CMU Advanced NLP 2024 (12) Reinforcement Learning/transcript.srt

1
00:00:00,840 --> 00:00:05,920
okay so uh let's get started um today

2
00:00:04,200 --> 00:00:08,000
I'm going to be talking about learning

3
00:00:05,920 --> 00:00:09,480
from Human feedback I wrote

4
00:00:08,000 --> 00:00:12,160
reinforcement learning from Human

5
00:00:09,480 --> 00:00:14,519
feedback because that's what um you know

6
00:00:12,160 --> 00:00:15,759
a lot of people talk about nowadays but

7
00:00:14,519 --> 00:00:18,880
actually there's other methods of

8
00:00:15,759 --> 00:00:21,840
learning from Human feedback so first

9
00:00:18,880 --> 00:00:24,760
I'm going to be talking about the ways

10
00:00:21,840 --> 00:00:27,920
we can get uh human feedback for the

11
00:00:24,760 --> 00:00:31,039
generations of models and mostly focus

12
00:00:27,920 --> 00:00:32,960
on generation tasks because is um

13
00:00:31,039 --> 00:00:35,800
generation tasks are harder than like

14
00:00:32,960 --> 00:00:38,559
classification tasks that we uh we deal

15
00:00:35,800 --> 00:00:40,000
with normally so I'll spend a fair

16
00:00:38,559 --> 00:00:42,239
amount of time talking about how we do

17
00:00:40,000 --> 00:00:45,760
that and then after I talk about how we

18
00:00:42,239 --> 00:00:48,360
do that we'll move into um how we

19
00:00:45,760 --> 00:00:51,160
actually learn from that

20
00:00:48,360 --> 00:00:53,399
signal so normally what we've done up

21
00:00:51,160 --> 00:00:56,399
until this point is maximum likelihood

22
00:00:53,399 --> 00:00:58,199
training uh this is just an overview

23
00:00:56,399 --> 00:00:59,559
slide so we what we want to do is we

24
00:00:58,199 --> 00:01:00,760
want to maximize the likelihood of

25
00:00:59,559 --> 00:01:03,280
predicting the next word and the

26
00:01:00,760 --> 00:01:05,960
reference given the previous words uh

27
00:01:03,280 --> 00:01:08,119
which gives us the loss of the output

28
00:01:05,960 --> 00:01:09,799
given the input uh where you know the

29
00:01:08,119 --> 00:01:13,960
input can be the prompt the output can

30
00:01:09,799 --> 00:01:16,080
be the answer to uh the output but

31
00:01:13,960 --> 00:01:18,360
there's uh lots of problems with

32
00:01:16,080 --> 00:01:20,439
learning from Maximum likelihood and I'm

33
00:01:18,360 --> 00:01:22,079
going to give three examples here I

34
00:01:20,439 --> 00:01:24,159
think all of these are actually real

35
00:01:22,079 --> 00:01:26,880
problems uh that we need to be worried

36
00:01:24,159 --> 00:01:30,240
about so the first one is that some

37
00:01:26,880 --> 00:01:32,439
mistakes are worse than others so um in

38
00:01:30,240 --> 00:01:33,560
the end we want good outputs and some

39
00:01:32,439 --> 00:01:36,520
mistaken

40
00:01:33,560 --> 00:01:38,200
predictions uh can be a bigger problem

41
00:01:36,520 --> 00:01:42,680
for the output being

42
00:01:38,200 --> 00:01:46,000
good so to give an example uh let's say

43
00:01:42,680 --> 00:01:47,600
what we actually wanted from like a

44
00:01:46,000 --> 00:01:49,320
speech recognition system or a

45
00:01:47,600 --> 00:01:54,040
translation system or something like

46
00:01:49,320 --> 00:01:54,040
that is uh please send this package to

47
00:01:54,280 --> 00:01:58,920
Pittsburgh if I write please send a

48
00:01:56,880 --> 00:02:01,560
package to Pittsburgh then this is not a

49
00:01:58,920 --> 00:02:03,560
huge problem

50
00:02:01,560 --> 00:02:06,479
if I write uh please send this package

51
00:02:03,560 --> 00:02:07,719
to Tokyo then that might be a big

52
00:02:06,479 --> 00:02:09,640
problem because the package you wanted

53
00:02:07,719 --> 00:02:12,760
to come to Pittsburgh goes to Tokyo

54
00:02:09,640 --> 00:02:13,680
instead and uh you might not want that

55
00:02:12,760 --> 00:02:16,080
to

56
00:02:13,680 --> 00:02:18,000
happen you might also have it say

57
00:02:16,080 --> 00:02:20,400
bleeping send this package to Pittsburgh

58
00:02:18,000 --> 00:02:22,200
instead of pleas um and that would be a

59
00:02:20,400 --> 00:02:24,200
problem in a customer service system

60
00:02:22,200 --> 00:02:28,400
right because your customer would uh

61
00:02:24,200 --> 00:02:28,400
leave and never come back

62
00:02:28,840 --> 00:02:32,040
so

63
00:02:30,360 --> 00:02:33,720
determiner like this is not going to

64
00:02:32,040 --> 00:02:35,640
cause a huge issue U messing up other

65
00:02:33,720 --> 00:02:37,519
things is going to cause a larger

66
00:02:35,640 --> 00:02:39,519
issue but from the point of view of

67
00:02:37,519 --> 00:02:42,680
Maximum likelihood all of these are just

68
00:02:39,519 --> 00:02:44,560
tokens and messing up one token is the

69
00:02:42,680 --> 00:02:47,519
same as messing up another token so

70
00:02:44,560 --> 00:02:50,040
that's uh you know an

71
00:02:47,519 --> 00:02:52,080
issue another problem is that the gold

72
00:02:50,040 --> 00:02:54,640
standard and maximum likelihood

73
00:02:52,080 --> 00:02:57,480
estimation can be bad it can be like not

74
00:02:54,640 --> 00:02:59,239
what you want and uh corpa are full of

75
00:02:57,480 --> 00:03:02,400
outputs that we wouldn't want a language

76
00:02:59,239 --> 00:03:05,400
model producing so for example uh toxic

77
00:03:02,400 --> 00:03:07,799
comments on Reddit uh

78
00:03:05,400 --> 00:03:09,959
disinformation um another thing that a

79
00:03:07,799 --> 00:03:13,000
lot of people don't think about uh quite

80
00:03:09,959 --> 00:03:15,640
as much is a lot of the data online is

81
00:03:13,000 --> 00:03:17,680
uh from is automatically generated

82
00:03:15,640 --> 00:03:19,720
nowadays for example from machine

83
00:03:17,680 --> 00:03:24,080
translation a lot of the translations

84
00:03:19,720 --> 00:03:25,720
online are from uh 2016 Google translate

85
00:03:24,080 --> 00:03:27,560
uh when Google translate was a lot less

86
00:03:25,720 --> 00:03:29,120
good than it is now and so you have like

87
00:03:27,560 --> 00:03:31,760
poor quality translations that were

88
00:03:29,120 --> 00:03:31,760
automatically

89
00:03:33,040 --> 00:03:37,959
a final problem is uh something that's

90
00:03:35,280 --> 00:03:40,360
called exposure bias and exposure bias

91
00:03:37,959 --> 00:03:44,000
basically what it means is mle training

92
00:03:40,360 --> 00:03:46,000
doesn't consider um the necessarity the

93
00:03:44,000 --> 00:03:48,599
necessity for generation and it relies

94
00:03:46,000 --> 00:03:51,360
on gold standard context so if we go

95
00:03:48,599 --> 00:03:54,159
back to the mle equation when we're

96
00:03:51,360 --> 00:03:57,200
calculating mle this y less than T is

97
00:03:54,159 --> 00:03:59,200
always correct it's always a good output

98
00:03:57,200 --> 00:04:01,439
and so what the model does is it learns

99
00:03:59,200 --> 00:04:04,280
to over rely on good

100
00:04:01,439 --> 00:04:06,079
outputs and one example of a problem

101
00:04:04,280 --> 00:04:08,360
that this causes is models tend to

102
00:04:06,079 --> 00:04:10,560
repeat themselves over and over again

103
00:04:08,360 --> 00:04:12,319
for example um when you use some

104
00:04:10,560 --> 00:04:15,079
generation algorithms and the reason why

105
00:04:12,319 --> 00:04:18,519
this happens is because in a gold

106
00:04:15,079 --> 00:04:22,079
standard output if a word has appeared

107
00:04:18,519 --> 00:04:25,840
previously that word is more likely to

108
00:04:22,079 --> 00:04:28,560
happen next so like if you say um like I

109
00:04:25,840 --> 00:04:29,759
am going um I am going to Pittsburgh

110
00:04:28,560 --> 00:04:31,880
you're much more likely to say

111
00:04:29,759 --> 00:04:33,000
Pittsburgh again in the future because

112
00:04:31,880 --> 00:04:35,720
you're talking about Pittsburgh

113
00:04:33,000 --> 00:04:37,400
topically as coherent so what you get is

114
00:04:35,720 --> 00:04:38,639
you get mle trained models saying I'm

115
00:04:37,400 --> 00:04:40,160
going to Pittsburgh I am going to

116
00:04:38,639 --> 00:04:41,680
Pittsburgh I am going to Pittsburgh I

117
00:04:40,160 --> 00:04:45,280
going to Pittsburgh you've probably seen

118
00:04:41,680 --> 00:04:47,320
this before uh at some point and so um

119
00:04:45,280 --> 00:04:49,320
exposure bias is basically that the

120
00:04:47,320 --> 00:04:51,039
model has never been exposed to mistakes

121
00:04:49,320 --> 00:04:55,240
in the past and so it can't deal with

122
00:04:51,039 --> 00:04:56,840
them so what this does is um if you have

123
00:04:55,240 --> 00:04:58,560
an alternative training algorithm you

124
00:04:56,840 --> 00:05:02,120
can fix this by generating a whole bunch

125
00:04:58,560 --> 00:05:04,880
of outputs uh down like scoring some of

126
00:05:02,120 --> 00:05:06,880
them poorly and penalizing the model for

127
00:05:04,880 --> 00:05:09,960
uh generating po outputs and so that can

128
00:05:06,880 --> 00:05:09,960
fix these problems as

129
00:05:10,800 --> 00:05:18,440
well uh any questions about this all

130
00:05:15,199 --> 00:05:20,800
good Okay cool so now I'd like to get

131
00:05:18,440 --> 00:05:23,919
into how we measure how good an output

132
00:05:20,800 --> 00:05:26,360
is and there's different ways of doing

133
00:05:23,919 --> 00:05:30,319
this um the first one is objective

134
00:05:26,360 --> 00:05:32,680
assessment so for some uh tasks or for

135
00:05:30,319 --> 00:05:35,400
many tasks there's kind of objectively a

136
00:05:32,680 --> 00:05:37,280
correct answer there's also human

137
00:05:35,400 --> 00:05:40,360
subjective annotations so you can ask

138
00:05:37,280 --> 00:05:42,919
humans to do annotation for you there's

139
00:05:40,360 --> 00:05:45,400
machine prediction of human

140
00:05:42,919 --> 00:05:48,319
preferences and there's also use in

141
00:05:45,400 --> 00:05:50,840
another system in a downstream

142
00:05:48,319 --> 00:05:52,960
task so the way objective assessment

143
00:05:50,840 --> 00:05:54,919
works is you have an annotated correct

144
00:05:52,960 --> 00:05:57,080
answer in match against this so like if

145
00:05:54,919 --> 00:06:00,600
you're solving math problems uh

146
00:05:57,080 --> 00:06:02,560
answering objective questions and and

147
00:06:00,600 --> 00:06:04,280
you know you can pick any arbitrary

148
00:06:02,560 --> 00:06:06,840
example you can pick your classification

149
00:06:04,280 --> 00:06:09,800
example from uh like your text

150
00:06:06,840 --> 00:06:11,880
classification tasks an even clearer

151
00:06:09,800 --> 00:06:13,880
example is if you have math problems

152
00:06:11,880 --> 00:06:15,639
there's kind of objectively one answer

153
00:06:13,880 --> 00:06:18,080
to any math problem and there's no other

154
00:06:15,639 --> 00:06:19,680
answer that could be correct so this

155
00:06:18,080 --> 00:06:21,160
makes your life easy if you're handling

156
00:06:19,680 --> 00:06:22,560
this type of problem but of course

157
00:06:21,160 --> 00:06:24,120
there's many other types of problems we

158
00:06:22,560 --> 00:06:26,039
want to handle that don't have objective

159
00:06:24,120 --> 00:06:29,039
answers like

160
00:06:26,039 --> 00:06:31,440
this so let's say we're handling a gener

161
00:06:29,039 --> 00:06:34,680
a generation task where we don't have an

162
00:06:31,440 --> 00:06:36,360
objective answer um in this Cas kind of

163
00:06:34,680 --> 00:06:39,440
one of our gold standards is human

164
00:06:36,360 --> 00:06:42,360
evaluation so we might have a source

165
00:06:39,440 --> 00:06:44,919
input like a prompt or an input text for

166
00:06:42,360 --> 00:06:47,240
machine translation we have one or

167
00:06:44,919 --> 00:06:49,960
several hypotheses and we ask a human

168
00:06:47,240 --> 00:06:53,280
annotator to basically give uh a score

169
00:06:49,960 --> 00:06:55,759
for them or do some sort of other

170
00:06:53,280 --> 00:06:59,759
annotation and the different varieties

171
00:06:55,759 --> 00:07:03,080
of annotation that we can give are um

172
00:06:59,759 --> 00:07:04,599
something called direct assessment so uh

173
00:07:03,080 --> 00:07:06,599
direct assessment is a term that comes

174
00:07:04,599 --> 00:07:09,280
from machine translation uh so you might

175
00:07:06,599 --> 00:07:11,039
not see it used uh lots of other places

176
00:07:09,280 --> 00:07:13,120
but it's basically just give a score

177
00:07:11,039 --> 00:07:15,759
directly to how good the output is so

178
00:07:13,120 --> 00:07:17,199
you can say like if you say please send

179
00:07:15,759 --> 00:07:18,960
this translation is please send this

180
00:07:17,199 --> 00:07:21,759
package to Tokyo we give it a score of

181
00:07:18,960 --> 00:07:24,360
two out of 10 or something like

182
00:07:21,759 --> 00:07:28,000
this

183
00:07:24,360 --> 00:07:30,840
so the the question here is like what

184
00:07:28,000 --> 00:07:32,400
does like let's say I gave a score of

185
00:07:30,840 --> 00:07:34,520
two out of 10 for please send this

186
00:07:32,400 --> 00:07:37,680
package to Tokyo what score should I

187
00:07:34,520 --> 00:07:40,240
give for please send a package to Tokyo

188
00:07:37,680 --> 00:07:42,360
anyone have any ideas the the correct

189
00:07:40,240 --> 00:07:46,520
answer is please send this package to

190
00:07:42,360 --> 00:07:48,000
take out of eight out of 10 yeah but you

191
00:07:46,520 --> 00:07:50,440
might disagree on that right it's kind

192
00:07:48,000 --> 00:07:52,159
of like subjective um one of the

193
00:07:50,440 --> 00:07:54,039
difficulties of direct assessment is

194
00:07:52,159 --> 00:07:55,520
giving a number like this is pretty

195
00:07:54,039 --> 00:07:57,800
difficult if you don't have a very clear

196
00:07:55,520 --> 00:07:59,720
rubric and very skilled annotators and

197
00:07:57,800 --> 00:08:02,879
it's hard to get consistency between

198
00:07:59,720 --> 00:08:04,400
people when you do this so the advantage

199
00:08:02,879 --> 00:08:05,599
is it kind of gives you an idea of how

200
00:08:04,400 --> 00:08:07,520
good things are overall but the

201
00:08:05,599 --> 00:08:09,280
