1
00:00:00,000 --> 00:00:04,030


2
00:00:04,030 --> 00:00:07,780
Introductions-- I'm
Heather, obviously.

3
00:00:07,780 --> 00:00:09,610
And I'm Candace Sidner.

4
00:00:09,610 --> 00:00:13,180
And you're working
right now at--

5
00:00:13,180 --> 00:00:16,720
At Mitsubishi Labs in Cambridge,
across from Tech Square.

6
00:00:16,720 --> 00:00:17,220
Awesome.

7
00:00:17,220 --> 00:00:18,820
Across from Tech Square, great.

8
00:00:18,820 --> 00:00:21,800
And how did you first get
into artificial intelligence

9
00:00:21,800 --> 00:00:23,540
or in robotics?

10
00:00:23,540 --> 00:00:27,220
Well, I started off in--

11
00:00:27,220 --> 00:00:30,370
as an undergraduate, I was
interested in how people

12
00:00:30,370 --> 00:00:32,380
communicate and use language.

13
00:00:32,380 --> 00:00:36,490
And I did my master's degree at
the University of Pittsburgh,

14
00:00:36,490 --> 00:00:39,710
and while I was there I took
courses with Herb Simon.

15
00:00:39,710 --> 00:00:40,210
Cool.

16
00:00:40,210 --> 00:00:42,560
And when I started doing
the master's thesis,

17
00:00:42,560 --> 00:00:45,070
Herb said, well, go see this
former student of mine whose

18
00:00:45,070 --> 00:00:47,200
name is [INAUDIBLE]
who's somewhere

19
00:00:47,200 --> 00:00:49,300
at the University of
Pittsburgh, and talk to him.

20
00:00:49,300 --> 00:00:52,090
And Harry Pople was at the
[INAUDIBLE] business school.

21
00:00:52,090 --> 00:00:56,750
Harry was doing what was called,
the INTERNIST medical system,

22
00:00:56,750 --> 00:00:59,970
which was a very complicated
and sophisticated AI

23
00:00:59,970 --> 00:01:03,640
program to mimic the common
decision-making that doctors

24
00:01:03,640 --> 00:01:07,920
did when they were
diagnosing patients.

25
00:01:07,920 --> 00:01:09,800
And so I did my master's
thesis with him.

26
00:01:09,800 --> 00:01:12,620
And then after my
thesis was finished,

27
00:01:12,620 --> 00:01:15,970
I was admitted to the doctoral
program at MIT at the AI lab.

28
00:01:15,970 --> 00:01:18,880


29
00:01:18,880 --> 00:01:23,273
And when did you
get into emotional--

30
00:01:23,273 --> 00:01:24,606
what are you mainly researching?

31
00:01:24,606 --> 00:01:25,840
So I don't missay.

32
00:01:25,840 --> 00:01:27,430


33
00:01:27,430 --> 00:01:31,360
So my work has largely been
in the area of human language

34
00:01:31,360 --> 00:01:33,820
and human conversation.

35
00:01:33,820 --> 00:01:36,030
So over the years,
I've looked a lot

36
00:01:36,030 --> 00:01:40,140
at how we model the
intentions of people

37
00:01:40,140 --> 00:01:42,060
when they're communicating
with one another,

38
00:01:42,060 --> 00:01:44,940
as well as what they do
with their language barrier.

39
00:01:44,940 --> 00:01:48,370
I've done a lot of work on
how we understand pronouns,

40
00:01:48,370 --> 00:01:51,340
how we understand all kinds
of referring expressions,

41
00:01:51,340 --> 00:01:56,950
and how we embed our
intentions in our utterances,

42
00:01:56,950 --> 00:02:00,590
and then convey that
over successive parts

43
00:02:00,590 --> 00:02:03,140
of the conversation.

44
00:02:03,140 --> 00:02:05,030
A few years ago,
I got interested

45
00:02:05,030 --> 00:02:08,120
in the problem of how we
use gestures and language.

46
00:02:08,120 --> 00:02:11,090
So obviously, one of them
is what I do with my hands.

47
00:02:11,090 --> 00:02:12,523
But equally
important is the fact

48
00:02:12,523 --> 00:02:13,940
that I'm looking
at you right now,

49
00:02:13,940 --> 00:02:16,730
and I'm looking at the
camera, for example.

50
00:02:16,730 --> 00:02:19,190
And that when I'm talking to
you, I actually talk to you.

