1
00:00:01,680 --> 00:00:04,080
The following content is
provided under a Creative

2
00:00:04,080 --> 00:00:05,620
Commons license.

3
00:00:05,620 --> 00:00:07,920
Your support will help
MIT OpenCourseWare

4
00:00:07,920 --> 00:00:12,280
continue to offer high quality
educational resources for free.

5
00:00:12,280 --> 00:00:14,910
To make a donation or
view additional materials

6
00:00:14,910 --> 00:00:18,870
from hundreds of MIT courses,
visit MIT OpenCourseWare

7
00:00:18,870 --> 00:00:20,010
at ocw.mit.edu.

8
00:00:22,095 --> 00:00:24,720
WINRICH FREIWALD: So my talk is
going to be mostly about faces.

9
00:00:24,720 --> 00:00:25,830
And in many ways
I'm going to connect

10
00:00:25,830 --> 00:00:27,900
to what Jim DiCarlo
was talking about today

11
00:00:27,900 --> 00:00:30,246
and what Nancy
talked about today.

12
00:00:30,246 --> 00:00:31,620
I just thought
preparing for this

13
00:00:31,620 --> 00:00:34,203
that I should say a few things
about primates and intelligence

14
00:00:34,203 --> 00:00:37,170
and how face recognition
might be connected

15
00:00:37,170 --> 00:00:39,960
both to the species
we are studying

16
00:00:39,960 --> 00:00:42,614
and to the overall
question of intelligence.

17
00:00:42,614 --> 00:00:44,280
And I thought the
most appropriate thing

18
00:00:44,280 --> 00:00:45,780
to do this at MBL
would be to start

19
00:00:45,780 --> 00:00:47,482
with this kind of creature.

20
00:00:47,482 --> 00:00:49,440
So you might have seen
in the paper Nature that

21
00:00:49,440 --> 00:00:51,210
came out last week--

22
00:00:51,210 --> 00:00:53,610
that the genome of the
octopus was sequenced.

23
00:00:53,610 --> 00:00:55,980
It was really heralded
in the public press

24
00:00:55,980 --> 00:00:58,260
that it's finally proving
their intelligence.

25
00:00:58,260 --> 00:01:01,320
The argument is that
there are 33,000 genes,

26
00:01:01,320 --> 00:01:02,957
10,000 more than humans.

27
00:01:02,957 --> 00:01:04,790
That's of course not a
very strong argument.

28
00:01:04,790 --> 00:01:06,870
There are plants
with 45,000 genes.

29
00:01:06,870 --> 00:01:09,480
So that really doesn't tell you
very much about intelligence.

30
00:01:09,480 --> 00:01:11,104
But amongst those
genes that were found

31
00:01:11,104 --> 00:01:14,130
were lots of genes that are
important for development

32
00:01:14,130 --> 00:01:14,970
of the brain.

33
00:01:14,970 --> 00:01:17,030
And there is very
high heterogeneity

34
00:01:17,030 --> 00:01:20,150
in certain gene families that
are controlling the development

35
00:01:20,150 --> 00:01:21,180
of the nervous system.

36
00:01:21,180 --> 00:01:24,030
So it's one example of
how a better understanding

37
00:01:24,030 --> 00:01:27,404
the intelligence of other
creatures in neural terms,

38
00:01:27,404 --> 00:01:29,820
even in creatures like the
octopus which is very difficult

39
00:01:29,820 --> 00:01:30,437
to study.

40
00:01:30,437 --> 00:01:32,520
The point I would like to
stress about the octopus

41
00:01:32,520 --> 00:01:35,692
is there's really nothing
social about the species.

42
00:01:35,692 --> 00:01:37,650
Actually almost simultaneously
with this paper,

43
00:01:37,650 --> 00:01:40,440
there was one report
about sexual reproduction

44
00:01:40,440 --> 00:01:43,260
in one particular species of
octopus where it's not clear

45
00:01:43,260 --> 00:01:45,600
if it's violence
or more affection.

46
00:01:45,600 --> 00:01:47,520
But this is really
the one exception

47
00:01:47,520 --> 00:01:50,830
to otherwise a life that's
pretty much non-social at all.

48
00:01:50,830 --> 00:01:53,880
So in the octopus you have
the egg stage and then

49
00:01:53,880 --> 00:01:56,280
this larva which are
hatching very early on.

50
00:01:56,280 --> 00:01:59,530
There's no interaction with
mom or anything like this.

51
00:01:59,530 --> 00:02:01,000
They go to the
surface of the ocean

52
00:02:01,000 --> 00:02:03,570
and they start feeding and try
to grow as fast as they can

53
00:02:03,570 --> 00:02:06,570
because only a few of
them are going to survive.

54
00:02:06,570 --> 00:02:09,139
Then reproduction is a
very scary enterprise.

55
00:02:09,139 --> 00:02:12,330
In many octopus,
the male has to be

56
00:02:12,330 --> 00:02:15,960
careful not to be eaten by
the female in the process.

57
00:02:15,960 --> 00:02:18,954
When it's successful, then
actually the male stops eating.

58
00:02:18,954 --> 00:02:21,120
He's going to die even
before most of the youngsters

59
00:02:21,120 --> 00:02:22,140
are going to hatch.

60
00:02:22,140 --> 00:02:26,250
And all that mom is going to
do is basically to make sure

61
00:02:26,250 --> 00:02:29,550
that fresh water is going
to be delivered to the eggs.

62
00:02:29,550 --> 00:02:31,710
And that's it in
terms of social life.

63
00:02:31,710 --> 00:02:37,230
So you can be very intelligent
like an octopus and, who knows,

64
00:02:37,230 --> 00:02:39,310
Gabriel just mentioned
extraterrestrial life.

65
00:02:39,310 --> 00:02:42,259
Maybe other species out there
might also not be that social.

66
00:02:42,259 --> 00:02:43,800
And so there doesn't
necessarily have

67
00:02:43,800 --> 00:02:46,920
to be a connection between
sociality and intelligence.

68
00:02:46,920 --> 00:02:49,816
Second thing is it
can be very social,

69
00:02:49,816 --> 00:02:51,690
get a warm fuzzy feeling
of being with others

70
00:02:51,690 --> 00:02:53,730
and you get lots of
protection being in the group,

71
00:02:53,730 --> 00:02:55,020
but following your
instincts in this way

72
00:02:55,020 --> 00:02:56,970
doesn't necessarily
make you very smart.

73
00:02:56,970 --> 00:02:59,490
So I don't want to argue
against the connection

74
00:02:59,490 --> 00:03:01,470
between sociality
and intelligence

75
00:03:01,470 --> 00:03:03,500
in the form of
social intelligence.

76
00:03:03,500 --> 00:03:06,030
But we have to be careful
there's no necessary connection

77
00:03:06,030 --> 00:03:07,470
between those.

78
00:03:07,470 --> 00:03:09,880
However for primates
there is this idea,

79
00:03:09,880 --> 00:03:11,870
this social
intelligence hypothesis

80
00:03:11,870 --> 00:03:14,210
that really what made
primates so intelligent

81
00:03:14,210 --> 00:03:15,234
is their sociality.

82
00:03:15,234 --> 00:03:17,650
And so let's consider a little
bit what the arguments are.

83
00:03:17,650 --> 00:03:21,480
It was most strongly put
forward by Nick Humphrey in '76,

84
00:03:21,480 --> 00:03:24,327
but there are similar
precursors to this idea.

85
00:03:24,327 --> 00:03:25,410
So who are these primates?

86
00:03:25,410 --> 00:03:28,260
The primates are a
small group of mammals

87
00:03:28,260 --> 00:03:30,525
with about 400 plus species.

88
00:03:30,525 --> 00:03:31,400
They're very diverse.

89
00:03:31,400 --> 00:03:33,060
You can have very
small primates,

90
00:03:33,060 --> 00:03:36,690
just 30 grams, all the way
to 200 kilogram animals.

91
00:03:36,690 --> 00:03:41,110
They evolved, studying
65, 85 million years ago,

92
00:03:41,110 --> 00:03:42,600
there was this mass extinction.

93
00:03:42,600 --> 00:03:45,360
And so you have this high
diversity within the mammals

94
00:03:45,360 --> 00:03:47,170
starting from this point on.

95
00:03:47,170 --> 00:03:49,570
So they have certain things
in common with other mammals,

96
00:03:49,570 --> 00:03:52,080
but they're also very
special in many ways.

97
00:03:52,080 --> 00:03:54,410
All of the primate
species are social.

98
00:03:54,410 --> 00:03:56,940
They're not [INAUDIBLE]
social, but they are social.

99
00:03:56,940 --> 00:03:58,330
They develop very slowly.

100
00:03:58,330 --> 00:04:01,280
So it's really very
different from the octopus.

101
00:04:01,280 --> 00:04:03,780
A lot of investment is made
into the offspring There are not

102
00:04:03,780 --> 00:04:05,819
very many offspring.

103
00:04:05,819 --> 00:04:07,860
And the lifespan of these
animals is pretty long.

104
00:04:07,860 --> 00:04:10,890
The octopus life span
is three to five years.

105
00:04:10,890 --> 00:04:12,880
They're very visual,
rather than other mammals,

106
00:04:12,880 --> 00:04:15,190
which are very
olfactory oriented.

107
00:04:15,190 --> 00:04:17,490
They have binocular vision,
many have color vision,

108
00:04:17,490 --> 00:04:19,980
so vision is very
important for primates.

109
00:04:19,980 --> 00:04:23,820
And they have on average larger
brains than other mammals have.

110
00:04:23,820 --> 00:04:26,010
So our understanding of
the anatomy of primates

111
00:04:26,010 --> 00:04:28,485
and what makes this special
is actually very rudimentary.

112
00:04:28,485 --> 00:04:31,886
But there are a few points
that should be of interest.

113
00:04:31,886 --> 00:04:33,510
So if you look at
the mammalian brains,

114
00:04:33,510 --> 00:04:35,130
obviously they're very, very
complex mammalian brains.

115
00:04:35,130 --> 00:04:36,720
They're not primate
[INAUDIBLE] ones.

116
00:04:36,720 --> 00:04:38,910
And the main factor
really is body mass.

117
00:04:38,910 --> 00:04:41,650
So the bigger an
animal is, this is

118
00:04:41,650 --> 00:04:46,780
elementary of body weight
versus brain weight.

119
00:04:46,780 --> 00:04:48,900
And you can see there's
a log-log relationship

120
00:04:48,900 --> 00:04:49,980
between the two.

121
00:04:49,980 --> 00:04:52,560
That's something that you find
all over the animal kingdom.

122
00:04:52,560 --> 00:04:56,910
But if you compare primates
to other non-primate mammals

123
00:04:56,910 --> 00:04:58,980
you can see that there
is a larger increase

124
00:04:58,980 --> 00:05:00,532
with body mass of brain mass.

125
00:05:00,532 --> 00:05:02,990
And if you count the number of
neurons, the number of brain

126
00:05:02,990 --> 00:05:05,390
neurons, it is increasing
more steeply with body mass

127
00:05:05,390 --> 00:05:07,109
than it is in other mammals.

128
00:05:07,109 --> 00:05:08,900
This is obviously a
very crude measurement.

129
00:05:08,900 --> 00:05:10,320
There are others.

130
00:05:10,320 --> 00:05:12,350
If you look for brains
roughly the same weight,

131
00:05:12,350 --> 00:05:14,610
you can again see
like in primates, you

132
00:05:14,610 --> 00:05:17,030
have many more neurons
than in non-primates

133
00:05:17,030 --> 00:05:18,849
just by these two examples here.

134
00:05:18,849 --> 00:05:21,390
So there seems to be something
different in the organization.

135
00:05:21,390 --> 00:05:23,098
There are other measures
you can look at.

136
00:05:23,098 --> 00:05:26,450
So for example, the
neuron size is increasing

137
00:05:26,450 --> 00:05:27,740
with brain size in rodents.

138
00:05:27,740 --> 00:05:29,870
So a larger brain
in rodents is not

139
00:05:29,870 --> 00:05:32,000
necessarily one that
has many more neurons,

140
00:05:32,000 --> 00:05:33,750
but these neurons are
just getting bigger.

141
00:05:33,750 --> 00:05:35,770
And in primates this is
really not so much the case.

142
00:05:35,770 --> 00:05:37,311
So the size of the
neuron pretty much

143
00:05:37,311 --> 00:05:39,980
stays constant even if you
decrease the size of the brain.

144
00:05:39,980 --> 00:05:42,499
Or how much white matter do
you actually need per neuron

145
00:05:42,499 --> 00:05:43,790
to connect other brain regions?

146
00:05:43,790 --> 00:05:46,225
In rodents, apparently
the fiber caliber

147
00:05:46,225 --> 00:05:48,240
is increasing with brain size.

148
00:05:48,240 --> 00:05:50,240
So again, your brain
might grow just

149
00:05:50,240 --> 00:05:53,750
because the anatomy
of the basic element

150
00:05:53,750 --> 00:05:55,431
requires that it
will need more space.

151
00:05:55,431 --> 00:05:56,930
But in primates
that's not the case.

152
00:05:56,930 --> 00:05:58,430
If you get more
white brain matter,

153
00:05:58,430 --> 00:06:02,240
that's likely because the
connectivity is more complex.

154
00:06:02,240 --> 00:06:04,490
Primate brains also fold
faster with increasing size

155
00:06:04,490 --> 00:06:05,450
than rodent neurons.

156
00:06:05,450 --> 00:06:07,640
So these are all just
very coarse indications

157
00:06:07,640 --> 00:06:09,140
that maybe there's
something special

158
00:06:09,140 --> 00:06:12,680
about the primate brain compared
to other mammalian brains.

159
00:06:12,680 --> 00:06:14,990
So along the anatomy, I
mentioned that primates

160
00:06:14,990 --> 00:06:16,370
have forward-facing eyes.

161
00:06:16,370 --> 00:06:18,940
They can do binocular vision,
they have color vision.

162
00:06:18,940 --> 00:06:20,690
They have skulls
with a large cranium,

163
00:06:20,690 --> 00:06:24,140
that's something that makes
them special from other mammals.

164
00:06:24,140 --> 00:06:27,807
They're also special in other
ways that are important.

165
00:06:27,807 --> 00:06:29,390
If you think about
embodied cognition,

166
00:06:29,390 --> 00:06:31,223
you don't have to buy
into all these points,

167
00:06:31,223 --> 00:06:33,742
but obviously if you have
a hand as complex as ours

168
00:06:33,742 --> 00:06:35,450
which we share with
many primate species,

169
00:06:35,450 --> 00:06:36,950
there are lots of
things you can do.

170
00:06:36,950 --> 00:06:39,500
And that requires you to
be able to control it.

171
00:06:39,500 --> 00:06:42,020
And this gives you a power to
interact with the environment

172
00:06:42,020 --> 00:06:44,090
that other animals
might not have.

173
00:06:44,090 --> 00:06:45,950
So the shoulder is
more mobile, there's

174
00:06:45,950 --> 00:06:48,100
an opposable thumb
in many species.

175
00:06:48,100 --> 00:06:50,810
And then in the face, there are
changes that you might already

176
00:06:50,810 --> 00:06:51,740
have seen here.

177
00:06:51,740 --> 00:06:55,790
So if you go up to the
more complex animals,

178
00:06:55,790 --> 00:06:59,060
the snout region is becoming
increasingly reduced.

179
00:06:59,060 --> 00:07:02,190
And I will tell you later
why this might be important.

180
00:07:02,190 --> 00:07:04,690
So these are anatomical
specialization in primates.

181
00:07:04,690 --> 00:07:06,794
Sociality is very
important as well.

182
00:07:06,794 --> 00:07:08,960
And so there are four main
organizational principles

183
00:07:08,960 --> 00:07:10,100
of sociality in primates.

184
00:07:10,100 --> 00:07:12,140
The dominating one is
the second one here.

185
00:07:12,140 --> 00:07:14,120
It's called the male
transfer system.

186
00:07:14,120 --> 00:07:19,470
So it's a polygamous,
multi-male organization.

187
00:07:19,470 --> 00:07:22,040
This is very important for
the social life of primates

188
00:07:22,040 --> 00:07:24,650
because what it means is
that the social behavior has

189
00:07:24,650 --> 00:07:25,574
to be complex.

190
00:07:25,574 --> 00:07:27,740
So there can be cooperation,
like grooming, defense,

191
00:07:27,740 --> 00:07:29,573
and hunting which all
the animals of a troop

192
00:07:29,573 --> 00:07:30,630
might engage in.

193
00:07:30,630 --> 00:07:33,000
But at the same time there's
competition for food,

194
00:07:33,000 --> 00:07:34,779
mates dominance, hierarchies.

195
00:07:34,779 --> 00:07:36,320
And it's a function
of the complexity

196
00:07:36,320 --> 00:07:39,340
of the social environment.

197
00:07:39,340 --> 00:07:41,482
So primate social
life was beautifully

198
00:07:41,482 --> 00:07:43,440
inscribed by Dorothy
Cheney and Robert Seyfarth

199
00:07:43,440 --> 00:07:45,300
in this wonderful book
Baboon Metaphysics.

200
00:07:45,300 --> 00:07:46,883
And I'm just going
to quote from that.

201
00:07:46,883 --> 00:07:49,806
So they studied baboon
monkeys in the wild and here's

202
00:07:49,806 --> 00:07:51,180
what they have to
say about this.

203
00:07:51,180 --> 00:07:53,550
"The domain of expertise
for baboons, and indeed

204
00:07:53,550 --> 00:07:55,920
for all monkeys and
apes, is social life.