disadvantage is it's more difficult to

202
00:08:07,520 --> 00:08:11,319
annotate and get

203
00:08:09,280 --> 00:08:13,159
consistency um another thing that I

204
00:08:11,319 --> 00:08:15,319
should point out is often scores are

205
00:08:13,159 --> 00:08:18,680
assigned separately based on desirable

206
00:08:15,319 --> 00:08:20,960
traits so um we don't necessarily just

207
00:08:18,680 --> 00:08:23,479
say how good is it we say how fluent is

208
00:08:20,960 --> 00:08:26,120
it like is it fluent uh

209
00:08:23,479 --> 00:08:28,159
English in Translation there's a concept

210
00:08:26,120 --> 00:08:30,720
called adequacy which is how well does

211
00:08:28,159 --> 00:08:34,599
the output reflect the input

212
00:08:30,720 --> 00:08:36,519
semantics um and if you're assessing

213
00:08:34,599 --> 00:08:38,440
translation systems actually it's common

214
00:08:36,519 --> 00:08:40,519
to assess fluency without even looking

215
00:08:38,440 --> 00:08:43,200
at the input because then you can just

216
00:08:40,519 --> 00:08:44,880
say how fluent is it but for adequacy

217
00:08:43,200 --> 00:08:46,320
you definitely need to understand the

218
00:08:44,880 --> 00:08:49,600
input so you need to be a bilingual

219
00:08:46,320 --> 00:08:54,680
speaker to be able to assess

220
00:08:49,600 --> 00:08:57,560
that um factuality um and so factuality

221
00:08:54,680 --> 00:09:00,160
is tricky um it can either be factuality

222
00:08:57,560 --> 00:09:03,880
grounded in a particular input text in

223
00:09:00,160 --> 00:09:05,600
which case um the facts would have to be

224
00:09:03,880 --> 00:09:07,680
you know things that were said in the

225
00:09:05,600 --> 00:09:09,399
input or it can be just kind of is the

226
00:09:07,680 --> 00:09:11,120
statement factual in general in which

227
00:09:09,399 --> 00:09:13,720
case you need to go online you need to

228
00:09:11,120 --> 00:09:16,480
search for things and like uh check

229
00:09:13,720 --> 00:09:18,480
whether the statement is factual or not

230
00:09:16,480 --> 00:09:20,480
um other things are like coherence does

231
00:09:18,480 --> 00:09:21,480
the output fit coherently within the

232
00:09:20,480 --> 00:09:23,680
larger

233
00:09:21,480 --> 00:09:25,680
discs um and there's many many other

234
00:09:23,680 --> 00:09:28,120
ones of these this is also task

235
00:09:25,680 --> 00:09:29,760
dependent so like the things you will

236
00:09:28,120 --> 00:09:31,000
evaluate for machine transl are

237
00:09:29,760 --> 00:09:32,880
different than the ones you would do for

238
00:09:31,000 --> 00:09:35,760
dialog which are different than the ones

239
00:09:32,880 --> 00:09:38,200
you would do for a general purpose

240
00:09:35,760 --> 00:09:41,279
chatot uh which is different kind things

241
00:09:38,200 --> 00:09:44,120
you would do for um summarization for

242
00:09:41,279 --> 00:09:46,320
example so if you're interested in doing

243
00:09:44,120 --> 00:09:47,519
something like this uh then I definitely

244
00:09:46,320 --> 00:09:48,800
encourage you to look at what other

245
00:09:47,519 --> 00:09:51,399
people have done for the tasks you're

246
00:09:48,800 --> 00:09:53,079
interested in uh previously and uh find

247
00:09:51,399 --> 00:09:54,880
out the different types of traits that

248
00:09:53,079 --> 00:09:58,320
did

249
00:09:54,880 --> 00:10:00,760
last uh any any questions about this

250
00:09:58,320 --> 00:10:03,079
also

251
00:10:00,760 --> 00:10:06,920
okay the next type of feedback is

252
00:10:03,079 --> 00:10:09,839
preference ratings um and so this is uh

253
00:10:06,920 --> 00:10:12,600
basically what you do is you have two or

254
00:10:09,839 --> 00:10:14,240
more outputs from different models or

255
00:10:12,600 --> 00:10:16,440
different Generations from an individual

256
00:10:14,240 --> 00:10:18,839
model and you ask a human which one is

257
00:10:16,440 --> 00:10:22,320
better like is one better than the other

258
00:10:18,839 --> 00:10:23,839
or are they tied and so in this case um

259
00:10:22,320 --> 00:10:26,320
you might have please send this package

260
00:10:23,839 --> 00:10:28,880
to Tokyo please send a package to

261
00:10:26,320 --> 00:10:31,040
Tokyo we might disagree on how like good

262
00:10:28,880 --> 00:10:33,959
or bad each of them are but I think most

263
00:10:31,040 --> 00:10:35,959
people would agree that this one is like

264
00:10:33,959 --> 00:10:37,480
despite the fact that it got this wrong

265
00:10:35,959 --> 00:10:40,160
the second one is better than the first

266
00:10:37,480 --> 00:10:42,240
one so this is a little bit of an easier

267
00:10:40,160 --> 00:10:45,040
task it's easier to uh get people to

268
00:10:42,240 --> 00:10:46,839
annotate these things

269
00:10:45,040 --> 00:10:50,519
consistently however it has the

270
00:10:46,839 --> 00:10:52,839
disadvantage that you can't really tell

271
00:10:50,519 --> 00:10:55,360
uh whether systems are really good or

272
00:10:52,839 --> 00:10:57,200
really bad so let's say you have a bunch

273
00:10:55,360 --> 00:11:00,279
of really bad systems that you're

274
00:10:57,200 --> 00:11:01,839
comparing with each other um you might

275
00:11:00,279 --> 00:11:03,680
find that one is better than the other

276
00:11:01,839 --> 00:11:06,000
but that still doesn't mean it's ready

277
00:11:03,680 --> 00:11:07,399
to be deployed or if you have a bunch of

278
00:11:06,000 --> 00:11:11,040
really good systems they're all

279
00:11:07,399 --> 00:11:13,000
basically you know very very similar to

280
00:11:11,040 --> 00:11:14,399
another but one is like slightly more

281
00:11:13,000 --> 00:11:18,639
fluent than the other you might still

282
00:11:14,399 --> 00:11:20,680
get a similar result um and so that also

283
00:11:18,639 --> 00:11:22,760
makes it uh you know a little bit

284
00:11:20,680 --> 00:11:24,880
difficult to use practically in some

285
00:11:22,760 --> 00:11:27,040
ways I didn't put it on the slide but

286
00:11:24,880 --> 00:11:30,680
there's another way you can kind of get

287
00:11:27,040 --> 00:11:33,920
the best of both worlds um which is a

288
00:11:30,680 --> 00:11:35,560
side by side assessment and side by-side

289
00:11:33,920 --> 00:11:38,440
assessment basically what you would do

290
00:11:35,560 --> 00:11:40,560
is you would say um please send this

291
00:11:38,440 --> 00:11:43,399
package to Tokyo please send a package

292
00:11:40,560 --> 00:11:47,279
to Pittsburgh give each of them a direct

293
00:11:43,399 --> 00:11:48,839
score um but you can use decimal places

294
00:11:47,279 --> 00:11:51,120
and you can't use the same score for all

295
00:11:48,839 --> 00:11:55,920
of them and so it's

296
00:11:51,120 --> 00:11:57,480
like five 500 and 4.99 out of five or

297
00:11:55,920 --> 00:11:59,519
something like that like you like one

298
00:11:57,480 --> 00:12:02,639
slightly better than the other or or

299
00:11:59,519 --> 00:12:04,480
something like that um so there are ways

300
00:12:02,639 --> 00:12:07,240
to kind of get Best of Both Worlds if

301
00:12:04,480 --> 00:12:11,720
you're interested in doing

302
00:12:07,240 --> 00:12:11,720
that um

303
00:12:14,920 --> 00:12:20,519
so one problem one other problem with

304
00:12:18,279 --> 00:12:22,519
preference rankings is that there's a

305
00:12:20,519 --> 00:12:24,440
limited number of things that humans can

306
00:12:22,519 --> 00:12:28,160
compare before they get really

307
00:12:24,440 --> 00:12:32,360
overwhelmed so if you say I

308
00:12:28,160 --> 00:12:35,560
want like I want to

309
00:12:32,360 --> 00:12:36,920
rate 15 systems or 20 systems with

310
00:12:35,560 --> 00:12:39,120
respect to how good they are with

311
00:12:36,920 --> 00:12:40,639
respect to each other it's going to be

312
00:12:39,120 --> 00:12:43,680
impossible for humans to come up with a

313
00:12:40,639 --> 00:12:46,959
good preference ranking between them and

314
00:12:43,680 --> 00:12:49,480
so the typical way around this um which

315
00:12:46,959 --> 00:12:52,360
is also used in uh things like the

316
00:12:49,480 --> 00:12:55,440
chatbot Arena by lmis and other things

317
00:12:52,360 --> 00:12:58,720
like this is to use uh something like an

318
00:12:55,440 --> 00:13:00,959
ELO or true skill ranking and what these

319
00:12:58,720 --> 00:13:03,079
are is these are things that were

320
00:13:00,959 --> 00:13:05,760
created for the ranking of like chess

321
00:13:03,079 --> 00:13:09,160
players or video game players or other

322
00:13:05,760 --> 00:13:11,720
things where they like b battle against

323
00:13:09,160 --> 00:13:13,920
each other in multiple matches uh

324
00:13:11,720 --> 00:13:16,440
pair-wise and then you put all of the

325
00:13:13,920 --> 00:13:18,399
wins and losses into these ranking

326
00:13:16,440 --> 00:13:20,600
algorithms and they give you a score

327
00:13:18,399 --> 00:13:22,920
about how good like each of the each of

328
00:13:20,600 --> 00:13:27,079
the players are so if you do something

329
00:13:22,920 --> 00:13:29,480
like this you can um get basically a

330
00:13:27,079 --> 00:13:32,120
ranking of systems despite the that you

331
00:13:29,480 --> 00:13:35,240
only did pairwise assessments so these

332
00:13:32,120 --> 00:13:35,240
are also a good thing to know

333
00:13:37,399 --> 00:13:43,839
about a final variety of human feedback

334
00:13:40,600 --> 00:13:45,320
uh that we create is uh air annotation

335
00:13:43,839 --> 00:13:47,519
and this can be useful for a number of

336
00:13:45,320 --> 00:13:49,839
reasons um but basically the way it

337
00:13:47,519 --> 00:13:53,839
works is you annotate individual errors

338
00:13:49,839 --> 00:13:55,639
within the outputs and um oh one thing I

339
00:13:53,839 --> 00:13:58,120
should mention is that um I'm giving a

340
00:13:55,639 --> 00:14:00,880
lot of examples from machine translation

341
00:13:58,120 --> 00:14:02,800
um I feel like machine translation has

342
00:14:00,880 --> 00:14:04,519
been doing evaluation of generated

343
00:14:02,800 --> 00:14:07,600
outputs for a lot longer than a lot of

344
00:14:04,519 --> 00:14:09,000
other uh fields of NLP have and

345
00:14:07,600 --> 00:14:11,800
therefore their methodology is more

346
00:14:09,000 --> 00:14:13,480
developed than a lot of other fields um

347
00:14:11,800 --> 00:14:16,199
but a lot of these things can also be

348
00:14:13,480 --> 00:14:18,079
applied to uh other uh other tasks as

349
00:14:16,199 --> 00:14:19,079
well but anyway getting back to this

350
00:14:18,079 --> 00:14:20,680
there's something for machine

351
00:14:19,079 --> 00:14:23,639
translation called multi-dimensional

352
00:14:20,680 --> 00:14:26,240
quality metrics and the multidimensional

353
00:14:23,639 --> 00:14:29,160
quality metrics basically what they do

354
00:14:26,240 --> 00:14:32,199
is they annotate spans in the output

355
00:14:29,160 --> 00:14:34,800
where each Span in the output is given a

356
00:14:32,199 --> 00:14:38,079
severity ranking of the error and it's

357
00:14:34,800 --> 00:14:40,199
given a type of the error and there's

358
00:14:38,079 --> 00:14:42,600
about eight different types of Errors

359
00:14:40,199 --> 00:14:44,839
like this doesn't violate or this

360
00:14:42,600 --> 00:14:47,399
violates linguistic conventions of using

361
00:14:44,839 --> 00:14:49,880
the word this instead of uh here by

362
00:14:47,399 --> 00:14:51,639
using the word uh instead of this here

363
00:14:49,880 --> 00:14:55,079
and then this is an accuracy error

364
00:14:51,639 --> 00:14:57,839
because it's not accurately con uh uh

365
00:14:55,079 --> 00:15:01,720
conveying the output and then this error

366
00:14:57,839 --> 00:15:04,600
is minor uh this error is Major um and

367
00:15:01,720 --> 00:15:06,399
then there's also like severe severe

368
00:15:04,600 --> 00:15:07,440
versus major but minor and major is a

369
00:15:06,399 --> 00:15:09,680
more important

370
00:15:07,440 --> 00:15:11,839
distinction um so the advantage of this

371
00:15:09,680 --> 00:15:14,279
is a couple fold number one it gives you

372
00:15:11,839 --> 00:15:16,440
more fine grained feedback uh in that

373
00:15:14,279 --> 00:15:19,199
you can say okay this system has a lot

374
00:15:16,440 --> 00:15:22,199
of uh accuracy errors this system has a

375
00:15:19,199 --> 00:15:24,880
lot of linguistic conventions errors um

376
00:15:22,199 --> 00:15:28,600
it also can be more consistent because

377
00:15:24,880 --> 00:15:29,839
if you just say to people which output

378
00:15:28,600 --> 00:15:31,800
is better

379
00:15:29,839 --> 00:15:34,560
or what is the score of this output

380
00:15:31,800 --> 00:15:36,360
people have trouble deciding about that

381
00:15:34,560 --> 00:15:39,560
because it's a more subjective

382
00:15:36,360 --> 00:15:41,680
evaluation but if I say is this word

383
00:15:39,560 --> 00:15:43,000
correct it's a little bit easier for

384
00:15:41,680 --> 00:15:44,759
people to do so you can get more

385
00:15:43,000 --> 00:15:46,920
consistent annotations

386
00:15:44,759 --> 00:15:49,720
here the problem with this is this can

387
00:15:46,920 --> 00:15:50,839
be very time consuming so um you know

388
00:15:49,720 --> 00:15:52,480
obviously you need to go through and

389
00:15:50,839 --> 00:15:56,440
annotate every single error if it's for

390
00:15:52,480 --> 00:15:56,440
a long outputs or something your

391
00:15:56,959 --> 00:16:03,519
problem so anyway these are just three

392
00:15:59,800 --> 00:16:05,680
uh ways of collecting human feedback um

393
00:16:03,519 --> 00:16:08,639
and then there's an alternative which is

394
00:16:05,680 --> 00:16:10,079
automatic evaluation of outputs and um

395
00:16:08,639 --> 00:16:14,399
there's a bunch of different ways we can

396
00:16:10,079 --> 00:16:16,800
do this the basic idea here is we have a

397
00:16:14,399 --> 00:16:20,199
source um we have a couple

398
00:16:16,800 --> 00:16:22,800
hypotheses and uh we have an automatic

399
00:16:20,199 --> 00:16:26,000
system that generates outputs uh like