51
00:02:19,190 --> 00:02:21,943
I don't talk to the wall
like this, because if I did,

52
00:02:21,943 --> 00:02:23,360
there's something
very importantly

53
00:02:23,360 --> 00:02:25,010
wrong with that interaction.

54
00:02:25,010 --> 00:02:26,960
I call that the
process of engagement.

55
00:02:26,960 --> 00:02:30,860
That is how we come
to perceive connection

56
00:02:30,860 --> 00:02:34,490
through nonverbal
means and interactions.

57
00:02:34,490 --> 00:02:35,660
There's also a verbal means.

58
00:02:35,660 --> 00:02:39,060
We do it, of course, as long as
we're having the conversation.

59
00:02:39,060 --> 00:02:41,000
Even if I'm doing
something else,

60
00:02:41,000 --> 00:02:43,340
you know that I
intend to be connected

61
00:02:43,340 --> 00:02:45,200
to you because I
continue to play

62
00:02:45,200 --> 00:02:46,600
my part in the conversation.

63
00:02:46,600 --> 00:02:53,320
But if I violate certain aspects
of our non-verbal connection,

64
00:02:53,320 --> 00:02:54,490
that can be very odd.

65
00:02:54,490 --> 00:02:55,870
Some of it's OK, however.

66
00:02:55,870 --> 00:02:59,308
So for example, if I'm
washing the dishes,

67
00:02:59,308 --> 00:03:00,850
I have to pay
attention to the dishes

68
00:03:00,850 --> 00:03:02,200
while I'm talking to you.

69
00:03:02,200 --> 00:03:05,730
I will pay attention to this,
and I'm only occasionally

70
00:03:05,730 --> 00:03:08,220
looking up as I have
something to say.

71
00:03:08,220 --> 00:03:10,920
And if you're talking,
instead I'll be doing "mhmm",

72
00:03:10,920 --> 00:03:12,780
and I'm doing when I
was doing the "mm".

73
00:03:12,780 --> 00:03:14,830
I may not look at you
at that point in time.

74
00:03:14,830 --> 00:03:19,350
I may be able to do this, but
you'll understand that I'm now

75
00:03:19,350 --> 00:03:21,260
multitasking as we are and--

76
00:03:21,260 --> 00:03:23,692
So can you learn those things?

77
00:03:23,692 --> 00:03:25,650
At this stage do they
have to be preprogrammed?

78
00:03:25,650 --> 00:03:26,020
No.

79
00:03:26,020 --> 00:03:28,400
Those are the sorts of things
we still have to program.

80
00:03:28,400 --> 00:03:32,270
Just getting clear about
what the [INAUDIBLE]..

81
00:03:32,270 --> 00:03:35,330
And the way I've been
testing a lot of those ideas

82
00:03:35,330 --> 00:03:39,110
is to have a robot that
can converse with the user

83
00:03:39,110 --> 00:03:43,420
and can both produce
proper behavior

84
00:03:43,420 --> 00:03:46,020
and then interpret the
behavior of the person.

85
00:03:46,020 --> 00:03:50,970
So that means the robot is not
so important that it's mobile.

86
00:03:50,970 --> 00:03:53,840
Up until now, we haven't really
done work on mobile robotics.

87
00:03:53,840 --> 00:03:58,512
But more importantly,

88
00:03:58,512 --> 00:04:00,220
it has to move in the
right sorts of way,

89
00:04:00,220 --> 00:04:01,930
and it also has to
have the ability

90
00:04:01,930 --> 00:04:05,250
to interpret people's
movements, as much

91
00:04:05,250 --> 00:04:07,330
as you can do that with
current vision systems.

92
00:04:07,330 --> 00:04:12,295
So is your main goal more to
interpret human expression

93
00:04:12,295 --> 00:04:16,290
or to have humanoid robots that
can replicate them, or both?

94
00:04:16,290 --> 00:04:18,089
Well, you have to do both.

95
00:04:18,089 --> 00:04:20,268
Conversation is a two way thing.

96
00:04:20,268 --> 00:04:22,560
You both have to produce it
properly, and you have to--

97
00:04:22,560 --> 00:04:24,700
I was just thinking of an
intelligent computer that

98
00:04:24,700 --> 00:04:26,740
was just trying to
interpret your work

99
00:04:26,740 --> 00:04:29,250
or something like that, where
you're getting frustrated.

100
00:04:29,250 --> 00:04:30,125
That might be useful.

