1
00:00:00,000 --> 00:00:06,970


2
00:00:06,970 --> 00:00:08,770
I guess I'll have you
introduce yourself.

3
00:00:08,770 --> 00:00:12,400
Yeah, I'm Bart Selman, currently
at the Department of Computer

4
00:00:12,400 --> 00:00:14,890
Science at Cornell University.

5
00:00:14,890 --> 00:00:15,730
Cool.

6
00:00:15,730 --> 00:00:18,580
What first got you
into the field of AI?

7
00:00:18,580 --> 00:00:20,920
I guess I studied physics.

8
00:00:20,920 --> 00:00:23,080
Actually, I did a masters
in physics in Holland

9
00:00:23,080 --> 00:00:28,360
in the early '80s,
and I took a course

10
00:00:28,360 --> 00:00:32,259
on artificial intelligence
in the philosophy department.

11
00:00:32,259 --> 00:00:35,140
So that got me familiar
with AI, and I was always

12
00:00:35,140 --> 00:00:39,550
interested in human thought
processes and cognition.

13
00:00:39,550 --> 00:00:42,630
And so after I finished my
masters in physics in Holland,

14
00:00:42,630 --> 00:00:44,890
I got a scholarship from
the Canadian government,

15
00:00:44,890 --> 00:00:45,640
and I studied--

16
00:00:45,640 --> 00:00:49,190
I went to Toronto and
did my masters and PhD

17
00:00:49,190 --> 00:00:52,607
in artificial intelligence.

18
00:00:52,607 --> 00:00:54,940
What kind of problems were
you working on when you first

19
00:00:54,940 --> 00:00:55,850
entered the field?

20
00:00:55,850 --> 00:00:57,267
So when I first
entered the field,

21
00:00:57,267 --> 00:00:58,940
I did a masters on
actually connections

22
00:00:58,940 --> 00:01:00,710
to natural language
understanding,

23
00:01:00,710 --> 00:01:01,780
so neural networks.

24
00:01:01,780 --> 00:01:06,580
And at the end of that,
actually, Hector Levesque

25
00:01:06,580 --> 00:01:11,110
came to Toronto, and he got
me interested in reasoning

26
00:01:11,110 --> 00:01:14,260
and complexity issues,
conditional complexity issues

27
00:01:14,260 --> 00:01:15,760
in artificial intelligence.

28
00:01:15,760 --> 00:01:18,820
And that's what I
started working on then.

29
00:01:18,820 --> 00:01:21,350
What have you worked on since?

30
00:01:21,350 --> 00:01:24,730
So my work started--

31
00:01:24,730 --> 00:01:28,632
in the mid '80s, there
was a movement to study--

32
00:01:28,632 --> 00:01:30,340
people started to
realize a lot of things

33
00:01:30,340 --> 00:01:32,298
we were trying to do in
artificial intelligence

34
00:01:32,298 --> 00:01:34,585
were computationally
intractable and complete,

35
00:01:34,585 --> 00:01:36,490
basically complete.

36
00:01:36,490 --> 00:01:41,780
And there was a whole effort
in identifying what could

37
00:01:41,780 --> 00:01:43,030
be done in a tractable manner.

38
00:01:43,030 --> 00:01:47,050
So people defined restricted
representation languages,

39
00:01:47,050 --> 00:01:51,610
restricted emotionalization
languages, and restricted forms

40
00:01:51,610 --> 00:01:53,660
of default reasoning.

41
00:01:53,660 --> 00:01:55,420
So this was fairly
formal work, where

42
00:01:55,420 --> 00:01:59,770
you would prove properties about
languages and about inference

43
00:01:59,770 --> 00:02:03,490
and try to find
tractable classes.

44
00:02:03,490 --> 00:02:06,140
And that was my PhD work.

45
00:02:06,140 --> 00:02:11,410
And then I finished 1990, and
I went to the laboratories

46
00:02:11,410 --> 00:02:13,410
and actually started
working with Henry Kautz.

47
00:02:13,410 --> 00:02:16,310


48
00:02:16,310 --> 00:02:19,000
Do you have any fun
stories or anecdotes

49
00:02:19,000 --> 00:02:20,590
you liked from Toronto?

50
00:02:20,590 --> 00:02:23,230
From Toronto?

51
00:02:23,230 --> 00:02:26,100
Fun stories.

52
00:02:26,100 --> 00:02:28,230
No, it was just a
very exciting time.

53
00:02:28,230 --> 00:02:31,070
I think, with the Canadian
government-- it was

54
00:02:31,070 --> 00:02:33,830
a private foundation, I think.