400
00:16:22,800 --> 00:16:28,279
scores and we optionally have a

401
00:16:26,000 --> 00:16:30,839
reference output so the reference output

402
00:16:28,279 --> 00:16:33,519
is a human created gold standard output

403
00:16:30,839 --> 00:16:35,120
with respect to how good that um uh with

404
00:16:33,519 --> 00:16:38,240
respect to like what the output should

405
00:16:35,120 --> 00:16:38,240
be in an ideal

406
00:16:38,279 --> 00:16:47,079
case and basically the goal of automatic

407
00:16:43,199 --> 00:16:50,199
evaluation is to

408
00:16:47,079 --> 00:16:52,839
predict human preferences or to predict

409
00:16:50,199 --> 00:16:56,240
what the human scores would be um

410
00:16:52,839 --> 00:16:58,600
because still at this point um we mostly

411
00:16:56,240 --> 00:16:59,480
view what humans think of the output to

412
00:16:58,600 --> 00:17:01,680
be

413
00:16:59,480 --> 00:17:03,280
uh kind of the

414
00:17:01,680 --> 00:17:06,199
standard

415
00:17:03,280 --> 00:17:08,439
and this is called a variety of things

416
00:17:06,199 --> 00:17:10,600
depending on what field you're in um in

417
00:17:08,439 --> 00:17:12,559
machine translation and summarization

418
00:17:10,600 --> 00:17:13,520
it's called automatic evaluation also a

419
00:17:12,559 --> 00:17:16,520
lot in

420
00:17:13,520 --> 00:17:18,400
dialogue um if you're talking about

421
00:17:16,520 --> 00:17:21,000
people from reinforcement learning or

422
00:17:18,400 --> 00:17:24,600
other things um or chat Bots or things

423
00:17:21,000 --> 00:17:28,240
like that uh a lot of people or uh like

424
00:17:24,600 --> 00:17:31,280
AGI or whatever um a lot of people call

425
00:17:28,240 --> 00:17:32,520
it uh word model um because that

426
00:17:31,280 --> 00:17:34,480
specifically comes from the point of

427
00:17:32,520 --> 00:17:36,440
view of like learning from this feedback

428
00:17:34,480 --> 00:17:37,960
but essentially they're the same thing

429
00:17:36,440 --> 00:17:41,080
uh from my point of view they're trying

430
00:17:37,960 --> 00:17:42,520
to predict how good an output is and how

431
00:17:41,080 --> 00:17:44,240
much you should reward the model for

432
00:17:42,520 --> 00:17:46,559
producing that

433
00:17:44,240 --> 00:17:48,679
output

434
00:17:46,559 --> 00:17:50,520
um so there's a bunch of different

435
00:17:48,679 --> 00:17:51,720
methods to do this I'm not going to

436
00:17:50,520 --> 00:17:53,799
cover all of them I'm just going to

437
00:17:51,720 --> 00:17:55,240
cover three paradigms for doing this so

438
00:17:53,799 --> 00:17:57,880
you know where to look further if you're

439
00:17:55,240 --> 00:18:00,039
interested in doing these things um the

440
00:17:57,880 --> 00:18:02,400
first one is embedding based

441
00:18:00,039 --> 00:18:04,679
evaluation and the way embedding based

442
00:18:02,400 --> 00:18:06,600
evaluation works is usually it's

443
00:18:04,679 --> 00:18:11,400
unsupervised calculation based on

444
00:18:06,600 --> 00:18:14,880
embeding similarity between um

445
00:18:11,400 --> 00:18:18,080
the output that the model generated and

446
00:18:14,880 --> 00:18:20,840
a reference output that uh you have

447
00:18:18,080 --> 00:18:23,400
created so sorry this is very small but

448
00:18:20,840 --> 00:18:25,559
we have a reference here that says the

449
00:18:23,400 --> 00:18:27,640
weather is cold today and we have a

450
00:18:25,559 --> 00:18:30,240
candidate that says it is freezing today

451
00:18:27,640 --> 00:18:33,000
so this is probably you know like a good

452
00:18:30,240 --> 00:18:35,480
um a reasonably good

453
00:18:33,000 --> 00:18:37,640
output and we run this through some

454
00:18:35,480 --> 00:18:39,120
embedding model uh it was called Bert

455
00:18:37,640 --> 00:18:40,679
score and so of course you can run it

456
00:18:39,120 --> 00:18:42,240
through Bert but basically it can be any

457
00:18:40,679 --> 00:18:43,799
embedding model that gives you embedding

458
00:18:42,240 --> 00:18:46,200
for each token in the

459
00:18:43,799 --> 00:18:47,640
sequence and so there are five tokens in

460
00:18:46,200 --> 00:18:49,720
this sequence four tokens in this

461
00:18:47,640 --> 00:18:51,960
sequence you get five tokens and then

462
00:18:49,720 --> 00:18:54,799
four sorry five embeddings and then four

463
00:18:51,960 --> 00:18:57,400
embeddings you calculate carewise cosine

464
00:18:54,799 --> 00:18:59,880
similarity between all of them and this

465
00:18:57,400 --> 00:19:03,480
gives you cosine

466
00:18:59,880 --> 00:19:06,480
similarity Matrix and then you take the

467
00:19:03,480 --> 00:19:09,120
ARG Max or you take the maximum

468
00:19:06,480 --> 00:19:11,280
similarity along either the

469
00:19:09,120 --> 00:19:15,799
rows or the

470
00:19:11,280 --> 00:19:19,559
columns and here the rows correspond

471
00:19:15,799 --> 00:19:22,400
to tokens in the reference and because

472
00:19:19,559 --> 00:19:24,039
the rows correspond to tokens in the

473
00:19:22,400 --> 00:19:26,960
reference

474
00:19:24,039 --> 00:19:28,320
the how well you find something that is

475
00:19:26,960 --> 00:19:31,679
similar to each of the tokens in the

476
00:19:28,320 --> 00:19:34,000
reference is like a recall based method

477
00:19:31,679 --> 00:19:35,919
because it's saying how many tokens in

478
00:19:34,000 --> 00:19:39,520
the reference have a good match in the

479
00:19:35,919 --> 00:19:41,120
output and then if you look at the

480
00:19:39,520 --> 00:19:42,799
columns if you look at the max and the

481
00:19:41,120 --> 00:19:44,960
columns this is like a precision based

482
00:19:42,799 --> 00:19:47,000
metric because it's saying how many of

483
00:19:44,960 --> 00:19:49,360
the things in the output are similar

484
00:19:47,000 --> 00:19:51,240
have a similar match in the reference so

485
00:19:49,360 --> 00:19:54,480
basically you can calculate recall and

486
00:19:51,240 --> 00:19:56,200
precision over all of the tokens and

487
00:19:54,480 --> 00:20:00,200
then feed this into something that looks

488
00:19:56,200 --> 00:20:02,400
like fmeasure and you can also use tfidf

489
00:20:00,200 --> 00:20:06,000
waiting um like what I talked about in

490
00:20:02,400 --> 00:20:07,799
the rag lecture uh to upweight low

491
00:20:06,000 --> 00:20:09,520
frequency words because low frequency

492
00:20:07,799 --> 00:20:11,440
words tend to be more content words and

493
00:20:09,520 --> 00:20:13,120
going back to my example you know if you

494
00:20:11,440 --> 00:20:14,280
make a mistake from Pittsburgh to Tokyo

495
00:20:13,120 --> 00:20:17,880
that's going to be more painful than

496
00:20:14,280 --> 00:20:21,000
making a mistake from this to um so

497
00:20:17,880 --> 00:20:22,520
actually if you'll uh if you were paying

498
00:20:21,000 --> 00:20:25,480
close attention to the rag lecture this

499
00:20:22,520 --> 00:20:27,360
looks really similar to the co bear um

500
00:20:25,480 --> 00:20:29,559
the co bear retrieval objective that I

501
00:20:27,360 --> 00:20:30,960
talked about in the r lecture um I don't

502
00:20:29,559 --> 00:20:32,840
think it's a coincidence they both came

503
00:20:30,960 --> 00:20:34,360
out around the same time uh so people

504
00:20:32,840 --> 00:20:36,360
were thinking about the same thing but

505
00:20:34,360 --> 00:20:37,600
um this is one method that's pretty

506
00:20:36,360 --> 00:20:40,200
widely

507
00:20:37,600 --> 00:20:43,480
use the bird Square code base is also

508
00:20:40,200 --> 00:20:45,440
really nice and easy to use so um if uh

509
00:20:43,480 --> 00:20:47,640
you want to try it out feel free to take

510
00:20:45,440 --> 00:20:47,640
a

511
00:20:48,159 --> 00:20:53,840
look cool um the next one I'd like to

512
00:20:51,600 --> 00:20:56,080
talk about is a regression based

513
00:20:53,840 --> 00:20:58,760
evaluation and the way this works is

514
00:20:56,080 --> 00:21:02,600
this is usually used in a supervised uh

515
00:20:58,760 --> 00:21:04,320
setting so uh the way what you have to

516
00:21:02,600 --> 00:21:07,600
do is you have to calculate a whole

517
00:21:04,320 --> 00:21:09,799
bunch of like actual human

518
00:21:07,600 --> 00:21:12,440
judgments and

519
00:21:09,799 --> 00:21:15,000
usually these judgments can either be

520
00:21:12,440 --> 00:21:16,960
direct assessment uh where you actually

521
00:21:15,000 --> 00:21:19,120
have a score or they can be pairwise

522
00:21:16,960 --> 00:21:20,840
judgments and then if you have direct

523
00:21:19,120 --> 00:21:23,640
assessment you use a regression based

524
00:21:20,840 --> 00:21:26,039
loss like uh minimum squared error if

525
00:21:23,640 --> 00:21:27,520
you have pairwise uh you use a ranking

526
00:21:26,039 --> 00:21:29,039
based loss that tries to upweight the

527
00:21:27,520 --> 00:21:31,360
ones that are higher scoring downward

528
00:21:29,039 --> 00:21:33,200
the ones that are lower scoring one

529
00:21:31,360 --> 00:21:35,720
typical example of this is Comet which

530
00:21:33,200 --> 00:21:37,200
is or has been at least for a very long

531
00:21:35,720 --> 00:21:39,880
time the state-of-the art and machine

532
00:21:37,200 --> 00:21:41,279
translation evaluation and the reason

533
00:21:39,880 --> 00:21:43,440
why it works so well is because we have

534
00:21:41,279 --> 00:21:44,720
a bunch of evaluations for machine

535
00:21:43,440 --> 00:21:46,080
translation they've been doing

536
00:21:44,720 --> 00:21:47,600
evaluation and machine translation

537
00:21:46,080 --> 00:21:50,480
systems for years and you can use that

538
00:21:47,600 --> 00:21:52,720
as lots of supervised training data so

539
00:21:50,480 --> 00:21:54,640
basically you just take um these

540
00:21:52,720 --> 00:21:56,440
evaluation data you have human

541
00:21:54,640 --> 00:21:59,080
annotations you have the output

542
00:21:56,440 --> 00:22:00,320
according to a model like comet um you

543
00:21:59,080 --> 00:22:02,679
calculate the difference between them

544
00:22:00,320 --> 00:22:05,640
and you update model

545
00:22:02,679 --> 00:22:07,080
parameters um the problem this is great

546
00:22:05,640 --> 00:22:08,520
if you have lots of training data the

547
00:22:07,080 --> 00:22:10,640
problem with this is for a lot of tasks

548
00:22:08,520 --> 00:22:12,360
we don't have lots of training data so

549
00:22:10,640 --> 00:22:14,720
um you know training these is a little

550
00:22:12,360 --> 00:22:14,720
bit less

551
00:22:15,400 --> 00:22:22,919
feasible and now recently uh what we

552
00:22:19,600 --> 00:22:25,279
have been moving into is is a QA based

553
00:22:22,919 --> 00:22:27,120
evaluation which is basically where we

554
00:22:25,279 --> 00:22:30,760
ask a language model how good the output

555
00:22:27,120 --> 00:22:32,279
is and so uh gmba is an example one of

556
00:22:30,760 --> 00:22:34,559
the early examples of this for machine

557
00:22:32,279 --> 00:22:37,320
translation evaluation uh where they

558
00:22:34,559 --> 00:22:39,840
basically just ask a g gp4 like score

559
00:22:37,320 --> 00:22:41,600
the following translation from Source

560
00:22:39,840 --> 00:22:44,000
language to target language with respect

561
00:22:41,600 --> 00:22:47,080
to the human reference um on a

562
00:22:44,000 --> 00:22:49,200
continuous scale from Z to 100 uh where

563
00:22:47,080 --> 00:22:51,320
the score of zero means no meaning

564
00:22:49,200 --> 00:22:54,039
preserved and the score of 100 means a

565
00:22:51,320 --> 00:22:56,880
perfect meaning in grammar uh you feed

566
00:22:54,039 --> 00:22:58,760
in the source um you feed in the T the

567
00:22:56,880 --> 00:23:01,000
human reference optionally if you have a

568
00:22:58,760 --> 00:23:03,320
human reference and then you feed in the

569
00:23:01,000 --> 00:23:06,760
Target um and you get a

570
00:23:03,320 --> 00:23:09,919
score and um so this this works pretty

571
00:23:06,760 --> 00:23:12,720
well this can give you uh better results

572
00:23:09,919 --> 00:23:15,159
um there's a especially if you have a

573
00:23:12,720 --> 00:23:16,960
strong language model the problem is

574
00:23:15,159 --> 00:23:18,279
it's very unpredictable whether this is

575
00:23:16,960 --> 00:23:20,120
going to work well and it's very

576
00:23:18,279 --> 00:23:23,039
dependent on the prompt that you're

577
00:23:20,120 --> 00:23:25,279
using so um right now A lot of people

578
00:23:23,039 --> 00:23:27,279
are using gp4 without actually

579
00:23:25,279 --> 00:23:29,039
validating whether it does a good job at

580
00:23:27,279 --> 00:23:33,080
evaluation and

581
00:23:29,039 --> 00:23:34,919
and my the results are all across the

582
00:23:33,080 --> 00:23:36,880
board it can be anywhere from very very

583
00:23:34,919 --> 00:23:38,640
good to very very bad at evaluating

584
00:23:36,880 --> 00:23:41,320
particular tasks so I would be at least

585
00:23:38,640 --> 00:23:43,559
a little bit suspicious of whether gp4

586
00:23:41,320 --> 00:23:45,679
is doing a good job evaluating for your

587
00:23:43,559 --> 00:23:49,320
task especially more complex

588
00:23:45,679 --> 00:23:51,960
tests um I would especially be

589
00:23:49,320 --> 00:23:54,000
suspicious if you're doing two uh any of

590
00:23:51,960 --> 00:23:56,760
the two following things number one if

591
00:23:54,000 --> 00:23:59,880
you're comparing gp4 or any model

592
00:23:56,760 --> 00:24:02,400
against itself in another model because

593
00:23:59,880 --> 00:24:05,200
gp4 really likes

594
00:24:02,400 --> 00:24:06,880
gp4 it really likes its own outputs and

595
00:24:05,200 --> 00:24:08,120
there are papers uh sorry I don't

596
00:24:06,880 --> 00:24:09,679
actually have the references here but I

597
00:24:08,120 --> 00:24:11,200
can follow up if people are interested

598
00:24:09,679 --> 00:24:13,080
but there are papers that demonstrate

599
00:24:11,200 --> 00:24:15,799
that gp4 likes it you know its own

600
00:24:13,080 --> 00:24:19,200
outputs more than others also if you're

601
00:24:15,799 --> 00:24:22,120
explicitly optimizing the outputs using

