1
00:00:00,399 --> 00:00:04,720
great um yeah so today we're going to be

2
00:00:03,320 --> 00:00:07,040
talking a little bit about generation

3
00:00:04,720 --> 00:00:08,639
algorithms um this will be sort of a

4
00:00:07,040 --> 00:00:10,160
tour through some of the most common

5
00:00:08,639 --> 00:00:12,080
methods and we're going to talk a little

6
00:00:10,160 --> 00:00:13,480
bit about the theory behind them as well

7
00:00:12,080 --> 00:00:15,080
um if you're looking at the slides on

8
00:00:13,480 --> 00:00:18,359
the website these might be ever so

9
00:00:15,080 --> 00:00:20,000
slightly different um but yeah I'll try

10
00:00:18,359 --> 00:00:21,640
to stop at each section boundary for

11
00:00:20,000 --> 00:00:23,840
questions also feel free to sort of

12
00:00:21,640 --> 00:00:25,720
interrupt at any point for

13
00:00:23,840 --> 00:00:27,720
clarifications so we're starting off

14
00:00:25,720 --> 00:00:29,560
today with some great news um let's say

15
00:00:27,720 --> 00:00:31,199
that you have some friend who maybe owns

16
00:00:29,560 --> 00:00:34,800
a giant tech company and they've gifted

17
00:00:31,199 --> 00:00:36,480
you this absolutely massive new model M

18
00:00:34,800 --> 00:00:38,079
um it's a great model it's pre-trained

19
00:00:36,480 --> 00:00:40,879
with the latest architecture it's

20
00:00:38,079 --> 00:00:42,920
pre-trained on um trillions of tokens of

21
00:00:40,879 --> 00:00:44,520
text it's got seven billion parameters

22
00:00:42,920 --> 00:00:46,399
it looks like a really promising new

23
00:00:44,520 --> 00:00:48,399
model you know it's the top of all these

24
00:00:46,399 --> 00:00:50,320
leaderboards um but if you actually take

25
00:00:48,399 --> 00:00:52,520
your new model M and you sort of open up

26
00:00:50,320 --> 00:00:53,719
this box and kind of Shake It Out maybe

27
00:00:52,520 --> 00:00:55,239
from last class you know a little bit

28
00:00:53,719 --> 00:00:57,000
architecturally what this model might

29
00:00:55,239 --> 00:00:58,239
look like but if you actually kind of

30
00:00:57,000 --> 00:01:00,320
take a closer look at it from a

31
00:00:58,239 --> 00:01:01,719
different angle what you see is that m

32
00:01:00,320 --> 00:01:04,920
is actually just a conditional

33
00:01:01,719 --> 00:01:07,200
probability distribution um you put some

34
00:01:04,920 --> 00:01:09,680
input X into your model and you get some

35
00:01:07,200 --> 00:01:10,680
probability out for any given sequence

36
00:01:09,680 --> 00:01:13,360
that you're sort of interested in

37
00:01:10,680 --> 00:01:14,960
evaluating right um and in particular M

38
00:01:13,360 --> 00:01:17,560
gives you a probability distribution

39
00:01:14,960 --> 00:01:19,439
over all tokens in its vocabulary to

40
00:01:17,560 --> 00:01:21,040
predict like what token you would output

41
00:01:19,439 --> 00:01:24,840
next right and so this is what this

42
00:01:21,040 --> 00:01:26,880
equation says um given some input X and

43
00:01:24,840 --> 00:01:29,520
everything that you've predicted so far

44
00:01:26,880 --> 00:01:32,399
you get the probability of the next

45
00:01:29,520 --> 00:01:33,600
token in YJ and if you multiply this out

46
00:01:32,399 --> 00:01:34,840
over all the probabilities in your

47
00:01:33,600 --> 00:01:37,159
sequence you can calculate the

48
00:01:34,840 --> 00:01:41,240
probability of any output y given your

49
00:01:37,159 --> 00:01:42,640
input X so what this like super fancy

50
00:01:41,240 --> 00:01:44,119
model that you spend a lot of money to

51
00:01:42,640 --> 00:01:46,280
train is really just a conditional

52
00:01:44,119 --> 00:01:47,920
probability distribution um but this

53
00:01:46,280 --> 00:01:49,600
turns out to be okay because you can use

54
00:01:47,920 --> 00:01:51,920
a conditional probability distribution

55
00:01:49,600 --> 00:01:54,399
to do sort of any task that we're really

56
00:01:51,920 --> 00:01:56,719
interested in in NLP um pretty much any

57
00:01:54,399 --> 00:01:58,680
task right so by changing what you

58
00:01:56,719 --> 00:02:01,360
consider your input X and your output y

59
00:01:58,680 --> 00:02:03,560
to be you can can get outputs from this

60
00:02:01,360 --> 00:02:06,479
model for things like translation for

61
00:02:03,560 --> 00:02:08,720
summarization for reasoning Tas um just

62
00:02:06,479 --> 00:02:10,520
by sort of changing what you consider

63
00:02:08,720 --> 00:02:12,760
your inputs and outputs in this

64
00:02:10,520 --> 00:02:14,239
setting but there's sort of both good

65
00:02:12,760 --> 00:02:15,920
and bad things about your model being a

66
00:02:14,239 --> 00:02:17,120
probability distribution instead of just

67
00:02:15,920 --> 00:02:20,599
an oracle that gives you sort of a

68
00:02:17,120 --> 00:02:22,080
single answer for every input um one

69
00:02:20,599 --> 00:02:24,480
kind of nice thing about this

70
00:02:22,080 --> 00:02:26,080
distribution um is that you can get at

71
00:02:24,480 --> 00:02:27,720
an idea of something like confidence

72
00:02:26,080 --> 00:02:30,120
right if you give your model the input 2

73
00:02:27,720 --> 00:02:32,480
plus 2 equals and almost all the

74
00:02:30,120 --> 00:02:34,200
probability mass is on the token of four

75
00:02:32,480 --> 00:02:35,760
you can say like the model predicts with

76
00:02:34,200 --> 00:02:38,319
pretty high confidence that 2 plus 2

77
00:02:35,760 --> 00:02:39,480
equals four um versus if you give it

78
00:02:38,319 --> 00:02:40,959
something that's maybe a little more

79
00:02:39,480 --> 00:02:43,120
open-ended like you ask it to predict

80
00:02:40,959 --> 00:02:44,640
Graham's favorite color and you see this

81
00:02:43,120 --> 00:02:47,040
distribution that's sort of a lot

82
00:02:44,640 --> 00:02:48,440
flatter you know the most likely output

83
00:02:47,040 --> 00:02:49,720
is green but maybe we don't have a lot

84
00:02:48,440 --> 00:02:51,560
of confidence that that's the correct

85
00:02:49,720 --> 00:02:53,040
answer um this is really closely tied

86
00:02:51,560 --> 00:02:55,200
into the idea of calibration which you

87
00:02:53,040 --> 00:02:58,879
guys talked about um I guess a couple of

88
00:02:55,200 --> 00:03:00,640
classes ago now the flip side of this

89
00:02:58,879 --> 00:03:03,680
though is that you know Noti that for

90
00:03:00,640 --> 00:03:06,760
this case like 2 plus 2al 4 not all of

91
00:03:03,680 --> 00:03:08,519
the probability mass is on four um and

92
00:03:06,760 --> 00:03:09,720
so models that are conditional

93
00:03:08,519 --> 00:03:11,560
probability distributions can

94
00:03:09,720 --> 00:03:13,560
hallucinate right um pretty much no

95
00:03:11,560 --> 00:03:15,799
matter what you do there's going to be

96
00:03:13,560 --> 00:03:17,680
some nonzero probability to some output

97
00:03:15,799 --> 00:03:19,920
that's incorrect or

98
00:03:17,680 --> 00:03:21,239
undesirable um in some cases maybe even

99
00:03:19,920 --> 00:03:23,760
offensive something that you don't want

100
00:03:21,239 --> 00:03:25,280
the model to Output um and this is sort

101
00:03:23,760 --> 00:03:27,840
of an artifact of the way these models

102
00:03:25,280 --> 00:03:29,280
are trained if there's some great work

103
00:03:27,840 --> 00:03:31,400
kind of more on the theory side here

104
00:03:29,280 --> 00:03:32,840
that shows that this is actually true

105
00:03:31,400 --> 00:03:35,120
even if everything in your input

106
00:03:32,840 --> 00:03:36,920
training data is sort of correct and

107
00:03:35,120 --> 00:03:38,439
factual and doesn't have any errors

108
00:03:36,920 --> 00:03:41,200
you'll still wind up with a situation

109
00:03:38,439 --> 00:03:44,480
where some nonzero probability mass is

110
00:03:41,200 --> 00:03:47,000
on some outputs that are undesirable or

111
00:03:44,480 --> 00:03:50,120
hallucinatory for sort of most inputs

112
00:03:47,000 --> 00:03:52,159
that you care about evaluating so if we

113
00:03:50,120 --> 00:03:55,079
have these issues how do we actually get

114
00:03:52,159 --> 00:03:56,519
a good output out of the model um and to

115
00:03:55,079 --> 00:03:58,640
do that we're first going to talk about

116
00:03:56,519 --> 00:04:00,079
some sampling methods um but I want to

117
00:03:58,640 --> 00:04:01,879
pause here in case there are of any

118
00:04:00,079 --> 00:04:04,159
questions on this idea of a model is a

119
00:04:01,879 --> 00:04:04,159
conditional

120
00:04:05,040 --> 00:04:11,680
distribution great so we can jump right

121
00:04:07,519 --> 00:04:13,560
in so we have this model right we know

122
00:04:11,680 --> 00:04:15,959
at each step at each token we might want

123
00:04:13,560 --> 00:04:17,919
to decode the distribution of likelihood

124
00:04:15,959 --> 00:04:18,959
over all vocabulary tokens right this

125
00:04:17,919 --> 00:04:21,680
conditional distribution we've been

126
00:04:18,959 --> 00:04:24,240
talking about um for the next time step

127
00:04:21,680 --> 00:04:26,400
and what we want out of this is a good

128
00:04:24,240 --> 00:04:28,000
