1
00:00:00,919 --> 00:00:05,879
so in my slides here I'm going to talk

2
00:00:03,760 --> 00:00:10,040
about debugging and understanding NLP

3
00:00:05,879 --> 00:00:12,400
models and this is how to tell uh when

4
00:00:10,040 --> 00:00:14,759
for example both your implementations

5
00:00:12,400 --> 00:00:17,320
are wrong and uh for example your

6
00:00:14,759 --> 00:00:19,000
underlying assumptions are wrong or your

7
00:00:17,320 --> 00:00:21,240
model is failing on particular segments

8
00:00:19,000 --> 00:00:23,439
of data or stuff like that so going to

9
00:00:21,240 --> 00:00:26,160
go through uh a variety of things that

10
00:00:23,439 --> 00:00:29,000
can go wrong with your experiments

11
00:00:26,160 --> 00:00:31,679
basically so a typical situation is

12
00:00:29,000 --> 00:00:33,399
you've implemented some NLP system you

13
00:00:31,679 --> 00:00:35,840
know based on neural networks of course

14
00:00:33,399 --> 00:00:36,920
because that's what we use nowadays um

15
00:00:35,840 --> 00:00:40,000
and you've looked at the code it

16
00:00:36,920 --> 00:00:42,000
basically looks okay um but it has low

17
00:00:40,000 --> 00:00:44,559
accuracy or it makes incomprehensible

18
00:00:42,000 --> 00:00:45,680
errors and you would like to uh fix

19
00:00:44,559 --> 00:00:47,440
these or you'd like to improve the

20
00:00:45,680 --> 00:00:49,120
accuracy or something like that and so

21
00:00:47,440 --> 00:00:52,000
what do I

22
00:00:49,120 --> 00:00:53,680
do and I think there's three dimensions

23
00:00:52,000 --> 00:00:56,239
of how you can understand your model and

24
00:00:53,680 --> 00:00:57,960
your Model Behavior um the first one is

25
00:00:56,239 --> 00:01:00,199
debugging the implementation so it's

26
00:00:57,960 --> 00:01:03,760
identifying problems that you have when

27
00:01:00,199 --> 00:01:05,880
you uh implemented something uh second

28
00:01:03,760 --> 00:01:07,759
thing is actionable evaluation so

29
00:01:05,880 --> 00:01:09,799
identifying typical error cases and how

30
00:01:07,759 --> 00:01:11,840
you what you can do to fix them and

31
00:01:09,799 --> 00:01:13,720
finally uh interpreting predictions or

32
00:01:11,840 --> 00:01:18,080
interpreting what's happening inside the

33
00:01:13,720 --> 00:01:19,920
model and uh this can maybe give you a

34
00:01:18,080 --> 00:01:21,520
deeper idea about what's happening in

35
00:01:19,920 --> 00:01:22,720
happening in individual cases and

36
00:01:21,520 --> 00:01:25,240
there's a lot of reasons why you might

37
00:01:22,720 --> 00:01:27,920
want to do that uh both like to make

38
00:01:25,240 --> 00:01:30,280
your models better and also for example

39
00:01:27,920 --> 00:01:31,840
if you want to be sure that your ition

40
00:01:30,280 --> 00:01:34,840
isn't doing something illegal like

41
00:01:31,840 --> 00:01:36,439
discriminating against people uh due to

42
00:01:34,840 --> 00:01:38,680
protected attributes or other things

43
00:01:36,439 --> 00:01:41,399
like that so um there's a number of

44
00:01:38,680 --> 00:01:42,920
reasons why you'd want to do that so I'm

45
00:01:41,399 --> 00:01:44,399
going to talk about the first two and

46
00:01:42,920 --> 00:01:48,840
Nishant is mainly going to talk about

47
00:01:44,399 --> 00:01:52,000
the second one so uh going right into

48
00:01:48,840 --> 00:01:55,159
it so in neural network models uh

49
00:01:52,000 --> 00:01:58,880
debugging is really important because

50
00:01:55,159 --> 00:02:00,920
they're opaque they're unpredictable and

51
00:01:58,880 --> 00:02:03,119
uh if you make little mistakes they can

52
00:02:00,920 --> 00:02:05,439
cause big problems with your

53
00:02:03,119 --> 00:02:07,399
output and another thing is that

54
00:02:05,439 --> 00:02:09,640
everything is a hyperparameter including

55
00:02:07,399 --> 00:02:11,239
your network size your model variations

56
00:02:09,640 --> 00:02:14,440
your batch size your strategy your

57
00:02:11,239 --> 00:02:18,120
Optimizer and your learning rate

58
00:02:14,440 --> 00:02:19,560
and finally unlike kind of more

59
00:02:18,120 --> 00:02:21,200
traditional machine learning methods

60
00:02:19,560 --> 00:02:23,000
like logistic progression or support

61
00:02:21,200 --> 00:02:25,160
Vector machines or something like that

62
00:02:23,000 --> 00:02:27,879
you might that you might have studied in

63
00:02:25,160 --> 00:02:30,160
your machine learning class um

64
00:02:27,879 --> 00:02:32,599
stochastic optimization has no guarantee

65
00:02:30,160 --> 00:02:34,239
about convergence um your loss might go

66
00:02:32,599 --> 00:02:35,720
down then it might go up and there might

67
00:02:34,239 --> 00:02:38,120
be absolutely nothing wrong with your

68
00:02:35,720 --> 00:02:40,200
training or it might be you know a

69
00:02:38,120 --> 00:02:42,319
serious problem so that's another issue

70
00:02:40,200 --> 00:02:45,440
you need to deal

71
00:02:42,319 --> 00:02:48,800
with so first I'd like to go into

72
00:02:45,440 --> 00:02:51,400
possible causes of problems with your

73
00:02:48,800 --> 00:02:53,440
implementation and I'm going to break

74
00:02:51,400 --> 00:02:55,040
them down into a typology and based on

75
00:02:53,440 --> 00:02:57,040
what part of the typology you're running

76
00:02:55,040 --> 00:02:59,200
into problems with you will need to fix

77
00:02:57,040 --> 00:03:00,800
them in different ways so your first

78
00:02:59,200 --> 00:03:02,599
goal when you're experiencing the

79
00:03:00,800 --> 00:03:04,720
problem is identifying why you're

80
00:03:02,599 --> 00:03:06,400
experiencing the problem uh because that

81
00:03:04,720 --> 00:03:08,760
will lead you to a

82
00:03:06,400 --> 00:03:10,440
solution so for training time problems

83
00:03:08,760 --> 00:03:12,560
there's a bunch of uh things that could

84
00:03:10,440 --> 00:03:14,360
be wrong uh the first is a lack of model

85
00:03:12,560 --> 00:03:16,280
capacity so your model is not able to

86
00:03:14,360 --> 00:03:18,599
model the phenomena that you want to

87
00:03:16,280 --> 00:03:20,000
model in the first place um you could

88
00:03:18,599 --> 00:03:22,080
have a poor training

89
00:03:20,000 --> 00:03:24,920
algorithm uh you could just have a bug

90
00:03:22,080 --> 00:03:27,080
in your code at training time another

91
00:03:24,920 --> 00:03:29,319
thing is uh test time problems and these

92
00:03:27,080 --> 00:03:30,599
can include a disconnect between what

93
00:03:29,319 --> 00:03:33,040
you're doing at training time and what

94
00:03:30,599 --> 00:03:35,640
you're testing at testing time uh

95
00:03:33,040 --> 00:03:37,959
failure of search

96
00:03:35,640 --> 00:03:39,920
algorithms and another thing you want to

97
00:03:37,959 --> 00:03:41,360
deal with is overfitting so you're

98
00:03:39,920 --> 00:03:44,319
actually doing well on the training set

99
00:03:41,360 --> 00:03:48,360
but you're doing poorly on the test

100
00:03:44,319 --> 00:03:50,400
Set uh finally you could have um optimiz

101
00:03:48,360 --> 00:03:52,640
a mismatch between the function you're

102
00:03:50,400 --> 00:03:54,920
optimizing at evaluation time and uh

103
00:03:52,640 --> 00:03:56,519
what you're actually evaluating sorry

104
00:03:54,920 --> 00:03:58,079
the fun the function that you're

105
00:03:56,519 --> 00:04:01,079
optimizing at training time and what

106
00:03:58,079 --> 00:04:03,720
you're actually evaluating at test time

107
00:04:01,079 --> 00:04:05,280
and my my best piece of advice for

108
00:04:03,720 --> 00:04:07,959
figuring out why things are going wrong

109
00:04:05,280 --> 00:04:11,040
is don't uh try to do all of them at

110
00:04:07,959 --> 00:04:12,560
once and rather uh start from the top

111
00:04:11,040 --> 00:04:15,239
and work it down because the ones at the

112
00:04:12,560 --> 00:04:17,600
top are often easier to uh diagnose and

113
00:04:15,239 --> 00:04:20,680
the ones at the

114
00:04:17,600 --> 00:04:23,000
bottom so looking at how you can debug

115
00:04:20,680 --> 00:04:25,919
systems at training time uh there's a

116
00:04:23,000 --> 00:04:27,360
number of ways you can do this uh but

117
00:04:25,919 --> 00:04:30,039
the most important thing for training

118
00:04:27,360 --> 00:04:33,479
time uh issues is looking at the loss

119
00:04:30,039 --> 00:04:36,759
function calculated on the training set

120
00:04:33,479 --> 00:04:38,960
and what I mean by this is don't look uh

121
00:04:36,759 --> 00:04:41,240
we talked about how we can't optimize

122
00:04:38,960 --> 00:04:45,039
error or accuracy easily so instead we

123
00:04:41,240 --> 00:04:47,120
optimize likelihood um and so you might

124
00:04:45,039 --> 00:04:49,080
want to look at accuracy to see whether

125
00:04:47,120 --> 00:04:50,759
your model is working well but I would

126
00:04:49,080 --> 00:04:53,039
urge you first to look at your

127
00:04:50,759 --> 00:04:55,080
likelihood or your loss function on the

128
00:04:53,039 --> 00:04:57,000
training set instead of your accuracy on

129
00:04:55,080 --> 00:04:58,479
the test set for example to diagnose

130
00:04:57,000 --> 00:05:00,600
these variety of

131
00:04:58,479 --> 00:05:02,919
problems and the sorts of things you

132
00:05:00,600 --> 00:05:05,840
want to look at are um is the loss

133
00:05:02,919 --> 00:05:10,639
function going down so is it you know

134
00:05:05,840 --> 00:05:14,199
converging into a good place

135
00:05:10,639 --> 00:05:16,280
um in general if this is your your

136
00:05:14,199 --> 00:05:18,600
loss um the first thing you should know

137
00:05:16,280 --> 00:05:20,440
is like what is a good loss uh in most

138
00:05:18,600 --> 00:05:22,280
cases a good loss is zero like log

139
00:05:20,440 --> 00:05:26,280
likelihood the best loss you can achieve

140
00:05:22,280 --> 00:05:28,639
is zero so you have zero down here um

141
00:05:26,280 --> 00:05:31,639
something

142
00:05:28,639 --> 00:05:31,639
like

143
00:05:31,919 --> 00:05:36,680
this is uh essentially a good loss

144
00:05:38,080 --> 00:05:43,120
function something like that uh

145
00:05:41,360 --> 00:05:45,120
especially if this is a relatively High

146
00:05:43,120 --> 00:05:47,759
number is usually a bad loss

147
00:05:45,120 --> 00:05:50,319
function

148
00:05:47,759 --> 00:05:52,680
um something like that on your training

149
00:05:50,319 --> 00:05:54,240
set is a very bad loss function uh

150
00:05:52,680 --> 00:05:55,840
something something is going seriously

151
00:05:54,240 --> 00:05:57,960
wrong if you see this on your Dev set

152
00:05:55,840 --> 00:05:59,800
that could be or your test set that

153
00:05:57,960 --> 00:06:01,199
could be uh overfitting but but if

154
00:05:59,800 --> 00:06:03,440
you're seeing that on your training set

155
00:06:01,199 --> 00:06:05,759
that's usually symptomatic of a problem

156
00:06:03,440 --> 00:06:09,160
so uh these are uh things that you

157
00:06:05,759 --> 00:06:10,960
should be uh knowing um is it going down

158
00:06:09,160 --> 00:06:13,520
basically to zero if you run training

159
00:06:10,960 --> 00:06:16,000
long enough um for many epochs over your

160
00:06:13,520 --> 00:06:17,479
training data so if it's not going down

161
00:06:16,000 --> 00:06:20,599
to zero and it's sticking up here then

162
00:06:17,479 --> 00:06:20,599
that's also an

163
00:06:21,120 --> 00:06:25,759
issue and um if it's not going down to

164
00:06:23,840 --> 00:06:27,919
close to zero on whatever training set

165
00:06:25,759 --> 00:06:30,199
you're training on um let's say you make

166
00:06:27,919 --> 00:06:31,840
your training set extremely small

167
00:06:30,199 --> 00:06:33,319
uh at least in that case it should go

168
00:06:31,840 --> 00:06:34,960
down to zero otherwise you might have a

169
00:06:33,319 --> 00:06:37,199
serious problem in your

170
00:06:34,960 --> 00:06:39,240
implementation so these are good things

171
00:06:37,199 --> 00:06:41,960
to check first when you're training a

172
00:06:39,240 --> 00:06:45,199
model um and there's a number of reasons

173
00:06:41,960 --> 00:06:47,759
why this might not be helping or why

174
00:06:45,199 --> 00:06:50,880
this might not be happening so um your

175
00:06:47,759 --> 00:06:53,120
Mo model might be too weak and so in

176
00:06:50,880 --> 00:06:55,440
general larger models tend to perform

177
00:06:53,120 --> 00:06:58,000
better uh especially if you're using a

178
00:06:55,440 --> 00:06:59,800
pre-trained model and um this is just an

179
00:06:58,000 --> 00:07:03,800
example from the T5 paper where they

180
00:06:59,800 --> 00:07:06,680
scale up the T5 model um from a

181
00:07:03,800 --> 00:07:09,319
relatively small model to what at the

182
00:07:06,680 --> 00:07:12,199
time was a very large model of 11

183
00:07:09,319 --> 00:07:14,360
billion parameters now this is you know

184
00:07:12,199 --> 00:07:17,479
a moderately sized model or maybe even

185
00:07:14,360 --> 00:07:20,879
small model by some standards but anyway

186
00:07:17,479 --> 00:07:23,800
you can see that it uh in continues to

187
00:07:20,879 --> 00:07:26,479
increase one really interesting

188
00:07:23,800 --> 00:07:30,080
phenomenon is uh that actually larger

189
00:07:26,479 --> 00:07:33,879
models can learn faster or at least with

190
00:07:30,080 --> 00:07:36,680
fewer steps than uh smaller

191
00:07:33,879 --> 00:07:40,199
models and so this

192
00:07:36,680 --> 00:07:42,240
is an interesting example this paper uh

193
00:07:40,199 --> 00:07:43,919
on neural scaling was it's a very

194
00:07:42,240 --> 00:07:48,000
influential paper but basically what

195
00:07:43,919 --> 00:07:51,000
they show is the darker purple ones are

196
00:07:48,000 --> 00:07:54,599
smaller models the yellow ones are

197
00:07:51,000 --> 00:07:57,159
bigger models and what you can see here

198
00:07:54,599 --> 00:07:59,639
is the purple model and on the left side

199
00:07:57,159 --> 00:08:02,120
they have the number of tokens processed

200
00:07:59,639 --> 00:08:05,759
the right side they have the number of

201
00:08:02,120 --> 00:08:08,159