205
00:07:55,920 --> 00:07:58,380
Most baboons live in multi-male
and multi-female groups

206
00:07:58,380 --> 00:08:02,130
that typically include eight
or nine metrolineal families."

207
00:08:02,130 --> 00:08:04,172
Which means that the
females stay in the group

208
00:08:04,172 --> 00:08:05,630
and they found
families and they're

209
00:08:05,630 --> 00:08:08,155
going to stay constant
over a long period of time.

210
00:08:08,155 --> 00:08:09,780
"They have a linear
dominance hierarchy

211
00:08:09,780 --> 00:08:12,750
of males the changes often and
the linear hierarchy of females

212
00:08:12,750 --> 00:08:15,660
and the offsprings that can
be stable for generations.

213
00:08:15,660 --> 00:08:18,840
Daily life in a baboon group
includes small scale alliances

214
00:08:18,840 --> 00:08:20,550
that may involve only
three individuals

215
00:08:20,550 --> 00:08:23,460
and occasional large scale
familiar battles that

216
00:08:23,460 --> 00:08:27,030
involve all of the members
of three or four metrolines.

217
00:08:27,030 --> 00:08:29,400
Males and females can
form short term bonds that

218
00:08:29,400 --> 00:08:32,309
lead to reproduction, or longer
term friendships that lead

219
00:08:32,309 --> 00:08:34,770
to cooperative child rearing."

220
00:08:34,770 --> 00:08:36,600
"The result of all
this social intrigue

221
00:08:36,600 --> 00:08:38,549
is a kind of Jane
Austen melodrama

222
00:08:38,549 --> 00:08:41,669
in which each individual must
predict the behavior of others

223
00:08:41,669 --> 00:08:43,200
and form those
relationships that

224
00:08:43,200 --> 00:08:45,310
return the greatest benefits.

225
00:08:45,310 --> 00:08:47,680
These are the problems that
the baboon mind must solve

226
00:08:47,680 --> 00:08:49,890
and this is the environment
in which it has evolved."

227
00:08:49,890 --> 00:08:51,660
Most of the problems
facing baboons

228
00:08:51,660 --> 00:08:55,357
can be expressed in two
words: other baboons.

229
00:08:55,357 --> 00:08:56,690
And so this is really important.

230
00:08:56,690 --> 00:08:59,174
So again, if you're social
you don't necessarily

231
00:08:59,174 --> 00:09:00,090
have to be very smart.

232
00:09:00,090 --> 00:09:01,620
You can be very smart
and not be social.

233
00:09:01,620 --> 00:09:03,286
But there's something
special apparently

234
00:09:03,286 --> 00:09:08,010
about primates that links our
intelligence to our sociality.

235
00:09:08,010 --> 00:09:10,020
So again, the Social
Intelligence Hypothesis,

236
00:09:10,020 --> 00:09:13,950
what works in its favor is that
the primates have large brains.

237
00:09:13,950 --> 00:09:15,770
And primates,
apparently group size

238
00:09:15,770 --> 00:09:18,570
is correlated with brain size
across different species,

239
00:09:18,570 --> 00:09:19,830
and not the home-range size.

240
00:09:19,830 --> 00:09:21,270
This is an
alternative hypothesis

241
00:09:21,270 --> 00:09:24,680
is that maybe you have to forage
in more complex environments.

242
00:09:24,680 --> 00:09:26,700
And so a similar proxy
for the complexity

243
00:09:26,700 --> 00:09:28,639
of your social life
and your physical life

244
00:09:28,639 --> 00:09:30,180
predicts there's a
better correlation

245
00:09:30,180 --> 00:09:33,406
of brain size with the social
life than physical life.

246
00:09:33,406 --> 00:09:34,780
The complexity of
an individual's

247
00:09:34,780 --> 00:09:37,321
social relationships increases
exponentially with group size.

248
00:09:37,321 --> 00:09:38,580
And groups are not small.

249
00:09:38,580 --> 00:09:40,550
And we're going to get back
to this point in a little bit.

250
00:09:40,550 --> 00:09:42,660
The baboons and other primates
know their peer's dominance,

251
00:09:42,660 --> 00:09:43,770
rank, and social relations.

252
00:09:43,770 --> 00:09:45,300
Everything that I mentioned
in the previous slide

253
00:09:45,300 --> 00:09:47,160
is good evidence
for behavioral work

254
00:09:47,160 --> 00:09:49,890
that actually primates
know something about it.

255
00:09:49,890 --> 00:09:52,410
This social knowledge
contrasts with surprising cases

256
00:09:52,410 --> 00:09:54,660
of ignorance outside
of the social domain.

257
00:09:54,660 --> 00:09:58,870
Even when it's something
as important as a predator.

258
00:09:58,870 --> 00:10:01,480
So Cheney and Seyfard, they
also studied vervet monkeys.

259
00:10:01,480 --> 00:10:04,110
And what they observed
is that vervet monkeys,

260
00:10:04,110 --> 00:10:06,930
one of the main predators
to them is a python.

261
00:10:06,930 --> 00:10:09,390
And the python, if
it crawls in the sand

262
00:10:09,390 --> 00:10:10,720
it leaves behind a trail.

263
00:10:10,720 --> 00:10:12,540
And there is absolutely no
indication that the vervet

264
00:10:12,540 --> 00:10:14,540
monkeys make the connection
between these trails

265
00:10:14,540 --> 00:10:16,002
and the presence of the python.

266
00:10:16,002 --> 00:10:17,460
Which is really
striking, you would

267
00:10:17,460 --> 00:10:20,010
imagine this is the first thing
that they would have to learn.

268
00:10:20,010 --> 00:10:21,000
So there are cases
where they actually

269
00:10:21,000 --> 00:10:23,374
follow the trail into the bush
and they're very surprised

270
00:10:23,374 --> 00:10:24,730
to find a python there.

271
00:10:24,730 --> 00:10:26,532
Another example
is, are leopards.

272
00:10:26,532 --> 00:10:27,990
So in the environment,
leopards are

273
00:10:27,990 --> 00:10:31,389
of course predators who also
feed on the vervet monkeys.

274
00:10:31,389 --> 00:10:33,680
Leopards have a way of putting
the carcasses of animals

275
00:10:33,680 --> 00:10:36,390
they hunted down into the trees
to protect them from larger

276
00:10:36,390 --> 00:10:37,959
predators like lions.

277
00:10:37,959 --> 00:10:39,750
So the presence of a
carcass would actually

278
00:10:39,750 --> 00:10:42,600
indicate to you that likely
there is a leopard around.

279
00:10:42,600 --> 00:10:44,416
And again, the
vervet monkeys they

280
00:10:44,416 --> 00:10:45,540
don't make that connection.

281
00:10:45,540 --> 00:10:47,190
So if there is a
carcas there, they're

282
00:10:47,190 --> 00:10:49,000
not particularly
scared about it.

283
00:10:49,000 --> 00:10:50,110
It doesn't mean that
they're ignorant.

284
00:10:50,110 --> 00:10:52,350
So they're even following
alarm calls from other species

285
00:10:52,350 --> 00:10:54,000
as to whether an
eagle is approaching,

286
00:10:54,000 --> 00:10:55,890
or if cattle is approaching.

287
00:10:55,890 --> 00:10:59,010
So it's not that they're
generally dumb in this way,

288
00:10:59,010 --> 00:11:02,070
but there's really this very big
contrast about all the details

289
00:11:02,070 --> 00:11:04,560
that they know about their
social world and the obliviance

290
00:11:04,560 --> 00:11:09,942
that they can express
for non-social factors.

291
00:11:09,942 --> 00:11:11,400
Then we have
specializations, we're

292
00:11:11,400 --> 00:11:14,050
going to talk about this
in the brain for processing

293
00:11:14,050 --> 00:11:15,534
social stimuli.

294
00:11:15,534 --> 00:11:16,950
And then there's
actually evidence

295
00:11:16,950 --> 00:11:19,283
that females who have better
social abilities, that they

296
00:11:19,283 --> 00:11:22,020
are less stressed and they have
better reproductive success.

297
00:11:22,020 --> 00:11:26,100
And so this all works to
say that if you are socially

298
00:11:26,100 --> 00:11:28,671
smart in a baboon environment
or in the environment

299
00:11:28,671 --> 00:11:30,420
of many primate species,
you actually have

300
00:11:30,420 --> 00:11:31,716
better success of reproducing.

301
00:11:31,716 --> 00:11:33,090
Therefore there's
a good argument

302
00:11:33,090 --> 00:11:36,675
to be made that your social
intelligence will therefore

303
00:11:36,675 --> 00:11:37,950
go to the next generation.

304
00:11:37,950 --> 00:11:40,161
This is how social
intelligence might improve.

305
00:11:40,161 --> 00:11:41,910
I think there's one
important point that's

306
00:11:41,910 --> 00:11:42,842
oftentimes not made.

307
00:11:42,842 --> 00:11:45,300
And that's if you become smarter
and smarter in interacting

308
00:11:45,300 --> 00:11:47,820
with your physical environment,
your physical environment

309
00:11:47,820 --> 00:11:49,302
does not change very much.

310
00:11:49,302 --> 00:11:51,510
But if you are interacting
with a social environment,

311
00:11:51,510 --> 00:11:53,310
and getting smarter
and smarter interacting

312
00:11:53,310 --> 00:11:55,260
with the social
environment, the elements

313
00:11:55,260 --> 00:11:57,130
in your social environment
you're interacting with,

314
00:11:57,130 --> 00:11:58,838
they're also getting
smarter and smarter.

315
00:11:58,838 --> 00:12:00,990
So you're actually
setting forth an arms race

316
00:12:00,990 --> 00:12:03,960
where you're not only improving
the situation by getting

317
00:12:03,960 --> 00:12:05,560
smarter, but we
have to get smart

318
00:12:05,560 --> 00:12:08,184
in order to keep pace with the
others who are outsmarting you.

319
00:12:08,184 --> 00:12:09,600
And so you can see
how there could

320
00:12:09,600 --> 00:12:12,294
be a connection between
sociality and intelligence.

321
00:12:12,294 --> 00:12:13,710
That there's really
this arms race

322
00:12:13,710 --> 00:12:15,870
does not occur for
physical interactions

323
00:12:15,870 --> 00:12:17,470
but for social interactions.

324
00:12:17,470 --> 00:12:19,740
So that you actually have
to be able to better predict

325
00:12:19,740 --> 00:12:21,960
the next move of someone
else in your group

326
00:12:21,960 --> 00:12:24,870
and you have to know something
about that individual for you

327
00:12:24,870 --> 00:12:26,551
to be successful.

328
00:12:26,551 --> 00:12:28,800
So there are arguments against
the Social Intelligence

329
00:12:28,800 --> 00:12:30,000
Hypothesis.

330
00:12:30,000 --> 00:12:31,770
So, in particular,
we're ignorant

331
00:12:31,770 --> 00:12:33,280
about many other species.

332
00:12:33,280 --> 00:12:35,970
So there are other social
species, hyenas for example

333
00:12:35,970 --> 00:12:38,740
have complex societies, but
hyenas have not been studied

334
00:12:38,740 --> 00:12:41,950
as much as monkeys have.

335
00:12:41,950 --> 00:12:43,780
We also don't know
about the complexity,

336
00:12:43,780 --> 00:12:45,870
mostly for this reason,
where they're really

337
00:12:45,870 --> 00:12:48,400
compared to whales,
whales for example,

338
00:12:48,400 --> 00:12:50,251
primate societies
are more complex.

339
00:12:50,251 --> 00:12:52,250
And this of course would
be a crucial conjecture

340
00:12:52,250 --> 00:12:55,709
of this Social
Intelligence Hypothesis.

341
00:12:55,709 --> 00:12:57,250
Then within the
primate orders, there

342
00:12:57,250 --> 00:12:58,780
are actually some
other correlations

343
00:12:58,780 --> 00:13:00,322
that do predict
brain size very well.

344
00:13:00,322 --> 00:13:02,863
And so within the primate order,
social learning, innovation,

345
00:13:02,863 --> 00:13:05,140
and tool use are strongly
correlated with brain size

346
00:13:05,140 --> 00:13:06,330
and not with group size.

347
00:13:06,330 --> 00:13:08,830
So you could imagine a scenario
where actually the evolution

348
00:13:08,830 --> 00:13:11,200
of basic, social
intelligence is something

349
00:13:11,200 --> 00:13:13,150
that's very basic to primates.

350
00:13:13,150 --> 00:13:15,010
But then if you go
to different species

351
00:13:15,010 --> 00:13:17,470
within the primate
order and ask like why

352
00:13:17,470 --> 00:13:20,740
did they become so smart,
like orangutans or chimps who

353
00:13:20,740 --> 00:13:22,270
can use tools, then
it really might

354
00:13:22,270 --> 00:13:23,811
be the tool use that
would be of more

355
00:13:23,811 --> 00:13:25,674
importance than the sociality.

356
00:13:25,674 --> 00:13:28,090
So I have a movie here that
actually illustrates these two

357
00:13:28,090 --> 00:13:29,314
different hypotheses.

358
00:13:29,314 --> 00:13:31,230
So you see social
interactions here in a group

359
00:13:31,230 --> 00:13:32,410
of Tonkin macaque monkeys.

360
00:13:32,410 --> 00:13:33,910
You can see the
facial displays, you

361
00:13:33,910 --> 00:13:35,810
can see that they are
tending to each other.

362
00:13:35,810 --> 00:13:38,268
And well you might think that
they are trying to figure out

363
00:13:38,268 --> 00:13:39,670
what's actually going on here.

364
00:13:39,670 --> 00:13:43,030
And here's the alternative
hypothesis of this.

365
00:13:43,030 --> 00:13:44,060
This is tool use.

366
00:13:44,060 --> 00:13:47,390
So you can see this guy
just invented a cool tool,

367
00:13:47,390 --> 00:13:48,430
a nose pick.

368
00:13:48,430 --> 00:13:50,380
And so it's anyone's
guess what's

369
00:13:50,380 --> 00:13:52,105
more important to
your intelligence,

370
00:13:52,105 --> 00:13:53,980
you been able to read
the social significance

371
00:13:53,980 --> 00:13:55,396
of other individuals
in your troop

372
00:13:55,396 --> 00:13:59,990
or your ability to
invent a nice nose pick.

373
00:13:59,990 --> 00:14:03,995
So the last point I wanted
to make is a question.

374
00:14:03,995 --> 00:14:06,370
So are the primates' abilities
in social knowledge really

375
00:14:06,370 --> 00:14:08,440
intelligent or is
it just more like

376
00:14:08,440 --> 00:14:09,800
idiot savant-like abilities?

377
00:14:09,800 --> 00:14:12,850
So a unique specialization
where they're good at.

378
00:14:12,850 --> 00:14:15,710
The argument that Cheney and
Seyfard made is the following.

379
00:14:15,710 --> 00:14:17,668
So the knowledge that
they have should actually

380
00:14:17,668 --> 00:14:19,835
be true knowledge and just
not learned associations.

381
00:14:19,835 --> 00:14:21,667
And the reason has to
do with the complexity

382
00:14:21,667 --> 00:14:22,880
of the social environment.

383
00:14:22,880 --> 00:14:25,060
If you have 80 different
individuals, which

384
00:14:25,060 --> 00:14:27,580
is the typical case for
these baboon monkeys,

385
00:14:27,580 --> 00:14:32,999
you have 3,160 pairs of
animals and 82,160 trails.

386
00:14:32,999 --> 00:14:34,540
It's going to be
virtually impossible

387
00:14:34,540 --> 00:14:36,415
for you to learn all
these different pairwise

388
00:14:36,415 --> 00:14:39,760
relationships and then behave
intelligently based upon it.

389
00:14:39,760 --> 00:14:42,640
Second, these relationships
can change very fast.

390
00:14:42,640 --> 00:14:44,862
So it would not be very
smart to try and make

391
00:14:44,862 --> 00:14:46,570
a list of all these
pairwise interactions

392
00:14:46,570 --> 00:14:48,370
and then act upon that.

393
00:14:48,370 --> 00:14:49,960
No single behavioral
metric seems

394
00:14:49,960 --> 00:14:52,270
to be necessary or sufficient
to recognize associations

395
00:14:52,270 --> 00:14:53,920
like matrilinear kin.

396
00:14:53,920 --> 00:14:55,510
So human observers
are apparently not

397
00:14:55,510 --> 00:14:57,843
very good to predict this if
they don't know the animals

398
00:14:57,843 --> 00:14:59,547
very well personally.

399
00:14:59,547 --> 00:15:01,130
Then you might think,
well, maybe it's

400
00:15:01,130 --> 00:15:03,088
not physically like a
list but you don't really

401
00:15:03,088 --> 00:15:05,440
learn this much and you
apply a simple rule to it.

402
00:15:05,440 --> 00:15:07,484
And that also doesn't
seem to work very well

403
00:15:07,484 --> 00:15:09,400
because social relationships
like friendships,

404
00:15:09,400 --> 00:15:10,684
they are intransitive.

405
00:15:10,684 --> 00:15:12,100
So if A and B are
friends, B and C

406
00:15:12,100 --> 00:15:13,900
are friends, that doesn't
mean that A and C necessarily

407
00:15:13,900 --> 00:15:15,110
have to be friends.

408
00:15:15,110 --> 00:15:17,080
Others like family
relationships are complex,

409
00:15:17,080 --> 00:15:18,080
they're non-associative.

410
00:15:18,080 --> 00:15:19,900
So if A is the mother
of B, that actually

411
00:15:19,900 --> 00:15:21,850
means that B is not
the mother of A.

412
00:15:21,850 --> 00:15:24,560
So there's a more complex
interaction there as well.

413
00:15:24,560 --> 00:15:28,240
And then finally, there can
be simultaneous membership

414
00:15:28,240 --> 00:15:29,800
in multiple classes.