602
00:24:19,200 --> 00:24:24,640
rlf um there is something called good

603
00:24:22,120 --> 00:24:27,120
Hearts law which is basically anytime

604
00:24:24,640 --> 00:24:29,520
you uh start optimizing towards a metric

605
00:24:27,120 --> 00:24:32,559
it becomes a bad metric and that also

606
00:24:29,520 --> 00:24:35,000
happens for gp4 based evaluations so if

607
00:24:32,559 --> 00:24:37,200
you start optimizing for gp4 based

608
00:24:35,000 --> 00:24:38,960
evaluations especially for reference

609
00:24:37,200 --> 00:24:41,679
list metrics that don't use a reference

610
00:24:38,960 --> 00:24:44,840
output then um you start basically

611
00:24:41,679 --> 00:24:47,440
exploiting the metric

612
00:24:44,840 --> 00:24:49,840
um another thing that you can do with QA

613
00:24:47,440 --> 00:24:53,279
based evaluation is ask about fine grade

614
00:24:49,840 --> 00:24:54,919
mistakes and so this is a paper by um uh

615
00:24:53,279 --> 00:24:56,480
Patrick Fernandez who's a student who's

616
00:24:54,919 --> 00:25:02,080
working with me and basically what we

617
00:24:56,480 --> 00:25:05,240
did is we asked the model to um not give

618
00:25:02,080 --> 00:25:07,360
a particular score but actually identify

619
00:25:05,240 --> 00:25:08,880
the mistakes in the output and when we

620
00:25:07,360 --> 00:25:10,559
asked it to identify the mistakes in the

621
00:25:08,880 --> 00:25:13,720
output we found that this gave more

622
00:25:10,559 --> 00:25:17,320
consistent uh results so kind of

623
00:25:13,720 --> 00:25:18,840
interestingly we ask humans to identify

624
00:25:17,320 --> 00:25:21,120
individual mistakes and the output that

625
00:25:18,840 --> 00:25:24,240
gives humans more consistent results

626
00:25:21,120 --> 00:25:25,559
it's the same thing for gp4 so um that

627
00:25:24,240 --> 00:25:27,320
that's another paper you can look at if

628
00:25:25,559 --> 00:25:29,640
you're

629
00:25:27,320 --> 00:25:32,679
interested

630
00:25:29,640 --> 00:25:38,000
cool um so I I mentioned that you could

631
00:25:32,679 --> 00:25:38,000
or could not uh trust uh yeah sorry go

632
00:25:44,679 --> 00:25:51,279
ahead uh correct so yeah B basically

633
00:25:47,360 --> 00:25:53,279
just what you do is you have the source

634
00:25:51,279 --> 00:25:54,960
um ideally you'll also have a reference

635
00:25:53,279 --> 00:25:57,840
output that was created by skilled

636
00:25:54,960 --> 00:25:59,720
humans and then you put in the Target

637
00:25:57,840 --> 00:26:02,279
you know output basically you have the

638
00:25:59,720 --> 00:26:08,000
input ideally a reference output created

639
00:26:02,279 --> 00:26:08,000
by Good by skilled humans and uh like

640
00:26:15,159 --> 00:26:20,240
hypothesis yeah I

641
00:26:17,919 --> 00:26:24,559
mean it's a good question and I don't

642
00:26:20,240 --> 00:26:26,919
know if we actually have a a very clear

643
00:26:24,559 --> 00:26:31,399
empirical like evidence of why this is

644
00:26:26,919 --> 00:26:33,320
the case but my hypothesis about this is

645
00:26:31,399 --> 00:26:36,159
yes we kind of would expect models to be

646
00:26:33,320 --> 00:26:38,200
more biased towards their own outputs

647
00:26:36,159 --> 00:26:40,919
and the reason why is because

648
00:26:38,200 --> 00:26:43,080
essentially you know models

649
00:26:40,919 --> 00:26:44,279
are within their embeddings they're

650
00:26:43,080 --> 00:26:45,760
encoding when they're in a high

651
00:26:44,279 --> 00:26:47,600
probability part of the space and when

652
00:26:45,760 --> 00:26:50,200
they're in a low probability part of the

653
00:26:47,600 --> 00:26:51,120
space and like the high probability part

654
00:26:50,200 --> 00:26:54,600
of the

655
00:26:51,120 --> 00:26:56,200
space is going to be the high

656
00:26:54,600 --> 00:26:58,600
probability part of the space is going

657
00:26:56,200 --> 00:27:02,559
to be associated with good outputs

658
00:26:58,600 --> 00:27:07,000
because like when

659
00:27:02,559 --> 00:27:08,600
models are more sure of their outputs

660
00:27:07,000 --> 00:27:11,960
they're more likely to be

661
00:27:08,600 --> 00:27:13,520
good just because that indicates that

662
00:27:11,960 --> 00:27:15,240
like they're closer to the training data

663
00:27:13,520 --> 00:27:17,760
that it had and other things like that

664
00:27:15,240 --> 00:27:21,600
so model probabilities are associated

665
00:27:17,760 --> 00:27:23,760
with outputs uh with uh with good

666
00:27:21,600 --> 00:27:26,600
outputs but just

667
00:27:23,760 --> 00:27:29,440
correla separately from

668
00:27:26,600 --> 00:27:32,120
that I believe a model can identify when

669
00:27:29,440 --> 00:27:33,320
it's in a high probability segment of

670
00:27:32,120 --> 00:27:35,799
the space and when it's in a low

671
00:27:33,320 --> 00:27:39,399
probability segment of the space and

672
00:27:35,799 --> 00:27:39,399
because of that I expect

673
00:27:39,519 --> 00:27:45,519
that I like there are segments of the

674
00:27:43,240 --> 00:27:47,120
embedding space where it's more likely

675
00:27:45,519 --> 00:27:48,360
to answer yes about something being good

676
00:27:47,120 --> 00:27:50,960
or not and those are going to be

677
00:27:48,360 --> 00:27:54,760
associated with high uh like high

678
00:27:50,960 --> 00:27:56,159
probability outbreaks as well and also

679
00:27:54,760 --> 00:27:57,760
models are more likely to generate

680
00:27:56,159 --> 00:28:00,240
outputs that are high probability

681
00:27:57,760 --> 00:28:02,320
according into their model by definition

682
00:28:00,240 --> 00:28:03,880
so all three of those effects together

683
00:28:02,320 --> 00:28:05,640
would basically go into a model being

684
00:28:03,880 --> 00:28:09,120
bios supports its own outputs compared

685
00:28:05,640 --> 00:28:11,559
to that puts in another model but um

686
00:28:09,120 --> 00:28:13,279
yeah this is a very handwavy explanation

687
00:28:11,559 --> 00:28:15,519
but like putting the two the three

688
00:28:13,279 --> 00:28:18,600
together models output high probability

689
00:28:15,519 --> 00:28:20,880
things from their own probability Space

690
00:28:18,600 --> 00:28:23,440
by definition

691
00:28:20,880 --> 00:28:25,760
um things that are high probability are

692
00:28:23,440 --> 00:28:27,519
associated with being good uh just

693
00:28:25,760 --> 00:28:29,279
because otherwise a model would be

694
00:28:27,519 --> 00:28:31,840
outputting garbage

695
00:28:29,279 --> 00:28:33,840
and um the final thing which is more

696
00:28:31,840 --> 00:28:35,679
tenuous is if the model is in a high

697
00:28:33,840 --> 00:28:37,919
probability segment of the space it's

698
00:28:35,679 --> 00:28:39,760
more likely to Output yes according to a

699
00:28:37,919 --> 00:28:41,480
question of it being good and I I think

700
00:28:39,760 --> 00:28:44,360
that's probably true but I'm not 100%

701
00:28:41,480 --> 00:28:44,360
sure about the the

702
00:28:45,559 --> 00:28:51,039
fin um maybe maybe someone wants to

703
00:28:49,000 --> 00:28:52,840
examinate examine that as a final

704
00:28:51,039 --> 00:28:54,200
project it seems like a interesting

705
00:28:52,840 --> 00:28:57,080
interesting

706
00:28:54,200 --> 00:29:00,039
question um cool uh were there any other

707
00:28:57,080 --> 00:29:00,039
questions about these methods

708
00:29:00,159 --> 00:29:07,120
here um okay so when I say like an

709
00:29:03,960 --> 00:29:11,080
evaluation metric is good or not what do

710
00:29:07,120 --> 00:29:13,200
I mean by this being good or not um or a

711
00:29:11,080 --> 00:29:16,880
reward model or whatever else and

712
00:29:13,200 --> 00:29:18,440
basically the um the way we typically do

713
00:29:16,880 --> 00:29:19,840
this is by doing something called meta

714
00:29:18,440 --> 00:29:22,440
evaluation so it's called meta

715
00:29:19,840 --> 00:29:25,799
evaluation because it's evaluation of

716
00:29:22,440 --> 00:29:29,279
evaluation and uh the way we do this is

717
00:29:25,799 --> 00:29:32,519
we have human uh scores and we have

718
00:29:29,279 --> 00:29:34,760
automatic scores and we usually

719
00:29:32,519 --> 00:29:38,640
calculate some sort of correlation

720
00:29:34,760 --> 00:29:41,000
between the scores so um typical ones

721
00:29:38,640 --> 00:29:46,440
are rank correlations like Pearson's

722
00:29:41,000 --> 00:29:48,799
correlation or tendle uh Tow and uh so

723
00:29:46,440 --> 00:29:51,200
the more Associated the automatic scores

724
00:29:48,799 --> 00:29:53,960
are with the human scores the higher

725
00:29:51,200 --> 00:29:55,159
these correlations are going to be um

726
00:29:53,960 --> 00:29:57,559
there's other things that you can

727
00:29:55,159 --> 00:30:00,080
calculate so if you're trying to figure

728
00:29:57,559 --> 00:30:01,640
out whether a model um matches human

729
00:30:00,080 --> 00:30:04,279
pairwise preferences you can just

730
00:30:01,640 --> 00:30:06,440
calculate accuracy so I didn't put that

731
00:30:04,279 --> 00:30:08,080
on um I didn't put that on the slide

732
00:30:06,440 --> 00:30:10,880
here but you can just calculate accuracy

733
00:30:08,080 --> 00:30:13,120
of pairwise preferences um you can also

734
00:30:10,880 --> 00:30:15,360
calculate the absolute error between the

735
00:30:13,120 --> 00:30:19,320
the judgments if you want to know uh

736
00:30:15,360 --> 00:30:21,720
whether the absolute error matches so um

737
00:30:19,320 --> 00:30:24,159
the these are good things to do if you

738
00:30:21,720 --> 00:30:25,600
want to use an evaluation metric but you

739
00:30:24,159 --> 00:30:27,200
aren't sure whether it's good or not I

740
00:30:25,600 --> 00:30:29,640
would check to see whether the authors

741
00:30:27,200 --> 00:30:32,000
have done this sort of meta evaluation

742
00:30:29,640 --> 00:30:33,760
if they haven't be a little bit

743
00:30:32,000 --> 00:30:36,960
suspicious if they have be a little bit

744
00:30:33,760 --> 00:30:39,799
less suspicious but um

745
00:30:36,960 --> 00:30:42,960
yeah how do people do this typically uh

746
00:30:39,799 --> 00:30:45,640
usually they create uh data sets like

747
00:30:42,960 --> 00:30:49,440
the WM they use data sets like the WMT

748
00:30:45,640 --> 00:30:53,960
shared tasks um or

749
00:30:49,440 --> 00:30:57,679
uh uh like some evl um but there's also

750
00:30:53,960 --> 00:30:59,960
other ways to create um uh there's also

751
00:30:57,679 --> 00:31:01,639
Lots other data sets but in order to do

752
00:30:59,960 --> 00:31:05,639
this reliably you need a fairly large

753
00:31:01,639 --> 00:31:05,639
data set so it's one thing to be aware

754
00:31:07,080 --> 00:31:10,760
of

755
00:31:08,720 --> 00:31:14,200
cool

756
00:31:10,760 --> 00:31:16,360
um then the final thing um all of the

757
00:31:14,200 --> 00:31:17,919
automatic evaluation methods that I

758
00:31:16,360 --> 00:31:20,240
talked about now are trying to match

759
00:31:17,919 --> 00:31:22,679
human preferences but that's not the

760
00:31:20,240 --> 00:31:24,960
only thing that you necessarily want to

761
00:31:22,679 --> 00:31:28,440
do the final thing that you might want

762
00:31:24,960 --> 00:31:30,840
to do is uh use the model outputs in a

763
00:31:28,440 --> 00:31:34,200
downstream system and see whether they

764
00:31:30,840 --> 00:31:36,399
are effective for that so there's two

765
00:31:34,200 --> 00:31:39,080
concepts of intrinsic evaluation and

766
00:31:36,399 --> 00:31:41,720
extrinsic evaluation so intrinsic

767
00:31:39,080 --> 00:31:44,159
evaluation um evaluates the quality of

768
00:31:41,720 --> 00:31:45,720
the output itself and so that would be

769
00:31:44,159 --> 00:31:48,639
like asking a human directly about how

770
00:31:45,720 --> 00:31:50,720
good is this output extrinsic evaluation

771
00:31:48,639 --> 00:31:53,679
is evaluating output quality by its

772
00:31:50,720 --> 00:31:57,000
utility um and so just to give one

773
00:31:53,679 --> 00:31:58,360
example um if you can evaluate large

774
00:31:57,000 --> 00:32:00,200
language model summary

775
00:31:58,360 --> 00:32:04,200
through question answering

776
00:32:00,200 --> 00:32:05,880
accuracy um and so you can take the

777
00:32:04,200 --> 00:32:07,399
output of an llm and feed it through a

778
00:32:05,880 --> 00:32:09,600
question answering model and see whether

779
00:32:07,399 --> 00:32:12,399
you're able to answer questions based on

780
00:32:09,600 --> 00:32:15,799
this and that kind of gives you a better

781
00:32:12,399 --> 00:32:18,279
idea of whether the summary require uh

782
00:32:15,799 --> 00:32:20,120
incorporates requisite information but

783
00:32:18,279 --> 00:32:22,120
if you think about anything an llm can

784
00:32:20,120 --> 00:32:23,760
be used for usually it's part of a

785
00:32:22,120 --> 00:32:26,679
bigger system so you can evaluate it as

786
00:32:23,760 --> 00:32:28,399
a part of that bigger system um the

787
00:32:26,679 --> 00:32:30,639
problem with this is it's a very

788
00:32:28,399 --> 00:32:33,960
indirect way of assessing things so like

789
00:32:30,639 --> 00:32:36,080
let's say your QA model is just bad uh

790
00:32:33,960 --> 00:32:38,480
how can you disentangle the effect of

791
00:32:36,080 --> 00:32:41,679
the L summary versus the QA model that's

792
00:32:38,480 --> 00:32:44,120
not a trivial thing to do so ideally

793
00:32:41,679 --> 00:32:47,000
like a combination of these two is

794
00:32:44,120 --> 00:32:47,000
practically the best way

795
00:32:48,039 --> 00:32:52,200
go cool so

796
00:32:56,039 --> 00:32:59,960
yeah yeah it wouldn't necessar

797
00:32:58,360 --> 00:33:05,679
say it's harder to do it might even be

798
00:32:59,960 --> 00:33:05,679
easier to do um which is like let's

799
00:33:06,679 --> 00:33:11,720
say Let me let me see if I can come up

800
00:33:09,360 --> 00:33:11,720
with

801
00:33:12,639 --> 00:33:17,600
example what let's

802
00:33:15,000 --> 00:33:19,670
say you

803
00:33:17,600 --> 00:33:22,979
are trying

804
00:33:19,670 --> 00:33:22,979
[Music]