uh compute or the amount of compute um

202
00:08:05,759 --> 00:08:10,080
and so what you can see is if you just

203
00:08:08,159 --> 00:08:12,240
look at the number of tokens processed

204
00:08:10,080 --> 00:08:14,280
the larger the model the faster it

205
00:08:12,240 --> 00:08:17,720
converges which

206
00:08:14,280 --> 00:08:21,400
is maybe a little bit surprising maybe a

207
00:08:17,720 --> 00:08:22,680
little bit you or maybe uh like some

208
00:08:21,400 --> 00:08:24,879
people have the intuition that this

209
00:08:22,680 --> 00:08:26,440
should be the case but when I first saw

210
00:08:24,879 --> 00:08:27,759
this I found it a little bit surprising

211
00:08:26,440 --> 00:08:29,000
because I thought it would be so large

212
00:08:27,759 --> 00:08:29,960
and noisy that the model would have

213
00:08:29,000 --> 00:08:32,320
trouble fit

214
00:08:29,960 --> 00:08:34,200
you know fitting the data as quickly but

215
00:08:32,320 --> 00:08:36,200
there's actually a good reason for this

216
00:08:34,200 --> 00:08:37,240
does anyone have a guess about why this

217
00:08:36,200 --> 00:08:39,719
is

218
00:08:37,240 --> 00:08:41,240
thee we've talked a little bit about the

219
00:08:39,719 --> 00:08:44,120
underlying phenomena for this in

220
00:08:41,240 --> 00:08:48,360
previous classes so you might be able to

221
00:08:44,120 --> 00:08:48,360
think back to some of the things you

222
00:08:50,480 --> 00:08:56,040
yeah yeah so um just to repeat there's a

223
00:08:54,160 --> 00:08:57,720
lot of different parameters so it can

224
00:08:56,040 --> 00:08:59,880
try to converge along a lot of different

225
00:08:57,720 --> 00:09:01,920
dimensions so if we think back to the

226
00:08:59,880 --> 00:09:04,079
like model pruning class and other stuff

227
00:09:01,920 --> 00:09:06,640
like that um part of the reason why we

228
00:09:04,079 --> 00:09:08,000
can prune large models so efficiently is

229
00:09:06,640 --> 00:09:10,200
because only like a small number of the

230
00:09:08,000 --> 00:09:12,440
parameters are actually useful and so if

231
00:09:10,200 --> 00:09:15,120
you start out with a much larger model

232
00:09:12,440 --> 00:09:17,720
it's more likely to have useful subsets

233
00:09:15,120 --> 00:09:20,320
of the parameters basically um which is

234
00:09:17,720 --> 00:09:21,560
called the lottery ticket hypothesis uh

235
00:09:20,320 --> 00:09:23,839
there there's a famous paper called the

236
00:09:21,560 --> 00:09:27,560
lottery ticket hypothesis examines this

237
00:09:23,839 --> 00:09:29,680
phenomenon so um one one interesting

238
00:09:27,560 --> 00:09:32,160
thing is you can see that even if you

239
00:09:29,680 --> 00:09:35,640
scale up the compute even if you measure

240
00:09:32,160 --> 00:09:37,640
based on compute the uh larger models

241
00:09:35,640 --> 00:09:38,959
eventually surpass the smaller models in

242
00:09:37,640 --> 00:09:41,920
terms of how efficient they are at

243
00:09:38,959 --> 00:09:44,680
modeling the data and that's just

244
00:09:41,920 --> 00:09:46,760
because models tend to learn well for a

245
00:09:44,680 --> 00:09:49,560
while and then they basically reach

246
00:09:46,760 --> 00:09:51,760
their capacity and stop learning well or

247
00:09:49,560 --> 00:09:53,680
they start learning very slowly and once

248
00:09:51,760 --> 00:09:57,120
you get to that point the larger models

249
00:09:53,680 --> 00:09:58,800
work better so there's a kind of

250
00:09:57,120 --> 00:10:00,640
counterintuitive thing that if you want

251
00:09:58,800 --> 00:10:04,160
to train faster you actually can train a

252
00:10:00,640 --> 00:10:06,839
larger model and uh that will that will

253
00:10:04,160 --> 00:10:08,000
uh get you to a good solution at some

254
00:10:06,839 --> 00:10:09,640
point that will get you to a good

255
00:10:08,000 --> 00:10:11,120
solution faster than a smaller model

256
00:10:09,640 --> 00:10:15,200
would you know of course you need memory

257
00:10:11,120 --> 00:10:15,200
and stuff but why are looking

258
00:10:20,040 --> 00:10:26,920
at so this is test loss training loss

259
00:10:22,760 --> 00:10:30,680
also looks like this um I think on

260
00:10:26,920 --> 00:10:34,360
this particular

261
00:10:30,680 --> 00:10:37,519
on this particular paper they never

262
00:10:34,360 --> 00:10:39,399
repeated data and if you never repeat

263
00:10:37,519 --> 00:10:42,560
data actually your training loss looks

264
00:10:39,399 --> 00:10:44,680
very similar to your test loss because

265
00:10:42,560 --> 00:10:46,079
it if you like actually if you can

266
00:10:44,680 --> 00:10:48,760
assume your training data set and your

267
00:10:46,079 --> 00:10:50,880
test data set are um uh identically

268
00:10:48,760 --> 00:10:52,279
distributed your training loss of new

269
00:10:50,880 --> 00:10:54,600
training data should be exactly the same

270
00:10:52,279 --> 00:10:55,959
as your test loss so I think that's

271
00:10:54,600 --> 00:10:57,760
basically why they were justified in

272
00:10:55,959 --> 00:11:01,000
doing that good but they probably did

273
00:10:57,760 --> 00:11:03,639
test loss to like I swashed the concern

274
00:11:01,000 --> 00:11:05,839
that this was overfitting comp or

275
00:11:03,639 --> 00:11:09,200
something but good

276
00:11:05,839 --> 00:11:11,279
question um cool so the these are are

277
00:11:09,200 --> 00:11:13,000
good things to know um so basically if

278
00:11:11,279 --> 00:11:14,839
you see your model doing something like

279
00:11:13,000 --> 00:11:16,279
this um plateauing out maybe your

280
00:11:14,839 --> 00:11:18,680
model's too small and you need to tr a

281
00:11:16,279 --> 00:11:20,920
big

282
00:11:18,680 --> 00:11:22,200
basically another uh piece of trouble

283
00:11:20,920 --> 00:11:26,800
that you can have is trouble with

284
00:11:22,200 --> 00:11:29,519
optimization and basically um you should

285
00:11:26,800 --> 00:11:31,600
check your Optimizer um usually people

286
00:11:29,519 --> 00:11:35,639
are using atom variants nowadays like

287
00:11:31,600 --> 00:11:37,839
atom or atom W so just use that um

288
00:11:35,639 --> 00:11:39,639
learning rate uh so make sure that the

289
00:11:37,839 --> 00:11:41,160
learning rate you're using is standard

290
00:11:39,639 --> 00:11:43,399
for kind of the model size that you're

291
00:11:41,160 --> 00:11:44,920
using and the best way to do this is uh

292
00:11:43,399 --> 00:11:46,000
look at previous papers and see what

293
00:11:44,920 --> 00:11:50,160
they're

294
00:11:46,000 --> 00:11:51,680
using um initialization most people

295
00:11:50,160 --> 00:11:53,440
nowadays will not be training from

296
00:11:51,680 --> 00:11:55,440
scratch but if you are training from

297
00:11:53,440 --> 00:11:58,040
scratch how you initialize your model is

298
00:11:55,440 --> 00:11:59,399
really important and normally the way

299
00:11:58,040 --> 00:12:03,320
you do this is you do this with some

300
00:11:59,399 --> 00:12:05,079
sort of uniform random noise and uh

301
00:12:03,320 --> 00:12:06,959
specifically you can pick the uniform

302
00:12:05,079 --> 00:12:08,800
random noise in intelligent ways based

303
00:12:06,959 --> 00:12:12,240
on the the data size which I'll talk

304
00:12:08,800 --> 00:12:13,920
about in a second um also mini batching

305
00:12:12,240 --> 00:12:15,639
um are you using sufficiently large

306
00:12:13,920 --> 00:12:17,480
batches of data if you're using small

307
00:12:15,639 --> 00:12:18,720
batches of data you might have too much

308
00:12:17,480 --> 00:12:21,279
noise in your training and it might

309
00:12:18,720 --> 00:12:23,839
diverge so uh these are things you need

310
00:12:21,279 --> 00:12:23,839
think about as

311
00:12:25,279 --> 00:12:30,560
well

312
00:12:27,519 --> 00:12:35,000
cool um so these are training time

313
00:12:30,560 --> 00:12:37,320
things um the next thing is debugging at

314
00:12:35,000 --> 00:12:37,320
test

315
00:12:38,160 --> 00:12:43,839
time and this is particularly important

316
00:12:41,240 --> 00:12:47,320
if you're doing any sort

317
00:12:43,839 --> 00:12:48,880
of like I guess a lot of this has kind

318
00:12:47,320 --> 00:12:51,360
of been commoditized and it's

319
00:12:48,880 --> 00:12:52,560
implemented in hugging face and stuff

320
00:12:51,360 --> 00:12:55,120
like that and as long as you're using

321
00:12:52,560 --> 00:12:57,279
the standard implementations you're less

322
00:12:55,120 --> 00:12:59,000
likely to run into these bugs but if you

323
00:12:57,279 --> 00:13:00,519
are implementing anything on your own

324
00:12:59,000 --> 00:13:03,040
this is actually really tricky and you

325
00:13:00,519 --> 00:13:07,880
can easily make mistakes so uh it's

326
00:13:03,040 --> 00:13:08,959
important to to know about it so um what

327
00:13:07,880 --> 00:13:10,680
one of the reasons why you can have

328
00:13:08,959 --> 00:13:12,240
training and test disconnects especially

329
00:13:10,680 --> 00:13:14,399
if you're doing something like text

330
00:13:12,240 --> 00:13:15,959
generation is that usually your loss

331
00:13:14,399 --> 00:13:17,720
calculation and prodiction functions

332
00:13:15,959 --> 00:13:20,480
will be implemented in different

333
00:13:17,720 --> 00:13:23,360
functions and like anything in software

334
00:13:20,480 --> 00:13:25,440
engineering Um this can be a source of

335
00:13:23,360 --> 00:13:26,760
bugs duplicated sour code can be a

336
00:13:25,440 --> 00:13:28,440
source of bugs because you might

337
00:13:26,760 --> 00:13:30,199
Implement one thing in one place in one

338
00:13:28,440 --> 00:13:33,000
way another thing in another place in

339
00:13:30,199 --> 00:13:35,560
another way so this is no exception to

340
00:13:33,000 --> 00:13:37,399
that um it's especially true for

341
00:13:35,560 --> 00:13:39,000
structured prediction models so anything

342
00:13:37,399 --> 00:13:40,399
where you're not just making a single

343
00:13:39,000 --> 00:13:42,079
prediction but you're making multiple

344
00:13:40,399 --> 00:13:43,839
predictions in a row so you need to be a

345
00:13:42,079 --> 00:13:46,959
little bit careful about

346
00:13:43,839 --> 00:13:49,880
that um another thing that you need to

347
00:13:46,959 --> 00:13:51,079
be pay attention about is often uh

348
00:13:49,880 --> 00:13:52,680
especially if you're doing your own

349
00:13:51,079 --> 00:13:55,880
implementation loss calculation it's

350
00:13:52,680 --> 00:13:59,800
mini batched and generation is not or in

351
00:13:55,880 --> 00:14:02,199
highly optimized versions of um of

352
00:13:59,800 --> 00:14:03,880
inference you might be doing inference

353
00:14:02,199 --> 00:14:05,360
with Dynamic batching and stuff like

354
00:14:03,880 --> 00:14:06,720
that and it might become complicated you

355
00:14:05,360 --> 00:14:09,800
might make

356
00:14:06,720 --> 00:14:12,160
mistakes um so how do

357
00:14:09,800 --> 00:14:15,839
we make sure that we're not making any

358
00:14:12,160 --> 00:14:18,560
mistakes here um there's a really simple

359
00:14:15,839 --> 00:14:21,199
way to debug any sort of mini batched

360
00:14:18,560 --> 00:14:24,199
loss calculation because normally when

361
00:14:21,199 --> 00:14:27,000
we mini batch loss calculations we're

362
00:14:24,199 --> 00:14:31,079
simultaneously calculating uh the loss

363
00:14:27,000 --> 00:14:35,600
for like uh four four or eight or

364
00:14:31,079 --> 00:14:37,560
whatever sequences at a time and so you

365
00:14:35,600 --> 00:14:40,279
can calculate the loss with a large

366
00:14:37,560 --> 00:14:42,000
batch size like 32 and then calculate

367
00:14:40,279 --> 00:14:44,920
the loss for each uh sentence

368
00:14:42,000 --> 00:14:47,720
individually and sum them together and

369
00:14:44,920 --> 00:14:49,480
these uh value should be the same and

370
00:14:47,720 --> 00:14:52,160
this can help make sure that you don't

371
00:14:49,480 --> 00:14:55,120
have any you know issues with your

372
00:14:52,160 --> 00:14:57,959
padding or your masking or other things

373
00:14:55,120 --> 00:14:59,800
like this um so this is particularly

374
00:14:57,959 --> 00:15:01,959
important if you're not just using out

375
00:14:59,800 --> 00:15:04,240
of the box things so you have a slightly

376
00:15:01,959 --> 00:15:06,240
unusually structured model with like

377
00:15:04,240 --> 00:15:08,880
hierarchical encoding or anything like

378
00:15:06,240 --> 00:15:11,680
that you need to be really careful about

379
00:15:08,880 --> 00:15:15,440
that um you can even create unit tests

380
00:15:11,680 --> 00:15:17,399
that test this so like um in machine

381
00:15:15,440 --> 00:15:18,959
learning code we don't write unit test

382
00:15:17,399 --> 00:15:20,160
or especially neural network based

383
00:15:18,959 --> 00:15:22,440
machine learning code we don't write

384
00:15:20,160 --> 00:15:24,160
unit tests that often because it's kind

385
00:15:22,440 --> 00:15:26,279
of hard to do there's lots of Randomness

386
00:15:24,160 --> 00:15:27,959
and other stuff like that um but this is

387
00:15:26,279 --> 00:15:30,959
one thing that you can easily test and

388
00:15:27,959 --> 00:15:30,959
and make sure that you don't hear the

389
00:15:32,440 --> 00:15:39,319
mistakes um any sort of uh generation

390
00:15:36,480 --> 00:15:43,199
algorithm uh so when you're generating

391
00:15:39,319 --> 00:15:44,639
or decoding um you can make sure that

392
00:15:43,199 --> 00:15:47,639
your decoding code is getting the same

393
00:15:44,639 --> 00:15:50,040
score is when you calculate the loss and

394
00:15:47,639 --> 00:15:52,959
an easy way to do this is you call the

395
00:15:50,040 --> 00:15:54,759
decoding function to generate an output

396
00:15:52,959 --> 00:15:57,399
and normally when you're doing any sort

397
00:15:54,759 --> 00:15:59,480
of search or sampling or something like

398
00:15:57,399 --> 00:16:02,120
that during the search or sampling

399
00:15:59,480 --> 00:16:05,000
you're calculating the logits or the log

400
00:16:02,120 --> 00:16:07,399