415
00:15:29,800 --> 00:15:32,050
And again, for you to be
able to keep track of this,

416
00:15:32,050 --> 00:15:33,424
you better have
a cognitive model

417
00:15:33,424 --> 00:15:36,190
of what's going on rather than
just a list of associations

418
00:15:36,190 --> 00:15:36,947
that you learned.

419
00:15:36,947 --> 00:15:38,530
It's very difficult
in experimentation

420
00:15:38,530 --> 00:15:40,000
to prove it's not
association, but I

421
00:15:40,000 --> 00:15:41,485
think these are very good
arguments to consider

422
00:15:41,485 --> 00:15:43,690
that actually these primates
have active knowledge

423
00:15:43,690 --> 00:15:45,950
of their social environment.

424
00:15:45,950 --> 00:15:48,220
So there's one example for
this where you can actually

425
00:15:48,220 --> 00:15:49,110
make the point very nicely.

426
00:15:49,110 --> 00:15:50,360
And this is the story of Ahla.

427
00:15:50,360 --> 00:15:56,020
Ahla is a baboon monkey
and she was actually

428
00:15:56,020 --> 00:15:58,580
living with farmers
in South West Africa.

429
00:15:58,580 --> 00:16:01,600
So there was a habit at the
time to actually replace

430
00:16:01,600 --> 00:16:05,080
dogs who were herding
goats with baboons.

431
00:16:05,080 --> 00:16:07,420
And so you can see
Ahla sitting here.

432
00:16:07,420 --> 00:16:10,220
You can see her
here adopting some

433
00:16:10,220 --> 00:16:11,560
of the behavior of the goats.

434
00:16:11,560 --> 00:16:14,590
So she's licking salt here,
which is something that baboons

435
00:16:14,590 --> 00:16:15,798
would naturally not do.

436
00:16:15,798 --> 00:16:18,006
She would continue to engage
in social behaviors that

437
00:16:18,006 --> 00:16:19,005
are typical for baboons.

438
00:16:19,005 --> 00:16:21,820
So she would groom the
goats, for example.

439
00:16:21,820 --> 00:16:24,140
But the most amazing
thing about her,

440
00:16:24,140 --> 00:16:26,290
which is a little hard to
see as you see her here,

441
00:16:26,290 --> 00:16:29,290
she's carrying one of the
little yearlings here,

442
00:16:29,290 --> 00:16:31,540
and brings it to its mother.

443
00:16:31,540 --> 00:16:34,000
And the description of
what this animal was doing

444
00:16:34,000 --> 00:16:36,370
was when the goats
were brought home

445
00:16:36,370 --> 00:16:39,820
and then sometimes they were
separating the mother goats

446
00:16:39,820 --> 00:16:43,210
from the offspring, she would
actually go manic and then

447
00:16:43,210 --> 00:16:44,980
try to pair them up
and she would not

448
00:16:44,980 --> 00:16:47,080
stop until she was finished.

449
00:16:47,080 --> 00:16:49,420
And this would even
happen as multiple goats

450
00:16:49,420 --> 00:16:52,250
were calling for the
yearlings and then vise versa.

451
00:16:52,250 --> 00:16:55,207
And so she would have to put
order into the social world

452
00:16:55,207 --> 00:16:57,040
that she was living in
and she wouldn't stop

453
00:16:57,040 --> 00:16:58,487
until this order was restored.

454
00:16:58,487 --> 00:17:00,070
And the farmer said
that they were not

455
00:17:00,070 --> 00:17:04,030
able to tell any of the
adult goals or the yearlings

456
00:17:04,030 --> 00:17:05,173
or know any of the pairs.

457
00:17:05,173 --> 00:17:06,339
But this was like her world.

458
00:17:06,339 --> 00:17:07,990
This was like a social
environment that she was in

459
00:17:07,990 --> 00:17:09,490
and she would
structure it according

460
00:17:09,490 --> 00:17:12,069
to her cognitive demands.

461
00:17:12,069 --> 00:17:13,599
So the point is
that primates have

462
00:17:13,599 --> 00:17:15,485
intricate social knowledge.

463
00:17:15,485 --> 00:17:17,359
So they know about the
status of individuals,

464
00:17:17,359 --> 00:17:19,119
like their age or their gender.

465
00:17:19,119 --> 00:17:21,369
They know about the
interactions of individuals.

466
00:17:21,369 --> 00:17:24,250
They recognize them very simply
like grooming and mothering.

467
00:17:24,250 --> 00:17:25,930
And then based on
these observations

468
00:17:25,930 --> 00:17:28,150
of the different individuals
in the social world,

469
00:17:28,150 --> 00:17:30,730
they built these
cognitive structures

470
00:17:30,730 --> 00:17:32,670
like friendship,
kinship, and hierarchy,

471
00:17:32,670 --> 00:17:35,900
have an interesting,
complicated structure to them.

472
00:17:35,900 --> 00:17:39,150
All of this is rooted in
the concept of the person.

473
00:17:39,150 --> 00:17:40,780
And this is very
important and as I'm

474
00:17:40,780 --> 00:17:41,810
going to be talking
about face recognition,

475
00:17:41,810 --> 00:17:43,939
I have to emphasize these
are two different things.

476
00:17:43,939 --> 00:17:45,730
You can recognize a
person from their face,

477
00:17:45,730 --> 00:17:47,146
but if you can
recognize a face it

478
00:17:47,146 --> 00:17:51,217
doesn't mean that you know who
it is, who is behind the face.

479
00:17:51,217 --> 00:17:53,550
So the person concept would
include something like this.

480
00:17:53,550 --> 00:17:55,124
It's a juvenile, female monkey.

481
00:17:55,124 --> 00:17:57,040
It's the daughter of x
and so on and so forth.

482
00:17:57,040 --> 00:17:58,250
So this is knowledge
that we have.

483
00:17:58,250 --> 00:18:00,100
It's actually been shown in
rhesus monkeys, the monkeys

484
00:18:00,100 --> 00:18:01,641
that we work with,
that they actually

485
00:18:01,641 --> 00:18:04,510
have this person knowledge.

486
00:18:04,510 --> 00:18:05,680
So why do we study faces?

487
00:18:05,680 --> 00:18:07,900
So for us faces really
are the ideal intersection

488
00:18:07,900 --> 00:18:10,990
between object
recognition, the study

489
00:18:10,990 --> 00:18:15,590
of which Jim DiCarlo talked
about, and social cognition.

490
00:18:15,590 --> 00:18:19,540
So as Jim was
alluding to yesterday,

491
00:18:19,540 --> 00:18:21,664
vision is really
important in primates.

492
00:18:21,664 --> 00:18:23,080
About a third of
the primate brain

493
00:18:23,080 --> 00:18:25,040
is thought to be involved
in visual function.

494
00:18:25,040 --> 00:18:25,750
So this is a lot.

495
00:18:25,750 --> 00:18:27,160
And this is
testament to the fact

496
00:18:27,160 --> 00:18:28,060
that there's a
lot of information

497
00:18:28,060 --> 00:18:30,160
to be gathered from the
outside visual world,

498
00:18:30,160 --> 00:18:32,800
but also that it's difficult
computationally to gather this.

499
00:18:32,800 --> 00:18:35,770
And so Jim was explaining some
of the computational challenges

500
00:18:35,770 --> 00:18:37,870
that object recognition
has to solve yesterday

501
00:18:37,870 --> 00:18:41,810
and then we'll come back to some
of his points a little later.

502
00:18:41,810 --> 00:18:43,150
So what is an object?

503
00:18:43,150 --> 00:18:45,700
So Jim was saying that it's
the basic unit of cognition.

504
00:18:45,700 --> 00:18:48,040
And just very quickly,
what it actually is it's

505
00:18:48,040 --> 00:18:49,534
more than a collection
of features.

506
00:18:49,534 --> 00:18:50,950
So the Gestalt
Rules of Perception

507
00:18:50,950 --> 00:18:51,960
actually emphasize this.

508
00:18:51,960 --> 00:18:53,918
So if you have proximity
of elements, you group

509
00:18:53,918 --> 00:18:54,520
them together.

510
00:18:54,520 --> 00:18:56,110
If these elements
share similarity,

511
00:18:56,110 --> 00:18:58,006
you group them together
to larger entities.

512
00:18:58,006 --> 00:18:59,380
If there's good
continuation, you

513
00:18:59,380 --> 00:19:01,669
group these local elements
together into lines.

514
00:19:01,669 --> 00:19:03,460
If that's common fate,
you group them again

515
00:19:03,460 --> 00:19:05,170
together to larger entities.

516
00:19:05,170 --> 00:19:07,277
And something similar
is true for faces.

517
00:19:07,277 --> 00:19:08,860
If you have different
face parts, then

518
00:19:08,860 --> 00:19:11,410
the wrong organization but
now you put them together

519
00:19:11,410 --> 00:19:13,510
correctly, you can
suddenly recognize a face.

520
00:19:13,510 --> 00:19:15,790
So there's a larger
scale organization to not

521
00:19:15,790 --> 00:19:18,140
that goes beyond just being
a collection of features.

522
00:19:18,140 --> 00:19:20,556
And maybe something similar
is going on for a higher order

523
00:19:20,556 --> 00:19:23,380
condition that I'm sure other
people are going to talk about.

524
00:19:23,380 --> 00:19:25,050
And it's in the
physical interactions,

525
00:19:25,050 --> 00:19:29,230
we infer causality from
just the sequence of events.

526
00:19:29,230 --> 00:19:32,030
Or social interactions,
like in Heider-Simmel movies

527
00:19:32,030 --> 00:19:34,319
where you are telling yourself
a complex social story

528
00:19:34,319 --> 00:19:35,860
unfold even when
there's just simple,

529
00:19:35,860 --> 00:19:37,780
geometric shapes moving around.

530
00:19:37,780 --> 00:19:41,030
So this creation of higher
order representations I think

531
00:19:41,030 --> 00:19:42,631
is essential for
object recognition.

532
00:19:42,631 --> 00:19:44,380
It's a constructive
process that the brain

533
00:19:44,380 --> 00:19:45,880
imposes on the
piece of information

534
00:19:45,880 --> 00:19:47,230
it gets from the eyes.

535
00:19:47,230 --> 00:19:49,180
It's not just a
collection of features.

536
00:19:49,180 --> 00:19:51,616
It's kind of the basis of
symbolic representations.

537
00:19:51,616 --> 00:19:52,990
It can create
meaning, especially

538
00:19:52,990 --> 00:19:55,490
if you think about the face,
if it's the face of someone you

539
00:19:55,490 --> 00:19:57,230
know that's very meaningful.

540
00:19:57,230 --> 00:19:59,350
And it makes
information actionable.

541
00:19:59,350 --> 00:20:01,400
And these are really,
I think, the import

542
00:20:01,400 --> 00:20:03,670
links between object recognition
and social cognition,

543
00:20:03,670 --> 00:20:07,449
and faces are smack
in the middle of this.

544
00:20:07,449 --> 00:20:08,740
So I already showed this movie.

545
00:20:08,740 --> 00:20:11,590
To use again I emphasize
the social communication

546
00:20:11,590 --> 00:20:12,779
that's taking place here.

547
00:20:12,779 --> 00:20:14,320
You can see the
facial displays here.

548
00:20:14,320 --> 00:20:16,860
So the older male who's
chasing the younger animal

549
00:20:16,860 --> 00:20:18,212
is making these facial displays.

550
00:20:18,212 --> 00:20:20,420
By the way, most of you will
never have seen a Tonkin

551
00:20:20,420 --> 00:20:22,586
macaque and still you cannot
understand what's going

552
00:20:22,586 --> 00:20:24,400
on there.

553
00:20:24,400 --> 00:20:26,170
This is something
very special, again

554
00:20:26,170 --> 00:20:28,870
that you don't have in all
animals, these facial displays.

555
00:20:28,870 --> 00:20:32,110
And Charles Darwin was actually
again one of the first people

556
00:20:32,110 --> 00:20:35,530
to notice it in 1872 is that
you use your face to express

557
00:20:35,530 --> 00:20:36,631
your emotional state.

558
00:20:36,631 --> 00:20:38,380
Otherwise your emotions
are private to you

559
00:20:38,380 --> 00:20:41,770
but you use body language
and then facial language

560
00:20:41,770 --> 00:20:43,155
to suppress your emotions.

561
00:20:43,155 --> 00:20:45,280
And oftentimes you do it
even if you don't want to.

562
00:20:45,280 --> 00:20:47,330
It just happens automatically.

563
00:20:47,330 --> 00:20:49,280
And that's not possible
in all animals.

564
00:20:49,280 --> 00:20:50,920
So if you are a fish
or a frog, there

565
00:20:50,920 --> 00:20:52,720
are lots of really
cool things you can do.

566
00:20:52,720 --> 00:20:54,970
You can sit on the front
porch and enjoy the day.

567
00:20:54,970 --> 00:20:57,760
So lots of things that you can
have in common with primates.

568
00:20:57,760 --> 00:21:00,220
But facial communication
really requires something

569
00:21:00,220 --> 00:21:02,940
more that's very
mammalian specific.

570
00:21:02,940 --> 00:21:04,900
In mammals you actually
have in the face

571
00:21:04,900 --> 00:21:07,390
musculature that it's not
attaching from bone to bone,

572
00:21:07,390 --> 00:21:09,310
but it's now
attaching to the skin.

573
00:21:09,310 --> 00:21:11,950
And so if you look at these two
rats here where their whiskers,

574
00:21:11,950 --> 00:21:14,200
I hope you can see it here,
in the end are labeled.

575
00:21:14,200 --> 00:21:16,830
You can see they're actively
exploring each other's faces

576
00:21:16,830 --> 00:21:18,510
in a somewhat sensory fashion.

577
00:21:18,510 --> 00:21:21,010
That's possible because they
can move their whiskers because

578
00:21:21,010 --> 00:21:22,264
of this musculature.

579
00:21:22,264 --> 00:21:23,680
And that's a
specialization that's

580
00:21:23,680 --> 00:21:25,155
becoming more and more
refined in primates.

581
00:21:25,155 --> 00:21:27,150
So in rhesus monkeys
and chimps and humans,

582
00:21:27,150 --> 00:21:29,124
we have 23 different
facial muscles.

583
00:21:29,124 --> 00:21:30,790
They're becoming more
and more flexible.

584
00:21:30,790 --> 00:21:33,123
I mentioned before that the
snout region is increasingly

585
00:21:33,123 --> 00:21:34,010
reduced in primates.

586
00:21:34,010 --> 00:21:35,050
So you have some
simpler primates

587
00:21:35,050 --> 00:21:36,820
where there's still
a strong snout, which

588
00:21:36,820 --> 00:21:38,887
is limiting the ability
of the face to move.

589
00:21:38,887 --> 00:21:41,470
But the more complex the primate
is getting, the more flexible

590
00:21:41,470 --> 00:21:43,303
these muscles become
and the more expressive

591
00:21:43,303 --> 00:21:44,080
the faces become.

592
00:21:44,080 --> 00:21:46,960
So the face now becomes richer
and richer with social signals

593
00:21:46,960 --> 00:21:48,160
that can read out.

594
00:21:48,160 --> 00:21:50,110
And in rhesus monkeys
which are shown here,

595
00:21:50,110 --> 00:21:51,860
you have a fixed set
of facial expressions

596
00:21:51,860 --> 00:21:54,715
that, again, for a system
that can analyze these,

597
00:21:54,715 --> 00:21:57,430
it's very important information
about the emotional state

598
00:21:57,430 --> 00:21:59,539
of another animal.

599
00:21:59,539 --> 00:22:01,330
Primates are also very
interested in faces.

600
00:22:01,330 --> 00:22:06,010
I'd very much like to show
this movie which is showing

601
00:22:06,010 --> 00:22:08,847
a three-day-old macaque monkey.

602
00:22:08,847 --> 00:22:10,930
And I'll tell you what the
point of this study is.

603
00:22:10,930 --> 00:22:12,920
So you can see that he's
attending very closely

604
00:22:12,920 --> 00:22:14,317
to the face of the experimenter.

605
00:22:14,317 --> 00:22:16,900
Of course, if there were bananas
you might think he would also

606
00:22:16,900 --> 00:22:19,150
be, this isn't proof that
there is this specialization

607
00:22:19,150 --> 00:22:21,250
for faces that Nancy
was alluding to before,

608
00:22:21,250 --> 00:22:22,990
but it's at least intuitive.

609
00:22:22,990 --> 00:22:25,180
The second thing why I
like to show this movie

610
00:22:25,180 --> 00:22:27,250
is exactly what
happened right now.

611
00:22:27,250 --> 00:22:28,810
You're all getting
really excited

612
00:22:28,810 --> 00:22:31,990
about this absolutely adorable,
little critter, right?

613
00:22:31,990 --> 00:22:33,980
And I've seen this movie
now hundreds of times

614
00:22:33,980 --> 00:22:35,021
and it's still like this.

615
00:22:35,021 --> 00:22:36,650
It's still very
emotionally charged.

616
00:22:36,650 --> 00:22:38,840
So here's the third reason.

617
00:22:38,840 --> 00:22:40,889
The experiment is based
on facial movements here.

618
00:22:40,889 --> 00:22:42,930
You can see it's getting
really excited about it,

619
00:22:42,930 --> 00:22:44,020
it's getting very active.

620
00:22:44,020 --> 00:22:45,978
And now he's reproducing
these facial movements

621
00:22:45,978 --> 00:22:47,189
as best as he can.

622
00:22:47,189 --> 00:22:48,730
There's a specific
facial interaction

623
00:22:48,730 --> 00:22:50,869
that's happening
in human babies,

624
00:22:50,869 --> 00:22:51,910
I think for three months.

