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
00:00:22,212 --> 00:00:23,920
LAURA SCHULZ: So what
you've heard by now

9
00:00:23,920 --> 00:00:26,570
is that the hard problem
of cognitive science

10
00:00:26,570 --> 00:00:29,810
turns out to be the problem
of commonsense reasoning.

11
00:00:29,810 --> 00:00:31,710
Our computers can drive.

12
00:00:31,710 --> 00:00:34,880
They can perform
fabulous calculations.

13
00:00:34,880 --> 00:00:38,480
They can beat us at Jeopardy!

14
00:00:38,480 --> 00:00:41,900
But when it comes to all
of the really hard problems

15
00:00:41,900 --> 00:00:45,530
of cognitive science, there's
only one organism that solves

16
00:00:45,530 --> 00:00:47,267
them, and that's a human child.

17
00:00:47,267 --> 00:00:49,850
And those are problems of face
recognition, scene recognition,

18
00:00:49,850 --> 00:00:53,000
motor planning, natural language
acquisition, causal reasoning,

19
00:00:53,000 --> 00:00:54,770
theory of mind, moral reasoning.

20
00:00:54,770 --> 00:00:57,500
All of those are solved in
largely unsupervised learning

21
00:00:57,500 --> 00:01:00,320
by children by the age of five.

22
00:01:00,320 --> 00:01:02,930
And those are the things
our computers don't do well.

23
00:01:02,930 --> 00:01:05,900
And that is largely because the
problem of human intelligence,

24
00:01:05,900 --> 00:01:08,030
common sense
intelligence, is a problem

25
00:01:08,030 --> 00:01:10,550
of drawing really rich
inferences that are massively

26
00:01:10,550 --> 00:01:12,032
underdetermined by the data.

27
00:01:12,032 --> 00:01:13,490
So to make that
point really clear,

28
00:01:13,490 --> 00:01:16,580
I'm to give you all a pop
intelligence quiz, OK?

29
00:01:16,580 --> 00:01:18,440
So here's the pop
intelligence quiz.

30
00:01:18,440 --> 00:01:19,940
Can I see a show
of hands, please--

31
00:01:19,940 --> 00:01:23,210
how many of you think
that I have a spleen?

32
00:01:23,210 --> 00:01:25,770
Can I see a show of hands?

33
00:01:25,770 --> 00:01:27,770
Excellent.

34
00:01:27,770 --> 00:01:30,890
How many of you would care--

35
00:01:30,890 --> 00:01:32,770
keep your hands down
if you're an M.D.

36
00:01:32,770 --> 00:01:34,700
But other than that,
how many of you

37
00:01:34,700 --> 00:01:36,470
would care to come
up here and diagram,

38
00:01:36,470 --> 00:01:38,900
for the class, a spleen,
and explain what it is,

39
00:01:38,900 --> 00:01:42,810
where it is, and what its exact
function is in the human body?

40
00:01:42,810 --> 00:01:43,970
Anyone?

41
00:01:43,970 --> 00:01:46,320
How many of you
have met me before?

42
00:01:46,320 --> 00:01:49,060
A few of you.

43
00:01:49,060 --> 00:01:51,950
But most of you, without
knowing anything about spleens,

44
00:01:51,950 --> 00:01:53,490
and not knowing
anything about me,

45
00:01:53,490 --> 00:01:56,930
are nonetheless extremely
confident that I have one.

46
00:01:56,930 --> 00:01:58,970
So you have a lot of
abstract knowledge

47
00:01:58,970 --> 00:02:02,480
in the absence of
really much in the way

48
00:02:02,480 --> 00:02:04,791
of specific, concrete facts.

49
00:02:04,791 --> 00:02:06,290
OK, let me give you
another problem.

50
00:02:06,290 --> 00:02:08,810
Of course, it's a kind
of classic one from--

51
00:02:08,810 --> 00:02:12,424
by the way, for those of you who
are curious, that's a spleen.

52
00:02:12,424 --> 00:02:15,020
All right, classic--
these problems

53
00:02:15,020 --> 00:02:17,964
crop up in every aspect of
human cognition, all right?

54
00:02:17,964 --> 00:02:19,130
What's behind the rectangle?

55
00:02:19,130 --> 00:02:20,835
AUDIENCE: [INAUDIBLE]

56
00:02:21,335 --> 00:02:22,800
LAURA SCHULZ: Right.

57
00:02:22,800 --> 00:02:24,560
You all know.

58
00:02:24,560 --> 00:02:27,050
You can't articulate
it, but you know.

59
00:02:27,050 --> 00:02:29,450
And the fact that
there are infinitely

60
00:02:29,450 --> 00:02:32,690
many other hypotheses
consistent with the data

61
00:02:32,690 --> 00:02:35,390
doesn't trouble you at
all, doesn't stop you

62
00:02:35,390 --> 00:02:38,787
from converging, collectively,
on a single answer, which

63
00:02:38,787 --> 00:02:41,120
means there has to be a lot
of constraints on how you're

64
00:02:41,120 --> 00:02:42,589
interpreting this kind of data.

65
00:02:42,589 --> 00:02:43,380
Here's another one.

66
00:02:43,380 --> 00:02:45,106
Complete the sentence.

67
00:02:45,106 --> 00:02:46,420
AUDIENCE: Very long neck.

68
00:02:46,420 --> 00:02:47,770
LAURA SCHULZ: Very long neck--

69
00:02:47,770 --> 00:02:50,650
could be, or temper,
or flight to Kenya.

70
00:02:50,650 --> 00:02:52,540
There are many things
that it could be.

71
00:02:52,540 --> 00:02:54,546
In this case, it looks
like a frequency issue.

72
00:02:54,546 --> 00:02:56,170
Oh, well, "neck" is
a very common word.

73
00:02:56,170 --> 00:02:56,950
Others aren't.

74
00:02:56,950 --> 00:03:00,610
But if I said, "giraffes are
really common on the African--"

75
00:03:00,610 --> 00:03:03,760
you would say "savannah,"
not "television," OK?

76
00:03:03,760 --> 00:03:06,850
So you're using a lot
of rich information

77
00:03:06,850 --> 00:03:09,600
to take a tiny bit of data
and draw rich inferences.

78
00:03:09,600 --> 00:03:11,672
And that is the problem,
and the hard problem,

79
00:03:11,672 --> 00:03:13,130
of common sense
intelligence, and I

80
00:03:13,130 --> 00:03:15,760
think, a real dissociation
between what we do

81
00:03:15,760 --> 00:03:17,140
and what our children do.

82
00:03:17,140 --> 00:03:20,200
Human intelligence uses abstract
structured representations

83
00:03:20,200 --> 00:03:22,732
to constrain the
hypotheses and make

84
00:03:22,732 --> 00:03:24,190
these kind of
outrageous inferences

85
00:03:24,190 --> 00:03:25,606
that we shouldn't
be able to make.

86
00:03:25,606 --> 00:03:27,670
That's all well
and good, but where

87
00:03:27,670 --> 00:03:30,640
do these abstract structured
representations come from?

88
00:03:30,640 --> 00:03:32,140
And there's only
two possibilities--

89
00:03:32,140 --> 00:03:34,200
we're born with them,
or we learn them.

90
00:03:34,200 --> 00:03:37,390
Liz Spelke has already told
you a lot about the reasons

91
00:03:37,390 --> 00:03:39,440
to think that we are
born with many of them,

92
00:03:39,440 --> 00:03:41,350
and so true for
anything that might

93
00:03:41,350 --> 00:03:43,180
be stable over
evolutionary time, that

94
00:03:43,180 --> 00:03:47,231
might emerge early in ontogeny
and broadly in phylogeny.

95
00:03:47,231 --> 00:03:49,480
And that's true for many
aspects of folk physics, folk

96
00:03:49,480 --> 00:03:52,220
psychology, causal reasoning,
navigation, number.

97
00:03:54,840 --> 00:03:57,319
But there's a lot of
other things you know.

98
00:03:57,319 --> 00:03:59,610
Shoes, ships, sealing wax--
basically everything else--

99
00:03:59,610 --> 00:04:02,170
they are not plausibly innate.

100
00:04:02,170 --> 00:04:04,230
And so how do you learn them?

101
00:04:04,230 --> 00:04:06,750
So Piaget, the
founder of our field

102
00:04:06,750 --> 00:04:10,080
of developmental psychology,
says you build these up

103
00:04:10,080 --> 00:04:11,370
from experience.

104
00:04:11,370 --> 00:04:14,760
You build them up starting with
sensory motor representations

105
00:04:14,760 --> 00:04:16,140
very, very gradually.

106
00:04:16,140 --> 00:04:19,420
You progress through a lot
of concrete information

107
00:04:19,420 --> 00:04:21,450
until finally, somewhere
around the age of 12,

108
00:04:21,450 --> 00:04:25,450
you get to abstract
representations of the world.

109
00:04:25,450 --> 00:04:27,820
But it turns out that just
like you and your spleens,

110
00:04:27,820 --> 00:04:30,160
children have a lot
of abstract knowledge

111
00:04:30,160 --> 00:04:34,150
before they have much of
this concrete information.

112
00:04:34,150 --> 00:04:35,740
They know, actually,
almost nothing--

113
00:04:35,740 --> 00:04:37,364
even less than you
know, it turns out--

114
00:04:37,364 --> 00:04:38,650
about anatomy and biology.

115
00:04:38,650 --> 00:04:40,090
But they know all of
these kinds of things.

116
00:04:40,090 --> 00:04:41,080
Animals have insides.

117
00:04:41,080 --> 00:04:43,150
Similar kinds of animals
have similar insides.

118
00:04:43,150 --> 00:04:44,500
Plants and objects don't.

119
00:04:44,500 --> 00:04:47,050
Removing those insides
is a bad idea usually.

120
00:04:47,050 --> 00:04:50,530
They can go on and on without
really understanding anything

121
00:04:50,530 --> 00:04:52,790
about anatomy or biology.

122
00:04:52,790 --> 00:04:53,560
And so for you.

123
00:04:53,560 --> 00:04:55,730
If I push you hard on
all kinds of things--

124
00:04:55,730 --> 00:04:57,460
how does a scissors
really work--

125
00:04:57,460 --> 00:05:00,160
you know, most of you would
be like, ahh, and [INAUDIBLE]

126
00:05:00,160 --> 00:05:02,290
pretty quick, OK?

127
00:05:02,290 --> 00:05:03,340
And we know this.

128
00:05:03,340 --> 00:05:06,877
So we have these
intuitive theories

129
00:05:06,877 --> 00:05:08,710
that seem to constrain
our hypothesis space.

130
00:05:08,710 --> 00:05:10,570
But it's a really hard
chicken and egg problem,

131
00:05:10,570 --> 00:05:12,910
because we've said we need
these rich abstract theories

132
00:05:12,910 --> 00:05:14,770
to constrain the
interpretation of data.

133
00:05:14,770 --> 00:05:16,270
But how do we learn
them if we don't

134
00:05:16,270 --> 00:05:17,410
have concrete information?

135
00:05:17,410 --> 00:05:19,172
And nonetheless, some
of them are learned.

136
00:05:19,172 --> 00:05:21,380
We're going to return to
that at the end of the talk.

137
00:05:21,380 --> 00:05:23,740
And really, I'm going to do that
to set up Josh and Tomer, who

138
00:05:23,740 --> 00:05:25,070
are going to talk about that.

139
00:05:25,070 --> 00:05:26,695
But first, what I'm
going to talk about

140
00:05:26,695 --> 00:05:29,170
is-- you know, OK, why
am I talking about this

141
00:05:29,170 --> 00:05:30,190
as an intuitive theory?

142
00:05:30,190 --> 00:05:32,200
Where is this
argument coming from?

143
00:05:32,200 --> 00:05:33,540
I'm interested in learning.

144
00:05:33,540 --> 00:05:35,350
And it's a research
program that emerged

145
00:05:35,350 --> 00:05:37,870
against the backdrop
of two revolutions

146
00:05:37,870 --> 00:05:39,970
in our understanding of
cognitive development.

147
00:05:39,970 --> 00:05:41,320
One was the infancy revolution.

148
00:05:41,320 --> 00:05:43,900
You've heard a lot about it, so
I'm going to go very quickly.

149
00:05:43,900 --> 00:05:45,441
Babies, it turns
out, know a lot more

150
00:05:45,441 --> 00:05:48,520
than we thought about
objects, and about number,

151
00:05:48,520 --> 00:05:52,600
and about agents, and their
goals, and their intentions.

152
00:05:52,600 --> 00:05:54,400
It's not just infants though.

153
00:05:54,400 --> 00:05:59,560
It turns out that very young
children, preschoolers, also

154
00:05:59,560 --> 00:06:02,440
represent knowledge
that is not plausibly

155
00:06:02,440 --> 00:06:07,780
innate in ways that are
abstract, that are coherent,

156
00:06:07,780 --> 00:06:11,560
that are causal, that support
prediction, and intervention,

157
00:06:11,560 --> 00:06:14,890
and explanation, and
counterfactual reasoning

158
00:06:14,890 --> 00:06:17,290
in ways that seem
to justify referring

159
00:06:17,290 --> 00:06:21,100
to them as intuitive theories.

160
00:06:21,100 --> 00:06:22,960
And together, these
two revolutions,

161
00:06:22,960 --> 00:06:25,134
the infancy revolution
and the revolution

162
00:06:25,134 --> 00:06:26,800
in our understanding
of early childhood,

163
00:06:26,800 --> 00:06:29,330
dismantled Piagetian
stage theory.

164
00:06:29,330 --> 00:06:31,220
There was never a time--

165
00:06:31,220 --> 00:06:32,980
there is never a
time-- in development

166
00:06:32,980 --> 00:06:36,490
when babies are only
sensory motor learners.

167
00:06:36,490 --> 00:06:39,040
There is never a time
when there is not

168
00:06:39,040 --> 00:06:43,780
some level of abstract
representation going on.

169
00:06:43,780 --> 00:06:47,380
But fundamentally, neither
of these revolutions

170
00:06:47,380 --> 00:06:49,267
was about learning per se.

171
00:06:49,267 --> 00:06:51,100
And I think I can make
that point most clear

172
00:06:51,100 --> 00:06:55,420
by pointing to a popular book
that came out at the time.

173
00:06:55,420 --> 00:06:58,669
The subtitle here is Minds,
Brains, and How Children Learn.

174
00:06:58,669 --> 00:07:00,210
But in fact, if you
look in the book,

175
00:07:00,210 --> 00:07:01,418
this was a publisher's title.

176
00:07:01,418 --> 00:07:02,640
There's nothing about brains.

177
00:07:02,640 --> 00:07:04,390
And there's very
little about learning.

178
00:07:04,390 --> 00:07:07,660
The titles are what children
know about objects, what

179
00:07:07,660 --> 00:07:10,080
children know about agents.