probabilities of each step that you

401
00:16:05,000 --> 00:16:09,120
sample so you keep track of that during

402
00:16:07,399 --> 00:16:12,279
your sampling

403
00:16:09,120 --> 00:16:14,319
algorithm and then after that you call

404
00:16:12,279 --> 00:16:16,800
the loss function on the generated

405
00:16:14,319 --> 00:16:18,639
output and you calculate the loss

406
00:16:16,800 --> 00:16:20,360
according to the loss function and the

407
00:16:18,639 --> 00:16:22,240
score of these two things should be the

408
00:16:20,360 --> 00:16:26,440
same uh

409
00:16:22,240 --> 00:16:26,440
so um you know you do your

410
00:16:27,920 --> 00:16:35,279
generate and that gives you an

411
00:16:32,000 --> 00:16:35,279
output in

412
00:16:35,600 --> 00:16:42,360
score and then you do um

413
00:16:39,319 --> 00:16:45,839
loss on the

414
00:16:42,360 --> 00:16:49,040
output and that gives you the score

415
00:16:45,839 --> 00:16:53,079
two and then you just compare these two

416
00:16:49,040 --> 00:16:56,360
things together and this can uh in in my

417
00:16:53,079 --> 00:17:01,120
experience this has allowed me to find

418
00:16:56,360 --> 00:17:03,240
the majority of the bugs in um these two

419
00:17:01,120 --> 00:17:04,679
things um have allowed me to find the

420
00:17:03,240 --> 00:17:06,600
majority of the bugs whenever I was

421
00:17:04,679 --> 00:17:09,199
doing any sort of like complex thing

422
00:17:06,600 --> 00:17:11,880
with respect to generation or models and

423
00:17:09,199 --> 00:17:13,360
stuff like that so um it's a very common

424
00:17:11,880 --> 00:17:15,439
place for bugs even if you're pretty

425
00:17:13,360 --> 00:17:17,280
familiar with models so I I would highly

426
00:17:15,439 --> 00:17:19,760
recommend

427
00:17:17,280 --> 00:17:21,319
that um this is particularly bad when

428
00:17:19,760 --> 00:17:25,559
you're doing something like a search

429
00:17:21,319 --> 00:17:28,400
algorithm like beam search um and

430
00:17:25,559 --> 00:17:30,400
so beam search uh as you know from the

431
00:17:28,400 --> 00:17:34,200
generation class instead of picking one

432
00:17:30,400 --> 00:17:37,080
high probability uh you know word in

433
00:17:34,200 --> 00:17:40,160
your next step you maintain several

434
00:17:37,080 --> 00:17:41,960
paths and one way that you can fix this

435
00:17:40,160 --> 00:17:44,320
is as you make search better the model

436
00:17:41,960 --> 00:17:45,760
score should get better so the log

437
00:17:44,320 --> 00:17:48,240
likelihood of the output should get

438
00:17:45,760 --> 00:17:50,280
better almost all of the time so you can

439
00:17:48,240 --> 00:17:51,840
search with varying beam sizes and make

440
00:17:50,280 --> 00:17:55,280
sure that you get a better overall model

441
00:17:51,840 --> 00:17:57,559
score at the end so um and you can even

442
00:17:55,280 --> 00:17:59,320
create a unit test testing this as well

443
00:17:57,559 --> 00:18:01,000
I don't think that that many people will

444
00:17:59,320 --> 00:18:02,480
be reimplementing beam search so you

445
00:18:01,000 --> 00:18:04,120
might not need to worry about that too

446
00:18:02,480 --> 00:18:05,679
much but in case you are doing anything

447
00:18:04,120 --> 00:18:08,159
with respect to search algorithms it's a

448
00:18:05,679 --> 00:18:08,159
good thing to

449
00:18:08,880 --> 00:18:15,159
know

450
00:18:10,480 --> 00:18:15,159
cool um any questions about these two so

451
00:18:16,919 --> 00:18:24,159
far no okay um so the second the next

452
00:18:22,600 --> 00:18:25,400
thing I want to talk about this is

453
00:18:24,159 --> 00:18:27,840
something that people think about a

454
00:18:25,400 --> 00:18:29,400
little bit less uh but it's actually

455
00:18:27,840 --> 00:18:31,280
something really important to know

456
00:18:29,400 --> 00:18:34,280
because it will affect you it will

457
00:18:31,280 --> 00:18:35,799
affect everybody uh to some extent it

458
00:18:34,280 --> 00:18:40,760
will affect you to a greater or lesser

459
00:18:35,799 --> 00:18:41,520
extent depending on um what uh type of

460
00:18:40,760 --> 00:18:44,480
you

461
00:18:41,520 --> 00:18:46,799
know system you're building but it will

462
00:18:44,480 --> 00:18:48,760
definitely affect everybody and that's

463
00:18:46,799 --> 00:18:50,960
the mismatch between the the function

464
00:18:48,760 --> 00:18:53,440
that you're optimizing at training time

465
00:18:50,960 --> 00:18:55,240
and the evaluation metric that you're

466
00:18:53,440 --> 00:18:58,000
evaluating and

467
00:18:55,240 --> 00:18:59,679
so uh like as I said in the

468
00:18:58,000 --> 00:19:01,679
reinforcement learning class it's very

469
00:18:59,679 --> 00:19:03,640
common to optimize for maximum

470
00:19:01,679 --> 00:19:06,039
likelihood for training uh but there's

471
00:19:03,640 --> 00:19:07,840
all kinds of problems with this you know

472
00:19:06,039 --> 00:19:09,640
um with respect to the mistake it not

473
00:19:07,840 --> 00:19:11,640
being sensitive to mistakes it not being

474
00:19:09,640 --> 00:19:14,799
sensitive to your generation

475
00:19:11,640 --> 00:19:16,520
algorithm um but even though your

476
00:19:14,799 --> 00:19:19,880
likelihood is getting better accuracy

477
00:19:16,520 --> 00:19:22,799
can get worse and this is a super simple

478
00:19:19,880 --> 00:19:25,080
example with uh image classification on

479
00:19:22,799 --> 00:19:27,919
mest and I I ran this experiment with

480
00:19:25,080 --> 00:19:30,880
like 10 lines of pytorch code or

481
00:19:27,919 --> 00:19:36,840
something like this uh maybe more like

482
00:19:30,880 --> 00:19:40,080
40 lines of P um and so here um on the

483
00:19:36,840 --> 00:19:43,120
left side we have the loss on the

484
00:19:40,080 --> 00:19:46,600
training set and the test set or the dev

485
00:19:43,120 --> 00:19:48,559
set and here we have accuracy on the

486
00:19:46,600 --> 00:19:50,799
training set in the test

487
00:19:48,559 --> 00:19:55,000
set

488
00:19:50,799 --> 00:19:56,159
and so oops I showed you the answer so I

489
00:19:55,000 --> 00:19:58,799
was going to do a quiz but I

490
00:19:56,159 --> 00:20:00,559
accidentally showed you the answer um

491
00:19:58,799 --> 00:20:04,440
but the problem here is basically

492
00:20:00,559 --> 00:20:06,320
because um the the loss you're

493
00:20:04,440 --> 00:20:09,400
calculating the likelihood of the

494
00:20:06,320 --> 00:20:11,120
correct answer and the likelihood of the

495
00:20:09,400 --> 00:20:12,440
correct answer is the probability of

496
00:20:11,120 --> 00:20:15,000
getting the correct

497
00:20:12,440 --> 00:20:17,240
answer the accuracy is the number of

498
00:20:15,000 --> 00:20:20,280
times you're getting the correct answer

499
00:20:17,240 --> 00:20:23,799
so as you train a model to get more and

500
00:20:20,280 --> 00:20:25,440
more confident it gets better it gets

501
00:20:23,799 --> 00:20:27,840
better and better at getting more

502
00:20:25,440 --> 00:20:30,039
answers correct but it also gets more

503
00:20:27,840 --> 00:20:33,360
and more confident in its answers and so

504
00:20:30,039 --> 00:20:36,200
if the you know there's any example that

505
00:20:33,360 --> 00:20:37,840
it's really bad at um it might get very

506
00:20:36,200 --> 00:20:42,320
confident in

507
00:20:37,840 --> 00:20:44,760
that answer that bad answer and the log

508
00:20:42,320 --> 00:20:47,320
likelihood of that answer will go up or

509
00:20:44,760 --> 00:20:49,679
sorry the log likelihood will go down so

510
00:20:47,320 --> 00:20:54,360
the negative log likelihood will go up

511
00:20:49,679 --> 00:20:56,720
is the loss so basically

512
00:20:54,360 --> 00:20:59,559
um the

513
00:20:56,720 --> 00:21:01,039
uh the loss that you're calculating and

514
00:20:59,559 --> 00:21:03,840
the thing that you care about in the end

515
00:21:01,039 --> 00:21:07,120
accuracy can be decorrelated

516
00:21:03,840 --> 00:21:09,520
um so there's also an interesting

517
00:21:07,120 --> 00:21:12,080
example um in text generation and this

518
00:21:09,520 --> 00:21:14,000
is part of the reason why uh we have all

519
00:21:12,080 --> 00:21:15,880
these other text generation algorithms

520
00:21:14,000 --> 00:21:20,080
like nucleus samp playing or topk samp

521
00:21:15,880 --> 00:21:23,039
playing or other things like this is um

522
00:21:20,080 --> 00:21:25,080
actually in a maximum likelihood trained

523
00:21:23,039 --> 00:21:27,799
model better

524
00:21:25,080 --> 00:21:29,559
search uh in in other words finding a

525
00:21:27,799 --> 00:21:32,159
better model scope

526
00:21:29,559 --> 00:21:36,120
doesn't necessarily give you a better

527
00:21:32,159 --> 00:21:37,840
generation result and this is an example

528
00:21:36,120 --> 00:21:39,080
uh from machine translation from a

529
00:21:37,840 --> 00:21:41,880
really long time

530
00:21:39,080 --> 00:21:44,000
ago uh but you know it still persists

531
00:21:41,880 --> 00:21:47,520
today which is they did beam search with

532
00:21:44,000 --> 00:21:53,600
a larger and larger beam

533
00:21:47,520 --> 00:21:56,640
and the be the best Beam for finding um

534
00:21:53,600 --> 00:21:59,640
the best scoring output basically was

535
00:21:56,640 --> 00:22:01,600
four and then the accuracy goes down and

536
00:21:59,640 --> 00:22:05,559
down and down as they find a better

537
00:22:01,600 --> 00:22:07,200
output and does anyone remember when we

538
00:22:05,559 --> 00:22:09,679
talked about the generation class where

539
00:22:07,200 --> 00:22:09,679
this comes

540
00:22:10,120 --> 00:22:15,000
from I don't know how explicitly we said

541
00:22:12,960 --> 00:22:18,600
we mentioned it in the generation class

542
00:22:15,000 --> 00:22:20,360
but basically the problem is um maximum

543
00:22:18,600 --> 00:22:22,559
likelihood train models like shorter

544
00:22:20,360 --> 00:22:25,240
outputs generally because if as we make

545
00:22:22,559 --> 00:22:27,760
the output longer uh the probability of

546
00:22:25,240 --> 00:22:29,679
the longer outputs goes down so as you

547
00:22:27,760 --> 00:22:32,039
improve the beam it will start

548
00:22:29,679 --> 00:22:34,799
generating shorter and shorter outputs

549
00:22:32,039 --> 00:22:36,480
and because of that the score goes down

550
00:22:34,799 --> 00:22:39,039
because blue score doesn't like outputs

551
00:22:36,480 --> 00:22:41,520
that are too short essentially so there

552
00:22:39,039 --> 00:22:44,039
are um there are hex around this for

553
00:22:41,520 --> 00:22:46,200
beam search where essentially what you

554
00:22:44,039 --> 00:22:48,559
do is you uh take the average log

555
00:22:46,200 --> 00:22:51,159
likelihood of each token instead of the

556
00:22:48,559 --> 00:22:52,760
overall log likelihood of each token um

557
00:22:51,159 --> 00:22:54,679
and that improves a little bit but still

558
00:22:52,760 --> 00:22:59,720
you can see as you search more the the

559
00:22:54,679 --> 00:23:01,440
accuracy goes down so um so that's the

560
00:22:59,720 --> 00:23:04,039
the general idea

561
00:23:01,440 --> 00:23:08,760
here there's a bunch of ways you can fix

562
00:23:04,039 --> 00:23:10,600
this um the most principled way is to

563
00:23:08,760 --> 00:23:12,760
use a method like reinforcement learning

564
00:23:10,600 --> 00:23:14,120
or something uh some sort of you know

565
00:23:12,760 --> 00:23:15,520
structured training algorithm that

566
00:23:14,120 --> 00:23:17,159
allows you to train your models so that

567
00:23:15,520 --> 00:23:20,159
you don't get these bad

568
00:23:17,159 --> 00:23:22,159
outputs um another way that's much

569
00:23:20,159 --> 00:23:25,640
easier is to do early stopping with the

570
00:23:22,159 --> 00:23:30,480
evaluation metric as opposed to um early

571
00:23:25,640 --> 00:23:32,840
stopping with the loss and by doing this

572
00:23:30,480 --> 00:23:34,520
you would stop here so you would stop

573
00:23:32,840 --> 00:23:37,159
where you get the highest evaluation

574
00:23:34,520 --> 00:23:42,600
metric uh that you care about instead of

575
00:23:37,159 --> 00:23:44,400
stopping here uh so that's um that's one

576
00:23:42,600 --> 00:23:46,600
way you can fix this

577
00:23:44,400 --> 00:23:49,760
problem does anyone have an idea about

578
00:23:46,600 --> 00:23:49,760
why this might be a bad

579
00:23:49,840 --> 00:23:57,159
idea why might it be a bad idea to stop

580
00:23:52,480 --> 00:23:57,159
here instead of stopping here for

581
00:23:57,440 --> 00:24:00,440
example

582
00:24:05,320 --> 00:24:10,200
yeah it's kind of overfitting it's

583
00:24:07,760 --> 00:24:13,640
overfitting in a particular way um but

584
00:24:10,200 --> 00:24:16,000
remember here this is still the accuracy

585
00:24:13,640 --> 00:24:18,400
on the dev set so we're not overfitting

586
00:24:16,000 --> 00:24:20,080
so much that the dev accuracy is going

587
00:24:18,400 --> 00:24:24,279
down that would be a different variety

588
00:24:20,080 --> 00:24:27,360
of overfitting but any any

589
00:24:24,279 --> 00:24:29,799
ideas go for it we don't want to be too

590
00:24:27,360 --> 00:24:31,600
confident yeah exactly we don't want it

591
00:24:29,799 --> 00:24:32,880
to be too confident in its wrong answers

592
00:24:31,600 --> 00:24:35,279
and we talked about

593
00:24:32,880 --> 00:24:38,000
calibration um where calibration is

594
00:24:35,279 --> 00:24:40,039
basically like how accurate are the

595
00:24:38,000 --> 00:24:41,480
probability estimates so this model over

596
00:24:40,039 --> 00:24:43,600
here is going to be really poorly

597
00:24:41,480 --> 00:24:45,159
calibrated it's going to be very

598
00:24:43,600 --> 00:24:46,240
confident regardless of whether it's

599
00:24:45,159 --> 00:24:49,440
correct or not and that could be a

600
00:24:46,240 --> 00:24:50,840
problem in dopram uh dopram tasks

601
00:24:49,440 --> 00:24:52,130
there's also another thing that I I

602
00:24:50,840 --> 00:24:55,189
forgot to put on

603
00:24:52,130 --> 00:24:55,189
[Music]