625
00:22:51,910 --> 00:22:54,119
It's happening in these
rhesus monkeys for two weeks.

626
00:22:54,119 --> 00:22:56,326
And you can see that there
is an intricate connection

627
00:22:56,326 --> 00:22:58,750
between what they're perceiving
and what they're acting on

628
00:22:58,750 --> 00:23:00,190
in an automatic fashion.

629
00:23:00,190 --> 00:23:03,070
But this emotional part I
think is really important.

630
00:23:03,070 --> 00:23:04,630
It's just at a
certain point that you

631
00:23:04,630 --> 00:23:06,220
can't control these things.

632
00:23:06,220 --> 00:23:09,010
Faces really get very deep into
your emotional and social brain

633
00:23:09,010 --> 00:23:09,750
automatically.

634
00:23:09,750 --> 00:23:11,680
And so one of the lines
of research in my lab

635
00:23:11,680 --> 00:23:12,940
is to try to figure
out the circuits

636
00:23:12,940 --> 00:23:14,740
that make that possible
and then to use

637
00:23:14,740 --> 00:23:17,280
this to get an inroad into
the social brain of function

638
00:23:17,280 --> 00:23:19,840
beyond face perception.

639
00:23:19,840 --> 00:23:22,210
So amongst the signals
the faces are sending,

640
00:23:22,210 --> 00:23:24,280
you recognize Charles
Darwin, so it's identity,

641
00:23:24,280 --> 00:23:26,300
they're getting their
social communication,

642
00:23:26,300 --> 00:23:28,970
there's emotional responses,
and there's also face following.

643
00:23:28,970 --> 00:23:31,110
So the direction that
the eyes are looking to,

644
00:23:31,110 --> 00:23:32,600
we are following automatically.

645
00:23:32,600 --> 00:23:34,808
We can control this later,
but this initial automatic

646
00:23:34,808 --> 00:23:35,390
response.

647
00:23:35,390 --> 00:23:36,765
Here's one very
nice illustration

648
00:23:36,765 --> 00:23:38,200
from the British TV show.

649
00:23:38,200 --> 00:23:40,240
You have people
wearing these glasses,

650
00:23:40,240 --> 00:23:43,870
actually it's a large background
to people wearing these glasses

651
00:23:43,870 --> 00:23:46,500
where their eyes are drawn
on these glasses that

652
00:23:46,500 --> 00:23:48,100
are going to one direction.

653
00:23:48,100 --> 00:23:50,417
And so you know that
these are not real eyes,

654
00:23:50,417 --> 00:23:52,000
but what's happening
is your attention

655
00:23:52,000 --> 00:23:54,082
is drawn constantly to
this upper right region.

656
00:23:54,082 --> 00:23:55,540
And it's getting
annoying over time

657
00:23:55,540 --> 00:23:57,130
because you know that
there's nothing there.

658
00:23:57,130 --> 00:23:58,830
You know that they're not
really paying attention there.

659
00:23:58,830 --> 00:24:00,490
But automatically your
attention is drawn there

660
00:24:00,490 --> 00:24:01,930
and then you're going back
again and your attention

661
00:24:01,930 --> 00:24:03,230
is going out there again.

662
00:24:03,230 --> 00:24:04,180
And so this is
another thing that

663
00:24:04,180 --> 00:24:05,596
comes from the
face that gets deep

664
00:24:05,596 --> 00:24:08,770
into your attentional
control system.

665
00:24:08,770 --> 00:24:11,440
So social perception can
start with faces, but faces

666
00:24:11,440 --> 00:24:15,790
are the most important visual
sources of information.

667
00:24:15,790 --> 00:24:17,580
We get gender and age,
of course identity,

668
00:24:17,580 --> 00:24:19,300
and things like
perceived trustworthiness

669
00:24:19,300 --> 00:24:21,919
or attractiveness from just a
very brief look at the face.

670
00:24:21,919 --> 00:24:23,710
And then there are
these dynamical signals,

671
00:24:23,710 --> 00:24:25,460
like mood and overt
direction of attention

672
00:24:25,460 --> 00:24:27,277
that we also get from the face.

673
00:24:27,277 --> 00:24:28,360
So how does this all work?

674
00:24:28,360 --> 00:24:31,720
So Jim was already explaining
some of the challenges

675
00:24:31,720 --> 00:24:33,065
of object recognition to you.

676
00:24:33,065 --> 00:24:34,690
And so here are some
of the challenges.

677
00:24:34,690 --> 00:24:36,940
So first of all, the social
scene like this one here,

678
00:24:36,940 --> 00:24:40,150
lighting conditions can
sometimes be non-optimal.

679
00:24:40,150 --> 00:24:43,570
And so the first thing for you
to analyze the facial signals

680
00:24:43,570 --> 00:24:46,390
which are in this scene, is to
localize where the faces are.

681
00:24:46,390 --> 00:24:47,620
And I'm going to tell you
a little bit about what

682
00:24:47,620 --> 00:24:49,685
we understand about
the mechanisms of that.

683
00:24:49,685 --> 00:24:51,310
Then once you know
where the faces are,

684
00:24:51,310 --> 00:24:52,685
you want to analyze
them further,

685
00:24:52,685 --> 00:24:54,880
you want to know who
these individuals are.

686
00:24:54,880 --> 00:24:57,280
And I just realized that the
images that I had from this,

687
00:24:57,280 --> 00:24:59,410
which are of course also
taken from The Godfather,

688
00:24:59,410 --> 00:25:01,645
might not be the best.

689
00:25:01,645 --> 00:25:03,520
Where is the other
picture of this individual

690
00:25:03,520 --> 00:25:05,552
here in this display
of these five faces?

691
00:25:08,390 --> 00:25:10,490
Upper right.

692
00:25:10,490 --> 00:25:12,050
And then there's
another individual,

693
00:25:12,050 --> 00:25:15,280
there's Don Corleone and
there's another person down here

694
00:25:15,280 --> 00:25:16,530
with two different directions.

695
00:25:16,530 --> 00:25:18,655
And then if the lights were
down a little bit more,

696
00:25:18,655 --> 00:25:19,790
you could see this better.

697
00:25:19,790 --> 00:25:21,950
The cool thing is that we have
a way of relating these two

698
00:25:21,950 --> 00:25:23,616
pictures to each other
knowing that they

699
00:25:23,616 --> 00:25:26,232
are from the same person, even
though physically on a pixel

700
00:25:26,232 --> 00:25:28,190
by pixel basis these two
actually are much more

701
00:25:28,190 --> 00:25:29,542
similar to each other.

702
00:25:29,542 --> 00:25:32,000
And so we'd like to figure out
how the brain is doing that,

703
00:25:32,000 --> 00:25:35,030
achieving object recognition,
in this case face recognition,

704
00:25:35,030 --> 00:25:37,760
in a manner that's invariant
to transformations that are not

705
00:25:37,760 --> 00:25:40,070
intrinsic to the object.

706
00:25:40,070 --> 00:25:42,320
This is just a reminder that
face recognition actually

707
00:25:42,320 --> 00:25:46,400
is very difficult. So this
is of course just made up

708
00:25:46,400 --> 00:25:49,550
from Curb Your Enthusiasm,
but there's a condition

709
00:25:49,550 --> 00:25:52,190
that many of you will have
heard about, prosopagnosia.

710
00:25:52,190 --> 00:25:54,380
And to a prosopagnotic
person who is face blind,

711
00:25:54,380 --> 00:25:56,370
the social world
might look like this.

712
00:25:56,370 --> 00:25:58,340
So a prosopagnotic has
great difficulty telling

713
00:25:58,340 --> 00:25:59,506
one individual from another.

714
00:25:59,506 --> 00:26:01,780
This is at least the
most typical condition.

715
00:26:01,780 --> 00:26:03,738
And you can imagine that
your social life would

716
00:26:03,738 --> 00:26:05,870
be really difficult
and your enthusiasm

717
00:26:05,870 --> 00:26:07,494
about socially
interacting would really

718
00:26:07,494 --> 00:26:09,740
be curbed if all the
individuals looked at this

719
00:26:09,740 --> 00:26:12,020
and looked all the same.

720
00:26:12,020 --> 00:26:16,550
So there must be something
about the new mechanisms that's

721
00:26:16,550 --> 00:26:17,420
very precise.

722
00:26:17,420 --> 00:26:19,378
So what's the neural
basis of face recognition?

723
00:26:19,378 --> 00:26:21,470
So the story really
starts with Charles Gross

724
00:26:21,470 --> 00:26:23,660
many years back in the
late '60s, early '70s.

725
00:26:23,660 --> 00:26:25,970
He was recording from the
inferotemporal cortex.

726
00:26:25,970 --> 00:26:28,310
He was showing pictures
of monkey faces,

727
00:26:28,310 --> 00:26:30,100
other social stimuli
like the monkey hand,

728
00:26:30,100 --> 00:26:31,670
he would scramble
the face, and then

729
00:26:31,670 --> 00:26:32,961
look at the responses of cells.

730
00:26:32,961 --> 00:26:35,370
And he was the first to find
a face selective neuron.

731
00:26:35,370 --> 00:26:36,110
Here's one.

732
00:26:36,110 --> 00:26:38,193
So these vertical lines
have the action potentials

733
00:26:38,193 --> 00:26:39,014
the cell is firing.

734
00:26:39,014 --> 00:26:40,930
This is the period of
time his face was shown.

735
00:26:40,930 --> 00:26:43,160
And this is the period of time
the control object, the hand,

736
00:26:43,160 --> 00:26:43,910
was shown.

737
00:26:43,910 --> 00:26:46,010
And you can see the cells
responding selectively

738
00:26:46,010 --> 00:26:47,960
to the face and not to the head.

739
00:26:47,960 --> 00:26:49,837
So this was a very nice finding.

740
00:26:49,837 --> 00:26:51,920
It actually took him some
time to convince himself

741
00:26:51,920 --> 00:26:53,420
that he could publish
it because he thought

742
00:26:53,420 --> 00:26:54,586
people would not believe it.

743
00:26:54,586 --> 00:26:57,590
It's recording a
anaesthetised animal.

744
00:26:57,590 --> 00:27:00,686
But luckily, he did publish
it many, many years later.

745
00:27:00,686 --> 00:27:02,060
And so this is
the first evidence

746
00:27:02,060 --> 00:27:04,311
that there is a specialization
in the brain for faces.

747
00:27:04,311 --> 00:27:06,310
He found many other cells
that like other things

748
00:27:06,310 --> 00:27:07,980
and then faces
and I think people

749
00:27:07,980 --> 00:27:09,920
thought that they
were intermingled

750
00:27:09,920 --> 00:27:10,780
with other objects.

751
00:27:10,780 --> 00:27:11,654
So this was the view.

752
00:27:11,654 --> 00:27:13,654
That this is the side
view of the macaque brain.

753
00:27:13,654 --> 00:27:16,112
This is the superior temporal
circuits, the one big circuit

754
00:27:16,112 --> 00:27:17,169
in the monkey brain.

755
00:27:17,169 --> 00:27:18,710
And all these symbols
here indicating

756
00:27:18,710 --> 00:27:20,750
positions where people found
face selective neurons.

757
00:27:20,750 --> 00:27:22,730
The thought was they
intermingled with object

758
00:27:22,730 --> 00:27:23,870
recognition hardware.

759
00:27:23,870 --> 00:27:25,286
It's basically the
view that there

760
00:27:25,286 --> 00:27:27,162
is a big IT cortex where
everything in object

761
00:27:27,162 --> 00:27:28,120
recognition can happen.

762
00:27:28,120 --> 00:27:29,090
And yes, of course
you would have

763
00:27:29,090 --> 00:27:30,230
some cells that
are face selective,

764
00:27:30,230 --> 00:27:32,521
you would have other cells
that are non-face selective.

765
00:27:32,521 --> 00:27:35,130
And the mixture and the
complex pattern of activity

766
00:27:35,130 --> 00:27:38,050
really is what gives you
the identity of the object.

767
00:27:38,050 --> 00:27:42,460
Then Nancy used fMRI to
discover face selective areas.

768
00:27:42,460 --> 00:27:44,210
So first, these are
views from face areas.

769
00:27:44,210 --> 00:27:47,060
We now know multiple face areas
that she was talking about

770
00:27:47,060 --> 00:27:47,920
before.

771
00:27:47,920 --> 00:27:50,366
So here are different
slices to this.

772
00:27:50,366 --> 00:27:51,740
And the thought
from these images

773
00:27:51,740 --> 00:27:54,680
really was that, no, that
maybe within this large expanse

774
00:27:54,680 --> 00:27:56,690
of object recognition
hardware there

775
00:27:56,690 --> 00:27:59,870
might be very specialized
regions that are really there

776
00:27:59,870 --> 00:28:02,330
selectively to process faces.

777
00:28:02,330 --> 00:28:04,970
And so you give the FFA, and so
the question you would ask is,

778
00:28:04,970 --> 00:28:07,939
is this really a region that's
devoted to face processing

779
00:28:07,939 --> 00:28:08,980
and face processing only?

780
00:28:11,760 --> 00:28:13,560
Are these regions
really face processing?

781
00:28:13,560 --> 00:28:15,981
Modules devoted
to face processing

782
00:28:15,981 --> 00:28:17,480
or is it just the
tip of the iceberg

783
00:28:17,480 --> 00:28:18,740
based on your
statistical analysis

784
00:28:18,740 --> 00:28:20,630
that this region just looks a
little bit more face selective

785
00:28:20,630 --> 00:28:22,250
than the neighboring regions?

786
00:28:22,250 --> 00:28:24,710
And second, do monkeys also
have these localized face

787
00:28:24,710 --> 00:28:25,460
areas like humans?

788
00:28:25,460 --> 00:28:26,876
And you've got the
answer already.

789
00:28:26,876 --> 00:28:28,075
Yes, they do.

790
00:28:28,075 --> 00:28:29,450
And then what is
the distribution

791
00:28:29,450 --> 00:28:32,660
of cells within these
regions versus outside?

792
00:28:32,660 --> 00:28:35,152
So this is really the
research Doris Tsao

793
00:28:35,152 --> 00:28:37,680
and I engaged on many years ago.

794
00:28:37,680 --> 00:28:39,680
We used fMRI on macaque
monkeys, same technology

795
00:28:39,680 --> 00:28:42,050
as in humans, slightly
different coils.

796
00:28:42,050 --> 00:28:43,550
And this is the
picture that we got.

797
00:28:43,550 --> 00:28:45,754
Very consistently across
different animals.

798
00:28:45,754 --> 00:28:47,420
Here in the temporal
lobe, you have six,

799
00:28:47,420 --> 00:28:50,500
face selective
regions that you find

800
00:28:50,500 --> 00:28:52,550
that anatomically
specific regions there's

801
00:28:52,550 --> 00:28:54,830
some variation from one
individual to the other.

802
00:28:54,830 --> 00:28:56,610
But with the exception of
the most posterior area,

803
00:28:56,610 --> 00:28:57,984
you actually find
all these areas

804
00:28:57,984 --> 00:29:00,015
in all individuals
on both hemispheres.

805
00:29:00,015 --> 00:29:02,390
There are also three areas in
the prefrontal cortex which

806
00:29:02,390 --> 00:29:03,742
are a little harder to find.

807
00:29:03,742 --> 00:29:05,200
But the one in
orbitofrontal cortex

808
00:29:05,200 --> 00:29:07,040
is actually is as
reproducible as the one

809
00:29:07,040 --> 00:29:08,370
in the temporal lobe.

810
00:29:08,370 --> 00:29:09,624
So, yes there are.

811
00:29:09,624 --> 00:29:11,540
Monkeys have localized
face areas like humans.

812
00:29:11,540 --> 00:29:13,250
And as Nancy was
alluding to, we actually

813
00:29:13,250 --> 00:29:14,420
have quite a bit
of evidence by now

814
00:29:14,420 --> 00:29:16,250
that these systems
might be homologous.

815
00:29:16,250 --> 00:29:18,722
Very, very difficult to prove
that they are homologous.

816
00:29:18,722 --> 00:29:20,180
But all the evidence
we have so far

817
00:29:20,180 --> 00:29:22,910
is really pointing
in this direction.

818
00:29:22,910 --> 00:29:24,710
So how selective are
these face patches?

819
00:29:24,710 --> 00:29:27,470
And what Doris and I did was
to lower recording electrodes

820
00:29:27,470 --> 00:29:30,350
into these face areas
and record from cells

821
00:29:30,350 --> 00:29:32,494
inside this fMRI
identified areas.

822
00:29:32,494 --> 00:29:34,160
And I'm going to show
you a movie of one

823
00:29:34,160 --> 00:29:37,137
of the first cells we recorded
from one of these regions.

824
00:29:37,137 --> 00:29:39,470
So it's actually a video we
took from a control monitor.

825
00:29:39,470 --> 00:29:41,510
So it shows the same
thing the monkey shows.

826
00:29:41,510 --> 00:29:44,400
The quality is not great because
it's an actual video camera

827
00:29:44,400 --> 00:29:45,692
we took to take this image.

828
00:29:45,692 --> 00:29:47,150
In addition to what
the animal saw,

829
00:29:47,150 --> 00:29:48,608
you will also see
this black square

830
00:29:48,608 --> 00:29:51,110
which is indicating where the
animal looked on the screen

831
00:29:51,110 --> 00:29:53,219
but the animal did not see this.

832
00:29:53,219 --> 00:29:55,760
And you're going to hear clicks
if everything works fine when

833
00:29:55,760 --> 00:29:58,334
the actual potential is fired.

834
00:29:58,334 --> 00:29:59,750
Anyway, here's the
quantification.