180
00:07:10,080 --> 00:07:12,850
And I can say this with great
reverence and some authority,

181
00:07:12,850 --> 00:07:15,160
because the first author's
my thesis advisor.

182
00:07:15,160 --> 00:07:16,810
And the book came
out in 1999, which

183
00:07:16,810 --> 00:07:18,850
is the year I started
graduate school.

184
00:07:18,850 --> 00:07:23,760
So literally and metaphorically,
this is where I began.

185
00:07:23,760 --> 00:07:27,010
I began with this metaphor
of the child as scientist.

186
00:07:27,010 --> 00:07:29,020
And it's a really
problematic metaphor,

187
00:07:29,020 --> 00:07:32,440
because science is a
historically and culturally

188
00:07:32,440 --> 00:07:35,290
specific practice
that is practiced

189
00:07:35,290 --> 00:07:37,600
by a tiny minority
of the human species

190
00:07:37,600 --> 00:07:40,000
and is difficult
even for us, right?

191
00:07:40,000 --> 00:07:42,190
So it seems a really
odd place to look

192
00:07:42,190 --> 00:07:46,750
for a universal metaphor
for cognitive development.

193
00:07:46,750 --> 00:07:49,610
But science has this
peculiar property,

194
00:07:49,610 --> 00:07:52,420
which is that it
gets the world right.

195
00:07:52,420 --> 00:07:55,570
And if you really want to
understand how new knowledge is

196
00:07:55,570 --> 00:07:58,480
possible, how you could
get the world right,

197
00:07:58,480 --> 00:08:01,990
you might want to understand
how scientists do it,

198
00:08:01,990 --> 00:08:04,420
what kinds of
epistemic practices

199
00:08:04,420 --> 00:08:07,390
might support learning
and discovery.

200
00:08:07,390 --> 00:08:10,771
And the answer to that is both
that we do and do not know,

201
00:08:10,771 --> 00:08:12,520
which is to say we can
say a lot of things

202
00:08:12,520 --> 00:08:13,720
about what scientists do.

203
00:08:13,720 --> 00:08:15,900
Here are some of them.

204
00:08:15,900 --> 00:08:17,410
They'll all be familiar to you.

205
00:08:17,410 --> 00:08:19,960
They would be familiar to
you if you were a physicist,

206
00:08:19,960 --> 00:08:22,690
if you were in aero-astro,
if you were in paleontology.

207
00:08:22,690 --> 00:08:24,700
These are the kinds of
scientific practices

208
00:08:24,700 --> 00:08:28,150
that cut across content
domains and arguably define

209
00:08:28,150 --> 00:08:30,947
what science is,
which is to say,

210
00:08:30,947 --> 00:08:33,280
if you did all of these things,
you couldn't necessarily

211
00:08:33,280 --> 00:08:34,210
do science.

212
00:08:34,210 --> 00:08:36,940
Science requires bringing
these inferential processes

213
00:08:36,940 --> 00:08:40,000
to bear on really
rich, specific,

214
00:08:40,000 --> 00:08:44,310
conceptual representations of
individual content domains.

215
00:08:44,310 --> 00:08:47,562
But arguably, if you had all
of that rich, specific content

216
00:08:47,562 --> 00:08:49,270
knowledge and you
didn't do these things,

217
00:08:49,270 --> 00:08:51,040
you couldn't learn
anything at all.

218
00:08:51,040 --> 00:08:53,380
These are the kinds
of epistemic practices

219
00:08:53,380 --> 00:08:56,220
that seem fundamental to
inquiry and discovery.

220
00:08:56,220 --> 00:08:58,480
And the argument that I've
made in my research program

221
00:08:58,480 --> 00:09:00,521
is they are fundamental
not just in science, they

222
00:09:00,521 --> 00:09:02,980
are fundamental in
cognitive development.

223
00:09:02,980 --> 00:09:07,030
These are the processes that
support learning and discovery.

224
00:09:07,030 --> 00:09:10,780
And there's good evidence that
each and every one of them

225
00:09:10,780 --> 00:09:15,190
emerges in some form in the
first few years of life.

226
00:09:15,190 --> 00:09:17,800
And that is because these are
the only rational processes

227
00:09:17,800 --> 00:09:22,090
we know of that can solve the
hard problem of learning, which

228
00:09:22,090 --> 00:09:25,150
is exactly the
problem of how to draw

229
00:09:25,150 --> 00:09:28,900
rich abstract entrances
rapidly and accurately

230
00:09:28,900 --> 00:09:31,720
from sparse, noisy data.

231
00:09:31,720 --> 00:09:34,570
I said we could characterize
these practices both formally

232
00:09:34,570 --> 00:09:36,380
and informally.

233
00:09:36,380 --> 00:09:38,115
And indeed, as you've
heard, I think,

234
00:09:38,115 --> 00:09:40,240
from some of our computational
modeling colleagues,

235
00:09:40,240 --> 00:09:42,167
for each and every one
of these practices,

236
00:09:42,167 --> 00:09:43,750
we can begin to
characterize something

237
00:09:43,750 --> 00:09:47,290
about what it means
to distinguish

238
00:09:47,290 --> 00:09:50,110
genuine causes from
spurious associations

239
00:09:50,110 --> 00:09:54,160
or to optimize information gain.

240
00:09:54,160 --> 00:09:57,300
But with all due respect to
my computational modeling

241
00:09:57,300 --> 00:10:01,270
colleagues, and much as we
want really simple models--

242
00:10:01,270 --> 00:10:04,930
Hebb's rule,
Rescorla-Wagner, Bayes' law--

243
00:10:04,930 --> 00:10:06,820
that would capture
it, none of these

244
00:10:06,820 --> 00:10:09,100
do justice to what
children can do.

245
00:10:09,100 --> 00:10:12,340
Because children can
do all of these things.

246
00:10:12,340 --> 00:10:15,190
And we don't yet have
a full formal theory

247
00:10:15,190 --> 00:10:18,160
of hypothesis, generation,
inquiry, and discovery.

248
00:10:18,160 --> 00:10:21,322
That remains a hard problem
of cognitive science.

249
00:10:21,322 --> 00:10:22,780
But it's a problem
to which I think

250
00:10:22,780 --> 00:10:24,580
our theories should aspire.

251
00:10:24,580 --> 00:10:26,920
Because there is
good empirical data

252
00:10:26,920 --> 00:10:29,440
that this is the kind of
learning that humans, including

253
00:10:29,440 --> 00:10:33,740
even very young
children, engage in.

254
00:10:33,740 --> 00:10:37,650
So normally, at this point,
what I would do is I'd say,

255
00:10:37,650 --> 00:10:39,650
and I'm going to show you
a few examples of this

256
00:10:39,650 --> 00:10:41,355
from my research program.

257
00:10:41,355 --> 00:10:43,730
But the talk I'm giving here
is a sort of funny throwback

258
00:10:43,730 --> 00:10:44,670
talk in some ways.

259
00:10:44,670 --> 00:10:45,920
What I was asked
to talk about today

260
00:10:45,920 --> 00:10:47,044
was the child as scientist.

261
00:10:47,044 --> 00:10:50,450
And that was a research program
from a few, few years ago.

262
00:10:50,450 --> 00:10:52,700
And you know, at that point,
I was a junior professor.

263
00:10:52,700 --> 00:10:54,230
And you know, when you're
a junior professor,

264
00:10:54,230 --> 00:10:56,230
like when you're a graduate
student or post-doc,

265
00:10:56,230 --> 00:10:59,140
it's all idealistic science
for the sake of knowledge,

266
00:10:59,140 --> 00:10:59,960
pure science.

267
00:10:59,960 --> 00:11:02,000
And then you get tenure.

268
00:11:02,000 --> 00:11:05,210
And it's all grants, and
money, and administration,

269
00:11:05,210 --> 00:11:07,370
and allocation of resources.

270
00:11:07,370 --> 00:11:10,340
And in the years
since, I have started

271
00:11:10,340 --> 00:11:15,140
to think not just about the
pure science of learning,

272
00:11:15,140 --> 00:11:18,500
but about the cost associated
with gaining information

273
00:11:18,500 --> 00:11:20,820
and the trade-offs between
those costs and rewards.

274
00:11:20,820 --> 00:11:24,680
So I'm going to take these same
practices now, and situate them

275
00:11:24,680 --> 00:11:27,890
in a world, a real world, that
has certain kinds of trade-offs

276
00:11:27,890 --> 00:11:30,380
in how you think about
information and talk,

277
00:11:30,380 --> 00:11:32,210
along with the
child as scientist,

278
00:11:32,210 --> 00:11:35,780
about what those costs
do and how they are also,

279
00:11:35,780 --> 00:11:38,386
in themselves, a source of
information about the world.

280
00:11:38,386 --> 00:11:40,010
So I'm going to talk
about this as sort

281
00:11:40,010 --> 00:11:42,200
of inferential economics.

282
00:11:42,200 --> 00:11:43,820
Children know
information is valuable.

283
00:11:43,820 --> 00:11:45,620
I'm going to show you
a couple of examples

284
00:11:45,620 --> 00:11:47,630
of how they reason about it.

285
00:11:47,630 --> 00:11:50,120
And children selectively
explore in ways

286
00:11:50,120 --> 00:11:51,371
that support information gain.

287
00:11:51,371 --> 00:11:52,994
So I'm going to show
you some old work,

288
00:11:52,994 --> 00:11:55,100
but I'm also going to
throw in a few new studies,

289
00:11:55,100 --> 00:11:57,070
because I can't resist.

290
00:11:57,070 --> 00:11:59,310
But information is also costly.

291
00:11:59,310 --> 00:12:02,120
And the costs themselves
are informative in a variety

292
00:12:02,120 --> 00:12:02,830
of ways.

293
00:12:02,830 --> 00:12:05,413
I'm not necessarily going to get
through all of these studies,

294
00:12:05,413 --> 00:12:07,620
although I might try to.

295
00:12:07,620 --> 00:12:10,970
But I want to give you a kind
of feel for the kinds of things

296
00:12:10,970 --> 00:12:12,180
children can do.

297
00:12:12,180 --> 00:12:14,690
All right, so let's
start by talking

298
00:12:14,690 --> 00:12:15,899
about a really basic problem.

299
00:12:15,899 --> 00:12:17,523
It's a problem that's
basic to science,

300
00:12:17,523 --> 00:12:20,060
but it's also a problem basic
to human learning, which is

301
00:12:20,060 --> 00:12:21,330
the problem of generalization.

302
00:12:21,330 --> 00:12:23,810
How do you generalize
from sparse data?

303
00:12:23,810 --> 00:12:25,732
In science, we do
this all the time.

304
00:12:25,732 --> 00:12:26,690
We have a small sample.

305
00:12:26,690 --> 00:12:29,990
We want to make a claim about
the population as a whole.

306
00:12:29,990 --> 00:12:33,820
And of course, we can use
feature similarity and category

307
00:12:33,820 --> 00:12:36,230
membership to say things
that look the same

308
00:12:36,230 --> 00:12:38,960
or belong to the same kind are
likely to share properties.

309
00:12:38,960 --> 00:12:40,370
So if some of
these Martian rocks

310
00:12:40,370 --> 00:12:42,860
have high concentrations of
silica, maybe they all do.

311
00:12:42,860 --> 00:12:46,130
If some of these needles
on Pacific silver fir trees

312
00:12:46,130 --> 00:12:49,230
grow flat on the branch,
maybe they all do.

313
00:12:52,270 --> 00:12:54,520
But in science, we can also
do something a little more

314
00:12:54,520 --> 00:12:56,140
fussy and suspicious.

315
00:12:56,140 --> 00:12:58,600
We can say, well,
you know, it kind of

316
00:12:58,600 --> 00:13:02,155
depends on how you got that
sample of evidence, right?

317
00:13:02,155 --> 00:13:04,612
If you randomly sampled from
the population, yeah, sure,

318
00:13:04,612 --> 00:13:06,070
the properties,
you can generalize.

319
00:13:06,070 --> 00:13:08,620
But if you cherry-picked
that data in some way,

320
00:13:08,620 --> 00:13:12,310
maybe the sample isn't going
to generalize quite as broadly.

321
00:13:12,310 --> 00:13:15,590
So do all Martian rocks have
high concentrations of silica

322
00:13:15,590 --> 00:13:17,620
or just the dusty
ones on the surface?

323
00:13:17,620 --> 00:13:20,620
Do all Pacific silver fir
needles lie flat or just

324
00:13:20,620 --> 00:13:22,870
those low on the canopy, right?

325
00:13:22,870 --> 00:13:24,950
These are the ones that
are easy to sample from.

326
00:13:24,950 --> 00:13:28,677
How do I know how
generalizable the property is?

327
00:13:28,677 --> 00:13:30,760
And if I think that you
cherry-picked your sample,

328
00:13:30,760 --> 00:13:33,970
I might constrain my inferences
only to things near the ground.

329
00:13:33,970 --> 00:13:35,470
So how far you're
going to extend

330
00:13:35,470 --> 00:13:37,780
a generalization in
science depends on

331
00:13:37,780 --> 00:13:39,940
whether you think that
the sampling process was

332
00:13:39,940 --> 00:13:41,110
random or selective.

333
00:13:41,110 --> 00:13:43,360
And we wanted to know whether
this was true for babies

334
00:13:43,360 --> 00:13:44,730
as well.

335
00:13:44,730 --> 00:13:45,820
So this is how we asked.

336
00:13:45,820 --> 00:13:47,950
We showed babies a
population, in this case,

337
00:13:47,950 --> 00:13:50,350
of blue and yellow dog toys.

338
00:13:50,350 --> 00:13:51,330
They're in a box.

339
00:13:51,330 --> 00:13:53,920
The box is transparent, has
a false front so it stays

340
00:13:53,920 --> 00:13:56,990
a less stable representation of
what looks like a lot of balls.

341
00:13:56,990 --> 00:13:58,900
And we're going to
reach into that box.

342
00:13:58,900 --> 00:14:00,739
And we're going to pull out--

343
00:14:00,739 --> 00:14:03,280
there are many more blue balls
in this box than yellow balls.

344
00:14:03,280 --> 00:14:05,620
We're going to pull out three
blue balls one at a time

345
00:14:05,620 --> 00:14:06,880
and squeak them--

346
00:14:06,880 --> 00:14:08,896
and squeeze them-- and
they're going to squeak.

347
00:14:08,896 --> 00:14:11,020
And then we're going to
hand the baby a yellow ball

348
00:14:11,020 --> 00:14:11,980
from the same box.

349
00:14:11,980 --> 00:14:14,170
And the question is, does
the baby squeeze the ball

350
00:14:14,170 --> 00:14:16,335
and expect it to squeak?

351
00:14:16,335 --> 00:14:18,460
Well, there's nothing very
suspicious about pulling

352
00:14:18,460 --> 00:14:22,180
three blue balls from a
box of mostly blue balls.

353
00:14:22,180 --> 00:14:25,160
And this has a lot of
feature properties in common.