101
00:04:30,125 --> 00:04:33,280


102
00:04:33,280 --> 00:04:35,640
I'm interested both in the
interpretation phase and also

103
00:04:35,640 --> 00:04:41,060
the ability to generate or
produce that kind of behavior.

104
00:04:41,060 --> 00:04:43,760
What do you think the
biggest challenges in AI

105
00:04:43,760 --> 00:04:47,930
historically have been, and
how has that changed, now?

106
00:04:47,930 --> 00:04:52,268
Well, one of the challenges
that really hasn't changed

107
00:04:52,268 --> 00:04:54,060
is something that I
talked about yesterday,

108
00:04:54,060 --> 00:04:58,180
which is the need for a deep
model of semantics of language.

109
00:04:58,180 --> 00:05:04,177
So we now are able to say things
about how a verb might have

110
00:05:04,177 --> 00:05:06,010
certain arguments, and
you can fill them in,

111
00:05:06,010 --> 00:05:08,710
and you can do some of
that adequate language

112
00:05:08,710 --> 00:05:09,730
understanding.

113
00:05:09,730 --> 00:05:13,750
But the much deeper problems
of how we know certain things

114
00:05:13,750 --> 00:05:18,460
about how sentences go together
in terms of their meaning,

115
00:05:18,460 --> 00:05:22,500
how we make deeper
inferences about what's

116
00:05:22,500 --> 00:05:26,240
going on between
utterances or even

117
00:05:26,240 --> 00:05:30,560
between practical sentences and
how that's connected onto what

118
00:05:30,560 --> 00:05:36,590
we know in general is still
relatively poorly understood.

119
00:05:36,590 --> 00:05:39,970
And so I think that's a
pretty important problem.

120
00:05:39,970 --> 00:05:42,480
Obviously, the work I've
started doing on gesture

121
00:05:42,480 --> 00:05:45,600
is only in its beginning
phases up until now.

122
00:05:45,600 --> 00:05:47,970
Until very recently-- and I'm
about the only person even

123
00:05:47,970 --> 00:05:50,100
working with Siri-- but
up until very recently,

124
00:05:50,100 --> 00:05:53,690
like maybe five years
ago, nobody thought to.

125
00:05:53,690 --> 00:05:55,440
And it's not that
nobody thought about it.

126
00:05:55,440 --> 00:05:58,070
Nobody really had a
good way to go about it,

127
00:05:58,070 --> 00:05:59,420
and that's gradually changed.

128
00:05:59,420 --> 00:06:01,480
What do you think of
Cynthia Breazeal's work?

129
00:06:01,480 --> 00:06:01,980
Oh!

130
00:06:01,980 --> 00:06:04,920
I like Cynthia's
work, obviously.

131
00:06:04,920 --> 00:06:08,665
In fact, some of her students
have been my summer interns.

132
00:06:08,665 --> 00:06:11,817
So we've happily
shared back and forth

133
00:06:11,817 --> 00:06:13,900
and I hope learned from
each other in the process.

134
00:06:13,900 --> 00:06:17,300


135
00:06:17,300 --> 00:06:21,570
So another area, of course, that
I think is largely unexplored

136
00:06:21,570 --> 00:06:23,730
is things that have to do
with social and emotional

137
00:06:23,730 --> 00:06:25,740
interactions.

138
00:06:25,740 --> 00:06:27,900
A number of years ago,
there were a few people

139
00:06:27,900 --> 00:06:33,445
in the language community who
explored emotional stuff, texts

140
00:06:33,445 --> 00:06:34,320
and things like that.

141
00:06:34,320 --> 00:06:39,090
But it kind of died off,
and looking at this stuff

142
00:06:39,090 --> 00:06:42,600
any further or any deeper
has really not happened.

143
00:06:42,600 --> 00:06:44,940
So the reason I
think it's important

144
00:06:44,940 --> 00:06:47,910
is that when we
communicate together

145
00:06:47,910 --> 00:06:52,350
in a human to human situation,
our emotional content

146
00:06:52,350 --> 00:06:54,300
is expressed both
in our language

147
00:06:54,300 --> 00:06:57,150
and in our gestural behavior--
mostly in our faces, I think.

148
00:06:57,150 --> 00:06:59,085
Although if we're really
upset, obviously it

149
00:06:59,085 --> 00:07:00,270
comes up in that way.

150
00:07:00,270 --> 00:07:03,900
And it's useful
to be able to take

151
00:07:03,900 --> 00:07:07,080
advantage of that information
on the part of the machine.

