1
00:00:01,640 --> 00:00:04,040
The following content is
provided under a Creative

2
00:00:04,040 --> 00:00:05,580
Commons license.

3
00:00:05,580 --> 00:00:07,880
Your support will help
MIT OpenCourseWare

4
00:00:07,880 --> 00:00:12,270
continue to offer high quality
educational resources for free.

5
00:00:12,270 --> 00:00:14,870
To make a donation, or
view additional materials

6
00:00:14,870 --> 00:00:18,830
from hundreds of MIT courses,
visit MIT OpenCourseWare

7
00:00:18,830 --> 00:00:20,000
at ocw.mit.edu.

8
00:00:23,327 --> 00:00:25,160
TOMASO POGGIO: This
problem of intelligence,

9
00:00:25,160 --> 00:00:27,950
it's one of those problems that
mankind has been busy with it

10
00:00:27,950 --> 00:00:30,710
for the last 2,000 years or so.

11
00:00:30,710 --> 00:00:35,120
But 50 years ago or
so, that was the start

12
00:00:35,120 --> 00:00:36,620
of artificial intelligence.

13
00:00:36,620 --> 00:00:43,160
It was a conference in
Dartmouth, '62 or so,

14
00:00:43,160 --> 00:00:46,460
with people like John
McCarthy and Marvin Minsky,

15
00:00:46,460 --> 00:00:49,570
who coined the term
artificial intelligence.

16
00:00:49,570 --> 00:00:53,540
And at that time,
progress was made.

17
00:00:53,540 --> 00:00:55,720
Progress has been made,
especially in the last 20

18
00:00:55,720 --> 00:00:56,219
years.

19
00:00:56,219 --> 00:00:57,070
I'll go through it.

20
00:00:59,840 --> 00:01:02,600
But they relied, really,
only on computer science

21
00:01:02,600 --> 00:01:04,430
and common sense.

22
00:01:04,430 --> 00:01:07,040
And in the meantime, there are
all these other disciplines

23
00:01:07,040 --> 00:01:09,470
which have made a
lot of progress,

24
00:01:09,470 --> 00:01:12,110
and that are very likely
to play a key role

25
00:01:12,110 --> 00:01:20,090
in the search for answers to
the problem of intelligence.

26
00:01:20,090 --> 00:01:25,850
So it was obvious that we
needed different expertises.

27
00:01:25,850 --> 00:01:29,200
Not all in computer
science, but in other ones.

28
00:01:29,200 --> 00:01:32,630
And so, this was the
people that we put together

29
00:01:32,630 --> 00:01:36,560
from different labs, from
neuroscience, from computer

30
00:01:36,560 --> 00:01:40,190
science, from cognitive
science, and from a number

31
00:01:40,190 --> 00:01:44,120
of institutions in the US.

32
00:01:44,120 --> 00:01:46,520
Especially MIT and Harvard.

33
00:01:46,520 --> 00:01:52,100
Let me tell you a bit more
about the background here.

34
00:01:52,100 --> 00:01:56,750
This idea of merging brain
research and computer

35
00:01:56,750 --> 00:02:02,240
science in the quest to
understand intelligence.

36
00:02:02,240 --> 00:02:07,250
Part of the reason for this
was progress and convergence

37
00:02:07,250 --> 00:02:09,139
we saw between
different disciplines.

38
00:02:09,139 --> 00:02:13,130
And one of them
was progress in AI.

39
00:02:13,130 --> 00:02:15,500
And this started,
really, with Deep Blue,

40
00:02:15,500 --> 00:02:17,650
I guess it was
called at the time.

41
00:02:17,650 --> 00:02:25,820
The machine IBM that managed
to beat Kasparov at chess

42
00:02:25,820 --> 00:02:27,770
for the world championship.

43
00:02:27,770 --> 00:02:32,810
And then, of course, there
was Watson beating champions

44
00:02:32,810 --> 00:02:34,310
in Jeopardy.

45
00:02:34,310 --> 00:02:39,620
And things like drones able
to land on aircraft carriers.

46
00:02:39,620 --> 00:02:45,150
So that's the most difficult
thing for the pilot to do.

47
00:02:45,150 --> 00:02:49,760
And in the meantime, things had
continued to go pretty fast.

48
00:02:49,760 --> 00:02:54,140
This was the cover of
Nature, probably eight months

49
00:02:54,140 --> 00:02:55,370
ago or so.

50
00:02:55,370 --> 00:02:58,940
DeepMind, which is one of
our industrial partners

51
00:02:58,940 --> 00:03:02,600
in the center, has developed
an artificial intelligence

52
00:03:02,600 --> 00:03:05,900
called DeepQ I
think, that learned

53
00:03:05,900 --> 00:03:11,870
to play better than humans,
49 classical Atari games.

