04,197
reflection on that topic. >> So,
I'm I'm basically just playing a bit

1701
01:40:04,280 --> 01:40:07,837
of devil's advocate
pushing back and wondering

1702
01:40:07,920 --> 01:40:09,517
is it truly possible to accept death?

1703
01:40:09,600 --> 01:40:13,397
And the flip side that's more
interesting, I think, for AI and

1704
01:40:13,480 --> 01:40:16,477
robotics, is how important is it to have

1705
01:40:16,560 --> 01:40:19,077
this as one of the suite of motivations

1706
01:40:19,160 --> 01:40:24,117
is to uh not just avoid

1707
01:40:24,200 --> 01:40:27,357
falling off the roof
or something like that, but

1708
01:40:27,440 --> 01:40:28,797
ponder

1709
01:40:28,880 --> 01:40:31,797
the the end of the ride.

1710
01:40:31,880 --> 01:40:34,557
Uh if you listen to the Stoics,

1711
01:40:34,640 --> 01:40:37,357
it's uh it's a great motivator. It adds

1712
01:40:37,440 --> 01:40:40,477
a sense of urgency. So, maybe a to truly

1713
01:40:40,560 --> 01:40:43,917
fear death or be cognizant of it might

1714
01:40:44,000 --> 01:40:49,597
give uh a deeper meaning
and urgency to the moment to live

1715
01:40:49,680 --> 01:40:54,317
fully. >> I may I may Maybe
I don't disagree with that.

1716
01:40:54,400 --> 01:40:58,797
I mean, I think what
motivates me here is uh

1717
01:40:58,880 --> 01:41:02,117
you know, knowing more about
about human nature. I mean, I think

1718
01:41:02,200 --> 01:41:05,837
human nature and human intelligence is
a big mystery. It's a scientific mystery.

1719
01:41:05,920 --> 01:41:08,437
>> Mhm. [snorts] >> Uh
in addition to, you know,

1720
01:41:08,520 --> 01:41:13,957
philosophical and etc., but, you know,
I'm a true believer in science, so

1721
01:41:14,040 --> 01:41:18,077
um I and and and I do
have kind of a belief

1722
01:41:18,160 --> 01:41:21,917
that for complex systems like
like the brain and the mind,

1723
01:41:22,000 --> 01:41:24,477
the the way to

1724
01:41:24,560 --> 01:41:28,997
understand it is to try to reproduce it
with, you know, artifacts that you build

1725
01:41:29,080 --> 01:41:32,157
because you know what's essential to it
when you try to build it. You know, the

1726
01:41:32,240 --> 01:41:36,317
same way um I've used this analogy
before with you, I believe. Um

1727
01:41:36,400 --> 01:41:39,437
the same way [clears throat] we
we only started to understand uh

1728
01:41:39,520 --> 01:41:42,357
aerodynamics when we started building
airplanes and that helped us understand

1729
01:41:42,440 --> 01:41:45,557
how birds fly. Uh, you know,
so I think there's kind of

1730
01:41:45,640 --> 01:41:49,997
a a similar process here where we
don't have a theory of a full theory

1731
01:41:50,080 --> 01:41:53,077
of intelligence but building,
you know, intelligent

1732
01:41:53,160 --> 01:41:56,397
artifacts will help
us perhaps develop some

1733
01:41:56,480 --> 01:42:00,677
you know, underlying theory
that encompasses not just artificial

1734
01:42:00,760 --> 01:42:05,797
implements but also uh human
and biological intelligence in general.

1735
01:42:05,880 --> 01:42:07,117
>> So, you're an interesting person to ask

1736
01:42:07,200 --> 01:42:09,997
this question about sort of all kinds of

1737
01:42:10,080 --> 01:42:13,317
different other intelligent entities

1738
01:42:13,400 --> 01:42:17,397
or intelligences. What are
your thoughts about kind of

1739
01:42:17,480 --> 01:42:19,117
like the Turing or the Chinese room

1740
01:42:19,200 --> 01:42:21,157
question?

1741
01:42:21,240 --> 01:42:24,837
If we create an AI system that exhibits

1742
01:42:24,920 --> 01:42:29,477
a lot of properties
of intelligence and consciousness,

1743
01:42:29,560 --> 01:42:34,277
how comfortable are you thinking of
that entity as intelligent or conscious?

1744
01:42:34,360 --> 01:42:37,597
So, so you're trying to build novel
systems that have intelligence and

1745
01:42:37,680 --> 01:42:42,237
there's metrics about their
performance but that metric is

1746
01:42:42,320 --> 01:42:46,957
external. >> Okay. >> So,
how are you are you okay calling a

1747
01:42:47,040 --> 01:42:50,957
thing intelligent or you
going to be like most humans

1748
01:42:51,040 --> 01:42:56,277
and be once again unhappy to be brought
down from a pedestal of consciousness /

1749
01:42:56,360 --> 01:43:02,157
intelligence? >> No, I'm I'll
be very happy to understand

1750
01:43:02,240 --> 01:43:05,677
uh more about uh human nature, human mind,

1751
01:43:05,760 --> 01:43:10,117
and human intelligence through
the construction of machines that

1752
01:43:10,200 --> 01:43:14,157
uh have similar abilities. And if a

1753
01:43:14,240 --> 01:43:18,157
consequence of this is to bring
down humanity one notch down from

1754
01:43:18,240 --> 01:43:21,437
its already low pedestal, I'm just
[laughter] fine with it. That's just

1755
01:43:21,520 --> 01:43:25,917
reality of life. Um, so I'm fine with
that. Now, you're asking me about things

1756
01:43:26,000 --> 01:43:29,797
that uh opinions I have that a lot
of people may disagree with.

1757
01:43:29,880 --> 01:43:34,797
I think uh if we think about the design of

1758
01:43:34,880 --> 01:43:38,717
autonomous intelligence systems, so
assuming that we are somewhat successful

1759
01:43:38,800 --> 01:43:41,957
at some at some level of getting
machines to learn models of

1760
01:43:42,040 --> 01:43:46,117
the world, predictive models of the world,
uh we have we build intrinsic

1761
01:43:46,200 --> 01:43:50,317
motivation objective functions to drive
the behavior of that system. Uh the

1762
01:43:50,400 --> 01:43:53,957
system also has perception modules that
allows it to estimate the state of the

1763
01:43:54,040 --> 01:43:57,757
world and then have some way of figuring
out a sequence of actions that, you

1764
01:43:57,840 --> 01:44:01,197
know, to optimize a particular objective.

1765
01:44:01,280 --> 01:44:05,357
If it has a critic of the type that I
was describing before, the thing that

1766
01:44:05,440 --> 01:44:10,477
makes you recoil your arm the second
time I tried to pinch you, um

1767
01:44:10,560 --> 01:44:13,597
intelligent autonomous
machine will have emotions.

1768
01:44:13,680 --> 01:44:16,237
I think emotions are an integral part of

1769
01:44:16,320 --> 01:44:18,317
autonomous intelligence.

1770
01:44:18,400 --> 01:44:20,917
If you have an intelligent system

1771
01:44:21,000 --> 01:44:25,917
that is driven by uh intrinsic
motivation, by objectives,

1772
01:44:26,000 --> 01:44:28,877
if it has a critic that
allows it to predict in

1773
01:44:28,960 --> 01:44:32,517
advance whether the outcome of a of
a situation is going to be good or

1774
01:44:32,600 --> 01:44:36,117
bad, it's going to have emotions.
It's going to have fear >> Yes.

1775
01:44:36,200 --> 01:44:40,037
>> uh when it predicts that the outcome
is going to is going to be bad

1776
01:44:40,120 --> 01:44:43,117
and and something to avoid, it's going
to have elation when it predicts that

1777
01:44:43,200 --> 01:44:44,837
it's going to be good. >> [snorts]

1778
01:44:44,920 --> 01:44:48,677
>> Um uh if it has drives to relate with

1779
01:44:48,760 --> 01:44:53,117
humans, um you know,
in some ways, the way humans have,

1780
01:44:53,200 --> 01:44:57,797
um you know, it's it's going to be social,
right? And so it's going to have

1781
01:44:57,880 --> 01:45:01,437
emotions about attachment
and and things of that type. So,

1782
01:45:01,520 --> 01:45:06,197
um so I think uh you know,
the the sort of sci-fi

1783
01:45:06,280 --> 01:45:09,957
thing where, you know, you see Commander
Data like having an emotion chip that he

1784
01:45:10,040 --> 01:45:13,637
can turn off, right?
Um I think that's ridiculous.

1785
01:45:13,720 --> 01:45:18,357
>> So, I mean, here's
the difficult philosophical social

1786
01:45:18,440 --> 01:45:21,997
question. Do you think
there will be a time

1787
01:45:22,080 --> 01:45:25,517
like a civil rights
movement for robots where

1788
01:45:25,600 --> 01:45:27,517
um okay, forget the movement, but a

1789
01:45:27,600 --> 01:45:31,677
discussion at the Supreme Court

1790
01:45:31,760 --> 01:45:34,797
that particular kinds of robots,

1791
01:45:34,880 --> 01:45:37,557
you know, particular kinds of systems,

1792
01:45:37,640 --> 01:45:42,957
um, deserve the same rights as humans
because they can suffer just as humans

1793
01:45:43,040 --> 01:45:46,677
can. Um, all those kinds of things.

1794
01:45:46,760 --> 01:45:48,077
>> Well,

1795
01:45:48,160 --> 01:45:49,917
perhaps, perhaps not. Like imagine that

1796
01:45:50,000 --> 01:45:53,797
humans were that that you could, uh,

1797
01:45:53,880 --> 01:45:57,677
you know, die and be restored. Like, you
know, you could be sort of, you know, be

1798
01:45:57,760 --> 01:46:01,757
3D reprinted and, you know, your brain
could be reconstructed in its finest

1799
01:46:01,840 --> 01:46:04,157
details. >> Our ideas
of rights will change in that

1800
01:46:04,240 --> 01:46:07,717
case. If you can always just >> [laughter]

1801
01:46:07,800 --> 01:46:11,637
>> there's always a backup, you can always
restore, maybe like the the importance

1802
01:46:11,720 --> 01:46:13,877
of murder will go down one notch.

1803
01:46:13,960 --> 01:46:17,717
>> That's right.
But also the, uh, your your

1804
01:46:17,800 --> 01:46:19,437
you know, desire to do dangerous things

1805
01:46:19,520 --> 01:46:20,877
like, you know,

1806
01:46:20,960 --> 01:46:25,197
you know, skydiving or you know, uh,

1807
01:46:25,280 --> 01:46:28,917
or or or, you know, race car driving,
you know, car racing or that kind of

1808
01:46:29,000 --> 01:46:31,357
stuff, you know, would probably increase.

1809
01:46:31,440 --> 01:46:34,757
Or or, you know, airplane aerobatics
or that kind of stuff, right? You would be

1810
01:46:34,840 --> 01:46:38,757
fine to do a lot of those things
or explore, you know, dangerous areas and

1811
01:46:38,840 --> 01:46:42,677
things like that. It would kind
of change your relationship. So, now, it's

1812
01:46:42,760 --> 01:46:47,197
very likely that robots would be like
that because, you know, they'll be

1813
01:46:47,280 --> 01:46:51,077
based on perhaps technology
that is somewhat similar to

1814
01:46:51,160 --> 01:46:54,197
uh, today's technology and you
can you can always have a backup.

1815
01:46:54,280 --> 01:46:57,597
>> So, it's possible I don't
know if you like video games,

1816
01:46:57,680 --> 01:47:01,237
but there's a there's
a game called Diablo and, um,

1817
01:47:01,320 --> 01:47:03,797
>> Oh, my my sons are huge fans of this.

1818
01:47:03,880 --> 01:47:07,517
>> Yes. Uh, and >> In fact,
they made a game that's

1819
01:47:07,600 --> 01:47:10,917
inspired by it. >> Awesome.
Like built a game?

1820
01:47:11,000 --> 01:47:14,557
>> My my three sons have a a game
design studio between them, yeah.

1821
01:47:14,640 --> 01:47:17,077
>> That's awesome. >> They came out with
a game >> Like just came out with a game?

1822
01:47:17,160 --> 01:47:20,077
>> Last year. No, this was last year,
early this year, about a year ago.

1823
01:47:20,160 --> 01:47:23,957
>> That's awesome. But so, in Diablo,
there's a something called hardcore mode

1824
01:47:24,040 --> 01:47:27,797
which if you die, there's no you're gone.

1825
01:47:27,880 --> 01:47:30,637
>> Right. >> That's it.
And so it's possible with AI

1826
01:47:30,720 --> 01:47:34,477
systems for them to be able to operate

1827
01:47:34,560 --> 01:47:38,237
successfully and for us to treat them in
a certain way for cuz if they have to be

1828
01:47:38,320 --> 01:47:40,197
integrated in human society,

1829
01:47:40,280 --> 01:47:43,077
they have to be able to die, no copies

1830
01:47:43,160 --> 01:47:45,757
allowed. In fact, copying is illegal.

1831
01:47:45,840 --> 01:47:49,397
It's possible with humans as well,
like cloning will be illegal even when

1832
01:47:49,480 --> 01:47:51,917
possible. >> but cloning
is not copying, right? I

1833
01:47:52,000 --> 01:47:56,357
mean, you don't reproduce the the
the mind of the person and experience.

1834
01:47:56,440 --> 01:47:58,237
>> It's just a delayed twin, so.

1835
01:47:58,320 --> 01:48:01,397
>> But then it's well but we were talking
about with computers that you will be

1836
01:48:01,480 --> 01:48:03,357
able to copy. You You will be able to

1837
01:48:03,440 --> 01:48:07,637
perfectly save, pickle the the the mind

1838
01:48:07,720 --> 01:48:10,357
state. And it's possible that that will

1839
01:48:10,440 --> 01:48:15,157
be illegal because that goes against um

1840
01:48:15,240 --> 01:48:17,797
that will destroy
the motivations of the system.

1841
01:48:17,880 --> 01:48:22,117
>> Okay, so let's say you you
have a domestic robot, okay,

1842
01:48:22,200 --> 01:48:24,277
sometime in the future. >> Yes.

1843
01:48:24,360 --> 01:48:27,197
>> And uh the domestic
robot, you know, comes to

1844
01:48:27,280 --> 01:48:30,957
you kind of somewhat pre-trained,
you know, it can do a bunch of things, but

1845
01:48:31,040 --> 01:48:33,597
it has a particular personality that
makes it slightly different from the

1846
01:48:33,680 --> 01:48:37,237
other robots because that makes them
more interesting. And then because it's,

1847
01:48:37,320 --> 01:48:40,877
you know, it's lived with you for 5
years, you've you've grown some

1848
01:48:40,960 --> 01:48:44,197
attachment to it and vice versa, and

1849
01:48:44,280 --> 01:48:48,877
it's learned a lot about you. Or maybe
it's not a household robot, maybe it's

1850
01:48:48,960 --> 01:48:52,037
maybe it's a virtual assistant
that lives in your, you know, augmented

1851
01:48:52,120 --> 01:48:53,757
reality glasses or whatever, right?

1852
01:48:53,840 --> 01:48:56,037
>> Yeah. >> Uh you know,
the the Her movie type

1853
01:48:56,120 --> 01:48:58,517
thing, right? Um

1854
01:48:58,600 --> 01:49:02,557
and that system to some extent the the

1855
01:49:02,640 --> 01:49:06,877
intelligence in that system is a bit
like your child or maybe your PhD

1856
01:49:06,960 --> 01:49:11,797
student in the sense that there's a lot
of you in that in that machine now, right?