152
00:07:07,080 --> 00:07:09,540
It may also turn out to be
useful to have the machine be

153
00:07:09,540 --> 00:07:10,830
able to express emotions.

154
00:07:10,830 --> 00:07:12,405
You can affect the [INAUDIBLE].

155
00:07:12,405 --> 00:07:14,880


156
00:07:14,880 --> 00:07:16,480
More subtle but
equally important

157
00:07:16,480 --> 00:07:18,040
is our social interactions.

158
00:07:18,040 --> 00:07:24,270
And recent researchers have
shown that when we have social

159
00:07:24,270 --> 00:07:29,100
conversations-- not just about
putting things together--

160
00:07:29,100 --> 00:07:34,650
with machines, people have
a markedly different sense

161
00:07:34,650 --> 00:07:38,178
of what that entity is.

162
00:07:38,178 --> 00:07:42,010
You can measure that on
standard psychological measures,

163
00:07:42,010 --> 00:07:43,640
which is pretty interesting.

164
00:07:43,640 --> 00:07:45,640
So it's interesting,
and I always

165
00:07:45,640 --> 00:07:47,590
write it as a kind
of practical thing

166
00:07:47,590 --> 00:07:49,020
you want to do with a computer.

167
00:07:49,020 --> 00:07:51,270
But from the point of view
of artificial intelligence,

168
00:07:51,270 --> 00:07:55,508
it's really kind of interesting
and important [INAUDIBLE]..

169
00:07:55,508 --> 00:07:57,300
What do you think the
most important skills

170
00:07:57,300 --> 00:08:01,230
for an artificial intelligence
researcher to have are today?

171
00:08:01,230 --> 00:08:02,010
Researchers?

172
00:08:02,010 --> 00:08:04,790
Wow.

173
00:08:04,790 --> 00:08:06,690
Really good
observational skills.

174
00:08:06,690 --> 00:08:08,570
So I think it's really
important to look

175
00:08:08,570 --> 00:08:11,180
at what the human-human
situation is

176
00:08:11,180 --> 00:08:15,030
and to be able to
observe that carefully.

177
00:08:15,030 --> 00:08:15,630
Very good

178
00:08:15,630 --> 00:08:18,210


179
00:08:18,210 --> 00:08:20,840
analytic skills
in terms of being

180
00:08:20,840 --> 00:08:22,880
able to quantify and
think about things

181
00:08:22,880 --> 00:08:26,570
and really analyze
what's going on.

182
00:08:26,570 --> 00:08:31,220
The ability to listen carefully
and read other people's work.

183
00:08:31,220 --> 00:08:33,559
And then, to be
creative and kind of

184
00:08:33,559 --> 00:08:36,718
be able to step off the edge.

185
00:08:36,718 --> 00:08:38,260
You can't make
progress if you're not

186
00:08:38,260 --> 00:08:41,650
willing to take risks.

187
00:08:41,650 --> 00:08:43,520
Well, we're sort
of short on time.

188
00:08:43,520 --> 00:08:47,140
But do you have any
anecdotes or stories

189
00:08:47,140 --> 00:08:49,360
that you remember
from your times at MIT

190
00:08:49,360 --> 00:08:52,000
or pathways along the way?

191
00:08:52,000 --> 00:08:52,750
Oh, wow.

192
00:08:52,750 --> 00:08:55,360
Well, a lot of my
anecdotes have to do

193
00:08:55,360 --> 00:08:58,570
with being a woman in a
highly male profession.

194
00:08:58,570 --> 00:09:02,440
And it's not always been
easy, then, to say the least.

195
00:09:02,440 --> 00:09:04,576
So I don't know if
I want to convey

196
00:09:04,576 --> 00:09:06,820
all them at this
particular point in time,

197
00:09:06,820 --> 00:09:08,910
but there's a certain
challenge to being

198
00:09:08,910 --> 00:09:11,490
a woman in an all male field.

199
00:09:11,490 --> 00:09:13,979
It has certainly colored
my professional life

200
00:09:13,979 --> 00:09:17,850
and my experiences
as a scientist.

201
00:09:17,850 --> 00:09:18,350
Cool.

202
00:09:18,350 --> 00:09:20,090
Well, a pleasure talking to you.

203
00:09:20,090 --> 00:09:23,690


204
00:09:23,690 --> 00:09:25,540
Let's go.