604
00:24:57,320 --> 00:25:00,320
um

605
00:25:02,919 --> 00:25:08,120
that I forgot to put on the slides but

606
00:25:04,520 --> 00:25:10,720
it's a um an interesting phenomenon that

607
00:25:08,120 --> 00:25:12,720
actually um kind of a lot of people in

608
00:25:10,720 --> 00:25:16,360
interpretability are interested in it's

609
00:25:12,720 --> 00:25:18,120
this uh generalization gring

610
00:25:16,360 --> 00:25:19,640
generalization Beyond overfitting on

611
00:25:18,120 --> 00:25:21,120
small algorithmic data sets and

612
00:25:19,640 --> 00:25:27,360
basically what they

613
00:25:21,120 --> 00:25:29,720
show is um you can be training for a

614
00:25:27,360 --> 00:25:31,320
very very long time

615
00:25:29,720 --> 00:25:34,279
um

616
00:25:31,320 --> 00:25:35,919
and uh like reducing the loss reducing

617
00:25:34,279 --> 00:25:40,399
the loss reducing the loss and reducing

618
00:25:35,919 --> 00:25:42,480
the loss and it's only after a very long

619
00:25:40,399 --> 00:25:43,840
time does your Model start generalizing

620
00:25:42,480 --> 00:25:48,240
well and getting good

621
00:25:43,840 --> 00:25:49,799
accuracy um the this paper the types of

622
00:25:48,240 --> 00:25:52,120
data sets it's talking about are data

623
00:25:49,799 --> 00:25:55,520
sets where you need to get many things

624
00:25:52,120 --> 00:25:58,640
in a row correct before you get the

625
00:25:55,520 --> 00:26:00,880
final answer correct so basically you

626
00:25:58,640 --> 00:26:02,320
need to get like 20 steps in a row or 50

627
00:26:00,880 --> 00:26:06,200
steps in a row correct before you get

628
00:26:02,320 --> 00:26:10,679
the final answer correct and um

629
00:26:06,200 --> 00:26:13,000
basically the reason why this happens is

630
00:26:10,679 --> 00:26:15,720
because this accuracy will keep going up

631
00:26:13,000 --> 00:26:17,760
but you only get the accuracy of each

632
00:26:15,720 --> 00:26:20,520
individual decision will keep going up

633
00:26:17,760 --> 00:26:22,880
but you only get marked like

634
00:26:20,520 --> 00:26:25,440
correct uh

635
00:26:22,880 --> 00:26:29,799
after you get like all 50 in a row

636
00:26:25,440 --> 00:26:31,200
correct so um it this difference can be

637
00:26:29,799 --> 00:26:33,039
even more Stark when you're talking

638
00:26:31,200 --> 00:26:35,399
about things that require like 50 steps

639
00:26:33,039 --> 00:26:37,399
of reasoning or like multiple steps of

640
00:26:35,399 --> 00:26:39,559
reasoning but like 50 token Generations

641
00:26:37,399 --> 00:26:42,679
correct before you get them right so um

642
00:26:39,559 --> 00:26:42,679
that's another thing to be aware

643
00:26:43,000 --> 00:26:49,240
of cool um so now I want to switch gears

644
00:26:46,960 --> 00:26:51,919
a little bit to actionable evaluation

645
00:26:49,240 --> 00:26:54,240
and how you can um evaluate your models

646
00:26:51,919 --> 00:26:56,640
in a way that makes it easy to find uh

647
00:26:54,240 --> 00:26:58,600
next steps to be

648
00:26:56,640 --> 00:27:00,159
improving uh are there any questions

649
00:26:58,600 --> 00:27:02,600
about the debugging part before we get

650
00:27:00,159 --> 00:27:02,600
into this

651
00:27:03,360 --> 00:27:10,120
part okay I'll

652
00:27:05,880 --> 00:27:12,840
go so um my first suggestion with

653
00:27:10,120 --> 00:27:15,559
respect to how you can actually you know

654
00:27:12,840 --> 00:27:17,440
improve systems is make sure that you're

655
00:27:15,559 --> 00:27:21,039
looking at the data that you're

656
00:27:17,440 --> 00:27:22,679
using and um both bugs and new research

657
00:27:21,039 --> 00:27:24,080
directions can be found by looking at

658
00:27:22,679 --> 00:27:27,159
your model

659
00:27:24,080 --> 00:27:31,640
outputs um

660
00:27:27,159 --> 00:27:33,279
so to give one example um of a very

661
00:27:31,640 --> 00:27:36,200
common mistake that you can make when

662
00:27:33,279 --> 00:27:40,159
you're creating a a generation algorithm

663
00:27:36,200 --> 00:27:41,600
it's these sort of off by one erors um

664
00:27:40,159 --> 00:27:43,919
so like let's say you implemented a

665
00:27:41,600 --> 00:27:46,039
translation system and it's generating

666
00:27:43,919 --> 00:27:49,440
outputs like went to the store yesterday

667
00:27:46,039 --> 00:27:51,080
bought a dog um you can immediately look

668
00:27:49,440 --> 00:27:53,440
at this and say hey this doesn't look

669
00:27:51,080 --> 00:27:58,360
like natural English what's going uh

670
00:27:53,440 --> 00:28:00,000
what's going on and the the problem here

671
00:27:58,360 --> 00:28:04,600
is

672
00:28:00,000 --> 00:28:04,600
you're um you're doing something

673
00:28:05,159 --> 00:28:12,720
like output uh

674
00:28:09,240 --> 00:28:14,600
one uh and you have a slice of like one

675
00:28:12,720 --> 00:28:17,399
instead of zero here or something like

676
00:28:14,600 --> 00:28:18,640
this and so this is a really silly error

677
00:28:17,399 --> 00:28:21,000
that you might just make a mistake on

678
00:28:18,640 --> 00:28:23,679
python on your you know pre-processing

679
00:28:21,000 --> 00:28:26,200
or postprocessing or something like this

680
00:28:23,679 --> 00:28:28,399
um but the problem is like if you look

681
00:28:26,200 --> 00:28:30,600
at your blue score based evaluation or

682
00:28:28,399 --> 00:28:32,840
something like that you'll have like

683
00:28:30,600 --> 00:28:34,760
you'll be one point worse or two points

684
00:28:32,840 --> 00:28:36,720
worse or something like that and you'll

685
00:28:34,760 --> 00:28:38,600
be like Oh I'm I'm two points worse why

686
00:28:36,720 --> 00:28:40,600
am I two point wor two points worse in

687
00:28:38,600 --> 00:28:43,760
the state of the art and it turns out it

688
00:28:40,600 --> 00:28:45,279
was a really like silly thing like this

689
00:28:43,760 --> 00:28:46,519
and immediately you'll see this if you

690
00:28:45,279 --> 00:28:47,960
look at your data but if you're doing

691
00:28:46,519 --> 00:28:49,600
all your experiments and just looking at

692
00:28:47,960 --> 00:28:51,519
the numbers it's really hard to tell you

693
00:28:49,600 --> 00:28:53,720
know why this is

694
00:28:51,519 --> 00:28:58,720
happening

695
00:28:53,720 --> 00:29:02,360
um another thing is uh if you

696
00:28:58,720 --> 00:29:04,799
have a good eye and can like just look

697
00:29:02,360 --> 00:29:07,799
through the data points

698
00:29:04,799 --> 00:29:09,640
um we as humans are pretty good uh

699
00:29:07,799 --> 00:29:14,200
pattern recognizers and especially you

700
00:29:09,640 --> 00:29:16,360
know CMU students uh you're uh very good

701
00:29:14,200 --> 00:29:18,519
and quick at picking up on things so if

702
00:29:16,360 --> 00:29:20,600
you look at the data and pour through

703
00:29:18,519 --> 00:29:22,880
things you can uh probably pick up

704
00:29:20,600 --> 00:29:24,880
patterns about why things are failing

705
00:29:22,880 --> 00:29:27,720
and so um you know you might look and

706
00:29:24,880 --> 00:29:29,919
see that uh compared to some other model

707
00:29:27,720 --> 00:29:31,679
your model is really bad at answering

708
00:29:29,919 --> 00:29:33,679
questions about people or something like

709
00:29:31,679 --> 00:29:36,480
that and then you figure out you'll need

710
00:29:33,679 --> 00:29:38,320
a better model of uh people or your rag

711
00:29:36,480 --> 00:29:40,519
systems uh that you're building for

712
00:29:38,320 --> 00:29:42,880
assignment two is maybe failing on all

713
00:29:40,519 --> 00:29:45,559
the research related questions so you

714
00:29:42,880 --> 00:29:47,080
need to come up with the research uh

715
00:29:45,559 --> 00:29:48,320
like scrape more research data or

716
00:29:47,080 --> 00:29:50,080
something like

717
00:29:48,320 --> 00:29:53,840
that

718
00:29:50,080 --> 00:29:55,760
um so there are methods to do this more

719
00:29:53,840 --> 00:29:58,039
systematically and this is something I

720
00:29:55,760 --> 00:29:59,720
picked up when I was doing an internship

721
00:29:58,039 --> 00:30:04,080
at Google and it really stuck with me

722
00:29:59,720 --> 00:30:09,080
for you know 14 uh 14 years now I guess

723
00:30:04,080 --> 00:30:10,960
13 years um so uh a very simple way to

724
00:30:09,080 --> 00:30:12,600
do this more systematically than just

725
00:30:10,960 --> 00:30:16,200
browsing through things is to randomly

726
00:30:12,600 --> 00:30:19,000
sample a 100 outputs and look at a 100

727
00:30:16,200 --> 00:30:21,840
erors and try to group them into some

728
00:30:19,000 --> 00:30:23,799
sort of typology and say oh uh this kind

729
00:30:21,840 --> 00:30:27,799
of air is particularly

730
00:30:23,799 --> 00:30:31,279
frequent and this is just one example of

731
00:30:27,799 --> 00:30:33,120
a typology that was defined by V at all

732
00:30:31,279 --> 00:30:37,320
um where they tried to take machine

733
00:30:33,120 --> 00:30:39,480
translation errors and group them into

734
00:30:37,320 --> 00:30:43,440
uh various varieties like correct words

735
00:30:39,480 --> 00:30:46,640
filler words local uh local range long

736
00:30:43,440 --> 00:30:48,440
range um uh sorry word word level word

737
00:30:46,640 --> 00:30:50,440
ordering erors local range long range

738
00:30:48,440 --> 00:30:54,279
phrase level local range long range and

739
00:30:50,440 --> 00:30:55,679
stuff like this um you can definitely

740
00:30:54,279 --> 00:30:58,399
look at previous work and see the

741
00:30:55,679 --> 00:31:00,559
typologies of errors that they used but

742
00:30:58,399 --> 00:31:02,440
the problem is like systems get better

743
00:31:00,559 --> 00:31:04,240
and actually I don't think this is a

744
00:31:02,440 --> 00:31:06,760
super relevant typology for machine

745
00:31:04,240 --> 00:31:10,120
translation anymore uh because machine

746
00:31:06,760 --> 00:31:12,159
translation systems like they don't make

747
00:31:10,120 --> 00:31:14,639
a whole lot of local range Word level

748
00:31:12,159 --> 00:31:16,159
errors anymore and rather we might want

749
00:31:14,639 --> 00:31:18,279
to know more fine grain like are they

750
00:31:16,159 --> 00:31:21,720
making mistakes on named entities or

751
00:31:18,279 --> 00:31:24,720
other things like that so actually

752
00:31:21,720 --> 00:31:24,720
we

753
00:31:26,919 --> 00:31:29,919
um

754
00:31:30,519 --> 00:31:36,279
did a re a more recent thing it's I

755
00:31:34,279 --> 00:31:39,159
guess four years ago now um but it was

756
00:31:36,279 --> 00:31:42,720
when uh people first started saying that

757
00:31:39,159 --> 00:31:46,200
machine translation systems are about as

758
00:31:42,720 --> 00:31:50,720
good as humans at doing a

759
00:31:46,200 --> 00:31:50,720
translation and when we did this we

760
00:31:52,480 --> 00:31:58,440
compared we compared machine translation

761
00:31:55,200 --> 00:31:59,960
systems to humans and we tried to find

762
00:31:58,440 --> 00:32:02,240
you know different types of things and

763
00:31:59,960 --> 00:32:03,919
we were inspired by V but we recreated

764
00:32:02,240 --> 00:32:06,159
our typology based on the things that we

765
00:32:03,919 --> 00:32:10,279
thought were you know the most important

766
00:32:06,159 --> 00:32:13,399
types of errors in like 2020 instead of

767
00:32:10,279 --> 00:32:16,799
2006 so this is really helpful the

768
00:32:13,399 --> 00:32:19,039
reason why it's really helpful is if you

769
00:32:16,799 --> 00:32:20,440
can do this even for a small sample of

770
00:32:19,039 --> 00:32:23,440
the outputs that you're looking at and

771
00:32:20,440 --> 00:32:25,279
identify the most like prominent types

772
00:32:23,440 --> 00:32:27,440
of eras that you're facing it often

773
00:32:25,279 --> 00:32:29,360
leads you to the most successful ways of

774
00:32:27,440 --> 00:32:31,519
improving the accuracy of your systems

775
00:32:29,360 --> 00:32:33,120
because you might if you don't do this

776
00:32:31,519 --> 00:32:35,000
you might be focusing on an air type

777
00:32:33,120 --> 00:32:38,000
that's not actually an error it's kind

778
00:32:35,000 --> 00:32:39,200
of like if you learned in uh programming

779
00:32:38,000 --> 00:32:40,799
you know software engineering or

780
00:32:39,200 --> 00:32:42,639
something like that you should never

781
00:32:40,799 --> 00:32:46,360
optimize your code until you run a

782
00:32:42,639 --> 00:32:47,799
profiler um because actually your code

783
00:32:46,360 --> 00:32:50,320
might be slow in a place that you never

784
00:32:47,799 --> 00:32:52,720
expected and so it's kind of the same

785
00:32:50,320 --> 00:32:56,600
principle here right so don't optimize

786
00:32:52,720 --> 00:32:58,720
your systems errors in a place uh where

787
00:32:56,600 --> 00:33:03,240
like actually it's not having in years

788
00:32:58,720 --> 00:33:06,440
so um that's a general principle

789
00:33:03,240 --> 00:33:09,440
here uh cool another thing you can do is

790
00:33:06,440 --> 00:33:11,760
quantitative analysis so um if you can

791
00:33:09,440 --> 00:33:13,880
think of the phenomenon that you choose

792
00:33:11,760 --> 00:33:17,480
to focus on um is that phenomenon

793
00:33:13,880 --> 00:33:19,159
getting better so if you focused on uh

794
00:33:17,480 --> 00:33:22,240
something that should improve the

795
00:33:19,159 --> 00:33:23,760
quality of low frequency words uh you

796
00:33:22,240 --> 00:33:26,200
can check if the accuracy on low

797
00:33:23,760 --> 00:33:27,399
frequency words is increasing if you

798
00:33:26,200 --> 00:33:29,600
focused on something that should be

799
00:33:27,399 --> 00:33:32,120
improving the syntax in a low resource

800
00:33:29,600 --> 00:33:36,080
language you can measure um whether it's

801
00:33:32,120 --> 00:33:37,360
doing better on word ordering or uh long

802
00:33:36,080 --> 00:33:41,840
distance

803
00:33:37,360 --> 00:33:44,360
dependencies um if you focused on

804
00:33:41,840 --> 00:33:46,039
improving a search algorithm for you

805
00:33:44,360 --> 00:33:47,519
know generation or something like that

806
00:33:46,039 --> 00:33:49,880
are the number of search errors that

807
00:33:47,519 --> 00:33:53,120
you're encountering being reduced so

808
00:33:49,880 --> 00:33:56,320
depending on what you planned on uh you

809
00:33:53,120 --> 00:33:57,919
know improving it's often a good idea to

810
00:33:56,320 --> 00:33:59,480
measure more directly whether it's

811
00:33:57,919 --> 00:34:00,559
improving the the thing that you think

812
00:33:59,480 --> 00:34:04,880
it should

813
00:34:00,559 --> 00:34:06,000
improve um one example of um so I I

814
00:34:04,880 --> 00:34:09,240
basically

815
00:34:06,000 --> 00:34:11,240
created since my experience doing this

816
00:34:09,240 --> 00:34:15,159
manually uh when I I was on an

817
00:34:11,240 --> 00:34:18,280
internship at Google um I've

818
00:34:15,159 --> 00:34:20,639
gradually improved my methodology for

819
00:34:18,280 --> 00:34:20,639
doing

820
00:34:21,679 --> 00:34:26,320
this and um and worked on automating

821
00:34:24,879 --> 00:34:30,599
things and

822
00:34:26,320 --> 00:34:33,839
so the first thing I had was a super

823
00:34:30,599 --> 00:34:35,560
hacky uh hacky script that basically

824
00:34:33,839 --> 00:34:37,720
writes out HTML

825
00:34:35,560 --> 00:34:39,320
files um and then I I had something

826
00:34:37,720 --> 00:34:42,320
called explainer board where we had a

827
00:34:39,320 --> 00:34:44,879
leader board and uh recently one of the

828
00:34:42,320 --> 00:34:47,800
things I've worked on is uh this uh

829
00:34:44,879 --> 00:34:53,200
together with um Alex Alex Cabrera who's

830
00:34:47,800 --> 00:34:56,760
a student here um is this toolkit called

831
00:34:53,200 --> 00:34:59,640
Zeno and um this is just an example from

832
00:34:56,760 --> 00:34:59,640
machine translation

833
00:35:03,440 --> 00:35:09,200
it's being a little bit

834
00:35:06,599 --> 00:35:11,079
slow um but basically what it does is it

835
00:35:09,200 --> 00:35:14,920
allows you to look at the data on the

836
00:35:11,079 --> 00:35:18,000
right side um and so these are just

837
00:35:14,920 --> 00:35:19,680
examples um but you can go in and do

838
00:35:18,000 --> 00:35:22,760
things like say Okay I want to look at

839
00:35:19,680 --> 00:35:24,640
all machine translation examples

840
00:35:22,760 --> 00:35:28,040
from

841
00:35:24,640 --> 00:35:30,920
uh housea and so it shows you the ones

842
00:35:28,040 --> 00:35:32,960
from housea I want to look

843
00:35:30,920 --> 00:35:36,240
at all

844
00:35:32,960 --> 00:35:38,880
examples let me clear that off I want to

845