1857
01:49:11,880 --> 01:49:15,757
>> Yeah. >> And so if it
were a living thing, you

1858
01:49:15,840 --> 01:49:19,677
would do this for free if you want,
right? If it's your child, your child

1859
01:49:19,760 --> 01:49:23,997
can, you know, then uh live
his or her own life and, you

1860
01:49:24,080 --> 01:49:26,997
know, the fact that they learned stuff
from you doesn't mean that you have any

1861
01:49:27,080 --> 01:49:29,877
ownership of it, right?
But if it's a robot

1862
01:49:29,960 --> 01:49:33,677
uh that you've trained,
perhaps you have some

1863
01:49:33,760 --> 01:49:37,597
intellectual property claim about >>
Intellectual property? Oh, I thought you

1864
01:49:37,680 --> 01:49:42,077
meant like uh permanent value
in the sense that it's part of you is in

1865
01:49:42,160 --> 01:49:43,597
>> Well, there is permanent value, right?

1866
01:49:43,680 --> 01:49:47,237
So, you would lose a lot if that robot
were to be destroyed and you you had no

1867
01:49:47,320 --> 01:49:49,877
backup, you would lose a lot. You know,
you would lose a lot of investment, you

1868
01:49:49,960 --> 01:49:53,517
know, kind of like a uh you know,
a person dying, you know?

1869
01:49:53,600 --> 01:49:57,917
Um that that uh a friend of a friend of
yours dying or or co-worker or something

1870
01:49:58,000 --> 01:50:00,397
like that. Um [snorts]

1871
01:50:00,480 --> 01:50:03,237
>> But also uh you have like intellectual

1872
01:50:03,320 --> 01:50:08,157
property rights in the sense that
that that system is fine-tuned to your

1873
01:50:08,240 --> 01:50:12,597
particular existence, so that's now
a very unique instantiation of that

1874
01:50:12,680 --> 01:50:16,197
original background model,
whatever it was, that arrived.

1875
01:50:16,280 --> 01:50:19,957
>> And then there are issues of privacy,
right? Because now imagine that that

1876
01:50:20,040 --> 01:50:24,197
robot has its own kind of volition
and decides to work for someone else.

1877
01:50:24,280 --> 01:50:29,237
>> Yes. >> Or kind of, you know, thinks
uh life with you is sort of untenable or

1878
01:50:29,320 --> 01:50:33,437
whatever. >> Right. >> Um >> [snorts]
>> Now, all the all the things that that

1879
01:50:33,520 --> 01:50:37,957
system learned from you, uh you know, how

1880
01:50:38,040 --> 01:50:40,557
can you like, you know, delete all
the personal information that that system

1881
01:50:40,640 --> 01:50:43,717
knows about you? >> Yeah. >> Uh I mean,
that would be kind of an ethical

1882
01:50:43,800 --> 01:50:45,317
question, like, you know, can you erase

1883
01:50:45,400 --> 01:50:51,717
the the mind of a of a intelligent
robot uh to protect your your privacy?

1884
01:50:51,800 --> 01:50:53,477
>> Yeah. >> You can't do this with humans.

1885
01:50:53,560 --> 01:50:57,557
You can ask them to shut up, but that you
don't have complete power over them.

1886
01:50:57,640 --> 01:51:00,637
>> Can't erase humans. Yeah, that's
the problem with the relationships, you

1887
01:51:00,720 --> 01:51:04,597
know? If they you break up,
you can't you can't erase the other human.

1888
01:51:04,680 --> 01:51:08,397
With robots, I think it'll have to be
the same thing with robots. That that

1889
01:51:08,480 --> 01:51:11,957
risk that that has to be um,

1890
01:51:12,040 --> 01:51:15,797
some risk to our interactions to truly

1891
01:51:15,880 --> 01:51:18,077
experience them deeply, it feels like.

1892
01:51:18,160 --> 01:51:21,517
So, you have to be able
to lose your robot friend.

1893
01:51:21,600 --> 01:51:25,597
And that robot friend to go tweeting
about how much of an you were.

1894
01:51:25,680 --> 01:51:30,077
>> But then, are you allowed to, you know,
murder the robot to protect your private

1895
01:51:30,160 --> 01:51:31,877
information if the robot decides to leave?

1896
01:51:31,960 --> 01:51:36,437
>> I have this intuition
that for robots with with certain

1897
01:51:36,520 --> 01:51:40,357
like it's it's almost like, uh,
regulation. If you declare your robot to

1898
01:51:40,440 --> 01:51:44,477
be, let's call it sentient or something
like that. Like, this this robot is

1899
01:51:44,560 --> 01:51:47,437
designed for human interaction,
then you're not allowed to murder these

1900
01:51:47,520 --> 01:51:50,077
robots, it's the same
as murdering other humans.

1901
01:51:50,160 --> 01:51:53,957
>> Well, but what about you do a backup of
the robot you do preserve on the on a on

1902
01:51:54,040 --> 01:51:55,757
a hard drive or the
equivalent in the future?

1903
01:51:55,840 --> 01:51:59,957
>> That might be illegal. Just like
it's like piracy piracy is illegal.

1904
01:52:00,040 --> 01:52:01,717
>> No, but it's your own
it's your own robot, right?

1905
01:52:01,800 --> 01:52:06,477
>> But you can't you don't >> But then
but then you can wipe out, >> [laughter]

1906
01:52:06,560 --> 01:52:09,757
>> its brain so the this robot doesn't
know anything about you anymore, but you

1907
01:52:09,840 --> 01:52:13,557
still have technically it's still
in existence because you backed it up.

1908
01:52:13,640 --> 01:52:17,317
>> And then there'll be these great
speeches at the Supreme Court by saying,

1909
01:52:17,400 --> 01:52:20,997
"Oh, sure you can erase the mind of the
robot just like you can erase the mind

1910
01:52:21,080 --> 01:52:22,917
of a human. We both can suffer."

1911
01:52:23,000 --> 01:52:24,757
There'll be some epic like Obama-type

1912
01:52:24,840 --> 01:52:27,837
character with a speech that we we like

1913
01:52:27,920 --> 01:52:30,357
the robots and the humans are the same.

1914
01:52:30,440 --> 01:52:34,717
Uh, we can both suffer,
we can both hope. We can both

1915
01:52:34,800 --> 01:52:36,757
all of those all those kinds of things.

1916
01:52:36,840 --> 01:52:39,117
Raise families, all that kind of stuff.

1917
01:52:39,200 --> 01:52:41,157
It's it's, uh, interesting for these

1918
01:52:41,240 --> 01:52:42,877
just like you said, emotion seems to be

1919
01:52:42,960 --> 01:52:46,797
a fascinatingly powerful aspect of human

1920
01:52:46,880 --> 01:52:50,237
human interaction, human robot
interaction. And if they're able to

1921
01:52:50,320 --> 01:52:54,197
exhibit emotions at the end
of the day, that's probably

1922
01:52:54,280 --> 01:52:57,797
going to have us deeply consider

1923
01:52:57,880 --> 01:53:02,237
human rights. Like what we value
in humans, what we value in other animals.

1924
01:53:02,320 --> 01:53:06,317
That's why robots and AI is great. It
makes us ask, uh, really good questions.

1925
01:53:06,400 --> 01:53:10,357
>> questions, yeah. But you I mean you
ask about you asked about the Chinese room

1926
01:53:10,440 --> 01:53:12,877
type argument, you know,
is it real if it looks real?

1927
01:53:12,960 --> 01:53:16,997
>> Yeah. >> I think the Chinese
room argument is a ridiculous one. So

1928
01:53:17,080 --> 01:53:19,637
>> So so for people don't
know, Chinese room is

1929
01:53:19,720 --> 01:53:21,917
uh you you can I don't even know how to

1930
01:53:22,000 --> 01:53:26,837
formulate it well, but basically
you can mimic the behavior of an

1931
01:53:26,920 --> 01:53:28,717
intelligent system by just following a

1932
01:53:28,800 --> 01:53:32,037
giant algorithm codebook that tells you

1933
01:53:32,120 --> 01:53:36,597
exactly how to respond in exactly each
case, but is that really intelligent?

1934
01:53:36,680 --> 01:53:40,437
It's like a giant lookup table. When
this person says this, you answer this.

1935
01:53:40,520 --> 01:53:42,917
When this person says
this, you answer this.

1936
01:53:43,000 --> 01:53:46,277
And um if you understand
how that works, you

1937
01:53:46,360 --> 01:53:50,677
have this giant nearly infinite lookup
table, is that really intelligence? Cuz

1938
01:53:50,760 --> 01:53:55,277
intelligence seems to be a mechanism
that's much more interesting and complex

1939
01:53:55,360 --> 01:53:56,517
than this lookup table.

1940
01:53:56,600 --> 01:53:58,997
>> I don't think so.
So the I mean the real

1941
01:53:59,080 --> 01:54:03,077
question comes down to do you think uh

1942
01:54:03,160 --> 01:54:06,957
you know, you can you can make an AI's
uh intelligence in some way, even if if

1943
01:54:07,040 --> 01:54:11,317
that involves uh learning.
And the answer is of course yes. There's

1944
01:54:11,400 --> 01:54:15,197
no question. Uh there's
a second question then, which is

1945
01:54:15,280 --> 01:54:19,397
uh assuming you can uh reproduce
intelligence in sort of different

1946
01:54:19,480 --> 01:54:24,317
hardware than biological hardware,
you know, what I computers, um

1947
01:54:24,400 --> 01:54:26,517
can you uh you know

1948
01:54:26,600 --> 01:54:30,237
match uh human intelligence in

1949
01:54:30,320 --> 01:54:34,757
all the domains in which
humans are intelligent?

1950
01:54:34,840 --> 01:54:39,037
Is it possible, right? So this is the
quite the hypothesis of strong AI. The

1951
01:54:39,120 --> 01:54:42,837
answer to this in my opinion
is a unqualified yes. This

1952
01:54:42,920 --> 01:54:44,517
will this will happen at some point.

1953
01:54:44,600 --> 01:54:47,997
There's no question that machines
at some point will become more intelligent

1954
01:54:48,080 --> 01:54:51,517
than humans in all domains where
humans are intelligent. This is not for

1955
01:54:51,600 --> 01:54:56,637
tomorrow. It's going to take a long time
uh [snorts] regardless of what, you know,

1956
01:54:56,720 --> 01:55:01,597
Elon and others have claimed or believed.
Um this is a lot a lot harder

1957
01:55:01,680 --> 01:55:04,477
than many of many
of those guys think it is.

1958
01:55:04,560 --> 01:55:08,757
Um and many of those guys who thought it
was simpler than that years you know, 5

1959
01:55:08,840 --> 01:55:13,197
years ago now think it's hard because
it's been 5 years and they realize it's

1960
01:55:13,280 --> 01:55:15,277
it's going to take a lot longer.

1961
01:55:15,360 --> 01:55:17,997
That includes [snorts] a bunch
of people at DeepMind for example, but um

1962
01:55:18,080 --> 01:55:21,197
>> Oh, interesting. I haven't actually
uh touched base with the DeepMind folks.

1963
01:55:21,280 --> 01:55:23,757
But some of it Elon or

1964
01:55:23,840 --> 01:55:26,637
uh Demis Hassabis, I mean sometimes in

1965
01:55:26,720 --> 01:55:30,037
your role you have to kind of create

1966
01:55:30,120 --> 01:55:35,077
deadlines that are nearer than farther
away to to kind of create an urgency cuz

1967
01:55:35,160 --> 01:55:37,437
you know, you have to believe
the impossible is possible in order to

1968
01:55:37,520 --> 01:55:41,037
accomplish it. And there's of course
a flip side to that coin, but it's it's a

1969
01:55:41,120 --> 01:55:44,357
weird You can't be too cynical
if you want to get something done.

1970
01:55:44,440 --> 01:55:48,597
>> Absolutely. I I I agree with that,
but um I mean you have to inspire people,

1971
01:55:48,680 --> 01:55:51,197
right? To work on certain
ambitious things.

1972
01:55:51,280 --> 01:55:55,437
Um Uh so, you know, it's it's uh it's

1973
01:55:55,520 --> 01:55:58,237
certainly uh a lot harder than we
believe, but there's no question in my

1974
01:55:58,320 --> 01:56:00,037
mind that this will this will happen.

1975
01:56:00,120 --> 01:56:03,637
And now, you know, people are kind
of worried about what does that mean for

1976
01:56:03,720 --> 01:56:07,077
humans? They they are going to be
brought down from their pedestal, you

1977
01:56:07,160 --> 01:56:11,757
know, a bunch of notches. With that and uh

1978
01:56:11,840 --> 01:56:14,397
you know, is that going to be good or bad?
I mean, it's just going to give

1979
01:56:14,480 --> 01:56:18,037
more power, right? To sort of amplify
your for human intelligence, really.

1980
01:56:18,120 --> 01:56:20,357
>> So, speaking of doing cool ambitious

1981
01:56:20,440 --> 01:56:24,437
things, FAIR, the Facebook AI research

1982
01:56:24,520 --> 01:56:28,397
group, has recently celebrated
its eighth birthday. Or

1983
01:56:28,480 --> 01:56:32,557
maybe you can correct me on that.
Uh looking back, what has been the

1984
01:56:32,640 --> 01:56:36,357
successes, the failures, the lessons
learned from the 8 years of FAIR? And

1985
01:56:36,440 --> 01:56:39,477
maybe you can also give context of

1986
01:56:39,560 --> 01:56:44,477
the newly minted Meta AI fit into
uh how does it relate to FAIR?

1987
01:56:44,560 --> 01:56:48,077
>> Right. So, let me tell you a little
bit about the organization of all this. Uh

1988
01:56:48,160 --> 01:56:52,597
uh Yeah, FAIR was created almost exactly
8 years ago. Uh it wasn't called FAIR

1989
01:56:52,680 --> 01:56:55,877
yet. It took that name
a few a few months later.

1990
01:56:55,960 --> 01:56:58,557
Um and at the time uh I joined uh

1991
01:56:58,640 --> 01:57:03,567
Facebook, um there was a a group called
the AI group that had about 12 engineers

1992
01:57:03,650 --> 01:57:06,717
[snorts] and a few sci- a few scientists,
like, you know, 10 engineers

1993
01:57:06,800 --> 01:57:08,157
and two scientists or something like that.

1994
01:57:08,240 --> 01:57:12,597
>> Mhm. >> I run it for 3
and 1/2 years as a director.

1995
01:57:12,680 --> 01:57:16,237
You know, hired the first few scientists
and kind of set up the culture and

1996
01:57:16,320 --> 01:57:20,837
organized it, you know, explained
to the Facebook leadership what what

1997
01:57:20,920 --> 01:57:24,677
fundamental research was about
and how it can work within uh

1998
01:57:24,760 --> 01:57:27,917
uh industry and how it needs
to be open and everything.

1999
01:57:28,000 --> 01:57:32,877
Um and uh I think it's
been an an qualified

2000
01:57:32,960 --> 01:57:35,797
success in the sense that

2001
01:57:35,880 --> 01:57:39,757
FAIR has simultaneously produced um you

2002
01:57:39,840 --> 01:57:43,517
know, top-level research and advanced
the the science and the technology,

2003
01:57:43,600 --> 01:57:45,477
provided tools, open-source tools like

2004
01:57:45,560 --> 01:57:47,837
PyTorch and many others.

2005
01:57:47,920 --> 01:57:51,237
Uh but at the same time has had a

2006
01:57:51,320 --> 01:57:57,397
a direct or mostly indirect impact
on uh Facebook at the time, now Meta.