805
00:33:24,639 --> 00:33:29,760
to let's say you're trying to

806
00:33:30,559 --> 00:33:33,559
guess

807
00:33:39,000 --> 00:33:45,399
whether let's say you're trying to guess

808
00:33:42,399 --> 00:33:46,559
whether a someone will be hired at a

809
00:33:45,399 --> 00:33:52,039
company or

810
00:33:46,559 --> 00:33:53,880
not based on an llm generated summary of

811
00:33:52,039 --> 00:33:58,880
their qualifications for a position or

812
00:33:53,880 --> 00:34:01,799
something like that um and

813
00:33:58,880 --> 00:34:03,080
you what actually maybe this is not a

814
00:34:01,799 --> 00:34:04,720
great example because whether you should

815
00:34:03,080 --> 00:34:06,960
be doing this ethically is a little bit

816
00:34:04,720 --> 00:34:08,159
unclear but let's say you were doing

817
00:34:06,960 --> 00:34:09,560
let's say you were doing something like

818
00:34:08,159 --> 00:34:11,520
that just because it's one example I can

819
00:34:09,560 --> 00:34:14,320
think of right now whether they will get

820
00:34:11,520 --> 00:34:16,320
hired or not is um is clear because you

821
00:34:14,320 --> 00:34:19,399
have a objective answer right whether

822
00:34:16,320 --> 00:34:21,480
they were hired or not um or maybe maybe

823
00:34:19,399 --> 00:34:23,800
another example would be like let's say

824
00:34:21,480 --> 00:34:26,320
um let's say you want to predict the

825
00:34:23,800 --> 00:34:29,599
diagnosis in a medical application based

826
00:34:26,320 --> 00:34:32,960
on an llm generated some of somebody's

827
00:34:29,599 --> 00:34:35,919
uh you know LM generated summary of

828
00:34:32,960 --> 00:34:38,480
somebody's you know past medical history

829
00:34:35,919 --> 00:34:40,839
and all this stuff and here you want the

830
00:34:38,480 --> 00:34:43,440
llm generated summary you definitely

831
00:34:40,839 --> 00:34:44,879
want the summary because the summary is

832
00:34:43,440 --> 00:34:47,560
going to be viewed by a doctor who will

833
00:34:44,879 --> 00:34:49,359
make the final decision but you also

834
00:34:47,560 --> 00:34:50,760
have information about the diagnoses of

835
00:34:49,359 --> 00:34:52,399
all the people in your medical system

836
00:34:50,760 --> 00:34:54,560
later because you know they went through

837
00:34:52,399 --> 00:34:56,480
your medical system for years and you

838
00:34:54,560 --> 00:34:58,200
know later like through lots of tests

839
00:34:56,480 --> 00:35:00,800
and stuff uh whether how they were

840
00:34:58,200 --> 00:35:02,320
diagnosed so you generate an LM based

841
00:35:00,800 --> 00:35:05,000
summary and then you predict the

842
00:35:02,320 --> 00:35:06,599
diagnosis from the summary so there the

843
00:35:05,000 --> 00:35:08,040
evaluation of the diagnosis is very

844
00:35:06,599 --> 00:35:11,480
clear because you kind of have a gold

845
00:35:08,040 --> 00:35:12,599
standard answer um but the EV intrinsic

846
00:35:11,480 --> 00:35:14,839
evaluation of whether it's a good

847
00:35:12,599 --> 00:35:16,839
summary or not is not as clear because

848
00:35:14,839 --> 00:35:19,400
you'd have pass do whether it's good and

849
00:35:16,839 --> 00:35:21,079
understandable summary so the extrinsic

850
00:35:19,400 --> 00:35:24,920
evaluation might be easier because it's

851
00:35:21,079 --> 00:35:26,480
clearer um so there are cases like that

852
00:35:24,920 --> 00:35:30,720
um the problem is you would have to have

853
00:35:26,480 --> 00:35:33,800
that data in order to do that um yeah do

854
00:35:30,720 --> 00:35:38,240
like evaluation yeah I was just

855
00:35:33,800 --> 00:35:40,800
wondering typically the

856
00:35:38,240 --> 00:35:42,880
like like how do you accomodate the

857
00:35:40,800 --> 00:35:47,160
diversity oh yeah that's a great that's

858
00:35:42,880 --> 00:35:50,240
a great question um so how do you how do

859
00:35:47,160 --> 00:35:50,240
you get these scores

860
00:35:50,720 --> 00:35:55,800
here there's a number of different

861
00:35:53,200 --> 00:35:59,160
things in the WMT shared tasks what they

862
00:35:55,800 --> 00:36:00,280
did is they did

863
00:35:59,160 --> 00:36:03,200
the first thing they do is they

864
00:36:00,280 --> 00:36:06,319
normalize by annotator and what they do

865
00:36:03,200 --> 00:36:10,400
is they basically take the zcore or Z

866
00:36:06,319 --> 00:36:12,240
score of the um of the human annotator's

867
00:36:10,400 --> 00:36:14,880
actual scores because some people are

868
00:36:12,240 --> 00:36:16,400
more harsh than other people and so what

869
00:36:14,880 --> 00:36:20,680
that means is you basically normalize to

870
00:36:16,400 --> 00:36:22,119
have zero mean in unit variance um and

871
00:36:20,680 --> 00:36:24,119
then after they've normalized to zero

872
00:36:22,119 --> 00:36:29,560
mean and unit variance then I think they

873
00:36:24,119 --> 00:36:29,560
average together different humans so um

874
00:36:30,160 --> 00:36:36,520
then for how do you deal with the fact

875
00:36:33,680 --> 00:36:38,040
that humans disagree on things and I

876
00:36:36,520 --> 00:36:39,480
think it's pretty varied I don't know if

877
00:36:38,040 --> 00:36:42,160
there's any gold standard way of doing

878
00:36:39,480 --> 00:36:43,839
it but sometimes you just average

879
00:36:42,160 --> 00:36:46,359
sometimes you throw away examples where

880
00:36:43,839 --> 00:36:47,960
humans disagree a lot um because like

881
00:36:46,359 --> 00:36:50,200
you can't get the humans to agree how

882
00:36:47,960 --> 00:36:53,319
could you expect how could you expect a

883
00:36:50,200 --> 00:36:55,119
machine to do well um so I think it it's

884
00:36:53,319 --> 00:36:59,200
a little bit test

885
00:36:55,119 --> 00:37:01,560
defending yeah so for

886
00:36:59,200 --> 00:37:04,560
generation inin

887
00:37:01,560 --> 00:37:06,280
andin yeah so for code generation that's

888
00:37:04,560 --> 00:37:08,200
I I I love this example because I've

889
00:37:06,280 --> 00:37:09,960
worked on code generation a lot of

890
00:37:08,200 --> 00:37:12,680
people only think about extrinsic

891
00:37:09,960 --> 00:37:14,400
evaluation of code Generation Um or I

892
00:37:12,680 --> 00:37:16,160
don't know if it's extrinsic but only

893
00:37:14,400 --> 00:37:19,160
think about execution based evaluation

894
00:37:16,160 --> 00:37:20,520
of code generation which is like you

895
00:37:19,160 --> 00:37:22,400
execute the code you see whether it

896
00:37:20,520 --> 00:37:25,040
passs unit tests and other things like

897
00:37:22,400 --> 00:37:26,839
this but in reality actually there's a

898
00:37:25,040 --> 00:37:28,599
lot of other important things for code

899
00:37:26,839 --> 00:37:30,560
like readability and other stuff like

900
00:37:28,599 --> 00:37:32,160
that and you should be evaluating those

901
00:37:30,560 --> 00:37:34,920
things but I think a lot of people like

902
00:37:32,160 --> 00:37:36,520
kind of ignore that so um there there

903
00:37:34,920 --> 00:37:38,880
are a few Pap that do that but most of

904
00:37:36,520 --> 00:37:41,000
the time people just execute the Cod

905
00:37:38,880 --> 00:37:45,520
process

906
00:37:41,000 --> 00:37:47,760
un cool okay um so yeah moving on to the

907
00:37:45,520 --> 00:37:51,160
learning part so now I'd like to talk

908
00:37:47,760 --> 00:37:55,280
about uh learning and the first thing

909
00:37:51,160 --> 00:37:59,480
I'll cover is error and risk and so

910
00:37:55,280 --> 00:38:02,280
basically um the way we calculate air is

911
00:37:59,480 --> 00:38:03,119
we generate an output and we calculate

912
00:38:02,280 --> 00:38:07,680
its

913
00:38:03,119 --> 00:38:09,480
Badness um and so generating the output

914
00:38:07,680 --> 00:38:13,160
could be argmax it could be sampling it

915
00:38:09,480 --> 00:38:15,800
could be anything else like that um and

916
00:38:13,160 --> 00:38:18,640
we calculate its Badness uh which is one

917
00:38:15,800 --> 00:38:21,040
minus in which could be like how bad is

918
00:38:18,640 --> 00:38:22,720
the output uh if you're you have a

919
00:38:21,040 --> 00:38:24,760
Badness measure or it could be one minus

920
00:38:22,720 --> 00:38:28,400
the evaluation Square to calculate its

921
00:38:24,760 --> 00:38:30,160
Badness and this is defined as error

922
00:38:28,400 --> 00:38:31,440
and generally what you want to do is you

923
00:38:30,160 --> 00:38:33,520
want to minimize

924
00:38:31,440 --> 00:38:36,800
error

925
00:38:33,520 --> 00:38:39,400
um because in the end you're going to be

926
00:38:36,800 --> 00:38:42,359
deploying A system that just outputs you

927
00:38:39,400 --> 00:38:46,079
know one thing and uh you're going to

928
00:38:42,359 --> 00:38:49,800
want that to be as good a thing as

929
00:38:46,079 --> 00:38:53,000
possible um but the problem with this is

930
00:38:49,800 --> 00:38:56,400
there's no easy way to actually optimize

931
00:38:53,000 --> 00:38:59,079
this value in especially in a text

932
00:38:56,400 --> 00:39:01,800
generation sty setting but even in the

933
00:38:59,079 --> 00:39:06,839
classification setting we can't easily

934
00:39:01,800 --> 00:39:06,839
maximize err because um if you look at

935
00:39:09,040 --> 00:39:14,200
the if you look at the surface of air uh

936
00:39:12,760 --> 00:39:15,960
at some point you're going to have a

937
00:39:14,200 --> 00:39:18,319
non-differentiable part when you take

938
00:39:15,960 --> 00:39:21,119
the argmax and or when you do sampling

939
00:39:18,319 --> 00:39:23,319
or anything like that so um you're not

940
00:39:21,119 --> 00:39:27,119
going to be able to do gradient based

941
00:39:23,319 --> 00:39:29,200
optimization so what we do normally is

942
00:39:27,119 --> 00:39:33,400
um

943
00:39:29,200 --> 00:39:37,000
we instead calculate something uh called

944
00:39:33,400 --> 00:39:38,560
risk and what risk looks like is uh we

945
00:39:37,000 --> 00:39:40,599
talked a little bit about minimum based

946
00:39:38,560 --> 00:39:43,520
risk for decoding but this is for uh

947
00:39:40,599 --> 00:39:46,160
training time and what it looks like is

948
00:39:43,520 --> 00:39:49,040
it's essentially the expected err of the

949
00:39:46,160 --> 00:39:52,359
output and the expected err of the

950
00:39:49,040 --> 00:39:54,760
output um includes a probability in the

951
00:39:52,359 --> 00:39:58,240
objective function here and that

952
00:39:54,760 --> 00:40:01,079
probability uh is differential basically

953
00:39:58,240 --> 00:40:02,319
so we can um uh we can easily do

954
00:40:01,079 --> 00:40:05,720
gradient based

955
00:40:02,319 --> 00:40:09,119
optimization through it um the problem

956
00:40:05,720 --> 00:40:12,200
with this is It's differentiable but for

957
00:40:09,119 --> 00:40:17,160
text generation for example the sum is

958
00:40:12,200 --> 00:40:20,319
intractable because we have a combinator

959
00:40:17,160 --> 00:40:23,880
large number of potential outputs um

960
00:40:20,319 --> 00:40:25,520
because you know if this is we've talked

961
00:40:23,880 --> 00:40:28,720
about this before but if this is like

962
00:40:25,520 --> 00:40:30,680
link you know 50 and we have a 30,000

963
00:40:28,720 --> 00:40:32,839
vocabul that's 30,000 to the 50

964
00:40:30,680 --> 00:40:34,599
possibilities we can't take a su over

965
00:40:32,839 --> 00:40:36,359
that many

966
00:40:34,599 --> 00:40:38,400
possibilities

967
00:40:36,359 --> 00:40:42,680
um

968
00:40:38,400 --> 00:40:45,839
so minimum R risk training uh tries to

969
00:40:42,680 --> 00:40:48,440
minimize risk reinforcement learning

970
00:40:45,839 --> 00:40:50,040
also many of the models especially

971
00:40:48,440 --> 00:40:53,599
policy gradient models are trying to

972
00:40:50,040 --> 00:40:55,240
minimize risk as well so um but the

973
00:40:53,599 --> 00:40:58,040
reason why I wanted to talk about risk

974
00:40:55,240 --> 00:41:00,440
first is because this is very simple to

975
00:40:58,040 --> 00:41:01,640
get to from the uh the point of view of

976
00:41:00,440 --> 00:41:06,560
like all the things that we've studied

977
00:41:01,640 --> 00:41:06,560
so so I think it's talking about

978
00:41:06,760 --> 00:41:11,800
that

979
00:41:08,319 --> 00:41:15,520
um one other thing that I should mention

980
00:41:11,800 --> 00:41:18,400
about is

981
00:41:15,520 --> 00:41:23,079
um or no sorry I'll I'll talk about that

982
00:41:18,400 --> 00:41:26,880
later so when we want to optimize risk

983
00:41:23,079 --> 00:41:30,560
um what we do is we sample in order to

984
00:41:26,880 --> 00:41:35,520
make this trct so a very simple way to

985
00:41:30,560 --> 00:41:37,640
minimize risk is instead of um instead

986
00:41:35,520 --> 00:41:39,359
of summing over all of the possible

987
00:41:37,640 --> 00:41:42,760
outputs we sum over a small number of

988
00:41:39,359 --> 00:41:46,079
possible outputs and we upgrade uh and

989
00:41:42,760 --> 00:41:47,359
we uh sorry normalize uh to make this

990
00:41:46,079 --> 00:41:51,200
all add up to

991
00:41:47,359 --> 00:41:52,839
one and so this normalizer here is

992
00:41:51,200 --> 00:41:55,319
basically the sum over all of the

993
00:41:52,839 --> 00:41:58,599
probabilities that we have uh on the top

994
00:41:55,319 --> 00:42:02,119
part here and and these samples can be

995
00:41:58,599 --> 00:42:05,480
created either using sampling or n best

996
00:42:02,119 --> 00:42:07,040
search we don't need to have from the

997
00:42:05,480 --> 00:42:11,040
point of view of doing this sort of

998
00:42:07,040 --> 00:42:13,960
minimum risk training the kind of

999
00:42:11,040 --> 00:42:16,880
correct way of doing this is sampling

1000
00:42:13,960 --> 00:42:19,880
using ancestral sampling uh like we

1001
00:42:16,880 --> 00:42:23,079
talked about before and um in minimizing

1002
00:42:19,880 --> 00:42:25,839
the output based on the the samples but

1003
00:42:23,079 --> 00:42:28,480
the problem with that is um as many of

1004
00:42:25,839 --> 00:42:31,440
you also might have seen when you were

1005
00:42:28,480 --> 00:42:33,599
sampling from your language model uh

1006
00:42:31,440 --> 00:42:35,160
from assignment one if you sample with

1007
00:42:33,599 --> 00:42:38,040
temperature one it gives you a lot of

1008
00:42:35,160 --> 00:42:40,720
like not very good outlets right and so

1009
00:42:38,040 --> 00:42:43,400
if you're sampling with temperature one

1010
00:42:40,720 --> 00:42:45,000
um you'll be exploring a a very large

1011
00:42:43,400 --> 00:42:47,880
part of the space that actually isn't

1012
00:42:45,000 --> 00:42:49,720
very good and so because of this uh some

1013
00:42:47,880 --> 00:42:51,480
other Alternatives that you can use is

1014
00:42:49,720 --> 00:42:53,400
you can just do endb search to find the

1015
00:42:51,480 --> 00:42:55,280
best outputs or you can sample with a

1016
00:42:53,400 --> 00:42:58,079
temperature that's not one or something

1017
00:42:55,280 --> 00:43:00,240
like that and basically create uh you

1018
00:42:58,079 --> 00:43:02,520
know a list of possible hypotheses and

1019
00:43:00,240 --> 00:43:04,079
then normalize other B so that's another

1020
00:43:02,520 --> 00:43:06,240
option and very often not using

1021
00:43:04,079 --> 00:43:11,200
temperature one is a better

1022
00:43:06,240 --> 00:43:15,280
way um if you're sampling with not

1023
00:43:11,200 --> 00:43:18,640
temperature one and you are um

1024
00:43:15,280 --> 00:43:20,920
potentially getting multiple outputs you

1025
00:43:18,640 --> 00:43:23,400
should try to D duplicate or sample

1026
00:43:20,920 --> 00:43:25,480
without replacement because if you get

1027
00:43:23,400 --> 00:43:27,559
multiple outputs here it messes up your

1028
00:43:25,480 --> 00:43:30,680
equations if you basically uh have the

1029
00:43:27,559 --> 00:43:30,680
same one in there multiple

1030
00:43:32,160 --> 00:43:37,800
times cool so so this is a really simple

1031
00:43:35,880 --> 00:43:40,079
example of how you can do minimal risk

1032
00:43:37,800 --> 00:43:42,119
training but now I want to get into uh

1033
00:43:40,079 --> 00:43:44,640
like reinforcement learning which is the

1034
00:43:42,119 --> 00:43:48,119
framing that most um

1035
00:43:44,640 --> 00:43:50,760
modern Works about this Paulo uh one

1036
00:43:48,119 --> 00:43:52,559
thing I should mention is there are

1037
00:43:50,760 --> 00:43:55,240
actually other alternatives to learning

1038
00:43:52,559 --> 00:43:57,359
from uh human feedback including like

1039
00:43:55,240 --> 00:43:59,359
margin loss margin based losses and

1040
00:43:57,359 --> 00:44:00,960
other stuff like that but most people

1041
00:43:59,359 --> 00:44:03,440
nowadays use reinforcement learning so

1042
00:44:00,960 --> 00:44:06,359
I'm only going to cover that

1043
00:44:03,440 --> 00:44:08,440
here so what is reinforcement learning

1044
00:44:06,359 --> 00:44:11,000