54
00:03:11,870 --> 00:03:14,285
By itself.

55
00:03:14,285 --> 00:03:18,280
And this was two or
three months ago,

56
00:03:18,280 --> 00:03:22,515
a cover of a Nature supplement,
on artificial intelligence

57
00:03:22,515 --> 00:03:23,390
and machine learning.

58
00:03:26,710 --> 00:03:30,280
This is showing a
system by Mobileye,

59
00:03:30,280 --> 00:03:36,030
this is an old video,
that gives vision to cars.

60
00:03:36,030 --> 00:03:38,290
There is a camera
looking outside,

61
00:03:38,290 --> 00:03:44,292
and is able to brake and
accelerate when needed.

62
00:03:44,292 --> 00:03:49,252
[AUDIO OUT]

63
00:04:49,380 --> 00:04:52,010
There have been,
there are, and there

64
00:04:52,010 --> 00:04:54,660
will be a lot of
significant advances in AI.

65
00:04:54,660 --> 00:04:58,590
I think it's a golden age
for intelligent applications.

66
00:04:58,590 --> 00:05:00,660
You know, if people want
to make a lot of money

67
00:05:00,660 --> 00:05:04,020
with useful things,
that's the time.

68
00:05:04,020 --> 00:05:07,320
But this is kind of engineering.

69
00:05:07,320 --> 00:05:10,770
Interesting one,
but engineering.

70
00:05:10,770 --> 00:05:15,660
And we are still very far from
understanding how people can

71
00:05:15,660 --> 00:05:18,040
answer questions about images.

72
00:05:18,040 --> 00:05:26,250
This is one of the main
focus in the center, really.

73
00:05:26,250 --> 00:05:29,070
How does your brain
answer simple questions

74
00:05:29,070 --> 00:05:31,230
about this image?

75
00:05:31,230 --> 00:05:32,890
About what is there?

76
00:05:32,890 --> 00:05:34,680
And what is this?

77
00:05:34,680 --> 00:05:35,850
Who is this person?

78
00:05:38,400 --> 00:05:40,860
What is she doing?

79
00:05:40,860 --> 00:05:43,320
What is she thinking?

80
00:05:43,320 --> 00:05:47,100
Please tell me a story
about this, what's going on?

81
00:05:47,100 --> 00:05:49,600
[INAUDIBLE]

82
00:05:52,701 --> 00:05:57,030
And we would like to know to
have a system that does that.

83
00:05:57,030 --> 00:06:02,070
But also, to know how
our brain does it.

84
00:06:02,070 --> 00:06:04,290
So that's the science part.

85
00:06:04,290 --> 00:06:09,620
It's not enough to
pass the Turing Test.

86
00:06:09,620 --> 00:06:11,820
In this case, to have
a system that does it.

87
00:06:11,820 --> 00:06:15,480
We want to have a system
that does it in the same way

88
00:06:15,480 --> 00:06:17,300
as our brain does it.

89
00:06:17,300 --> 00:06:23,160
And we want to compare
your model, our system,

90
00:06:23,160 --> 00:06:28,440
with measurements on the brain
of people, or monkeys, also

91
00:06:28,440 --> 00:06:31,420
during the same task.

92
00:06:31,420 --> 00:06:37,890
So that's what we call
Turing plus, plus questions.

93
00:06:37,890 --> 00:06:40,650
And part of the
rationale about it

94
00:06:40,650 --> 00:06:43,680
is, this is kind of a more
philosophical discussion.

95
00:06:43,680 --> 00:06:45,330
I personally think
that it's very

96
00:06:45,330 --> 00:06:49,440
difficult to have a definition
of intelligence, in general.

97
00:06:49,440 --> 00:06:53,210
There are many different
forms of intelligence.

98
00:06:53,210 --> 00:06:57,300
What we can ask is
questions about,

99
00:06:57,300 --> 00:06:59,490
what is human intelligence?

100
00:06:59,490 --> 00:07:02,310
Because we can study that.

101
00:07:02,310 --> 00:07:02,810
Right.

102
00:07:06,280 --> 00:07:10,270
You know, it is, I don't
know, the ENIAC computers

103
00:07:10,270 --> 00:07:14,510
in the '50, more or less
intelligent than a person.

104
00:07:14,510 --> 00:07:17,450
You know, it can do
things a person cannot do.

105
00:07:17,450 --> 00:07:18,600
And so on.

106
00:07:18,600 --> 00:07:23,560
There are certain things ants
or bees do, are pretty amazing.

107
00:07:23,560 --> 00:07:24,700
Is this intelligence?

108
00:07:24,700 --> 00:07:27,870
Yeah, in a certain sense is.

109
00:07:27,870 --> 00:07:32,440
So I think, in terms of
a well-defined question,

110
00:07:32,440 --> 00:07:36,628
the real question is
about human intelligence.