2007
01:57:57,480 --> 01:58:00,477
>> Mhm. >> Uh in the sense
that a lot of systems

2008
01:58:00,560 --> 01:58:05,037
that are that Meta
is built around um now are

2009
01:58:05,120 --> 01:58:10,237
are based on research projects
that uh started at uh at FAIR.

2010
01:58:10,320 --> 01:58:13,557
So, if [snorts] you were to take out,
you know, deep learning out of uh

2011
01:58:13,640 --> 01:58:18,717
Facebook services now and and Meta more
generally, um I mean, the company would

2012
01:58:18,800 --> 01:58:21,557
would literally crumble.
I mean, it's completely built around

2013
01:58:21,640 --> 01:58:25,877
uh around AI these days.
It's really essential to the operations.

2014
01:58:25,960 --> 01:58:30,117
So, what happened uh after
3 and 1/2 years is that uh

2015
01:58:30,200 --> 01:58:33,277
I changed role. I became chief
scientist, so I'm I'm not doing

2016
01:58:33,360 --> 01:58:36,757
day-to-day management
of uh of FAIR anymore.

2017
01:58:36,840 --> 01:58:40,117
I'm more of a kind of you know,
think about strategy and

2018
01:58:40,200 --> 01:58:43,797
things like that and and I carry my I
I conduct my own research. I have, you

2019
01:58:43,880 --> 01:58:46,437
know, my own kind of research group
working on self-supervised learning and

2020
01:58:46,520 --> 01:58:49,677
things like this, which I didn't
have time to do when I was director.

2021
01:58:49,760 --> 01:58:53,117
>> [snorts] >> Um so,
now uh FAIR is run by, uh, Joelle

2022
01:58:53,200 --> 01:58:57,437
Pineau and, uh, Antoine Bord >> [snorts]
>> uh, together because FAIR is kind of

2023
01:58:57,520 --> 01:59:01,277
split in two now. There's something
called FAIR Labs, which is sort of

2024
01:59:01,360 --> 01:59:04,917
bottom-up scientist-driven research,
and FAIR Excel, which is slightly more

2025
01:59:05,000 --> 01:59:09,957
organized for bigger projects that require
a little more kind of focus and more

2026
01:59:10,040 --> 01:59:13,677
engineering support and things like that.
So, Joelle leads FAIR Lab and

2027
01:59:13,760 --> 01:59:14,837
Antoine Bord leads FAIR Excel.

2028
01:59:14,920 --> 01:59:19,957
>> Where are they located? >> Uh,
it's decentralized all over. Um,

2029
01:59:20,040 --> 01:59:23,277
so there's there's there's no
question that the leadership, uh,

2030
01:59:23,360 --> 01:59:25,997
of the company believes that

2031
01:59:26,080 --> 01:59:30,357
this was a very worthwhile investment,
and what that means is that

2032
01:59:30,440 --> 01:59:34,757
uh, it's it's there for the long run.

2033
01:59:34,840 --> 01:59:38,437
Right? So, there is, uh,
if you if you want to talk in these

2034
01:59:38,520 --> 01:59:41,437
terms, which I don't like, this is a
this is a business model, if you want,

2035
01:59:41,520 --> 01:59:46,157
where where uh, FAIR, despite being a very
fundamental research lab, brings a lot

2036
01:59:46,240 --> 01:59:49,797
of value to the company, either
mostly indirectly through other groups.

2037
01:59:49,880 --> 01:59:53,197
>> Mhm. >> Um, now,
what happened 3 and 1/2 years

2038
01:59:53,280 --> 01:59:55,037
ago when I stepped down is, uh, was also

2039
01:59:55,120 --> 02:00:00,317
the creation of Facebook AI, which
was basically a larger organization that,

2040
02:00:00,400 --> 02:00:05,877
uh, covers FAIR, so FAIR is included in
it, but also has other organizations that

2041
02:00:05,960 --> 02:00:10,357
are uh, focused on applied
research or advanced

2042
02:00:10,440 --> 02:00:15,757
development uh, of AI technology that is
more, you know, focused on the >> Right.

2043
02:00:15,840 --> 02:00:18,597
>> products of the company. >> So,
less emphasis on fundamental research.

2044
02:00:18,680 --> 02:00:21,037
>> Less fundamental, but it's still
a research. I mean, there's a lot of

2045
02:00:21,120 --> 02:00:26,077
papers coming out of those organizations
and uh, the people are some awesome and,

2046
02:00:26,160 --> 02:00:30,917
you know, wonderful to interact with,
and but it it serves, uh, as kind of,

2047
02:00:31,000 --> 02:00:34,957
uh, uh, a way to kind of scale up, if you

2048
02:00:35,040 --> 02:00:38,237
want, um, sort of AI
technology which, you know,

2049
02:00:38,320 --> 02:00:41,917
may be very experimental and and sort
of lab prototypes into things that are

2050
02:00:42,000 --> 02:00:44,157
usable. >> So,
FAIR is [snorts] a subset of Meta

2051
02:00:44,240 --> 02:00:47,237
AI. Is FAIR become like KFC? It it'll

2052
02:00:47,320 --> 02:00:49,717
just keep the F. Nobody cares what the F

2053
02:00:49,800 --> 02:00:51,317
stands for.

2054
02:00:51,400 --> 02:00:55,157
>> We'll know soon enough by probably

2055
02:00:55,240 --> 02:00:57,597
probably by the end of the of 2021.

2056
02:00:57,680 --> 02:01:00,277
>> not a giant change, mayor fair.

2057
02:01:00,360 --> 02:01:04,077
>> Well, mayor doesn't sound too good, but
you know, the the brand people are kind

2058
02:01:04,160 --> 02:01:08,397
of deciding on this. And they've been
hesitating for for a while now, and they

2059
02:01:08,480 --> 02:01:10,877
you know, they tell us they're going
to come up with an answer as to whether

2060
02:01:10,960 --> 02:01:13,797
fair is going to change name or whether
we're going to change just the meaning

2061
02:01:13,880 --> 02:01:16,557
of the F. >> Oh, that's
a good call. I would keep

2062
02:01:16,640 --> 02:01:18,037
fair and change the meaning of the F.

2063
02:01:18,120 --> 02:01:20,197
That's a good one. >> be my preference,
you know, I would turn

2064
02:01:20,280 --> 02:01:22,757
I would turn the F into fundamental.

2065
02:01:22,840 --> 02:01:24,237
>> Oh, that's good.
>> Fundamentally I research.

2066
02:01:24,320 --> 02:01:25,157
>> Oh, that's really good. Yeah, yeah.

2067
02:01:25,240 --> 02:01:28,357
>> Within meta AI. So,
this would be sort of meta fair.

2068
02:01:28,440 --> 02:01:30,157
>> Yeah. But, you know,
people will call it fair, right?

2069
02:01:30,240 --> 02:01:34,197
>> Yeah, exactly. I like it. >> meta AI

2070
02:01:34,280 --> 02:01:38,997
uh is part of the reality lab. So, you

2071
02:01:39,080 --> 02:01:43,157
know, uh uh meta now
the new Facebook right is

2072
02:01:43,240 --> 02:01:45,317
called meta and it's kind of

2073
02:01:45,400 --> 02:01:51,107
divided into you know,
Facebook, Instagram, WhatsApp.

2074
02:01:51,190 --> 02:01:54,157
>> [snorts] >> And

2075
02:01:54,240 --> 02:01:58,797
reality lab. And reality lab
is about, you know, AR, VR, um

2076
02:01:58,880 --> 02:02:02,197
you know, telepresence, communication
part of uh technology and stuff like

2077
02:02:02,280 --> 02:02:06,117
that. It's kind of the you can
think of it as the sort of a

2078
02:02:06,200 --> 02:02:10,957
combination of um sort of new
products and and technology

2079
02:02:11,040 --> 02:02:15,557
part of uh of uh meta.
>> Is that where the touch sensing for

2080
02:02:15,640 --> 02:02:18,197
robots? I saw that you were
posting about. That's that's

2081
02:02:18,280 --> 02:02:21,317
>> sensing for robot is part of fair,
actually. That's that's a fair part.

2082
02:02:21,400 --> 02:02:25,877
>> Okay, cool. >> Yeah. This is also
the No, but there is the the other way the

2083
02:02:25,960 --> 02:02:29,517
the haptic glove, right?
That that has like >> more reality lab.

2084
02:02:29,600 --> 02:02:32,557
>> That's that's reality lab research.

2085
02:02:32,640 --> 02:02:35,997
>> Reality lab research. But, by the way,
the touch sensors are super interesting.

2086
02:02:36,080 --> 02:02:39,237
Uh like integrating
that modality into the whole

2087
02:02:39,320 --> 02:02:42,037
uh sensing uh suite is very interesting.

2088
02:02:42,120 --> 02:02:45,597
So, uh what do you think
about the metaverse?

2089
02:02:45,680 --> 02:02:50,317
What do you think about this whole
this whole kind of expansion of the view

2090
02:02:50,400 --> 02:02:52,837
of the role of Facebook
and Meta in the world?

2091
02:02:52,920 --> 02:02:56,197
>> Well, a metaverse really should
be thought of as the next step in the

2092
02:02:56,280 --> 02:03:00,237
internet, right? Sort of trying to

2093
02:03:00,320 --> 02:03:03,157
uh kind of, you know, make the

2094
02:03:03,240 --> 02:03:07,597
experience more compelling
um of, you know, being

2095
02:03:07,680 --> 02:03:10,517
connected either with other
people or with content.

2096
02:03:10,600 --> 02:03:15,437
Um and you know,
we are evolved and trying to

2097
02:03:15,520 --> 02:03:20,677
evolve in, you know, 3D environments where
uh you know, we can see other people, we

2098
02:03:20,760 --> 02:03:24,157
can talk to them when
you're when when near them or

2099
02:03:24,240 --> 02:03:26,397
you know, and other people who are far
away can't hear us, you know, things

2100
02:03:26,480 --> 02:03:30,517
like that, right? So, it it's uh there's
a lot of social conventions that exist

2101
02:03:30,600 --> 02:03:33,997
in the real world that we can try
to transpose. Now, what is going to be

2102
02:03:34,080 --> 02:03:38,117
eventually the the uh how
compelling is it going to be?

2103
02:03:38,200 --> 02:03:40,797
Like are, you know,
uh is it going to be the case that

2104
02:03:40,880 --> 02:03:44,557
people are going to be willing to do
this if they have to wear, you know,

2105
02:03:44,640 --> 02:03:49,477
a huge pair of goggles all day? Maybe not.
Um >> But then again, if the experience is

2106
02:03:49,560 --> 02:03:52,237
sufficiently compelling, maybe so.

2107
02:03:52,320 --> 02:03:55,877
>> Or if the device that you have to wear
is just basically a pair of glasses, you

2108
02:03:55,960 --> 02:03:58,877
know, technology makes
sufficient progress for that.

2109
02:03:58,960 --> 02:04:04,357
Um um you know, AR is a much easier
concept to grasp that you're going to

2110
02:04:04,440 --> 02:04:09,517
have, you know, augmented reality glasses
that basically contain some sort of, you

2111
02:04:09,600 --> 02:04:12,157
know, virtual assistant that can
help you in your daily lives.

2112
02:04:12,240 --> 02:04:14,637
>> But at the same at the same time
with the AR, you have to contend with

2113
02:04:14,720 --> 02:04:17,917
reality. With VR, you can completely
detach yourself from reality, so it

2114
02:04:18,000 --> 02:04:22,237
gives you freedom. It might be
easier to design worlds in in VR.

2115
02:04:22,320 --> 02:04:26,597
>> Yeah, but you you you can imagine
how, you know, the metaverse being uh

2116
02:04:26,680 --> 02:04:29,637
>> a mix. >> A mix, right?
Or or like you can have

2117
02:04:29,720 --> 02:04:32,597
objects that exist in the metaverse
that, you know, pop up on top of the

2118
02:04:32,680 --> 02:04:36,277
real world or only
exist in virtual reality.

2119
02:04:36,360 --> 02:04:38,917
>> Okay, let me ask the hard question.

2120
02:04:39,000 --> 02:04:40,437
>> Oh, because all
of this was easy so far.

2121
02:04:40,520 --> 02:04:42,517
>> This was easy. Uh

2122
02:04:42,600 --> 02:04:45,237
the Facebook, now Meta, the social

2123
02:04:45,320 --> 02:04:48,477
network has been painted by the media as

2124
02:04:48,560 --> 02:04:52,957
a net negative for society,
even destructive and evil at times. You've

2125
02:04:53,040 --> 02:04:54,877
pushed back against this, defending

2126
02:04:54,960 --> 02:04:58,477
Facebook. Can you explain your defense?

2127
02:04:58,560 --> 02:05:02,917
>> Yeah, so the the description,
the company that is being described in the

2128
02:05:03,000 --> 02:05:07,797
in some media uh is not
the company we know when we

2129
02:05:07,880 --> 02:05:11,357
work inside. And you know

2130
02:05:11,440 --> 02:05:14,997
it could be claimed that a lot
of employees are uninformed about

2131
02:05:15,080 --> 02:05:18,747
what really goes on in the company, but
you know, I'm a vice president. I mean,

2132
02:05:18,830 --> 02:05:21,197
[snorts] I have a pretty good vision
of what goes on, you know, I don't know

2133
02:05:21,280 --> 02:05:25,077
everything obviously. I'm not involved
in in in everything, but certainly not

2134
02:05:25,160 --> 02:05:28,837
in decision about like, you know, content
moderation or anything like this, but

2135
02:05:28,920 --> 02:05:32,077
but I have, you know, some decent
vision of what what goes on.

2136
02:05:32,160 --> 02:05:35,557
And this evil that is being
described, I just don't see it.

2137
02:05:35,640 --> 02:05:40,077
And then, you know, I think
there is an easy story to buy

2138
02:05:40,160 --> 02:05:44,157
which is that, you know, all the bad
things in the in the world and you know,

2139
02:05:44,240 --> 02:05:47,757
the the reason your friend
believe crazy stuff.

2140
02:05:47,840 --> 02:05:51,557
Um you know, there's an easy
scapegoat, right? In the

2141
02:05:51,640 --> 02:05:56,397
uh in in in in social media
in general, Facebook in particular.

2142
02:05:56,480 --> 02:05:59,157
But you have to look at the data.
Like, is it the case

2143
02:05:59,240 --> 02:06:04,197
that uh Facebook, for example,
uh polarizes people politically?

2144
02:06:04,280 --> 02:06:07,157
Um are there academic
studies that show this?

2145
02:06:07,240 --> 02:06:12,517
Is it the case that uh you know, teenagers
uh think of themselves less if they use

2146
02:06:12,600 --> 02:06:16,517
Instagram more? Is it the case that uh

2147
02:06:16,600 --> 02:06:20,717
you know, people get
more riled up against uh

2148
02:06:20,800 --> 02:06:25,117
you know, opposite sides in a in a
debate or political opinion if they if

2149
02:06:25,200 --> 02:06:27,597
they are more on Facebook
or if they are less?

2150
02:06:27,680 --> 02:06:32,797
And uh study after study show
that none of this is true.

2151
02:06:32,880 --> 02:06:36,957
This is independent studies by academic
that are not funded by Facebook or Meta.

2152
02:06:37,040 --> 02:06:40,557
Um, you know, studied by Stanford,
by some of my colleagues at NYU actually

2153
02:06:40,640 --> 02:06:42,317
with whom I have no connection.