354
00:14:25,160 --> 00:14:26,729
So it looks like the others.

355
00:14:26,729 --> 00:14:28,520
We predict that children
should generalize.

356
00:14:28,520 --> 00:14:31,330
They should try squeezing this
ball and should squeeze often.

357
00:14:31,330 --> 00:14:32,704
And the question
is, what happens

358
00:14:32,704 --> 00:14:36,490
if you do exactly the same
sample in exactly the same way

359
00:14:36,490 --> 00:14:39,054
from a different population?

360
00:14:39,054 --> 00:14:41,470
Now, it is very unlikely that
you sampled three blue balls

361
00:14:41,470 --> 00:14:44,140
from a population of
mostly yellow balls.

362
00:14:44,140 --> 00:14:45,610
In this case, it's
much more likely

363
00:14:45,610 --> 00:14:47,800
that you were
sampling selectively.

364
00:14:47,800 --> 00:14:51,350
So maybe only the blue
balls had the property.

365
00:14:51,350 --> 00:14:51,850
Yeah?

366
00:14:51,850 --> 00:14:52,920
AUDIENCE: And they can
see the population?

367
00:14:52,920 --> 00:14:55,419
LAURA SCHULZ: They can see the
population-- transparent box,

368
00:14:55,419 --> 00:14:56,670
transparent front.

369
00:14:56,670 --> 00:15:00,160
So they can see the population.

370
00:15:00,160 --> 00:15:01,670
And if children
understand that it's

371
00:15:01,670 --> 00:15:04,210
not just about the
property similarity

372
00:15:04,210 --> 00:15:07,120
but something about how that
evidence was generated, then,

373
00:15:07,120 --> 00:15:09,125
in this case, children
should say, well look,

374
00:15:09,125 --> 00:15:11,500
you just looked like you were
cherry-picking your sample.

375
00:15:11,500 --> 00:15:13,270
Maybe it doesn't
generalize predictions

376
00:15:13,270 --> 00:15:15,340
that fewer children
should try squeaking

377
00:15:15,340 --> 00:15:17,234
and children should
squeeze less often.

378
00:15:17,234 --> 00:15:19,150
So I'm going to show you
what this looks like.

379
00:15:19,150 --> 00:15:21,680
By the way, the yellow one has
that funny thing at the end

380
00:15:21,680 --> 00:15:24,230
so that children could do
something else with the ball,

381
00:15:24,230 --> 00:15:24,730
right?

382
00:15:24,730 --> 00:15:26,170
So they can bang it,
or throw it around,

383
00:15:26,170 --> 00:15:27,290
or other things like that.

384
00:15:27,290 --> 00:15:28,961
So let me show you
what it looks like.

385
00:15:28,961 --> 00:15:31,210
Kids are always going to see
three squeaky blue balls.

386
00:15:31,210 --> 00:15:32,918
They're always going
to get a yellow one.

387
00:15:32,918 --> 00:15:35,200
But you'll see that they do
different things depending

388
00:15:35,200 --> 00:15:36,010
on whether they think--

389
00:15:36,010 --> 00:15:36,370
[VIDEO PLAYBACK]

390
00:15:36,370 --> 00:15:37,930
LAURA SCHULZ: --the evidence
was randomly sampled

391
00:15:37,930 --> 00:15:39,460
and possibly
generalizable or not.

392
00:15:46,404 --> 00:15:48,740
So the child at the
top is squeezing,

393
00:15:48,740 --> 00:16:04,992
and squeezing, and
squeezing, and squeezing.

394
00:16:12,864 --> 00:16:14,360
[END PLAYBACK]

395
00:16:14,360 --> 00:16:16,970
So these were-- this was
true both of the mean number

396
00:16:16,970 --> 00:16:19,077
of squeezes and the number
of individual children

397
00:16:19,077 --> 00:16:19,910
who squeezed at all.

398
00:16:19,910 --> 00:16:21,760
I'm just going to show you
the mean number of squeezes.

399
00:16:21,760 --> 00:16:23,843
But what you'll see is
that children are much more

400
00:16:23,843 --> 00:16:26,000
likely to squeeze, and
squeeze persistently,

401
00:16:26,000 --> 00:16:27,830
in the condition where
the evidence looks

402
00:16:27,830 --> 00:16:31,202
like it was randomly sampled
than selectively sampled.

403
00:16:31,202 --> 00:16:32,660
But what you could
worry about here

404
00:16:32,660 --> 00:16:34,130
is children are
sensitive to something

405
00:16:34,130 --> 00:16:36,230
about the relationship of the
sample and the population.

406
00:16:36,230 --> 00:16:37,700
But maybe they will
just generalize

407
00:16:37,700 --> 00:16:42,770
from a majority object to a
minority, but not the reverse.

408
00:16:42,770 --> 00:16:44,990
Maybe they won't generalize
from a minority object

409
00:16:44,990 --> 00:16:46,107
to the majority.

410
00:16:46,107 --> 00:16:48,440
So they don't really care
about whether the evidence was

411
00:16:48,440 --> 00:16:50,120
randomly sampled
or not, they just

412
00:16:50,120 --> 00:16:53,060
care about that aspect of
the sample and population.

413
00:16:53,060 --> 00:16:55,730
So we ran a replication of
the yellow ball condition.

414
00:16:55,730 --> 00:16:58,520
Again, we're going to pull an
unlikely sample from that box,

415
00:16:58,520 --> 00:16:59,900
three blue balls in a row.

416
00:16:59,900 --> 00:17:02,420
And we're going to compare
it with a sample that's

417
00:17:02,420 --> 00:17:03,470
not that improbable.

418
00:17:03,470 --> 00:17:06,579
You could easily randomly
sample just one blue ball

419
00:17:06,579 --> 00:17:07,710
from the box.

420
00:17:07,710 --> 00:17:10,369
So in this case,
children are going

421
00:17:10,369 --> 00:17:12,038
to see much less
squeezing, right?

422
00:17:12,038 --> 00:17:14,079
They're only going to see
one blue ball squeezed.

423
00:17:14,079 --> 00:17:16,020
And we squeeze it both
once and three times,

424
00:17:16,020 --> 00:17:18,944
but it's only one blue ball
in two different conditions.

425
00:17:18,944 --> 00:17:21,319
And the prediction there is
that even though the children

426
00:17:21,319 --> 00:17:23,169
themselves are seeing
much less squeezing,

427
00:17:23,169 --> 00:17:25,460
they should say, well, that's
not an improbable sample.

428
00:17:25,460 --> 00:17:28,940
And they, themselves,
should squeeze more.

429
00:17:28,940 --> 00:17:34,040
And that's exactly what we found
in both of those conditions.

430
00:17:34,040 --> 00:17:35,220
It's graded, by the way.

431
00:17:35,220 --> 00:17:36,440
If you do two balls,
they're intermediate.

432
00:17:36,440 --> 00:17:37,464
And if it's a model--

433
00:17:37,464 --> 00:17:39,380
well, not going to talk
about-- but yeah, what

434
00:17:39,380 --> 00:17:42,227
happens if you just pour them
upside down and drop them?

435
00:17:42,227 --> 00:17:43,810
Now this is a really
improbable sample

436
00:17:43,810 --> 00:17:45,601
I said, three blue
balls from a yellow box.

437
00:17:45,601 --> 00:17:47,360
But you've just given
positive evidence

438
00:17:47,360 --> 00:17:49,340
that you shook the ball, and
it just happened to fall out.

439
00:17:49,340 --> 00:17:50,881
And they don't know
we're MIT, and we

440
00:17:50,881 --> 00:17:54,460
can do sneaky technological
things like have a trap door.

441
00:17:54,460 --> 00:17:56,510
So in this case, it's
an improbable sample,

442
00:17:56,510 --> 00:17:57,920
but it was randomly generated.

443
00:17:57,920 --> 00:17:59,660
And the prediction
is, in this case,

444
00:17:59,660 --> 00:18:01,960
the babies themselves
should squeeze more.

445
00:18:01,960 --> 00:18:04,460
Because as I say, if you can
pour out any balls that squeak,

446
00:18:04,460 --> 00:18:06,080
probably everything squeaks.

447
00:18:06,080 --> 00:18:07,137
And indeed, they do.

448
00:18:07,137 --> 00:18:08,420
Indeed, they do.

449
00:18:08,420 --> 00:18:09,260
All right?

450
00:18:09,260 --> 00:18:12,950
So 15-month-old
babies' generalizations

451
00:18:12,950 --> 00:18:15,230
take into account more
than category membership

452
00:18:15,230 --> 00:18:17,180
and the perceptual
similarity of objects.

453
00:18:17,180 --> 00:18:19,160
They make graded inferences
that are sensitive

454
00:18:19,160 --> 00:18:21,410
both to the amount of
evidence they observe

455
00:18:21,410 --> 00:18:24,140
and to the process by which
that evidence is sampled.

456
00:18:24,140 --> 00:18:26,081
Is that clear?

457
00:18:26,081 --> 00:18:27,830
All right, let me show
you another example

458
00:18:27,830 --> 00:18:29,110
of sort of child as scientist.

459
00:18:29,110 --> 00:18:31,610
It's going to start with a
hard problem of confounding

460
00:18:31,610 --> 00:18:33,985
that we all have, which is
that we are part of the world.

461
00:18:33,985 --> 00:18:35,760
So in one-offs, when
things go wrong,

462
00:18:35,760 --> 00:18:37,580
we may not know if
we were responsible

463
00:18:37,580 --> 00:18:39,110
or the world was
responsible, right?

464
00:18:39,110 --> 00:18:42,280
This is a chronic problem
in relationships, right--

465
00:18:42,280 --> 00:18:43,730
you or me?

466
00:18:43,730 --> 00:18:45,710
So it's a hard problem
of confounding,

467
00:18:45,710 --> 00:18:47,842
and you might need some
data to disambiguate it.

468
00:18:47,842 --> 00:18:50,300
So here we're going to give
babies a case where they cannot

469
00:18:50,300 --> 00:18:51,630
do something.

470
00:18:51,630 --> 00:18:53,570
And the question
is, can we give them

471
00:18:53,570 --> 00:18:57,332
a little bit of data to
unconfound that problem

472
00:18:57,332 --> 00:18:59,540
and convince them either
that the problem is probably

473
00:18:59,540 --> 00:19:02,219
with the toy or the
problem is with themselves?

474
00:19:02,219 --> 00:19:04,510
And the argument is, if they
think that it's themselves

475
00:19:04,510 --> 00:19:05,343
that's the problem--

476
00:19:05,343 --> 00:19:07,670
it's the agent state and
not the environment state--

477
00:19:07,670 --> 00:19:11,730
they should hold the toy
constant and change the agent.

478
00:19:11,730 --> 00:19:13,760
But if they think it's
the toy, then they

479
00:19:13,760 --> 00:19:15,830
should just go ahead and
reach for another toy.

480
00:19:15,830 --> 00:19:17,746
So in both cases, they're
going to have access

481
00:19:17,746 --> 00:19:19,560
to another person, their mom.

482
00:19:19,560 --> 00:19:21,980
And they're going to have
access to another toy.

483
00:19:21,980 --> 00:19:25,070
And so the question
is, what do they

484
00:19:25,070 --> 00:19:29,280
do if we give them
minimal statistical data

485
00:19:29,280 --> 00:19:31,880
to disambiguate these, OK?

486
00:19:31,880 --> 00:19:34,430
So this is the setup.

487
00:19:34,430 --> 00:19:37,340
We're going to show
the babies two agents.

488
00:19:37,340 --> 00:19:43,220
In one condition, I am going
to succeed one time at making

489
00:19:43,220 --> 00:19:44,805
that toy go and fail one time.

490
00:19:45,305 --> 00:19:48,770
And Hyowon Gweon, my
collaborator on this project,

491
00:19:48,770 --> 00:19:50,490
is also going to fail
and succeed once.

492
00:19:50,490 --> 00:19:53,419
So this looks like this
toy has maybe faulty wiring

493
00:19:53,419 --> 00:19:53,960
or something.

494
00:19:53,960 --> 00:19:55,793
It works some of the
time, not all the time.

495
00:19:55,793 --> 00:19:57,480
It's just not a great toy.

496
00:19:57,480 --> 00:19:59,240
The babies can have another toy.

497
00:19:59,240 --> 00:20:00,710
The parents are
going to be there.

498
00:20:00,710 --> 00:20:03,420
If they think it's the toy,
they should change the object.

499
00:20:03,420 --> 00:20:07,410
In the other condition, Hyowon
is always going to succeed,

500
00:20:07,410 --> 00:20:11,050
which is generally
true in my experience.

501
00:20:11,050 --> 00:20:13,490
And as is always true of my
experience in technology,

502
00:20:13,490 --> 00:20:14,690
I am always going to fail.

503
00:20:14,690 --> 00:20:17,030
And in this case,
the children should

504
00:20:17,030 --> 00:20:20,000
conclude that there's something
wrong with the person.

505
00:20:20,000 --> 00:20:22,850
And if that's the case, they
should hold the object constant

506
00:20:22,850 --> 00:20:23,735
and change the agent.

507
00:20:23,735 --> 00:20:24,860
This is what it looks like.

508
00:20:24,860 --> 00:20:27,245
[VIDEO PLAYBACK]

509
00:20:27,245 --> 00:20:32,969
- [INAUDIBLE]

510
00:20:32,969 --> 00:20:34,260
LAURA SCHULZ: We've lost sound.

511
00:20:41,700 --> 00:20:45,210
Well there's an audio
here, but [INAUDIBLE] we

512
00:20:45,210 --> 00:20:47,434
are going to want that later.

513
00:20:47,434 --> 00:20:49,469
- Cool What happened to my toys?

514
00:20:49,469 --> 00:20:51,510
LAURA SCHULZ: In any case,
were just showing them

515
00:20:51,510 --> 00:20:52,470
the data at this point.

516
00:20:52,470 --> 00:20:55,120
And the babies are going to
get the toy in each condition.

517
00:20:55,120 --> 00:20:56,334
One toy is on the mat.

518
00:20:56,334 --> 00:20:58,500
By the way, there are lots
of individual differences

519
00:20:58,500 --> 00:21:01,620
in any two sets of clips
between the exact positioning

520
00:21:01,620 --> 00:21:03,810
of the parent, the
child, the toy.

521
00:21:03,810 --> 00:21:06,750
All of these were
coded blind conditions

522
00:21:06,750 --> 00:21:08,520
for all of these other
variables to make

523
00:21:08,520 --> 00:21:10,870
sure those were evenly
matched across conditions,

524
00:21:10,870 --> 00:21:11,820
and they were.

525
00:21:11,820 --> 00:21:16,120
So what you're going
to see now though is

526
00:21:16,120 --> 00:21:18,950
that in the condition where
the babies think it is probably

527
00:21:18,950 --> 00:21:23,730
the toy, they engage in
a very different behavior

528
00:21:23,730 --> 00:21:30,190
than when they think
it is probably agent.