output um for some definition of good

129
00:04:26,400 --> 00:04:30,919
that we can sort of develop as we go

130
00:04:28,000 --> 00:04:32,479
here so maybe the natural first thing to

131
00:04:30,919 --> 00:04:34,880
try is we have a probability

132
00:04:32,479 --> 00:04:36,600
distribution can we just sample from it

133
00:04:34,880 --> 00:04:39,600
right and this is something called

134
00:04:36,600 --> 00:04:41,639
ancestral sampling so at each time step

135
00:04:39,600 --> 00:04:43,560
we're going to draw a token from this

136
00:04:41,639 --> 00:04:45,039
distribution sort of according to its

137
00:04:43,560 --> 00:04:47,199
relative probability right so if

138
00:04:45,039 --> 00:04:48,639
something has twice as much probability

139
00:04:47,199 --> 00:04:51,280
Mass according to the model we'll draw

140
00:04:48,639 --> 00:04:54,000
it twice as often um and we can sample

141
00:04:51,280 --> 00:04:55,560
from this distribution at each time step

142
00:04:54,000 --> 00:04:58,080
and this is sort of this is sort of a

143
00:04:55,560 --> 00:05:00,199
nice setup um we get exact samples from

144
00:04:58,080 --> 00:05:02,639
the model distribution so using the

145
00:05:00,199 --> 00:05:04,479
setup if you can you imagine like

146
00:05:02,639 --> 00:05:06,680
drawing an almost infinite number of

147
00:05:04,479 --> 00:05:08,320
samples like a ridiculously large number

148
00:05:06,680 --> 00:05:10,160
and you look at their probabilities

149
00:05:08,320 --> 00:05:11,840
you'd sort of get something from this

150
00:05:10,160 --> 00:05:13,039
distribution with exactly the

151
00:05:11,840 --> 00:05:15,720
probability that the real model

152
00:05:13,039 --> 00:05:17,280
distribution is given you um so this is

153
00:05:15,720 --> 00:05:19,039
great this gives us an exact sample from

154
00:05:17,280 --> 00:05:21,400
the model this seems to be exactly what

155
00:05:19,039 --> 00:05:22,880
we want um but you can guess probably by

156
00:05:21,400 --> 00:05:24,639
the fact that we're only like 10 minutes

157
00:05:22,880 --> 00:05:27,000
into class here this is not really the

158
00:05:24,639 --> 00:05:28,280
end of the story um and there's actually

159
00:05:27,000 --> 00:05:30,800
a couple of problems with sampling

160
00:05:28,280 --> 00:05:32,560
directly from our model distribu

161
00:05:30,800 --> 00:05:35,280
the one that we're really going to focus

162
00:05:32,560 --> 00:05:37,919
on first here is this idea of a long

163
00:05:35,280 --> 00:05:41,400
tail so a model like llama and maybe our

164
00:05:37,919 --> 00:05:43,639
new model M um has 32,000 vocabulary

165
00:05:41,400 --> 00:05:46,280
tokens and you can imagine maybe out of

166
00:05:43,639 --> 00:05:48,000
those tokens there might be one or even

167
00:05:46,280 --> 00:05:49,720
2,000 of those tokens that are sort of a

168
00:05:48,000 --> 00:05:51,919
reasonable next thing to predict for a

169
00:05:49,720 --> 00:05:53,479
really open-ended task right but there's

170
00:05:51,919 --> 00:05:55,440
going to be all kinds of things in that

171
00:05:53,479 --> 00:05:57,039
distribution um that are maybe like

172
00:05:55,440 --> 00:05:58,440
punctuation there maybe tokens that

173
00:05:57,039 --> 00:06:00,280
won't actually lead to the correct

174
00:05:58,440 --> 00:06:01,840
answer like there's a lot of things in

175
00:06:00,280 --> 00:06:04,560
this distribution that would be all

176
00:06:01,840 --> 00:06:06,160
really low likelihood and this is fine

177
00:06:04,560 --> 00:06:08,759
these things just get low probability

178
00:06:06,160 --> 00:06:11,039
Mass but the problem is if you give sort

179
00:06:08,759 --> 00:06:13,639
of a small amount of probability Mass to

180
00:06:11,039 --> 00:06:16,599
30,000 different things that mass will

181
00:06:13,639 --> 00:06:19,360
add up pretty quickly um and to see this

182
00:06:16,599 --> 00:06:20,360
we have sort of this illustration here

183
00:06:19,360 --> 00:06:21,560
um I don't know if you can see the

184
00:06:20,360 --> 00:06:23,280
difference between the green and the

185
00:06:21,560 --> 00:06:25,720
yellow but I've also drawn a little bar

186
00:06:23,280 --> 00:06:27,800
between them this is a really longtailed

187
00:06:25,720 --> 00:06:29,720
distribution and the green part of the

188
00:06:27,800 --> 00:06:31,960
distribution which is a lot of tokens

189
00:06:29,720 --> 00:06:34,000
with high likelihood has 50% of the

190
00:06:31,960 --> 00:06:35,560
total probability the Yellow Part which

191
00:06:34,000 --> 00:06:37,360
is all a lot of things that are all

192
00:06:35,560 --> 00:06:40,280
individually not super likely is the

193
00:06:37,360 --> 00:06:41,720
other 50% of the probability and so what

194
00:06:40,280 --> 00:06:44,360
that means is if you're doing something

195
00:06:41,720 --> 00:06:46,120
like ancestral sampling 50% of the time

196
00:06:44,360 --> 00:06:49,160
you'll be sampling something really

197
00:06:46,120 --> 00:06:51,520
unlikely from this long tail um that

198
00:06:49,160 --> 00:06:53,759
seems sort of not like what we want

199
00:06:51,520 --> 00:06:56,080
right um so is there anything we can do

200
00:06:53,759 --> 00:06:58,080
about this and the obvious for solution

201
00:06:56,080 --> 00:06:59,400
here is can we just cut off that tail

202
00:06:58,080 --> 00:07:01,680
like if we know these tokens are not

203
00:06:59,400 --> 00:07:03,039
super likely can we just ignore them and

204
00:07:01,680 --> 00:07:05,039
there's a couple of different ways to do

205
00:07:03,039 --> 00:07:07,919
that um the first of these is something

206
00:07:05,039 --> 00:07:10,080
called topk sampling where we say okay

207
00:07:07,919 --> 00:07:12,479
you know maybe we think there are 10

208
00:07:10,080 --> 00:07:14,000
reasonable like outputs is right maybe

209
00:07:12,479 --> 00:07:17,280
we'll just sample from the 10 most

210
00:07:14,000 --> 00:07:19,759
probable tokens um here maybe we say if

211
00:07:17,280 --> 00:07:21,479
we want to pick top six sampling we'll

212
00:07:19,759 --> 00:07:23,919
sample from just the six most probable

213
00:07:21,479 --> 00:07:26,240
tokens and so in this example you can

214
00:07:23,919 --> 00:07:27,680
see we originally had 10 tokens and

215
00:07:26,240 --> 00:07:30,560
we're going to sample from just the blue

216
00:07:27,680 --> 00:07:32,919
ones just the six most likely tokens

217
00:07:30,560 --> 00:07:34,360
um in this example this distribution is

218
00:07:32,919 --> 00:07:37,280
pretty flat there's a lot of things that

219
00:07:34,360 --> 00:07:40,120
are like kind of likely right so that

220
00:07:37,280 --> 00:07:43,000
those six tokens are only 68% of the

221
00:07:40,120 --> 00:07:45,360
total probability Mass um if we go like

222
00:07:43,000 --> 00:07:47,240
one time step further here we might have

223
00:07:45,360 --> 00:07:49,360
a distribution that's a lot peier most

224
00:07:47,240 --> 00:07:51,759
of the mass is on just a single token

225
00:07:49,360 --> 00:07:53,919
and so sampling from just the top six

226
00:07:51,759 --> 00:07:56,400
tokens actually captures 99% of the

227
00:07:53,919 --> 00:07:58,360
probability mes maybe we say that seems

228
00:07:56,400 --> 00:08:01,199
a little excessive right we don't really

229
00:07:58,360 --> 00:08:03,400
need um maybe all of these tokens that

230
00:08:01,199 --> 00:08:05,479
are all kind of low probability maybe we

231
00:08:03,400 --> 00:08:07,000
just want to sort of sample from the top

232
00:08:05,479 --> 00:08:08,080
half of our distribution or something or

233
00:08:07,000 --> 00:08:10,840
the top

234
00:08:08,080 --> 00:08:12,919
90% um so instead of choosing a top

235
00:08:10,840 --> 00:08:15,560
number of tokens to sample from you

236
00:08:12,919 --> 00:08:17,400
could choose a top amount of probability

237
00:08:15,560 --> 00:08:20,000
and this is something called top P or

238
00:08:17,400 --> 00:08:21,520
nucleus sampling so P here is the amount

239
00:08:20,000 --> 00:08:24,039
of probability from your distribution

240
00:08:21,520 --> 00:08:26,639
you want to consider so if you decide

241
00:08:24,039 --> 00:08:29,280
your p is about like 94% of the

242
00:08:26,639 --> 00:08:31,639
probability Mass you in this first examp

243
00:08:29,280 --> 00:08:33,719
example here would choose almost all of

244
00:08:31,639 --> 00:08:35,440
the tokens you keep adding tokens in

245
00:08:33,719 --> 00:08:37,159
until you reach an amount of total

246
00:08:35,440 --> 00:08:39,479
probability that's about

247
00:08:37,159 --> 00:08:40,880
094 but then when you get to the Second

248
00:08:39,479 --> 00:08:43,240
Step where you have a couple of really

249
00:08:40,880 --> 00:08:45,959
highly probable tokens you'd only need a

250
00:08:43,240 --> 00:08:47,959
couple of tokens to add up to 094 or

251
00:08:45,959 --> 00:08:50,320
even higher than 0.94 and so you would

252
00:08:47,959 --> 00:08:52,200
just sample from a smaller set of tokens

253
00:08:50,320 --> 00:08:54,600
so in top K sampling the total amount of

254