2154
02:06:42,400 --> 02:06:44,477
Um, you know, there's a study recently

2155
02:06:44,560 --> 02:06:47,237
they they they paid people, I think it

2156
02:06:47,320 --> 02:06:50,077
was in um

2157
02:06:50,160 --> 02:06:54,357
in in in the former Yugoslavia, I'm not
exactly sure in what what part, but they

2158
02:06:54,440 --> 02:06:57,757
paid people to not use
Facebook for a while in the

2159
02:06:57,840 --> 02:07:04,197
period um uh before
the anniversary of the Srebrenica

2160
02:07:04,280 --> 02:07:07,477
>> Mhm. >> massacres, right? So, you know,
people get riled up like should you know,

2161
02:07:07,560 --> 02:07:11,437
should we have a celebration,
I mean a memorial kind of

2162
02:07:11,520 --> 02:07:14,437
celebration for it or not. So,
they paid [snorts] a bunch of people

2163
02:07:14,520 --> 02:07:18,157
to not use Facebook for a few weeks.

2164
02:07:18,240 --> 02:07:22,277
It turns out that those
people ended up being more

2165
02:07:22,360 --> 02:07:25,357
polarized than they were at the
beginning and the people who were more

2166
02:07:25,440 --> 02:07:27,757
on Facebook were less polarized.

2167
02:07:27,840 --> 02:07:31,157
Um, there's a study,
you know, from Stanford of

2168
02:07:31,240 --> 02:07:35,877
uh economists at Stanford that tried
to identify the causes of increasing

2169
02:07:35,960 --> 02:07:39,717
polarization in the US.
And it's been going on for 40 years

2170
02:07:39,800 --> 02:07:42,437
before, you know,
Mark Zuckerberg was born.

2171
02:07:42,520 --> 02:07:46,837
>> Yeah. >> Uh continuously.
And um and uh so the if

2172
02:07:46,920 --> 02:07:50,077
there is a cause, it's not Facebook
or social media. So, you could say if

2173
02:07:50,160 --> 02:07:54,277
social media just accelerated but no,
I mean it's basically a continuous

2174
02:07:54,360 --> 02:07:57,797
uh evolution by some measure
of uh polarization in the US.

2175
02:07:57,880 --> 02:08:01,517
And then you compare this with other
countries like uh >> [snorts]

2176
02:08:01,600 --> 02:08:05,757
>> uh the the the west half of Germany
because you can go 40 years in the east

2177
02:08:05,840 --> 02:08:09,237
east side or Denmark
or or other countries.

2178
02:08:09,320 --> 02:08:12,557
Uh and they use Facebook just as much
and they're not getting more polarized,

2179
02:08:12,640 --> 02:08:16,397
they're getting less polarized. So,
if you want to look for, you know, uh a

2180
02:08:16,480 --> 02:08:21,077
causal relationship there um you
can find a scapegoat, but you can't

2181
02:08:21,160 --> 02:08:24,397
find a cause. Now, if you want to fix
the problem, you have to find the right

2182
02:08:24,480 --> 02:08:26,837
cause. And what riled me up is that

2183
02:08:26,920 --> 02:08:31,477
people now are accusing Facebook of bad
deeds that are done by others and those

2184
02:08:31,560 --> 02:08:33,557
others are we're not
doing anything about them.

2185
02:08:33,640 --> 02:08:36,477
>> Mhm. >> And by the way,
those others include the

2186
02:08:36,560 --> 02:08:39,557
owner of the Wall Street Journal in which
all of those papers were published.

2187
02:08:39,640 --> 02:08:44,077
>> So, I should mention that I'm talking
to Sheryl Mike Sheryl for on this podcast

2188
02:08:44,160 --> 02:08:46,837
and also Mark Zuckerberg and probably
these are conversations you can have

2189
02:08:46,920 --> 02:08:49,837
with them. Uh cuz it's
very interesting to me even

2190
02:08:49,920 --> 02:08:53,957
if Facebook has an has some
measurable negative effect. You

2191
02:08:54,040 --> 02:08:55,677
can't just consider that in isolation.

2192
02:08:55,760 --> 02:08:58,797
You have to consider about all
the positive ways that it connects us. So,

2193
02:08:58,880 --> 02:09:00,757
like every technology

2194
02:09:00,840 --> 02:09:03,237
>> It's a question.
You can't just say like

2195
02:09:03,320 --> 02:09:08,117
uh there's an increase in division.
Yes, probably Google search engine has

2196
02:09:08,200 --> 02:09:11,797
created increase in division. We have
to consider about how much information are

2197
02:09:11,880 --> 02:09:15,757
brought to the world. Like I'm sure
Wikipedia created more division. If you

2198
02:09:15,840 --> 02:09:19,517
just look at the division, we have
to look at the full context of the world

2199
02:09:19,600 --> 02:09:22,837
and did it make a better world? You have
to >> the printing press has created more

2200
02:09:22,920 --> 02:09:26,677
division, right? >> Exactly.
>> So, you know, when the

2201
02:09:26,760 --> 02:09:30,757
printing press was invented,
the first books that were that were

2202
02:09:30,840 --> 02:09:34,677
printed were things like the Bible and
that allowed people to read the Bible by

2203
02:09:34,760 --> 02:09:37,117
themselves, not get the message uniquely

2204
02:09:37,200 --> 02:09:41,757
from priests in Europe.
And that created, you know, the Protestant

2205
02:09:41,840 --> 02:09:46,637
movement and 200 years of religious
persecution and wars. So, that's a bad

2206
02:09:46,720 --> 02:09:49,677
side effect of the printing press.
You know, social networks aren't being

2207
02:09:49,760 --> 02:09:52,597
nearly as bad as the printing press, but
nobody would say the printing press was

2208
02:09:52,680 --> 02:09:57,197
a bad idea. >> [sighs and gasps] >>
Yeah, a lot of it is perception and

2209
02:09:57,280 --> 02:10:00,197
there's a lot of different
incentives operating here.

2210
02:10:00,280 --> 02:10:02,757
Um maybe a quick comment since you're

2211
02:10:02,840 --> 02:10:07,477
one of the top leaders at Facebook
and and at meta, sorry, that's in the tech

2212
02:10:07,560 --> 02:10:10,997
space. I'm sure Facebook involves a lot of

2213
02:10:11,080 --> 02:10:15,077
incredible technological challenges
that need to be solved. A lot

2214
02:10:15,160 --> 02:10:18,117
of it probably is on the computer
infrastructure, the hardware, the I

2215
02:10:18,200 --> 02:10:20,437
mean, it's just a huge amount.

2216
02:10:20,520 --> 02:10:23,317
Uh maybe can you give me context about

2217
02:10:23,400 --> 02:10:28,277
how much of Sherp's life is AI and how
much of it is low-level compute, how

2218
02:10:28,360 --> 02:10:31,477
much of it is flying all
around doing business stuff

2219
02:10:31,560 --> 02:10:33,917
in the same way as Zuckerberg,
Mark Zuckerberg.

2220
02:10:34,000 --> 02:10:37,237
>> They really focus on AI.
I mean, certainly, uh,

2221
02:10:37,320 --> 02:10:41,397
in the, uh, in the run-up
of the creation of Fair

2222
02:10:41,480 --> 02:10:45,557
and for, you know, at least
a year after that, if not more,

2223
02:10:45,640 --> 02:10:47,717
uh, Mark was was very, very much focused

2224
02:10:47,800 --> 02:10:52,677
on on AI and was spending quite a lot of
effort, um, on it. And that's his style.

2225
02:10:52,760 --> 02:10:55,797
When he gets interested in something,
he reads everything about it.

2226
02:10:55,880 --> 02:10:57,237
You know, he read some of my papers, for

2227
02:10:57,320 --> 02:11:03,397
example, before he joined. Um,
um, and, uh, so he he learned a lot

2228
02:11:03,480 --> 02:11:08,357
about it. >> notes.
>> [laughter] >> Right? Um, >> [snorts]

2229
02:11:08,440 --> 02:11:13,357
>> and, uh, uh, you know, Sherp
was really tweeting also. I mean, Sherp

2230
02:11:13,440 --> 02:11:18,237
is really kind of, um, you know, has

2231
02:11:18,320 --> 02:11:22,237
something I've tried to preserve also,
despite my, uh, not so young age, which

2232
02:11:22,320 --> 02:11:25,997
is a a sense of wonder about, uh,
science and technology. And he certainly

2233
02:11:26,080 --> 02:11:29,357
he certainly has that. Um,
he's also a wonderful person. I

2234
02:11:29,440 --> 02:11:33,237
mean, in terms of, uh, like, uh,
as a manager, like, dealing with people

2235
02:11:33,320 --> 02:11:35,517
and everything. Mark also, actually.

2236
02:11:35,600 --> 02:11:39,677
Um, so I mean, they're very like,
you know, very human people.

2237
02:11:39,760 --> 02:11:44,077
For in the case of Mark, it's, uh,
shockingly human, you know, given his, uh,

2238
02:11:44,160 --> 02:11:49,077
his trajectory. Uh, I mean,
the personality of of him that

2239
02:11:49,160 --> 02:11:51,477
is painted in the press
is just completely wrong.

2240
02:11:51,560 --> 02:11:55,197
>> Yeah. But you have to know
how to play the press. So that's

2241
02:11:55,280 --> 02:11:56,997
I I put some of that responsibility on

2242
02:11:57,080 --> 02:12:00,197
him, too. You have to

2243
02:12:00,280 --> 02:12:04,717
It's like, um, you know,
like, the director, the

2244
02:12:04,800 --> 02:12:08,877
conductor of an orchestra,
you have to play the press and the public

2245
02:12:08,960 --> 02:12:11,997
in a certain kind of way where you
convey your true self to them. If there

2246
02:12:12,080 --> 02:12:13,597
is a depth and kindness to you.

2247
02:12:13,680 --> 02:12:17,957
>> And he's probably And he's
probably not the best at it. So, yeah.

2248
02:12:18,040 --> 02:12:20,557
>> [laughter] >> You have to learn.
And it's it's sad to

2249
02:12:20,640 --> 02:12:25,957
see I'll talk to him about it,
but the strap is slowly stepping down.

2250
02:12:26,040 --> 02:12:31,317
It's always sad to see folks sort
of be there for a long time and slowly

2251
02:12:31,400 --> 02:12:34,357
I guess time is up. >> You know,
I think he's I think he's done

2252
02:12:34,440 --> 02:12:38,477
the thing he set out to do
and, you know, he's he's

2253
02:12:38,560 --> 02:12:44,637
got, you know, uh family
priorities and stuff like that and

2254
02:12:44,720 --> 02:12:49,797
um I I understand, you know,
after 13 years or something.

2255
02:12:49,880 --> 02:12:53,337
>> It's been a good run.
>> Which in Silicon Valley is basically a

2256
02:12:53,420 --> 02:12:56,917
[laughter] lifetime. >> Yeah.
>> You know, cuz you know, it's dog years.

2257
02:12:57,000 --> 02:13:00,317
>> So, NeurIPS the conference
just wrapped up.

2258
02:13:00,400 --> 02:13:03,717
Let [clears throat] me just go back
to something else. You posted a paper you

2259
02:13:03,800 --> 02:13:06,357
co-authored was rejected from NeurIPS.

2260
02:13:06,440 --> 02:13:09,697
As you said, proudly in quotes, rejected.

2261
02:13:09,780 --> 02:13:12,837
>> [laughter] >> No joke. >> Yeah, I know.

2262
02:13:12,920 --> 02:13:15,837
Uh can you describe this paper and like

2263
02:13:15,920 --> 02:13:20,397
what what what was the idea in it?
And also, maybe this is a good opportunity

2264
02:13:20,480 --> 02:13:24,277
to ask what are the pros and cons, what
works and what doesn't about the review

2265
02:13:24,360 --> 02:13:26,837
process. >> Yeah, let me
talk about the paper first.

2266
02:13:26,920 --> 02:13:30,557
I'll talk about the review talk
about the review process afterwards.

2267
02:13:30,640 --> 02:13:34,477
Um the paper is called VICReg.
So, this is I mentioned that before,

2268
02:13:34,560 --> 02:13:38,277
variance-invariance covariance
regularization. And it's a technique in

2269
02:13:38,360 --> 02:13:42,557
non-contrastive learning technique
for what I call joint embedding

2270
02:13:42,640 --> 02:13:44,797
architecture. So, Siamese nets are an

2271
02:13:44,880 --> 02:13:46,757
example of joint embedding architecture.

2272
02:13:46,840 --> 02:13:51,157
So, joint embedding architecture is uh uh

2273
02:13:51,240 --> 02:13:54,837
Let me back up a little bit, right? So, if
you want to do self-supervised learning,

2274
02:13:54,920 --> 02:13:57,917
uh you can you can do it by prediction.

2275
02:13:58,000 --> 02:14:00,917
Um so, let's say you want to train
a system to predict video, right? You show

2276
02:14:01,000 --> 02:14:05,557
it a video clip and uh and you train
the system to predict the next the

2277
02:14:05,640 --> 02:14:09,677
continuation of that video clip. Now,
because you need to handle uncertainty,

2278
02:14:09,760 --> 02:14:13,557
cuz there There of you, you know,
many continuations that are plausible.

2279
02:14:13,640 --> 02:14:15,397
You need to have You need to handle this

2280
02:14:15,480 --> 02:14:17,917
in some way. You need to have a way for

2281
02:14:18,000 --> 02:14:22,407
the system to be able
to produce multiple predictions.

2282
02:14:22,490 --> 02:14:25,437
>> [snorts] >> And the way
The only way I know to do this

2283
02:14:25,520 --> 02:14:29,597
is through what's called a latent
variable. So, you have some sort of

2284
02:14:29,680 --> 02:14:33,677
hidden vector of a variable that you
can vary over a set or draw from a

2285
02:14:33,760 --> 02:14:37,037
distribution. And as you vary this
vector over a set, the output, the

2286
02:14:37,120 --> 02:14:39,997
prediction varies over a set
of plausible predictions.

2287
02:14:40,080 --> 02:14:42,677
>> Mhm. >> Okay. So, [clears
throat] that's called I call this a

2288
02:14:42,760 --> 02:14:46,037
a generative latent variable model.

2289
02:14:46,120 --> 02:14:48,477
>> Got it. >> Okay? Now,
there is an alternative to

2290
02:14:48,560 --> 02:14:50,597
this to handle uncertainty.

2291
02:14:50,680 --> 02:14:54,077
And instead of directly predicting the

2292
02:14:54,160 --> 02:14:57,757
the next frames of the of the of the clip,

2293
02:14:57,840 --> 02:15:02,637
you also run those
through another neural net.

2294
02:15:02,720 --> 02:15:07,437
Uh so, that you now have two neural nets.
One that looks at the the

2295
02:15:07,520 --> 02:15:11,437
you know, the initial segment of the
video clip. And another one that looks

2296
02:15:11,520 --> 02:15:14,807
at the the continuation
during training, right?

2297
02:15:14,890 --> 02:15:19,557
>> [snorts] >> And what what you're
trying to do is learn a representation

2298
02:15:19,640 --> 02:15:23,757
of those two video clips that is
maximally informative about the video

2299
02:15:23,840 --> 02:15:28,117
clips themselves, but is such
that you can predict the

2300
02:15:28,200 --> 02:15:32,357
representation of the second video clip
from the representation of the first one

2301
02:15:32,440 --> 02:15:35,517
easily. >> Okay. >> And you
can sort of formalize this in

2302
02:15:35,600 --> 02:15:38,357
terms of maximizing mutual information
and some stuff like that, but it doesn't

2303
02:15:38,440 --> 02:15:41,517
matter. Uh what you want is

2304
02:15:41,600 --> 02:15:43,837
uh informative representative represent

2305
02:15:43,920 --> 02:15:48,637
you know, informative representations
of the two video clips that are mutually

2306
02:15:48,720 --> 02:15:51,997
predictable. >> Mhm. >> [clears throat] >>
What that means is that there's a lot of

2307
02:15:52,080 --> 02:15:55,397
details in the second video
clip that are irrelevant.