55
00:02:33,830 --> 00:02:37,450
They brought back a number
of prominent researchers,

56
00:02:37,450 --> 00:02:42,530
kind of an answer of a
brain drain going to the US.

57
00:02:42,530 --> 00:02:47,810
So they set up these chairs
in artificial intelligence,

58
00:02:47,810 --> 00:02:52,100
and so they brought
back Hector Levesque.

59
00:02:52,100 --> 00:02:54,560
Actually, Geoff
Hinton came later on,

60
00:02:54,560 --> 00:02:58,490
and he attracted some
big names to Toronto.

61
00:02:58,490 --> 00:03:01,440
So it was an exciting time, and
they didn't have to teach much,

62
00:03:01,440 --> 00:03:05,150
so we could spend all
our time on research.

63
00:03:05,150 --> 00:03:06,200
So it was a very--

64
00:03:06,200 --> 00:03:09,320
these were the highlights,
the high points

65
00:03:09,320 --> 00:03:10,730
of artificial intelligence.

66
00:03:10,730 --> 00:03:17,120
I remember, I think it
was the '84 and '85 IJCAI

67
00:03:17,120 --> 00:03:20,420
when Hector won the
Computer Thought Award.

68
00:03:20,420 --> 00:03:22,750
I think he gave a
talk for 5,000 people,

69
00:03:22,750 --> 00:03:26,348
and that was quite
something in the hall.

70
00:03:26,348 --> 00:03:28,640
This was in the same place
where they had the Olympics,

71
00:03:28,640 --> 00:03:30,320
I guess, the year before.

72
00:03:30,320 --> 00:03:32,600
The whole UCLA campus
was full, and this

73
00:03:32,600 --> 00:03:36,340
was a lot of press and a lot
of hype basically about AI.

74
00:03:36,340 --> 00:03:38,810
So it was the high point.

75
00:03:38,810 --> 00:03:39,693
Cool.

76
00:03:39,693 --> 00:03:42,110
So I've seen all the ups and
downs, I guess, on the field.

77
00:03:42,110 --> 00:03:44,930
And it's interesting how the
field now is in an upswing

78
00:03:44,930 --> 00:03:46,480
again.

79
00:03:46,480 --> 00:03:48,320
So arriving in Toronto
when there were all

80
00:03:48,320 --> 00:03:54,200
these other real artificial
intelligence minds coming,

81
00:03:54,200 --> 00:03:56,400
where did you guys see
artificial intelligence going

82
00:03:56,400 --> 00:03:59,045
in the next 20 years
or so from then?

83
00:03:59,045 --> 00:04:01,970
It was moving fast.

84
00:04:01,970 --> 00:04:05,330
There was a strong belief at
that point in sort of logical

85
00:04:05,330 --> 00:04:10,130
and the reasoning-based
approaches and people

86
00:04:10,130 --> 00:04:14,130
and cognitive reasoning,
and those kind of areas.

87
00:04:14,130 --> 00:04:16,610
Geoff Hinton, of course, the
neural network perspective

88
00:04:16,610 --> 00:04:18,890
and the learning perspective.

89
00:04:18,890 --> 00:04:22,520
I think sort of what people
realized in the next decade,

90
00:04:22,520 --> 00:04:26,300
roughly from '85 to '95,
is the limitations of some

91
00:04:26,300 --> 00:04:30,790
of these approaches and that it
wasn't as easy as we thought.

92
00:04:30,790 --> 00:04:33,110
And so people went
back to the basics,

93
00:04:33,110 --> 00:04:36,840
again, trying to fix
some of these problems.

94
00:04:36,840 --> 00:04:38,215
So it's all very
hard to predict.

95
00:04:38,215 --> 00:04:40,048
I think one of the key
things in artificial,

96
00:04:40,048 --> 00:04:42,870
it's very hard to predict where
the next breakthrough will be.

97
00:04:42,870 --> 00:04:43,370
Sure.

98
00:04:43,370 --> 00:04:46,160
And I'm sure we're actually
going again at current times.

99
00:04:46,160 --> 00:04:48,500
But those people always try.

100
00:04:48,500 --> 00:04:50,570
And I think what brought
us all together is

101
00:04:50,570 --> 00:04:53,070
sort of we come to
science with a desire

102
00:04:53,070 --> 00:04:55,910
to figure out thinking,
how people think,

103
00:04:55,910 --> 00:04:59,390
and understanding mental
images, how people manipulate

104
00:04:59,390 --> 00:05:00,775
mental images to do thinking.