529
00:21:30,190 --> 00:21:32,615
- [CRYING]

530
00:21:34,070 --> 00:21:36,980
- [INAUDIBLE]

531
00:21:41,830 --> 00:21:42,756
[END PLAYBACK]

532
00:21:42,756 --> 00:21:44,130
And that's what
we found overall.

533
00:21:44,130 --> 00:21:46,410
The distribution, overall,
of children's tendency

534
00:21:46,410 --> 00:21:48,190
to perform one
action versus another

535
00:21:48,190 --> 00:21:51,420
differed across conditions
depending on the pattern

536
00:21:51,420 --> 00:21:53,140
of data that they observed.

537
00:21:56,900 --> 00:21:59,900
So 16-month-olds track
the statistical dependence

538
00:21:59,900 --> 00:22:01,490
between agents,
objects, and outcomes.

539
00:22:01,490 --> 00:22:04,440
They can use minimal data
to make attributions.

540
00:22:04,440 --> 00:22:06,830
And they help them choose
between seeking help

541
00:22:06,830 --> 00:22:10,920
from others or
exploring on their own.

542
00:22:10,920 --> 00:22:12,624
Clear?

543
00:22:12,624 --> 00:22:15,560
OK.

544
00:22:15,560 --> 00:22:17,900
I've just shown you kids'
sensitivity to the data

545
00:22:17,900 --> 00:22:18,680
that they see.

546
00:22:18,680 --> 00:22:20,990
But of course, data isn't
handed out all the time.

547
00:22:20,990 --> 00:22:22,950
At least disambiguating
data isn't handed out.

548
00:22:22,950 --> 00:22:24,964
One of the really
important, hard things

549
00:22:24,964 --> 00:22:26,630
that you have to do
if you want to learn

550
00:22:26,630 --> 00:22:28,940
is sometimes actually
figure out what data would

551
00:22:28,940 --> 00:22:31,320
be informative and go get it.

552
00:22:31,320 --> 00:22:33,230
And that is a really
characteristic thing

553
00:22:33,230 --> 00:22:33,860
about science.

554
00:22:33,860 --> 00:22:35,760
And the question is,
is it, in some sense,

555
00:22:35,760 --> 00:22:37,755
a characteristic thing
about common sense.

556
00:22:37,755 --> 00:22:39,380
So that's what we're
going to ask here.

557
00:22:39,380 --> 00:22:42,530
We're going to jump to
much older children here.

558
00:22:42,530 --> 00:22:45,110
These are four and
five-year-olds.

559
00:22:45,110 --> 00:22:47,660
And we're going to give
them a problem where

560
00:22:47,660 --> 00:22:49,910
instead of showing them
the disambiguating data,

561
00:22:49,910 --> 00:22:52,277
we're going to ask if the
kids themselves will find it.

562
00:22:52,277 --> 00:22:54,360
So what we showed children--
when you were little,

563
00:22:54,360 --> 00:22:55,700
you possibly played
with some beads

564
00:22:55,700 --> 00:22:57,430
that snapped together
and pulled apart.

565
00:22:57,430 --> 00:22:58,960
These are like toddler toys.

566
00:22:58,960 --> 00:23:00,710
So we gave them these
snap together beads.

567
00:23:00,710 --> 00:23:02,120
They're each uniquely colored.

568
00:23:05,530 --> 00:23:08,220
We place each bead, one
at a time, on a toy.

569
00:23:08,220 --> 00:23:12,110
And in one condition, every
bead makes the toy play music

570
00:23:12,110 --> 00:23:13,610
to each bead you put on.

571
00:23:13,610 --> 00:23:16,615
And the other condition,
only half the beads did.

572
00:23:16,615 --> 00:23:18,740
So the only difference
between these two conditions

573
00:23:18,740 --> 00:23:20,990
is basically the base rate
of the candidate causes.

574
00:23:20,990 --> 00:23:23,060
One works for every bead,
the other only works

575
00:23:23,060 --> 00:23:25,880
for some of the beads.

576
00:23:25,880 --> 00:23:28,084
And then we took all
these training toys away,

577
00:23:28,084 --> 00:23:29,750
and we showed the
children either a pair

578
00:23:29,750 --> 00:23:31,460
that we had epoxied
together-- it was stuck.

579
00:23:31,460 --> 00:23:33,290
We tried to pull it apart,
we showed we couldn't.

580
00:23:33,290 --> 00:23:34,700
The children tried to pull
it apart, they couldn't.

581
00:23:34,700 --> 00:23:36,230
It's a stuck pair
of beads, or it's

582
00:23:36,230 --> 00:23:38,810
an ordinary, separable
pair of beads.

583
00:23:38,810 --> 00:23:41,240
And then, as a pair, we placed
each pair, one at a time,

584
00:23:41,240 --> 00:23:44,120
on the toy, and the
toy played music.

585
00:23:44,120 --> 00:23:46,400
In principle, this evidence
is always confounded.

586
00:23:46,400 --> 00:23:49,580
One bead in each pair might
be the responsible party

587
00:23:49,580 --> 00:23:51,267
activating the toy.

588
00:23:51,267 --> 00:23:52,850
But as a practical
matter, if you just

589
00:23:52,850 --> 00:23:55,529
learn the base rate is that
every single bead activates

590
00:23:55,529 --> 00:23:58,070
this toy, there's not a lot of
information to be gained here.

591
00:23:58,070 --> 00:24:00,980
You should just assume that
all of these beads work.

592
00:24:00,980 --> 00:24:02,870
And in that
condition, we expected

593
00:24:02,870 --> 00:24:05,060
that kids would play
indiscriminately

594
00:24:05,060 --> 00:24:06,352
with the two toys.

595
00:24:06,352 --> 00:24:09,350
But in the condition where
only some of the beads work,

596
00:24:09,350 --> 00:24:10,830
there's genuine uncertainty.

597
00:24:10,830 --> 00:24:11,330
Right?

598
00:24:11,330 --> 00:24:12,900
Maybe only one of
these beads work.

599
00:24:12,900 --> 00:24:14,480
Maybe they both did.

600
00:24:14,480 --> 00:24:15,980
And if that's true,
and if kids are

601
00:24:15,980 --> 00:24:18,440
sensitive to the possibility
of information gain,

602
00:24:18,440 --> 00:24:21,360
only one of these beads affords
the possibility of finding out.

603
00:24:21,360 --> 00:24:24,330
On only one can you
isolate the variables.

604
00:24:24,330 --> 00:24:25,820
And that's with
the separable pair.

605
00:24:25,820 --> 00:24:28,070
So we thought on this condition,
the kids should selectively

606
00:24:28,070 --> 00:24:29,278
play with the separable pair.

607
00:24:29,278 --> 00:24:32,120
And in particular, they should
place each bead, one at a time,

608
00:24:32,120 --> 00:24:34,220
on the toy.

609
00:24:34,220 --> 00:24:35,657
So that's, in
fact, what we find.

610
00:24:35,657 --> 00:24:37,490
In the obvious condition,
the kids basically

611
00:24:37,490 --> 00:24:39,164
never separated the pair.

612
00:24:39,164 --> 00:24:41,330
And in the some beads
condition, about half the kids

613
00:24:41,330 --> 00:24:46,280
did it and performed the
exhaustive intervention.

614
00:24:46,280 --> 00:24:48,620
That was cool, but my graduate
student at the time said,

615
00:24:48,620 --> 00:24:50,370
they're doing something
really interesting

616
00:24:50,370 --> 00:24:52,610
even with the stuck
pair of beads.

617
00:24:52,610 --> 00:24:54,129
We should look at
the stuck pair.

618
00:24:54,129 --> 00:24:56,170
And I said, what can they
do with the stuck pair?

619
00:24:56,170 --> 00:24:57,730
There's nothing to be
done with the stuck pair.

620
00:24:57,730 --> 00:24:58,230
It's stuck.

621
00:24:58,230 --> 00:25:00,021
And she said, well,
let's just try it again

622
00:25:00,021 --> 00:25:00,869
with the stuck pair.

623
00:25:00,869 --> 00:25:01,910
So we did the same thing.

624
00:25:01,910 --> 00:25:03,980
They got introduced the fact
that either every bead worked,

625
00:25:03,980 --> 00:25:05,000
or only some of
the beads worked.

626
00:25:05,000 --> 00:25:07,190
And this time we introduced
just the stuck pair,

627
00:25:07,190 --> 00:25:10,180
and we placed it on the
toy, and the toy made music.

628
00:25:10,180 --> 00:25:11,930
And let me show you
what the children did.

629
00:25:11,930 --> 00:25:12,200
[VIDEO PLAYBACK]

630
00:25:12,200 --> 00:25:14,156
- All right, I'm
going to do this one,

631
00:25:14,156 --> 00:25:16,112
and then it'll be your
turn a little later.

632
00:25:16,112 --> 00:25:19,140
But now, can you just watch
and see watch happens?

633
00:25:19,140 --> 00:25:19,895
All right.

634
00:25:24,178 --> 00:25:27,154
This one makes the machine go.

635
00:25:27,154 --> 00:25:30,240
How about this one?

636
00:25:30,240 --> 00:25:33,350
This one doesn't
make the machine go.

637
00:25:33,350 --> 00:25:36,130
What about this one?

638
00:25:36,130 --> 00:25:39,634
This one doesn't
make the machine go.

639
00:25:39,634 --> 00:25:42,898
Let's try this one.

640
00:25:42,898 --> 00:25:44,182
This one makes the machine go.

641
00:25:44,182 --> 00:25:46,140
LAURA SCHULZ: She goes
over that a second time,

642
00:25:46,140 --> 00:25:49,750
and then she hands
the child the toy.

643
00:25:49,750 --> 00:25:52,690
- Just a minute.

644
00:25:52,690 --> 00:25:54,557
- Look at that.

645
00:25:54,557 --> 00:25:55,140
[END PLAYBACK]

646
00:25:55,140 --> 00:25:57,639
LAURA SCHULZ: The child plays
around, does just what we did.

647
00:26:04,630 --> 00:26:07,651
And then she does
something we'd never done.

648
00:26:07,651 --> 00:26:09,150
She rotates the
position of the bead

649
00:26:09,150 --> 00:26:13,158
so that only one makes contact
with the toy at a time.

650
00:26:13,158 --> 00:26:17,898
And if you have a folk
theory of contact causality,

651
00:26:17,898 --> 00:26:21,400
that is a pretty good way
to isolate your variables.

652
00:26:21,400 --> 00:26:23,420
And not one that had
occurred to, say, the PI

653
00:26:23,420 --> 00:26:24,657
on this investigation.

654
00:26:24,657 --> 00:26:26,240
But in fact, it
occurred to about half

655
00:26:26,240 --> 00:26:28,370
the kids again,
in that condition.

656
00:26:28,370 --> 00:26:31,010
In the some beads condition,
where there was uncertainty,

657
00:26:31,010 --> 00:26:33,960
the kids were more likely to
design their own intervention

658
00:26:33,960 --> 00:26:36,950
to try to isolate the variables
than in the condition where

659
00:26:36,950 --> 00:26:38,951
all the beads worked.

660
00:26:38,951 --> 00:26:40,910
So preschoolers are
using information

661
00:26:40,910 --> 00:26:42,680
about the base rate
of candidate causes

662
00:26:42,680 --> 00:26:44,732
to distinguish the
ambiguity of the evidence,

663
00:26:44,732 --> 00:26:46,190
and they're selecting
and designing

664
00:26:46,190 --> 00:26:48,740
potentially informative
interventions to isolate

665
00:26:48,740 --> 00:26:52,350
these causal variables.

666
00:26:52,350 --> 00:26:54,200
All right.

667
00:26:54,200 --> 00:26:55,750
I'm going to show
you some new work

668
00:26:55,750 --> 00:26:59,010
now, kind of on the same theme.

669
00:26:59,010 --> 00:27:01,010
One way that investigations
can be uninformative

670
00:27:01,010 --> 00:27:02,691
is because evidence
is confounded.

671
00:27:02,691 --> 00:27:04,190
We're all familiar
with that, right?

672
00:27:04,190 --> 00:27:05,960
We think we did the perfect
experiment, then we're like,

673
00:27:05,960 --> 00:27:07,580
oh, well, it really could have
been because of this really

674
00:27:07,580 --> 00:27:09,140
silly, boring reason.

675
00:27:09,140 --> 00:27:10,760
And that's disappointing.

676
00:27:10,760 --> 00:27:12,150
And we have to run it again.

677
00:27:12,150 --> 00:27:14,240
But another reason
that investigations

678
00:27:14,240 --> 00:27:17,060
can be uninformative is because
they generate outcomes that

679
00:27:17,060 --> 00:27:20,120
are super hard to distinguish.

680
00:27:20,120 --> 00:27:22,460
So if I have a
handkerchief in one pocket,

681
00:27:22,460 --> 00:27:24,704
and a candy cane in
the other, then a child

682
00:27:24,704 --> 00:27:26,120
who wants that
candy cane is going

683
00:27:26,120 --> 00:27:28,670
to have no trouble patting
you down and finding the candy

684
00:27:28,670 --> 00:27:29,240
cane.

685
00:27:29,240 --> 00:27:31,550
But if I have a
pen in one pocket

686
00:27:31,550 --> 00:27:33,770
and a candy cane in
the other, that's

687
00:27:33,770 --> 00:27:35,082
going to be a harder problem.

688
00:27:35,082 --> 00:27:37,040
And this might be more
salient to you if I say,

689
00:27:37,040 --> 00:27:39,860
you're going to go in and have
a lab test for a fatal disease,

690
00:27:39,860 --> 00:27:41,792
or potentially a benign disease.

691
00:27:41,792 --> 00:27:44,000
And you know what the results
are going to look like?

692
00:27:44,000 --> 00:27:45,860
One is going to
be reddish maroon,

693
00:27:45,860 --> 00:27:48,620
and the other is going
to be maroonish red.

694
00:27:48,620 --> 00:27:50,330
OK, that's not the
kind of test you want.

695
00:27:50,330 --> 00:27:54,470
You want yellow, blue, right?

696
00:27:54,470 --> 00:27:55,420
So this is important.

697
00:27:55,420 --> 00:27:58,040
We care about how
much uncertainty there

698
00:27:58,040 --> 00:28:01,230
is over interpreting
the outcome as well.

699
00:28:01,230 --> 00:28:06,230
So if children are sensitive
to how useful actions are

700
00:28:06,230 --> 00:28:07,910
for information gain,
then they should

701
00:28:07,910 --> 00:28:10,610
prefer interventions that
generate distinctive patterns

702
00:28:10,610 --> 00:28:12,260
of evidence.

703
00:28:12,260 --> 00:28:17,510
So Max Siegel in my lab has
been running some experiments

704
00:28:17,510 --> 00:28:18,455
like this.