00:08:52,200 --> 00:08:56,560
probability your sampling from can move

255
00:08:54,600 --> 00:08:58,120
around in top P sampling the total

256
00:08:56,560 --> 00:08:59,839
number of tokens you're sampling from

257
00:08:58,120 --> 00:09:01,959
might change

258
00:08:59,839 --> 00:09:04,760
um but maybe we sort of don't want to

259
00:09:01,959 --> 00:09:07,279
impose a strong constraint like we want

260
00:09:04,760 --> 00:09:09,279
like 94% here maybe just what we really

261
00:09:07,279 --> 00:09:11,040
care about is saying that we're not

262
00:09:09,279 --> 00:09:14,000
going to sample anything that's really

263
00:09:11,040 --> 00:09:16,800
really unlikely right another way of

264
00:09:14,000 --> 00:09:18,560
doing this is called Epsilon sampling

265
00:09:16,800 --> 00:09:20,519
where we just sample tokens that have at

266
00:09:18,560 --> 00:09:22,920
least some minimum amount of probability

267
00:09:20,519 --> 00:09:24,720
to them right so maybe we just want

268
00:09:22,920 --> 00:09:29,519
tokens that have probability of at least

269
00:09:24,720 --> 00:09:31,240
0.05 here um in this first um example

270
00:09:29,519 --> 00:09:32,640
everything has at least some reasonable

271
00:09:31,240 --> 00:09:34,240
amount of probability so we're actually

272
00:09:32,640 --> 00:09:36,240
going to sample from our full

273
00:09:34,240 --> 00:09:37,720
distribution and then in the second

274
00:09:36,240 --> 00:09:39,279
example when we have a lot of things

275
00:09:37,720 --> 00:09:41,160
that are really unlikely we'll only

276
00:09:39,279 --> 00:09:43,800
sample from sort of the more likely part

277
00:09:41,160 --> 00:09:45,240
of the distribution um so all three of

278
00:09:43,800 --> 00:09:47,000
these methods are sort of different ways

279
00:09:45,240 --> 00:09:49,399
of trying to cut off the long tail using

280
00:09:47,000 --> 00:09:51,480
sort of different

281
00:09:49,399 --> 00:09:53,000
characteristics the tail of the

282
00:09:51,480 --> 00:09:55,680
distribution though isn't the only thing

283
00:09:53,000 --> 00:09:58,000
we could choose to modify um we could

284
00:09:55,680 --> 00:09:59,880
also choose to modify this sort of

285
00:09:58,000 --> 00:10:02,120
peakiness of the distribution

286
00:09:59,880 --> 00:10:03,880
so if you look here at the middle of

287
00:10:02,120 --> 00:10:06,600
these diagrams say this is your original

288
00:10:03,880 --> 00:10:08,519
distribution over next tokens and maybe

289
00:10:06,600 --> 00:10:11,040
you want to modify some properties of

290
00:10:08,519 --> 00:10:12,640
this distribution like you say I want an

291
00:10:11,040 --> 00:10:14,200
output that's really diverse and

292
00:10:12,640 --> 00:10:15,680
interesting and open-ended like maybe

293
00:10:14,200 --> 00:10:17,920
this is something like story generation

294
00:10:15,680 --> 00:10:20,120
where you want to have sort of a lot of

295
00:10:17,920 --> 00:10:21,279
maybe surprising things in your output

296
00:10:20,120 --> 00:10:23,480
you could say I want to sort of

297
00:10:21,279 --> 00:10:26,440
distribute my probability Mass more over

298
00:10:23,480 --> 00:10:28,399
the token space and you can do this um

299
00:10:26,440 --> 00:10:32,720
by sort of flattening this distribution

300
00:10:28,399 --> 00:10:34,240
like you see on the the right here um

301
00:10:32,720 --> 00:10:36,800
where now there's sort of more

302
00:10:34,240 --> 00:10:39,040
probability Mass spread over this um

303
00:10:36,800 --> 00:10:40,320
like wider set of tokens you could also

304
00:10:39,040 --> 00:10:42,720
say the opposite right you could say

305
00:10:40,320 --> 00:10:44,120
maybe I'm doing something like math

306
00:10:42,720 --> 00:10:45,519
where there shouldn't really be a lot of

307
00:10:44,120 --> 00:10:47,800
correct answers there should be really

308
00:10:45,519 --> 00:10:50,399
only one or maybe only like a few

309
00:10:47,800 --> 00:10:52,320
potential reasonable next answers and so

310
00:10:50,399 --> 00:10:54,160
you can make your distribution peier or

311
00:10:52,320 --> 00:10:56,639
sharper so that more of the probability

312
00:10:54,160 --> 00:11:00,200
mass is on the things at the very top um

313
00:10:56,639 --> 00:11:02,000
the way you do this is you modify y your

314
00:11:00,200 --> 00:11:04,320
loges your outputs of the last layer of

315
00:11:02,000 --> 00:11:06,399
the model before you apply softn so when

316
00:11:04,320 --> 00:11:08,360
you're predicting you get your outputs

317
00:11:06,399 --> 00:11:10,040
of the last layer of the model and then

318
00:11:08,360 --> 00:11:11,560
you apply softmax which turns those

319
00:11:10,040 --> 00:11:15,240
outputs into a distribution right they

320
00:11:11,560 --> 00:11:17,399
all sum up the um like Mass over all

321
00:11:15,240 --> 00:11:18,839
vocabulary tokens sums to one and so

322
00:11:17,399 --> 00:11:21,920
that is sort of a distribution you could

323
00:11:18,839 --> 00:11:23,519
sample from if you divide those Logics

324
00:11:21,920 --> 00:11:26,000
by some number before you apply that

325
00:11:23,519 --> 00:11:27,880
softmax you can make that distribution

326
00:11:26,000 --> 00:11:30,760
flatter by using a number greater than

327
00:11:27,880 --> 00:11:32,440
one or peier by using a number less than

328
00:11:30,760 --> 00:11:35,079
one and this is this type of parameter

329
00:11:32,440 --> 00:11:36,839
is called temperature um you can apply

330
00:11:35,079 --> 00:11:38,480
this with any of the other methods for

331
00:11:36,839 --> 00:11:40,279
sort of cutting off the long tail but

332
00:11:38,480 --> 00:11:41,920
what people will often do is just apply

333
00:11:40,279 --> 00:11:43,639
a temperature and then sample from that

334
00:11:41,920 --> 00:11:45,320
distribution and that's what we call

335
00:11:43,639 --> 00:11:48,720
temperature

336
00:11:45,320 --> 00:11:49,920
sampling so these I think most of you

337
00:11:48,720 --> 00:11:51,320
might already have been at least a

338
00:11:49,920 --> 00:11:53,000
little bit familiar with some of these

339
00:11:51,320 --> 00:11:56,079
methods I want to touch briefly on a

340
00:11:53,000 --> 00:11:58,160
couple of other ideas for modifying this

341
00:11:56,079 --> 00:11:59,680
distribution maybe some more complex and

342
00:11:58,160 --> 00:12:01,839
more recent ideas and the one that I

343
00:11:59,680 --> 00:12:04,279
want to talk about in more detail is

344
00:12:01,839 --> 00:12:05,399
something called contrastive decoding so

345
00:12:04,279 --> 00:12:07,360
the idea here is that we could

346
00:12:05,399 --> 00:12:10,800
incorporate some extra information at

347
00:12:07,360 --> 00:12:12,760
decoding time um using some other

348
00:12:10,800 --> 00:12:15,320
distribution some other data or in this

349
00:12:12,760 --> 00:12:17,320
case some other model so if you've ever

350
00:12:15,320 --> 00:12:19,240
played around with a really like

351
00:12:17,320 --> 00:12:21,800
relatively small language model maybe

352
00:12:19,240 --> 00:12:23,320
something like gbt2 small um You

353
00:12:21,800 --> 00:12:26,560
probably noticed you try to give it some

354
00:12:23,320 --> 00:12:28,240
inputs and maybe it degenerates into

355
00:12:26,560 --> 00:12:30,160
just repeating the same sequence over

356
00:12:28,240 --> 00:12:31,720
and over maybe it gives you outputs that

357
00:12:30,160 --> 00:12:33,399
are just completely incorrect like you

358
00:12:31,720 --> 00:12:35,320
ask it a factual question and it gets it

359
00:12:33,399 --> 00:12:37,120
wrong um and you don't see those

360
00:12:35,320 --> 00:12:39,519
problems if you look at sort of a larger

361
00:12:37,120 --> 00:12:41,399
model that's trained on more data so the

362
00:12:39,519 --> 00:12:43,199
question here is can you use what that

363
00:12:41,399 --> 00:12:46,480
smaller model is getting wrong to make

364
00:12:43,199 --> 00:12:49,120
your larger model even better um and the

365
00:12:46,480 --> 00:12:51,360
way we do this is by sort of the

366
00:12:49,120 --> 00:12:52,880
intuition that if the smaller model

367
00:12:51,360 --> 00:12:55,079
doesn't have a lot of probability on

368
00:12:52,880 --> 00:12:57,160
some answer but the the larger model

369
00:12:55,079 --> 00:12:58,519
does it's likely because that larger

370
00:12:57,160 --> 00:13:02,279
model has learned something with the

371
00:12:58,519 --> 00:13:04,000
smaller model didn't know and so here we

372
00:13:02,279 --> 00:13:06,199
modify the probability distribution

373
00:13:04,000 --> 00:13:08,199
coming out of the larger model to choose

374
00:13:06,199 --> 00:13:11,120
outputs that that model thinks are very

375
00:13:08,199 --> 00:13:12,600
likely and the amateur or the the weaker

376
00:13:11,120 --> 00:13:15,480
model thinks are not

377
00:13:12,600 --> 00:13:20,000
likely so in this example here from

378
00:13:15,480 --> 00:13:22,560
their paper um if you have sort of a

379
00:13:20,000 --> 00:13:27,199
input like Barack Obama was born in

380
00:13:22,560 --> 00:13:29,720
Hawaii he was born in L um the smaller