um learning reinforcement learning is

1045
00:44:08,440 --> 00:44:14,559
learning where we have an environment uh

1046
00:44:11,000 --> 00:44:16,079
x uh ability to make actions a and get a

1047
00:44:14,559 --> 00:44:20,160
delayed reward

1048
00:44:16,079 --> 00:44:21,880
R and um there's a really nice example

1049
00:44:20,160 --> 00:44:24,400
uh if you're not familiar with the

1050
00:44:21,880 --> 00:44:27,480
basics of policy gradient by Andre

1051
00:44:24,400 --> 00:44:28,800
karpathy which I linked in the um in the

1052
00:44:27,480 --> 00:44:29,680
recommended reading so you can take a

1053
00:44:28,800 --> 00:44:34,680
look at

1054
00:44:29,680 --> 00:44:37,240
that um but in that example gives an

1055
00:44:34,680 --> 00:44:39,440
example of pong uh where you're playing

1056
00:44:37,240 --> 00:44:42,640
the game pong where X is your observed

1057
00:44:39,440 --> 00:44:45,640
image a is up or down and R is the wind

1058
00:44:42,640 --> 00:44:47,480
loss at the end of the game uh does

1059
00:44:45,640 --> 00:44:50,559
anyone have an idea about uh what this

1060
00:44:47,480 --> 00:44:52,119
looks like for any arbitrary NLP task

1061
00:44:50,559 --> 00:44:56,520
that we might want to do reinforcement

1062
00:44:52,119 --> 00:44:59,040
learning for so what what is X what is a

1063
00:44:56,520 --> 00:44:59,040
and what is

1064
00:45:00,040 --> 00:45:04,680
are pick your favorite uh your favorite

1065
00:45:06,920 --> 00:45:09,920
Trask

1066
00:45:10,960 --> 00:45:18,400
anybody

1067
00:45:12,520 --> 00:45:18,400
yeah be or what what's X first

1068
00:45:19,680 --> 00:45:28,720
yeah you have generate okay is the

1069
00:45:24,440 --> 00:45:29,720
next be like the Buton like whether or

1070
00:45:28,720 --> 00:45:32,520
not

1071
00:45:29,720 --> 00:45:35,240
you okay yeah I I think this is very

1072
00:45:32,520 --> 00:45:37,119
close just to repeat it it's like X is

1073
00:45:35,240 --> 00:45:39,599
what you've generated so far a is the

1074
00:45:37,119 --> 00:45:41,559
next token and R is the button that the

1075
00:45:39,599 --> 00:45:45,400
user clicks about whether it's good or

1076
00:45:41,559 --> 00:45:46,920
not um I think that's reasonably good

1077
00:45:45,400 --> 00:45:48,760
although I don't know if we'd expect

1078
00:45:46,920 --> 00:45:52,960
them to click the button every token we

1079
00:45:48,760 --> 00:45:54,880
generate right so um it might be that X

1080
00:45:52,960 --> 00:45:57,880
is the conversational history up till

1081
00:45:54,880 --> 00:46:02,319
this point um a

1082
00:45:57,880 --> 00:46:04,280
a could be a next token generation and

1083
00:46:02,319 --> 00:46:06,520
then R is a reward we get in an

1084
00:46:04,280 --> 00:46:08,280
arbitrary time point it might not be

1085
00:46:06,520 --> 00:46:09,960
like immediately after generating the

1086
00:46:08,280 --> 00:46:12,040
next token but it might be later and

1087
00:46:09,960 --> 00:46:13,480
that's actually really really important

1088
00:46:12,040 --> 00:46:15,040
from the point of view of reinforcement

1089
00:46:13,480 --> 00:46:19,599
learning and I'll I'll talk about that

1090
00:46:15,040 --> 00:46:23,040
in a second um anyone have an idea from

1091
00:46:19,599 --> 00:46:24,960
I don't know uh code generation or

1092
00:46:23,040 --> 00:46:28,119
translation or some other

1093
00:46:24,960 --> 00:46:31,160
things C generation maybe s is a

1094
00:46:28,119 --> 00:46:33,040
compiler or like the gra scpt and then

1095
00:46:31,160 --> 00:46:37,000
the

1096
00:46:33,040 --> 00:46:42,520
is the actual code that right and reward

1097
00:46:37,000 --> 00:46:44,839
is yep um so X could be the compiler

1098
00:46:42,520 --> 00:46:47,559
it's probably the compiler and all of

1099
00:46:44,839 --> 00:46:50,200
the surrounding code context like what

1100
00:46:47,559 --> 00:46:52,520
what is the natural language output and

1101
00:46:50,200 --> 00:46:53,960
it's also um you know what is the

1102
00:46:52,520 --> 00:46:57,280
project that you're you're working on

1103
00:46:53,960 --> 00:47:00,079
and stuff like that um a i think

1104
00:46:57,280 --> 00:47:02,800
typically we would treat each token in

1105
00:47:00,079 --> 00:47:04,160
the code to be an action um and then R

1106
00:47:02,800 --> 00:47:06,599
would be the reward after a long

1107
00:47:04,160 --> 00:47:08,640
sequence of actions um and it could be

1108
00:47:06,599 --> 00:47:11,119
the reward from the compiler it could be

1109
00:47:08,640 --> 00:47:13,160
the reward from a code readability model

1110
00:47:11,119 --> 00:47:15,720
it could be the reward from a speed

1111
00:47:13,160 --> 00:47:17,079
execution speed and stuff like that so

1112
00:47:15,720 --> 00:47:18,839
like one of the interesting things about

1113
00:47:17,079 --> 00:47:22,640
R is you can be really creative about

1114
00:47:18,839 --> 00:47:25,400
how you form R um which is not easy to

1115
00:47:22,640 --> 00:47:27,319
do uh if you're just doing maximum

1116
00:47:25,400 --> 00:47:29,240
likelihood also so you can come up with

1117
00:47:27,319 --> 00:47:32,920
a r that really matches with like what

1118
00:47:29,240 --> 00:47:36,559
you want um what you want in an output

1119
00:47:32,920 --> 00:47:40,079
so why reinforcement learning in NLP um

1120
00:47:36,559 --> 00:47:42,599
and I think there's basically three um

1121
00:47:40,079 --> 00:47:44,240
three answers the first one is you have

1122
00:47:42,599 --> 00:47:49,000
a typical reinforcement learning

1123
00:47:44,240 --> 00:47:51,119
scenario um where you have a dialogue

1124
00:47:49,000 --> 00:47:52,720
where you get lots of responses and then

1125
00:47:51,119 --> 00:47:54,559
you get a reward at the end so the

1126
00:47:52,720 --> 00:47:57,359
thumbs up and thumbs down from humans is

1127
00:47:54,559 --> 00:47:59,839
a very typical example of

1128
00:47:57,359 --> 00:48:02,800
uh reinforcement learning because you

1129
00:47:59,839 --> 00:48:05,000
get a delayed reward uh at some point in

1130
00:48:02,800 --> 00:48:07,599
the dialogue when a human presses up or

1131
00:48:05,000 --> 00:48:09,280
down um another like actually more

1132
00:48:07,599 --> 00:48:11,680
technical scenario where reinforcement

1133
00:48:09,280 --> 00:48:14,960
learning has been used um for a long

1134
00:48:11,680 --> 00:48:17,400
time is call centers so we've had

1135
00:48:14,960 --> 00:48:20,680
dialogue systems for call centers and

1136
00:48:17,400 --> 00:48:23,160
then if you complete a ticket purchase

1137
00:48:20,680 --> 00:48:24,839
um or you complete resolve a ticket

1138
00:48:23,160 --> 00:48:27,480
without ever having to go to a human

1139
00:48:24,839 --> 00:48:30,800
operator you get a really big reward

1140
00:48:27,480 --> 00:48:33,640
if you have to go to the human operator

1141
00:48:30,800 --> 00:48:36,400
you get maybe a smaller reward and if

1142
00:48:33,640 --> 00:48:39,200
the person yells at you and hangs up

1143
00:48:36,400 --> 00:48:41,640
then you get a really negative reward so

1144
00:48:39,200 --> 00:48:43,040
um this is kind of the typical example

1145
00:48:41,640 --> 00:48:45,599
reinforcement learning has been used for

1146
00:48:43,040 --> 00:48:48,520
a long time there another example is if

1147
00:48:45,599 --> 00:48:53,280
you have like latent variables uh chains

1148
00:48:48,520 --> 00:48:55,799
of thought where um you decide the

1149
00:48:53,280 --> 00:48:58,839
latent variable and then get a reward um

1150
00:48:55,799 --> 00:49:02,799
you get a reward based Bas on how those

1151
00:48:58,839 --> 00:49:03,920
latent variables affect the output so um

1152
00:49:02,799 --> 00:49:07,200
this

1153
00:49:03,920 --> 00:49:09,799
is uh this is another example

1154
00:49:07,200 --> 00:49:12,599
because the Chain of Thought itself

1155
00:49:09,799 --> 00:49:13,880
might not actually be good you might

1156
00:49:12,599 --> 00:49:15,839
have a bad Chain of Thought and still

1157
00:49:13,880 --> 00:49:17,760
get the correct answer so you don't

1158
00:49:15,839 --> 00:49:19,640
actually know for sure that a chain of

1159
00:49:17,760 --> 00:49:22,359
thought that was automatically generated

1160
00:49:19,640 --> 00:49:24,799
is good or not but um that so that kind

1161
00:49:22,359 --> 00:49:27,000
of makes it a reinforcement learning

1162
00:49:24,799 --> 00:49:29,520
problem and another thing is you might

1163
00:49:27,000 --> 00:49:32,520
have a sequence level evaluation metric

1164
00:49:29,520 --> 00:49:34,240
um so that you can't optimize the

1165
00:49:32,520 --> 00:49:36,839
evaluation metric without uh first

1166
00:49:34,240 --> 00:49:38,480
generating the whole like sequence so

1167
00:49:36,839 --> 00:49:40,880
that would be any of the evaluation

1168
00:49:38,480 --> 00:49:42,400
metrics that I talked about before so um

1169
00:49:40,880 --> 00:49:44,720
these are three scenarios where you can

1170
00:49:42,400 --> 00:49:47,079
use reinforcement

1171
00:49:44,720 --> 00:49:50,000
planning so

1172
00:49:47,079 --> 00:49:51,400
um I'm going to make a few steps through

1173
00:49:50,000 --> 00:49:54,640
but like let's start again with our

1174
00:49:51,400 --> 00:49:57,359
supervised mle loss and uh that's just

1175
00:49:54,640 --> 00:50:01,799
the log probability here um in the

1176
00:49:57,359 --> 00:50:04,160
context of reinforcement learning this

1177
00:50:01,799 --> 00:50:07,079
is also called imitation

1178
00:50:04,160 --> 00:50:08,880
learning because um essentially you're

1179
00:50:07,079 --> 00:50:12,680
learning how to perform actions by

1180
00:50:08,880 --> 00:50:14,559
imitating a teacher um and imitation

1181
00:50:12,680 --> 00:50:15,960
learning is not just supervised mle

1182
00:50:14,559 --> 00:50:18,440
there's also other varieties of

1183
00:50:15,960 --> 00:50:21,440
imitation learning but um this is one

1184
00:50:18,440 --> 00:50:21,440
variety of imitation

1185
00:50:22,520 --> 00:50:27,640
learning the next thing I'd like to talk

1186
00:50:24,599 --> 00:50:30,079
about is self-training and basically

1187
00:50:27,640 --> 00:50:31,760
self-training the idea is that you

1188
00:50:30,079 --> 00:50:33,720
sample or argmax according to the

1189
00:50:31,760 --> 00:50:36,119
current model so you have your current

1190
00:50:33,720 --> 00:50:38,000
model and you get a sample from it and

1191
00:50:36,119 --> 00:50:41,520
then you use the sample or samples to

1192
00:50:38,000 --> 00:50:43,680
maximize likelihood so um basically

1193
00:50:41,520 --> 00:50:47,520
instead of doing maximum likelihood with

1194
00:50:43,680 --> 00:50:49,520
respect to the a gold standard output

1195
00:50:47,520 --> 00:50:51,280
you're doing it with respect to your own

1196
00:50:49,520 --> 00:50:55,280
output

1197
00:50:51,280 --> 00:50:55,280
so does this seem like a good

1198
00:50:55,640 --> 00:51:03,880
idea I see a few people shaking heads um

1199
00:51:00,480 --> 00:51:03,880
any ideas why this is not a good

1200
00:51:04,680 --> 00:51:07,680
idea

1201
00:51:15,040 --> 00:51:20,599
yeah yeah exactly so if you don't have

1202
00:51:17,720 --> 00:51:23,760
any access to any notion well it's good

1203
00:51:20,599 --> 00:51:27,480
um this will be optimizing towards good

1204
00:51:23,760 --> 00:51:28,839
outputs and bad outputs right so um your

1205
00:51:27,480 --> 00:51:30,200
model might be outputting bad outputs

1206
00:51:28,839 --> 00:51:32,839
and you're just reinforcing the errors

1207
00:51:30,200 --> 00:51:35,160
set the model R already nonetheless like

1208
00:51:32,839 --> 00:51:37,799
self trining actually improves your

1209
00:51:35,160 --> 00:51:39,680
accuracy somewhat in some cases like for

1210
00:51:37,799 --> 00:51:43,040
example if your accuracy is if your

1211
00:51:39,680 --> 00:51:45,520
model is Right more often than not um

1212
00:51:43,040 --> 00:51:49,119
basically optimizing towards the more

1213
00:51:45,520 --> 00:51:51,720
often the not right outputs can actually

1214
00:51:49,119 --> 00:51:53,640
um due to the implicit regularization

1215
00:51:51,720 --> 00:51:55,000
that models have and early stopping and

1216
00:51:53,640 --> 00:51:56,559
other things like that it can actually

1217
00:51:55,000 --> 00:51:59,280
move you in the right direction and

1218
00:51:56,559 --> 00:52:01,559
improve accuracy

1219
00:51:59,280 --> 00:52:05,000
um

1220
00:52:01,559 --> 00:52:06,640
so there are alternatives to this that

1221
00:52:05,000 --> 00:52:09,520
further improve accuracy so like for

1222
00:52:06,640 --> 00:52:12,720
example if you have multiple models and

1223
00:52:09,520 --> 00:52:16,200
um you only generate sentences where the

1224
00:52:12,720 --> 00:52:17,760
models agree then this can improve your

1225
00:52:16,200 --> 00:52:20,000
uh overall accuracy

1226
00:52:17,760 --> 00:52:24,240
further um this is called code training

1227
00:52:20,000 --> 00:52:27,799
it was actually uh created by uh uh

1228
00:52:24,240 --> 00:52:30,160
people at at CMU as well and another

1229
00:52:27,799 --> 00:52:32,280
successful alternative uh is adding

1230
00:52:30,160 --> 00:52:34,920
noise to the input to match the noise

1231
00:52:32,280 --> 00:52:38,760
that you find in the output so if you uh

1232
00:52:34,920 --> 00:52:40,720
add like word uh word-based Dropout or

1233
00:52:38,760 --> 00:52:44,000
other things like that this can also

1234
00:52:40,720 --> 00:52:47,400
help uh accommodate these things but

1235
00:52:44,000 --> 00:52:48,920
anyway um so self trining is is useful

1236
00:52:47,400 --> 00:52:50,480
but there are better Alternatives if you

1237
00:52:48,920 --> 00:52:54,079
can get a reward

1238
00:52:50,480 --> 00:52:55,559
function so um the simplest variety of

1239
00:52:54,079 --> 00:52:56,960
this is something called policy gradient

1240
00:52:55,559 --> 00:52:59,720
or reinforce

1241
00:52:56,960 --> 00:53:02,319
um or more specifically reinforce and

1242
00:52:59,720 --> 00:53:06,280
basically what this does is this adds a

1243
00:53:02,319 --> 00:53:08,359
term that scales the loss by the reward

1244
00:53:06,280 --> 00:53:12,400
so if you can get a reward for each

1245
00:53:08,359 --> 00:53:15,680
output basically this

1246
00:53:12,400 --> 00:53:18,119
um you uh instead of doing self trining

1247
00:53:15,680 --> 00:53:21,760
entirely by itself you multiply it by a

1248
00:53:18,119 --> 00:53:23,119
reward and this allows you to increase

1249
00:53:21,760 --> 00:53:24,640
the likelihood of things that get a high

1250
00:53:23,119 --> 00:53:28,440
reward decrease the likelihood of things

1251
00:53:24,640 --> 00:53:28,440
that get a low reward

1252
00:53:29,680 --> 00:53:34,960
so uh a brief quiz here under what

1253
00:53:32,440 --> 00:53:37,599
conditions is this equal equivalent to

1254
00:53:34,960 --> 00:53:41,480
ml or essentially equivalent to maximum

1255
00:53:37,599 --> 00:53:43,079
leg uh estimation and so like in order

1256
00:53:41,480 --> 00:53:45,480
to make this quiz easier I'll go back to

1257
00:53:43,079 --> 00:53:47,720
maximum likelihood estimation so it

1258
00:53:45,480 --> 00:53:50,359
looked a bit like this um you calculated

1259
00:53:47,720 --> 00:53:53,440
the log probability of the true output

1260
00:53:50,359 --> 00:53:55,440
and now let me go uh to

1261
00:53:53,440 --> 00:53:56,960
here any

1262
00:53:55,440 --> 00:54:00,119
ideas

1263
00:53:56,960 --> 00:54:05,040
yeah when your reward equals to

1264
00:54:00,119 --> 00:54:05,040
one some sometimes in zero other times

1265
00:54:07,760 --> 00:54:10,960
what any

1266
00:54:12,760 --> 00:54:17,520
ideas what when when does your reward

1267
00:54:15,280 --> 00:54:19,640
need to be equal to one in order to make

1268
00:54:17,520 --> 00:54:23,400
this

1269
00:54:19,640 --> 00:54:23,400
equation equivalent this

1270
00:54:24,960 --> 00:54:31,680
equation yeah when Y and Y hat are the

1271
00:54:27,319 --> 00:54:36,119
same so um basically

1272
00:54:31,680 --> 00:54:38,880
this objective is equivalent to the mle

1273
00:54:36,119 --> 00:54:43,160
objective when you're using a zero1

1274
00:54:38,880 --> 00:54:44,480
loss um where or you're using an

1275
00:54:43,160 --> 00:54:46,359
evaluation function that gives you a

1276
00:54:44,480 --> 00:54:50,920
score of one when it's exact match and

1277
00:54:46,359 --> 00:54:51,720
zero when it's not exact match so um but

1278
00:54:50,920 --> 00:54:54,480
that

1279
00:54:51,720 --> 00:54:56,440
also demonstrates that this can be more

1280
00:54:54,480 --> 00:54:58,400
flexible because you can have other

1281
00:54:56,440 --> 00:55:00,160
rewards that are not just one and zero

1282
00:54:58,400 --> 00:55:02,599
for exact match but you can use things

1283
00:55:00,160 --> 00:55:05,359
that give you partial credit you can use

1284
00:55:02,599 --> 00:55:06,880
things that uplate multiple potential uh