2308
02:15:55,480 --> 02:15:57,757
Uh you know, um

2309
02:15:57,840 --> 02:16:00,757
I let's say video clip consists in, you

2310
02:16:00,840 --> 02:16:03,597
know, a camera panning the scene.

2311
02:16:03,680 --> 02:16:06,797
Uh there's going to be a piece of that
room that is going to be revealed, and I

2312
02:16:06,880 --> 02:16:10,277
can somewhat predict what the what
that room is going to look like, but I may

2313
02:16:10,360 --> 02:16:14,637
not be able to predict the details
of the texture of the ground and where the

2314
02:16:14,720 --> 02:16:16,877
tiles are ending and stuff and stuff
like that, right? So, those are

2315
02:16:16,960 --> 02:16:21,637
irrelevant details that perhaps
my representation will eliminate.

2316
02:16:21,720 --> 02:16:24,797
And so, what I need
is to train this second

2317
02:16:24,880 --> 02:16:27,157
neural net in such a way that

2318
02:16:27,240 --> 02:16:33,117
uh whenever the the the the continuation
video clip varies over all the possible

2319
02:16:33,200 --> 02:16:35,517
continuations, >> Mhm.

2320
02:16:35,600 --> 02:16:37,557
>> the representation doesn't change.

2321
02:16:37,640 --> 02:16:40,437
>> Got it. So,
it's the Yeah, yeah, got it.

2322
02:16:40,520 --> 02:16:44,197
Over the space of representations,
doing the same kind of thing as you're

2323
02:16:44,280 --> 02:16:46,237
doing with similarity learning.

2324
02:16:46,320 --> 02:16:49,957
>> Right. >> Yeah. >> So,
so these are two ways to handle

2325
02:16:50,040 --> 02:16:53,797
multimodality in a prediction, right?
In the first way, you parametrize the

2326
02:16:53,880 --> 02:16:57,717
prediction with a latent variable,
but you predict pixels essentially, right?

2327
02:16:57,800 --> 02:17:01,237
In the second one, you don't predict
pixels, you predict an abstract

2328
02:17:01,320 --> 02:17:03,877
representation of pixels,
and you guarantee that this abstract

2329
02:17:03,960 --> 02:17:06,957
representation has as much
information as possible

2330
02:17:07,040 --> 02:17:10,397
about the input, but sort of, you know,
drops all the stuff that you really

2331
02:17:10,480 --> 02:17:15,157
can't predict, essentially.
Uh I used to be a big fan of the first

2332
02:17:15,240 --> 02:17:19,037
approach, and in fact, in this paper with
Ishan Mishra, this this blog post, >> Mhm.

2333
02:17:19,120 --> 02:17:21,877
>> uh the dark matter intelligence, I was
kind of advocating for this. And in the

2334
02:17:21,960 --> 02:17:24,237
last [clears throat] year and a half,
I've completely changed

2335
02:17:24,320 --> 02:17:26,277
my mind. I'm now a big
fan of the second one.

2336
02:17:26,360 --> 02:17:30,117
>> Mhm. >> Uh and it's because of uh

2337
02:17:30,200 --> 02:17:34,597
uh a small collection of algorithms
that have been proposed over the last uh

2338
02:17:34,680 --> 02:17:39,717
year and a half or so, 2 years,
uh to do this, uh including VICReg, uh

2339
02:17:39,800 --> 02:17:43,957
its predecessor called Barlow
Twins, which I mentioned, uh

2340
02:17:44,040 --> 02:17:47,077
a method from our friends
at DeepMind called BYOL,

2341
02:17:47,160 --> 02:17:51,637
um and and and there's a bunch of others
now that kind of work similarly. So,

2342
02:17:51,720 --> 02:17:55,157
they're they're all based on this idea
of joint embedding. Some of them have an

2343
02:17:55,240 --> 02:17:58,397
explicit criterion that is an
approximation of mutual information.

2344
02:17:58,480 --> 02:18:02,037
Some others like BYOL work, but we don't
really know why. And there's been like

2345
02:18:02,120 --> 02:18:05,117
lots of theoretical papers about why
Barlow Twins works. No, it's not that

2346
02:18:05,200 --> 02:18:08,277
because we take it out and it still works
and you know, blah blah blah. I mean, so

2347
02:18:08,360 --> 02:18:12,557
there's like a big a big debate, but um
uh but the important point is that we

2348
02:18:12,640 --> 02:18:16,317
now have a collection of non-contrastive
joint embedding methods, which I think

2349
02:18:16,400 --> 02:18:18,357
is the best thing since sliced bread.

2350
02:18:18,440 --> 02:18:23,117
So, I'm super excited about this
because I think it's our best shot

2351
02:18:23,200 --> 02:18:26,757
for techniques that would
allow us to kind of build uh

2352
02:18:26,840 --> 02:18:30,157
predictive world models and at
the same time learn how I pick up

2353
02:18:30,240 --> 02:18:33,877
representations of the world where what
matters about the world is preserved and

2354
02:18:33,960 --> 02:18:36,317
what is irrelevant is eliminated.

2355
02:18:36,400 --> 02:18:39,877
>> Uh by the way,
the representations before and after is

2356
02:18:39,960 --> 02:18:43,957
in the space in a sequence
of images or is it for single images?

2357
02:18:44,040 --> 02:18:46,637
>> Uh it would be either for a single
image, for a sequence. It could It

2358
02:18:46,720 --> 02:18:49,277
doesn't have to be images. This could
be applied to text. This could be applied

2359
02:18:49,360 --> 02:18:53,197
to just about any signal.
I'm looking at you know, I'm looking for

2360
02:18:53,280 --> 02:18:56,837
methods that are generally applicable,
that are not specific to you know, one

2361
02:18:56,920 --> 02:18:59,517
particular modality. You know,
it could be audio or whatever.

2362
02:18:59,600 --> 02:19:03,517
>> Got it. So, what's the story behind
this paper? This This paper is what? Is Is

2363
02:19:03,600 --> 02:19:05,437
describing one of the one such method?

2364
02:19:05,520 --> 02:19:09,157
>> This this VicReg method. So, this is
co-authored uh the first author is uh a

2365
02:19:09,240 --> 02:19:11,677
student uh called Adrien Bard, who is a

2366
02:19:11,760 --> 02:19:14,597
resident PhD student at Fair Paris.

2367
02:19:14,680 --> 02:19:19,637
He's co-advised by me and Jean Ponce,
um who's a professor at École Normale

2368
02:19:19,720 --> 02:19:23,037
Supérieure, also uh
research director at Inria.

2369
02:19:23,120 --> 02:19:27,317
Um so, this is wonderful program
in France where PhD students can basically

2370
02:19:27,400 --> 02:19:29,117
do their PhD in industry, and that's

2371
02:19:29,200 --> 02:19:31,477
kind of what what's happening here.

2372
02:19:31,560 --> 02:19:35,717
Um and this paper is a follow-up on uh

2373
02:19:35,800 --> 02:19:37,797
the the this Barlow Twins paper by yeah,

2374
02:19:37,880 --> 02:19:41,437
my former postdoc now Stefano De uh with

2375
02:19:41,520 --> 02:19:43,917
Li Jing and and Yurish Montal and bunch

2376
02:19:44,000 --> 02:19:46,557
of other people from from from Fair.

2377
02:19:46,640 --> 02:19:49,157
And one of the main criticism from

2378
02:19:49,240 --> 02:19:50,997
reviewers is that VicReg is not

2379
02:19:51,080 --> 02:19:54,877
different enough from Barlow Twins. But,

2380
02:19:54,960 --> 02:19:58,797
you know, my impression
is that it's, you know,

2381
02:19:58,880 --> 02:20:02,797
bottom twins with a few
bugs fixed essentially. And

2382
02:20:02,880 --> 02:20:05,037
in the end, this is what people will use.

2383
02:20:05,120 --> 02:20:08,957
>> Right. So, >> But, you know,
I'm used to stuff >> Yeah.

2384
02:20:09,040 --> 02:20:10,997
>> that I submit being
rejected for a while.

2385
02:20:11,080 --> 02:20:14,157
>> might be rejected and actually
accepted and well-cited cuz people use it.

2386
02:20:14,240 --> 02:20:16,317
>> Well, it's already
decided like a bunch of times.

2387
02:20:16,400 --> 02:20:18,517
>> So, I mean,
the the the question is then

2388
02:20:18,600 --> 02:20:22,637
to the deeper question about peer
review and conferences. I mean, computer

2389
02:20:22,720 --> 02:20:26,877
science as a field is kind of unique
that the conference is highly prized.

2390
02:20:26,960 --> 02:20:29,997
That's one. >> Right.
>> And it's interesting because the peer

2391
02:20:30,080 --> 02:20:33,517
review process there is similar,
I suppose, to the journals, but it's

2392
02:20:33,600 --> 02:20:38,437
accelerated significantly. Well,
not significantly, but it goes fast.

2393
02:20:38,520 --> 02:20:43,437
It's a nice way to get stuff out quickly,
to peer review quickly, go to present it

2394
02:20:43,520 --> 02:20:46,997
quickly to the community.
So, not quickly, but quicker.

2395
02:20:47,080 --> 02:20:48,957
>> Yeah. >> But, nevertheless,
it has many of the

2396
02:20:49,040 --> 02:20:53,357
same flaws of peer review cuz it's
a limited number of people look at it.

2397
02:20:53,440 --> 02:20:57,397
There's bias in the following like
if you if you want to do new ideas,

2398
02:20:57,480 --> 02:21:01,637
you're going to get pushback.
Um there's self-interested people that

2399
02:21:01,720 --> 02:21:05,517
kind of can infer who submitted it

2400
02:21:05,600 --> 02:21:09,637
and kind of you know, be cranky
about it, all that kind of stuff.

2401
02:21:09,720 --> 02:21:12,397
>> Yeah, I mean, there's a lot of,
you know, social [snorts] phenomena there.

2402
02:21:12,480 --> 02:21:16,397
Um there's one social phenomenon which
is that because the field has been

2403
02:21:16,480 --> 02:21:20,437
growing exponentially, the vast
majority of people in the field

2404
02:21:20,520 --> 02:21:22,477
are extremely junior. >> Yeah.

2405
02:21:22,560 --> 02:21:26,757
>> So, as a consequence, and that's just
a consequence of the field growing, right?

2406
02:21:26,840 --> 02:21:31,437
So, as the number of as the size of the
field kind of start saturating, you you

2407
02:21:31,520 --> 02:21:35,877
have less of that problem
of reviewers being very uh

2408
02:21:35,960 --> 02:21:42,077
inexperienced. A consequence of this
is that, you know, young reviewers

2409
02:21:42,160 --> 02:21:46,517
I mean, there's a phenomenon which is that
reviewers try to make their life easy.

2410
02:21:46,600 --> 02:21:49,397
And to make their life
easy when reviewing a paper

2411
02:21:49,480 --> 02:21:53,157
is very simple. You just have to find
a flaw in the paper, right? So, basically

2412
02:21:53,240 --> 02:21:57,917
they see that task as finding flaws
in papers and most papers have flaws,

2413
02:21:58,000 --> 02:21:59,557
even the good ones.

2414
02:21:59,640 --> 02:22:02,637
So it's it's easy to you know

2415
02:22:02,720 --> 02:22:06,717
to do that. You you your job
is easier as a reviewer

2416
02:22:06,800 --> 02:22:10,877
if you just focus on this.
But what's important is

2417
02:22:10,960 --> 02:22:16,077
like is there a new idea in that
paper that is likely to influence

2418
02:22:16,160 --> 02:22:19,317
It doesn't matter if the experiments are
not that great, if the protocol is, you

2419
02:22:19,400 --> 02:22:23,757
know, so-so, you know,
things like that. As long as

2420
02:22:23,840 --> 02:22:28,677
there is a worthy idea in it that
will influence the way people think

2421
02:22:28,760 --> 02:22:32,517
about the problem >> [snorts] >>
even if they make it better, you know,

2422
02:22:32,600 --> 02:22:36,437
eventually. I think that's that's
really what what makes a paper

2423
02:22:36,520 --> 02:22:37,437
useful.

2424
02:22:37,520 --> 02:22:40,157
And so this combination of

2425
02:22:40,240 --> 02:22:43,677
social phenomena creates

2426
02:22:43,760 --> 02:22:47,477
a disease that has plagued, you know,
other fields in the past like speech

2427
02:22:47,560 --> 02:22:51,877
recognition where basically, you know,
people chase these numbers on uh

2428
02:22:51,960 --> 02:22:55,917
on benchmarks and and it's
much easier to get a paper

2429
02:22:56,000 --> 02:23:00,117
accepted if it brings
an incremental improvement on a

2430
02:23:00,200 --> 02:23:04,317
sort of mainstream well-accepted uh method

2431
02:23:04,400 --> 02:23:07,997
or problem. And those
are to me boring papers. I

2432
02:23:08,080 --> 02:23:12,957
mean they're not useless, right? Because
industry, you know, strives on on those

2433
02:23:13,040 --> 02:23:15,157
kind of progress.
But they're not the one that I'm

2434
02:23:15,240 --> 02:23:18,397
interested in in terms of like
new concepts and new ideas. So

2435
02:23:18,480 --> 02:23:23,077
papers that are really trying
to strike kind of new advances

2436
02:23:23,160 --> 02:23:26,117
generally don't make it.
Now thankfully we have archive.

2437
02:23:26,200 --> 02:23:30,757
>> Archive, exactly. And then there's
open review type of situations where you

2438
02:23:30,840 --> 02:23:34,917
And then I mean Twitter is a kind
of open review. I I'm a huge believer that

2439
02:23:35,000 --> 02:23:38,277
review should be done by thousands
of people, not two people.

2440
02:23:38,360 --> 02:23:41,997
>> And so archive Like do you
see a future where a lot of

2441
02:23:42,080 --> 02:23:45,797
really strong papers, it's already
the present, but a growing future where

2442
02:23:45,880 --> 02:23:49,877
it'll just be archive and

2443
02:23:49,960 --> 02:23:52,437
you're presenting an ongoing continuous

2444
02:23:52,520 --> 02:23:55,717
conference called Twitter / the internet

2445
02:23:55,800 --> 02:23:59,557
/ archive sanity.
Andre just released a new version.

2446
02:23:59,640 --> 02:24:02,197
So just not you know not being so

2447
02:24:02,280 --> 02:24:05,437
elitist about this particular gating.

2448
02:24:05,520 --> 02:24:06,837
>> not a question of being elitist or not.

2449
02:24:06,920 --> 02:24:11,157
It's a question of being basically

2450
02:24:11,240 --> 02:24:15,557
recommendation and sort of approvals
for people who don't see themselves as

2451
02:24:15,640 --> 02:24:18,997
having the ability to do so by themselves,
right? And so it saves time,

2452
02:24:19,080 --> 02:24:20,837
right? If you rely on other people's

2453
02:24:20,920 --> 02:24:23,997
opinion and you trust those people or

2454
02:24:24,080 --> 02:24:29,637
those groups to evaluate a paper for you.

2455
02:24:29,720 --> 02:24:34,477
Um that saves you time cuz you know
you don't have to like scrutinize

2456
02:24:34,560 --> 02:24:37,357
the paper as much, you know, it
is brought to your attention. I mean it's

2457
02:24:37,440 --> 02:24:40,637
the whole idea of sort of you know
collective recommendation system, right?