00:35:36,240 --> 00:35:40,800
look at all examples where the accuracy

846
00:35:38,880 --> 00:35:43,440
is

847
00:35:40,800 --> 00:35:45,280
low um and so now I can look at all the

848
00:35:43,440 --> 00:35:49,640
examples where the accuracy is low and I

849
00:35:45,280 --> 00:35:52,640
I can go in and uh uh examine them so uh

850
00:35:49,640 --> 00:35:54,880
you can also go in and build charts like

851
00:35:52,640 --> 00:35:58,280
this so like what is the overall

852
00:35:54,880 --> 00:36:02,200
performance um what is is the

853
00:35:58,280 --> 00:36:05,960
performance what is the performance

854
00:36:02,200 --> 00:36:07,520
um on different scripts so you can see

855
00:36:05,960 --> 00:36:10,880
which model which model is doing better

856
00:36:07,520 --> 00:36:13,960
at scripts and stuff like that so um or

857
00:36:10,880 --> 00:36:16,000
you can put things side by side and say

858
00:36:13,960 --> 00:36:20,720
okay I want to find all the examples

859
00:36:16,000 --> 00:36:21,800
where uh chat GPT is doing much worse

860
00:36:20,720 --> 00:36:25,280
than GPT

861
00:36:21,800 --> 00:36:28,240
4 uh or like GPT 3.5 is doing much worse

862
00:36:25,280 --> 00:36:29,680
than gp4 and here we can see that oh in

863
00:36:28,240 --> 00:36:31,520
this case it's generating something in

864
00:36:29,680 --> 00:36:34,079
the wrong script or something like that

865
00:36:31,520 --> 00:36:37,839
so um there's also tooling that you can

866
00:36:34,079 --> 00:36:40,480
use to make this easier as

867
00:36:37,839 --> 00:36:43,520
well and the way uh the way you use this

868
00:36:40,480 --> 00:36:46,079
is you basically

869
00:36:43,520 --> 00:36:48,000
um uh create a pandas data frame with

870
00:36:46,079 --> 00:36:49,680
all of your data in it and you upload

871
00:36:48,000 --> 00:36:52,400
the pandas data frame with any metadata

872
00:36:49,680 --> 00:36:54,280
you want to use and you can uh use and I

873
00:36:52,400 --> 00:36:56,520
think VJ will be having a recitation on

874
00:36:54,280 --> 00:37:02,560
this if you're interested in taking a

875
00:36:56,520 --> 00:37:04,680
look cool um so that is the my part and

876
00:37:02,560 --> 00:37:07,760
then we'll be doing Nishant next while

877
00:37:04,680 --> 00:37:09,480
Nishant comes up to set up are there any

878
00:37:07,760 --> 00:37:10,520
questions about the thing that I talked

879
00:37:09,480 --> 00:37:14,079
about

880
00:37:10,520 --> 00:37:14,079
here yeah

881
00:37:14,359 --> 00:37:18,200
so that when I

882
00:37:26,200 --> 00:37:30,079
regular um

883
00:37:28,160 --> 00:37:32,560
is that does that make a difference in

884
00:37:30,079 --> 00:37:35,400
terms of like what we're expecting when

885
00:37:32,560 --> 00:37:38,800
we're evaluating the model

886
00:37:35,400 --> 00:37:41,720
model yeah so just to repeat the

887
00:37:38,800 --> 00:37:43,680
question it's a a great question so if

888
00:37:41,720 --> 00:37:49,440
you apply

889
00:37:43,680 --> 00:37:49,440
regularization um will that change the

890
00:37:49,640 --> 00:37:54,079
overall expectation for the model loss

891
00:37:52,040 --> 00:37:55,680
so I was saying loss should converge to

892
00:37:54,079 --> 00:37:57,200
zero once you start applying

893
00:37:55,680 --> 00:37:59,079
regularization or weight Decay or

894
00:37:57,200 --> 00:38:02,640
something like that it definitely might

895
00:37:59,079 --> 00:38:04,520
not converge to Z um and the reason why

896
00:38:02,640 --> 00:38:06,520
is because once you start applying

897
00:38:04,520 --> 00:38:09,319
regularization there is no zero loss

898
00:38:06,520 --> 00:38:11,480
solion um because in order to reduce the

899
00:38:09,319 --> 00:38:14,960
loss you need to make move things away

900
00:38:11,480 --> 00:38:16,359
move weights away from zero um but when

901
00:38:14,960 --> 00:38:19,560
you move weights away from zero the

902
00:38:16,359 --> 00:38:22,200
regularization L becomes n negative so

903
00:38:19,560 --> 00:38:24,599
one thing you can do however is measure

904
00:38:22,200 --> 00:38:26,880
the losses separately so measure the

905
00:38:24,599 --> 00:38:27,960
regularization component of the loss and

906
00:38:26,880 --> 00:38:29,760
the um

907
00:38:27,960 --> 00:38:31,920
the log like we had component with the

908
00:38:29,760 --> 00:38:33,560
loss and with any reasonable

909
00:38:31,920 --> 00:38:35,280
regularization and a reasonably

910
00:38:33,560 --> 00:38:38,000
parameterized model I do think the loss

911
00:38:35,280 --> 00:38:39,760
should be getting closer to Zer like the

912
00:38:38,000 --> 00:38:41,920
actual likely should be getting closer

913
00:38:39,760 --> 00:38:41,920
to

914
00:38:42,200 --> 00:38:46,520
zero uh you were using an extremely

915
00:38:44,480 --> 00:38:49,240
small model in the mini L signed though

916
00:38:46,520 --> 00:38:53,680
so that might make it more

917
00:38:49,240 --> 00:38:56,440
difficult yeah and any other

918
00:38:53,680 --> 00:38:59,440
things okay if not

919
00:38:56,440 --> 00:38:59,440
I'll

920
00:39:13,720 --> 00:39:19,160
all right can everyone hear

921
00:39:15,319 --> 00:39:21,440
me sweet okay move this it looks like

922
00:39:19,160 --> 00:39:24,200
I'm talking to someone instead of

923
00:39:21,440 --> 00:39:24,200
between both of

924
00:39:26,359 --> 00:39:29,359
you

925
00:39:33,319 --> 00:39:37,680
all right so hi everyone um I'm going to

926
00:39:35,720 --> 00:39:39,400
talk about model interpretability for

927
00:39:37,680 --> 00:39:41,680
for those who don't know me I'm one of

928
00:39:39,400 --> 00:39:44,359
your Tas I'm a first year PhD student

929
00:39:41,680 --> 00:39:47,359
working with Mona diab on model

930
00:39:44,359 --> 00:39:47,359
interpretability

931
00:39:48,800 --> 00:39:55,400
um where what do I

932
00:39:51,839 --> 00:39:59,119
click your your mouse should be there

933
00:39:55,400 --> 00:40:01,599
yeah just

934
00:39:59,119 --> 00:40:04,160
cool okay um

935
00:40:01,599 --> 00:40:06,079
so what I want you to take away if you

936
00:40:04,160 --> 00:40:08,359
if you fall asleep this is too boring

937
00:40:06,079 --> 00:40:09,839
here are sort of the two main takeaways

938
00:40:08,359 --> 00:40:12,040
one I want to convince you that model

939
00:40:09,839 --> 00:40:14,720
interpretability is important to study

940
00:40:12,040 --> 00:40:16,720
and two I want I want you to find this

941
00:40:14,720 --> 00:40:18,880
interesting um and something you want to

942
00:40:16,720 --> 00:40:20,079
explore more there's a bunch of details

943
00:40:18,880 --> 00:40:21,800
here this is going to be kind of a

944
00:40:20,079 --> 00:40:24,599
whirlwind tour you're not going to get

945
00:40:21,800 --> 00:40:27,440
super deep into anything um so hopefully

946
00:40:24,599 --> 00:40:28,839
this acts as a starting point um then

947
00:40:27,440 --> 00:40:33,800
than anything

948
00:40:28,839 --> 00:40:37,040
else so interpretability in AI um the

949
00:40:33,800 --> 00:40:38,480
the definition is it's the study of

950
00:40:37,040 --> 00:40:40,440
understanding the decisions that AI

951
00:40:38,480 --> 00:40:42,640
systems make and putting them into

952
00:40:40,440 --> 00:40:44,280
easily human understandable terms this

953
00:40:42,640 --> 00:40:47,640
can mean a lot of different things and

954
00:40:44,280 --> 00:40:49,280
this is often really hard um and the why

955
00:40:47,640 --> 00:40:51,319
is to use that understanding to

956
00:40:49,280 --> 00:40:54,040
iteratively better Design Systems that

957
00:40:51,319 --> 00:40:56,240
are better they're more more performant

958
00:40:54,040 --> 00:40:59,240
but also those that are more human

959
00:40:56,240 --> 00:40:59,240
understandable

960
00:41:00,119 --> 00:41:06,599
um so interpretability is this big blah

961
00:41:03,720 --> 00:41:08,440
but there's a bunch of other uh spheres

962
00:41:06,599 --> 00:41:11,920
that intersect with it this is a super

963
00:41:08,440 --> 00:41:14,920
incomplete list uh so bear with me the

964
00:41:11,920 --> 00:41:16,560
causality and data integrate with this

965
00:41:14,920 --> 00:41:19,000
there's aspects that are interpretable

966
00:41:16,560 --> 00:41:20,480
there's aspects that matter here um

967
00:41:19,000 --> 00:41:22,400
explainable AI is another thing that

968
00:41:20,480 --> 00:41:24,440
you've probably heard this sits firmly

969
00:41:22,400 --> 00:41:27,800
in the interpretability blob and

970
00:41:24,440 --> 00:41:30,520
connects with ideas and causality and uh

971
00:41:27,800 --> 00:41:32,680
in data too um model interpretability

972
00:41:30,520 --> 00:41:34,200
sits on this kind of other side of

973
00:41:32,680 --> 00:41:37,680
things it intersects a little bit with

974
00:41:34,200 --> 00:41:40,000
causality and explainable AI but uh is a

975
00:41:37,680 --> 00:41:42,280
little bit separate for it um and from

976
00:41:40,000 --> 00:41:43,880
it and mechanistic interpretability

977
00:41:42,280 --> 00:41:45,400
which which you've probably heard of

978
00:41:43,880 --> 00:41:47,680
it's gotten a lot of Buzz recently kind

979
00:41:45,400 --> 00:41:48,880
of sits inside of model interpretability

980
00:41:47,680 --> 00:41:51,680
it's a special case of model

981
00:41:48,880 --> 00:41:53,160
interpretability I hope the mech people

982
00:41:51,680 --> 00:41:56,640
agree with me

983
00:41:53,160 --> 00:41:58,040
but um so yeah so historically we've

984
00:41:56,640 --> 00:42:00,880
been dealing with really really really

985
00:41:58,040 --> 00:42:03,680
small models you had Bas Nets this is a

986
00:42:00,880 --> 00:42:07,560
this is very small model um if all these

987
00:42:03,680 --> 00:42:10,000
are binary variables this is uh eight

988
00:42:07,560 --> 00:42:12,680
total parameters and only four of which

989
00:42:10,000 --> 00:42:14,880
are independent uh we also used to work

990
00:42:12,680 --> 00:42:18,160
with linear regression a lot and in the

991
00:42:14,880 --> 00:42:20,680
first case that's a nice line can be two

992
00:42:18,160 --> 00:42:23,240
parameters the multivariate case again

993
00:42:20,680 --> 00:42:25,880
that's a a small number of parameters

994
00:42:23,240 --> 00:42:27,880
we've moved to of more things we've

995
00:42:25,880 --> 00:42:30,400
moved to

996
00:42:27,880 --> 00:42:32,160
MLPs that have larger weight matrices

997
00:42:30,400 --> 00:42:33,920
but all these are kind of digestible and

998
00:42:32,160 --> 00:42:37,200
interpretable so the interpretability

999
00:42:33,920 --> 00:42:40,160
world was sort of uh not super concerned

1000
00:42:37,200 --> 00:42:41,280
with large ginormous things but we're

1001
00:42:40,160 --> 00:42:44,800
not there

1002
00:42:41,280 --> 00:42:47,000
anymore uh this is a language model this

1003
00:42:44,800 --> 00:42:50,839
is part of still part of a language

1004
00:42:47,000 --> 00:42:51,960
model now it's getting more and more and

1005
00:42:50,839 --> 00:42:55,119
more

1006
00:42:51,960 --> 00:42:57,920
hairing and this is just not

1007
00:42:55,119 --> 00:43:00,520
interpretable um I mentioned

1008
00:42:57,920 --> 00:43:03,280
on on the first day of class that I hate

1009
00:43:00,520 --> 00:43:05,240
when we update parameters of models also

1010
00:43:03,280 --> 00:43:07,720
hate when models are this big and this

1011
00:43:05,240 --> 00:43:10,000
is a six layer Transformer this is way

1012
00:43:07,720 --> 00:43:15,920
smaller than basically anything that we

1013
00:43:10,000 --> 00:43:18,040
have um and this makes things very very

1014
00:43:15,920 --> 00:43:20,920
uninterpretable um so we'll talk about

1015
00:43:18,040 --> 00:43:22,880
one one way that people sort of uh five

1016
00:43:20,920 --> 00:43:24,599
years ago started addressing this

1017
00:43:22,880 --> 00:43:25,680
problem and this is and this is the idea

1018
00:43:24,599 --> 00:43:28,000
of

1019
00:43:25,680 --> 00:43:30,880
probing so how do we make sense of a

1020
00:43:28,000 --> 00:43:35,160
giant model this is one way so we take

1021
00:43:30,880 --> 00:43:38,200
our giant model we cut the top off

1022
00:43:35,160 --> 00:43:40,520
basically um and now we have this thing

1023
00:43:38,200 --> 00:43:42,119
we stick a probe which actually in a lot

1024
00:43:40,520 --> 00:43:44,559
of cases looks very similar to a

1025
00:43:42,119 --> 00:43:47,280
language modeling head uh usually it's a

1026
00:43:44,559 --> 00:43:51,640
small two layer or one layer

1027
00:43:47,280 --> 00:43:54,319
MLP um and we basically treat the model

1028
00:43:51,640 --> 00:43:56,760
as something that uh that exists and we

1029
00:43:54,319 --> 00:44:00,240
only really care about the output of of

1030
00:43:56,760 --> 00:44:03,240
the model so more specifically what is a

1031
00:44:00,240 --> 00:44:05,720
probe it's a classifier this this green

1032
00:44:03,240 --> 00:44:07,680
thing here uh that is specifically

1033
00:44:05,720 --> 00:44:09,200
trained to predict some specific

1034
00:44:07,680 --> 00:44:11,480
property from the pre-trained models

1035
00:44:09,200 --> 00:44:16,440
representations

1036
00:44:11,480 --> 00:44:18,480
alone so um in 2019 Ian Tenny and folks

1037
00:44:16,440 --> 00:44:21,319
introduced Edge probing so this is a

1038
00:44:18,480 --> 00:44:23,240
general method um it works to probe

1039
00:44:21,319 --> 00:44:27,559
different types of information out of a

1040
00:44:23,240 --> 00:44:29,960
model so this bottom part here uh yeah

1041
00:44:27,559 --> 00:44:33,160
this bottom part here it you pass it in

1042
00:44:29,960 --> 00:44:36,520
a sequence you pass it into a model this

1043
00:44:33,160 --> 00:44:38,839
is Burt in their experiments often uh

1044
00:44:36,520 --> 00:44:40,960
and that outputs a set of contextual

1045
00:44:38,839 --> 00:44:44,359
vectors these contextual vectors can be

1046
00:44:40,960 --> 00:44:45,920
at any layer um often it's near the

1047
00:44:44,359 --> 00:44:49,280
often it's near the top but we'll talk

1048
00:44:45,920 --> 00:44:51,079
about uh the the fact that this can work

1049
00:44:49,280 --> 00:44:53,359
kind of across layers and different

1050
00:44:51,079 --> 00:44:55,599
layers and code different information

1051
00:44:53,359 --> 00:44:58,960
and on top of this you have this MLP

1052
00:44:55,599 --> 00:45:02,480
that you train to Output a prediction

1053
00:44:58,960 --> 00:45:05,599
your model is always always fixed um in

1054
00:45:02,480 --> 00:45:08,079
these cases so you can do things like

1055
00:45:05,599 --> 00:45:09,880
part of speech tagging where each

1056
00:45:08,079 --> 00:45:12,400
specific word you try to determine what

1057
00:45:09,880 --> 00:45:16,640
its part of speech is and in that case

1058
00:45:12,400 --> 00:45:18,000
this these S1 and S2 spans here uh only

1059
00:45:16,640 --> 00:45:19,440
one of them is active because you're

1060
00:45:18,000 --> 00:45:21,440
predicting for every single

1061
00:45:19,440 --> 00:45:23,240
contextualized Vector you're predicting

1062
00:45:21,440 --> 00:45:25,359
whether that thing is a noun or a verb

1063
00:45:23,240 --> 00:45:27,440
or something like this you can have

1064
00:45:25,359 --> 00:45:29,599
other sorts of tasks too like in ailment

1065
00:45:27,440 --> 00:45:32,520
where you have two sequences and two

1066
00:45:29,599 --> 00:45:35,079
spans um and you use the embeddings for

1067
00:45:32,520 --> 00:45:37,359
those spans um for like sentence one and

1068
00:45:35,079 --> 00:45:39,319
sentence two you pull them together in

1069
00:45:37,359 --> 00:45:43,359
some way and then you pass them to this

1070
00:45:39,319 --> 00:45:47,480
MLP and you see whether the MLP can uh

1071
00:45:43,359 --> 00:45:49,680
solve that test so they did this uh in

1072
00:45:47,480 --> 00:45:52,559
another paper uh Bert rediscovers the

1073
00:45:49,680 --> 00:45:54,280
NLP Pipeline and this there's a lot

1074
00:45:52,559 --> 00:45:57,079
going on in this figure the the only

1075
00:45:54,280 --> 00:45:59,599
major thing here to take away um is

1076
00:45:57,079 --> 00:46:02,720
these numbers that are in this like pink

1077
00:45:59,599 --> 00:46:05,359
purple color um so these are a bunch of

1078
00:46:02,720 --> 00:46:07,960
different uh properties supp part of

1079
00:46:05,359 --> 00:46:11,319
speech uh and and a bunch of other

1080
00:46:07,960 --> 00:46:13,520
things um and what they basically find

1081
00:46:11,319 --> 00:46:15,640
is that at earlier layers in the model

1082
00:46:13,520 --> 00:46:18,760
the things that are closer to the Token

1083
00:46:15,640 --> 00:46:21,480
level representation are more um

1084
00:46:18,760 --> 00:46:23,400