381
00:13:27,199 --> 00:13:31,360
model would often do something like

382
00:13:29,720 --> 00:13:35,399
start repeating and actually if you

383
00:13:31,360 --> 00:13:36,720
sample sort of naively from the um

384
00:13:35,399 --> 00:13:38,560
larger model you can wind up in these

385
00:13:36,720 --> 00:13:40,000
situations as well right so if you just

386
00:13:38,560 --> 00:13:41,959
choose the most likely thing at each

387
00:13:40,000 --> 00:13:43,399
step you wind up in this Loop where it's

388
00:13:41,959 --> 00:13:45,560
like he was born in Hawaii he was born

389
00:13:43,399 --> 00:13:48,199
in Hawaii he was born in Hawaii um and

390
00:13:45,560 --> 00:13:51,320
this is behavior we generally don't want

391
00:13:48,199 --> 00:13:52,680
um if you do something like nucleus or

392
00:13:51,320 --> 00:13:53,720
top PE sampling you can wind up with

393
00:13:52,680 --> 00:13:55,880
things that are actually completely

394
00:13:53,720 --> 00:13:58,839
incorrect like he was born in Washington

395
00:13:55,880 --> 00:14:01,480
DC um but if you use contrastive

396
00:13:58,839 --> 00:14:04,120
decoding you take the outputs coming out

397
00:14:01,480 --> 00:14:05,720
of your expert model here and you

398
00:14:04,120 --> 00:14:07,680
subtract out the probabilities coming

399
00:14:05,720 --> 00:14:10,160
out of the weaker model and you can wind

400
00:14:07,680 --> 00:14:11,880
up with things that the higher model the

401
00:14:10,160 --> 00:14:13,759
stronger model ascribed probability to

402
00:14:11,880 --> 00:14:15,480
but the weaker model did not likely

403
00:14:13,759 --> 00:14:16,920
because these are sort of facts that the

404
00:14:15,480 --> 00:14:18,959
larger model knows that the smaller

405
00:14:16,920 --> 00:14:20,800
model does not so here we actually get

406
00:14:18,959 --> 00:14:23,199
the year Barack Obama was born which is

407
00:14:20,800 --> 00:14:25,800
maybe a fact that the larger model knows

408
00:14:23,199 --> 00:14:27,639
and the smaller model didn't know um and

409
00:14:25,800 --> 00:14:29,759
so this is just one of sort of a broad

410
00:14:27,639 --> 00:14:32,560
class of methods where you use external

411
00:14:29,759 --> 00:14:35,199
information to improve your decoding by

412
00:14:32,560 --> 00:14:38,720
modifying this distribution at each

413
00:14:35,199 --> 00:14:40,720
set um those are sort of a brief tour of

414
00:14:38,720 --> 00:14:43,920
a couple of different sampling methods

415
00:14:40,720 --> 00:14:43,920
before we move into search

416
00:14:44,600 --> 00:14:50,440
yeah

417
00:14:46,279 --> 00:14:54,880
yeah is it going to improve upon just

418
00:14:50,440 --> 00:14:57,240
the yeah it generally does um and the

419
00:14:54,880 --> 00:14:59,800
intuition for why this might be I think

420
00:14:57,240 --> 00:15:01,680
is that there are sort of these

421
00:14:59,800 --> 00:15:04,560
degenerate cases like just repeating

422
00:15:01,680 --> 00:15:06,120
over and over that both the expert and

423
00:15:04,560 --> 00:15:09,000
the weak model would give relatively

424
00:15:06,120 --> 00:15:10,880
high probability to um maybe the expert

425
00:15:09,000 --> 00:15:13,199
model is like slightly less likely to do

426
00:15:10,880 --> 00:15:14,959
these things but it's still like sort of

427
00:15:13,199 --> 00:15:16,639
an easy case for the model to learn and

428
00:15:14,959 --> 00:15:18,120
so both of those models will have high

429
00:15:16,639 --> 00:15:20,079
probability for those things but the

430
00:15:18,120 --> 00:15:21,800
things that are genuinely like good

431
00:15:20,079 --> 00:15:23,880
outputs that only the expert would get

432
00:15:21,800 --> 00:15:25,519
right those will have low probability

433
00:15:23,880 --> 00:15:27,600
under the weak model and so you're sort

434
00:15:25,519 --> 00:15:30,880
of subtracting out all the degenerate

435
00:15:27,600 --> 00:15:33,759
behaviors and keeping to really good out

436
00:15:30,880 --> 00:15:35,240
this if you're generating a longer

437
00:15:33,759 --> 00:15:37,440
sequence with with

438
00:15:35,240 --> 00:15:40,759
contacing how do you know which steps

439
00:15:37,440 --> 00:15:45,120
you want to bring out yeah this is a

440
00:15:40,759 --> 00:15:48,560
great question so for this particular

441
00:15:45,120 --> 00:15:50,560
case oh yeah sorry so this was if you're

442
00:15:48,560 --> 00:15:52,279
doing contrastive decoding over a really

443
00:15:50,560 --> 00:15:54,399
long sequence like when do you choose to

444
00:15:52,279 --> 00:15:55,800
bring in the expert right and for

445
00:15:54,399 --> 00:15:58,600
contrastive decoding we're actually

446
00:15:55,800 --> 00:16:00,759
going to do this at every individual

447
00:15:58,600 --> 00:16:02,440
time step so we're going to use the

448
00:16:00,759 --> 00:16:04,800
expert model to decode and we're going

449
00:16:02,440 --> 00:16:07,000
to bring in the amateur to sort of

450
00:16:04,800 --> 00:16:09,079
subtract out probabilities at each next

451
00:16:07,000 --> 00:16:10,399
token prediction um you don't have to do

452
00:16:09,079 --> 00:16:12,800
that I think that's that's what they do

453
00:16:10,399 --> 00:16:15,000
in the paper um you could also decide to

454
00:16:12,800 --> 00:16:16,680
only do this sort of if you have high

455
00:16:15,000 --> 00:16:19,639
uncertainty or something if you don't

456
00:16:16,680 --> 00:16:22,639
have a really sharp probability

457
00:16:19,639 --> 00:16:22,639
distribution

458
00:16:23,160 --> 00:16:28,160
yeah yeah how weak should the weak

459
00:16:25,399 --> 00:16:30,199
predictor be um in the in the paper what

460
00:16:28,160 --> 00:16:31,600
they're look at is actually not a huge

461
00:16:30,199 --> 00:16:34,560
difference between the two models so you

462
00:16:31,600 --> 00:16:35,800
can see here this is gpd2 XL and small

463
00:16:34,560 --> 00:16:37,319
so there's a difference in parameter

464
00:16:35,800 --> 00:16:39,519
counts and like a bit of a difference in

465
00:16:37,319 --> 00:16:42,160
data I think here but these are actually

466
00:16:39,519 --> 00:16:44,959
not like gpd2 XL is certainly not like a

467
00:16:42,160 --> 00:16:48,399
super strong model now um I think they

468
00:16:44,959 --> 00:16:50,920
try a couple of different settings and

469
00:16:48,399 --> 00:16:52,319
the general intuition I think if I'm

470
00:16:50,920 --> 00:16:54,880
remembering it correctly is that you

471
00:16:52,319 --> 00:16:56,319
want a model that's not like so close in

472
00:16:54,880 --> 00:16:58,000
performance to your expert that you're

473
00:16:56,319 --> 00:16:59,839
basically just subtracting out useful

474
00:16:58,000 --> 00:17:02,240
things but you also don't want a model

475
00:16:59,839 --> 00:17:03,519
that's like so degenerate that it is not

476
00:17:02,240 --> 00:17:04,959
hasn't learned anything useful about

477
00:17:03,519 --> 00:17:06,839
your task at all so I think it might

478
00:17:04,959 --> 00:17:09,600
depend on what task you're looking

479
00:17:06,839 --> 00:17:12,919
at

480
00:17:09,600 --> 00:17:14,559
yes this is for inference um so actually

481
00:17:12,919 --> 00:17:17,640
everything we look at today will not

482
00:17:14,559 --> 00:17:17,640
require aning of the

483
00:17:19,360 --> 00:17:26,559
model Okay cool so now we're going to

484
00:17:24,000 --> 00:17:30,039
step into sort of a slightly different

485
00:17:26,559 --> 00:17:31,280
um set of strategies here which is maybe

486
00:17:30,039 --> 00:17:33,039
we don't just want something from the

487
00:17:31,280 --> 00:17:35,160
model distribution or something from a

488
00:17:33,039 --> 00:17:37,760
modified distribution maybe we actually

489
00:17:35,160 --> 00:17:39,840
just want the quote unquote best thing

490
00:17:37,760 --> 00:17:42,960
the single most likely output given our

491
00:17:39,840 --> 00:17:45,200
input right and here this would be the Y

492
00:17:42,960 --> 00:17:48,039
hat the single sequence that satisfies

493
00:17:45,200 --> 00:17:51,919
that has the highest score py given X

494
00:17:48,039 --> 00:17:54,240
for the X that we gave the model um this

495
00:17:51,919 --> 00:17:56,000
is this section is called mode seeking

496
00:17:54,240 --> 00:17:58,039
search because this is the mode of the

497
00:17:56,000 --> 00:18:00,440
distribution over outputs if you sampled

498
00:17:58,039 --> 00:18:01,760
a huge huge number of times and you

499
00:18:00,440 --> 00:18:04,720
looked at the single most likely

500
00:18:01,760 --> 00:18:06,720
sequence you got it would be this y hat

501
00:18:04,720 --> 00:18:09,280
and so how do we find this

502
00:18:06,720 --> 00:18:11,600
thing well one idea is we know the

503
00:18:09,280 --> 00:18:13,159
distribution at each individual setep

504
00:18:11,600 --> 00:18:16,000
can we just pick the most likely thing

505
00:18:13,159 --> 00:18:18,960
from that distribution and so in Greedy

506
00:18:16,000 --> 00:18:21,080
decoding we take the argmax the single

507
00:18:18,960 --> 00:18:22,720
highest probability token at each step