111
00:07:36,628 --> 00:07:40,900
And so that's what, from
the scientific part,

112
00:07:40,900 --> 00:07:42,400
we are focused on.

113
00:07:42,400 --> 00:07:48,190
And would like to be
able to answer how

114
00:07:48,190 --> 00:07:51,310
people do understand images.

115
00:07:51,310 --> 00:07:52,600
We start with vision.

116
00:07:52,600 --> 00:07:54,790
We are not limited,
eventually, to vision.

117
00:07:54,790 --> 00:07:57,400
But in the first five
years of the center,

118
00:07:57,400 --> 00:07:59,406
that's the main focus.

119
00:07:59,406 --> 00:08:02,110
And answer the
question about images.

120
00:08:02,110 --> 00:08:08,500
And we want to understand
how the answers are produced

121
00:08:08,500 --> 00:08:12,580
by our brain at the
computational, psychophysical,

122
00:08:12,580 --> 00:08:15,240
and neural level.

123
00:08:15,240 --> 00:08:17,920
It's ambitious.

124
00:08:17,920 --> 00:08:21,040
And I think there are
probably, in terms

125
00:08:21,040 --> 00:08:24,500
of having all these
different levels, levels

126
00:08:24,500 --> 00:08:35,200
of really understanding from the
what, where, the neuroscience,

127
00:08:35,200 --> 00:08:36,669
to the behavior.

128
00:08:39,250 --> 00:08:41,950
We are not yet at
the point in which

129
00:08:41,950 --> 00:08:43,710
we can answer all
those kind of questions

130
00:08:43,710 --> 00:08:45,190
at all these different levels.

131
00:08:45,190 --> 00:08:47,080
But some, we are.

132
00:08:47,080 --> 00:08:50,150
One example is, who is there?

133
00:08:50,150 --> 00:08:53,380
It's essentially
face recognition.

134
00:08:53,380 --> 00:08:56,830
And this is an
interesting problem.

135
00:08:56,830 --> 00:09:02,380
Because we know from work,
originally in the monkeys,

136
00:09:02,380 --> 00:09:06,640
and then with fMRI in humans.

137
00:09:06,640 --> 00:09:09,700
Shown here, parts of
the brains of cortex,

138
00:09:09,700 --> 00:09:13,210
which are involved in
face recognition and face

139
00:09:13,210 --> 00:09:14,440
perception.

140
00:09:14,440 --> 00:09:19,870
And then, it's possible
to identify analog regions

141
00:09:19,870 --> 00:09:21,430
in the monkey.

142
00:09:21,430 --> 00:09:29,830
And record from the different
patches in the monkeys,

143
00:09:29,830 --> 00:09:35,590
each one probably around 100,000
neurons, maybe 200,000 or so.

144
00:09:35,590 --> 00:09:39,240
And look at their
properties when the monkeys

145
00:09:39,240 --> 00:09:41,950
is looking at the face.

146
00:09:41,950 --> 00:09:45,270
And make models of
what's going on.

147
00:09:45,270 --> 00:09:47,510
And, of course, we
want these models

148
00:09:47,510 --> 00:09:54,550
to respect the neural
data, ideally the MRI data.

149
00:09:54,550 --> 00:10:00,010
And do the job of recognizing
faces as well as human do.

150
00:10:00,010 --> 00:10:03,520
So we are getting there.

151
00:10:03,520 --> 00:10:06,080
I'm not saying we
have the answers,

152
00:10:06,080 --> 00:10:10,880
but we have at least
models that can be tested

153
00:10:10,880 --> 00:10:12,760
at all these different levels.

154
00:10:12,760 --> 00:10:15,520
So that's kind of
the ideal situation,

155
00:10:15,520 --> 00:10:18,800
from the point of view of what
we want to do in the center.

156
00:10:18,800 --> 00:10:20,710
Now as I said that,
not all problems

157
00:10:20,710 --> 00:10:24,040
are mature at this level.

158
00:10:24,040 --> 00:10:26,780
There are certain
like telling a story.

159
00:10:26,780 --> 00:10:28,060
We don't know exactly.

160
00:10:28,060 --> 00:10:31,330
We cannot record yet from
neurons in the monkey,

161
00:10:31,330 --> 00:10:33,610
when the monkey is
telling a story.

162
00:10:33,610 --> 00:10:37,180
Because the monkey has not been
able to tell its story, right.

163
00:10:37,180 --> 00:10:41,770
And so there are other
questions that are not

164
00:10:41,770 --> 00:10:44,210
as advanced as this one.

165
00:10:44,210 --> 00:10:49,960
But other type of studies can
be done on them, should be done.

166
00:10:49,960 --> 00:10:53,880
And this is what
we'll hear about.