extractable using a probe and the things

1085
00:46:21,480 --> 00:46:26,440
that require more contextualized

1086
00:46:23,400 --> 00:46:29,440
information um is extractable later from

1087
00:46:26,440 --> 00:46:32,359
later layers in the model and so here's

1088
00:46:29,440 --> 00:46:34,599
sort of a brief uh description of what

1089
00:46:32,359 --> 00:46:37,599
these tasks are so the ones on the

1090
00:46:34,599 --> 00:46:40,040
bottom are more semantic more

1091
00:46:37,599 --> 00:46:42,040
contextualized like uh semantic prot

1092
00:46:40,040 --> 00:46:43,880
roles and relation relation

1093
00:46:42,040 --> 00:46:45,839
classification and then the first few

1094
00:46:43,880 --> 00:46:48,200
are more you know chunking in part of

1095
00:46:45,839 --> 00:46:51,880
speech tagging and um dependency

1096
00:46:48,200 --> 00:46:51,880
labeling in these in these sorts of

1097
00:46:52,040 --> 00:46:57,200
tests um so there's a bunch of issues

1098
00:46:54,480 --> 00:46:59,520
with probing um and there aren't many

1099
00:46:57,200 --> 00:47:03,559
probing papers now as there were many

1100
00:46:59,520 --> 00:47:05,960
years ago um and so if your probe let's

1101
00:47:03,559 --> 00:47:07,960
say your probe

1102
00:47:05,960 --> 00:47:09,920
works it's possible that the

1103
00:47:07,960 --> 00:47:12,200
representation actually encodes that

1104
00:47:09,920 --> 00:47:14,520
information it's also possible that it

1105
00:47:12,200 --> 00:47:16,359
doesn't and the probe solved the task by

1106
00:47:14,520 --> 00:47:18,119
itself uh keep in mind that you're

1107
00:47:16,359 --> 00:47:20,640
learning this probe you're training this

1108
00:47:18,119 --> 00:47:22,720
Probe on labeled data uh let's say your

1109
00:47:20,640 --> 00:47:24,599
probe doesn't work does that tell you

1110
00:47:22,720 --> 00:47:27,119
anything Maybe not maybe the

1111
00:47:24,599 --> 00:47:30,280
representation lacks the information or

1112
00:47:27,119 --> 00:47:31,800
maybe your probe doesn't doesn't

1113
00:47:30,280 --> 00:47:33,800
actually isn't actually able to

1114
00:47:31,800 --> 00:47:35,240
disentangle that information from your

1115
00:47:33,800 --> 00:47:36,720
representation maybe the probe is not

1116
00:47:35,240 --> 00:47:38,359
the right function class maybe you

1117
00:47:36,720 --> 00:47:40,839
poorly trained your probe there's

1118
00:47:38,359 --> 00:47:42,280
hyperparameters for your probe so often

1119
00:47:40,839 --> 00:47:43,000
times your probe doesn't give you that

1120
00:47:42,280 --> 00:47:46,119
much

1121
00:47:43,000 --> 00:47:49,040
information there's more problems too so

1122
00:47:46,119 --> 00:47:50,800
often we want to probe tasks themselves

1123
00:47:49,040 --> 00:47:53,240
and that requires a lot of supervised

1124
00:47:50,800 --> 00:47:55,880
data um but we can't collect a lot of

1125
00:47:53,240 --> 00:47:58,440
supervised data so we collect some of it

1126
00:47:55,880 --> 00:48:00,040
and then that instead produces this

1127
00:47:58,440 --> 00:48:02,480
convenient sample that we have that's a

1128
00:48:00,040 --> 00:48:04,119
data set that is a convenient sample of

1129
00:48:02,480 --> 00:48:07,000
your task so really what you're probing

1130
00:48:04,119 --> 00:48:10,040
is the data set and so with all these

1131
00:48:07,000 --> 00:48:11,800
limitations it's it's fallen out of

1132
00:48:10,040 --> 00:48:13,599
favor a little bit it's still very very

1133
00:48:11,800 --> 00:48:16,400
useful but it's fallen out of favor as

1134
00:48:13,599 --> 00:48:20,000
like a core model interpretability

1135
00:48:16,400 --> 00:48:22,160
idea um also probes designed in this way

1136
00:48:20,000 --> 00:48:26,079
are correlated they're correlative not

1137
00:48:22,160 --> 00:48:27,880
really causitive so your your underlying

1138
00:48:26,079 --> 00:48:29,640
model is trained in a specific way all

1139
00:48:27,880 --> 00:48:31,359
of that information is disentangled and

1140
00:48:29,640 --> 00:48:32,920
kind of thrown away and you're only

1141
00:48:31,359 --> 00:48:34,599
looking at the output representation and

1142
00:48:32,920 --> 00:48:36,559
you're saying is my output

1143
00:48:34,599 --> 00:48:39,200
representation correlated to the thing

1144
00:48:36,559 --> 00:48:42,400
that I'm training this probe for there's

1145
00:48:39,200 --> 00:48:44,960
no notion of intervening on this lat and

1146
00:48:42,400 --> 00:48:46,559
space there's no notion of of causation

1147
00:48:44,960 --> 00:48:49,119
really so you're just seeing whether

1148
00:48:46,559 --> 00:48:52,559
your representation is correlated with

1149
00:48:49,119 --> 00:48:54,480
your property that you're probing for um

1150
00:48:52,559 --> 00:48:56,200
and with these limitations the

1151
00:48:54,480 --> 00:48:58,720
community's moved a little bit away from

1152
00:48:56,200 --> 00:48:58,720
this area

1153
00:48:59,040 --> 00:49:02,200
uh there's a bunch of other probing

1154
00:49:00,240 --> 00:49:04,920
works so a bunch of people aim to solve

1155
00:49:02,200 --> 00:49:06,000
a bunch of these problems um and uh for

1156
00:49:04,920 --> 00:49:09,200
the sake of time I'm not going to go

1157
00:49:06,000 --> 00:49:12,599
into all of these but uh I'd encourage

1158
00:49:09,200 --> 00:49:14,000
you to look into these they for for some

1159
00:49:12,599 --> 00:49:17,319
of these problems they're able to

1160
00:49:14,000 --> 00:49:19,520
control for um control for like the

1161
00:49:17,319 --> 00:49:22,200
complexity of the of the probe and

1162
00:49:19,520 --> 00:49:24,359
things like this um but even despite

1163
00:49:22,200 --> 00:49:25,720
that probing is sort of slowly kind of

1164
00:49:24,359 --> 00:49:28,160
falling out of

1165
00:49:25,720 --> 00:49:29,640
favor uh so before I move into model

1166
00:49:28,160 --> 00:49:31,920
interpretability are there any questions

1167
00:49:29,640 --> 00:49:31,920
on

1168
00:49:35,520 --> 00:49:40,599
probing all right so what is model

1169
00:49:38,680 --> 00:49:44,000
interpretability so this is my

1170
00:49:40,599 --> 00:49:45,400
definition here uh this is the study of

1171
00:49:44,000 --> 00:49:46,599
understanding the internals of models

1172
00:49:45,400 --> 00:49:49,079
for example their weights and

1173
00:49:46,599 --> 00:49:51,160
activations putting those insights in

1174
00:49:49,079 --> 00:49:53,319
human intelligible terms and using that

1175
00:49:51,160 --> 00:49:55,920
insight to both patch current models and

1176
00:49:53,319 --> 00:49:57,359
develop better ones for not sort of able

1177
00:49:55,920 --> 00:49:58,760
to do both of these things patching

1178
00:49:57,359 --> 00:50:00,160
current models and develop better ones

1179
00:49:58,760 --> 00:50:02,440
we're kind of doing interpretability for

1180
00:50:00,160 --> 00:50:04,960
interpretability sake that's nice and

1181
00:50:02,440 --> 00:50:08,079
fun but it's not as applicable for the

1182
00:50:04,960 --> 00:50:09,720
for the community so you've probably

1183
00:50:08,079 --> 00:50:12,240
heard of the term mechanistic

1184
00:50:09,720 --> 00:50:14,480
interpretability it's in my opinion a

1185
00:50:12,240 --> 00:50:16,559
subfield of model interpretability and

1186
00:50:14,480 --> 00:50:19,319
this is sort of my definition I it

1187
00:50:16,559 --> 00:50:21,440
aligns reasonably well to the core

1188
00:50:19,319 --> 00:50:22,720
mechanistic interpretability people um

1189
00:50:21,440 --> 00:50:24,880
but it's the study of reverse

1190
00:50:22,720 --> 00:50:26,280
engineering parametric models often

1191
00:50:24,880 --> 00:50:28,839
neural networks because that's what we

1192
00:50:26,280 --> 00:50:31,400
use from their learned weights into more

1193
00:50:28,839 --> 00:50:32,839
human interpretable algorithmic units uh

1194
00:50:31,400 --> 00:50:36,839
and often they call these things

1195
00:50:32,839 --> 00:50:39,440
circuits um and and these are basically

1196
00:50:36,839 --> 00:50:42,880
functions that uh you can describe in a

1197
00:50:39,440 --> 00:50:45,000
human interpretable way that sit inside

1198
00:50:42,880 --> 00:50:46,760
models um there's a bunch of notable

1199
00:50:45,000 --> 00:50:50,720
work again for the sake of time I'm

1200
00:50:46,760 --> 00:50:54,319
going to just briefly talk about them

1201
00:50:50,720 --> 00:50:56,839
um so the first one is they they look

1202
00:50:54,319 --> 00:50:58,440
into analyzing small MLPs and

1203
00:50:56,839 --> 00:51:01,400
Transformers to build out the intuition

1204
00:50:58,440 --> 00:51:04,119
of what circuits exist um and this a lot

1205
00:51:01,400 --> 00:51:06,559
of this work came out of earlier work on

1206
00:51:04,119 --> 00:51:08,480
on lstms and doing similar sorts of

1207
00:51:06,559 --> 00:51:11,880
things with with

1208
00:51:08,480 --> 00:51:14,319
lstms um and they find a bunch of things

1209
00:51:11,880 --> 00:51:15,839
one thing that they find is this idea of

1210
00:51:14,319 --> 00:51:19,599
induction heads and these induction

1211
00:51:15,839 --> 00:51:21,760
heads they say is sort of sort of helps

1212
00:51:19,599 --> 00:51:24,680
prove why models can do in context

1213
00:51:21,760 --> 00:51:26,599
learning so so an induction head is

1214
00:51:24,680 --> 00:51:28,839
something that it's it's a specific

1215
00:51:26,599 --> 00:51:32,440
attention head that kind of allows you

1216
00:51:28,839 --> 00:51:35,599
to um when given a prefix allow you to

1217
00:51:32,440 --> 00:51:37,559
kind of copy the necessary resulting

1218
00:51:35,599 --> 00:51:39,640
token from the underlying training data

1219
00:51:37,559 --> 00:51:41,720
that the model seen before so in context

1220
00:51:39,640 --> 00:51:44,599
learning what you generally provide is

1221
00:51:41,720 --> 00:51:46,440
some sort of prefix and then you uh

1222
00:51:44,599 --> 00:51:48,480
provide some example and hopefully you

1223
00:51:46,440 --> 00:51:51,040
know you can classify the thing or

1224
00:51:48,480 --> 00:51:53,280
something like this um it's it's saying

1225
00:51:51,040 --> 00:51:56,200
that there's these attention heads

1226
00:51:53,280 --> 00:51:59,400
Loosely uh that exist that are able to

1227
00:51:56,200 --> 00:52:00,680
copy unearth that information um for for

1228
00:51:59,400 --> 00:52:03,319
a specific

1229
00:52:00,680 --> 00:52:07,200
context um other things that they' that

1230
00:52:03,319 --> 00:52:09,880
they've done is um on neurons so uh this

1231
00:52:07,200 --> 00:52:13,160
poly semanticity so what this this kind

1232
00:52:09,880 --> 00:52:15,240
of means is that your your neuron is a

1233
00:52:13,160 --> 00:52:18,000
uh you have a set of neurons in your

1234
00:52:15,240 --> 00:52:20,880
activation space so let's say at layer

1235
00:52:18,000 --> 00:52:23,200
10 in your model you have an output um

1236
00:52:20,880 --> 00:52:26,280
and so your activations is let's say a

1237
00:52:23,200 --> 00:52:28,400
thousand dimensional here those each of

1238
00:52:26,280 --> 00:52:31,319
those thousand individual neurons may

1239
00:52:28,400 --> 00:52:35,839
represent more than one specific

1240
00:52:31,319 --> 00:52:37,839
feature um and so they they talk about

1241
00:52:35,839 --> 00:52:41,280
this in that context and this is kind of

1242
00:52:37,839 --> 00:52:43,240
a theory but you can think about um

1243
00:52:41,280 --> 00:52:46,359
trying to process

1244
00:52:43,240 --> 00:52:49,400
input and when you're processing a a

1245
00:52:46,359 --> 00:52:50,960
vocab of size 50,000 or 250,000 at some

1246
00:52:49,400 --> 00:52:52,359
point in the model we're actually

1247
00:52:50,960 --> 00:52:55,400
compressing it down to the hidden

1248
00:52:52,359 --> 00:52:58,119
Dimension and so in some cases that

1249
00:52:55,400 --> 00:53:00,319
looks like you're going to compress a

1250
00:52:58,119 --> 00:53:03,440
much richer feature representation down

1251
00:53:00,319 --> 00:53:06,359
into a smaller set of neurons so it is

1252
00:53:03,440 --> 00:53:08,319
reasonable to believe that um a specific

1253
00:53:06,359 --> 00:53:10,799
neuron will represent multiple of those

1254
00:53:08,319 --> 00:53:15,480
features and given the structure of our

1255
00:53:10,799 --> 00:53:18,720
weight matrices um it it is the case

1256
00:53:15,480 --> 00:53:21,839
that if they are representing more

1257
00:53:18,720 --> 00:53:23,960
features than uh number of elements in

1258
00:53:21,839 --> 00:53:26,000
the actual or number of neurons in the

1259
00:53:23,960 --> 00:53:28,680
activation space then many of these

1260
00:53:26,000 --> 00:53:30,880
features linearly dependent and so we're

1261
00:53:28,680 --> 00:53:35,400
not really able to utilize them that

1262
00:53:30,880 --> 00:53:37,960
well um they they they talk about this

1263
00:53:35,400 --> 00:53:42,200
they don't talk about this in the the

1264
00:53:37,960 --> 00:53:44,799
most uh the the best way but uh it seems

1265
00:53:42,200 --> 00:53:48,040
kind of clear to me that um since you

1266
00:53:44,799 --> 00:53:50,880
have embedding matrices that are um not

1267
00:53:48,040 --> 00:53:53,599
square that you're that these neurons

1268
00:53:50,880 --> 00:53:56,400
have to exist um and they have to

1269
00:53:53,599 --> 00:53:59,200
incorporate multiple features at once m

1270
00:53:56,400 --> 00:54:02,559
multiple redundant features at

1271
00:53:59,200 --> 00:54:04,680
once um so before I move on to the rest

1272
00:54:02,559 --> 00:54:07,839
of model interpretability any questions

1273
00:54:04,680 --> 00:54:07,839
about mechanistic

1274
00:54:09,880 --> 00:54:12,880
interpretability

1275
00:54:21,480 --> 00:54:28,040
yeah so most of their studies are for uh

1276
00:54:24,920 --> 00:54:29,720
a very small set of of of models and

1277
00:54:28,040 --> 00:54:32,040
most of these are old GPT models there

1278
00:54:29,720 --> 00:54:34,160
have been a few works like in the last

1279
00:54:32,040 --> 00:54:36,760
couple of months on doing this for the

1280
00:54:34,160 --> 00:54:39,720
Llama based models um it seems like this

1281
00:54:36,760 --> 00:54:42,040
is a general more General phenomenal for

1282
00:54:39,720 --> 00:54:43,760
for language models it also is the case

1283
00:54:42,040 --> 00:54:46,839
that certain attention heads specialize

1284
00:54:43,760 --> 00:54:49,480
and talk about them a little bit um in

1285
00:54:46,839 --> 00:54:51,599
in the activations part um but yeah

1286
00:54:49,480 --> 00:54:53,799
there's they're not like all attention

1287
00:54:51,599 --> 00:54:56,400
heads aren't created equal they start

1288
00:54:53,799 --> 00:55:00,280
this way and it seems to be a general

1289
00:54:56,400 --> 00:55:01,799
princip of and one one other thing I you

1290
00:55:00,280 --> 00:55:04,520
might know about this better than I do

1291
00:55:01,799 --> 00:55:06,520
but I think there are some preliminary

1292
00:55:04,520 --> 00:55:09,160
words that say that Transformers seem to

1293
00:55:06,520 --> 00:55:11,720
be particularly good at doing things

1294
00:55:09,160 --> 00:55:15,160
like induction heads compared

1295
00:55:11,720 --> 00:55:17,200
to uh H for current models and there was

1296
00:55:15,160 --> 00:55:20,720
a paper really recently about comparing

1297
00:55:17,200 --> 00:55:23,599
like Mamba and um in Transformer based

1298
00:55:20,720 --> 00:55:26,400
models Mamba being a uh kind of more

1299
00:55:23,599 --> 00:55:30,280
like closer to a with network which we

1300
00:55:26,400 --> 00:55:33,119
also going to talk about us but um so I

1301
00:55:30,280 --> 00:55:37,319
I think there's some indication that

1302
00:55:33,119 --> 00:55:39,920
Transformers actually kind of are key or

1303
00:55:37,319 --> 00:55:43,680
are at least like better at kind of in

1304
00:55:39,920 --> 00:55:46,760
Contex learning than otheres are so

1305
00:55:43,680 --> 00:55:48,920
there is some

1306
00:55:46,760 --> 00:55:50,839
interesting implications of that which

1307
00:55:48,920 --> 00:55:53,240
is like well if Transformers are good

1308
00:55:50,839 --> 00:55:57,359
what's better than Transformer yeah like

1309
00:55:53,240 --> 00:55:58,799
naturally learning this s of thing so um

1310
00:55:57,359 --> 00:56:00,720
they're good at yeah they're like really

1311
00:55:58,799 --> 00:56:04,039
good at copying and like maintaining

1312
00:56:00,720 --> 00:56:06,799
information like more so um and yeah I

1313
00:56:04,039 --> 00:56:08,200
think it'd be cool to like be able to I

1314
00:56:06,799 --> 00:56:09,839
don't know how to do this but be able to

1315
00:56:08,200 --> 00:56:11,440
extract that kind of information like