508
00:18:21,080 --> 00:18:24,840
and we continue generating until the

509
00:18:22,720 --> 00:18:26,600
single highest most the single highest

510
00:18:24,840 --> 00:18:28,840
probability token is the stop token

511
00:18:26,600 --> 00:18:31,559
right the end of sequence token

512
00:18:28,840 --> 00:18:33,400
um for an individual token right if we

513
00:18:31,559 --> 00:18:35,559
only want a single token output this is

514
00:18:33,400 --> 00:18:38,320
exactly what we want this is the single

515
00:18:35,559 --> 00:18:40,400
most likely output um and that's great

516
00:18:38,320 --> 00:18:44,000
but if we're looking at something that

517
00:18:40,400 --> 00:18:45,120
is maybe several tokens long are we

518
00:18:44,000 --> 00:18:47,360
actually going to get the highest

519
00:18:45,120 --> 00:18:49,720
probability thing and if you kind of

520
00:18:47,360 --> 00:18:52,159
squint at this you can see that maybe we

521
00:18:49,720 --> 00:18:54,120
have a problem here where the highest

522
00:18:52,159 --> 00:18:56,320
probability sequence that you get from

523
00:18:54,120 --> 00:18:58,039
multiplying across multiple steps

524
00:18:56,320 --> 00:18:59,559
doesn't necessarily start with the token

525
00:18:58,039 --> 00:19:01,600
that was highest probability at time

526
00:18:59,559 --> 00:19:03,200
step one right maybe if you're doing

527
00:19:01,600 --> 00:19:04,720
something like unconditional generation

528
00:19:03,200 --> 00:19:06,720
the highest probability token at time

529
00:19:04,720 --> 00:19:08,360
step one is always the but there could

530
00:19:06,720 --> 00:19:09,919
be a really probable sentence that just

531
00:19:08,360 --> 00:19:11,480
doesn't happen to start with the the

532
00:19:09,919 --> 00:19:12,720
word the' and you would never find it

533
00:19:11,480 --> 00:19:15,080
using GRE

534
00:19:12,720 --> 00:19:17,360
decoding so this isn't going to give us

535
00:19:15,080 --> 00:19:19,799
the highest probability output over a

536
00:19:17,360 --> 00:19:22,000
sequence that's more than one token one

537
00:19:19,799 --> 00:19:23,360
can we do anything better to try to find

538
00:19:22,000 --> 00:19:25,640
this um

539
00:19:23,360 --> 00:19:27,559
output and here we get into sort of one

540
00:19:25,640 --> 00:19:29,520
of the most popular decoding methods the

541
00:19:27,559 --> 00:19:32,600
one that you maybe heard of before which

542
00:19:29,520 --> 00:19:35,080
is beam search the idea here is that we

543
00:19:32,600 --> 00:19:36,559
don't want to miss a high probability

544
00:19:35,080 --> 00:19:38,880
token that's hidden behind a lower

545
00:19:36,559 --> 00:19:40,200
probability prefix so we want to kind of

546
00:19:38,880 --> 00:19:42,000
search through a couple of different

547
00:19:40,200 --> 00:19:43,760
options so that we don't discard

548
00:19:42,000 --> 00:19:47,120
something too early that might have high

549
00:19:43,760 --> 00:19:49,360
probability um later on in generation

550
00:19:47,120 --> 00:19:50,919
and this is a type of bread first search

551
00:19:49,360 --> 00:19:53,200
so we're going to look at a wide variety

552
00:19:50,919 --> 00:19:54,600
of options at a given time step we're

553
00:19:53,200 --> 00:19:55,600
going to pick some set of them to

554
00:19:54,600 --> 00:19:57,120
continue and then we're going to look at

555
00:19:55,600 --> 00:19:58,919
a wide variety of options for the next

556
00:19:57,120 --> 00:19:59,960
time step instead of generating all the

557
00:19:58,919 --> 00:20:02,200
way through a sequence and then

558
00:19:59,960 --> 00:20:04,320
generating all the way through another

559
00:20:02,200 --> 00:20:05,760
sequence um and how this works is we're

560
00:20:04,320 --> 00:20:07,559
going to pick sort of a number of

561
00:20:05,760 --> 00:20:09,400
candidates we'd like to explore a beam

562
00:20:07,559 --> 00:20:11,039
with so in this example we're going to

563
00:20:09,400 --> 00:20:12,799
pick three and we're going to say all

564
00:20:11,039 --> 00:20:15,480
right here are maybe three options for

565
00:20:12,799 --> 00:20:17,640
time step one for if we pick each of

566
00:20:15,480 --> 00:20:19,760
those three options what would be the

567
00:20:17,640 --> 00:20:21,799
three most likely things for time step

568
00:20:19,760 --> 00:20:23,200
two right rather than choosing just the

569
00:20:21,799 --> 00:20:24,520
single most likely thing in Greedy

570
00:20:23,200 --> 00:20:26,960
decoding we're going to pick three

571
00:20:24,520 --> 00:20:29,120
options and so now we have three options

572
00:20:26,960 --> 00:20:32,559
for time step one three options for time

573
00:20:29,120 --> 00:20:34,280
step two we now have nine options um

574
00:20:32,559 --> 00:20:36,320
here right three options and then three

575
00:20:34,280 --> 00:20:37,679
more for each of these and we don't want

576
00:20:36,320 --> 00:20:40,159
to continue doing this because this is

577
00:20:37,679 --> 00:20:41,960
going to sort of combinator explode so

578
00:20:40,159 --> 00:20:44,080
we need to choose some subset of these

579
00:20:41,960 --> 00:20:45,880
to continue with and the way we do that

580
00:20:44,080 --> 00:20:47,799
is we look at the probability over this

581
00:20:45,880 --> 00:20:49,240
two token sequence and we choose the two

582
00:20:47,799 --> 00:20:51,520
that have the highest probability

583
00:20:49,240 --> 00:20:53,400
overall so in this instance we've chosen

584
00:20:51,520 --> 00:20:55,679
sort of one thing from this first group

585
00:20:53,400 --> 00:20:57,760
and two things from the second group and

586
00:20:55,679 --> 00:20:59,760
now we're back down to three hypotheses

587
00:20:57,760 --> 00:21:02,120
each now two tokens long and we'll

588
00:20:59,760 --> 00:21:04,000
continue generating to time step three

589
00:21:02,120 --> 00:21:05,600
we'll get nine options we'll pre it back

590
00:21:04,000 --> 00:21:07,760
down to three and we'll continue until

591
00:21:05,600 --> 00:21:09,159
the end of generation where we now have

592
00:21:07,760 --> 00:21:10,679
three sequences and we'll just pick the

593
00:21:09,159 --> 00:21:14,000
one that's highest probability out of

594
00:21:10,679 --> 00:21:15,679
those three to return um this is not

595
00:21:14,000 --> 00:21:17,360
guaranteed to get you the highest

596
00:21:15,679 --> 00:21:18,480
probability thing right you still have

597
00:21:17,360 --> 00:21:20,039
this risk that you could be sort of

598
00:21:18,480 --> 00:21:22,279
pruning out something that's high

599
00:21:20,039 --> 00:21:24,159
probability but in general this sort of

600
00:21:22,279 --> 00:21:26,600
works um much better than greedy

601
00:21:24,159 --> 00:21:28,520
decoding and this is if you have a

602
00:21:26,600 --> 00:21:31,120
language model and you're sort of not

603
00:21:28,520 --> 00:21:32,440
what um decoding method it's using outs

604
00:21:31,120 --> 00:21:34,200
are pretty good it's either beam search

605
00:21:32,440 --> 00:21:37,120
or temperature samping right this is

606
00:21:34,200 --> 00:21:40,039
very effective this is used um pretty

607
00:21:37,120 --> 00:21:41,760
broadly there are however some issues

608
00:21:40,039 --> 00:21:43,760
with beam search and one of the biggest

609
00:21:41,760 --> 00:21:46,159
ones is that when you're doing this

610
00:21:43,760 --> 00:21:47,679
maximum likelihood sampling you really

611
00:21:46,159 --> 00:21:50,080
or the sampling to search for something

612
00:21:47,679 --> 00:21:51,760
that's very high likelihood um you

613
00:21:50,080 --> 00:21:53,679
really sacrifice a lot of diversity in

614
00:21:51,760 --> 00:21:55,320
your outputs and in particular you could

615
00:21:53,679 --> 00:21:57,279
wind up at the end of beam search with

616
00:21:55,320 --> 00:21:58,919
three different outputs to choose from

617
00:21:57,279 --> 00:22:00,120
that are all pretty pretty much the same

618
00:21:58,919 --> 00:22:02,640
like they're slightly different token

619
00:22:00,120 --> 00:22:04,559
sequences but they look very similar and

620
00:22:02,640 --> 00:22:07,480
so maybe you want to S get sort of a

621
00:22:04,559 --> 00:22:08,919
more diverse set um there's a couple of

622
00:22:07,480 --> 00:22:10,640
different methods in this category I'm

623
00:22:08,919 --> 00:22:12,679
going to very briefly shout out two of

624
00:22:10,640 --> 00:22:14,200
them um but the idea here is to sort of

625
00:22:12,679 --> 00:22:16,440
reintroduce some of the benefits of

626
00:22:14,200 --> 00:22:19,120
sampling while still doing this kind of

627
00:22:16,440 --> 00:22:20,919
search for high probability things um

628
00:22:19,120 --> 00:22:22,600
diverse beam search is one of these

629
00:22:20,919 --> 00:22:25,520
methods and here the idea is that we

630
00:22:22,600 --> 00:22:27,279
want to modify that scoring step when we

631
00:22:25,520 --> 00:22:28,600
choose which three out of our nine beams

632
00:22:27,279 --> 00:22:30,200
we want to continue

633
00:22:28,600 --> 00:22:32,000
to avoid choosing things that are really