105
00:05:00,775 --> 00:05:04,080


106
00:05:04,080 --> 00:05:06,860
But at that point, it was
a very optimistic period.

107
00:05:06,860 --> 00:05:10,520
And I think that went
away for a while.

108
00:05:10,520 --> 00:05:12,170
And now, it's coming back.

109
00:05:12,170 --> 00:05:14,510
I was going to ask if
you've done any work

110
00:05:14,510 --> 00:05:17,540
in computational neuroscience.

111
00:05:17,540 --> 00:05:20,390
Also, for my master's, not
literally neuroscience,

112
00:05:20,390 --> 00:05:24,020
but we did build a parser that
was actually using a Boltzmann

113
00:05:24,020 --> 00:05:26,690
machine to disambiguate.

114
00:05:26,690 --> 00:05:28,430
You give a little
input sentence,

115
00:05:28,430 --> 00:05:30,460
and then it would
settle on one or two.

116
00:05:30,460 --> 00:05:33,090
If it had multiple
parses, it would settle

117
00:05:33,090 --> 00:05:34,970
on possible parsing states.

118
00:05:34,970 --> 00:05:35,480
I see.

119
00:05:35,480 --> 00:05:39,700
And on small signs and
basic examples, it worked.

120
00:05:39,700 --> 00:05:40,700
So that was sort of fun.

121
00:05:40,700 --> 00:05:42,080
OK.

122
00:05:42,080 --> 00:05:43,340
That's cool.

123
00:05:43,340 --> 00:05:44,900
What are you working
on these days?

124
00:05:44,900 --> 00:05:47,430
So basically, when
I went to Bell Labs,

125
00:05:47,430 --> 00:05:49,020
we were sort of in front of it.

126
00:05:49,020 --> 00:05:52,070
That was actually the
hey day for Bell Labs.

127
00:05:52,070 --> 00:05:54,770
Ron Brachman started the
artificial intelligence center

128
00:05:54,770 --> 00:05:55,370
there.

129
00:05:55,370 --> 00:05:58,960
And many, many very
good people went there

130
00:05:58,960 --> 00:06:01,940
in a short period
of time, Bell Labs.

131
00:06:01,940 --> 00:06:02,660
Arno Penzias.

132
00:06:02,660 --> 00:06:06,080
I remember we presented that
work to the Nobel Prize winner

133
00:06:06,080 --> 00:06:08,320
in charge of Bell
Labs, Arno Penzias.

134
00:06:08,320 --> 00:06:11,750
And he liked AI and wanted to
see sort of the latest things.

135
00:06:11,750 --> 00:06:13,430
I remember him showing--

136
00:06:13,430 --> 00:06:16,020
we had very stochastic
search algorithms,

137
00:06:16,020 --> 00:06:19,190
and we did sort of standard
deterministic search of them

138
00:06:19,190 --> 00:06:21,200
for logical inference.

139
00:06:21,200 --> 00:06:24,135
We actually did stochastic
local search management

140
00:06:24,135 --> 00:06:25,220
for inference.

141
00:06:25,220 --> 00:06:27,890
And he could solve
very large problems.

142
00:06:27,890 --> 00:06:30,880
And we would show it to him, and
he would take a little video.

143
00:06:30,880 --> 00:06:33,410
We had a little graphical
demonstration of that.

144
00:06:33,410 --> 00:06:36,430
And we would show the general
audience to understand science.

145
00:06:36,430 --> 00:06:39,340
And so that was
an exciting time.

146
00:06:39,340 --> 00:06:41,740
And I shifted sort
of my emphasis

147
00:06:41,740 --> 00:06:46,893
from looking at restricted
formalisms and [INAUDIBLE]

148
00:06:46,893 --> 00:06:48,310
and started working
on [INAUDIBLE]

149
00:06:48,310 --> 00:06:49,980
justafiability testing.

150
00:06:49,980 --> 00:06:52,680
So it was propositional
reasoning.

151
00:06:52,680 --> 00:06:56,202
And that's what I have
focused most on since then.

152
00:06:56,202 --> 00:06:58,035
And there, we have made
tremendous progress.

153
00:06:58,035 --> 00:07:00,220
So from a few hundred
variable problems

154
00:07:00,220 --> 00:07:03,940
in the early '90s to now a
million variable problems, five

155
00:07:03,940 --> 00:07:06,010
million constraint problems.

156
00:07:06,010 --> 00:07:08,080
And the area has
nicely branched out.

157
00:07:08,080 --> 00:07:11,122
So now, my interest is beyond.