705
00:28:18,455 --> 00:28:20,390
He started with a very
simple one, basically

706
00:28:20,390 --> 00:28:21,950
the equivalent of the
handkerchief and the candy

707
00:28:21,950 --> 00:28:22,450
cane.

708
00:28:22,450 --> 00:28:24,680
He said, OK, there's either
a bean bag in this box

709
00:28:24,680 --> 00:28:26,971
that I'm going to put in
here, or a pencil in this box.

710
00:28:26,971 --> 00:28:29,150
It is a shiny, cool,
hologram, sparkly pencil.

711
00:28:29,150 --> 00:28:30,632
You'll want it.

712
00:28:30,632 --> 00:28:32,340
That's going to go in
this box over here.

713
00:28:32,340 --> 00:28:33,890
And in this box,
either the really cool,

714
00:28:33,890 --> 00:28:35,970
shiny hologram pencil, or the
really boring yellow pencil

715
00:28:35,970 --> 00:28:36,690
is going to go in this box.

716
00:28:36,690 --> 00:28:37,790
And guess what I'm going to do?

717
00:28:37,790 --> 00:28:39,890
I'm going to take each box,
and I'm going to shake it.

718
00:28:39,890 --> 00:28:40,580
So he does that.

719
00:28:40,580 --> 00:28:42,070
And you know what
you hear both times?

720
00:28:42,070 --> 00:28:43,270
Ka-thunk, ka-thunk, ka-thunk.

721
00:28:43,270 --> 00:28:44,478
Ka-thunk, ka-thunk, ka-thunk.

722
00:28:44,478 --> 00:28:46,100
Indistinguishable sounds.

723
00:28:46,100 --> 00:28:49,100
And now the question is,
which box do you want to open?

724
00:28:49,100 --> 00:28:50,710
Which box you want to open?

725
00:28:50,710 --> 00:28:51,290
Right?

726
00:28:51,290 --> 00:28:52,850
And if you're sensitive
to the ambiguity,

727
00:28:52,850 --> 00:28:54,710
you should say, well,
listen, if it were a beanbag,

728
00:28:54,710 --> 00:28:57,160
I would really know, so that
must be the sparkly pencil,

729
00:28:57,160 --> 00:28:57,410
right?

730
00:28:57,410 --> 00:28:58,820
But I'm never going
to know in this box,

731
00:28:58,820 --> 00:29:00,736
because both pencils are
going to sound alike,

732
00:29:00,736 --> 00:29:03,280
so I really better
choose this box.

733
00:29:03,280 --> 00:29:07,100
And then he's going to
do a harder problem.

734
00:29:07,100 --> 00:29:11,884
He's going to say, there are
eight shiny, colorful marbles--

735
00:29:11,884 --> 00:29:12,800
you really want them--

736
00:29:12,800 --> 00:29:15,400
in this box, or two
really boring white ones.

737
00:29:15,400 --> 00:29:18,800
Or there are eight colorful,
shiny marbles in this box,

738
00:29:18,800 --> 00:29:22,130
or six really boring white ones.

739
00:29:22,130 --> 00:29:24,980
In each case, they
each get hidden, you're

740
00:29:24,980 --> 00:29:26,627
going to hear the box.

741
00:29:26,627 --> 00:29:29,210
In each case, it's going to make
exactly the same sound, which

742
00:29:29,210 --> 00:29:31,220
is actually the sound of
eight marbles in a box.

743
00:29:31,220 --> 00:29:33,420
And the question is, which
box do you want to open?

744
00:29:33,420 --> 00:29:34,400
So let me show you
how that works.

745
00:29:34,400 --> 00:29:35,296
[VIDEO PLAYBACK]

746
00:29:35,296 --> 00:29:37,490
- [INAUDIBLE] and I
also have some marbles.

747
00:29:37,490 --> 00:29:40,622
Well, you see these marbles
right here, the white ones?

748
00:29:40,622 --> 00:29:42,068
These are Bunny's.

749
00:29:42,068 --> 00:29:45,442
Oh, six of my marbles, yay.

750
00:29:45,442 --> 00:29:48,345
Oh, two of my marbles, yay.

751
00:29:48,345 --> 00:29:50,190
Guess what, Taylor?

752
00:29:50,190 --> 00:29:52,810
These marbles with lots
of different colors, those

753
00:29:52,810 --> 00:29:54,840
are yours for right now.

754
00:29:54,840 --> 00:29:57,244
That's pretty cool, right?

755
00:29:57,244 --> 00:29:59,232
Those are your marbles.

756
00:29:59,232 --> 00:30:00,512
That's awesome.

757
00:30:00,512 --> 00:30:03,990
And in this game, I'm
going to hide either

758
00:30:03,990 --> 00:30:07,990
your marbles or Bunny's
marbles inside of this box.

759
00:30:07,990 --> 00:30:09,850
And then I'm going
to hide either

760
00:30:09,850 --> 00:30:13,570
your or Bunny's marbles
inside of this box.

761
00:30:13,570 --> 00:30:14,769
Does that sound like fun?

762
00:30:14,769 --> 00:30:16,852
And then we're going to
look for your marbles, OK?

763
00:30:16,852 --> 00:30:19,497
If you find them, you
get another sticker.

764
00:30:19,497 --> 00:30:21,580
All right, so I'm going
to put Bunny's right here,

765
00:30:21,580 --> 00:30:23,090
and we're going to
do the hide game.

766
00:30:23,090 --> 00:30:25,665
So first, I'm going to choose
either your marbles or Bunny's

767
00:30:25,665 --> 00:30:27,705
marbles and put them in here.

768
00:30:27,705 --> 00:30:28,890
I'm going to pour them in.

769
00:30:33,174 --> 00:30:36,020
Look, somebody's
marbles are in here.

770
00:30:36,020 --> 00:30:38,190
Now I'm going to do the
same thing with this box.

771
00:30:38,190 --> 00:30:39,959
Either your marbles
or Bunny's marbles

772
00:30:39,959 --> 00:30:41,408
are going to go in this box.

773
00:30:45,255 --> 00:30:45,755
All right.

774
00:30:45,755 --> 00:30:48,750
Are you ready to begin
shaking and listening?

775
00:30:48,750 --> 00:30:52,000
So remember, over here,
there's either your marbles

776
00:30:52,000 --> 00:30:54,340
or Bunny's marbles.

777
00:30:54,340 --> 00:30:55,580
OK, let's listen.

778
00:30:55,580 --> 00:30:57,404
[CLUNKING NOISES]

779
00:31:00,101 --> 00:31:00,600
All right.

780
00:31:00,600 --> 00:31:03,572
And over here, there's
either your marbles

781
00:31:03,572 --> 00:31:05,378
or Bunny's marbles.

782
00:31:05,378 --> 00:31:06,346
Let's listen.

783
00:31:06,346 --> 00:31:08,282
[CLUNKING NOISES]

784
00:31:10,686 --> 00:31:11,186
Cool.

785
00:31:11,186 --> 00:31:12,510
Let's do it one more time.

786
00:31:12,510 --> 00:31:12,780
[END PLAYBACK]

787
00:31:12,780 --> 00:31:14,400
LAURA SCHULZ: We'll skip
the one more time, but--

788
00:31:14,400 --> 00:31:15,261
oops.

789
00:31:15,261 --> 00:31:15,760
Sorry.

790
00:31:23,450 --> 00:31:26,220
You can at the
general principle.

791
00:31:26,220 --> 00:31:28,700
The children are overwhelmingly
good at this kind of task,

792
00:31:28,700 --> 00:31:30,590
it turns out, in
both of these cases

793
00:31:30,590 --> 00:31:33,110
and in many, many other
iterations they went.

794
00:31:33,110 --> 00:31:35,347
They're very confident
about which box

795
00:31:35,347 --> 00:31:37,430
they should pick, which
means they're representing

796
00:31:37,430 --> 00:31:39,950
to themselves something about
the ambiguity of the evidence,

797
00:31:39,950 --> 00:31:41,870
and their own
ability to perceive

798
00:31:41,870 --> 00:31:44,760
these kinds of distinctions.

799
00:31:44,760 --> 00:31:46,570
So with Max, we've
been talking about this

800
00:31:46,570 --> 00:31:48,590
as a kind of intuitive
psychophysics,

801
00:31:48,590 --> 00:31:50,690
where they can represent
their own discrimination

802
00:31:50,690 --> 00:31:54,770
threshold to make these
kinds of distinctions.

803
00:31:54,770 --> 00:31:56,570
And they prefer
interventions that

804
00:31:56,570 --> 00:31:58,640
generate distinctive
patterns of evidence

805
00:31:58,640 --> 00:32:00,790
and maximize the possibility
of information gain.

806
00:32:00,790 --> 00:32:02,440
Is that clear?

807
00:32:02,440 --> 00:32:03,920
OK.

808
00:32:03,920 --> 00:32:06,740
So there's a lot of
ways in which children

809
00:32:06,740 --> 00:32:09,530
seem to be using intuitive
theories, some kind

810
00:32:09,530 --> 00:32:11,800
of abstract, higher
order of knowledge,

811
00:32:11,800 --> 00:32:19,130
to make inferences from data.

812
00:32:19,130 --> 00:32:23,600
But information is also
costly to generate.

813
00:32:23,600 --> 00:32:25,890
And the costs themselves
are informative.

814
00:32:25,890 --> 00:32:28,920
So I'm going to talk a little
bit about that piece now.

815
00:32:28,920 --> 00:32:30,440
It's costly both
for the learner,

816
00:32:30,440 --> 00:32:32,390
who cannot learn everything.

817
00:32:32,390 --> 00:32:35,582
And in a cultural context,
where you're not just

818
00:32:35,582 --> 00:32:37,040
learning and
exploring on your own,

819
00:32:37,040 --> 00:32:39,500
but you're actually getting
information from other people,

820
00:32:39,500 --> 00:32:47,220
it's costly also for the
teacher or for the informant.

821
00:32:47,220 --> 00:32:48,989
And these kinds
of costs, and how

822
00:32:48,989 --> 00:32:50,780
we negotiate these
kinds of costs, I think,

823
00:32:50,780 --> 00:32:53,750
are really fundamental
to a lot of hard problems

824
00:32:53,750 --> 00:32:56,920
in communication and language.

825
00:32:56,920 --> 00:32:59,540
Lots of the field of
pragmatics deals with problems

826
00:32:59,540 --> 00:33:00,869
of underdetermination.

827
00:33:00,869 --> 00:33:02,910
We say these sentences,
we understand each other.

828
00:33:02,910 --> 00:33:05,690
We understand each other in all
kinds of ambiguous contexts.

829
00:33:05,690 --> 00:33:08,000
We use a lot of social
cues and other information

830
00:33:08,000 --> 00:33:09,140
to disambiguate.

831
00:33:09,140 --> 00:33:11,960
But part of what we do
is, we make inferences

832
00:33:11,960 --> 00:33:13,790
about how much this
person is going

833
00:33:13,790 --> 00:33:17,150
to communicate in
this context, given

834
00:33:17,150 --> 00:33:19,820
how much I need to understand.

835
00:33:19,820 --> 00:33:22,505
And we use this to resolve
these kinds of ambiguities.

836
00:33:22,505 --> 00:33:26,431
I'm going to give you a
few examples of that here.

837
00:33:26,431 --> 00:33:28,430
Now, again, I'm going to
start with the study we

838
00:33:28,430 --> 00:33:30,929
did a while ago then show you
a little bit more recent work.

839
00:33:38,170 --> 00:33:39,920
I'm going to skip
over some of this just

840
00:33:39,920 --> 00:33:42,489
to be able to cover
all of this and say,

841
00:33:42,489 --> 00:33:44,780
Because there's a cost of
information for both teachers

842
00:33:44,780 --> 00:33:47,150
and learners, it
predicts some trade-offs

843
00:33:47,150 --> 00:33:49,260
in the kinds of inferences
you should make.

844
00:33:49,260 --> 00:33:51,650
So for instance, if a
knowledgeable informant

845
00:33:51,650 --> 00:33:55,530
shows you a toy and says, here,
this toy has a single function.

846
00:33:55,530 --> 00:33:59,240
Then if you, the learner, think
that that teacher is trying

847
00:33:59,240 --> 00:34:01,295
to generate a true sample
from the hypothesis

848
00:34:01,295 --> 00:34:02,690
based on what actually
is going to get

849
00:34:02,690 --> 00:34:04,314
the right idea to
your head, you should

850
00:34:04,314 --> 00:34:07,850
assume that there is only
one function and not two,

851
00:34:07,850 --> 00:34:08,976
or three, or four, or five.

852
00:34:08,976 --> 00:34:11,600
Because if there were more, they
should have shown them to you,

853
00:34:11,600 --> 00:34:12,109
right?

854
00:34:12,109 --> 00:34:13,564
Because if they know
the true hypothesis,

855
00:34:13,564 --> 00:34:15,105
and they could just
demonstrate that,

856
00:34:15,105 --> 00:34:18,110
they can rule out
all of that for you.

857
00:34:21,090 --> 00:34:24,860
But if you just stumble upon
a single function of a toy,

858
00:34:24,860 --> 00:34:27,860
or a not knowledgeable teacher
generates it accidentally,

859
00:34:27,860 --> 00:34:30,739
or if the teacher, as
Liz Spelke pointed out,

860
00:34:30,739 --> 00:34:33,800
is interrupted in the middle
of that demonstration,

861
00:34:33,800 --> 00:34:36,560
then you shouldn't assume that
that evidence is exhaustive.

862
00:34:36,560 --> 00:34:38,620
It suspends that
inference, right?

863
00:34:38,620 --> 00:34:40,010
Now, OK, well,
you showed me one,

864
00:34:40,010 --> 00:34:41,935
but maybe there are
two, three, or four.

865
00:34:41,935 --> 00:34:43,310
So it's only in
a condition where

866
00:34:43,310 --> 00:34:45,920
I think you are a fully
informed, freely acting

867
00:34:45,920 --> 00:34:48,560
teacher that I should assume,
well, look, if you are helpful,

868
00:34:48,560 --> 00:34:50,850
knowledgeable teacher,
then the information

869
00:34:50,850 --> 00:34:53,969
you give me should not only
be true of the hypothesis,

870
00:34:53,969 --> 00:34:56,300
it should help me
distinguish that hypothesis

871
00:34:56,300 --> 00:34:59,190
from available alternatives.

872
00:34:59,190 --> 00:35:01,670
And what that means is,
there's a specific trade-off

873
00:35:01,670 --> 00:35:03,860
between instruction
and exploration.

874
00:35:03,860 --> 00:35:05,540
Because if I'm
instructed that there's

875
00:35:05,540 --> 00:35:08,462
one function of the toy, I don't
need to explore any further.

876
00:35:08,462 --> 00:35:10,670
But if I just happen to find
one function of the toy,

877
00:35:10,670 --> 00:35:12,500
maybe I do.