634
00:22:30,200 --> 00:22:34,320
really close to each other right so

635
00:22:32,000 --> 00:22:36,039
maybe our highest probability thing is

636
00:22:34,320 --> 00:22:37,559
some sequence a and then if we look at

637
00:22:36,039 --> 00:22:39,520
the other sequences there's one that's

638
00:22:37,559 --> 00:22:41,279
pretty high probability but very similar

639
00:22:39,520 --> 00:22:43,600
to that sequence and there's one that's

640
00:22:41,279 --> 00:22:45,320
like slightly lower probability but very

641
00:22:43,600 --> 00:22:47,200
different and so maybe we would choose a

642
00:22:45,320 --> 00:22:49,679
sequence that is a little lower

643
00:22:47,200 --> 00:22:51,760
probability to maximize diversity in our

644
00:22:49,679 --> 00:22:53,799
set to try to get like sort of a wider

645
00:22:51,760 --> 00:22:56,200
range of options to choose from later in

646
00:22:53,799 --> 00:22:58,200
generation so this modifies the scoring

647
00:22:56,200 --> 00:23:00,120
to not just take into account likelihood

648
00:22:58,200 --> 00:23:03,200
but also similarity to other

649
00:23:00,120 --> 00:23:05,400
KS another option down this path is

650
00:23:03,200 --> 00:23:07,640
stochastic beam search where we're going

651
00:23:05,400 --> 00:23:09,279
to keep the scoring the same but rather

652
00:23:07,640 --> 00:23:11,679
than choosing just the top three most

653
00:23:09,279 --> 00:23:13,279
likely tokens to expand out each beam

654
00:23:11,679 --> 00:23:15,200
we're actually going to sample from some

655
00:23:13,279 --> 00:23:17,000
distribution and you could sample from

656
00:23:15,200 --> 00:23:18,760
the model distribution directly using

657
00:23:17,000 --> 00:23:20,200
ancestral sampling or you could use any

658
00:23:18,760 --> 00:23:22,679
of our sampling methods we talked about

659
00:23:20,200 --> 00:23:24,200
in the last section to do this and the

660
00:23:22,679 --> 00:23:25,799
the idea here is sort of similar to

661
00:23:24,200 --> 00:23:29,279
diverse beam search we want to get sort

662
00:23:25,799 --> 00:23:31,240
of a wider exploration of our models

663
00:23:29,279 --> 00:23:33,520
like output space you know we want to

664
00:23:31,240 --> 00:23:35,360
sort of explore more things instead of

665
00:23:33,520 --> 00:23:36,760
just seeking winding up with a bunch of

666
00:23:35,360 --> 00:23:39,679
outputs that look very similar at the

667
00:23:36,760 --> 00:23:41,120
end of beam search um if folks are

668
00:23:39,679 --> 00:23:43,679
interested in these I think these are

669
00:23:41,120 --> 00:23:46,159
both linked on the website um the the

670
00:23:43,679 --> 00:23:48,679
papers that both of these ideas came

671
00:23:46,159 --> 00:23:51,480
from

672
00:23:48,679 --> 00:23:54,400
Yes um for stochastic

673
00:23:51,480 --> 00:23:57,039
resarch the sampl probability takes into

674
00:23:54,400 --> 00:23:59,039
account the current part that we already

675
00:23:57,039 --> 00:24:02,000
travel okay

676
00:23:59,039 --> 00:24:04,320
yeah exactly so it's this um like

677
00:24:02,000 --> 00:24:05,640
selection step here but we're instead of

678
00:24:04,320 --> 00:24:07,760
just doing greedy selection we're going

679
00:24:05,640 --> 00:24:11,760
to do

680
00:24:07,760 --> 00:24:17,520
assembling yes my question was on the T

681
00:24:11,760 --> 00:24:23,200
yeah like you for something super simple

682
00:24:17,520 --> 00:24:26,520
like if both of them have a high are you

683
00:24:23,200 --> 00:24:28,120
like yeah so you would if it has a

684
00:24:26,520 --> 00:24:30,080
really high probability under both

685
00:24:28,120 --> 00:24:32,880
models it would have a lower probability

686
00:24:30,080 --> 00:24:35,080
after doing this sort of contrasted

687
00:24:32,880 --> 00:24:36,600
de right so if the if the smaller

688
00:24:35,080 --> 00:24:38,799
model's really good at your task this

689
00:24:36,600 --> 00:24:40,960
might not work very

690
00:24:38,799 --> 00:24:43,360
well yeah I think in the paper they're

691
00:24:40,960 --> 00:24:45,320
generally evaluating on these sort of

692
00:24:43,360 --> 00:24:48,279
like open ended generation task I bet

693
00:24:45,320 --> 00:24:51,279
this works a lot worse for

694
00:24:48,279 --> 00:24:51,279
now

695
00:24:56,760 --> 00:24:59,760
yes

696
00:25:02,440 --> 00:25:08,120
you yeah this is a great question um and

697
00:25:05,960 --> 00:25:11,559
so the question is how do we measure

698
00:25:08,120 --> 00:25:14,120
similar beams um you can sort of Define

699
00:25:11,559 --> 00:25:15,559
any kind of similarity function you like

700
00:25:14,120 --> 00:25:17,520
here um anything that you'd use to

701
00:25:15,559 --> 00:25:20,440
evaluate like how similar something is

702
00:25:17,520 --> 00:25:22,360
to a gold reference right um I think in

703
00:25:20,440 --> 00:25:25,039
the original diverse beam search they do

704
00:25:22,360 --> 00:25:27,760
this by looking at like exact token

705
00:25:25,039 --> 00:25:30,640
match across the two right like if these

706
00:25:27,760 --> 00:25:33,880
beams are the same in all but one of the

707
00:25:30,640 --> 00:25:35,600
tokens or they have like you know 50% of

708
00:25:33,880 --> 00:25:37,120
the tokens are shared across the beams

709
00:25:35,600 --> 00:25:38,559
and maybe these are really similar and

710
00:25:37,120 --> 00:25:40,559
they should try to choose two things

711
00:25:38,559 --> 00:25:42,600
that are different um but you could swap

712
00:25:40,559 --> 00:25:46,200
that out for any

713
00:25:42,600 --> 00:25:49,440
metc yes so

714
00:25:46,200 --> 00:25:50,960
the there's kind of like a that's Happ

715
00:25:49,440 --> 00:25:53,360
at

716
00:25:50,960 --> 00:25:55,000
every for the stochastic be search

717
00:25:53,360 --> 00:25:57,720
there's like a shering what do you mean

718
00:25:55,000 --> 00:26:00,520
by a shepher so it says modify the next

719
00:25:57,720 --> 00:26:03,000
sech selection because they're like um

720
00:26:00,520 --> 00:26:06,919
it is searching at a different space and

721
00:26:03,000 --> 00:26:09,679
it's not searching within the same 3D

722
00:26:06,919 --> 00:26:14,080
SE is it searching in a different space

723
00:26:09,679 --> 00:26:15,799
yeah so it's um in the same probability

724
00:26:14,080 --> 00:26:18,399
distribution but it'll see a different

725
00:26:15,799 --> 00:26:20,840
part of the distribution so when you're

726
00:26:18,399 --> 00:26:22,640
doing the grey search you'll only ever

727
00:26:20,840 --> 00:26:24,559
look at the top three tokens in the next

728
00:26:22,640 --> 00:26:27,120
token distribution because you're just

729
00:26:24,559 --> 00:26:29,840
selecting like the maximums um but in

730
00:26:27,120 --> 00:26:31,360
sampling you could you could get the

731
00:26:29,840 --> 00:26:32,880
same tokens right if they're really high

732
00:26:31,360 --> 00:26:35,720
likelihood but you could also sample

733
00:26:32,880 --> 00:26:38,399
something that's further down in the

734
00:26:35,720 --> 00:26:42,760
distribution yeah as a followup to that

735
00:26:38,399 --> 00:26:44,880
like into uh our stamping we take into

736
00:26:42,760 --> 00:26:46,960
account the probability of the prefix

737
00:26:44,880 --> 00:26:50,679
like the current hypothesis right

738
00:26:46,960 --> 00:26:51,760
because otherwise it is the same as just

739
00:26:50,679 --> 00:26:54,279
uh

740
00:26:51,760 --> 00:26:57,159
in yeah so in the sampling we're taking

741
00:26:54,279 --> 00:27:00,120
into account the previous the prefix

742
00:26:57,159 --> 00:27:02,600
yeah so so it we will take into account

743
00:27:00,120 --> 00:27:06,200
the prefix but this sampling mechanism

744
00:27:02,600 --> 00:27:08,320
here could be ancestral sampling um the

745
00:27:06,200 --> 00:27:10,480
only the difference here is that we're

746
00:27:08,320 --> 00:27:12,600
also doing a sort of search step on top

747
00:27:10,480 --> 00:27:14,679
of that to choose the maximum likelihood

748
00:27:12,600 --> 00:27:18,080
things across multiple

749
00:27:14,679 --> 00:27:20,559
me another important thing um is you

750
00:27:18,080 --> 00:27:22,279
sample without replacement and so

751
00:27:20,559 --> 00:27:24,120
normally you sample with replacement and

752
00:27:22,279 --> 00:27:25,840
you might get exactly the same thing but

753
00:27:24,120 --> 00:27:28,000
when you're doing stasic beam search you

754
00:27:25,840 --> 00:27:30,240
sample without replacement so you get

755
00:27:28,000 --> 00:27:33,279
like three ones according to the

756
00:27:30,240 --> 00:27:36,080
probability but they're guaranteed to be

757
00:27:33,279 --> 00:27:37,799
different right so beam search like one

758
00:27:36,080 --> 00:27:39,559
of the characteristics of beam search is

759
00:27:37,799 --> 00:27:41,640
you always get three different things

760
00:27:39,559 --> 00:27:44,240
because you're picking the three top

761
00:27:41,640 --> 00:27:45,760
when you do sampling uh like stochastic

762
00:27:44,240 --> 00:27:47,399
Bean shirts you get three different

763