2458
02:24:40,720 --> 02:24:43,957
So I actually thought about this a lot um

2459
02:24:44,040 --> 02:24:46,157
you know, about 10 15 years ago. Of

2460
02:24:46,240 --> 02:24:48,597
course the what discussions at

2461
02:24:48,680 --> 02:24:53,597
NIPS and you know and we're about
to create ICLR with Yoshua Bengio and so I

2462
02:24:53,680 --> 02:24:58,157
wrote a document um kind
of describing a reviewing system

2463
02:24:58,240 --> 02:25:01,837
which basically was you know, you post
your paper on some repository, let's say

2464
02:25:01,920 --> 02:25:06,997
archive or now could be open review.
Um and then you can form

2465
02:25:07,080 --> 02:25:11,037
a reviewing entity which is equivalent
to a reviewing board, you know, of a

2466
02:25:11,120 --> 02:25:15,837
journal or um program
committee of a conference. Um

2467
02:25:15,920 --> 02:25:19,917
you have to list
the members and then um that

2468
02:25:20,000 --> 02:25:24,437
group reviewing entity can choose
to review a particular paper

2469
02:25:24,520 --> 02:25:28,277
spontaneously or not. There is no
exclusive relationship anymore between a

2470
02:25:28,360 --> 02:25:31,037
paper and a venue or reviewing entity.

2471
02:25:31,120 --> 02:25:33,597
Any reviewing entity can review any paper.

2472
02:25:33,680 --> 02:25:38,517
Um or may choose not to.
Um and then you know, given evaluation

2473
02:25:38,600 --> 02:25:42,437
it's not published not published, it's
just an evaluation and a comment which

2474
02:25:42,520 --> 02:25:45,597
would be public, signed
by the reviewing entity.

2475
02:25:45,680 --> 02:25:47,757
And if it's signed
by the reviewing entity,

2476
02:25:47,840 --> 02:25:50,517
you know it's one of the members
of reviewing entity. So, if the reviewing

2477
02:25:50,600 --> 02:25:54,877
entity is, you know,
Lex Fridman's, you know, preferred

2478
02:25:54,960 --> 02:25:57,517
papers, right? You know it's
Lex Fridman writing the review.

2479
02:25:57,600 --> 02:26:02,077
>> Yes. What So,
for me one that's a beautiful

2480
02:26:02,160 --> 02:26:06,437
system, I think, but what in addition
to that it feels like there should be a

2481
02:26:06,520 --> 02:26:08,877
reputation system for the reviewers.

2482
02:26:08,960 --> 02:26:10,957
>> Absolutely. For the reviewing entities.

2483
02:26:11,040 --> 02:26:12,397
Not the reviewers individually.

2484
02:26:12,480 --> 02:26:13,677
>> The reviewing entities, sure.

2485
02:26:13,760 --> 02:26:17,157
>> But even that within that
the reviewers too, because uh

2486
02:26:17,240 --> 02:26:19,397
I There's another thing here. It's not

2487
02:26:19,480 --> 02:26:22,197
just the reputation, it's an incentive

2488
02:26:22,280 --> 02:26:24,677
for an individual person to do great.

2489
02:26:24,760 --> 02:26:29,117
Right now in in the academic setting,
the incentive is kind of uh

2490
02:26:29,200 --> 02:26:32,677
internal, just wanting to do a good job,
but honestly that's not a strong enough

2491
02:26:32,760 --> 02:26:36,237
incentive to do a really good job
in reading a paper, in finding the

2492
02:26:36,320 --> 02:26:39,997
beautiful amidst the mistakes and the
flaws and all that kind of stuff. Like

2493
02:26:40,080 --> 02:26:44,397
if you're the person that first
discovered a powerful paper,

2494
02:26:44,480 --> 02:26:49,517
and you get to be proud of that discovery,
then that gives a huge incentive to you.

2495
02:26:49,600 --> 02:26:50,837
>> That's That's a big
part of my proposal,

2496
02:26:50,920 --> 02:26:52,877
actually. You know, I describe that as

2497
02:26:52,960 --> 02:26:58,117
you know, if if your evaluation
of papers is predictive of future success,

2498
02:26:58,200 --> 02:26:59,437
>> Yes. Yes. That's a good one.

2499
02:26:59,520 --> 02:27:03,477
>> then your reputation should
go up as a reviewing entity.

2500
02:27:03,560 --> 02:27:06,957
Um So, yeah, exactly. I mean, that that um

2501
02:27:07,040 --> 02:27:11,397
I even had a master's student who was um
a master's student in uh library science

2502
02:27:11,480 --> 02:27:14,797
and computer science. Actually,
kind of worked out exactly how should that

2503
02:27:14,880 --> 02:27:16,997
should work with formulas
and everything, but

2504
02:27:17,080 --> 02:27:20,517
>> So, in terms of implementation, do
you think that's something that's doable?

2505
02:27:20,600 --> 02:27:23,477
>> I mean, I've been sort of, you know,
talking about this to sort of various

2506
02:27:23,560 --> 02:27:27,757
people like, you know, Andrew
McCallum who started OpenReview.

2507
02:27:27,840 --> 02:27:29,637
>> [snorts] >> And the reason
why we picked OpenReview

2508
02:27:29,720 --> 02:27:31,517
for ICLR initially, even though it was

2509
02:27:31,600 --> 02:27:33,957
very early for them, um is because my

2510
02:27:34,040 --> 02:27:38,557
hope was that ICLR was it
was eventually going to kind of

2511
02:27:38,640 --> 02:27:40,917
inaugurate this type of uh system. So, I

2512
02:27:41,000 --> 02:27:43,797
clear kept the idea of open reviews.

2513
02:27:43,880 --> 02:27:47,517
Uh, so whether reviews are you know,
published uh with the paper, which I

2514
02:27:47,600 --> 02:27:51,677
think is very useful. But,
in many ways has kind of reverted

2515
02:27:51,760 --> 02:27:56,157
to kind of more of a conventional type
conferences for everything else. And

2516
02:27:56,240 --> 02:28:00,677
that I mean, I I don't
run I clear. I'm just the

2517
02:28:00,760 --> 02:28:06,357
president of the uh of the foundation, but
um you know, people who run it should make

2518
02:28:06,440 --> 02:28:09,597
decisions about how to run it and I'm
not going to tell them um because

2519
02:28:09,680 --> 02:28:13,957
they're volunteers and I'm really
thankful that they do that. So, but I'm

2520
02:28:14,040 --> 02:28:15,997
saddened by the fact that uh we're not

2521
02:28:16,080 --> 02:28:18,877
being um innovative enough.

2522
02:28:18,960 --> 02:28:23,197
>> Yeah, me too. I hope that changes.
Uh, yeah. Cuz the communication [snorts]

2523
02:28:23,280 --> 02:28:27,237
science broadly,
but communication computer science ideas

2524
02:28:27,320 --> 02:28:30,317
is is is how you make those
ideas have impact, I think.

2525
02:28:30,400 --> 02:28:31,677
>> Yeah, and I think you know, a lot of

2526
02:28:31,760 --> 02:28:35,077
this is um because people have in their

2527
02:28:35,160 --> 02:28:38,877
mind kind of an objective which is you

2528
02:28:38,960 --> 02:28:43,677
know, fairness for authors
and the ability to count points

2529
02:28:43,760 --> 02:28:46,037
basically and and give credits accurately.

2530
02:28:46,120 --> 02:28:50,357
>> Mhm. >> But, that comes at the
expense of the progress of science.

2531
02:28:50,440 --> 02:28:53,917
>> Yeah. >> So, to some extent we're
slowing down the progress of science.

2532
02:28:54,000 --> 02:28:56,197
>> And are we actually achieving fairness?

2533
02:28:56,280 --> 02:29:00,077
>> And we're not achieving fairness,
you know, we have biases, you know, we're

2534
02:29:00,160 --> 02:29:05,797
doing you know, a double blind review, but
uh you know, the the the biases are are

2535
02:29:05,880 --> 02:29:08,477
still there. They're different
different kinds of biases.

2536
02:29:08,560 --> 02:29:12,677
>> You write that the phenomenon of
emergence, collective behavior exhibited

2537
02:29:12,760 --> 02:29:17,197
by large collection of simple elements
in interaction is one of the things that

2538
02:29:17,280 --> 02:29:19,197
got you into neural nets in the first

2539
02:29:19,280 --> 02:29:24,517
place. I love cellular automata. I love
simple interacting elements and uh the

2540
02:29:24,600 --> 02:29:27,197
things that emerge from them. Do you think

2541
02:29:27,280 --> 02:29:31,677
we understand how complex systems can
emerge from such simple components that

2542
02:29:31,760 --> 02:29:34,197
interact simply? >> No, we don't.

2543
02:29:34,280 --> 02:29:36,997
It's a big mystery. It also is a mystery
for physicists. It's a mystery for

2544
02:29:37,080 --> 02:29:42,037
biologists. Um you know,
how is it that uh >> [snorts]

2545
02:29:42,120 --> 02:29:45,757
>> the uh universe around us seems
to be increasing in complexity and not

2546
02:29:45,840 --> 02:29:49,197
decreasing? I mean,
that that that is a kind of

2547
02:29:49,280 --> 02:29:52,717
curious property of uh physics
that despite the second law of

2548
02:29:52,800 --> 02:29:54,277
thermodynamics,

2549
02:29:54,360 --> 02:29:57,037
we seem to be, you know, evolution and

2550
02:29:57,120 --> 02:30:00,437
and learning and uh etc. seems to be

2551
02:30:00,520 --> 02:30:04,757
kind of at least locally uh to increase uh

2552
02:30:04,840 --> 02:30:07,917
complexity and not decrease it. So,
perhaps the ultimate purpose of the

2553
02:30:08,000 --> 02:30:10,957
universe is uh to just get more complex.

2554
02:30:11,040 --> 02:30:16,637
>> Have these uh I mean, uh small
pockets of beautiful complexity.

2555
02:30:16,720 --> 02:30:19,317
Uh does that the cellular automata, these

2556
02:30:19,400 --> 02:30:21,677
kinds of emergence and complex systems

2557
02:30:21,760 --> 02:30:26,997
give you some intuition or guide
your understanding of machine learning

2558
02:30:27,080 --> 02:30:28,597
systems and neural networks and so on?

2559
02:30:28,680 --> 02:30:31,397
Or are these for you right
now disparate concepts?

2560
02:30:31,480 --> 02:30:36,477
>> Well, you got you got me into it. You
know, I I discovered the existence of uh

2561
02:30:36,560 --> 02:30:40,357
the perceptron when I
was a college student. Um

2562
02:30:40,440 --> 02:30:44,797
you know, by reading a book on it was
a debate between Chomsky and Piaget and

2563
02:30:44,880 --> 02:30:48,637
Seymour Papert from MIT was kind
of singing the praise of the perceptron in

2564
02:30:48,720 --> 02:30:52,037
that in the book and I I first time I
heard about Turing machine. Right? So, I

2565
02:30:52,120 --> 02:30:55,757
started digging the literature and I
found those paper those books which were

2566
02:30:55,840 --> 02:31:00,477
basically transcri- transcription of,
you know, workshops or conferences

2567
02:31:00,560 --> 02:31:03,877
uh from the '50s and '60s
about self-organizing systems.

2568
02:31:03,960 --> 02:31:07,277
Mhm. So, there were there was a series
of conferences on self-organizing

2569
02:31:07,360 --> 02:31:11,917
systems and the these books on this um
some of them are you can actually get

2570
02:31:12,000 --> 02:31:16,077
them at the internet archive,
you know, the digital version.

2571
02:31:16,160 --> 02:31:21,077
Uh and there are like uh fascinating
articles in there by this guy whose name

2572
02:31:21,160 --> 02:31:24,677
has been largely forgotten,
Heinz von Foerster.

2573
02:31:24,760 --> 02:31:29,357
So, it's a German physicist who
immigrated to the US and uh worked on

2574
02:31:29,440 --> 02:31:33,277
self-organizing systems in the in the 50s

2575
02:31:33,360 --> 02:31:36,917
and in the 60s he created at University
of Illinois Urbana-Champaign he created

2576
02:31:37,000 --> 02:31:41,877
the Biological Computer Laboratory
BCL which was you know all about neural

2577
02:31:41,960 --> 02:31:44,757
nets. Unfortunately
that was kind of towards

2578
02:31:44,840 --> 02:31:49,117
the end of the popularity of neural nets
so that that lab never kind of strive

2579
02:31:49,200 --> 02:31:52,557
very much but but he wrote a bunch of
papers about self-organization and about

2580
02:31:52,640 --> 02:31:55,477
the the mystery of self-organization. An

2581
02:31:55,560 --> 02:32:00,077
example he has is you take imagine you
are in space there's no gravity you have

2582
02:32:00,160 --> 02:32:04,037
a big box with magnets in it okay

2583
02:32:04,120 --> 02:32:07,477
you know kind of rectangular
magnets with north pole on one end

2584
02:32:07,560 --> 02:32:11,197
south pole on the other end. You shake
the box gently and the magnets will kind

2585
02:32:11,280 --> 02:32:17,117
of stick to themselves and probably form
a complex structure you know spontaneously

2586
02:32:17,200 --> 02:32:19,597
you know that could be an example
of self-organization but you know you have

2587
02:32:19,680 --> 02:32:23,157
lots of examples neural nets are
an example of self-organization too

2588
02:32:23,240 --> 02:32:27,597
you know in many respect.
And it's it's a bit of a mystery

2589
02:32:27,680 --> 02:32:31,117
you know how like what
what is possible with this

2590
02:32:31,200 --> 02:32:33,877
you know pattern formation
in physical systems

2591
02:32:33,960 --> 02:32:37,997
in chaotic system and things like that
you know you know the emergence of life

2592
02:32:38,080 --> 02:32:41,437
you know things like that so you
know how does does that happen?

2593
02:32:41,520 --> 02:32:44,477
It's a it's big puzzle
for for physicists as well.

2594
02:32:44,560 --> 02:32:47,117
>> It feels like
understanding this the the

2595
02:32:47,200 --> 02:32:52,317
mathematics of emergence in some
constrained situations might help us

2596
02:32:52,400 --> 02:32:57,837
create intelligence. Like help us
add a little spice to the systems

2597
02:32:57,920 --> 02:33:01,437
because you seem to be able to

2598
02:33:01,520 --> 02:33:04,077
in complex systems with emergence to be

2599
02:33:04,160 --> 02:33:06,517
able to get a lot from little.

2600
02:33:06,600 --> 02:33:11,637
And so that seems like a shortcut
to get big leaps in performance.

2601
02:33:11,720 --> 02:33:14,317
But um >> But there's
there's a missing concept

2602
02:33:14,400 --> 02:33:16,877
theoretical concept
that we are we don't have.

2603
02:33:16,960 --> 02:33:19,437
>> Yeah. >> And it's it's
something also I've been

2604
02:33:19,520 --> 02:33:22,677
fascinated by since my undergrad days

2605
02:33:22,760 --> 02:33:25,557
And it's how you measure complexity.

2606
02:33:25,640 --> 02:33:29,717
Right. So, we don't actually have good
ways of measuring or at least we don't

2607
02:33:29,800 --> 02:33:33,197
have good ways of interpreting
the measures that we we have at our

2608
02:33:33,280 --> 02:33:36,477
disposal. Like, how do you measure
the complexity of something, right?

2609
02:33:36,560 --> 02:33:38,317
So, there's all those things,
you know, like, you know,

2610
02:33:38,400 --> 02:33:42,317
Kolmogorov-Chaitin-Solomonoff complexity
of, you know, the length of the shortest

2611
02:33:42,400 --> 02:33:46,357
program that would generate a bit string
can be thought of as the complexity of

2612
02:33:46,440 --> 02:33:47,837
that bit string, right?