878
00:35:12,500 --> 00:35:14,930
So let me show you what
we did to test this.

879
00:35:14,930 --> 00:35:15,890
We had a novel toy.

880
00:35:15,890 --> 00:35:18,620
It actually had four interesting
properties, a squeaker,

881
00:35:18,620 --> 00:35:20,930
a light, a mirror, and music.

882
00:35:20,930 --> 00:35:23,120
And we demonstrated
a single function

883
00:35:23,120 --> 00:35:27,686
of the toy, the squeaker,
in three conditions.

884
00:35:27,686 --> 00:35:28,810
And we also had a baseline.

885
00:35:28,810 --> 00:35:31,320
In the pedagogical condition,
we said, watch this,

886
00:35:31,320 --> 00:35:33,290
I'm going to show you--

887
00:35:33,290 --> 00:35:35,510
sorry, the alignment's off--

888
00:35:35,510 --> 00:35:37,580
but watch this, I'm
going to show you my toy.

889
00:35:37,580 --> 00:35:40,250
They pulled the toy and
then said, wow, see that.

890
00:35:40,250 --> 00:35:42,080
OK, the accidental
condition was,

891
00:35:42,080 --> 00:35:44,030
look at this neat
toy I found here,

892
00:35:44,030 --> 00:35:47,446
accidentally pulled this tube
in the same way, wow see that.

893
00:35:47,446 --> 00:35:49,820
The baseline was, just look
at this neat toy I have here,

894
00:35:49,820 --> 00:35:51,170
with no demonstration.

895
00:35:51,170 --> 00:35:52,970
And the interrupted
condition was identical

896
00:35:52,970 --> 00:35:55,460
to the pedagogical condition,
except the teacher was

897
00:35:55,460 --> 00:35:58,160
interrupted immediately
after pulling the tube,

898
00:35:58,160 --> 00:36:00,122
and then she said,
wow, see that.

899
00:36:00,122 --> 00:36:00,830
So is that clear?

900
00:36:00,830 --> 00:36:04,490
I'm sorry for the
slide misalignment.

901
00:36:04,490 --> 00:36:08,030
And the prediction is that
in the first condition,

902
00:36:08,030 --> 00:36:10,520
children should constrain their
exploration relative to all

903
00:36:10,520 --> 00:36:11,464
the other conditions.

904
00:36:11,464 --> 00:36:13,130
So let me show you
what that looks like.

905
00:36:23,210 --> 00:36:23,710
Or not.

906
00:36:23,710 --> 00:36:26,209
Which is too bad, because this
is a really super cute slide.

907
00:36:26,209 --> 00:36:27,650
But it's not going to work.

908
00:36:27,650 --> 00:36:29,870
In this condition,
what we found was this.

909
00:36:29,870 --> 00:36:32,180
We found a child in the
children's' museum with a toy

910
00:36:32,180 --> 00:36:33,888
with all of these kind
of wow properties.

911
00:36:33,888 --> 00:36:35,049
We say, wow, see this.

912
00:36:35,049 --> 00:36:36,590
We show them the
property of the toy,

913
00:36:36,590 --> 00:36:41,210
and the child spends 90 seconds
pulling only the squeaker.

914
00:36:41,210 --> 00:36:43,481
He then says, I'm very
smart for a five-year-old.

915
00:36:43,481 --> 00:36:45,980
And when she asks for all of
the other functions of the toy,

916
00:36:45,980 --> 00:36:47,771
he doesn't know any of
the other functions,

917
00:36:47,771 --> 00:36:48,922
because he hasn't explored.

918
00:36:48,922 --> 00:36:51,380
And what we think is, he is
very smart for a five-year-old.

919
00:36:51,380 --> 00:36:53,180
Because it's a completely
rational inference that,

920
00:36:53,180 --> 00:36:54,980
if there were more
functions of the toy,

921
00:36:54,980 --> 00:36:57,650
then they should have
been demonstrated.

922
00:36:57,650 --> 00:36:59,960
And so what we find
overall is, in fact,

923
00:36:59,960 --> 00:37:02,330
that children do fewer
actions, and they discover

924
00:37:02,330 --> 00:37:04,490
fewer functions of this toy.

925
00:37:04,490 --> 00:37:07,430
This isn't just true, it
turns out, we now know,

926
00:37:07,430 --> 00:37:09,890
because we live in a
hyperpedagogical culture.

927
00:37:09,890 --> 00:37:13,010
Because Laura Shneidman
and Amanda Woodward

928
00:37:13,010 --> 00:37:15,490
just replicated this
study with Yucatec Mayan

929
00:37:15,490 --> 00:37:19,394
toddlers and found the same
kind of effect, constraints

930
00:37:19,394 --> 00:37:20,810
in the pedagogical
condition, even

931
00:37:20,810 --> 00:37:22,268
though it's a
culture that's pretty

932
00:37:22,268 --> 00:37:26,460
limited in their pedagogy.

933
00:37:26,460 --> 00:37:30,810
So information is costly,
and pedagogical contexts

934
00:37:30,810 --> 00:37:33,200
strengthen the inference that
the absence of evidence--

935
00:37:33,200 --> 00:37:36,050
a teacher's failure to go on
and teach you more information--

936
00:37:36,050 --> 00:37:39,308
is, in fact, evidence
of its absence.

937
00:37:39,308 --> 00:37:41,300
Is that clear?

938
00:37:41,300 --> 00:37:44,030
And this is a very
sensible inductive bias,

939
00:37:44,030 --> 00:37:47,355
but it predicts that instruction
will, for better or worse,

940
00:37:47,355 --> 00:37:48,230
constrain expression.

941
00:37:48,230 --> 00:37:50,220
Because that's what
it's supposed to do.

942
00:37:50,220 --> 00:37:52,850
It's supposed to constrain
the hypotheses you consider.

943
00:37:52,850 --> 00:37:54,399
And indeed, it works quite well.

944
00:37:54,399 --> 00:37:56,690
And that's good if you're
right about the world, right?

945
00:37:56,690 --> 00:37:58,956
And it constrains it
to efficient learning.

946
00:37:58,956 --> 00:38:00,830
But it's bad if you're
wrong about the world.

947
00:38:00,830 --> 00:38:03,350
Because the unknown unknowns,
the things you don't know

948
00:38:03,350 --> 00:38:05,020
are true that you
failed to teach,

949
00:38:05,020 --> 00:38:09,440
are going to potentially
mislead a learner.

950
00:38:09,440 --> 00:38:10,820
How much is enough information?

951
00:38:10,820 --> 00:38:12,320
Well, there are
lots of good reasons

952
00:38:12,320 --> 00:38:14,794
why teachers ought to provide
very limited information.

953
00:38:14,794 --> 00:38:17,210
First of all, as I showed you
in the first set of studies,

954
00:38:17,210 --> 00:38:19,090
evidence often supports
generalization.

955
00:38:19,090 --> 00:38:19,590
Right?

956
00:38:19,590 --> 00:38:21,173
One dog toy squeaks,
probably they all

957
00:38:21,173 --> 00:38:25,356
squeak, barring how
generalizable that sample is.

958
00:38:25,356 --> 00:38:27,230
So I don't need to show
you every single toy.

959
00:38:27,230 --> 00:38:28,760
I don't need to
show a child, this

960
00:38:28,760 --> 00:38:30,170
is a cup, and that's a
cup, and this a cup too,

961
00:38:30,170 --> 00:38:31,970
and that's a cup,
and that's a cup.

962
00:38:31,970 --> 00:38:33,650
Once the child has
a cup, I can assume

963
00:38:33,650 --> 00:38:35,870
that that child herself
will be able to make

964
00:38:35,870 --> 00:38:37,282
the rational generalization.

965
00:38:37,282 --> 00:38:39,740
Or sometimes I know you're not
going to be able to make it,

966
00:38:39,740 --> 00:38:41,630
but the additional information
is just too costly.

967
00:38:41,630 --> 00:38:43,338
I'm working on teaching
you two plus two,

968
00:38:43,338 --> 00:38:46,280
I'm not going to teach you
linear algebra right now.

969
00:38:46,280 --> 00:38:47,667
It's a waste of our time.

970
00:38:47,667 --> 00:38:49,250
So that's another
reason why you might

971
00:38:49,250 --> 00:38:51,270
provide limited information.

972
00:38:51,270 --> 00:38:57,890
So what are the contexts in
which omitting information is

973
00:38:57,890 --> 00:39:01,010
a reasonable thing to do,
and when is it misleading?

974
00:39:01,010 --> 00:39:04,020
When is this a real problem?

975
00:39:04,020 --> 00:39:06,680
And the answer turns out to
be, if I'm the informant,

976
00:39:06,680 --> 00:39:09,080
and I know I'm
providing information

977
00:39:09,080 --> 00:39:12,184
that is going to lead you
to the wrong hypothesis,

978
00:39:12,184 --> 00:39:13,850
then we consider that
a sin of omission.

979
00:39:13,850 --> 00:39:15,020
Right?

980
00:39:15,020 --> 00:39:18,390
If I'm omitting information,
and I'm not doing that,

981
00:39:18,390 --> 00:39:20,014
then maybe that's not a problem.

982
00:39:20,014 --> 00:39:21,680
So one of the questions
is, can children

983
00:39:21,680 --> 00:39:23,210
distinguish these contexts?

984
00:39:23,210 --> 00:39:26,810
Can they tell when the
teacher is providing

985
00:39:26,810 --> 00:39:30,560
too little information, and it
is going to cost the learner

986
00:39:30,560 --> 00:39:34,970
something in terms of what they
can gain, and when they're not.

987
00:39:34,970 --> 00:39:36,470
So to test this--

988
00:39:36,470 --> 00:39:38,780
this is again,
Hyowon Gweon's work--

989
00:39:38,780 --> 00:39:40,310
we introduced a toy.

990
00:39:40,310 --> 00:39:42,462
And it had one function,
this wind up mechanism.

991
00:39:42,462 --> 00:39:44,420
And the kids got to
explore, and they found out

992
00:39:44,420 --> 00:39:46,000
the toy did one thing.

993
00:39:46,000 --> 00:39:48,330
In the other condition,
the toy looked the same.

994
00:39:48,330 --> 00:39:50,420
But in fact, the toy
had lots of functions,

995
00:39:50,420 --> 00:39:52,460
and the children knew that.

996
00:39:52,460 --> 00:39:54,420
So the children always
knew the ground truth.

997
00:39:54,420 --> 00:39:57,290
The toy either had
one function or four.

998
00:39:57,290 --> 00:39:59,910
And then there was a
teacher who taught Elmo.

999
00:39:59,910 --> 00:40:01,520
The teacher always
did the same thing.

1000
00:40:01,520 --> 00:40:03,970
The teacher always
taught just one function.

1001
00:40:03,970 --> 00:40:06,830
And the first question
was, the teacher's always

1002
00:40:06,830 --> 00:40:09,370
doing the same thing with
an identical looking toy.

1003
00:40:09,370 --> 00:40:11,229
Do the kids penalize
the teacher?

1004
00:40:11,229 --> 00:40:13,020
Do they think he's a
bad teacher if he only

1005
00:40:13,020 --> 00:40:15,080
teaches one function when
there are really four,

1006
00:40:15,080 --> 00:40:17,960
compared to when he teaches one
function and there's only one.

1007
00:40:17,960 --> 00:40:20,840
So the first thing we did was,
ask kids to rate that teacher.

1008
00:40:20,840 --> 00:40:22,640
And indeed, they think
that this teacher

1009
00:40:22,640 --> 00:40:24,520
is a terrible teacher when
there's four functions

1010
00:40:24,520 --> 00:40:25,520
and he only teaches one.

1011
00:40:25,520 --> 00:40:29,890
They think he's a good teacher
when he teaches one of one.

1012
00:40:29,890 --> 00:40:33,380
But the really interesting
question was, what would

1013
00:40:33,380 --> 00:40:35,300
the children do to
compensate if they

1014
00:40:35,300 --> 00:40:36,800
knew they had a bad teacher?

1015
00:40:36,800 --> 00:40:38,960
So we ran exactly
the same set up

1016
00:40:38,960 --> 00:40:41,750
where the teacher shows Elmo
the toy in the one function case

1017
00:40:41,750 --> 00:40:44,000
and the four function case.

1018
00:40:44,000 --> 00:40:46,209
And for reasons that
will become clear,

1019
00:40:46,209 --> 00:40:47,750
we also ran a control
condition where

1020
00:40:47,750 --> 00:40:49,670
there were four
active functions,

1021
00:40:49,670 --> 00:40:53,030
and the teacher taught all four.

1022
00:40:53,030 --> 00:40:54,890
In all cases, the
teacher then goes on

1023
00:40:54,890 --> 00:40:56,780
and runs that experiment
I just showed you.

1024
00:40:56,780 --> 00:40:58,490
The teacher shows
just the squeaker toy

1025
00:40:58,490 --> 00:41:00,832
here of this single function.

1026
00:41:00,832 --> 00:41:02,540
The question is, what
should the kids do?

1027
00:41:02,540 --> 00:41:04,456
So it's a complicated
set up, so I'll walk you

1028
00:41:04,456 --> 00:41:06,630
through it a little bit.

1029
00:41:06,630 --> 00:41:09,199
When you teach one
of one function,

1030
00:41:09,199 --> 00:41:11,240
you should infer the toy
probably does one thing,

1031
00:41:11,240 --> 00:41:12,470
it's a good teacher.

1032
00:41:12,470 --> 00:41:13,910
And so when they show
you one function of this,

1033
00:41:13,910 --> 00:41:15,570
you should constrain
your exploration,

1034
00:41:15,570 --> 00:41:17,987
say that was a
sensible inference.

1035
00:41:17,987 --> 00:41:19,820
When they teach one of
four, you should say,

1036
00:41:19,820 --> 00:41:20,720
that's a bad teacher.

1037
00:41:20,720 --> 00:41:21,920
The toy probably does
more than one thing,

1038
00:41:21,920 --> 00:41:23,170
the new toy probably does too.

1039
00:41:23,170 --> 00:41:24,954
I'm going to explore
more broadly.

1040
00:41:24,954 --> 00:41:27,120
But we don't know, in that
case, if they're doing it

1041
00:41:27,120 --> 00:41:29,607
because they just saw a
toy with one function,

1042
00:41:29,607 --> 00:41:31,440
and so they think this
toy has one function.

1043
00:41:31,440 --> 00:41:32,820
And they just saw a toy
with four functions,

1044
00:41:32,820 --> 00:41:34,570
so they think this toy
has four functions.

1045
00:41:34,570 --> 00:41:36,930
So we can disambiguate
those with this condition.

1046
00:41:36,930 --> 00:41:38,970
Now if they're just
generalizing from the toy,

1047
00:41:38,970 --> 00:41:40,230
this toy has four
functions, well,

1048
00:41:40,230 --> 00:41:41,970
then they should think this
toy has four functions.