00:27:45,760 --> 00:27:49,440
things they're not guaranteed to be the

764
00:27:47,399 --> 00:27:51,760
top they could be distributed according

765
00:27:49,440 --> 00:27:54,360
to the prob distribution but they're

766
00:27:51,760 --> 00:27:55,840
guaranteed so um you can take a look at

767
00:27:54,360 --> 00:27:58,039
the paper for more details of exactly

768
00:27:55,840 --> 00:28:00,159
how it looks but that that's

769
00:27:58,039 --> 00:28:03,039
so then is the main difference that

770
00:28:00,159 --> 00:28:05,120
compared to plus temping that we have n

771
00:28:03,039 --> 00:28:08,519
options that we're cheing tet instead of

772
00:28:05,120 --> 00:28:10,320
going with the going with only one and

773
00:28:08,519 --> 00:28:11,200
you can't yeah you can't simple the same

774
00:28:10,320 --> 00:28:14,960
thing

775
00:28:11,200 --> 00:28:16,919
right yeah so just uh repeat recording

776
00:28:14,960 --> 00:28:19,159
is that n options we're keeping track of

777
00:28:16,919 --> 00:28:22,240
and they're all going to be unique token

778
00:28:19,159 --> 00:28:24,240
sequences at least um you can actually

779
00:28:22,240 --> 00:28:26,200
get the same output sequence from two

780
00:28:24,240 --> 00:28:28,120
different toen sequences if you tokenize

781
00:28:26,200 --> 00:28:32,360
slightly differently um but these will

782
00:28:28,120 --> 00:28:37,840
always be unique tokens

783
00:28:32,360 --> 00:28:39,279
Le so that was sort of a a why like a a

784
00:28:37,840 --> 00:28:41,320
set of methods that we've developed to

785
00:28:39,279 --> 00:28:43,600
try to find the most probable sequence

786
00:28:41,320 --> 00:28:44,480
out of the model um but in the next

787
00:28:43,600 --> 00:28:46,039
section here we're going to sort of

788
00:28:44,480 --> 00:28:50,240
think about whether that's actually what

789
00:28:46,039 --> 00:28:51,679
we want to do at all um so what is like

790
00:28:50,240 --> 00:28:54,240
is do we really want the highest

791
00:28:51,679 --> 00:28:56,880
probability thing um we know that

792
00:28:54,240 --> 00:28:58,600
outputs with really low probability tend

793
00:28:56,880 --> 00:29:00,640
to be really like worse than outfits

794
00:28:58,600 --> 00:29:03,240
with high probability right maybe I'm

795
00:29:00,640 --> 00:29:05,840
trying to predict like what the next

796
00:29:03,240 --> 00:29:08,640
sentence should be after the cat saw the

797
00:29:05,840 --> 00:29:11,240
dog right the cat sat down is way higher

798
00:29:08,640 --> 00:29:12,559
probability than the cat grew wings and

799
00:29:11,240 --> 00:29:14,039
at least with the cats I've met that

800
00:29:12,559 --> 00:29:15,679
sounds pretty that sounds pretty much

801
00:29:14,039 --> 00:29:19,559
right right like this is a much better

802
00:29:15,679 --> 00:29:21,720
output than the cat gr wings but if you

803
00:29:19,559 --> 00:29:24,159
look at just the outputs with relatively

804
00:29:21,720 --> 00:29:25,960
high probability it's sort of less clear

805
00:29:24,159 --> 00:29:27,880
that this defines an exact ranking

806
00:29:25,960 --> 00:29:30,559
between those outputs right

807
00:29:27,880 --> 00:29:32,600
is the cat sat down necessarily better

808
00:29:30,559 --> 00:29:34,519
than the cat ran away these both seem

809
00:29:32,600 --> 00:29:35,720
like pretty reasonable outputs to me

810
00:29:34,519 --> 00:29:40,200
even though one of them is slightly

811
00:29:35,720 --> 00:29:42,799
higher probability and so we do we

812
00:29:40,200 --> 00:29:45,240
really like necessarily need to recover

813
00:29:42,799 --> 00:29:47,200
the cat that down um and this gets a

814
00:29:45,240 --> 00:29:49,399
little a little more complicated still

815
00:29:47,200 --> 00:29:51,120
if we look at sort of a range of outputs

816
00:29:49,399 --> 00:29:53,120
so say there's sort of six outputs that

817
00:29:51,120 --> 00:29:55,240
our model could give us um and here

818
00:29:53,120 --> 00:29:57,559
we're looking at sort of full sequences

819
00:29:55,240 --> 00:30:00,120
not individual tokens just for clarity

820
00:29:57,559 --> 00:30:02,640
so maybe our outputs in order of

821
00:30:00,120 --> 00:30:05,840
probability are the cat sat down it ran

822
00:30:02,640 --> 00:30:08,240
away it sprinted off it got out of there

823
00:30:05,840 --> 00:30:09,720
it's very small and it grew Wings right

824
00:30:08,240 --> 00:30:11,440
so we're definitely sure that the cat

825
00:30:09,720 --> 00:30:13,159
sat down is a better output than the cat

826
00:30:11,440 --> 00:30:15,360
grew wings and if we're doing a mod

827
00:30:13,159 --> 00:30:17,600
seeking search we would find that as our

828
00:30:15,360 --> 00:30:19,440
most likely thing if we're if we you

829
00:30:17,600 --> 00:30:21,440
know do a good job searching and we'd

830
00:30:19,440 --> 00:30:23,519
return that as our output but if you

831
00:30:21,440 --> 00:30:25,919
look at the rest of this distribution

832
00:30:23,519 --> 00:30:27,880
you see that there's actually a whole

833
00:30:25,919 --> 00:30:29,240
set of outputs after that all say

834
00:30:27,880 --> 00:30:31,720
something that kind of means the cat

835
00:30:29,240 --> 00:30:33,480
left the area right it's just that this

836
00:30:31,720 --> 00:30:35,200
probability is split over these three

837
00:30:33,480 --> 00:30:37,080
different generations and if you

838
00:30:35,200 --> 00:30:39,120
actually add up the probability mass of

839
00:30:37,080 --> 00:30:40,880
all three of these sequences this is

840
00:30:39,120 --> 00:30:42,919
double the probability mass of the cat

841
00:30:40,880 --> 00:30:44,360
sat down but because none of these

842
00:30:42,919 --> 00:30:45,960
individual sequences is higher

843
00:30:44,360 --> 00:30:47,399
probability if you're doing mode seeking

844
00:30:45,960 --> 00:30:50,640
search you wouldn't you wouldn't be able

845
00:30:47,399 --> 00:30:52,480
to see this effect right so do we really

846
00:30:50,640 --> 00:30:53,760
want to return the cat sat down or do we

847
00:30:52,480 --> 00:30:55,200
want to return something that means the

848
00:30:53,760 --> 00:30:57,559
cat left the

849
00:30:55,200 --> 00:30:59,200
area the question then is like if it's

850
00:30:57,559 --> 00:31:03,120
not probability that makes an output

851
00:30:59,200 --> 00:31:04,679
good what is it so we have this one

852
00:31:03,120 --> 00:31:06,039
output that's really high probability

853
00:31:04,679 --> 00:31:09,000
but it's very different from everything

854
00:31:06,039 --> 00:31:10,720
else in our set and then we have a

855
00:31:09,000 --> 00:31:13,200
couple of outputs that are all pretty

856
00:31:10,720 --> 00:31:15,080
high probability and similar to a bunch

857
00:31:13,200 --> 00:31:17,840
of other relatively high probability

858
00:31:15,080 --> 00:31:19,720
things so maybe it's sort of less risky

859
00:31:17,840 --> 00:31:21,399
to return one of these right are thing

860
00:31:19,720 --> 00:31:23,200
that's higher probability but different

861
00:31:21,399 --> 00:31:24,600
than everything else could be different

862
00:31:23,200 --> 00:31:26,840
because it's way better or it could be

863
00:31:24,600 --> 00:31:29,000
different because it's way worse um

864
00:31:26,840 --> 00:31:31,120
another way to think about this is you

865
00:31:29,000 --> 00:31:32,600
know maybe if you and your friends were

866
00:31:31,120 --> 00:31:34,200
cheating on a test which you shouldn't

867
00:31:32,600 --> 00:31:35,480
do but if you were going to do it and

868
00:31:34,200 --> 00:31:37,519
all of your friends sent you their

869
00:31:35,480 --> 00:31:39,240
answers um maybe one of your friends has

870
00:31:37,519 --> 00:31:40,960
a slightly higher score in the class

871
00:31:39,240 --> 00:31:42,519
than everyone else but they said the

872
00:31:40,960 --> 00:31:44,480
answer was answer a and everyone else

873
00:31:42,519 --> 00:31:45,799
said the answer was B right you still

874
00:31:44,480 --> 00:31:48,480
might go with the answer that everyone

875
00:31:45,799 --> 00:31:50,679
else said because like what there's it

876
00:31:48,480 --> 00:31:52,679
sort of feels less risky like maybe

877
00:31:50,679 --> 00:31:54,440
everyone else got the answer get that

878
00:31:52,679 --> 00:31:55,880
answer and so your one friend could be

879
00:31:54,440 --> 00:31:56,919
right when everyone else is wrong or

880
00:31:55,880 --> 00:31:59,679
they could have made a mistake that no

881
00:31:56,919 --> 00:32:01,240
one El else is making so this is sort of

882
00:31:59,679 --> 00:32:03,519
the same concept right we want an output

883
00:32:01,240 --> 00:32:06,320
that's relatively high probability but

884
00:32:03,519 --> 00:32:09,399
also relatively low

885
00:32:06,320 --> 00:32:11,320
risk and so here maybe if we were using

886
00:32:09,399 --> 00:32:13,679
this criteria we'd return the cat ran

887
00:32:11,320 --> 00:32:14,720
away as our sort of as our sort of

888
00:32:13,679 --> 00:32:16,720
single

889
00:32:14,720 --> 00:32:19,440
output so how do you find something

890