2613
02:33:47,920 --> 02:33:49,957
Um I've been fascinated by that concept.

2614
02:33:50,040 --> 02:33:53,437
But, the problem with that is that

2615
02:33:53,520 --> 02:33:57,837
that complexity is defined up to a
constant, which can be very large.

2616
02:33:57,920 --> 02:34:00,957
Right? Uh there's there are similar
concepts that are derived from, you

2617
02:34:01,040 --> 02:34:05,317
know, you know,
Bayesian probability theory,

2618
02:34:05,400 --> 02:34:09,597
where, you know, the complexity
of something is uh negative log of its

2619
02:34:09,680 --> 02:34:12,717
probability, essentially, right? And you
have a complete equivalence between the

2620
02:34:12,800 --> 02:34:15,157
two things. And there you
would think, you know, the

2621
02:34:15,240 --> 02:34:18,677
probability is something that's well
defined mathematically, which means

2622
02:34:18,760 --> 02:34:22,517
complexity is well defined. But, it's
not true. You need to have a model of uh

2623
02:34:22,600 --> 02:34:25,797
of of the distribution.
Uh you may need to have a prior if

2624
02:34:25,880 --> 02:34:29,317
you're doing Bayesian inference. And the
prior plays the same role as the choice

2625
02:34:29,400 --> 02:34:32,477
of the computer with which you
measure your Kolmogorov complexity.

2626
02:34:32,560 --> 02:34:37,197
And so, every measure of complexity
we have has some arbitrariness in it.

2627
02:34:37,280 --> 02:34:40,757
Um you know, an additive constant which

2628
02:34:40,840 --> 02:34:45,717
is can be arbitrarily large. And so,
you know, how can we come up with a good

2629
02:34:45,800 --> 02:34:48,317
theory of how things become more complex
if we don't have a good measure of

2630
02:34:48,400 --> 02:34:52,317
complexity? >> Yeah.
Which we need for this one way

2631
02:34:52,400 --> 02:34:56,397
that people study this in space biology,
the people that study the origin of life

2632
02:34:56,480 --> 02:34:59,757
or try to recreate
the life in in laboratory.

2633
02:34:59,840 --> 02:35:03,957
And the the more interesting one is the
alien one is when we go to other planets,

2634
02:35:04,040 --> 02:35:06,637
how do we recognize that it's life?

2635
02:35:06,720 --> 02:35:10,997
Cuz you know, complexity we associate
complexity maybe some level of mobility

2636
02:35:11,080 --> 02:35:14,277
with life. You know,
we have to be able to like

2637
02:35:14,360 --> 02:35:16,677
have concrete

2638
02:35:16,760 --> 02:35:21,877
algorithms for like um measuring
the level of complexity we

2639
02:35:21,960 --> 02:35:25,277
see to in order to know the difference
between life and non-life.

2640
02:35:25,360 --> 02:35:26,717
>> And the problem
is that complexity is in

2641
02:35:26,800 --> 02:35:29,077
the eye of the beholder. So, let me give

2642
02:35:29,160 --> 02:35:31,357
you an example. If I um

2643
02:35:31,440 --> 02:35:34,997
if I give you uh an image of the MNIST

2644
02:35:35,080 --> 02:35:39,877
digits, right? And I I flip through
MNIST digits. There is some obviously some

2645
02:35:39,960 --> 02:35:43,237
structure to it because local
structure, you know, neighboring

2646
02:35:43,320 --> 02:35:45,797
pixels are correlated uh

2647
02:35:45,880 --> 02:35:48,037
across the entire data set.

2648
02:35:48,120 --> 02:35:53,837
Now, imagine that um I I apply
a random permutation to all the pixels.

2649
02:35:53,920 --> 02:35:56,477
>> Mhm. >> A fixed random permutation.

2650
02:35:56,560 --> 02:36:01,397
Now, I show you those images. They will
look you know, really disorganized to you.

2651
02:36:01,480 --> 02:36:04,717
More complex. In fact,
they're not more complex in

2652
02:36:04,800 --> 02:36:08,397
absolute terms. They're exactly the same
as originally, right? And if you knew

2653
02:36:08,480 --> 02:36:11,957
what the permutation was, you know,
you could undo the permutation.

2654
02:36:12,040 --> 02:36:16,597
Now, imagine I give you special
glasses that undo that permutation.

2655
02:36:16,680 --> 02:36:19,477
Now, all of a sudden what looked
complicated becomes simple.

2656
02:36:19,560 --> 02:36:22,637
>> Right. >> So, if you
have two if you have, you

2657
02:36:22,720 --> 02:36:26,597
know, humans on one end and then another
race of aliens that sees the universe

2658
02:36:26,680 --> 02:36:29,397
with permutation glasses, >> Yeah,
with the permutation glasses.

2659
02:36:29,480 --> 02:36:32,357
>> Okay. >> [laughter] >> What
we perceive as simple to them is

2660
02:36:32,440 --> 02:36:34,237
hardly complicated. It's probably heat.

2661
02:36:34,320 --> 02:36:39,477
>> Yeah. Heat, yeah. >> Okay. And what
they perceive as simple to us is is random

2662
02:36:39,560 --> 02:36:42,157
uh fluctuation. It's heat. >> Yeah.

2663
02:36:42,240 --> 02:36:44,717
>> Uh so >> in the eye of the beholder.

2664
02:36:44,800 --> 02:36:46,877
>> Depends what kind
of glasses you're wearing.

2665
02:36:46,960 --> 02:36:50,317
>> Depends what kind of algorithm
you're running in your perception system.

2666
02:36:50,400 --> 02:36:54,117
>> So, I don't think we'll have a theory
of intelligence, self-organization,

2667
02:36:54,200 --> 02:36:57,077
evolution, things like this, until we

2668
02:36:57,160 --> 02:37:01,397
have a good handle on a notion
of complexity, which we know is in the eye

2669
02:37:01,480 --> 02:37:05,837
of the eye of the beholder. >> [sighs]
>> Yeah, it's sad to think that we might

2670
02:37:05,920 --> 02:37:10,357
not be able to detect or interact
with alien species because we're wearing

2671
02:37:10,440 --> 02:37:13,517
different glasses. >> Um because
their notion of locality may

2672
02:37:13,600 --> 02:37:14,837
be different from ours. >> Yeah.

2673
02:37:14,920 --> 02:37:17,757
>> This actually connects with uh
fascinating questions in physics at the

2674
02:37:17,840 --> 02:37:21,597
moment, like modern physics.
Uh quantum physics, like you know,

2675
02:37:21,680 --> 02:37:25,037
questions about like you know, can we
recover the information that's lost in a

2676
02:37:25,120 --> 02:37:26,477
black hole and things like that, right?

2677
02:37:26,560 --> 02:37:30,957
And uh and that relies
on notions of complexity. Um

2678
02:37:31,040 --> 02:37:33,597
Yeah, which you know,
I find I find it fascinating.

2679
02:37:33,680 --> 02:37:35,397
>> Can you describe your personal quest to

2680
02:37:35,480 --> 02:37:39,237
build an expressive electronic wind

2681
02:37:39,320 --> 02:37:44,077
instrument, EWI?
What is it? What does it take

2682
02:37:44,160 --> 02:37:47,557
to uh to build it? >> Well,
I'm a tinkerer. I like building

2683
02:37:47,640 --> 02:37:51,517
things. Uh I like building things
with combinations of electronics and, you

2684
02:37:51,600 --> 02:37:55,597
know, mechanical stuff.
Um you know, I have a bunch of different

2685
02:37:55,680 --> 02:37:59,517
hobbies, but um you know,
probably my first one was

2686
02:37:59,600 --> 02:38:02,397
little was building model airplanes and
stuff like that, and I still do that to

2687
02:38:02,480 --> 02:38:05,157
some extent. Uh but also
electronics. I taught myself

2688
02:38:05,240 --> 02:38:09,517
electronics before I studied it.
Um and the reason I taught myself

2689
02:38:09,600 --> 02:38:13,477
electronics is because
of music. Um my cousin

2690
02:38:13,560 --> 02:38:17,397
was an inspiring electronic musician and
he had an analog synthesizer, and I was,

2691
02:38:17,480 --> 02:38:19,277
you know, basically modifying it for him.

2692
02:38:19,360 --> 02:38:21,237
>> Mhm. [snorts] >> Uh
and building sequencers and stuff

2693
02:38:21,320 --> 02:38:24,557
like that, right, for him. I was I
was in high school when I was doing this.

2694
02:38:24,640 --> 02:38:28,557
>> And the interest in like progressive
rock, like '80s, like what's what's the

2695
02:38:28,640 --> 02:38:31,397
greatest band of all time
according to Yann LeCun?

2696
02:38:31,480 --> 02:38:35,357
>> Um there's too there's too many of
them, but you know, it's a combination of

2697
02:38:35,440 --> 02:38:38,757
uh uh you [clears throat] know, Mahavishnu

2698
02:38:38,840 --> 02:38:44,437
Orchestra, uh Weather Report,
Yes, Genesis, uh you

2699
02:38:44,520 --> 02:38:47,877
know, >> Genesis.
>> pre-Peter Gabriel, >> Uh-huh.

2700
02:38:47,960 --> 02:38:51,037
>> um >> [laughter] >> Gentle Giant,
you know, things like that.

2701
02:38:51,120 --> 02:38:55,197
>> Great. Okay, so this this love
of electronics and this love of music

2702
02:38:55,280 --> 02:38:59,277
combine together. >> Right. So,
I was actually trying to play uh

2703
02:38:59,360 --> 02:39:03,437
uh Baroque and Renaissance uh
music, and I I played in a uh

2704
02:39:03,520 --> 02:39:07,677
orchestra when I was in high
school and uh first year of college.

2705
02:39:07,760 --> 02:39:11,517
And I played the recorder, crumhorn,
uh uh a little bit of oboe, you know,

2706
02:39:11,600 --> 02:39:14,997
things like that. So, I'm a I'm a wind
instrument player. But, I always wanted

2707
02:39:15,080 --> 02:39:18,677
to play improvised music, even though I
don't know anything about it. Uh and the

2708
02:39:18,760 --> 02:39:23,197
only way I figured, you know, short
of like learning to play the saxophone was

2709
02:39:23,280 --> 02:39:27,277
to play electronic wind instrument.
So, they behave you know, the fingering is

2710
02:39:27,360 --> 02:39:31,157
similar to a saxophone, but you know,
you have a wide variety of sound because

2711
02:39:31,240 --> 02:39:34,997
you control the synthesizer with it.
So, I I had a a bunch of those, you know,

2712
02:39:35,080 --> 02:39:37,077
going back to the late '80s >> [snorts]

2713
02:39:37,160 --> 02:39:41,157
>> uh from either Yamaha or Akai. They

2714
02:39:41,240 --> 02:39:44,757
They're both kind of the main
manufacturers of those. So, they were

2715
02:39:44,840 --> 02:39:47,357
classically, you know,
going back several decades.

2716
02:39:47,440 --> 02:39:50,277
Uh but, I've never been completely
satisfied with them because of lack of

2717
02:39:50,360 --> 02:39:54,357
expressivity. And you know,
those things, you know,

2718
02:39:54,440 --> 02:39:56,757
are somewhat expressive. I mean,
they measure the breath pressure, they

2719
02:39:56,840 --> 02:40:00,437
measure the lip pressure, and you know, uh

2720
02:40:00,520 --> 02:40:04,317
you have like various parameters you can
you can vary with with fingers, but they

2721
02:40:04,400 --> 02:40:08,997
they they're not really as expressive
as a acoustic instrument, right? You

2722
02:40:09,080 --> 02:40:12,237
you hear John Coltrane play two
notes and you hear you know it's John

2723
02:40:12,320 --> 02:40:14,157
Coltrane, you know
it's got a unique sound.

2724
02:40:14,240 --> 02:40:18,117
>> Mhm. >> Uh or or Miles Davis,
right? You can hear it's Miles Davis

2725
02:40:18,200 --> 02:40:22,357
uh playing the trumpet
because the the sound

2726
02:40:22,440 --> 02:40:27,077
reflects their you know,
physiognomy, basically, the

2727
02:40:27,160 --> 02:40:32,077
shape of the vocal tract um kind
of shapes the the sound. So, how

2728
02:40:32,160 --> 02:40:35,917
how do you do this with uh electronic
wind instrument? And it was um many

2729
02:40:36,000 --> 02:40:39,837
years ago I met a guy called David
Wessel. He he was a professor at

2730
02:40:39,920 --> 02:40:44,077
Berkeley and created the the
center for like, you know, music

2731
02:40:44,160 --> 02:40:46,917
technology there.
And he was interested in that question.

2732
02:40:47,000 --> 02:40:49,277
>> Mhm. >> [clears throat] >> And so,
I kept kind of thinking about

2733
02:40:49,360 --> 02:40:53,717
this for many years and finally, because
of COVID, you know, I was at home. I was

2734
02:40:53,800 --> 02:40:58,597
in my workshop. My workshop serves also
as my kind of zoom uh room and and and

2735
02:40:58,680 --> 02:41:00,677
home office. And uh >>
This is in New Jersey?

2736
02:41:00,760 --> 02:41:03,917
>> In New Jersey. And um I started uh

2737
02:41:04,000 --> 02:41:07,757
really being serious about you
know building my own EWI instrument.

2738
02:41:07,840 --> 02:41:09,557
>> What else is going
on in that New Jersey

2739
02:41:09,640 --> 02:41:11,397
workshop? Is there some Is there some

2740
02:41:11,480 --> 02:41:16,597
crazy stuff you built that just or or
like left on the workshop floor? Left

2741
02:41:16,680 --> 02:41:19,477
behind. >> Uh a lot
of crazy stuff is uh you know

2742
02:41:19,560 --> 02:41:21,397
electronics we built with

2743
02:41:21,480 --> 02:41:23,997
microcontrollers of various kinds and

2744
02:41:24,080 --> 02:41:28,137
you know weird flying contraptions.

2745
02:41:28,220 --> 02:41:30,637
>> [laughter] >> So you still love flying?

2746
02:41:30,720 --> 02:41:35,397
>> It's a family disease.
My My dad got me into it when I was a kid.

2747
02:41:35,480 --> 02:41:39,637
And he was building model
airplanes when he was a kid and

2748
02:41:39,720 --> 02:41:43,717
and he was a mechanical engineer. He
taught himself electronics also so he he

2749
02:41:43,800 --> 02:41:46,917
built his early radio
control systems in the

2750
02:41:47,000 --> 02:41:51,117
late '60s, early '70s.
Um and so that that that's what got me

2751
02:41:51,200 --> 02:41:54,517
into I mean he got me into kind
of you know engineering and science and

2752
02:41:54,600 --> 02:41:56,597
technology. >> Do you
also have an interest and

2753
02:41:56,680 --> 02:42:00,637
appreciation of flight in other forms
like with drones, quadrocopters or do

2754
02:42:00,720 --> 02:42:04,277
you Is it Is it Is it model
airplane the the thing that

2755
02:42:04,360 --> 02:42:08,357
>> I you know before
drones were you know kind of

2756
02:42:08,440 --> 02:42:10,037
consumer products

2757
02:42:10,120 --> 02:42:12,517
um you know I built my own you know with

2758
02:42:12,600 --> 02:42:17,357
also building a microcontroller
with gyroscopes and accelerometers for

2759
02:42:17,440 --> 02:42:20,557
stabilization writing the firmware for it
you know and then when it became kind

2760
02:42:20,640 --> 02:42:23,037
of a standard thing you could buy it
was boring you know I stopped doing it it

2761
02:42:23,120 --> 02:42:26,597
was not fun anymore. Um >> Yeah.