1049
00:41:41,970 --> 00:41:44,100
But if they're generalizing
from the teacher,

1050
00:41:44,100 --> 00:41:45,234
this is a good teacher.

1051
00:41:45,234 --> 00:41:47,400
So when the teacher now
tells you about the new toy,

1052
00:41:47,400 --> 00:41:50,220
that it has one
function, the kids

1053
00:41:50,220 --> 00:41:51,700
should constrain
their exploration.

1054
00:41:51,700 --> 00:41:55,260
So does everyone understand
the logic of the design here?

1055
00:41:55,260 --> 00:41:59,430
And in fact, that's
exactly what we find.

1056
00:41:59,430 --> 00:42:02,346
The children compensated
with additional exploration

1057
00:42:02,346 --> 00:42:03,720
when they thought
the teacher had

1058
00:42:03,720 --> 00:42:07,227
provided insufficient
information to the learner.

1059
00:42:07,227 --> 00:42:07,810
Is that clear?

1060
00:42:10,960 --> 00:42:11,520
OK.

1061
00:42:11,520 --> 00:42:14,820
So information is costly
to teachers and learners.

1062
00:42:14,820 --> 00:42:16,440
If teachers minimize
their own costs

1063
00:42:16,440 --> 00:42:18,330
and provide too
little information,

1064
00:42:18,330 --> 00:42:20,200
children think
they're poor teachers.

1065
00:42:20,200 --> 00:42:22,200
They suspend the inference
that that information

1066
00:42:22,200 --> 00:42:23,160
is representative.

1067
00:42:23,160 --> 00:42:27,900
They compensate with
additional exploration.

1068
00:42:27,900 --> 00:42:32,580
I'm going to go ahead and
show you just a couple

1069
00:42:32,580 --> 00:42:33,600
more examples here.

1070
00:42:37,300 --> 00:42:39,160
There's too little information.

1071
00:42:39,160 --> 00:42:40,940
But because
information is costly,

1072
00:42:40,940 --> 00:42:42,820
you can also provide
too much information.

1073
00:42:42,820 --> 00:42:44,890
I might be doing that right now.

1074
00:42:44,890 --> 00:42:46,810
Too much information is costly.

1075
00:42:46,810 --> 00:42:49,600
It's taking a toll on the
learner to absorb all of that.

1076
00:42:49,600 --> 00:42:51,058
And you have to
know, well, are you

1077
00:42:51,058 --> 00:42:53,420
providing me too much
information, or just the right

1078
00:42:53,420 --> 00:42:53,920
amount?

1079
00:42:53,920 --> 00:42:55,961
And at the risk of falling
into this trap myself,

1080
00:42:55,961 --> 00:42:58,600
I am going to show you
quickly this study.

1081
00:42:58,600 --> 00:43:00,620
Because how much
information is too much

1082
00:43:00,620 --> 00:43:02,620
information depends on a
hard question which is,

1083
00:43:02,620 --> 00:43:03,940
how much do you already know?

1084
00:43:03,940 --> 00:43:04,587
Right?

1085
00:43:04,587 --> 00:43:05,920
You've all been here all summer.

1086
00:43:05,920 --> 00:43:07,090
You know a lot of things.

1087
00:43:07,090 --> 00:43:09,160
I'm not a very good
estimate of what you know

1088
00:43:09,160 --> 00:43:11,240
or how much this
information is going,

1089
00:43:11,240 --> 00:43:13,040
so it's a little
hard for me titrate

1090
00:43:13,040 --> 00:43:16,750
what's the right amount of
information to give you.

1091
00:43:16,750 --> 00:43:18,340
And the question
is, can children

1092
00:43:18,340 --> 00:43:20,200
take these kinds of
theory of mind problems

1093
00:43:20,200 --> 00:43:23,110
into account in order to
estimate what information they

1094
00:43:23,110 --> 00:43:24,670
should be getting?

1095
00:43:24,670 --> 00:43:27,580
To test this, we give
kids a 20-button toy.

1096
00:43:27,580 --> 00:43:30,850
If I push a single button and
it makes music, how many of you

1097
00:43:30,850 --> 00:43:33,839
think that all the other
buttons make music?

1098
00:43:33,839 --> 00:43:35,380
Because you can
generalize from data,

1099
00:43:35,380 --> 00:43:36,970
and that is a really good
inductive inference there.

1100
00:43:36,970 --> 00:43:38,261
They look the same, it's a toy.

1101
00:43:38,261 --> 00:43:40,090
One makes music,
they probably all do.

1102
00:43:40,090 --> 00:43:41,880
But suppose I go on
now to show you--

1103
00:43:41,880 --> 00:43:44,400
so that's your prior
expectation, they all work.

1104
00:43:44,400 --> 00:43:46,750
But that one doesn't work,
and that one doesn't work,

1105
00:43:46,750 --> 00:43:48,958
and that one doesn't work,
and that one doesn't work,

1106
00:43:48,958 --> 00:43:51,562
and that one doesn't work,
and that-- oh, that one works.

1107
00:43:51,562 --> 00:43:53,770
And that one doesn't work,
and that one doesn't work,

1108
00:43:53,770 --> 00:43:55,450
and that one doesn't work,
and that one doesn't work.

1109
00:43:55,450 --> 00:43:56,440
I'm doing this for a reason.

1110
00:43:56,440 --> 00:43:57,648
I know this is really boring.

1111
00:43:57,648 --> 00:44:00,790
It's partly to give you a
break, but it is also-- oh,

1112
00:44:00,790 --> 00:44:03,200
that one works.

1113
00:44:03,200 --> 00:44:06,370
OK, so now what you have
learned about this toy

1114
00:44:06,370 --> 00:44:10,350
is that actually, only
three of these buttons work.

1115
00:44:10,350 --> 00:44:13,019
And suppose I show you this
across a couple of toys.

1116
00:44:13,019 --> 00:44:14,560
You've just changed
your expectation.

1117
00:44:14,560 --> 00:44:16,854
Now if I show you
that one button works,

1118
00:44:16,854 --> 00:44:18,270
you don't think
all the rest work.

1119
00:44:18,270 --> 00:44:20,830
You think probably two
others work, right?

1120
00:44:20,830 --> 00:44:25,410
So if I bring out a brand new
toy, and I push this button,

1121
00:44:25,410 --> 00:44:28,300
you'll probably be
relieved if I just go ahead

1122
00:44:28,300 --> 00:44:30,934
and push these three, right?

1123
00:44:30,934 --> 00:44:32,350
And I don't go
around and show you

1124
00:44:32,350 --> 00:44:34,660
all of the inert
buttons on the toy.

1125
00:44:34,660 --> 00:44:36,400
Because information is costly.

1126
00:44:36,400 --> 00:44:39,820
You have to sit there through
all those demonstrations.

1127
00:44:39,820 --> 00:44:44,110
In this experiment, I'll
show you Gweon's work.

1128
00:44:44,110 --> 00:44:46,210
We gave kids a common
ground condition

1129
00:44:46,210 --> 00:44:49,420
where everybody shared prior
knowledge, this abstract theory

1130
00:44:49,420 --> 00:44:52,810
you can use to constrain your
interpretation of this data.

1131
00:44:52,810 --> 00:44:56,519
In this condition, there
were two toy makers,

1132
00:44:56,519 --> 00:44:58,810
who are the informants, and
there's two naive learners,

1133
00:44:58,810 --> 00:45:00,520
Ernie and Bert.

1134
00:45:00,520 --> 00:45:02,410
And in the common
ground condition,

1135
00:45:02,410 --> 00:45:04,450
Ernie, and Bert,
and the toy makers

1136
00:45:04,450 --> 00:45:07,330
are all there while the
child explores and finds out

1137
00:45:07,330 --> 00:45:09,870
that only three buttons
work on these toys.

1138
00:45:09,870 --> 00:45:10,750
OK.

1139
00:45:10,750 --> 00:45:14,980
And then one teacher
shows of a brand new toy

1140
00:45:14,980 --> 00:45:17,550
just like this every single
button, the inert ones

1141
00:45:17,550 --> 00:45:18,620
and the non-inert ones.

1142
00:45:18,620 --> 00:45:22,010
And the other teacher shows
just the three working buttons.

1143
00:45:22,010 --> 00:45:22,566
Right?

1144
00:45:22,566 --> 00:45:24,190
And then we say, hey,
kids, guess what?

1145
00:45:24,190 --> 00:45:25,810
We have a whole closet
full of more of those toys.

1146
00:45:25,810 --> 00:45:27,601
One of these teachers
can show them to you.

1147
00:45:27,601 --> 00:45:29,440
Which one do you want?

1148
00:45:29,440 --> 00:45:31,480
Which one do you want, OK?

1149
00:45:31,480 --> 00:45:35,080
The other condition is almost
the same, but guess what?

1150
00:45:35,080 --> 00:45:39,100
The child explores
on their own, right?

1151
00:45:39,100 --> 00:45:42,910
And so there's no common ground
about what these toys do.

1152
00:45:42,910 --> 00:45:44,920
And then the teachers
do the same thing.

1153
00:45:44,920 --> 00:45:47,680
One teacher shows
exhaustive information,

1154
00:45:47,680 --> 00:45:51,970
and the other teacher pushes
only the three working buttons.

1155
00:45:51,970 --> 00:45:55,960
So in this case, that efficient
information, that less costly

1156
00:45:55,960 --> 00:45:58,330
simple demonstration
could mislead the learners

1157
00:45:58,330 --> 00:45:59,890
about what the
true hypothesis is.

1158
00:45:59,890 --> 00:46:02,940
They have a prior that all
these buttons ought to work.

1159
00:46:02,940 --> 00:46:05,680
Which toy maker would
you rather learn

1160
00:46:05,680 --> 00:46:08,800
from depends on whether
the learners share

1161
00:46:08,800 --> 00:46:13,120
that prior background
information or not.

1162
00:46:13,120 --> 00:46:16,690
And that was true not only
when the children were

1163
00:46:16,690 --> 00:46:20,950
judging the informants, but when
they were teaching themselves.

1164
00:46:20,950 --> 00:46:23,650
So again, these are
four and five-year-olds.

1165
00:46:23,650 --> 00:46:27,320
The children had
a condition where

1166
00:46:27,320 --> 00:46:31,350
Elmo got to see that only three
buttons worked on the toy,

1167
00:46:31,350 --> 00:46:32,920
and the condition
where Elmo didn't

1168
00:46:32,920 --> 00:46:36,220
get to see how many of those
buttons worked on the toy.

1169
00:46:36,220 --> 00:46:40,210
And then the children got
to teach Elmo the toy.

1170
00:46:40,210 --> 00:46:43,840
And the children were much more
likely to press more buttons

1171
00:46:43,840 --> 00:46:46,269
and provide exhaustive evidence
in the no common ground

1172
00:46:46,269 --> 00:46:48,060
condition than the
common ground condition.

1173
00:46:48,060 --> 00:46:49,780
So children themselves
are adjusting

1174
00:46:49,780 --> 00:46:52,339
the cost of the
information they provide

1175
00:46:52,339 --> 00:46:54,130
based on their prior
expectation about what

1176
00:46:54,130 --> 00:46:56,710
they think the learner is
going to learn from the data.

1177
00:46:56,710 --> 00:46:59,680
What did we do is, lastly,
tell you a little bit about how

1178
00:46:59,680 --> 00:47:05,500
the costs of information
are informative

1179
00:47:05,500 --> 00:47:08,184
not just in figuring out
what data to learn from

1180
00:47:08,184 --> 00:47:09,975
and how you should
communicate information,

1181
00:47:09,975 --> 00:47:12,310
but in actually figuring
out what people are doing,

1182
00:47:12,310 --> 00:47:14,650
and actually grounding out
ordinary, everyday theory

1183
00:47:14,650 --> 00:47:16,067
of mind.

1184
00:47:16,067 --> 00:47:18,400
I'm going to start with an
example you're familiar with.

1185
00:47:18,400 --> 00:47:19,850
I know you've seen this before.

1186
00:47:19,850 --> 00:47:23,710
This is an experiment by Gergely
and Csibra, a rational action.

1187
00:47:23,710 --> 00:47:27,010
This little ball jumps over the
wall to get to the momma ball--

1188
00:47:27,010 --> 00:47:28,390
you've all seen this?

1189
00:47:28,390 --> 00:47:30,190
I think Josh was
presenting it maybe?

1190
00:47:30,190 --> 00:47:30,690
OK.

1191
00:47:30,690 --> 00:47:32,314
And when you took
the wall away, babies

1192
00:47:32,314 --> 00:47:34,150
expect that ball to
take an efficient route.

1193
00:47:34,150 --> 00:47:37,210
So we think that rational agents
should take the most efficient

1194
00:47:37,210 --> 00:47:43,420
route to the goal, they should
maximize their overall utility.

1195
00:47:43,420 --> 00:47:46,720
But there's a lot of reasons
why that ball might have jumped

1196
00:47:46,720 --> 00:47:49,690
over the wall, and they
all have to do with costs

1197
00:47:49,690 --> 00:47:50,590
of rewards of action.

1198
00:47:50,590 --> 00:47:53,290
One might be, it was really
hard to get over the wall,

1199
00:47:53,290 --> 00:47:55,954
but really rewarding to do so.

1200
00:47:55,954 --> 00:47:57,370
Another reason,
though, is that it

1201
00:47:57,370 --> 00:47:59,828
was really easy to get over
the wall, so she might as well,

1202
00:47:59,828 --> 00:48:01,960
but she didn't care that
much about the reward.

1203
00:48:01,960 --> 00:48:04,300
These could have the
same net utility,

1204
00:48:04,300 --> 00:48:06,806
but psychologically, you really
care about the difference.

1205
00:48:06,806 --> 00:48:08,680
If we're talking about
the internal structure

1206
00:48:08,680 --> 00:48:10,600
of an agent's
motivations, you want

1207
00:48:10,600 --> 00:48:13,310
to decompose this
simple argument

1208
00:48:13,310 --> 00:48:15,460
about rational,
goal-directed action

1209
00:48:15,460 --> 00:48:18,670
into what the particular
costs and rewards are.

1210
00:48:18,670 --> 00:48:21,310
So Julian Jara-Ettinger
in our lab

1211
00:48:21,310 --> 00:48:23,200
has developed this
account he calls

1212
00:48:23,200 --> 00:48:26,650
the naive utility calculus,
which is our way of reasoning

1213
00:48:26,650 --> 00:48:28,030
about other peoples' actions.

1214
00:48:28,030 --> 00:48:31,500
Which is, we assume other agents
are acting to maximize utility,

1215
00:48:31,500 --> 00:48:34,180
but we care about the
internal structure also.

1216
00:48:34,180 --> 00:48:36,760
There are agent invariant
rewards and costs.

1217
00:48:36,760 --> 00:48:38,920
Two cookies is always
more than one cookie.