1285
00:55:05,359 --> 00:55:08,880
potentially correct outputs and other

1286
00:55:06,880 --> 00:55:13,400
things like

1287
00:55:08,880 --> 00:55:17,160
that so one problem with these methods

1288
00:55:13,400 --> 00:55:21,799
is um how do we know which action led to

1289
00:55:17,160 --> 00:55:24,720
the reward so the best scenario is after

1290
00:55:21,799 --> 00:55:26,359
each action you get a reward so after

1291
00:55:24,720 --> 00:55:28,960
each token that you generated you get

1292
00:55:26,359 --> 00:55:31,240
get a thumbs up or thumbs down uh from

1293
00:55:28,960 --> 00:55:34,280
the user about whether they like that

1294
00:55:31,240 --> 00:55:36,000
token or not um and how much happier

1295
00:55:34,280 --> 00:55:37,720
they are after you generated that token

1296
00:55:36,000 --> 00:55:42,400
than they were before you generated that

1297
00:55:37,720 --> 00:55:44,200
token um the problem with this is that

1298
00:55:42,400 --> 00:55:45,799
that's completely infeasible right like

1299
00:55:44,200 --> 00:55:47,039
every time after you use chat GPD you're

1300
00:55:45,799 --> 00:55:50,480
not going to press thumbs up and thumbs

1301
00:55:47,039 --> 00:55:52,559
down after each token so um in reality

1302
00:55:50,480 --> 00:55:55,559
what we get is usually we get it at the

1303
00:55:52,559 --> 00:55:57,000
end of uh roll out of many many

1304
00:55:55,559 --> 00:55:58,640
different actions and we're not sure

1305
00:55:57,000 --> 00:55:59,720
which action is responsible for giving

1306
00:55:58,640 --> 00:56:02,559
us the

1307
00:55:59,720 --> 00:56:05,440
reward and

1308
00:56:02,559 --> 00:56:08,000
so there's a few typical ways of dealing

1309
00:56:05,440 --> 00:56:09,640
with this um the most typical way of

1310
00:56:08,000 --> 00:56:13,359
dealing with this right now is just not

1311
00:56:09,640 --> 00:56:15,440
dealing with it um and just hoping that

1312
00:56:13,359 --> 00:56:17,200
your optimization algorithm internally

1313
00:56:15,440 --> 00:56:21,480
will be able to do credit

1314
00:56:17,200 --> 00:56:24,520
assignment um and so what that entails

1315
00:56:21,480 --> 00:56:27,319
is essentially you um give an equal

1316
00:56:24,520 --> 00:56:29,880
reward for each token in the output

1317
00:56:27,319 --> 00:56:32,480
other ways that you can deal with it are

1318
00:56:29,880 --> 00:56:35,640
um you can assign decaying rewards from

1319
00:56:32,480 --> 00:56:37,559
future events so like let's say let's

1320
00:56:35,640 --> 00:56:41,839
say you're talking about a chat bot for

1321
00:56:37,559 --> 00:56:44,119
example maybe this is the the most uh

1322
00:56:41,839 --> 00:56:46,599
kind of intuitive way of thinking about

1323
00:56:44,119 --> 00:56:50,400
it but you you have a chat bot you have

1324
00:56:46,599 --> 00:56:52,599
like 20 chat turns and you have the user

1325
00:56:50,400 --> 00:56:55,640
give a thumbs up or a thumbs down on the

1326
00:56:52,599 --> 00:56:58,920
20th chat turn there you would assign a

1327
00:56:55,640 --> 00:57:01,440
reward of um like let's say it gave a

1328
00:56:58,920 --> 00:57:03,640
thumbs up there you would re assign a

1329
00:57:01,440 --> 00:57:06,559
reward of one for the previous chat turn

1330
00:57:03,640 --> 00:57:09,839
a reward of like 0.5 for the second to

1331
00:57:06,559 --> 00:57:11,720
previous chat term a reward of 0.25 for

1332
00:57:09,839 --> 00:57:14,319
the third to previous chat term to

1333
00:57:11,720 --> 00:57:16,160
basically say yeah like the user is

1334
00:57:14,319 --> 00:57:18,240
feeling good at the moment they gave the

1335
00:57:16,160 --> 00:57:20,359
thumbs up and that's probably more

1336
00:57:18,240 --> 00:57:23,400
likely due to the things that happened

1337
00:57:20,359 --> 00:57:23,400
recently so

1338
00:57:23,559 --> 00:57:28,119
yeah we have a

1339
00:57:26,680 --> 00:57:32,280
like not

1340
00:57:28,119 --> 00:57:34,160
learning so the reward model can be any

1341
00:57:32,280 --> 00:57:35,839
of the methods that I talked about

1342
00:57:34,160 --> 00:57:37,480
before so it can be human feedback

1343
00:57:35,839 --> 00:57:39,000
directly like a thumbs up or a thumbs

1344
00:57:37,480 --> 00:57:42,200
down it could also be from a reward

1345
00:57:39,000 --> 00:57:44,599
model uh that was pre-trained you could

1346
00:57:42,200 --> 00:57:47,680
also theoretically learn the reward

1347
00:57:44,599 --> 00:57:52,720
model simultaneously but you'd have to

1348
00:57:47,680 --> 00:57:55,200
simultaneously with the model itself um

1349
00:57:52,720 --> 00:57:57,280
so yeah I'm going to talk a little bit

1350
00:57:55,200 --> 00:58:00,359
about DP which kind of does that a

1351
00:57:57,280 --> 00:58:01,720
little bit but um I I would basically

1352
00:58:00,359 --> 00:58:03,160
say that wherever you're getting your

1353
00:58:01,720 --> 00:58:06,280
reward is probably from one of the

1354
00:58:03,160 --> 00:58:06,280
things I talked about earlier

1355
00:58:06,359 --> 00:58:14,960
today cool any other

1356
00:58:09,319 --> 00:58:17,720
questions okay um so that's the basic

1357
00:58:14,960 --> 00:58:20,640
the basic idea the very simplest thing

1358
00:58:17,720 --> 00:58:23,359
that you can do is you can just sample

1359
00:58:20,640 --> 00:58:26,079
um optimize the subjective function this

1360
00:58:23,359 --> 00:58:28,359
is dead easy you it's not hard to imp

1361
00:58:26,079 --> 00:58:30,799
imp it all as long as you have some

1362
00:58:28,359 --> 00:58:32,760
source of reward signal um but the

1363
00:58:30,799 --> 00:58:35,559
problem is uh reinforcement learning can

1364
00:58:32,760 --> 00:58:38,599
be very unstable and it's hard to get it

1365
00:58:35,559 --> 00:58:40,160
to uh you know work properly if you uh

1366
00:58:38,599 --> 00:58:42,400
don't do some additional tricks so I'd

1367
00:58:40,160 --> 00:58:45,720
like to talk about this

1368
00:58:42,400 --> 00:58:45,720
next oh yeah

1369
00:58:48,880 --> 00:58:51,880
sir

1370
00:58:55,039 --> 00:58:58,039
yeah

1371
00:59:03,280 --> 00:59:08,960
yeah the typical the typical way is you

1372
00:59:05,440 --> 00:59:12,960
just have an exponential decay um so you

1373
00:59:08,960 --> 00:59:16,200
you multiply each time by what 0.5 0. or

1374
00:59:12,960 --> 00:59:19,400
something like that

1375
00:59:16,200 --> 00:59:19,400
um from

1376
00:59:20,319 --> 00:59:27,720
A6 um cool okay

1377
00:59:25,039 --> 00:59:30,720
so

1378
00:59:27,720 --> 00:59:33,319
and that's one option and sorry just to

1379
00:59:30,720 --> 00:59:35,760
clarify the most common option nowadays

1380
00:59:33,319 --> 00:59:37,920
um at least from the point of view of

1381
00:59:35,760 --> 00:59:39,839
models is not to Decay it at all and

1382
00:59:37,920 --> 00:59:43,880
just assign the same amount for each

1383
00:59:39,839 --> 00:59:45,319
token um I'm not actually 100% sure what

1384
00:59:43,880 --> 00:59:47,319
people are doing with respect to like

1385
00:59:45,319 --> 00:59:49,280
long chat things I think probably

1386
00:59:47,319 --> 00:59:51,720
they're only assigning it to the current

1387
00:59:49,280 --> 00:59:54,240
like utterance and then not optimizing

1388
00:59:51,720 --> 00:59:57,240
the previous utterances so like if they

1389
00:59:54,240 --> 00:59:59,039
get a thumbs up or thumbs down signal um

1390
00:59:57,240 --> 01:00:00,720
then they they would assign an

1391
00:59:59,039 --> 01:00:02,440
equivalent reward for all of the tokens

1392
01:00:00,720 --> 01:00:04,640
and the current utterance and zero

1393
01:00:02,440 --> 01:00:06,119
reward for the previous ones but I'm not

1394
01:00:04,640 --> 01:00:08,480
100% sure about that there might be

1395
01:00:06,119 --> 01:00:11,200
other methods that people are

1396
01:00:08,480 --> 01:00:13,960
using um

1397
01:00:11,200 --> 01:00:16,680
cool so uh stabilizing reinforcement

1398
01:00:13,960 --> 01:00:18,520
learning so um stabilizing reinforcement

1399
01:00:16,680 --> 01:00:21,839
learning there's a lot of reasons why

1400
01:00:18,520 --> 01:00:23,880
it's unstable um the first reason is

1401
01:00:21,839 --> 01:00:27,200
you're sampling an individual output and

1402
01:00:23,880 --> 01:00:30,160
calculating the um uh calculating based

1403
01:00:27,200 --> 01:00:32,039
on the S individual sampled output and

1404
01:00:30,160 --> 01:00:33,440
then there's an Infinity of other

1405
01:00:32,039 --> 01:00:36,480
outputs that you could be optimizing

1406
01:00:33,440 --> 01:00:39,119
over for mle this is not a problem

1407
01:00:36,480 --> 01:00:41,319
because for mle you're always

1408
01:00:39,119 --> 01:00:45,359
contrasting the gold standard output to

1409
01:00:41,319 --> 01:00:46,599
all of the other outputs in the space um

1410
01:00:45,359 --> 01:00:48,280
and you're saying I want to upweight the

1411
01:00:46,599 --> 01:00:51,200
gold standard output and down we all of

1412
01:00:48,280 --> 01:00:53,039
the other ones but for reinforcement

1413
01:00:51,200 --> 01:00:54,760
learning you only have a single sampled

1414
01:00:53,039 --> 01:00:57,520
output that output might be wrong and

1415
01:00:54,760 --> 01:00:59,359
that's a source of inst ility this is

1416
01:00:57,520 --> 01:01:02,079
particularly a problem when using bigger

1417
01:00:59,359 --> 01:01:05,960
output spaces like all of the in the

1418
01:01:02,079 --> 01:01:07,920
vocabul another problem is uh anytime

1419
01:01:05,960 --> 01:01:11,599
you start using negative

1420
01:01:07,920 --> 01:01:15,160
rewards um because if you start using

1421
01:01:11,599 --> 01:01:17,559
negative rewards those rewards will be

1422
01:01:15,160 --> 01:01:19,520
downweighting the probability of a

1423
01:01:17,559 --> 01:01:20,680
particular output sequence and that

1424
01:01:19,520 --> 01:01:22,440
might be a good idea maybe you're

1425
01:01:20,680 --> 01:01:24,319
getting a toxic output or something like

1426
01:01:22,440 --> 01:01:25,960
that and you want to down it but at the

1427
01:01:24,319 --> 01:01:28,280
same time in addition to that toxic

1428
01:01:25,960 --> 01:01:30,000
output there's like you know a

1429
01:01:28,280 --> 01:01:31,599
combinatorial number of completely

1430
01:01:30,000 --> 01:01:33,880
nonsense outputs that aren't even

1431
01:01:31,599 --> 01:01:36,599
English and so basically you can start

1432
01:01:33,880 --> 01:01:38,920
diverge from the N starting start to

1433
01:01:36,599 --> 01:01:40,799
diverge from the natural like language

1434
01:01:38,920 --> 01:01:44,720
modeling distribution that you have

1435
01:01:40,799 --> 01:01:49,079
before so this is a big uh a big

1436
01:01:44,720 --> 01:01:51,880
problem so a number of uh strategies can

1437
01:01:49,079 --> 01:01:53,880
be used to stabilize the first one is

1438
01:01:51,880 --> 01:01:55,480
this is completely obvious right now and

1439
01:01:53,880 --> 01:01:57,240
nobody in their right mind would avoid

1440
01:01:55,480 --> 01:02:00,119
doing this but the first one is

1441
01:01:57,240 --> 01:02:02,839
pre-training with mle and so you start

1442
01:02:00,119 --> 01:02:04,920
with a pre-trained model um and then

1443
01:02:02,839 --> 01:02:09,359
switch over to RL after you finished

1444
01:02:04,920 --> 01:02:11,520
pre-training the model um and so

1445
01:02:09,359 --> 01:02:13,279
this makes a lot of sense if you're

1446
01:02:11,520 --> 01:02:14,960
training a language model which I assume

1447
01:02:13,279 --> 01:02:17,039
that almost everybody in this class is

1448
01:02:14,960 --> 01:02:20,279
going to be doing but it does only work

1449
01:02:17,039 --> 01:02:22,720
in scenarios where you can run mle and

1450
01:02:20,279 --> 01:02:24,359
so it doesn't work if you're predicting

1451
01:02:22,720 --> 01:02:27,240
like latent variables that aren't

1452
01:02:24,359 --> 01:02:28,760
included in the original space

1453
01:02:27,240 --> 01:02:31,960
um it

1454
01:02:28,760 --> 01:02:34,279
also doesn't work in a setting where

1455
01:02:31,960 --> 01:02:36,640
like you want to learn a

1456
01:02:34,279 --> 01:02:40,799
chatbot you want to learn a chatbot for

1457
01:02:36,640 --> 01:02:44,200
customer service for a

1458
01:02:40,799 --> 01:02:48,039
company that

1459
01:02:44,200 --> 01:02:49,960
has like for example a product catalog

1460
01:02:48,039 --> 01:02:53,559
that the language model has never seen

1461
01:02:49,960 --> 01:02:56,000
before and so if the language model has

1462
01:02:53,559 --> 01:02:57,359
no information about the product catalog

1463
01:02:56,000 --> 01:02:59,920
whatsoever you don't provide it through

1464
01:02:57,359 --> 01:03:02,440
rag or something like that it's going to

1465
01:02:59,920 --> 01:03:04,039
have to explore infinitely or not

1466
01:03:02,440 --> 01:03:05,599
infinitely but it's going to have to

1467
01:03:04,039 --> 01:03:08,359
explore too large of a space and you're

1468
01:03:05,599 --> 01:03:10,000
never going to converge with um with

1469
01:03:08,359 --> 01:03:12,359
your language modeling objectives so you

1470
01:03:10,000 --> 01:03:15,000
need to basically be able to create at

1471
01:03:12,359 --> 01:03:16,079
least some supervised training data to

1472
01:03:15,000 --> 01:03:19,279
train with

1473
01:03:16,079 --> 01:03:20,720
mle um but assuming you can do that I'm

1474
01:03:19,279 --> 01:03:22,920
assuming that almost everybody is going

1475
01:03:20,720 --> 01:03:26,400
to do some sort of pre-training with

1476
01:03:22,920 --> 01:03:27,880
ML um The Next Step that people use uh

1477
01:03:26,400 --> 01:03:30,520
in reinforcement learning that's really

1478
01:03:27,880 --> 01:03:34,319
important to stabilize is regularization

1479
01:03:30,520 --> 01:03:35,880
to an existing model and you have an

1480
01:03:34,319 --> 01:03:39,039
existing model and you want to prevent

1481
01:03:35,880 --> 01:03:40,559
it from getting too far away and the

1482
01:03:39,039 --> 01:03:42,279
reason why you want to do this is like

1483
01:03:40,559 --> 01:03:45,720
let's say you start assigning a negative

1484
01:03:42,279 --> 01:03:47,440
reward to toxic utterances for example

1485
01:03:45,720 --> 01:03:49,200
if your model stops being a language

1486
01:03:47,440 --> 01:03:51,920
model whatsoever that's a bad idea so

1487
01:03:49,200 --> 01:03:53,400
you want to keep it as a language model

1488
01:03:51,920 --> 01:03:55,599
keep it close enough to still being a

1489
01:03:53,400 --> 01:03:57,559
competent language model while you know

1490
01:03:55,599 --> 01:03:59,599
like removing the toxic

1491
01:03:57,559 --> 01:04:03,039
utterances so there's a number of

1492
01:03:59,599 --> 01:04:05,680
methods that people use to do this um uh

1493
01:04:03,039 --> 01:04:08,359
the most prominent ones are kale

1494
01:04:05,680 --> 01:04:10,279
regularization uh well so the the first

1495
01:04:08,359 --> 01:04:13,119
most prominent one is K regularization

1496
01:04:10,279 --> 01:04:15,839
and the way this works is basically in

1497
01:04:13,119 --> 01:04:19,400
addition you add you have two

1498
01:04:15,839 --> 01:04:22,279
terms the first term is a term that

1499
01:04:19,400 --> 01:04:25,760
improves your reward so you have your

1500
01:04:22,279 --> 01:04:28,039
old model where your old model is

1501
01:04:25,760 --> 01:04:31,279
creating a

1502
01:04:28,039 --> 01:04:32,440
probability uh it has a probability here

1503
01:04:31,279 --> 01:04:34,960
and then you have the probability

1504
01:04:32,440 --> 01:04:38,160
assigned by your new model and then you

1505
01:04:34,960 --> 01:04:41,200
have your reward signal here and so this

1506
01:04:38,160 --> 01:04:43,599
is basically improving the log odds or

1507
01:04:41,200 --> 01:04:46,960
improving the odds of getting a good

1508
01:04:43,599 --> 01:04:49,720
reward for high reward

1509
01:04:46,960 --> 01:04:52,920
sequences separately from this you have

1510
01:04:49,720 --> 01:04:55,920
this K regularization term and this K

1511
01:04:52,920 --> 01:04:58,119
regularization term is keeping the

1512
01:04:55,920 --> 01:05:00,279
scores of or it's keeping the

1513
01:04:58,119 --> 01:05:02,400
probability distribution of your new

1514
01:05:00,279 --> 01:05:03,960
model similar to the probability

1515
01:05:02,400 --> 01:05:09,200
distribution of your old

1516
01:05:03,960 --> 01:05:11,359
model and this beta parameter basically

1517
01:05:09,200 --> 01:05:15,240
you can increase it or decrease it based

1518
01:05:11,359 --> 01:05:18,400
on how similar you want to keep the um

1519
01:05:15,240 --> 01:05:18,400
how similar you want to keep the

1520
01:05:20,720 --> 01:05:24,640
model another method that people use is

1521
01:05:23,160 --> 01:05:29,279
something called proximal policy

1522
01:05:24,640 --> 01:05:30,920
optimization or or Po and this is a

1523
01:05:29,279 --> 01:05:33,920