835
00:29:59,750 --> 00:30:04,500
So with 96 different stimuli in
this image set, 16 faces and 18

836
00:30:04,500 --> 00:30:07,010
non-face stimuli, this
is the average response

837
00:30:07,010 --> 00:30:09,282
which is normalized
between minus 1 and 1.

838
00:30:09,282 --> 00:30:10,990
And you can see that
the biggest response

839
00:30:10,990 --> 00:30:14,277
of this particular cell actually
of course to the 16 faces

840
00:30:14,277 --> 00:30:15,860
and not to any of
the control objects.

841
00:30:15,860 --> 00:30:18,200
You can see though that
there are some stimuli here

842
00:30:18,200 --> 00:30:19,700
in the gadget
category, for example,

843
00:30:19,700 --> 00:30:21,991
that are eliciting responses
that are quite respectable

844
00:30:21,991 --> 00:30:23,090
relative to the faces.

845
00:30:23,090 --> 00:30:25,840
But really the biggest responses
recorded were to the faces.

846
00:30:25,840 --> 00:30:28,270
So then we color coded
this so you have a response

847
00:30:28,270 --> 00:30:31,060
vector of the sale, where red
is now symbolizing response

848
00:30:31,060 --> 00:30:33,070
enhancement and blue
is symbolizing response

849
00:30:33,070 --> 00:30:34,427
suppression below baseline.

850
00:30:34,427 --> 00:30:36,010
And the advantage
of using this format

851
00:30:36,010 --> 00:30:37,810
is you can now stick
all the responses you

852
00:30:37,810 --> 00:30:39,226
get from all the
cells that you're

853
00:30:39,226 --> 00:30:42,040
recording day after day after
day from this one face area.

854
00:30:42,040 --> 00:30:44,240
And you get a population
response matrix.

855
00:30:44,240 --> 00:30:46,700
And the way that this
works is cell numbers

856
00:30:46,700 --> 00:30:49,300
are organized from top to
bottom, page number from left

857
00:30:49,300 --> 00:30:50,170
to right.

858
00:30:50,170 --> 00:30:52,900
And you can see very quickly
that most of the cells

859
00:30:52,900 --> 00:30:55,930
here are either selectively
enhanced or selectively

860
00:30:55,930 --> 00:30:57,052
suppressed by faces.

861
00:30:57,052 --> 00:30:58,510
There's a small
group here between,

862
00:30:58,510 --> 00:31:00,520
something like 10% of
the cells, where it's not

863
00:31:00,520 --> 00:31:01,780
so clear what they are doing.

864
00:31:01,780 --> 00:31:03,747
But if you do the
population average,

865
00:31:03,747 --> 00:31:05,830
you can see much bigger
responses to all the faces

866
00:31:05,830 --> 00:31:07,870
rather than on face objects.

867
00:31:07,870 --> 00:31:10,570
If you look more closely, what
these pictures are eliciting

868
00:31:10,570 --> 00:31:12,740
in these intermediate
responses, these

869
00:31:12,740 --> 00:31:14,680
are like clock
faces, apples, pears,

870
00:31:14,680 --> 00:31:17,050
there are things that have
physical properties in common

871
00:31:17,050 --> 00:31:17,794
with faces.

872
00:31:17,794 --> 00:31:19,210
So you can kind
of fool the system

873
00:31:19,210 --> 00:31:20,950
to give a partial response.

874
00:31:20,950 --> 00:31:23,200
And this is one clue to what
this area might be doing.

875
00:31:23,200 --> 00:31:24,699
It should be doing
a visual analysis

876
00:31:24,699 --> 00:31:26,590
of the incoming stimuli
to try and figure out

877
00:31:26,590 --> 00:31:29,810
if these are faces are not.

878
00:31:29,810 --> 00:31:32,071
So these are cells in
the middle face patches.

879
00:31:32,071 --> 00:31:33,820
I was actually going
over this pretty fast

880
00:31:33,820 --> 00:31:36,200
but I'm going to use
this later quite a bit.

881
00:31:36,200 --> 00:31:37,450
And so let's wind back.

882
00:31:37,450 --> 00:31:41,710
We have one posterior area
here, to middle face areas,

883
00:31:41,710 --> 00:31:44,290
see the middle face patches,
and then three anterior ones.

884
00:31:44,290 --> 00:31:46,450
I'm mostly going to talk
about this one here, AL,

885
00:31:46,450 --> 00:31:48,210
and this one here, M, in
addition to the middle face

886
00:31:48,210 --> 00:31:48,710
patches.

887
00:31:54,160 --> 00:31:57,740
So we think that actually
this is another automatic face

888
00:31:57,740 --> 00:31:59,170
recognition feat.

889
00:31:59,170 --> 00:32:01,880
We can't stop feeling
sorry for these peppers.

890
00:32:01,880 --> 00:32:03,830
They've been just cut in half.

891
00:32:03,830 --> 00:32:05,340
And so they seem
to be screaming,

892
00:32:05,340 --> 00:32:08,030
and then you know
they are OK but still

893
00:32:08,030 --> 00:32:11,420
you feel like something
really bad just happened.

894
00:32:11,420 --> 00:32:15,410
And so we can't stop having
these inferences about peppers

895
00:32:15,410 --> 00:32:17,150
where they look like faces.

896
00:32:17,150 --> 00:32:19,780
And one reason could be that we
have this specialized circuitry

897
00:32:19,780 --> 00:32:21,740
that's just getting active
with right features,

898
00:32:21,740 --> 00:32:24,860
even if you know
these are not faces.

899
00:32:24,860 --> 00:32:28,202
OK so when the faces were
discovered by Charles Gross,

900
00:32:28,202 --> 00:32:29,910
this really fell on
very fertile grounds.

901
00:32:29,910 --> 00:32:32,156
And I should just discuss
some of the implications.

902
00:32:32,156 --> 00:32:33,530
So David Hubel
and Torsten Weisel

903
00:32:33,530 --> 00:32:36,980
just discovered a few years
before orientation selectivity.

904
00:32:36,980 --> 00:32:39,140
So it was a big jump
from early processing

905
00:32:39,140 --> 00:32:41,970
where I could see how
selectivity of cells

906
00:32:41,970 --> 00:32:43,550
was getting more complex.

907
00:32:43,550 --> 00:32:45,782
More concentric representations
to elongated ones,

908
00:32:45,782 --> 00:32:47,240
from simple cells
to complex cells,

909
00:32:47,240 --> 00:32:50,510
but complex cells are as
selective as simple cells

910
00:32:50,510 --> 00:32:52,220
but don't really have
a special location.

911
00:32:52,220 --> 00:32:54,620
All the way up to the opposite
end of the visual system

912
00:32:54,620 --> 00:32:57,050
and now you find a
face selective neuron.

913
00:32:57,050 --> 00:32:59,940
Jerome Lettvin just had
coined the term grandmother

914
00:32:59,940 --> 00:33:02,284
neuron which some of you
brought up yesterday.

915
00:33:02,284 --> 00:33:04,700
The idea is that there should
be one neuron in your brain,

916
00:33:04,700 --> 00:33:07,040
or this is the hypothetical
situation you came up with,

917
00:33:07,040 --> 00:33:09,290
one neuron in your the brain
that's firing if and only

918
00:33:09,290 --> 00:33:11,210
if you see your grandmother,
no matter what she's wearing,

919
00:33:11,210 --> 00:33:12,650
which direction you see from.

920
00:33:12,650 --> 00:33:14,630
That's the neural
correlate of you

921
00:33:14,630 --> 00:33:17,000
perceiving your grandmother
is the activity of this one

922
00:33:17,000 --> 00:33:17,750
neuron.

923
00:33:17,750 --> 00:33:19,960
And there were other
concepts like Jerzy Konorski

924
00:33:19,960 --> 00:33:21,852
gnostic unit that
made the same point.

925
00:33:21,852 --> 00:33:23,810
Then Horace Barlow came
up with this idea maybe

926
00:33:23,810 --> 00:33:26,360
it's not one cell,
but multiple cells.

927
00:33:26,360 --> 00:33:27,920
But gave us a sparse
representation

928
00:33:27,920 --> 00:33:29,990
of pontifical
cells, a few of them

929
00:33:29,990 --> 00:33:31,820
at the top of a
processing hierarchy.

930
00:33:31,820 --> 00:33:33,612
And that's actually
how we recognize faces.

931
00:33:33,612 --> 00:33:36,111
And then of course there's the
opposite view of Donald Hebb.

932
00:33:36,111 --> 00:33:37,920
He talked about cell
assembly since there's

933
00:33:37,920 --> 00:33:40,760
no-- things are there like
large assembles] of cells.

934
00:33:40,760 --> 00:33:42,860
Or Karl Lashley who
talked about mass action

935
00:33:42,860 --> 00:33:44,360
and actually were
completely against

936
00:33:44,360 --> 00:33:45,944
functional specialization.

937
00:33:45,944 --> 00:33:47,360
If you look at a
plot like this, I

938
00:33:47,360 --> 00:33:49,235
think one of the things
you want to emphasize

939
00:33:49,235 --> 00:33:50,990
is that these cells
really don't fall

940
00:33:50,990 --> 00:33:52,300
into any of these categories.

941
00:33:52,300 --> 00:33:54,508
You can have cells that are
very, very face selective

942
00:33:54,508 --> 00:33:56,270
but they don't have
to be very sparse.

943
00:33:56,270 --> 00:33:58,160
They will appear sparse if
you poke them over and over

944
00:33:58,160 --> 00:33:59,820
and over with non-face
stimuli because they're not

945
00:33:59,820 --> 00:34:01,096
going to respond to those.

946
00:34:01,096 --> 00:34:02,720
But within the domain
of faces, they're

947
00:34:02,720 --> 00:34:05,814
going to respond to
pretty much all faces.

948
00:34:05,814 --> 00:34:07,980
There are differences between
these different cells.

949
00:34:07,980 --> 00:34:09,800
I'm going to come
back to that as well.

950
00:34:09,800 --> 00:34:11,550
But it's one example
where we can actually

951
00:34:11,550 --> 00:34:14,020
ask them these deep questions
about what is the neural code

952
00:34:14,020 --> 00:34:15,860
and quantitative
matter by focusing

953
00:34:15,860 --> 00:34:18,987
on the right stimulus and the
right place to look at it.

954
00:34:18,987 --> 00:34:21,320
So we have some evidence that
monkeys, like humans, have

955
00:34:21,320 --> 00:34:23,510
face regions, and
the monkey face

956
00:34:23,510 --> 00:34:26,810
patches appear to be dedicated
domain specific modules.

957
00:34:26,810 --> 00:34:28,370
The practical
implications of this

958
00:34:28,370 --> 00:34:30,650
is that now we have
unprecedented access

959
00:34:30,650 --> 00:34:32,570
to function homogeneous
populations of cells

960
00:34:32,570 --> 00:34:34,760
coding for one high
level object category.

961
00:34:34,760 --> 00:34:36,780
And we know this category,
we can make stimuli.

962
00:34:36,780 --> 00:34:38,690
And we can modify
the stimuli sometimes

963
00:34:38,690 --> 00:34:39,698
in parametric fashions.

964
00:34:39,698 --> 00:34:41,239
And so we can have
very deep insights

965
00:34:41,239 --> 00:34:43,572
into how these cells actually
are processing interfaces,

966
00:34:43,572 --> 00:34:46,100
how they are restricting
properties from these faces.

967
00:34:46,100 --> 00:34:49,130
And we can do causal
tests and actually show

968
00:34:49,130 --> 00:34:52,147
whether these cells are involved
in face recognition behavior.

969
00:34:52,147 --> 00:34:54,230
And we're just going to
go over this very quickly.

970
00:34:54,230 --> 00:34:56,477
This is work of
Srivatsun Sadagopan,

971
00:34:56,477 --> 00:34:58,310
he actually gave me
this picture of himself.

972
00:34:58,310 --> 00:35:01,280
This combines the front
view and a profile view.

973
00:35:01,280 --> 00:35:02,322
The logic is very simple.

974
00:35:02,322 --> 00:35:04,780
So we wanted to inactivate one
particular region in a male.

975
00:35:04,780 --> 00:35:06,920
I'm going to tell you
in a second why a male.

976
00:35:06,920 --> 00:35:08,780
While the monkey would
be engaged in a task

977
00:35:08,780 --> 00:35:11,071
like this where it is to find
a face in a visual scene.

978
00:35:11,071 --> 00:35:12,634
The visual scene
that we constructed

979
00:35:12,634 --> 00:35:13,550
looks a bit like this.

980
00:35:13,550 --> 00:35:15,216
It's displayed on a
touch screen monitor

981
00:35:15,216 --> 00:35:16,760
so the animal's
free to move around.

982
00:35:16,760 --> 00:35:18,259
It has to find the
face in the scene

983
00:35:18,259 --> 00:35:20,270
and the scene is
composed of a pink noise

984
00:35:20,270 --> 00:35:22,926
background embedded in which
there are 24 different objects.

985
00:35:22,926 --> 00:35:24,800
And the target object,
in this case the face,

986
00:35:24,800 --> 00:35:26,480
is going to be
varying in visibility

987
00:35:26,480 --> 00:35:28,460
across 10 different levels.

988
00:35:28,460 --> 00:35:31,657
We would have other tasks where
the monkey body was included,

989
00:35:31,657 --> 00:35:33,240
which you will not
be able to see here

990
00:35:33,240 --> 00:35:35,360
but there is a monkey
body here or the monkey

991
00:35:35,360 --> 00:35:37,580
was looking for a shoe.

992
00:35:37,580 --> 00:35:39,590
Then we would infuse
muscimol which

993
00:35:39,590 --> 00:35:42,407
is a pharmacological agent
that is inactivating cells

994
00:35:42,407 --> 00:35:44,240
along with a contrast
agent gadolinium which

995
00:35:44,240 --> 00:35:45,472
you can measure an MRI.

996
00:35:45,472 --> 00:35:47,680
And then this yellow region
here is the face area now

997
00:35:47,680 --> 00:35:49,763
and this white region is
the actual injection site

998
00:35:49,763 --> 00:35:50,780
that we used.

999
00:35:50,780 --> 00:35:53,460
And so this gives it a way for
every experiment to control,

1000
00:35:53,460 --> 00:35:55,580
are we inside the face
area or are we outside.

1001
00:35:55,580 --> 00:35:58,570
And we can use the outside
injections as controls.

1002
00:35:58,570 --> 00:36:00,220
What's been found is shown here.

1003
00:36:00,220 --> 00:36:01,900
So we have a psychometric curve.

1004
00:36:01,900 --> 00:36:04,150
So in normal behavior you're
getting better and better

1005
00:36:04,150 --> 00:36:09,770
at finding the face in the scene
as you increase its visibility.

1006
00:36:09,770 --> 00:36:12,500
If you inactivate, you get
a reduced face detection

1007
00:36:12,500 --> 00:36:14,180
behavior.

1008
00:36:14,180 --> 00:36:18,290
I should emphasize that we only
inactivating one face area out

1009
00:36:18,290 --> 00:36:19,610
of 12 in the temporal lobe.

1010
00:36:19,610 --> 00:36:21,351
We are only activating
one hemisphere.

1011
00:36:21,351 --> 00:36:22,850
And what Jim DiCarlo
was emphasizing

1012
00:36:22,850 --> 00:36:25,590
yesterday this retinopathy
at this level of processing.

1013
00:36:25,590 --> 00:36:27,710
So the animal can actually
use a scanning strategy

1014
00:36:27,710 --> 00:36:31,280
to go one direction and
overcome this deficit.

1015
00:36:31,280 --> 00:36:34,132
And so likely this effect
would be much stronger

1016
00:36:34,132 --> 00:36:35,840
if we had inactivated
on both hemispheres

1017
00:36:35,840 --> 00:36:38,260
or would have controlled
precisely for the eye movement.

1018
00:36:38,260 --> 00:36:40,250
But we are here for
natural behavior.

1019
00:36:40,250 --> 00:36:41,790
And the controls
of bodies and shoes

1020
00:36:41,790 --> 00:36:43,040
are, there is no effect there.

1021
00:36:43,040 --> 00:36:45,510
We put lots of controls and
we went to the next state

1022
00:36:45,510 --> 00:36:46,570
of the behavior.

1023
00:36:46,570 --> 00:36:48,480
We did the injections
outside as I mentioned,

1024
00:36:48,480 --> 00:36:51,200
it's very specific for
inactivation inside the face

1025
00:36:51,200 --> 00:36:54,970
area that the most basic of
face recognition abilities, face

1026
00:36:54,970 --> 00:36:57,250
detection is impaired.

1027
00:36:57,250 --> 00:36:59,422
And so there's a way
you might visualize it.

1028
00:36:59,422 --> 00:37:00,880
So one way to
explain this behavior

1029
00:37:00,880 --> 00:37:03,070
would be that the visibility
of a face like this

1030
00:37:03,070 --> 00:37:04,552
would actually,
with an activation,

1031
00:37:04,552 --> 00:37:06,010
look something like
this where it's

1032
00:37:06,010 --> 00:37:09,867
going to be harder to detect.

1033
00:37:09,867 --> 00:37:11,700
The second way we can
take advantage of this

1034
00:37:11,700 --> 00:37:13,230
is that we have now access
to individual cells.

1035
00:37:13,230 --> 00:37:14,813
We actually ask more
precise questions

1036
00:37:14,813 --> 00:37:16,650
about how they're
processing faces.

1037
00:37:16,650 --> 00:37:17,880
And actually there's
an activation study

1038
00:37:17,880 --> 00:37:20,421
that was motivated by earlier
work we had done on selectivity

1039
00:37:20,421 --> 00:37:22,380
of these cells for
features that should

1040
00:37:22,380 --> 00:37:24,780
be relevant for face detection.