1316
00:56:09,839 --> 00:56:13,359
what of the Transformers is actually

1317
00:56:11,440 --> 00:56:15,119
helping it do this copying mechanism or

1318
00:56:13,359 --> 00:56:17,799
like being a better in context learner

1319
00:56:15,119 --> 00:56:20,039
then we can develop a better structure a

1320
00:56:17,799 --> 00:56:23,119
slightly better structure than than a

1321
00:56:20,039 --> 00:56:26,000
Transformer um hopefully someone comes

1322
00:56:23,119 --> 00:56:28,240
up with that soon but cool any other

1323
00:56:26,000 --> 00:56:28,240
question

1324
00:56:29,799 --> 00:56:34,359
questions all right so let's move into

1325
00:56:32,240 --> 00:56:35,880
model interpretability so there are

1326
00:56:34,359 --> 00:56:37,480
weights and their activations I

1327
00:56:35,880 --> 00:56:39,160
mentioned these are these are the two

1328
00:56:37,480 --> 00:56:41,119
things these are the two things that

1329
00:56:39,160 --> 00:56:43,440
we're going to look at so what can you

1330
00:56:41,119 --> 00:56:45,480
do with the weights of an RD train model

1331
00:56:43,440 --> 00:56:47,799
really you can just edit them and then

1332
00:56:45,480 --> 00:56:49,200
kind of see what happens activations

1333
00:56:47,799 --> 00:56:51,240
similarly you can look at the

1334
00:56:49,200 --> 00:56:52,720
activations for different inputs you can

1335
00:56:51,240 --> 00:56:54,520
poke them with a stick and see what

1336
00:56:52,720 --> 00:56:56,359
happens a lot of my research is poking

1337
00:56:54,520 --> 00:56:58,559
models with a stick and looking at the

1338
00:56:56,359 --> 00:57:00,920
activations it's like predominantly what

1339
00:56:58,559 --> 00:57:02,240
I've done so we'll talk about that um

1340
00:57:00,920 --> 00:57:04,359
and the technical term for this is

1341
00:57:02,240 --> 00:57:06,599
intervening on them by adding some

1342
00:57:04,359 --> 00:57:07,839
vector or other sort of manipulation to

1343
00:57:06,599 --> 00:57:09,440
the lat space but really what you're

1344
00:57:07,839 --> 00:57:13,960
doing is like

1345
00:57:09,440 --> 00:57:17,599
Pok um so when you look at weights uh

1346
00:57:13,960 --> 00:57:19,920
one one class of methods or or area is

1347
00:57:17,599 --> 00:57:21,920
on model editing fine-tuning is like the

1348
00:57:19,920 --> 00:57:23,480
most extreme version of model editing

1349
00:57:21,920 --> 00:57:26,599
usually these things are much more

1350
00:57:23,480 --> 00:57:29,640
targeted um so in the model editing sort

1351
00:57:26,599 --> 00:57:32,160
of landscape your goal or your target is

1352
00:57:29,640 --> 00:57:35,119
you have a concept or a specific fact

1353
00:57:32,160 --> 00:57:37,440
that needs to be changed in the model um

1354
00:57:35,119 --> 00:57:39,640
and your approach here is you update or

1355
00:57:37,440 --> 00:57:41,359
edit the weights of the model to edit

1356
00:57:39,640 --> 00:57:43,640
the model's belief of that factor

1357
00:57:41,359 --> 00:57:45,599
concept and ideally you do this without

1358
00:57:43,640 --> 00:57:47,319
changing any of the other behavior of

1359
00:57:45,599 --> 00:57:49,760
the model so for example let's say

1360
00:57:47,319 --> 00:57:51,920
you're trying to say that Graham is no

1361
00:57:49,760 --> 00:57:54,559
longer a professor at CMU but is a

1362
00:57:51,920 --> 00:57:57,319
professor at Stanford you don't want

1363
00:57:54,559 --> 00:57:59,960
every single person at CMU to now be a

1364
00:57:57,319 --> 00:58:02,920
professor or uh now be affiliated with

1365
00:57:59,960 --> 00:58:07,839
Stanford right um gr pleas not to

1366
00:58:02,920 --> 00:58:09,039
Stanford um so here's one approach paper

1367
00:58:07,839 --> 00:58:11,720
that came out a couple years ago there's

1368
00:58:09,039 --> 00:58:13,559
a lot of work down here uh in in the

1369
00:58:11,720 --> 00:58:15,799
model editing World I'll give you sort

1370
00:58:13,559 --> 00:58:17,440
of a really brief overview of this but

1371
00:58:15,799 --> 00:58:20,520
basically they have facts that they want

1372
00:58:17,440 --> 00:58:22,400
to they want to manipulate um so for

1373
00:58:20,520 --> 00:58:24,680
example the the example that they give

1374
00:58:22,400 --> 00:58:26,640
in the figure is they want to associate

1375
00:58:24,680 --> 00:58:30,960
the Space Needle with Paris the Space

1376
00:58:26,640 --> 00:58:32,520
Needle is a a cool needle in in Seattle

1377
00:58:30,960 --> 00:58:36,000
has nothing to do with Paris but Paris

1378
00:58:32,520 --> 00:58:38,400
also has a tower so it's close um so

1379
00:58:36,000 --> 00:58:40,920
they use causal tracing to isolate the

1380
00:58:38,400 --> 00:58:43,839
causal effect uh of the individual

1381
00:58:40,920 --> 00:58:45,799
hidden States for this fact so they

1382
00:58:43,839 --> 00:58:47,839
basically continuously perturb the input

1383
00:58:45,799 --> 00:58:49,760
do a bunch of forward passes and

1384
00:58:47,839 --> 00:58:51,720
sequentially find the specific hidden

1385
00:58:49,760 --> 00:58:55,280
states that are associated kind of with

1386
00:58:51,720 --> 00:58:56,839
this fact um then they make an edit and

1387
00:58:55,280 --> 00:58:59,119
their edit

1388
00:58:56,839 --> 00:59:02,039
uh looks like this thing on the right um

1389
00:58:59,119 --> 00:59:05,280
so they treat this pair Space Needle and

1390
00:59:02,039 --> 00:59:07,240
Paris as this uh key value pair where

1391
00:59:05,280 --> 00:59:10,359
Space Needle is the key you pass this

1392
00:59:07,240 --> 00:59:12,480
into um into this weight Matrix this

1393
00:59:10,359 --> 00:59:14,640
original part of the model you want this

1394
00:59:12,480 --> 00:59:16,599
now instead of outputting Seattle to

1395
00:59:14,640 --> 00:59:19,119
Output Paris and they have some nice

1396
00:59:16,599 --> 00:59:21,599
math and a closed form solution to to

1397
00:59:19,119 --> 00:59:23,880
identify this this is super expensive

1398
00:59:21,599 --> 00:59:25,359
because they have to the causal tracing

1399
00:59:23,880 --> 00:59:27,680
part have to do a bunch of forward

1400
00:59:25,359 --> 00:59:30,680
passes um and they make this a little

1401
00:59:27,680 --> 00:59:33,480
bit better in future future work they

1402
00:59:30,680 --> 00:59:37,920
also do sort of a more

1403
00:59:33,480 --> 00:59:40,160
comprehensive um edit um so these are

1404
00:59:37,920 --> 00:59:44,599
kind like some of the things you can do

1405
00:59:40,160 --> 00:59:46,799
um I'm less excited about model editing

1406
00:59:44,599 --> 00:59:49,039
um there's there's some work on model

1407
00:59:46,799 --> 00:59:51,319
editing sort of it's it's hard to

1408
00:59:49,039 --> 00:59:53,160
control what other things break there's

1409
00:59:51,319 --> 00:59:56,240
a and there's some work with when you

1410
00:59:53,160 --> 01:00:00,000
edit a specific fact things start being

1411
00:59:56,240 --> 01:00:02,680
weird and being biased in other ways um

1412
01:00:00,000 --> 01:00:05,760
and so

1413
01:00:02,680 --> 01:00:09,119
yeah do all kind of seual information

1414
01:00:05,760 --> 01:00:11,880
like X is and Y would they Alles to the

1415
01:00:09,119 --> 01:00:14,319
same layer is it just with the

1416
01:00:11,880 --> 01:00:16,920
specific for this specific example it

1417
01:00:14,319 --> 01:00:19,039
looks at this specific point uh for

1418
01:00:16,920 --> 01:00:21,039
every example they'll probably find

1419
01:00:19,039 --> 01:00:22,119
different regions in a different degree

1420
01:00:21,039 --> 01:00:25,680
of

1421
01:00:22,119 --> 01:00:27,960
manipulation um and yeah that it gets a

1422
01:00:25,680 --> 01:00:30,920
little unprincipled kind of quickly it's

1423
01:00:27,960 --> 01:00:33,000
not like they're able to find you know a

1424
01:00:30,920 --> 01:00:35,680
specific attention head that or a

1425
01:00:33,000 --> 01:00:38,240
specific layer or specific weight Matrix

1426
01:00:35,680 --> 01:00:42,400
that corresponds to like

1427
01:00:38,240 --> 01:00:46,720
all yeah relations of a specific

1428
01:00:42,400 --> 01:00:49,160
type any for questions for yeah this is

1429
01:00:46,720 --> 01:00:51,119
actually just a question if you know um

1430
01:00:49,160 --> 01:00:53,200
it seems like more frequent facts might

1431
01:00:51,119 --> 01:00:55,240
appear in both places in the model is do

1432
01:00:53,200 --> 01:00:59,280
you know if that's actually the I have

1433
01:00:55,240 --> 01:01:02,440
no idea but uh I would imagine that um

1434
01:00:59,280 --> 01:01:06,240
it probably could occur in more places

1435
01:01:02,440 --> 01:01:08,160
but also um a lot of the information is

1436
01:01:06,240 --> 01:01:10,119
redundant anyway in the model especially

1437
01:01:08,160 --> 01:01:11,720
for larger models so you might have to

1438
01:01:10,119 --> 01:01:13,599
make targeted interventions in multiple

1439
01:01:11,720 --> 01:01:15,480
places but it's possible that one

1440
01:01:13,599 --> 01:01:17,680
intervention in one place sufficiently

1441
01:01:15,480 --> 01:01:21,039
destroys like contextualized information

1442
01:01:17,680 --> 01:01:22,680
in other places if it's close um it

1443
01:01:21,039 --> 01:01:24,839
depends on how big this intervention is

1444
01:01:22,680 --> 01:01:28,200
if it's like hitting it with a hammer

1445
01:01:24,839 --> 01:01:30,520
rather than some like nice fine grain

1446
01:01:28,200 --> 01:01:33,359
thing but that'd be a good be a good

1447
01:01:30,520 --> 01:01:36,839
experiment to see

1448
01:01:33,359 --> 01:01:36,839
um any other

1449
01:01:37,240 --> 01:01:41,559
questions all right so we'll move into

1450
01:01:39,760 --> 01:01:43,680
the stuff that I'm most familiar with

1451
01:01:41,559 --> 01:01:46,319
and some of my work so looking at

1452
01:01:43,680 --> 01:01:48,319
activations um so this is this is work

1453
01:01:46,319 --> 01:01:50,480
I've been doing for a while uh this idea

1454
01:01:48,319 --> 01:01:52,799
of steering vectors so I mentioned I

1455
01:01:50,480 --> 01:01:54,480
poke models so it's thick steering

1456
01:01:52,799 --> 01:01:57,000
Vector is that thick so it's basically a

1457
01:01:54,480 --> 01:01:59,000
fix length vector that steers a language

1458
01:01:57,000 --> 01:02:00,920
model to generate a specific sequence

1459
01:01:59,000 --> 01:02:02,720
exactly when added to the hidden sites

1460
01:02:00,920 --> 01:02:06,319
of a model at a specific

1461
01:02:02,720 --> 01:02:09,000
location um and I'll I'll read this

1462
01:02:06,319 --> 01:02:11,400
again it's there's a very like specific

1463
01:02:09,000 --> 01:02:13,319
form that I wrote this in so uh it's

1464
01:02:11,400 --> 01:02:15,359
it's a fix length Vector that steers a

1465
01:02:13,319 --> 01:02:17,640
language model to generate a specific

1466
01:02:15,359 --> 01:02:19,359
sequence exactly when added to the

1467
01:02:17,640 --> 01:02:22,559
hidden states of a model at a specific

1468
01:02:19,359 --> 01:02:24,480
point so this is different than um a

1469
01:02:22,559 --> 01:02:26,839
soft prompt or different than a model

1470
01:02:24,480 --> 01:02:29,520
editing sort of approach

1471
01:02:26,839 --> 01:02:31,400
um in this case there is a vector that

1472
01:02:29,520 --> 01:02:32,960
corresponds to a sequence and that

1473
01:02:31,400 --> 01:02:35,359
Vector doesn't correspond to any other

1474
01:02:32,960 --> 01:02:36,640
sequence there could be multiple vectors

1475
01:02:35,359 --> 01:02:39,079
and it turns out there are multiple

1476
01:02:36,640 --> 01:02:41,799
vectors that correspond to that sequence

1477
01:02:39,079 --> 01:02:44,160
it'll be a little bit clearer um based

1478
01:02:41,799 --> 01:02:46,279
on how we extract these

1479
01:02:44,160 --> 01:02:48,839
things um so this is the stick that

1480
01:02:46,279 --> 01:02:52,000
we're talking the language

1481
01:02:48,839 --> 01:02:53,599
model um so how do we extract them so

1482
01:02:52,000 --> 01:02:57,400
this is

1483
01:02:53,599 --> 01:03:00,200
gpt2 um basically this Z steer thing on

1484
01:02:57,400 --> 01:03:03,240
the left this is the steering Vector

1485
01:03:00,200 --> 01:03:05,799
this gets initialized randomly um with

1486
01:03:03,240 --> 01:03:09,520
small like in a in a reasonable way

1487
01:03:05,799 --> 01:03:11,440
uniformly and small um and for any

1488
01:03:09,520 --> 01:03:14,000
sequence a specific sequence that we

1489
01:03:11,440 --> 01:03:17,680
want the model to generate we

1490
01:03:14,000 --> 01:03:19,400
optimize this steering Vector Z steer uh

1491
01:03:17,680 --> 01:03:21,559
to generate that sequence keeping the

1492
01:03:19,400 --> 01:03:23,960
rest of the model entirely fixed so

1493
01:03:21,559 --> 01:03:26,200
think about it as we're nudging an a

1494
01:03:23,960 --> 01:03:29,880
frozen model to be able to generate a

1495
01:03:26,200 --> 01:03:31,680
specific sequence at a specific time um

1496
01:03:29,880 --> 01:03:33,880
and we have a lot of different options

1497
01:03:31,680 --> 01:03:35,559
on where to inject the steering intive

1498
01:03:33,880 --> 01:03:37,520
can put it basically anywhere in the

1499
01:03:35,559 --> 01:03:41,799
model we can put it at any time step any

1500
01:03:37,520 --> 01:03:43,839
number of these things in practice um

1501
01:03:41,799 --> 01:03:45,839
providing it just at the first time step

1502
01:03:43,839 --> 01:03:48,039
and somewhere in the middle of the model

1503
01:03:45,839 --> 01:03:52,480
basically not the first layer and not

1504
01:03:48,039 --> 01:03:56,240
the last layer works pretty well um and

1505
01:03:52,480 --> 01:04:00,279
so more formally um forget the kind of

1506
01:03:56,240 --> 01:04:03,640
notation um but right here we initialize

1507
01:04:00,279 --> 01:04:06,559
um this Z steer and for a few iterations

1508
01:04:03,640 --> 01:04:08,039
um we do forward passes first this

1509
01:04:06,559 --> 01:04:09,599
starts as random and then this gets

1510
01:04:08,039 --> 01:04:11,960
closer and closer and closer to being

1511
01:04:09,599 --> 01:04:14,279
able to generate this sequence and

1512
01:04:11,960 --> 01:04:16,599
eventually we get to a point uh and this

1513
01:04:14,279 --> 01:04:18,400
n is pretty small it's eight or 10 or

1514
01:04:16,599 --> 01:04:20,160
something like that um for most

1515
01:04:18,400 --> 01:04:22,200
sequences we get to a point where we

1516
01:04:20,160 --> 01:04:23,920
have found this stick that is allowed to

1517
01:04:22,200 --> 01:04:26,079
poke this model to generate that

1518
01:04:23,920 --> 01:04:29,319
sequence exactly now when we greedy

1519
01:04:26,079 --> 01:04:32,480
decode from the model we pass in just a

1520
01:04:29,319 --> 01:04:34,920
beginning of sequence token and this Z

1521
01:04:32,480 --> 01:04:37,119
steer the steering vector and it's able

1522
01:04:34,920 --> 01:04:39,720
to uncover a whole sequence that whole

1523
01:04:37,119 --> 01:04:41,319
sequence that we had at the beginning

1524
01:04:39,720 --> 01:04:44,240
entirely

1525
01:04:41,319 --> 01:04:46,640
um this is weird and interesting because

1526
01:04:44,240 --> 01:04:48,880
in a lot of cases um in like the

1527
01:04:46,640 --> 01:04:52,039
prompting world in the soft prompt World

1528
01:04:48,880 --> 01:04:54,640
usually you need a pretty large uh width

1529
01:04:52,039 --> 01:04:57,880
of a prompt to be able to do things with

1530
01:04:54,640 --> 01:05:00,400
um and this this generally in in that

1531
01:04:57,880 --> 01:05:02,000
structure you're doing a specific task

1532
01:05:00,400 --> 01:05:04,200
and you're providing kind of a large a

1533
01:05:02,000 --> 01:05:06,720
large prompt to do this with large soft

1534
01:05:04,200 --> 01:05:10,520
prompt to do this with this is often H

1535
01:05:06,720 --> 01:05:13,200
this often has a width of 50 and a and a

1536
01:05:10,520 --> 01:05:15,520
length of the the hidden size or the

1537
01:05:13,200 --> 01:05:17,160
embedding size of the model in our cases

1538
01:05:15,520 --> 01:05:20,079
all of our steering vectors are withth

1539
01:05:17,160 --> 01:05:21,440
one and they're of just the hidden size

1540
01:05:20,079 --> 01:05:24,039
of the

1541
01:05:21,440 --> 01:05:26,520
model um so what ends

1542
01:05:24,039 --> 01:05:29,559
up happening

1543
01:05:26,520 --> 01:05:31,520
um actually before I go to results any

1544
01:05:29,559 --> 01:05:34,720
questions this is a this is a weird

1545
01:05:31,520 --> 01:05:38,160
setup and weird relative to what other

1546
01:05:34,720 --> 01:05:39,310
people do so happy to take any

1547
01:05:38,160 --> 01:05:42,480
questions

1548
01:05:39,310 --> 01:05:42,480
[Music]