method that is based on

1524
01:05:30,920 --> 01:05:38,160
clipping uh the

1525
01:05:33,920 --> 01:05:40,920
outputs and we Define uh this ratio

1526
01:05:38,160 --> 01:05:43,880
here so this ratio is equivalent to this

1527
01:05:40,920 --> 01:05:46,160
here so it's basically um kind of the

1528
01:05:43,880 --> 01:05:47,839
amount that you're learning or the

1529
01:05:46,160 --> 01:05:51,720
amount that the new model up weights

1530
01:05:47,839 --> 01:05:54,039
High reward sequences and so here we

1531
01:05:51,720 --> 01:05:58,200
have the same thing that we had

1532
01:05:54,039 --> 01:06:01,200
above so it it looks like this but over

1533
01:05:58,200 --> 01:06:03,720
here we have a clipped version of this

1534
01:06:01,200 --> 01:06:07,000
where essentially what we do is we

1535
01:06:03,720 --> 01:06:07,000
clip this

1536
01:06:21,119 --> 01:06:27,880
ratio this ratio to be within uh a

1537
01:06:24,720 --> 01:06:32,160
certain range of the original ratio and

1538
01:06:27,880 --> 01:06:37,880
what this is doing is this is

1539
01:06:32,160 --> 01:06:41,400
essentially forcing the model to um not

1540
01:06:37,880 --> 01:06:44,000
reward large jumps in the space um

1541
01:06:41,400 --> 01:06:47,559
because if you take the

1542
01:06:44,000 --> 01:06:49,160
minimum and actually I'm I'm sorry I

1543
01:06:47,559 --> 01:06:50,720
just realized I I might have done

1544
01:06:49,160 --> 01:06:52,520
something confusing here because this is

1545
01:06:50,720 --> 01:06:53,960
actually higher as better so this isn't

1546
01:06:52,520 --> 01:06:56,079
really a loss function this is something

1547
01:06:53,960 --> 01:06:57,680
you're attempting to maximize so

1548
01:06:56,079 --> 01:06:59,839
in contrast to all of the other things I

1549
01:06:57,680 --> 01:07:01,680
was talking about before um this is

1550
01:06:59,839 --> 01:07:04,400
something where higher is better instead

1551
01:07:01,680 --> 01:07:07,599
of lower is better but anyway basically

1552
01:07:04,400 --> 01:07:09,599
by taking the minimum of this you're

1553
01:07:07,599 --> 01:07:11,960
encouraging the model

1554
01:07:09,599 --> 01:07:16,279
to

1555
01:07:11,960 --> 01:07:18,559
uh keep examining the space where you

1556
01:07:16,279 --> 01:07:20,799
don't diverge much from the original

1557
01:07:18,559 --> 01:07:22,920
model and if the space where the

1558
01:07:20,799 --> 01:07:25,240
original model was in is better than the

1559
01:07:22,920 --> 01:07:27,440
new space that your model has moved into

1560
01:07:25,240 --> 01:07:30,920
you move back towards the original model

1561
01:07:27,440 --> 01:07:33,000
so basically like if you had um if you

1562
01:07:30,920 --> 01:07:34,960
learned a model if you started learning

1563
01:07:33,000 --> 01:07:37,960
a model that looked like it was

1564
01:07:34,960 --> 01:07:40,279
optimizing uh your your reward but then

1565
01:07:37,960 --> 01:07:43,119
suddenly the model went off the rails

1566
01:07:40,279 --> 01:07:45,000
and um it starts generating completely

1567
01:07:43,119 --> 01:07:47,319
nonsense outputs that get really bad

1568
01:07:45,000 --> 01:07:49,119
reward this will push it back towards

1569
01:07:47,319 --> 01:07:50,920
the original policy and that's the basic

1570
01:07:49,119 --> 01:07:54,279
idea behind

1571
01:07:50,920 --> 01:07:57,640
P um in terms of what I see people using

1572
01:07:54,279 --> 01:07:59,799
um po was like really really popular for

1573
01:07:57,640 --> 01:08:01,880
a while but I've started to see people

1574
01:07:59,799 --> 01:08:04,799
use alternative strategies that use K

1575
01:08:01,880 --> 01:08:06,880
regularization so I don't I don't think

1576
01:08:04,799 --> 01:08:08,520
either one of them is like particularly

1577
01:08:06,880 --> 01:08:10,039
more popular than any of the others and

1578
01:08:08,520 --> 01:08:13,720
this one's a little bit simpler

1579
01:08:10,039 --> 01:08:13,720
conceptually so I like the the

1580
01:08:14,880 --> 01:08:19,279
one cool um any questions about

1581
01:08:20,359 --> 01:08:26,759
this okay um and actually one thing I

1582
01:08:24,640 --> 01:08:29,679
should mention is um all of these things

1583
01:08:26,759 --> 01:08:32,120
are implemented uh in you know whatever

1584
01:08:29,679 --> 01:08:33,759
libraries you use like hugging face TRL

1585
01:08:32,120 --> 01:08:35,679
Transformer reinforcement learning as an

1586
01:08:33,759 --> 01:08:37,040
example Library all of these methods are

1587
01:08:35,679 --> 01:08:38,400
implemented there so if you actually

1588
01:08:37,040 --> 01:08:40,600
want to use these in practice that's

1589
01:08:38,400 --> 01:08:40,600
good

1590
01:08:40,839 --> 01:08:46,359
place the next thing is adding a

1591
01:08:42,920 --> 01:08:48,679
Baseline and so the basic idea is that

1592
01:08:46,359 --> 01:08:52,199
you have ex expectations about your

1593
01:08:48,679 --> 01:08:54,640
reward for a particular sentence and um

1594
01:08:52,199 --> 01:08:56,560
like let's say we wanted to uh translate

1595
01:08:54,640 --> 01:08:58,400
a sentence and we have uh something like

1596
01:08:56,560 --> 01:09:01,279
this is an easy sentence and buffalo

1597
01:08:58,400 --> 01:09:02,920
buffalo buffalo which is a harder

1598
01:09:01,279 --> 01:09:07,799
sentence to

1599
01:09:02,920 --> 01:09:09,679
translate and so we have a reward um if

1600
01:09:07,799 --> 01:09:11,759
if you're not familiar with this example

1601
01:09:09,679 --> 01:09:13,480
you can search on Wikipedia for buffalo

1602
01:09:11,759 --> 01:09:16,759
buffalo buffalo and you'll you'll find

1603
01:09:13,480 --> 01:09:19,520
out what I'm talking about um but uh

1604
01:09:16,759 --> 01:09:21,440
there's a reward uh and let's say you

1605
01:09:19,520 --> 01:09:24,359
got a reward of 0.8 for the first one

1606
01:09:21,440 --> 01:09:29,679
and a reward of 0.3 for the second

1607
01:09:24,359 --> 01:09:31,679
one but the problem is if um the first

1608
01:09:29,679 --> 01:09:33,640
one actually is really easy and the

1609
01:09:31,679 --> 01:09:36,120
second one is really hard getting a

1610
01:09:33,640 --> 01:09:37,799
reward of 0.8 for the second one for

1611
01:09:36,120 --> 01:09:40,080
like a translation or something is

1612
01:09:37,799 --> 01:09:41,120
actually bad right and a reward of 0.3

1613
01:09:40,080 --> 01:09:45,239
is good because you're moving in the

1614
01:09:41,120 --> 01:09:49,359
right direction and so you basically um

1615
01:09:45,239 --> 01:09:52,239
you have uh the Baseline uh minus reward

1616
01:09:49,359 --> 01:09:54,960
or sorry reward minus Baseline and this

1617
01:09:52,239 --> 01:09:56,520
would give you a negative value for this

1618
01:09:54,960 --> 01:09:59,320
first one a positive value for the

1619
01:09:56,520 --> 01:10:01,360
second one and so the basic idea is can

1620
01:09:59,320 --> 01:10:04,400
we predict a priori how difficult this

1621
01:10:01,360 --> 01:10:05,440
example is and then uh adjust our reward

1622
01:10:04,400 --> 01:10:08,360
based on

1623
01:10:05,440 --> 01:10:10,960
that and

1624
01:10:08,360 --> 01:10:13,679
so that's the basic idea you just have

1625
01:10:10,960 --> 01:10:15,560
kind of like a baseline model um you

1626
01:10:13,679 --> 01:10:19,320
have a baseline model that predicts this

1627
01:10:15,560 --> 01:10:19,320
and uh you adjust uh

1628
01:10:19,760 --> 01:10:25,000
appropriately um there's two major ways

1629
01:10:22,719 --> 01:10:27,600
you can do this the first one um the

1630
01:10:25,000 --> 01:10:29,800
Baseline doesn't need to be anything um

1631
01:10:27,600 --> 01:10:32,960
the only hope is that it decreases the

1632
01:10:29,800 --> 01:10:35,960
variance in your reward uh and makes

1633
01:10:32,960 --> 01:10:38,239
learning more stable um there's two

1634
01:10:35,960 --> 01:10:40,159
options that I see done pretty widely

1635
01:10:38,239 --> 01:10:43,000
the first one is predicting the final

1636
01:10:40,159 --> 01:10:47,360
reward um predicting the final reward

1637
01:10:43,000 --> 01:10:50,960
using a model that doesn't look at

1638
01:10:47,360 --> 01:10:53,400
all at the answer that you provided it

1639
01:10:50,960 --> 01:10:55,880
only looks at the input or it only looks

1640
01:10:53,400 --> 01:10:58,840
at the intermediate States of uh you

1641
01:10:55,880 --> 01:11:00,480
know a model or something and so at the

1642
01:10:58,840 --> 01:11:03,280
sentence level you can have one Baseline

1643
01:11:00,480 --> 01:11:04,719
per sentence um you can also do it at

1644
01:11:03,280 --> 01:11:10,560
each decoder

1645
01:11:04,719 --> 01:11:11,640
State and this is uh basically you can

1646
01:11:10,560 --> 01:11:13,040
do this anytime you're doing

1647
01:11:11,640 --> 01:11:15,199
reinforcement learning by just training

1648
01:11:13,040 --> 01:11:18,199
a regression model that does this for

1649
01:11:15,199 --> 01:11:19,679
you based on the rewards you get the

1650
01:11:18,199 --> 01:11:21,040
important thing is the Baseline is not

1651
01:11:19,679 --> 01:11:22,640
allowed to use any of your actual

1652
01:11:21,040 --> 01:11:25,679
predictions because once you start using

1653
01:11:22,640 --> 01:11:26,640
the predictions then um your uh it's not

1654
01:11:25,679 --> 01:11:28,679
a

1655
01:11:26,640 --> 01:11:30,840
baseline another option which is

1656
01:11:28,679 --> 01:11:33,440
relatively easy to implement but can

1657
01:11:30,840 --> 01:11:36,320
still be effective is you calculate the

1658
01:11:33,440 --> 01:11:38,719
mean of the rewards in a batch and so if

1659
01:11:36,320 --> 01:11:40,880
you have a big batch of data and your

1660
01:11:38,719 --> 01:11:44,440
average reward in the batch is like

1661
01:11:40,880 --> 01:11:46,480
0.4 uh then you just subtract that 0.4

1662
01:11:44,440 --> 01:11:50,080
uh and calculate your reward based on

1663
01:11:46,480 --> 01:11:50,080
that so that's another option that can

1664
01:11:51,800 --> 01:11:57,800
use

1665
01:11:53,639 --> 01:12:00,000
um a kind of extreme example of this uh

1666
01:11:57,800 --> 01:12:01,199
of creating a baseline is contrasting

1667
01:12:00,000 --> 01:12:03,639
pairwise

1668
01:12:01,199 --> 01:12:05,880
examples um or

1669
01:12:03,639 --> 01:12:08,280
contrasting different outputs for the

1670
01:12:05,880 --> 01:12:12,040
same input

1671
01:12:08,280 --> 01:12:13,920
and you can easily learn uh directly

1672
01:12:12,040 --> 01:12:16,239
from pairwise Human

1673
01:12:13,920 --> 01:12:18,199
preferences uh which can provide more

1674
01:12:16,239 --> 01:12:20,760
stability because you know one is better

1675
01:12:18,199 --> 01:12:23,880
than the other so you essentially can be

1676
01:12:20,760 --> 01:12:26,199
sure that uh you're upweighting a better

1677
01:12:23,880 --> 01:12:29,560
one and down weting a worse one

1678
01:12:26,199 --> 01:12:31,400
um this is the idea behind DPO which is

1679
01:12:29,560 --> 01:12:33,719
a recently pretty popular model but

1680
01:12:31,400 --> 01:12:36,800
there's also other previous methods that

1681
01:12:33,719 --> 01:12:40,199
did similar things and the way DPO works

1682
01:12:36,800 --> 01:12:45,040
is it basically calculates this ratio of

1683
01:12:40,199 --> 01:12:49,280
uh the probability of the new uh the new

1684
01:12:45,040 --> 01:12:51,639
model to the old model but it UPS this

1685
01:12:49,280 --> 01:12:53,639
probability for a good output and it

1686
01:12:51,639 --> 01:12:56,280
downweights this probability for a bad

1687
01:12:53,639 --> 01:12:57,679
output and so

1688
01:12:56,280 --> 01:13:00,120
here we have our better outputs over

1689
01:12:57,679 --> 01:13:02,040
here here we have our worse outputs and

1690
01:13:00,120 --> 01:13:03,600
you just it's basically learning to

1691
01:13:02,040 --> 01:13:05,639
upate the probability and downweight

1692
01:13:03,600 --> 01:13:09,320
probability

1693
01:13:05,639 --> 01:13:09,320
accordingly so

1694
01:13:09,360 --> 01:13:15,040
um you can notice that DPO is very

1695
01:13:12,280 --> 01:13:18,040
similar to PO um and that it's learning

1696
01:13:15,040 --> 01:13:19,679
uh it's using these ratios but the

1697
01:13:18,040 --> 01:13:21,520
disadvantage of this is you obviously

1698
01:13:19,679 --> 01:13:23,120
require pairwise judgments and you can't

1699
01:13:21,520 --> 01:13:26,120
learn a model if you don't have these

1700
01:13:23,120 --> 01:13:28,080
pawise judgments so

1701
01:13:26,120 --> 01:13:30,760
the

1702
01:13:28,080 --> 01:13:33,159
beta yeah so the beta term is is

1703
01:13:30,760 --> 01:13:35,840
basically a normalization term it's a

1704
01:13:33,159 --> 01:13:39,960
hyper parameter um

1705
01:13:35,840 --> 01:13:41,840
for DPO sorry I read the paper right

1706
01:13:39,960 --> 01:13:43,639
when it came out and I don't remember if

1707
01:13:41,840 --> 01:13:45,600
it's a direct derivation from the K

1708
01:13:43,639 --> 01:13:47,960
Divergence term or not but I think it

1709
01:13:45,600 --> 01:13:49,800
might be um I'd have to go back and look

1710
01:13:47,960 --> 01:13:50,480
at the look at the paper but basically

1711
01:13:49,800 --> 01:13:53,600
the

1712
01:13:50,480 --> 01:13:56,760
more the larger this is the larger

1713
01:13:53,600 --> 01:13:59,320
gradient steps you'll be

1714
01:13:56,760 --> 01:14:00,639
it also um like you'll notice there

1715
01:13:59,320 --> 01:14:03,400
sorry I didn't mention this but you'll

1716
01:14:00,639 --> 01:14:06,120
notice there's a sigmoid term here so

1717
01:14:03,400 --> 01:14:09,000
the the

1718
01:14:06,120 --> 01:14:10,080
beta the larger you increase the beta

1719
01:14:09,000 --> 01:14:13,239
the

1720
01:14:10,080 --> 01:14:16,600
more small differences in these

1721
01:14:13,239 --> 01:14:18,719
values like it basically like stretches

1722
01:14:16,600 --> 01:14:22,280
or shrinks the sigmoid with respect to

1723
01:14:18,719 --> 01:14:24,120
how beak the it is so it will um it will

1724
01:14:22,280 --> 01:14:25,800
affect how much like small differences

1725
01:14:24,120 --> 01:14:27,960
in this will affect

1726
01:14:25,800 --> 01:14:30,120
but I I think this was derived from the

1727
01:14:27,960 --> 01:14:31,760
K regularization term that we had

1728
01:14:30,120 --> 01:14:34,400
previously in

1729
01:14:31,760 --> 01:14:35,800
um in this slide here but I have to go

1730
01:14:34,400 --> 01:14:40,520
back and double check unless somebody

1731
01:14:35,800 --> 01:14:43,239
knows it is okay good yeah

1732
01:14:40,520 --> 01:14:45,000
so I don't want to say wrong things but

1733
01:14:43,239 --> 01:14:48,239
I also don't want

1734
01:14:45,000 --> 01:14:50,920
to okay cool um and so then increasing

1735
01:14:48,239 --> 01:14:55,080
batch size

1736
01:14:50,920 --> 01:14:57,360
um because each uh another thing is um

1737
01:14:55,080 --> 01:14:58,440
kind of NE necessarily reinforcement

1738
01:14:57,360 --> 01:14:59,920
learning is going to have higher

1739
01:14:58,440 --> 01:15:01,400
variance and maximum likelihood

1740
01:14:59,920 --> 01:15:04,199
estimation just because we're doing samp

1741
01:15:01,400 --> 01:15:07,840
playing and other things like this and

1742
01:15:04,199 --> 01:15:09,440
um so one very simple thing you can do

1743
01:15:07,840 --> 01:15:11,280
is just increase the number of examples

1744
01:15:09,440 --> 01:15:13,679
or rollouts that you do before an update

1745
01:15:11,280 --> 01:15:15,800
to stabilize and so I I would definitely

1746
01:15:13,679 --> 01:15:17,480
suggest that if you're seeing any

1747
01:15:15,800 --> 01:15:18,679
stability after doing all of the tricks

1748
01:15:17,480 --> 01:15:20,400
that I mentioned before that you

1749
01:15:18,679 --> 01:15:23,040
increase your batch size and often that

1750
01:15:20,400 --> 01:15:25,480
can just resolve your problems

1751
01:15:23,040 --> 01:15:28,760
um another uh

1752
01:15:25,480 --> 01:15:30,560
thing that people often do is um save

1753
01:15:28,760 --> 01:15:32,040
many many previous rollouts because

1754
01:15:30,560 --> 01:15:34,199
generally doing rollouts is more

1755
01:15:32,040 --> 01:15:37,840
expensive doing rollouts and collecting

1756
01:15:34,199 --> 01:15:39,560
rewards is more expensive and so um you

1757
01:15:37,840 --> 01:15:42,360
can save the roll outs that you have

1758
01:15:39,560 --> 01:15:43,840
done before and uh keep them around so

1759
01:15:42,360 --> 01:15:46,600
you can update parameters with larger

1760
01:15:43,840 --> 01:15:50,800
batches in a more efficient

1761
01:15:46,600 --> 01:15:53,120
way cool so that's all I have uh I just

1762
01:15:50,800 --> 01:15:54,400
realized we're exactly at time so uh I

1763
01:15:53,120 --> 01:15:56,440
should finish up here but I'll be happy

1764
01:15:54,400 --> 01:15:59,440
to take any

1765
01:15:56,440 --> 01:15:59,440
for

1766
01:16:01,679 --> 01:16:04,679
thanks