158
00:07:11,122 --> 00:07:13,330
I work with people in hardware
verification, software

159
00:07:13,330 --> 00:07:16,090
verification, because there
these tools are really

160
00:07:16,090 --> 00:07:18,400
used to support hardware
and software design.

161
00:07:18,400 --> 00:07:20,440
So I sort of see
your reasoning now,

162
00:07:20,440 --> 00:07:23,030
as sort of augmenting
the human reasoning.

163
00:07:23,030 --> 00:07:25,190
So human designers, engineers.

164
00:07:25,190 --> 00:07:27,250
And I think that
that's sort of where

165
00:07:27,250 --> 00:07:29,190
I see the future, one
part of the future of AI

166
00:07:29,190 --> 00:07:32,360
going to be, I think,
augmenting human intelligence

167
00:07:32,360 --> 00:07:35,260
and actually strengthening
it in many ways.

168
00:07:35,260 --> 00:07:39,960
Where do you see your work
specifically, your season, so

169
00:07:39,960 --> 00:07:41,300
the next progression?

170
00:07:41,300 --> 00:07:44,230
So that's sort of where
I've always been interested.

171
00:07:44,230 --> 00:07:46,420
For example,
mathematical discoveries

172
00:07:46,420 --> 00:07:47,980
and mathematical
theorem proving.

173
00:07:47,980 --> 00:07:50,020
And so I work with
some people at Cornell

174
00:07:50,020 --> 00:07:53,120
who have actually studied
very sophisticated

175
00:07:53,120 --> 00:07:56,115
theorem-proving systems.

176
00:07:56,115 --> 00:07:58,240
The search space is too
large for actually doing it

177
00:07:58,240 --> 00:07:58,970
automatically.

178
00:07:58,970 --> 00:08:03,070
So basically, they reformulate
proofs by mathematicians

179
00:08:03,070 --> 00:08:05,532
and then verify
them automatically.

180
00:08:05,532 --> 00:08:07,240
But what we now want
to do is we actually

181
00:08:07,240 --> 00:08:10,390
want to have these proofs
automatically discovered.

182
00:08:10,390 --> 00:08:12,160
So the one thing
we're seeing is that I

183
00:08:12,160 --> 00:08:15,190
think computers will do a
slightly different kind of math

184
00:08:15,190 --> 00:08:17,360
than humans do.

185
00:08:17,360 --> 00:08:19,210
But that can be
nicely complementary.

186
00:08:19,210 --> 00:08:21,400
So you can conjecture
some things

187
00:08:21,400 --> 00:08:24,730
about mathematical objects,
and a computer can be used

188
00:08:24,730 --> 00:08:28,578
and hopefully will be used
to verify these properties

189
00:08:28,578 --> 00:08:29,870
in a reasonable amount of time.

190
00:08:29,870 --> 00:08:32,110
And that's sort of
an interactive setup

191
00:08:32,110 --> 00:08:34,360
that could actually help
the human scientist.

192
00:08:34,360 --> 00:08:37,627
Can you go into a little
more detail about that?

193
00:08:37,627 --> 00:08:38,210
So basically--

194
00:08:38,210 --> 00:08:39,653
The differences.

195
00:08:39,653 --> 00:08:41,320
Oh, the differences
in what they can do.

196
00:08:41,320 --> 00:08:43,690
Well, I think it's clearest
in something like chip

197
00:08:43,690 --> 00:08:45,880
design or software design.

198
00:08:45,880 --> 00:08:47,440
And chip design is already used.

199
00:08:47,440 --> 00:08:49,510
So the chip designer has
sort of a varying thing,

200
00:08:49,510 --> 00:08:52,960
as a human, fairly good
creative insights of what

201
00:08:52,960 --> 00:08:55,300
a reasonably good design is.

202
00:08:55,300 --> 00:08:59,570
However, when you use
millions of gates,

203
00:08:59,570 --> 00:09:03,640
millions of components, subtle
interactions you may miss.

204
00:09:03,640 --> 00:09:07,000
So you may want to know, can
this register ever overflow?

205
00:09:07,000 --> 00:09:08,620
And so what the
chip designers do

206
00:09:08,620 --> 00:09:10,400
is then they actually
send that query.

207
00:09:10,400 --> 00:09:12,400
It's actually going
to be translated

208
00:09:12,400 --> 00:09:14,890
by [INAUDIBLE] techniques,
[INAUDIBLE] liability problem

209
00:09:14,890 --> 00:09:17,245
and ask the computer, can
you check whether this really

210
00:09:17,245 --> 00:09:19,150
will work if I put it together?