1041
00:37:24,780 --> 00:37:26,280
This is what Shay
Ohayon did when he

1042
00:37:26,280 --> 00:37:27,690
was a grad student with Doris.

1043
00:37:27,690 --> 00:37:30,420
He's actually now a
post-doc with Jim.

1044
00:37:30,420 --> 00:37:33,870
And again, the question is
how can you detect faces even

1045
00:37:33,870 --> 00:37:35,980
when the lighting conditions
are very difficult.

1046
00:37:35,980 --> 00:37:37,896
There's beautiful work
from Pavan Sinha that's

1047
00:37:37,896 --> 00:37:39,719
emphasizing that
coarse, contrast

1048
00:37:39,719 --> 00:37:41,760
relationships in the face
of very good heuristics

1049
00:37:41,760 --> 00:37:42,730
to do that.

1050
00:37:42,730 --> 00:37:44,790
The reason is that the
3D structure of our face

1051
00:37:44,790 --> 00:37:46,956
stays the same even when
the lighting conditions are

1052
00:37:46,956 --> 00:37:47,490
changing.

1053
00:37:47,490 --> 00:37:49,530
And so no matter whether
light is shining from,

1054
00:37:49,530 --> 00:37:51,420
typically the eye
regions because they're

1055
00:37:51,420 --> 00:37:53,400
receded relative to
nose and forehead,

1056
00:37:53,400 --> 00:37:54,817
are darker than
nose and forehead.

1057
00:37:54,817 --> 00:37:56,775
And so he found that in
the human psychophysics

1058
00:37:56,775 --> 00:37:59,010
you have 12 heuristics like
this, forehead brighter

1059
00:37:59,010 --> 00:38:00,690
than left eye, forehead
brighter than right eye,

1060
00:38:00,690 --> 00:38:02,731
nose brighter than mouth,
and so on and so forth.

1061
00:38:02,731 --> 00:38:04,680
Twelve of these
characteristics that actually

1062
00:38:04,680 --> 00:38:06,580
together can allow you
to detect the face.

1063
00:38:06,580 --> 00:38:08,880
And in fact, your face
detector on your cell phone

1064
00:38:08,880 --> 00:38:10,749
is using a very similar
strategy as well.

1065
00:38:10,749 --> 00:38:12,790
Trying to find the coarse,
contrast relationships

1066
00:38:12,790 --> 00:38:14,100
in the scene.

1067
00:38:14,100 --> 00:38:16,680
So what Shay did was he
started with a real face,

1068
00:38:16,680 --> 00:38:19,050
he would parse it into
11 parts, and then

1069
00:38:19,050 --> 00:38:20,940
randomly assigned 11
different luminance

1070
00:38:20,940 --> 00:38:25,230
values to these 11 different
parts and change this rapidly.

1071
00:38:25,230 --> 00:38:27,149
And then the analysis
would look like this.

1072
00:38:27,149 --> 00:38:29,190
So no matter what the
overall pattern looks like,

1073
00:38:29,190 --> 00:38:31,380
he's going to look for
a particular contrast

1074
00:38:31,380 --> 00:38:33,644
relationship with the
forehead versus the left eye.

1075
00:38:33,644 --> 00:38:36,060
And he's going to ask is the
neuron responding differently

1076
00:38:36,060 --> 00:38:38,143
in these conditions where
the forehead is brighter

1077
00:38:38,143 --> 00:38:40,170
than the left eye versus
in these conditions

1078
00:38:40,170 --> 00:38:42,990
where the left eye is brighter
than the left forehead.

1079
00:38:42,990 --> 00:38:46,650
And you can do this for
all pairs of combinations

1080
00:38:46,650 --> 00:38:48,390
of 11 different parts.

1081
00:38:48,390 --> 00:38:50,940
And in these 55
different combinations,

1082
00:38:50,940 --> 00:38:52,920
we can mark by
arrows the prediction

1083
00:38:52,920 --> 00:38:54,430
from human phycophysics.

1084
00:38:54,430 --> 00:38:57,270
So human psychophysics told
us 12 of these constrast pairs

1085
00:38:57,270 --> 00:38:58,446
are going to be important.

1086
00:38:58,446 --> 00:38:59,820
It also told us
what the polarity

1087
00:38:59,820 --> 00:39:02,327
was that was going to be
important for detecting a face.

1088
00:39:02,327 --> 00:39:04,410
Again, forehead brighter
than eye, or eye brighter

1089
00:39:04,410 --> 00:39:06,380
than forehead.

1090
00:39:06,380 --> 00:39:08,800
OK and this is what
we actually found.

1091
00:39:08,800 --> 00:39:11,040
So what this shows is
a population diagram

1092
00:39:11,040 --> 00:39:14,370
of all the cells that Shay found
and it was half of the cells

1093
00:39:14,370 --> 00:39:17,120
that he recorded from that
showed some selectivity

1094
00:39:17,120 --> 00:39:19,050
to some of these
contrast features

1095
00:39:19,050 --> 00:39:21,060
and to some of these
contrast polarities.

1096
00:39:21,060 --> 00:39:23,220
What's plotted here
upwards is for one contrast

1097
00:39:23,220 --> 00:39:26,220
polarity and one
particular contrast pair,

1098
00:39:26,220 --> 00:39:28,622
the number of cells he
found that were selective.

1099
00:39:28,622 --> 00:39:30,330
And as you go through
the entire diagram,

1100
00:39:30,330 --> 00:39:33,061
you see that there are only a
very few examples of contrast

1101
00:39:33,061 --> 00:39:34,560
pairs where actually
different cells

1102
00:39:34,560 --> 00:39:36,480
like different polarities.

1103
00:39:36,480 --> 00:39:38,460
Like here for example
you have more than 60,

1104
00:39:38,460 --> 00:39:40,920
70 cells that all like
one polarity and not

1105
00:39:40,920 --> 00:39:43,110
a single one that likes
the opposite polarity.

1106
00:39:43,110 --> 00:39:45,190
And this is true for
all these polarities.

1107
00:39:45,190 --> 00:39:47,040
So it's a very
consistent pattern.

1108
00:39:47,040 --> 00:39:50,700
Second, we can explain all the
human psychophysics preferences

1109
00:39:50,700 --> 00:39:51,890
here.

1110
00:39:51,890 --> 00:39:53,970
Not only that, these
are important dimensions

1111
00:39:53,970 --> 00:39:54,979
for these cells.

1112
00:39:54,979 --> 00:39:56,520
But in all these
cases the cells care

1113
00:39:56,520 --> 00:39:58,936
for the exact same polarity
that they would have predicted

1114
00:39:58,936 --> 00:40:00,000
from human psychophysics.

1115
00:40:00,000 --> 00:40:02,550
In addition, there are other
contrast pairs that apparently

1116
00:40:02,550 --> 00:40:04,470
don't matter this much
in human psychophysics,

1117
00:40:04,470 --> 00:40:06,300
but these cells also care about.

1118
00:40:06,300 --> 00:40:08,750
So they seem to be using these
coarse contrast features.

1119
00:40:08,750 --> 00:40:10,834
And again, they're very
useful for face detection.

1120
00:40:10,834 --> 00:40:13,333
Now we got the behavior that
we know that the areas involved

1121
00:40:13,333 --> 00:40:14,970
in the face detection
with stimuli

1122
00:40:14,970 --> 00:40:17,449
where it's actually hard
to make out the detail.

1123
00:40:17,449 --> 00:40:18,990
Then Shay did a
control and I thought

1124
00:40:18,990 --> 00:40:21,180
this was really the
coolest thing ever.

1125
00:40:21,180 --> 00:40:22,930
He's a computer scientist
and so of course

1126
00:40:22,930 --> 00:40:24,990
he knew about databases
and how to use them.

1127
00:40:24,990 --> 00:40:26,531
And he would say,
OK, can we actually

1128
00:40:26,531 --> 00:40:30,390
fool these cells into responding
to non-face stimuli that

1129
00:40:30,390 --> 00:40:33,540
comply by the rules of this
coarse contrast of faces?

1130
00:40:33,540 --> 00:40:36,630
So here are some examples.

1131
00:40:36,630 --> 00:40:38,460
This is just a
pattern where there

1132
00:40:38,460 --> 00:40:40,810
are some dark regions where
the human face eyes might be

1133
00:40:40,810 --> 00:40:41,540
and so on and so forth.

1134
00:40:41,540 --> 00:40:43,498
This is a pattern that
only has one of these 12

1135
00:40:43,498 --> 00:40:44,620
contrasts correct.

1136
00:40:44,620 --> 00:40:46,650
But also in human faces,
you can find some--

1137
00:40:46,650 --> 00:40:48,732
when a person is
smiling, wearing glasses,

1138
00:40:48,732 --> 00:40:50,190
most of this
contrasts are actually

1139
00:40:50,190 --> 00:40:52,645
not in the face that
should be there.

1140
00:40:52,645 --> 00:40:54,990
So there are some very contrast
correct faces and there

1141
00:40:54,990 --> 00:40:57,962
are some faces that are not
very very contrast correct.

1142
00:40:57,962 --> 00:40:59,670
And now you can ask
how does the response

1143
00:40:59,670 --> 00:41:02,490
of the cells in the middle
face patch change as you are

1144
00:41:02,490 --> 00:41:04,500
increasing the number
of correct contrasts,

1145
00:41:04,500 --> 00:41:07,092
either on the face
or non-face stimuli.

1146
00:41:07,092 --> 00:41:08,050
And the answer is this.

1147
00:41:08,050 --> 00:41:10,620
If you increase the number
of contrasts in the face,

1148
00:41:10,620 --> 00:41:12,294
the cells respond more and more.

1149
00:41:12,294 --> 00:41:14,210
If you change the contrast
of non-face objects

1150
00:41:14,210 --> 00:41:15,402
the cells don't care.

1151
00:41:15,402 --> 00:41:17,360
So there is something
else that caused contrast

1152
00:41:17,360 --> 00:41:19,235
that the cells care
about, they're not easily

1153
00:41:19,235 --> 00:41:21,540
fooled to respond to
things that clearly

1154
00:41:21,540 --> 00:41:25,890
aren't faces when the coarse
contrasts are correct.

1155
00:41:25,890 --> 00:41:27,914
And we could actually
have predicted something

1156
00:41:27,914 --> 00:41:29,830
like this should happen
from an earlier study.

1157
00:41:29,830 --> 00:41:31,350
So the first study
we did where we

1158
00:41:31,350 --> 00:41:33,810
took advantage of the fact
that in a face selective area

1159
00:41:33,810 --> 00:41:36,120
we can record from faces
over and over again

1160
00:41:36,120 --> 00:41:37,840
and that they have
similar properties

1161
00:41:37,840 --> 00:41:40,440
was a study where we looked at
the effect of part and whole.

1162
00:41:40,440 --> 00:41:42,975
This is one of the central
features in psychophysical

1163
00:41:42,975 --> 00:41:45,930
of human faces is that you can
get information from the face

1164
00:41:45,930 --> 00:41:48,600
without any detail just
from the gist of the face.

1165
00:41:48,600 --> 00:41:50,940
An example is again
from Pavan Sinha.

1166
00:41:50,940 --> 00:41:54,270
If you have a blurred face
of a familiar individual,

1167
00:41:54,270 --> 00:41:57,391
like some of you might
recognize Woody Allen here,

1168
00:41:57,391 --> 00:41:59,640
of course with the glasses
it's a little bit cheating,

1169
00:41:59,640 --> 00:42:01,290
but you can recognize him.

1170
00:42:01,290 --> 00:42:02,970
And the other
examples in the study

1171
00:42:02,970 --> 00:42:04,530
were people who
don't wear glasses,

1172
00:42:04,530 --> 00:42:06,630
so you can recognize a famous
face just from the gist of it.

1173
00:42:06,630 --> 00:42:07,880
You don't need the details.

1174
00:42:07,880 --> 00:42:09,630
On the other hand, we
can process details.

1175
00:42:09,630 --> 00:42:11,310
We can focus on details.

1176
00:42:11,310 --> 00:42:14,880
And so how do these two
things relate to each other?

1177
00:42:14,880 --> 00:42:18,360
What we did was we
constructed a face space,

1178
00:42:18,360 --> 00:42:21,210
a cartoon face space, based
on very, very simple geometric

1179
00:42:21,210 --> 00:42:21,930
shapes.

1180
00:42:21,930 --> 00:42:24,310
So these faces are just made
out of ovals, and triangles,

1181
00:42:24,310 --> 00:42:25,080
and lines.

1182
00:42:25,080 --> 00:42:27,760
So very simple geometric shapes.

1183
00:42:27,760 --> 00:42:30,390
But if they are put together
they actually look like faces.

1184
00:42:30,390 --> 00:42:32,139
Now we can parameterize
this face space,

1185
00:42:32,139 --> 00:42:33,430
we can vary certain parameters.

1186
00:42:33,430 --> 00:42:37,070
So we had faces that
change in aspect ratio.

1187
00:42:37,070 --> 00:42:40,230
They go from Sesame
character Ernie here to Bert.

1188
00:42:40,230 --> 00:42:44,100
We have pupil size, like no
pupil here, very big pupils.

1189
00:42:44,100 --> 00:42:46,060
We have inter-eye distance here.

1190
00:42:46,060 --> 00:42:49,440
So these eyes are close together
almost like cyclopian fashion

1191
00:42:49,440 --> 00:42:52,070
or they can be very far
apart from each other,

1192
00:42:52,070 --> 00:42:55,450
stretching the outside of the
face and so on and so forth.

1193
00:42:55,450 --> 00:42:57,694
And we would now randomly
change these features,

1194
00:42:57,694 --> 00:42:59,360
all these different
features dimensions,

1195
00:42:59,360 --> 00:43:01,819
randomly choosing a new value
every time we show this face.

1196
00:43:01,819 --> 00:43:03,443
It looks a bit like
a cartoon character

1197
00:43:03,443 --> 00:43:04,809
who is trying to talk to you.

1198
00:43:04,809 --> 00:43:06,600
But the way we analyze
this is very simple.

1199
00:43:06,600 --> 00:43:09,270
So we just asked no matter
what these other features are,

1200
00:43:09,270 --> 00:43:11,330
for the first feature
dimension mention does

1201
00:43:11,330 --> 00:43:13,830
the firing of the cell change
as we're changing this feature

1202
00:43:13,830 --> 00:43:14,380
dimension.

1203
00:43:14,380 --> 00:43:15,630
Then we asked this for
the second dimension,

1204
00:43:15,630 --> 00:43:18,005
the third dimension, and so
on and so forth for all these

1205
00:43:18,005 --> 00:43:19,410
slightly different dimensions.

1206
00:43:19,410 --> 00:43:21,840
What we found is shown
here for one example cell.

1207
00:43:21,840 --> 00:43:23,954
So we had 19 different
tuning curves.

1208
00:43:23,954 --> 00:43:25,620
And of these 19
different tuning curves,

1209
00:43:25,620 --> 00:43:27,422
for four are
significantly tuned.

1210
00:43:27,422 --> 00:43:29,880
For this particular cell it
was face aspect ratio so didn't

1211
00:43:29,880 --> 00:43:31,062
like Ernie, it liked Bert.

1212
00:43:31,062 --> 00:43:33,270
It liked the eyes very close
together, not far apart.

1213
00:43:33,270 --> 00:43:36,120
It likes the eyes a little
bit narrow, not wide.

1214
00:43:36,120 --> 00:43:38,252
And it liked big
irises, not small ones.

1215
00:43:38,252 --> 00:43:39,960
What's very typical
for how the cells are

1216
00:43:39,960 --> 00:43:42,810
processing these features are
these ram shaped tuning curves.

1217
00:43:42,810 --> 00:43:45,750
So more than 2/3 of the tuning
curves have this ramp shape.

1218
00:43:45,750 --> 00:43:48,282
Which means that these cells
are relaying the information

1219
00:43:48,282 --> 00:43:50,490
that they are measuring
almost in one to one fashion.

1220
00:43:53,500 --> 00:43:55,590
This is not what the
cells are actually doing.

1221
00:43:55,590 --> 00:43:57,030
It's just a metaphor.

1222
00:43:57,030 --> 00:43:59,280
But it's almost like they're
taking a ruler, measuring

1223
00:43:59,280 --> 00:44:00,696
eye distance, and
they're relaying

1224
00:44:00,696 --> 00:44:03,900
this feature in almost one to
one fashion in their output.

1225
00:44:03,900 --> 00:44:07,320
Another implication is that
most of your coding capacity

1226
00:44:07,320 --> 00:44:09,522
is actually at the extremes
because many cells have

1227
00:44:09,522 --> 00:44:11,730
big responses there, many
cells have small responses.

1228
00:44:11,730 --> 00:44:13,630
So most of the
capacity is there.

1229
00:44:13,630 --> 00:44:15,390
That's the range where
caricatures live.

1230
00:44:15,390 --> 00:44:17,940
And oftentimes we are better
able to recognize individuals

1231
00:44:17,940 --> 00:44:21,950
based on caricatures than
the individuals themselves.

1232
00:44:21,950 --> 00:44:23,640
So in the middle
the face patches,

1233
00:44:23,640 --> 00:44:26,181
they're causally and selectively
relevant for face detection.

1234
00:44:26,181 --> 00:44:27,990
The cells are virtually
all face selective.

1235
00:44:27,990 --> 00:44:30,030
Based on these two findings
we actually suggest

1236
00:44:30,030 --> 00:44:33,000
and it's a little like Nancy
said as a strong signal

1237
00:44:33,000 --> 00:44:35,989
you're putting out that you
get backlash if there is.