1218
00:48:38,920 --> 00:48:42,175
Higher hills are always higher
than lower hills for all of us.

1219
00:48:42,175 --> 00:48:44,505
But some of us are more
motivated to get cookies

1220
00:48:44,505 --> 00:48:45,880
than others of
us, and some of us

1221
00:48:45,880 --> 00:48:48,910
find hills more costly
than others of us.

1222
00:48:48,910 --> 00:48:50,980
So in addition to
these agents invariant

1223
00:48:50,980 --> 00:48:52,900
aspects of costs
and rewards, there's

1224
00:48:52,900 --> 00:48:55,420
also these internal
subjective things that

1225
00:48:55,420 --> 00:48:56,680
are harder to judge, right?

1226
00:48:56,680 --> 00:49:00,040
How competent you are,
what your values are,

1227
00:49:00,040 --> 00:49:01,600
and your preferences.

1228
00:49:01,600 --> 00:49:05,220
And so understanding how
all of these worked together

1229
00:49:05,220 --> 00:49:08,170
lets you take a very,
very simple analysis

1230
00:49:08,170 --> 00:49:11,420
and make surprisingly
powerful inferences about what

1231
00:49:11,420 --> 00:49:12,791
other agents are doing.

1232
00:49:12,791 --> 00:49:14,290
I'm going to show
you a few examples

1233
00:49:14,290 --> 00:49:17,380
and actually connect
it back to how children

1234
00:49:17,380 --> 00:49:19,150
are scientists in this regard.

1235
00:49:19,150 --> 00:49:21,340
So here is an
example experiment.

1236
00:49:21,340 --> 00:49:25,810
Here is Grover, and there
is a cracker and a cookie

1237
00:49:25,810 --> 00:49:27,300
down on this low shelf.

1238
00:49:27,300 --> 00:49:31,030
And Grover goes ahead,
and he chooses the cookie.

1239
00:49:31,030 --> 00:49:33,580
And now there's a cracker
and a cookie and this shelf,

1240
00:49:33,580 --> 00:49:36,850
and Grover goes ahead
and chooses the cracker.

1241
00:49:36,850 --> 00:49:40,360
And the question is, what
does Grover like better?

1242
00:49:40,360 --> 00:49:42,526
Well, if your read
out of preferences,

1243
00:49:42,526 --> 00:49:43,650
it's your actions you take.

1244
00:49:43,650 --> 00:49:44,990
It's your goal-directed action.

1245
00:49:44,990 --> 00:49:46,130
You should be at chance--

1246
00:49:46,130 --> 00:49:48,350
you chose a cracker
once and a cookie once.

1247
00:49:48,350 --> 00:49:50,090
But if even young
children understand,

1248
00:49:50,090 --> 00:49:52,120
no, it's not that simple, right?

1249
00:49:52,120 --> 00:49:55,180
You're not just acting
to maximize reward,

1250
00:49:55,180 --> 00:49:56,710
you're acting to
maximize utility.

1251
00:49:56,710 --> 00:49:58,480
You have to take the
costs into account.

1252
00:49:58,480 --> 00:50:00,670
Then clearly his
preference is what

1253
00:50:00,670 --> 00:50:03,520
he chose when the costs
were matched, right?

1254
00:50:03,520 --> 00:50:05,200
Not when the costs
were mismatched.

1255
00:50:05,200 --> 00:50:07,075
And the children should
say, no, no, no, it's

1256
00:50:07,075 --> 00:50:08,890
what he chose on the low box.

1257
00:50:08,890 --> 00:50:10,960
Which treat does he like best?

1258
00:50:10,960 --> 00:50:12,530
Is that clear?

1259
00:50:12,530 --> 00:50:16,540
And indeed, that is
what children do.

1260
00:50:16,540 --> 00:50:20,050
You can also introduce
a couple of characters.

1261
00:50:20,050 --> 00:50:22,390
Cookie Monster, who has a
strong preference for cookies,

1262
00:50:22,390 --> 00:50:25,170
and Grover, who is indifferent,
who likes them both.

1263
00:50:25,170 --> 00:50:27,850
So for Cookie Monster, the
reward value of cookies

1264
00:50:27,850 --> 00:50:28,890
is much higher.

1265
00:50:28,890 --> 00:50:30,980
For Grover, they're equivalent.

1266
00:50:30,980 --> 00:50:32,440
And now you can
set up a situation

1267
00:50:32,440 --> 00:50:36,340
where there's crackers in a low
box and cookies at a high box.

1268
00:50:36,340 --> 00:50:38,680
And you say, guys, go on,
you can make a choice.

1269
00:50:38,680 --> 00:50:40,510
And Grover chooses
the cracker and Cookie

1270
00:50:40,510 --> 00:50:43,140
Monster chooses the cracker.

1271
00:50:43,140 --> 00:50:47,120
And you say to the kids,
which puppet can't climb?

1272
00:50:47,120 --> 00:50:49,000
Now, no puppet has climbed.

1273
00:50:49,000 --> 00:50:50,380
No puppet has failed to climb.

1274
00:50:50,380 --> 00:50:52,660
No puppet has even
thought about climbing.

1275
00:50:52,660 --> 00:50:54,597
But the kids can know
the answer, right?

1276
00:50:54,597 --> 00:50:56,180
Because if Cookie
Monster could climb,

1277
00:50:56,180 --> 00:50:58,540
and he had a high reward,
then who would do it.

1278
00:50:58,540 --> 00:51:00,250
So you don't really
know about Grover,

1279
00:51:00,250 --> 00:51:03,220
but you can make an inference
that the costs were not

1280
00:51:03,220 --> 00:51:05,736
equivalent.

1281
00:51:05,736 --> 00:51:08,110
And by the way, in case you're
worried that kids are just

1282
00:51:08,110 --> 00:51:09,555
saying, well,
Cookie Monster, I've

1283
00:51:09,555 --> 00:51:10,930
been listening to
Michelle Obama.

1284
00:51:10,930 --> 00:51:13,210
You know, obesity and
fitness-- cookies aren't good,

1285
00:51:13,210 --> 00:51:14,160
I can't climb.

1286
00:51:14,160 --> 00:51:17,170
We ran the same experiment
with Grover and clover,

1287
00:51:17,170 --> 00:51:19,690
and Grover really likes
clovers, and Cookie Monster

1288
00:51:19,690 --> 00:51:21,280
likes them both.

1289
00:51:21,280 --> 00:51:23,830
And you now flip the
inference around.

1290
00:51:23,830 --> 00:51:31,300
OK, so they can consider how
the costs affect the inferences

1291
00:51:31,300 --> 00:51:33,459
they make.

1292
00:51:33,459 --> 00:51:35,500
All right, so let's bring
this all back together.

1293
00:51:35,500 --> 00:51:36,640
I've thrown a lot
of information--

1294
00:51:36,640 --> 00:51:38,500
probably too much
information-- at you

1295
00:51:38,500 --> 00:51:40,300
about how kids can
reason from sparse data,

1296
00:51:40,300 --> 00:51:41,710
about how they use
their theories to make

1297
00:51:41,710 --> 00:51:43,918
these inferences, about how
they use this in teaching

1298
00:51:43,918 --> 00:51:46,520
and learning in social context.

1299
00:51:46,520 --> 00:51:48,760
But if kids are sensitive
to these utilities

1300
00:51:48,760 --> 00:51:51,100
and the trade-offs,
then the kinds of things

1301
00:51:51,100 --> 00:51:54,140
that I showed you them doing
with beads and with machines,

1302
00:51:54,140 --> 00:51:56,930
they should also be able
to do in social context.

1303
00:51:56,930 --> 00:51:59,410
We believe psychology to
be something of a science,

1304
00:51:59,410 --> 00:52:01,049
as well as all of
these other sciences,

1305
00:52:01,049 --> 00:52:02,590
and they should be
able to apply some

1306
00:52:02,590 --> 00:52:04,840
of the same principles--
holding some things constant,

1307
00:52:04,840 --> 00:52:07,060
manipulating others-- in
order to gain information.

1308
00:52:07,060 --> 00:52:10,330
So we basically ask that
question of the children here.

1309
00:52:10,330 --> 00:52:13,270
Can they distinguish agents'
different competencies

1310
00:52:13,270 --> 00:52:17,280
and rewards by manipulating
the contexts that they see

1311
00:52:17,280 --> 00:52:19,100
and gaining information?

1312
00:52:19,100 --> 00:52:22,300
So here, we don't know if
Cookie Monster can climb.

1313
00:52:22,300 --> 00:52:24,910
So let's put one
treat on each box.

1314
00:52:24,910 --> 00:52:26,890
Where should we put
the treats to find out

1315
00:52:26,890 --> 00:52:28,720
if Cookie Monster can climb?

1316
00:52:28,720 --> 00:52:30,890
Now, only one of these
interventions is informative.

1317
00:52:30,890 --> 00:52:32,306
If the cookie is
down low, and you

1318
00:52:32,306 --> 00:52:34,490
know Cookie Monster
prefers cookies,

1319
00:52:34,490 --> 00:52:36,420
then you're not going
to get any information.

1320
00:52:36,420 --> 00:52:38,260
But if the cookie
is up high, then you

1321
00:52:38,260 --> 00:52:40,360
are going to get
information, right?

1322
00:52:40,360 --> 00:52:44,460
And in fact, the kids are
overwhelmingly good at this.

1323
00:52:44,460 --> 00:52:46,840
In case you think, well,
treats should be put up high--

1324
00:52:46,840 --> 00:52:48,798
and this is also true,
by the way, for clovers,

1325
00:52:48,798 --> 00:52:49,880
but it's all right.

1326
00:52:49,880 --> 00:52:51,820
In case that you think they
just have a heuristic like,

1327
00:52:51,820 --> 00:52:52,810
oh, well, let's
put treats up high,

1328
00:52:52,810 --> 00:52:54,518
you can ask the question
a different way.

1329
00:52:54,518 --> 00:52:57,430
You can say, both of our
friends can climb the short box.

1330
00:52:57,430 --> 00:53:00,210
But only one of our friends
can climb the tall box,

1331
00:53:00,210 --> 00:53:01,870
and we don't know which one.

1332
00:53:01,870 --> 00:53:04,492
So let's put the cookie up
here and the cracker down here.

1333
00:53:04,492 --> 00:53:06,700
And if we want to figure
out which one of our friends

1334
00:53:06,700 --> 00:53:10,330
can climb, which friend
should we send in?

1335
00:53:10,330 --> 00:53:12,512
Well, Grover has no
particular incentive to climb.

1336
00:53:12,512 --> 00:53:13,720
He could just do the cracker.

1337
00:53:13,720 --> 00:53:16,210
But Cookie Monster, he
has an incentive to climb.

1338
00:53:16,210 --> 00:53:18,820
You should probably
send in Cookie Monster.

1339
00:53:18,820 --> 00:53:21,370
And again, these are
the kinds of inferences

1340
00:53:21,370 --> 00:53:23,723
that young children can make.

1341
00:53:23,723 --> 00:53:24,569
Is that clear?

1342
00:53:24,569 --> 00:53:25,840
All right.

1343
00:53:25,840 --> 00:53:31,000
So, end of a long winded talk.

1344
00:53:31,000 --> 00:53:32,890
I want to return it to this.

1345
00:53:32,890 --> 00:53:35,560
So I framed it in terms of
the way I've been increasingly

1346
00:53:35,560 --> 00:53:38,085
thinking, and the projects
that we're increasingly

1347
00:53:38,085 --> 00:53:39,460
moving towards,
which is thinking

1348
00:53:39,460 --> 00:53:43,080
not just about the pure pursuit
of knowledge and information,

1349
00:53:43,080 --> 00:53:45,634
but how do you pursue
information in a complex world

1350
00:53:45,634 --> 00:53:47,800
where it's not just your
own individual exploration.

1351
00:53:47,800 --> 00:53:49,060
You get it in a
social context, you

1352
00:53:49,060 --> 00:53:50,351
get in interaction with others.

1353
00:53:50,351 --> 00:53:55,120
The information is costly both
to deliver and to process.

1354
00:53:55,120 --> 00:53:57,850
But that, those costs
themselves, are information,

1355
00:53:57,850 --> 00:54:01,030
and you can use them to
make sense of the world.

1356
00:54:01,030 --> 00:54:04,750
And so I want to sort of bring
these together and come back

1357
00:54:04,750 --> 00:54:08,620
to a problem that I posed a
bit earlier, which is, I think,

1358
00:54:08,620 --> 00:54:12,130
I hope I've made the case
that what kids are doing

1359
00:54:12,130 --> 00:54:15,700
is not reasoning about
huge sets of data.

1360
00:54:15,700 --> 00:54:19,300
What kids are doing is, they're
taking some very abstract

1361
00:54:19,300 --> 00:54:22,480
structure knowledge and using
it to constrain their inferences

1362
00:54:22,480 --> 00:54:24,550
about tiny amounts of data.

1363
00:54:24,550 --> 00:54:27,100
Cookie Monster this, a
couple of trials of evidence.

1364
00:54:27,100 --> 00:54:29,560
And then they make
good inductive guesses,

1365
00:54:29,560 --> 00:54:34,420
which are sometimes
wrong, but they are good.

1366
00:54:34,420 --> 00:54:38,380
And I said, these
abstract representations,

1367
00:54:38,380 --> 00:54:40,100
not all of them
are innate, right?

1368
00:54:40,100 --> 00:54:42,670
You've seen beautiful
evidence of the many that are,

1369
00:54:42,670 --> 00:54:44,020
but a lot of them aren't.

1370
00:54:44,020 --> 00:54:45,130
A lot of them aren't.

1371
00:54:45,130 --> 00:54:47,588
A lot of the things that govern
your common sense knowledge

1372
00:54:47,588 --> 00:54:49,210
every day, how do you get those?

1373
00:54:49,210 --> 00:54:52,610
And how do you get those
from tiny amounts of data?

1374
00:54:52,610 --> 00:54:57,320
This is a problem that bugged
me deeply for a very long time,

1375
00:54:57,320 --> 00:55:03,670
and I think there's been a real
leap and a very exciting ascent

1376
00:55:03,670 --> 00:55:07,210
of how it is actually possible
to use tiny amounts of data

1377
00:55:07,210 --> 00:55:09,154
to make really rich
abstract inferences, which

1378
00:55:09,154 --> 00:55:10,570
then constrain
your interpretation

1379
00:55:10,570 --> 00:55:11,290
of subsequent data.

1380
00:55:11,290 --> 00:55:13,000
I think that is a really
important problem.

1381
00:55:13,000 --> 00:55:14,230
And with that, I'm
going to turn it over

1382
00:55:14,230 --> 00:55:16,490
to Josh and Tomer, who
maybe can tell you how

1383
00:55:16,490 --> 00:55:17,990
that is going to actually work.

1384
00:55:17,990 --> 00:55:20,830
So thanks to everyone
here at Woods Hole.