1549
01:05:42,880 --> 01:05:50,640
yeah similarly if your prompt was um of

1550
01:05:47,440 --> 01:05:53,440
a specific type so the prompt here is a

1551
01:05:50,640 --> 01:05:55,720
continuous Vector passed in it's a

1552
01:05:53,440 --> 01:05:59,760
single length width hidden size

1553
01:05:55,720 --> 01:06:02,799
continuous Vector so um it's kind of

1554
01:05:59,760 --> 01:06:05,559
like maybe collapsing your prompt into

1555
01:06:02,799 --> 01:06:08,480
this compressing it into this tiny

1556
01:06:05,559 --> 01:06:12,119
VOR you can think of that way

1557
01:06:08,480 --> 01:06:16,920
yeah any other questions

1558
01:06:12,119 --> 01:06:16,920
yeah this would be like

1559
01:06:18,160 --> 01:06:23,359
I'm

1560
01:06:20,880 --> 01:06:28,279
things potentially um this is something

1561
01:06:23,359 --> 01:06:30,640
that I want to work on it uh like a year

1562
01:06:28,279 --> 01:06:32,119
ago and didn't get didn't get this

1563
01:06:30,640 --> 01:06:34,559
sufficient Buy in and then had to apply

1564
01:06:32,119 --> 01:06:36,880
to grad school and all these things so

1565
01:06:34,559 --> 01:06:40,160
it went by the wayside but but

1566
01:06:36,880 --> 01:06:43,440
definitely something to something to

1567
01:06:40,160 --> 01:06:45,920
pursue um there's a lot of scope there

1568
01:06:43,440 --> 01:06:45,920
any other

1569
01:06:47,640 --> 01:06:54,480
questions all right so move over to

1570
01:06:51,319 --> 01:06:56,119
results so we can find steering vectors

1571
01:06:54,480 --> 01:06:58,520
and that's and that's interesting thing

1572
01:06:56,119 --> 01:07:00,559
um and we can find them pretty easily

1573
01:06:58,520 --> 01:07:02,559
and for most sequences even sequences

1574
01:07:00,559 --> 01:07:04,559
that the model hasn't seen before the

1575
01:07:02,559 --> 01:07:06,400
underlying language model hasn't seen

1576
01:07:04,559 --> 01:07:09,640
before

1577
01:07:06,400 --> 01:07:13,160
um it also works for and this is kind of

1578
01:07:09,640 --> 01:07:16,799
a negative but it also works for random

1579
01:07:13,160 --> 01:07:20,039
sequences of very small length but it's

1580
01:07:16,799 --> 01:07:22,359
harder to find so you can imagine if

1581
01:07:20,039 --> 01:07:24,760
your uh steering Vector is basically a

1582
01:07:22,359 --> 01:07:26,279
giant bulldozer it doesn't matter what

1583
01:07:24,760 --> 01:07:28,640
your model is learning learned similar

1584
01:07:26,279 --> 01:07:30,160
to the probe situation if you can

1585
01:07:28,640 --> 01:07:32,559
compress all that information of that

1586
01:07:30,160 --> 01:07:35,400
sequence into the vector you don't

1587
01:07:32,559 --> 01:07:37,400
really need the language model um so

1588
01:07:35,400 --> 01:07:39,559
there are cases when you're looking at

1589
01:07:37,400 --> 01:07:40,760
sequences of length like five seven

1590
01:07:39,559 --> 01:07:43,079
eight something like this you can

1591
01:07:40,760 --> 01:07:45,520
uniformly sample from the vocabulary at

1592
01:07:43,079 --> 01:07:47,359
random with replacement generate utter

1593
01:07:45,520 --> 01:07:49,799
garbage and find steering vectors for

1594
01:07:47,359 --> 01:07:53,200
them takes a little while but your model

1595
01:07:49,799 --> 01:07:55,520
is complex enough that you can basically

1596
01:07:53,200 --> 01:07:57,960
bulldo your model to be able to do this

1597
01:07:55,520 --> 01:08:00,200
even if that sequence is incredibly low

1598
01:07:57,960 --> 01:08:01,480
likelihood under the model but it works

1599
01:08:00,200 --> 01:08:05,319
better for things that are higher

1600
01:08:01,480 --> 01:08:07,760
likelihood under the model um

1601
01:08:05,319 --> 01:08:09,920
predictably the I think the thing that

1602
01:08:07,760 --> 01:08:12,760
surprised me the most was these steering

1603
01:08:09,920 --> 01:08:15,319
vectors themselves have interpretable

1604
01:08:12,760 --> 01:08:17,960
properties U so distances in steering

1605
01:08:15,319 --> 01:08:20,759
Vector space reflect semantic similarity

1606
01:08:17,960 --> 01:08:23,640
so if you have two sentences that are

1607
01:08:20,759 --> 01:08:26,719
close um they're also close in steering

1608
01:08:23,640 --> 01:08:29,759
Vector space that's kind of nice

1609
01:08:26,719 --> 01:08:32,359
um it does better than for example the

1610
01:08:29,759 --> 01:08:34,520
representations one would use for for

1611
01:08:32,359 --> 01:08:37,159
probing so mean pooling Bert hidden

1612
01:08:34,520 --> 01:08:39,600
States like we looked at before those do

1613
01:08:37,159 --> 01:08:42,080
actually worse than steering vectors um

1614
01:08:39,600 --> 01:08:45,799
just a bit

1615
01:08:42,080 --> 01:08:47,880
surprising um style transfer is possible

1616
01:08:45,799 --> 01:08:49,719
with simple Vector arithmetic so it' be

1617
01:08:47,880 --> 01:08:52,799
nice to say that I have a sequence I

1618
01:08:49,719 --> 01:08:56,000
want to subtract you know negativity and

1619
01:08:52,799 --> 01:08:58,799
add positivity for for sentiment or

1620
01:08:56,000 --> 01:09:00,520
other sorts of Styles um we can do this

1621
01:08:58,799 --> 01:09:02,159
and we can do this reasonably well in

1622
01:09:00,520 --> 01:09:05,319
steering VOR

1623
01:09:02,159 --> 01:09:07,920
space um we can also decode from

1624
01:09:05,319 --> 01:09:10,600
interpolations in the Laten space so you

1625
01:09:07,920 --> 01:09:12,759
take two steering vectors for two

1626
01:09:10,600 --> 01:09:14,759
sequences you look in the middle of them

1627
01:09:12,759 --> 01:09:17,400
you linearly interpolate between them

1628
01:09:14,759 --> 01:09:20,600
and you decode um if the space is kind

1629
01:09:17,400 --> 01:09:22,080
of weirdly peaky then you would have

1630
01:09:20,600 --> 01:09:23,839
issues and what you would generate is

1631
01:09:22,080 --> 01:09:25,080
garbage and there's no guarantee that

1632
01:09:23,839 --> 01:09:27,199
the space should be reasonable in

1633
01:09:25,080 --> 01:09:30,480
between but it turns out it

1634
01:09:27,199 --> 01:09:33,719
is um here's an example of one of these

1635
01:09:30,480 --> 01:09:36,359
style transfer cases so very very simple

1636
01:09:33,719 --> 01:09:39,239
easy easy sentence we found steering

1637
01:09:36,359 --> 01:09:41,679
vectors for The Taste is excellent and

1638
01:09:39,239 --> 01:09:43,640
and we took a sample of 100 positive

1639
01:09:41,679 --> 01:09:45,359
sentences and 100 negative sentences

1640
01:09:43,640 --> 01:09:47,159
found their steering vectors took the

1641
01:09:45,359 --> 01:09:48,960
mean and thought that you know that

1642
01:09:47,159 --> 01:09:51,400
looks like the positive concept steering

1643
01:09:48,960 --> 01:09:54,040
Vector negative concept steering Vector

1644
01:09:51,400 --> 01:09:56,600
we just did Vector arithmetic just did

1645
01:09:54,040 --> 01:09:59,880
uh current steering

1646
01:09:56,600 --> 01:10:02,440
Vector uh plus negative minus positive

1647
01:09:59,880 --> 01:10:03,520
and we got the taste is unpleasant um

1648
01:10:02,440 --> 01:10:06,960
and

1649
01:10:03,520 --> 01:10:08,880
similarly um in the reverse

1650
01:10:06,960 --> 01:10:12,520
directions it turns out that the

1651
01:10:08,880 --> 01:10:15,199
magnitude matters because um for every

1652
01:10:12,520 --> 01:10:17,800
single sequence there's kind of an end

1653
01:10:15,199 --> 01:10:20,640
dimensional ball around that steering

1654
01:10:17,800 --> 01:10:23,640
Vector that we found that also decodes

1655
01:10:20,640 --> 01:10:25,920
that specific sequence and so that shows

1656
01:10:23,640 --> 01:10:28,880
that the space is kind of reasonably

1657
01:10:25,920 --> 01:10:32,320
well formed there's there's of course uh

1658
01:10:28,880 --> 01:10:34,280
a lot of weird sort of areas um and so

1659
01:10:32,320 --> 01:10:37,120
if you go poke around in steering Vector

1660
01:10:34,280 --> 01:10:38,760
space and sort of try to sample from it

1661
01:10:37,120 --> 01:10:41,280
eventually you'll find some weird edge

1662
01:10:38,760 --> 01:10:43,320
cases and some garbage and repeated text

1663
01:10:41,280 --> 01:10:46,159
and little things like

1664
01:10:43,320 --> 01:10:50,520
this any questions here before I kind of

1665
01:10:46,159 --> 01:10:50,520
Rapid Fire through the the last few

1666
01:10:50,920 --> 01:10:57,239
things yeah like here

1667
01:10:57,400 --> 01:11:01,400
yeah so we went uh Beyond this um there

1668
01:11:00,199 --> 01:11:04,280
was

1669
01:11:01,400 --> 01:11:07,440
so in in these specific experiments we

1670
01:11:04,280 --> 01:11:09,600
looked at the middle of gpt2 um so this

1671
01:11:07,440 --> 01:11:12,679
was like layer six layer seven and at

1672
01:11:09,600 --> 01:11:15,280
the first time step we didn't do any um

1673
01:11:12,679 --> 01:11:17,239
like magnitude scaling and so you can

1674
01:11:15,280 --> 01:11:19,480
imagine if you put a giant Vector in

1675
01:11:17,239 --> 01:11:21,040
there the models never the rest of the

1676
01:11:19,480 --> 01:11:24,679
model has never seen something of that

1677
01:11:21,040 --> 01:11:26,159
magnitude so it's now in a weird State

1678
01:11:24,679 --> 01:11:28,280
and it's just going to break so if you

1679
01:11:26,159 --> 01:11:30,560
put this to like I don't know 500 or

1680
01:11:28,280 --> 01:11:32,960
something like this it break it just has

1681
01:11:30,560 --> 01:11:35,239
no idea it's like basically like telling

1682
01:11:32,960 --> 01:11:37,199
the rest your model hey it's like a

1683
01:11:35,239 --> 01:11:38,760
completely untrained model be it looks

1684
01:11:37,199 --> 01:11:42,000
similar to like random performance you

1685
01:11:38,760 --> 01:11:43,840
get repeats and things like this smaller

1686
01:11:42,000 --> 01:11:45,800
you end up staying in this ball for the

1687
01:11:43,840 --> 01:11:47,920
sequence two two seemed pretty

1688
01:11:45,800 --> 01:11:50,199
reasonable but we didn't spend a lot of

1689
01:11:47,920 --> 01:11:53,560
time just like the day before the paper

1690
01:11:50,199 --> 01:11:56,600
was do we were two seems reasonable we

1691
01:11:53,560 --> 01:11:59,159
went to three we went to five 10 broke

1692
01:11:56,600 --> 01:12:01,199
five somewhat broke two seems

1693
01:11:59,159 --> 01:12:03,440
reasonable

1694
01:12:01,199 --> 01:12:06,400
um decent signings

1695
01:12:03,440 --> 01:12:08,639
hopefully um cool so I'll talk about uh

1696
01:12:06,400 --> 01:12:10,920
a similar type of work uh that came out

1697
01:12:08,639 --> 01:12:13,000
more recently on inference time

1698
01:12:10,920 --> 01:12:14,159
intervention so basically they use some

1699
01:12:13,000 --> 01:12:16,719
of the ideas that we talked about

1700
01:12:14,159 --> 01:12:18,840
earlier they use linear probes um to

1701
01:12:16,719 --> 01:12:20,560
find a tension head that correspond to a

1702
01:12:18,840 --> 01:12:23,600
desired attribute they did this for

1703
01:12:20,560 --> 01:12:26,440
truthful QA so uh their Hope was to find

1704
01:12:23,600 --> 01:12:28,639
truthful directions in Len space

1705
01:12:26,440 --> 01:12:31,639
um and then they shifted the attention

1706
01:12:28,639 --> 01:12:33,199
head activations um during inference

1707
01:12:31,639 --> 01:12:35,280
along the directions determined by the

1708
01:12:33,199 --> 01:12:38,280
probes um so what this kind of looks

1709
01:12:35,280 --> 01:12:40,280
like is you take your attention heads

1710
01:12:38,280 --> 01:12:42,440
you probe them so you stick classify on

1711
01:12:40,280 --> 01:12:44,360
top um this classifier learns to

1712
01:12:42,440 --> 01:12:47,679
disentangle sort of truthful and

1713
01:12:44,360 --> 01:12:50,239
untruthful and now you have um now you

1714
01:12:47,679 --> 01:12:52,080
have a hyperplane and then you can move

1715
01:12:50,239 --> 01:12:54,320
orthogonally to this hyper plane in the

1716
01:12:52,080 --> 01:12:55,920
direction depending on which way you

1717
01:12:54,320 --> 01:12:58,080
want to shift so if you want to move

1718
01:12:55,920 --> 01:13:02,040
towards truthful you can move in that

1719
01:12:58,080 --> 01:13:04,400
direction or or away um and they do this

1720
01:13:02,040 --> 01:13:07,560
it works pretty well um I think they do

1721
01:13:04,400 --> 01:13:09,679
this for GPT model and maybe a llama

1722
01:13:07,560 --> 01:13:12,960
model um but can't can't remember the

1723
01:13:09,679 --> 01:13:15,960
exact details um and it's a similar

1724
01:13:12,960 --> 01:13:21,040
intervention um they basically add this

1725
01:13:15,960 --> 01:13:23,400
Vector um that they found and they they

1726
01:13:21,040 --> 01:13:25,679
have a little note on scaling they if

1727
01:13:23,400 --> 01:13:27,719
they scale if that if the magnitude of

1728
01:13:25,679 --> 01:13:30,000
the thing is too much things break so

1729
01:13:27,719 --> 01:13:33,880
they have a they like hyper parameter

1730
01:13:30,000 --> 01:13:36,800
search for the sort of magnitude of

1731
01:13:33,880 --> 01:13:38,840
activation um but it's sort of a very

1732
01:13:36,800 --> 01:13:41,520
similar approach to what we did but this

1733
01:13:38,840 --> 01:13:43,040
focuses on specific attention heads and

1734
01:13:41,520 --> 01:13:44,440
they don't do this for all the attention

1735
01:13:43,040 --> 01:13:46,600
heads so back to like your question

1736
01:13:44,440 --> 01:13:49,080
earlier do attention heads specialize it

1737
01:13:46,600 --> 01:13:52,360
seems like they do and so there are many

1738
01:13:49,080 --> 01:13:54,320
of them that uh have like no probing

1739
01:13:52,360 --> 01:13:57,719
accuracy or limited probing accuracy and

1740
01:13:54,320 --> 01:13:59,400
actually um are like distractors for the

1741
01:13:57,719 --> 01:14:03,400
CH FL

1742
01:13:59,400 --> 01:14:03,400
Direction any questions

1743
01:14:06,040 --> 01:14:11,760
here cool so more activation

1744
01:14:09,120 --> 01:14:14,760
manipulation so there's uh some work

1745
01:14:11,760 --> 01:14:17,600
recently on contrastive steering vectors

1746
01:14:14,760 --> 01:14:19,480
so the way we did this like sentiment

1747
01:14:17,600 --> 01:14:21,080
steering was we had some positive

1748
01:14:19,480 --> 01:14:23,040
sentences some negative sentences they

1749
01:14:21,080 --> 01:14:24,520
weren't tied together in any reasonable

1750
01:14:23,040 --> 01:14:26,360
way we found their steering vectors

1751
01:14:24,520 --> 01:14:30,040
separately you could imagine the case

1752
01:14:26,360 --> 01:14:33,159
and maybe a more useful case um with two

1753
01:14:30,040 --> 01:14:36,280
prompts that um you can design that go

1754
01:14:33,159 --> 01:14:38,000
two different ways you can sort of find

1755
01:14:36,280 --> 01:14:42,280
their representations and do the

1756
01:14:38,000 --> 01:14:45,679
manipulation the differences here um

1757
01:14:42,280 --> 01:14:48,800
like individually rather than um for a

1758
01:14:45,679 --> 01:14:52,400
whole concept or a whole attribute and

1759
01:14:48,800 --> 01:14:54,400
the value here is your context is um

1760
01:14:52,400 --> 01:14:56,600
preserved so if you're doing something

1761
01:14:54,400 --> 01:14:58,239
like you know you're doing retrieval

1762
01:14:56,600 --> 01:15:00,440
based things now you have some sort of

1763
01:14:58,239 --> 01:15:03,360
document and then you have a question if

1764
01:15:00,440 --> 01:15:05,040
your question sort of uh if you want to

1765
01:15:03,360 --> 01:15:07,560
ask it in two different ways for two

1766
01:15:05,040 --> 01:15:08,880
different things this would be a much

1767
01:15:07,560 --> 01:15:11,239
better approach if you want to use

1768
01:15:08,880 --> 01:15:14,600
steering vectors than the stuff I was

1769
01:15:11,239 --> 01:15:16,159
doing um and it seems to work a little

1770
01:15:14,600 --> 01:15:17,880
bit better they didn't compare against

1771
01:15:16,159 --> 01:15:19,400
our our things because it's not like an

1772
01:15:17,880 --> 01:15:21,880
Apples to Apples comparison but it seems

1773
01:15:19,400 --> 01:15:23,960
to work better and be more General um

1774
01:15:21,880 --> 01:15:25,560
and be more

1775
01:15:23,960 --> 01:15:27,840
useful

1776
01:15:25,560 --> 01:15:27,840
any

1777
01:15:31,400 --> 01:15:37,679
questions cool so what can model

1778
01:15:35,080 --> 01:15:40,080
interpretability give us these are these

1779
01:15:37,679 --> 01:15:41,960
are my concluding remarks so hopefully

1780
01:15:40,080 --> 01:15:43,920
we get a better understanding of how

1781
01:15:41,960 --> 01:15:46,840
language models work their their

1782
01:15:43,920 --> 01:15:49,520
internals their structure um we get to

1783
01:15:46,840 --> 01:15:52,800
understand uh kind of why they do really

1784
01:15:49,520 --> 01:15:55,239
well this is still like very very

1785
01:15:52,800 --> 01:15:57,320
unclear um and hopefully we find

1786
01:15:55,239 --> 01:15:59,400
lightweight methods to control and steer

1787
01:15:57,320 --> 01:16:03,360
models as models become more and more

1788
01:15:59,400 --> 01:16:05,280
useful um and and impact more more users

1789
01:16:03,360 --> 01:16:09,360
we need better ways to control and steer

1790
01:16:05,280 --> 01:16:13,120
them um and it's unclear how much

1791
01:16:09,360 --> 01:16:15,360
industry will devote to these things um

1792
01:16:13,120 --> 01:16:18,080
so it might be the role of Academia to

1793
01:16:15,360 --> 01:16:21,239
do more science in in order to figure

1794
01:16:18,080 --> 01:16:23,920
out how to control and steer these

1795
01:16:21,239 --> 01:16:25,520
better um and hopefully we can also find

1796
01:16:23,920 --> 01:16:29,199
potential Al Alternatives or

1797
01:16:25,520 --> 01:16:34,840
complimentary methods to to do alignment

1798
01:16:29,199 --> 01:16:37,480
um rhf is kind of expensive um and if if

1799
01:16:34,840 --> 01:16:40,080
we could do this with limited data and

1800
01:16:37,480 --> 01:16:42,760
um exploit structure um and information

1801
01:16:40,080 --> 01:16:46,400
that's already in the model more so than

1802
01:16:42,760 --> 01:16:48,600
than these methods um maybe maybe we can

1803
01:16:46,400 --> 01:16:50,920
align them better and these things don't

1804
01:16:48,600 --> 01:16:52,480
have to be uh Alternatives they can be

1805
01:16:50,920 --> 01:16:53,840
complimentary to to

1806
01:16:52,480 --> 01:16:57,159
[Music]

1807
01:16:53,840 --> 01:17:00,040
rhm um here's some resources this is an

1808
01:16:57,159 --> 01:17:01,280
extremely incomplete group but here are

1809
01:17:00,040 --> 01:17:04,080
some folks that work on model

1810
01:17:01,280 --> 01:17:07,040
interoperability there's many of these

1811
01:17:04,080 --> 01:17:09,120
um I cited some some work from some of

1812
01:17:07,040 --> 01:17:11,280
these teams but um there's a lot of

1813
01:17:09,120 --> 01:17:13,280
people working on it and in the last

1814
01:17:11,280 --> 01:17:15,040
like year there's been kind of an

1815
01:17:13,280 --> 01:17:17,480
explosion especially in the mechanistic

1816
01:17:15,040 --> 01:17:21,639
interpretability kind of World um Sasha

1817
01:17:17,480 --> 01:17:23,800
Rush had a recent tweet that uh asked

1818
01:17:21,639 --> 01:17:25,320
like prospective grad students what is

1819
01:17:23,800 --> 01:17:27,239
the topic that they're most excited

1820
01:17:25,320 --> 01:17:29,880
about and mechanistic interpretability

1821
01:17:27,239 --> 01:17:33,960
was a thing that seemed to have won out

1822
01:17:29,880 --> 01:17:37,040
um so I encourage you to to kind of dive

1823
01:17:33,960 --> 01:17:38,719
into this literature and read some of

1824
01:17:37,040 --> 01:17:41,679
the papers if you're if you're excited

1825
01:17:38,719 --> 01:17:45,199
about it and yeah thanks for your

1826
01:17:41,679 --> 01:17:45,199
attention and that's all I

1827
01:17:45,400 --> 01:17:48,400
have