1238
00:44:35,989 --> 00:44:37,530
So we do think that
there are modules

1239
00:44:37,530 --> 00:44:39,780
that are there for face
processing and face processing

1240
00:44:39,780 --> 00:44:40,585
only.

1241
00:44:40,585 --> 00:44:42,960
The gain of the tuning curve
is modulated by the presence

1242
00:44:42,960 --> 00:44:44,320
of the entire face.

1243
00:44:44,320 --> 00:44:46,741
There's this ramp shaped
tuning, which is very useful.

1244
00:44:46,741 --> 00:44:48,990
The cells are sensitive to
contrast relations which is

1245
00:44:48,990 --> 00:44:51,700
very useful for face detection.

1246
00:44:51,700 --> 00:44:55,060
So we can really get
mechanistic about understanding

1247
00:44:55,060 --> 00:44:55,860
face recognition.

1248
00:44:55,860 --> 00:44:57,330
It's not just that we
can say, OK, these cells

1249
00:44:57,330 --> 00:44:58,830
are responding more
to faces but we

1250
00:44:58,830 --> 00:45:00,871
can say why they're
responding more to some faces

1251
00:45:00,871 --> 00:45:01,780
than to others.

1252
00:45:01,780 --> 00:45:03,821
In fact, you can predict
from the cartoon results

1253
00:45:03,821 --> 00:45:07,200
how the cells are responding
to pictures of actual people

1254
00:45:07,200 --> 00:45:09,064
with the very fine
details physically.

1255
00:45:09,064 --> 00:45:10,980
And so at the level of
the middle face patches

1256
00:45:10,980 --> 00:45:12,060
we already have some
of the requirements

1257
00:45:12,060 --> 00:45:13,232
for face recognition system.

1258
00:45:13,232 --> 00:45:14,940
So we have mechanisms
for face detection,

1259
00:45:14,940 --> 00:45:16,860
we have some encoding
of facial features,

1260
00:45:16,860 --> 00:45:18,832
and we have encoding
of configurations.

1261
00:45:21,810 --> 00:45:23,362
Nancy said I should
talk about this.

1262
00:45:23,362 --> 00:45:24,570
I'm going to talk about this.

1263
00:45:24,570 --> 00:45:29,025
Sebastian Miller was a wonderful
grad student with Doris and me.

1264
00:45:29,025 --> 00:45:30,400
He asked the
following questions.

1265
00:45:30,400 --> 00:45:32,996
So whether the face pictures are
related to each other or not.

1266
00:45:32,996 --> 00:45:34,620
If you look at the
overall organization

1267
00:45:34,620 --> 00:45:37,420
these face errors are very
far apart from each other.

1268
00:45:37,420 --> 00:45:39,170
So the most posterior
to the most anterior

1269
00:45:39,170 --> 00:45:41,336
is one inch apart, it's a
third of the entire extent

1270
00:45:41,336 --> 00:45:42,271
of the primate brain.

1271
00:45:42,271 --> 00:45:44,520
They live in different
cytoarchitectonic environments.

1272
00:45:44,520 --> 00:45:46,728
So you could also imagine
that maybe the connectivity

1273
00:45:46,728 --> 00:45:47,842
is mostly local.

1274
00:45:47,842 --> 00:45:49,800
On the other hand, they
are interested in faces

1275
00:45:49,800 --> 00:45:51,510
and so you would
imagine that maybe there

1276
00:45:51,510 --> 00:45:54,084
are specialized
connections between them.

1277
00:45:54,084 --> 00:45:56,250
The way we addressed this
was with micro stimulation

1278
00:45:56,250 --> 00:45:57,041
inside the scanner.

1279
00:45:57,041 --> 00:45:58,950
So we would first
image the face areas.

1280
00:45:58,950 --> 00:46:02,570
We would then lower an electrode
to one of the face areas here,

1281
00:46:02,570 --> 00:46:05,220
we record from cells to make
sure it's face selective,

1282
00:46:05,220 --> 00:46:07,110
but then we would
use this electrode

1283
00:46:07,110 --> 00:46:09,107
to pass a current through
inside the scanner.

1284
00:46:09,107 --> 00:46:10,690
Passing a current
through an electrode

1285
00:46:10,690 --> 00:46:11,760
is going to activate cells.

1286
00:46:11,760 --> 00:46:14,218
That in turn is going to change
blood flow and oxygenation.

1287
00:46:14,218 --> 00:46:16,000
So things we can pick
up on the scanner.

1288
00:46:16,000 --> 00:46:18,541
And so yes, if this worked you
should get a swath of activity

1289
00:46:18,541 --> 00:46:19,980
around your simulation site.

1290
00:46:19,980 --> 00:46:21,930
But if these cells at
your stimulation site

1291
00:46:21,930 --> 00:46:24,554
have predictions that are strong
and focal enough to drive down

1292
00:46:24,554 --> 00:46:26,670
some neurons, you might
also find activation

1293
00:46:26,670 --> 00:46:29,190
that spatially remote
locations and they can then

1294
00:46:29,190 --> 00:46:32,742
see where these locations are
related to the face areas.

1295
00:46:32,742 --> 00:46:34,200
So here is a computer
flattened map

1296
00:46:34,200 --> 00:46:36,990
of the brain of one
macaque monkeys.

1297
00:46:36,990 --> 00:46:39,360
The green areas
indicate by outline

1298
00:46:39,360 --> 00:46:41,070
the extent of the face areas.

1299
00:46:41,070 --> 00:46:42,510
We placed our
simulation electrode

1300
00:46:42,510 --> 00:46:43,992
in one of the face
areas and this

1301
00:46:43,992 --> 00:46:45,700
is the map we got from
micro stimulation,

1302
00:46:45,700 --> 00:46:46,876
versus no micro stimulation.

1303
00:46:46,876 --> 00:46:48,750
So there's no visual
stimulus there, actually

1304
00:46:48,750 --> 00:46:51,090
that works during sleep
during complete darkness.

1305
00:46:51,090 --> 00:46:52,560
So yes, we get a
swath of activity

1306
00:46:52,560 --> 00:46:54,120
around the stimulation
site, and we

1307
00:46:54,120 --> 00:46:57,385
get multiple spatially disjunct
regions that are activated.

1308
00:46:57,385 --> 00:46:59,010
So they're strongly
driven by the cells

1309
00:46:59,010 --> 00:47:02,464
in this region and these
overlap with the face areas.

1310
00:47:02,464 --> 00:47:04,380
And this was either be
found very consistently

1311
00:47:04,380 --> 00:47:06,630
across different phase areas
if you stimulated outside

1312
00:47:06,630 --> 00:47:08,760
you also got this patchy
pattern of connectivity,

1313
00:47:08,760 --> 00:47:11,155
but now it's outside
of the face system.

1314
00:47:11,155 --> 00:47:12,780
And so this is the
picture that we got.

1315
00:47:12,780 --> 00:47:14,488
Is that yes these face
areas are actually

1316
00:47:14,488 --> 00:47:17,970
part of a network of face
areas that are strongly

1317
00:47:17,970 --> 00:47:19,330
interconnected with each other.

1318
00:47:19,330 --> 00:47:21,580
There's now data from
retrograde tracer studies,

1319
00:47:21,580 --> 00:47:23,670
we find that 90% of
the cell bodies that

1320
00:47:23,670 --> 00:47:26,530
are labeled after an
injection inside a face area

1321
00:47:26,530 --> 00:47:29,740
are inside other face areas
or in the same face areas.

1322
00:47:29,740 --> 00:47:33,150
So it's a surprisingly
anatomically specialized

1323
00:47:33,150 --> 00:47:35,310
and closed network.

1324
00:47:35,310 --> 00:47:37,560
So what's happening
in these areas,

1325
00:47:37,560 --> 00:47:40,540
and again, my movie
isn't going to work.

1326
00:47:40,540 --> 00:47:44,461
So in this area AL was more
anterior, also virtually

1327
00:47:44,461 --> 00:47:45,960
all of the cells
are face selective.

1328
00:47:45,960 --> 00:47:47,668
But you have a property
emerge here which

1329
00:47:47,668 --> 00:47:48,670
you didn't have before.

1330
00:47:48,670 --> 00:47:50,040
And that is mere
symmetric confusion.

1331
00:47:50,040 --> 00:47:51,270
And it's something
that we did not expect.

1332
00:47:51,270 --> 00:47:52,770
We're still puzzled by it.

1333
00:47:52,770 --> 00:47:56,456
We have no explanation
why it's happening.

1334
00:47:56,456 --> 00:47:58,830
But in this area you have
cells that like a profile view.

1335
00:47:58,830 --> 00:48:00,390
And if they like
one profile view,

1336
00:48:00,390 --> 00:48:02,265
they also like the
opposite profile view.

1337
00:48:02,265 --> 00:48:04,140
And this region here as
I mentioned initially

1338
00:48:04,140 --> 00:48:06,485
that some of the cells
did not really seem to be

1339
00:48:06,485 --> 00:48:07,110
face selective.

1340
00:48:07,110 --> 00:48:08,280
It's a small percentage.

1341
00:48:08,280 --> 00:48:10,885
But actually these cells are
selective for facial profile.

1342
00:48:10,885 --> 00:48:13,260
But if they like one profile
right, they don't like left.

1343
00:48:13,260 --> 00:48:14,968
If they like left,
they don't like right.

1344
00:48:14,968 --> 00:48:17,620
In AL, this is being confused.

1345
00:48:17,620 --> 00:48:19,620
And then if you go
to AM, you have cells

1346
00:48:19,620 --> 00:48:21,245
that respond to all faces.

1347
00:48:21,245 --> 00:48:22,620
It doesn't matter
where they are,

1348
00:48:22,620 --> 00:48:24,119
doesn't matter who
they are, doesn't

1349
00:48:24,119 --> 00:48:25,910
matter how big they are.

1350
00:48:25,910 --> 00:48:29,650
And the other cells that also
don't care where they are

1351
00:48:29,650 --> 00:48:32,130
and how big they are,
but they care exquisitely

1352
00:48:32,130 --> 00:48:33,517
about identity.

1353
00:48:33,517 --> 00:48:35,100
So they can be very,
very finely tuned

1354
00:48:35,100 --> 00:48:37,680
to identity in
particular to people

1355
00:48:37,680 --> 00:48:39,750
that the animals never
see in real life.

1356
00:48:39,750 --> 00:48:42,060
So there seems to be a
computation going on from here

1357
00:48:42,060 --> 00:48:48,390
to here where in Jim
DiCarlo's conceptual framework

1358
00:48:48,390 --> 00:48:51,330
you could imagine that there's
a manifold that's now becoming

1359
00:48:51,330 --> 00:48:55,020
sort of flatter and more like
a more explicit representation.

1360
00:48:55,020 --> 00:48:56,520
And for some reason
creating this

1361
00:48:56,520 --> 00:48:59,439
has to go through this
mirror symmetric confusion.

1362
00:48:59,439 --> 00:49:00,480
I just want to highlight.

1363
00:49:00,480 --> 00:49:02,195
So we meant to touch
upon the question

1364
00:49:02,195 --> 00:49:04,320
whether a face area should
do different coputations

1365
00:49:04,320 --> 00:49:06,090
from non-face areas.

1366
00:49:06,090 --> 00:49:07,590
Actually, my
intuition about this

1367
00:49:07,590 --> 00:49:09,100
was actually quite the opposite.

1368
00:49:09,100 --> 00:49:11,350
So I thought they would
likely do the same computation

1369
00:49:11,350 --> 00:49:12,766
or hopefully the
same computations

1370
00:49:12,766 --> 00:49:15,500
as outside the areas just
in different material.

1371
00:49:15,500 --> 00:49:16,792
So why in other non-face areas?

1372
00:49:16,792 --> 00:49:18,374
Why don't you want
to mix these cells?

1373
00:49:18,374 --> 00:49:20,460
We have one study that's
a little too complicated

1374
00:49:20,460 --> 00:49:22,532
to explain here that
gives some clues.

1375
00:49:22,532 --> 00:49:24,240
But some computation
work from Joel Leibo

1376
00:49:24,240 --> 00:49:25,823
was a grad student
with Tommy actually

1377
00:49:25,823 --> 00:49:27,180
gave some clues to that.

1378
00:49:27,180 --> 00:49:30,840
So Joel and Tommy were
thinking about invariance.

1379
00:49:30,840 --> 00:49:32,370
And Tommy told me
he's going to talk

1380
00:49:32,370 --> 00:49:35,520
about this at some later
point in the course.

1381
00:49:35,520 --> 00:49:37,590
The different kinds of
transformations, easy ones

1382
00:49:37,590 --> 00:49:38,430
and difficult ones.

1383
00:49:38,430 --> 00:49:41,010
The easy ones are
affine transformations.

1384
00:49:41,010 --> 00:49:44,430
We're just shifting something
in space or in size,

1385
00:49:44,430 --> 00:49:45,990
or we rotate it in plane.

1386
00:49:45,990 --> 00:49:47,982
And so if you learn
how to correct

1387
00:49:47,982 --> 00:49:50,190
for this transformation for
just three dots of light,

1388
00:49:50,190 --> 00:49:52,320
you can look for any image
that you can ever see.

1389
00:49:52,320 --> 00:49:53,470
So this is relatively easy.

1390
00:49:53,470 --> 00:49:55,470
But then they're non-affine
transformations that

1391
00:49:55,470 --> 00:49:56,845
are actually
changing the picture

1392
00:49:56,845 --> 00:49:58,560
and they are very
difficult. So if you

1393
00:49:58,560 --> 00:50:00,730
change your facial
expression for example,

1394
00:50:00,730 --> 00:50:02,370
or if lighting
conditions are changing,

1395
00:50:02,370 --> 00:50:05,231
or if you're turning
your head in depth,

1396
00:50:05,231 --> 00:50:06,730
this is a non-affine
transformation.

1397
00:50:06,730 --> 00:50:08,736
So it's not predictable
from just three dots.

1398
00:50:08,736 --> 00:50:10,110
And you can learn
something there

1399
00:50:10,110 --> 00:50:13,020
that could tell you
how this picture would

1400
00:50:13,020 --> 00:50:15,030
look like under this
non-affine transformation.

1401
00:50:15,030 --> 00:50:17,040
And one of the insights from
Joel was that if you learn this

1402
00:50:17,040 --> 00:50:19,610
non-affine transformation on
a particular object category--

1403
00:50:19,610 --> 00:50:20,700
let's say faces--

1404
00:50:20,700 --> 00:50:22,949
you actually have learned
nothing about another object

1405
00:50:22,949 --> 00:50:24,720
category like cars.

1406
00:50:24,720 --> 00:50:26,352
That's actually quite
surprising to me

1407
00:50:26,352 --> 00:50:27,810
but there could be
a reason why you

1408
00:50:27,810 --> 00:50:29,351
might want to have
all the cells that

1409
00:50:29,351 --> 00:50:31,890
have to learn representations
across one transformation put

1410
00:50:31,890 --> 00:50:34,884
them all in one location.

1411
00:50:34,884 --> 00:50:36,300
The second insight
they had, and I

1412
00:50:36,300 --> 00:50:38,490
think it's still
very surprising to me

1413
00:50:38,490 --> 00:50:40,120
that it actually
works so easily,

1414
00:50:40,120 --> 00:50:42,630
is to give the computation
a count for the system

1415
00:50:42,630 --> 00:50:44,880
that I just described to
you in qualitative terms.

1416
00:50:44,880 --> 00:50:47,620
So we have three
levels of processing.

1417
00:50:47,620 --> 00:50:49,530
So we have a front
end where cells

1418
00:50:49,530 --> 00:50:51,360
are useful for face
detection, they're

1419
00:50:51,360 --> 00:50:52,860
all very face selective.

1420
00:50:52,860 --> 00:50:55,560
So you could think of this
like a three level processing

1421
00:50:55,560 --> 00:50:58,800
hierarchy where the level one
is like a face filter that's

1422
00:50:58,800 --> 00:51:01,590
just going to tell you
it's their face or not.

1423
00:51:01,590 --> 00:51:03,450
The top level you want
an identification.

1424
00:51:03,450 --> 00:51:04,650
And I didn't show
the examples, again

1425
00:51:04,650 --> 00:51:06,390
I hope with a connection
to the monitor I

1426
00:51:06,390 --> 00:51:07,804
can show you the actual movies.

1427
00:51:07,804 --> 00:51:09,720
You have some cells that
are very, very finely

1428
00:51:09,720 --> 00:51:11,322
selective of facial identity.

1429
00:51:11,322 --> 00:51:12,780
With pattern readout
techniques you

1430
00:51:12,780 --> 00:51:17,149
can read out identity
extremely reliably.

1431
00:51:17,149 --> 00:51:18,940
If you now have like
Hebbian Learning Rule,

1432
00:51:18,940 --> 00:51:20,940
maybe Tommy is going to
explain this to you more,

1433
00:51:20,940 --> 00:51:22,606
you actually get
something pretty magic.

1434
00:51:22,606 --> 00:51:24,976
So you do get invariance
at level number three,

1435
00:51:24,976 --> 00:51:27,600
which is kind of what you wanted
and might not be surprised by.

1436
00:51:27,600 --> 00:51:29,724
But as a byproduct, you
are getting mirror symmetry

1437
00:51:29,724 --> 00:51:30,402
at level two.

1438
00:51:30,402 --> 00:51:32,610
And that's something you
didn't stick into the system

1439
00:51:32,610 --> 00:51:34,674
and it just happens,
not like magic,

1440
00:51:34,674 --> 00:51:36,090
but there's an
explanation for why

1441
00:51:36,090 --> 00:51:38,310
this happens, out of
very general assumptions

1442
00:51:38,310 --> 00:51:39,670
about the system.