211
00:09:19,150 --> 00:09:21,340
And I sort of, what
I would like to see

212
00:09:21,340 --> 00:09:24,880
is an application
like this, and that's

213
00:09:24,880 --> 00:09:27,380
what we're starting to work on
in, for example, mathematics.

214
00:09:27,380 --> 00:09:30,790
So we can say, I think,
that this observation

215
00:09:30,790 --> 00:09:32,350
might have certain properties.

216
00:09:32,350 --> 00:09:36,110
Can you check whether this
property really whole or not?

217
00:09:36,110 --> 00:09:36,610
Yes.

218
00:09:36,610 --> 00:09:37,360
Wow.

219
00:09:37,360 --> 00:09:39,430
And I think this is
becoming possible.

220
00:09:39,430 --> 00:09:43,645
So that's in the
next 5 to 20 years.

221
00:09:43,645 --> 00:09:45,645
What do you think about
the new sort of emphasis

222
00:09:45,645 --> 00:09:49,530
on tying artificial
intelligence into sort

223
00:09:49,530 --> 00:09:53,660
of biology, nanotechnology,
stuff in that direction?

224
00:09:53,660 --> 00:09:54,160
Yeah.

225
00:09:54,160 --> 00:09:56,040
So those, I think, are
interesting direction.

226
00:09:56,040 --> 00:09:59,500
I mean, my own work, it has,
for example, connections

227
00:09:59,500 --> 00:10:00,510
to statistical physics.

228
00:10:00,510 --> 00:10:05,710
So I think a big change
in the last 10, 15 years

229
00:10:05,710 --> 00:10:07,540
is the sophistication
of techniques.

230
00:10:07,540 --> 00:10:09,040
So we're looking
at other fields,

231
00:10:09,040 --> 00:10:13,270
like statistics,
probability, biology,

232
00:10:13,270 --> 00:10:15,490
where we reach out
to other sciences.

233
00:10:15,490 --> 00:10:17,110
And I think that can
be very fruitful.

234
00:10:17,110 --> 00:10:19,090
So I'm mostly really
working as a physicist,

235
00:10:19,090 --> 00:10:22,330
a physical physicist,
who actually have come up

236
00:10:22,330 --> 00:10:25,160
with new kinds of [INAUDIBLE]
truth and dare techniques.

237
00:10:25,160 --> 00:10:27,370
So I think that outreach
to the community

238
00:10:27,370 --> 00:10:30,130
is going to be very
important, and that I

239
00:10:30,130 --> 00:10:33,746
think will accelerate.

240
00:10:33,746 --> 00:10:37,090
So I guess we have time
for one final question?

241
00:10:37,090 --> 00:10:38,470
Besides your own work, which--

242
00:10:38,470 --> 00:10:40,310
because you're working
on, is obviously

243
00:10:40,310 --> 00:10:42,940
the coolest thing
in the field, what

244
00:10:42,940 --> 00:10:44,470
do you think the
next coolest thing

245
00:10:44,470 --> 00:10:46,450
is that's being
worked on right now?

246
00:10:46,450 --> 00:10:52,270
I think the whole integration
with sensors and data.

247
00:10:52,270 --> 00:10:56,800
So the fact that we have
so much more data available

248
00:10:56,800 --> 00:11:00,910
makes these statistical methods
and the statistical approaches

249
00:11:00,910 --> 00:11:03,490
more promising.

250
00:11:03,490 --> 00:11:06,000
Similarly, we mine
data on the web.

251
00:11:06,000 --> 00:11:09,250
I think that might actually
help at this moment.

252
00:11:09,250 --> 00:11:11,980
Many years ago, sort of
common-sense reasoning

253
00:11:11,980 --> 00:11:14,110
may become more feasible
by actually going back

254
00:11:14,110 --> 00:11:15,978
to the web [INAUDIBLE].

255
00:11:15,978 --> 00:11:16,850
Cool.

256
00:11:16,850 --> 00:11:18,120
[INAUDIBLE] on the web.

257
00:11:18,120 --> 00:11:18,990
Makes sense.

258
00:11:18,990 --> 00:11:20,820
Unfortunately, we have
to wrap things up.

259
00:11:20,820 --> 00:11:22,020
But thanks a lot
for talking to me.

260
00:11:22,020 --> 00:11:22,250
Thanks.

261
00:11:22,250 --> 00:11:22,600
I really enjoyed it.

262
00:11:22,600 --> 00:11:23,517
It's good to meet you.

263
00:11:23,517 --> 00:11:24,800
OK.

264
00:11:24,800 --> 00:11:27,000