2762
02:42:26,680 --> 02:42:28,157
You were doing it before it was cool.

2763
02:42:28,240 --> 02:42:31,477
>> Yeah. >> What advice would
you give to a young person

2764
02:42:31,560 --> 02:42:34,197
today in high school in college that

2765
02:42:34,280 --> 02:42:37,077
dreams of doing something big like Yann

2766
02:42:37,160 --> 02:42:42,317
LeCun. Like let's talk in the space
of intelligence. Dreams of having a chance

2767
02:42:42,400 --> 02:42:45,917
to solve some fundamental
problem in space of intelligence

2768
02:42:46,000 --> 02:42:50,637
both for their career and just in life
being somebody who was a part of

2769
02:42:50,720 --> 02:42:55,637
creating something special.
>> So try to get

2770
02:42:55,720 --> 02:43:00,797
interested by big questions. Things like,
you know, what is intelligence? Uh

2771
02:43:00,880 --> 02:43:05,397
what is the universe made of?
What's life all about? Things like that.

2772
02:43:05,480 --> 02:43:09,557
Um Like even like crazy big questions like

2773
02:43:09,640 --> 02:43:13,277
um what's time?
Like nobody knows what time is.

2774
02:43:13,360 --> 02:43:18,077
Um And uh and then

2775
02:43:18,160 --> 02:43:23,357
learn basic things like basic methods
either from math from physics or from

2776
02:43:23,440 --> 02:43:27,157
engineering. Uh things
that have a long shelf life.

2777
02:43:27,240 --> 02:43:30,597
Um like if you have a choice
between like, you know,

2778
02:43:30,680 --> 02:43:34,517
learning uh you know,
mobile programming on iPhone

2779
02:43:34,600 --> 02:43:38,360
or quantum mechanics,
take quantum mechanics.

2780
02:43:38,520 --> 02:43:40,357
Um because you're going to learn things

2781
02:43:40,440 --> 02:43:45,477
that you have no idea exist. It and it
you may not you you you may never be a

2782
02:43:45,560 --> 02:43:48,677
quantum physicist but you
will learn about path integrals

2783
02:43:48,760 --> 02:43:51,837
and path integrals are used um everywhere.
It's the same formula that

2784
02:43:51,920 --> 02:43:55,277
you use for, you know, Bayesian
integration and stuff like that.

2785
02:43:55,360 --> 02:43:58,037
>> So the ideas the little ideas

2786
02:43:58,120 --> 02:44:03,397
within quantum mechanics or within some
of these kind of more solidified fields

2787
02:44:03,480 --> 02:44:08,277
will have a longer shelf life. They will
you somehow use indirectly in in your

2788
02:44:08,360 --> 02:44:11,917
work. >> Learn classical mechanics like
you learn about Lagrangians for example.

2789
02:44:12,000 --> 02:44:16,357
>> Mhm. >> Um which is like
a a huge hugely useful

2790
02:44:16,440 --> 02:44:19,237
concept, you know, for all
kinds of different things. Uh

2791
02:44:19,320 --> 02:44:22,517
learn uh statistical physics because uh

2792
02:44:22,600 --> 02:44:26,877
all the math that comes out
of you know, for machine learning

2793
02:44:26,960 --> 02:44:30,997
uh basically comes out of was figured
out by statistical physicist in the you

2794
02:44:31,080 --> 02:44:35,637
know, late 19th early 20th century, right?
So um and for some of them actually more

2795
02:44:35,720 --> 02:44:39,997
recently for by people like like Giorgio
Parisi who just got the Nobel Prize for

2796
02:44:40,080 --> 02:44:42,077
the replica method among other things.

2797
02:44:42,160 --> 02:44:45,957
Um it's used for a lot of different
things. Uh you know, variational

2798
02:44:46,040 --> 02:44:48,917
inference uh that math
comes from statistical

2799
02:44:49,000 --> 02:44:53,837
physics. Um so um so
a lot of those kind of, you

2800
02:44:53,920 --> 02:44:56,957
know, basic courses,
you know, you'll you'll

2801
02:44:57,040 --> 02:44:59,717
If you do electrical engineering,
you take signal processing, you'll you'll

2802
02:44:59,800 --> 02:45:01,757
learn about Fourier transforms.

2803
02:45:01,840 --> 02:45:04,037
Again, something super useful. It's at

2804
02:45:04,120 --> 02:45:06,037
the basis of things like graph neural

2805
02:45:06,120 --> 02:45:09,517
nets, which is a an entirely new

2806
02:45:09,600 --> 02:45:13,397
sub-area of, you know, AI machine
learning deep learning, which I think is

2807
02:45:13,480 --> 02:45:15,797
super promising for all
kinds of applications.

2808
02:45:15,880 --> 02:45:18,797
Um, something very promising if you're
more interested in applications is the

2809
02:45:18,880 --> 02:45:23,397
applications of, uh, AI machine
learning and deep learning to science.

2810
02:45:23,480 --> 02:45:26,997
Or to science that can
help solve big problems

2811
02:45:27,080 --> 02:45:29,437
in the world. I've I've colleagues at,

2812
02:45:29,520 --> 02:45:34,117
uh, at Meta at FAIR who started this
project called Open Catalyst, and it's,

2813
02:45:34,200 --> 02:45:36,437
um, it's an open project collaborative.

2814
02:45:36,520 --> 02:45:39,757
And the idea is to use,
uh, deep learning to

2815
02:45:39,840 --> 02:45:43,117
help design, uh, new chemical compounds

2816
02:45:43,200 --> 02:45:47,677
or materials that would facilitate
the separation of hydrogen from oxygen.

2817
02:45:47,760 --> 02:45:49,797
If you can efficiently separate oxygen

2818
02:45:49,880 --> 02:45:52,837
from hydrogen with electricity,

2819
02:45:52,920 --> 02:45:55,397
you you solve climate change.

2820
02:45:55,480 --> 02:45:59,957
It's as simple as that, because, um,
you you cover, you know, some random

2821
02:46:00,040 --> 02:46:03,957
desert with, uh, solar panels, uh,
and you have it have them work all

2822
02:46:04,040 --> 02:46:07,397
day, produce hydrogen, and then you ship
the hydrogen wherever it's needed. You

2823
02:46:07,480 --> 02:46:09,357
don't need anything else.

2824
02:46:09,440 --> 02:46:13,037
Um, you know, you you you have

2825
02:46:13,120 --> 02:46:15,917
controllable power that's, you know,

2826
02:46:16,000 --> 02:46:17,557
what can be transported, uh, anywhere.

2827
02:46:17,640 --> 02:46:20,917
So, if we if we have a large-scale

2828
02:46:21,000 --> 02:46:25,437
efficient, uh, energy storage
technology like producing

2829
02:46:25,520 --> 02:46:28,717
hydrogen, uh, we we solve
climate change. Here's

2830
02:46:28,800 --> 02:46:31,837
another way to solve climate change is,
uh, if you can figure out how to make

2831
02:46:31,920 --> 02:46:34,677
fusion work. Now, the problem with
fusion is that you make a super hot

2832
02:46:34,760 --> 02:46:38,637
plasma, and the plasma is unstable,
and you can't control it. Maybe with deep

2833
02:46:38,720 --> 02:46:41,517
learning you can find controllers
that will stabilize plasma and make, you

2834
02:46:41,600 --> 02:46:45,477
know, practical fusion reactors. I mean,
that's very speculative, but, you know,

2835
02:46:45,560 --> 02:46:50,197
it's worth trying because, um,
you know, it uh, the payoff is huge.

2836
02:46:50,280 --> 02:46:53,117
There's a group at Google
working on this led by John Platt.

2837
02:46:53,200 --> 02:46:55,997
>> So, control convert as many problems in

2838
02:46:56,080 --> 02:47:00,437
science and physics
and biology and chemistry into a

2839
02:47:00,520 --> 02:47:03,477
into a learnable problem and see
if a machine can learn it.

2840
02:47:03,560 --> 02:47:07,677
>> Right. I mean, there's properties of
you know, complex materials that we don't

2841
02:47:07,760 --> 02:47:10,277
understand from first
principle, for example.

2842
02:47:10,360 --> 02:47:14,677
Right? So, you know,
if we could design new

2843
02:47:14,760 --> 02:47:18,317
you know, new materials,
we could make more efficient batteries.

2844
02:47:18,400 --> 02:47:21,477
You know, we could make maybe faster
electronics. We could I mean, there's a

2845
02:47:21,560 --> 02:47:25,797
lot of things we can imagine
doing or you know, lighter

2846
02:47:25,880 --> 02:47:28,997
materials for for cars or airplanes
or things like that. Maybe better fuel

2847
02:47:29,080 --> 02:47:32,397
cells. There's all kinds of stuff we
can imagine. If we had good fuel cells,

2848
02:47:32,480 --> 02:47:36,677
hydrogen fuel cells, we could use
them to power airplanes and you know,

2849
02:47:36,760 --> 02:47:39,237
transportation wouldn't be or or cars.

2850
02:47:39,320 --> 02:47:43,237
We wouldn't have emission
problem CO2 emission problems

2851
02:47:43,320 --> 02:47:47,397
for for air transportation
anymore. So, this

2852
02:47:47,480 --> 02:47:52,037
there's a lot of those things I think
where AI you know, can be used it

2853
02:47:52,120 --> 02:47:55,877
and and this is not even talking about
all the sort of medicine, biology and

2854
02:47:55,960 --> 02:47:57,597
and everything like that, right?

2855
02:47:57,680 --> 02:48:01,197
You know, like protein folding,
you know, figuring out

2856
02:48:01,280 --> 02:48:03,877
like how can you design a protein
that it sticks to another protein that a

2857
02:48:03,960 --> 02:48:07,277
particular sites because that's
how you design drugs in the end.

2858
02:48:07,360 --> 02:48:10,637
So, you know, deep learning would be
useful and those are and those are kind

2859
02:48:10,720 --> 02:48:14,117
of you know, would be sort
of enormous progress if we could

2860
02:48:14,200 --> 02:48:17,477
use it for that.
Here's an example. If you take

2861
02:48:17,560 --> 02:48:19,397
this is like from recent material

2862
02:48:19,480 --> 02:48:23,197
physics. You take a monoatomic layer of

2863
02:48:23,280 --> 02:48:25,637
graphene, right? So, it's just carbon on

2864
02:48:25,720 --> 02:48:31,037
a hexagonal mesh and then make
this single single atom thick.

2865
02:48:31,120 --> 02:48:34,237
You put another one on top.
You twist them by some magic

2866
02:48:34,320 --> 02:48:36,717
number of degrees, 3 degrees or something.

2867
02:48:36,800 --> 02:48:40,437
It becomes superconductor.
Nobody has any idea why.

2868
02:48:40,520 --> 02:48:44,357
>> [laughter] >> I want
to know how that was discovered,

2869
02:48:44,440 --> 02:48:46,637
but that's the kind of thing that
machine learning can actually discover

2870
02:48:46,720 --> 02:48:50,037
these kinds of things.
>> maybe not, but but there is

2871
02:48:50,120 --> 02:48:51,997
a hint perhaps that with machine

2872
02:48:52,080 --> 02:48:54,797
learning we could train a system to

2873
02:48:54,880 --> 02:48:59,477
basically be a phenomenological model
of some complex emergent phenomenon which,

2874
02:48:59,560 --> 02:49:03,037
you know, superconductivity
is one of those.

2875
02:49:03,120 --> 02:49:07,757
Uh where, you know, think this collective
phenomenon is too difficult to describe

2876
02:49:07,840 --> 02:49:11,957
from first principles with the current,
you know, the usual sort of, you know,

2877
02:49:12,040 --> 02:49:15,757
reductionist type method.
But we could have

2878
02:49:15,840 --> 02:49:19,037
uh deep learning systems
that predict the properties

2879
02:49:19,120 --> 02:49:23,317
of a system from a description of it uh
after being trained with sufficiently uh

2880
02:49:23,400 --> 02:49:26,117
many samples. Um

2881
02:49:26,200 --> 02:49:28,517
There's um this guy Pascal Fua at EPFL.

2882
02:49:28,600 --> 02:49:32,517
He has a startup company um
that >> [snorts] >> where he he

2883
02:49:32,600 --> 02:49:36,557
basically trained a um
a convolutional net essentially to

2884
02:49:36,640 --> 02:49:39,877
predict the aerodynamic
properties of solids.

2885
02:49:39,960 --> 02:49:42,957
And you can generate as much data as you
want by just running uh computational

2886
02:49:43,040 --> 02:49:47,597
fluid dynamics, right?
So, you give a like a

2887
02:49:47,680 --> 02:49:51,677
a wing airfoil or something
shape of some kind,

2888
02:49:51,760 --> 02:49:55,957
and you run computational fluid dynamics,
you get as a result the drag

2889
02:49:56,040 --> 02:49:59,677
and, you know, uh lift
and all that stuff, right? And

2890
02:49:59,760 --> 02:50:03,077
you can you can generate lots of data,
train a neural net to make those

2891
02:50:03,160 --> 02:50:05,957
predictions, and now what you
have is a differentiable model

2892
02:50:06,040 --> 02:50:10,677
of, let's say, drag and and lift as a
function of the shape of that solid, and

2893
02:50:10,760 --> 02:50:13,837
so you can do back gradient descent,
you can optimize the shape so you get the

2894
02:50:13,920 --> 02:50:18,117
properties you want. >> [snorts] >>
Um >> Yeah, that's incredible. That's

2895
02:50:18,200 --> 02:50:21,597
incredible. And then
on top of all that, probably

2896
02:50:21,680 --> 02:50:22,597
you should read a little bit of

2897
02:50:22,680 --> 02:50:25,517
literature and a little bit of history

2898
02:50:25,600 --> 02:50:30,717
for inspiration and for wisdom, cuz
after all all of these technologies will

2899
02:50:30,800 --> 02:50:32,237
have to work in a human world.

2900
02:50:32,320 --> 02:50:34,477
>> Yes. >> And the human
world is complicated.

2901
02:50:34,560 --> 02:50:39,117
>> It is. So, yeah.
>> Yeah, and this is uh

2902
02:50:39,200 --> 02:50:41,957
an amazing conversation. I'm really
honored that you would talk with me

2903
02:50:42,040 --> 02:50:46,077
today. Thank you for all the amazing
work you're doing at FAIR at Meta. And

2904
02:50:46,160 --> 02:50:50,117
thank you for being so passionate after
all these years about everything that's

2905
02:50:50,200 --> 02:50:53,877
going on. You're you're a beacon of hope
for the machine learning community. And

2906
02:50:53,960 --> 02:50:56,557
thank you so much for spending your
valuable time with me today. That was

2907
02:50:56,640 --> 02:51:00,717
awesome. >> Thanks for having me on.
That was It was a pleasure.

2908
02:51:00,800 --> 02:51:03,757
>> Thanks for listening to this
conversation with Yann LeCun. To support

2909
02:51:03,840 --> 02:51:07,637
this podcast, please check out
our sponsors in the description.

2910
02:51:07,720 --> 02:51:12,517
And now, let me leave you
with some words from Isaac Asimov.

2911
02:51:12,600 --> 02:51:15,677
Your assumptions are
your windows on the world.

2912
02:51:15,760 --> 02:51:20,677
Scrub them off every once in a while,
or the light won't come in.

2913
02:51:20,760 --> 02:51:25,520
Thank you for listening,
and hope to see you next time.