1443
00:51:39,670 --> 00:51:41,520
So the point I
want to make here,

1444
00:51:41,520 --> 00:51:43,200
this particular
model could be wrong,

1445
00:51:43,200 --> 00:51:45,927
but it's something about
how knowing something

1446
00:51:45,927 --> 00:51:47,760
about the overall
organization of the system

1447
00:51:47,760 --> 00:51:51,644
might actually reveal underlying
relationships that you

1448
00:51:51,644 --> 00:51:52,560
might not think about.

1449
00:51:52,560 --> 00:51:54,240
So the fact that there are
three levels of processing

1450
00:51:54,240 --> 00:51:56,430
and not four or five
or six might actually

1451
00:51:56,430 --> 00:51:58,389
impact whether you find
mirror symmetry or not.

1452
00:51:58,389 --> 00:51:59,846
Or whether you find
mirrir symmetry

1453
00:51:59,846 --> 00:52:01,110
at one level or another level.

1454
00:52:01,110 --> 00:52:04,380
And you don't necessarily
get this automatically.

1455
00:52:04,380 --> 00:52:07,140
Just put any processing
system together.

1456
00:52:07,140 --> 00:52:09,180
Becuase I was mentioning
a facial motion,

1457
00:52:09,180 --> 00:52:12,810
I would like to give a
brief vignette of that.

1458
00:52:12,810 --> 00:52:17,115
And so I was emphasizing
here transformations

1459
00:52:17,115 --> 00:52:17,990
along this direction.

1460
00:52:17,990 --> 00:52:20,640
But you can see that at least
two levels of processing

1461
00:52:20,640 --> 00:52:22,710
there are actually
two face areas here,

1462
00:52:22,710 --> 00:52:25,240
one lateral to the STS, and
the other one deep inside.

1463
00:52:25,240 --> 00:52:27,739
And so one of the questions we
had was what's going on here?

1464
00:52:27,739 --> 00:52:29,110
How are they different?

1465
00:52:29,110 --> 00:52:31,740
And so one way you
might think about this

1466
00:52:31,740 --> 00:52:34,060
is again, connecting faces
to social perception.

1467
00:52:34,060 --> 00:52:35,697
So there are some
faces out there

1468
00:52:35,697 --> 00:52:36,780
that are not really faces.

1469
00:52:36,780 --> 00:52:39,074
And so the faces of dolls
are just one example.

1470
00:52:39,074 --> 00:52:41,115
So physically they are
faces but you can actually

1471
00:52:41,115 --> 00:52:43,899
tell that they are not
really real, they're agents.

1472
00:52:43,899 --> 00:52:46,440
And people like Thalia Wheatley
are wondering about questions

1473
00:52:46,440 --> 00:52:47,940
like why dolls are creepy.

1474
00:52:47,940 --> 00:52:51,120
So there is an
expectation that the face

1475
00:52:51,120 --> 00:52:52,604
should belong to a real agent.

1476
00:52:52,604 --> 00:52:54,770
And there are different
clues that can give it away.

1477
00:52:54,770 --> 00:52:56,728
So if the face is on top
of a body, more likely

1478
00:52:56,728 --> 00:52:59,190
it's an agent just like
an artificial stimulus.

1479
00:52:59,190 --> 00:53:02,110
If a face is moving that's
another clue it's an agent.

1480
00:53:02,110 --> 00:53:03,540
And this will
change fundamentally

1481
00:53:03,540 --> 00:53:05,010
how you interacting with it.

1482
00:53:05,010 --> 00:53:06,676
Your interaction with
the doll is likely

1483
00:53:06,676 --> 00:53:08,720
going to be very different
than with a baby.

1484
00:53:08,720 --> 00:53:10,365
And so again, objects
and their meaning

1485
00:53:10,365 --> 00:53:12,240
are making them actionable
in different ways.

1486
00:53:12,240 --> 00:53:13,860
And we have to understand what
the circuits are that actually

1487
00:53:13,860 --> 00:53:15,222
make this possible.

1488
00:53:15,222 --> 00:53:16,680
So one way to look
at this again is

1489
00:53:16,680 --> 00:53:17,900
to think about these
facial displays.

1490
00:53:17,900 --> 00:53:19,792
I showed the Tonkin
macaques and Clark

1491
00:53:19,792 --> 00:53:22,230
Fisher was an M.D. PhD student
in the lab was actually

1492
00:53:22,230 --> 00:53:23,850
addressing this question.

1493
00:53:23,850 --> 00:53:25,730
And so he made movies
like this one here.

1494
00:53:25,730 --> 00:53:29,070
Luckily there's no sound so
we can actually play them.

1495
00:53:29,070 --> 00:53:31,170
These are movies of
macaque monkeys making

1496
00:53:31,170 --> 00:53:33,170
facial movements of
all different kinds

1497
00:53:33,170 --> 00:53:34,701
of facial expressions.

1498
00:53:34,701 --> 00:53:36,450
And we then also have
stills that are just

1499
00:53:36,450 --> 00:53:38,177
changing from time to time.

1500
00:53:38,177 --> 00:53:40,260
And we have controls of
toys that the animals also

1501
00:53:40,260 --> 00:53:42,000
know that are either
moving or that

1502
00:53:42,000 --> 00:53:44,242
are jumping every second
from one state to the other.

1503
00:53:44,242 --> 00:53:46,200
And we would ask as we
had in an earlier study,

1504
00:53:46,200 --> 00:53:48,330
are these areas
responding to this motion

1505
00:53:48,330 --> 00:53:52,101
differently than to
the static images.

1506
00:53:52,101 --> 00:53:53,100
So here's what he found.

1507
00:53:53,100 --> 00:53:55,450
So he had six different face
areas he was looking at.

1508
00:53:55,450 --> 00:53:57,284
If you're looking at
static form selectivity

1509
00:53:57,284 --> 00:53:59,741
we're just reproducing the way
that the area was found just

1510
00:53:59,741 --> 00:54:01,000
with a different stimulus.

1511
00:54:01,000 --> 00:54:03,720
So all the six areas respond
more to faces than to objects.

1512
00:54:03,720 --> 00:54:06,774
If we now compare moving
faces to static faces,

1513
00:54:06,774 --> 00:54:08,190
all the arrows are
responding more

1514
00:54:08,190 --> 00:54:10,050
to the moving faces
than the static ones.

1515
00:54:10,050 --> 00:54:12,630
Some quantitative differences,
but overall the same pattern,

1516
00:54:12,630 --> 00:54:15,222
more responsive to moving
than static stimuli.

1517
00:54:15,222 --> 00:54:17,430
If you now compare on the
right inside the modulation

1518
00:54:17,430 --> 00:54:19,140
by moving objects
through static objects.

1519
00:54:19,140 --> 00:54:21,902
You can see that also all
the areas or almost all

1520
00:54:21,902 --> 00:54:23,610
have a slight advantage
of moving objects

1521
00:54:23,610 --> 00:54:25,140
over static objects.

1522
00:54:25,140 --> 00:54:28,440
There seems to be a general
motion sensitivity there.

1523
00:54:28,440 --> 00:54:31,624
But if you do the interaction
of shape and motion,

1524
00:54:31,624 --> 00:54:33,540
you can see that all the
areas are selectively

1525
00:54:33,540 --> 00:54:38,520
more enhanced by face motion
than by non-face motion.

1526
00:54:38,520 --> 00:54:40,909
But they all look
pretty similar.

1527
00:54:40,909 --> 00:54:42,450
So now you can
actually wonder if you

1528
00:54:42,450 --> 00:54:44,850
have a contrast like this,
like moving versus still,

1529
00:54:44,850 --> 00:54:46,780
there a couple of things
that are different.

1530
00:54:46,780 --> 00:54:50,207
So is it really about motion or
is it just about the content?

1531
00:54:50,207 --> 00:54:52,540
If you just show a picture
every one second you can say,

1532
00:54:52,540 --> 00:54:54,060
well, there's less
content there,

1533
00:54:54,060 --> 00:54:56,185
therefore you might have
more adaptation, therefore

1534
00:54:56,185 --> 00:54:57,640
less response across the board.

1535
00:54:57,640 --> 00:54:59,250
Is it about update frequency?

1536
00:54:59,250 --> 00:55:02,250
You know, a fast update
versus a slow update.

1537
00:55:02,250 --> 00:55:09,867
So what Clark did, he was
creating another stimulus.

1538
00:55:09,867 --> 00:55:11,700
If you think about
creepiness, it's actually

1539
00:55:11,700 --> 00:55:12,780
like a little bit creepy.

1540
00:55:12,780 --> 00:55:16,150
So like a scrambled version
of a motion that's shown here.

1541
00:55:16,150 --> 00:55:18,750
Shows us the same frames of
the movie but just randomly

1542
00:55:18,750 --> 00:55:19,876
associated with each other.

1543
00:55:19,876 --> 00:55:22,000
So if anything, now the
motion energy in this thing

1544
00:55:22,000 --> 00:55:23,579
is higher than in this one here.

1545
00:55:23,579 --> 00:55:25,620
And we can look now for
the contrast of those two

1546
00:55:25,620 --> 00:55:28,715
and also the contrast
through here.

1547
00:55:28,715 --> 00:55:30,090
And now what he
finds is actually

1548
00:55:30,090 --> 00:55:33,120
something where the face areas
are qualitatively different.

1549
00:55:33,120 --> 00:55:35,430
He finds two face
areas here which

1550
00:55:35,430 --> 00:55:37,320
are responding more
to the natural motion

1551
00:55:37,320 --> 00:55:39,670
than unscrambled motion,
and three face areas

1552
00:55:39,670 --> 00:55:41,670
that are responding more
to the scrambled motion

1553
00:55:41,670 --> 00:55:43,080
than the natural motion.

1554
00:55:43,080 --> 00:55:45,300
So they shift
opposite preferences.

1555
00:55:45,300 --> 00:55:47,040
What we think is
going in the areas

1556
00:55:47,040 --> 00:55:49,950
here that remember also the
benefit for facial motion

1557
00:55:49,950 --> 00:55:52,260
over static faces,
that they just

1558
00:55:52,260 --> 00:55:53,914
like a fast update of content.

1559
00:55:53,914 --> 00:55:55,580
Ideally in a way
that's not predictable.

1560
00:55:55,580 --> 00:55:57,120
If you show me something
new I'm going to respond.

1561
00:55:57,120 --> 00:55:59,520
And if it's something that
I can't predict even better.

1562
00:55:59,520 --> 00:56:00,670
So this is what
these guys are doing.

1563
00:56:00,670 --> 00:56:02,280
But these ones here,
they seem to be

1564
00:56:02,280 --> 00:56:04,410
really sensitive to facial
movement and naturalness

1565
00:56:04,410 --> 00:56:05,860
of facial movement.

1566
00:56:05,860 --> 00:56:09,160
And that was not smart.

1567
00:56:09,160 --> 00:56:11,820
And these areas are located
more deep inside the STS,

1568
00:56:11,820 --> 00:56:15,760
more dorsally, and these areas
are located more ventrally.

1569
00:56:15,760 --> 00:56:18,075
So there's an organization
and to discover the new face

1570
00:56:18,075 --> 00:56:19,283
area they didn't know before.

1571
00:56:19,283 --> 00:56:22,380
It's a seventh face area which
he called MD which is really

1572
00:56:22,380 --> 00:56:23,899
like a face motion area.

1573
00:56:23,899 --> 00:56:26,190
There are lots of reasons
why we're excited about this.

1574
00:56:26,190 --> 00:56:29,310
I mentioned the link between
face perception and agency

1575
00:56:29,310 --> 00:56:29,970
interpretation.

1576
00:56:29,970 --> 00:56:33,540
This is one possible
link, there are more.

1577
00:56:33,540 --> 00:56:35,480
It might be a second
phase processing system,

1578
00:56:35,480 --> 00:56:37,290
I'm going to go through
all the evidence.

1579
00:56:37,290 --> 00:56:39,539
It's only indirect, though
that this area might not

1580
00:56:39,539 --> 00:56:40,830
be connected to the other ones.

1581
00:56:40,830 --> 00:56:43,205
I told you before that the
six face areas are intricately

1582
00:56:43,205 --> 00:56:44,820
integrated to be one network.

1583
00:56:44,820 --> 00:56:47,430
This area never showed
up in stimulation,

1584
00:56:47,430 --> 00:56:48,722
therefore it might be separate.

1585
00:56:48,722 --> 00:56:50,888
And this is kind of nice
because it fits very nicely

1586
00:56:50,888 --> 00:56:51,870
to the human situation.

1587
00:56:51,870 --> 00:56:54,660
So in the human brain you
have the posterior STS face

1588
00:56:54,660 --> 00:56:57,600
area which is exquisitely
sensitive to facial motion.

1589
00:56:57,600 --> 00:57:00,160
Actually you often don't
even get it for static faces.

1590
00:57:00,160 --> 00:57:01,830
But it's very sensitive
to facial motion

1591
00:57:01,830 --> 00:57:03,960
and actually Nancy has a
beautiful study on that.

1592
00:57:03,960 --> 00:57:05,730
And this area by
several accounts

1593
00:57:05,730 --> 00:57:07,620
is not like the other face area.

1594
00:57:07,620 --> 00:57:10,630
So it seems to be
like a specialization.

1595
00:57:10,630 --> 00:57:12,430
That's another
thing, another reason

1596
00:57:12,430 --> 00:57:16,966
why we think these systems might
be connected to each other.

1597
00:57:16,966 --> 00:57:18,340
Just a cool thing
in the end, who

1598
00:57:18,340 --> 00:57:19,810
can recognize this actor here?

1599
00:57:19,810 --> 00:57:22,000
Show of hands.

1600
00:57:22,000 --> 00:57:23,320
OK, who can recognize him now?

1601
00:57:25,960 --> 00:57:30,270
So facial motion gives away a
lot of things like identity.

1602
00:57:30,270 --> 00:57:34,860
Jack Nicholson has very
typical facial movements.

1603
00:57:34,860 --> 00:57:37,350
So it's not just agency, it's
not just facial expressions,

1604
00:57:37,350 --> 00:57:39,641
it's also identity and lots
of things that can go away.

1605
00:57:39,641 --> 00:57:44,170
So we actually don't know yet
what these areas are doing.

1606
00:57:44,170 --> 00:57:47,830
So my summary, and I'm
sorry I'm going into lunch.

1607
00:57:47,830 --> 00:57:50,436
So we can do fMRI on macaque
monkeys just as in humans.

1608
00:57:50,436 --> 00:57:51,810
What we find we
can apply to lots

1609
00:57:51,810 --> 00:57:53,309
of domains within
attention studies,

1610
00:57:53,309 --> 00:57:54,492
found new attention areas.

1611
00:57:54,492 --> 00:57:55,950
Here we applied to
face processing,

1612
00:57:55,950 --> 00:57:58,530
we find face selected
areas for face processing.

1613
00:57:58,530 --> 00:58:00,400
Recording the micro stimulate--

1614
00:58:00,400 --> 00:58:03,780
so inactivating these regions
is supporting the notion

1615
00:58:03,780 --> 00:58:05,940
that these are
likely modules that

1616
00:58:05,940 --> 00:58:10,590
are selective for processing
faces and faces only.

1617
00:58:10,590 --> 00:58:13,800
These are interconnected into
a face processing network.

1618
00:58:13,800 --> 00:58:16,050
It looks like all these areas
have different functions

1619
00:58:16,050 --> 00:58:16,980
and specializations.

1620
00:58:16,980 --> 00:58:19,320
So fMRI experiments
are notoriously

1621
00:58:19,320 --> 00:58:23,576
under powered in the way of
number of different dimensions.

1622
00:58:23,576 --> 00:58:25,200
So if we call them
face areas it is not

1623
00:58:25,200 --> 00:58:26,880
to say that they're
all doing the same,

1624
00:58:26,880 --> 00:58:28,910
but they likely all have
different functions and likely

1625
00:58:28,910 --> 00:58:30,618
sub regions of these
different functions.

1626
00:58:30,618 --> 00:58:33,240
So again, there's no
contradiction here

1627
00:58:33,240 --> 00:58:35,380
to the view of a
fine organization.

1628
00:58:35,380 --> 00:58:36,964
Then there's a seventh
face area which

1629
00:58:36,964 --> 00:58:38,171
doesn't seem to be connected.

1630
00:58:38,171 --> 00:58:40,640
Could be separate and separate
face processing system.

1631
00:58:40,640 --> 00:58:43,170
We have evidence for processing
that we can understand now

1632
00:58:43,170 --> 00:58:44,950
in computational terms.

1633
00:58:44,950 --> 00:58:46,860
And this is one way
that it can link,

1634
00:58:46,860 --> 00:58:49,830
sometimes causally sometimes
correlationally, activity

1635
00:58:49,830 --> 00:58:50,580
of single cells.

1636
00:58:50,580 --> 00:58:52,410
Two different levels
of organization

1637
00:58:52,410 --> 00:58:54,380
to a very complex
social behavior.

1638
00:58:54,380 --> 00:58:56,670
And that's, again, I think
a very cool opportunity

1639
00:58:56,670 --> 00:58:59,169
to have in the domain
of social cognition

1640
00:58:59,169 --> 00:59:01,710
that you can actually control
stimuli very well because faces

1641
00:59:01,710 --> 00:59:04,200
are so powerful to get
into your social brain,

1642
00:59:04,200 --> 00:59:06,870
you can likely take this
approach deeper and get insight

1643
00:59:06,870 --> 00:59:09,450
into actual social intelligence
beyond face perception

1644
00:59:09,450 --> 00:59:11,190
this way.