00:32:16,720 --> 00:32:21,000
that's high probability and low risk

891
00:32:19,440 --> 00:32:22,480
there's sort of two questions here right

892
00:32:21,000 --> 00:32:24,399
we have to figure out how to estimate

893
00:32:22,480 --> 00:32:26,120
probability and if we're looking at a

894
00:32:24,399 --> 00:32:28,519
set of outputs like the six we saw

895
00:32:26,120 --> 00:32:29,880
before maybe we can just do this by

896
00:32:28,519 --> 00:32:31,720
counting right we could sample

897
00:32:29,880 --> 00:32:34,000
everything from the model and get exact

898
00:32:31,720 --> 00:32:35,200
probability or we could take a sample

899
00:32:34,000 --> 00:32:38,080
from the model and just look at

900
00:32:35,200 --> 00:32:40,200
probabilities in that set and from there

901
00:32:38,080 --> 00:32:41,840
from that sample um sort of one

902
00:32:40,200 --> 00:32:43,559
reasonable thing to do is just count

903
00:32:41,840 --> 00:32:45,320
frequency right if something's in our

904
00:32:43,559 --> 00:32:47,919
sample twice as often we just say it's

905
00:32:45,320 --> 00:32:49,799
twice as frequent or it's twice as

906
00:32:47,919 --> 00:32:52,880
probable um this is something called

907
00:32:49,799 --> 00:32:54,440
Monte Carlos sampling if you do this um

908
00:32:52,880 --> 00:32:56,039
enough times like if you sample an

909
00:32:54,440 --> 00:32:58,279
infinite set this is would give you

910
00:32:56,039 --> 00:33:00,880
exactly the model distri distribution um

911
00:32:58,279 --> 00:33:02,840
but for the sort of reasonable size sets

912
00:33:00,880 --> 00:33:04,200
we're working with maybe like a 100

913
00:33:02,840 --> 00:33:06,320
samples this gives us a sort of

914
00:33:04,200 --> 00:33:09,440
reasonable approximation for what we for

915
00:33:06,320 --> 00:33:10,840
what we need to do here at least so

916
00:33:09,440 --> 00:33:12,000
we're just going to take a sample to get

917
00:33:10,840 --> 00:33:13,440
probability and we're just going to

918
00:33:12,000 --> 00:33:15,519
count things in that sample to see how

919
00:33:13,440 --> 00:33:17,320
likely things are that doesn't seem too

920
00:33:15,519 --> 00:33:20,080
bad how do we estimate

921
00:33:17,320 --> 00:33:21,679
risk the idea here is that we have a

922
00:33:20,080 --> 00:33:24,080
bunch of other things in this set of

923
00:33:21,679 --> 00:33:26,080
outputs and we can treat those as sort

924
00:33:24,080 --> 00:33:27,880
of like pseudo references right we can

925
00:33:26,080 --> 00:33:29,840
evaluate agreement between the thing

926
00:33:27,880 --> 00:33:31,519
we're looking at and each of those other

927
00:33:29,840 --> 00:33:33,480
references and this is sort of the same

928
00:33:31,519 --> 00:33:35,519
idea of calculating similarity in

929
00:33:33,480 --> 00:33:37,159
diverse beam search we're going to use

930
00:33:35,519 --> 00:33:39,639
some kind of metric to compare how

931
00:33:37,159 --> 00:33:41,279
similar these things are um this metric

932
00:33:39,639 --> 00:33:43,080
could be anything you use Downstream it

933
00:33:41,279 --> 00:33:44,840
could be like an engram overlap metric

934
00:33:43,080 --> 00:33:48,600
like Rouge or blue or it could also be

935
00:33:44,840 --> 00:33:51,120
something um neural or semantic like um

936
00:33:48,600 --> 00:33:54,799
something like BT score or Bart

937
00:33:51,120 --> 00:33:56,600
score and so this concept um is a type

938
00:33:54,799 --> 00:33:57,919
of decoding called minimum based risk

939
00:33:56,600 --> 00:33:59,600
decoding

940
00:33:57,919 --> 00:34:01,840
and what this equation captures is

941
00:33:59,600 --> 00:34:03,919
exactly the intuition that we were um

942
00:34:01,840 --> 00:34:06,600
sort of talking about just a slide ago

943
00:34:03,919 --> 00:34:08,159
where we're going to choose something

944
00:34:06,600 --> 00:34:09,919
that is low risk which means it's

945
00:34:08,159 --> 00:34:11,960
similar to a lot of other things in this

946
00:34:09,919 --> 00:34:12,800
set of outputs we've sampled and we're

947
00:34:11,960 --> 00:34:14,800
going to choose something that's

948
00:34:12,800 --> 00:34:17,560
relatively high probability which means

949
00:34:14,800 --> 00:34:19,159
that sort of when we sum up over this if

950
00:34:17,560 --> 00:34:21,399
something occurs in our set a bunch of

951
00:34:19,159 --> 00:34:23,320
times it's going to have pretty strong

952
00:34:21,399 --> 00:34:25,800
weight in picking which um of these

953
00:34:23,320 --> 00:34:27,000
outputs are similar right if sort of

954
00:34:25,800 --> 00:34:28,399
there's one thing in the set that

955
00:34:27,000 --> 00:34:29,919
appears a bunch of times it's going to

956
00:34:28,399 --> 00:34:32,040
have a strong influence on which thing

957
00:34:29,919 --> 00:34:34,119
we pick and so that kind of captures

958
00:34:32,040 --> 00:34:38,520
high probability in this

959
00:34:34,119 --> 00:34:41,119
setting so to see how this works we can

960
00:34:38,520 --> 00:34:44,639
look at an example um in

961
00:34:41,119 --> 00:34:47,399
summarization so we choose some Metric

962
00:34:44,639 --> 00:34:49,639
maybe we choose um Rouge which is an

963
00:34:47,399 --> 00:34:51,399
engram overlap metric for summarization

964
00:34:49,639 --> 00:34:52,879
and we say we're going to sample 100

965
00:34:51,399 --> 00:34:55,960
things and we're going to use this

966
00:34:52,879 --> 00:35:00,359
equation to choose the one that has the

967
00:34:55,960 --> 00:35:03,960
sort of lower EST risk according to MBR

968
00:35:00,359 --> 00:35:06,480
um so if we do that and we look at this

969
00:35:03,960 --> 00:35:07,560
sort of table of results here um you can

970
00:35:06,480 --> 00:35:09,680
see that this

971
00:35:07,560 --> 00:35:11,320
outperforms the other sampling methods

972
00:35:09,680 --> 00:35:13,720
that we've looked at before so greedy

973
00:35:11,320 --> 00:35:15,640
decoding here is just sampling the

974
00:35:13,720 --> 00:35:18,760
single most likely thing in each step

975
00:35:15,640 --> 00:35:21,800
beam search here is the BS with five or

976
00:35:18,760 --> 00:35:24,359
10 beams and DBS is the diverse beam

977
00:35:21,800 --> 00:35:27,040
search we were talking about um if we

978
00:35:24,359 --> 00:35:29,440
use minimum based risk and we use grou

979
00:35:27,040 --> 00:35:31,240
is the sort of determiner of similarity

980
00:35:29,440 --> 00:35:32,680
we do way better across all of our

981
00:35:31,240 --> 00:35:33,960
metrics but we especially do really good

982
00:35:32,680 --> 00:35:36,680
at Rouge because that's sort of the

983
00:35:33,960 --> 00:35:38,119
metric that we've been using to evaluate

984
00:35:36,680 --> 00:35:40,240
and then if we swap this out for other

985
00:35:38,119 --> 00:35:43,599
metrics you still see an performance

986
00:35:40,240 --> 00:35:46,440
improvement over these um search methods

987
00:35:43,599 --> 00:35:48,119
here um what's the sort of catch here

988
00:35:46,440 --> 00:35:49,920
the catch here is that MBR requires you

989
00:35:48,119 --> 00:35:51,599
to sample a hundred things and so this

990
00:35:49,920 --> 00:35:54,760
is a lot more expensive it's a lot

991
00:35:51,599 --> 00:35:54,760
slower at infin

992
00:35:54,800 --> 00:35:58,800
time um yes

993
00:36:04,200 --> 00:36:10,040
yes a great question why does the beam

994
00:36:07,000 --> 00:36:14,000
search with more beams perform worse um

995
00:36:10,040 --> 00:36:16,720
this is a well a relatively welln

996
00:36:14,000 --> 00:36:19,359
phenomena called the cursive beam search

997
00:36:16,720 --> 00:36:21,640
which is we actually lost your M so you

998
00:36:19,359 --> 00:36:24,599
mic and we can speak okay yeah so this

999
00:36:21,640 --> 00:36:26,079
is called the cursive beam search um and

1000
00:36:24,599 --> 00:36:27,760
the idea here is that beam search is

1001
00:36:26,079 --> 00:36:29,359
like an approxim search right so if you

1002
00:36:27,760 --> 00:36:31,200
add more beams you should be doing

1003
00:36:29,359 --> 00:36:33,319
better and better at finding the maximum

1004
00:36:31,200 --> 00:36:34,800
likelihood thing and generally you are

1005
00:36:33,319 --> 00:36:37,160
you get something that is higher

1006
00:36:34,800 --> 00:36:39,160
probability but as you add more beams

1007
00:36:37,160 --> 00:36:42,319
you also often get something that does

1008
00:36:39,160 --> 00:36:42,319
worse on your Downstream

1009
00:36:44,160 --> 00:36:47,560
metrics back up

1010
00:36:54,240 --> 00:36:58,680
there is that back online

1011
00:36:59,119 --> 00:37:06,520
yeah is that back is that any louder no

1012
00:37:03,520 --> 00:37:06,520
it

1013
00:37:07,000 --> 00:37:12,640
question oh there we go is that better

1014
00:37:09,599 --> 00:37:13,760
great um yeah so why why does this

1015
00:37:12,640 --> 00:37:16,040
happen right why do you get something

1016
00:37:13,760 --> 00:37:18,560
that's higher likelihood but um lower

1017
00:37:16,040 --> 00:37:22,040
performance Downstream um and this is

1018
00:37:18,560 --> 00:37:24,000
like another sort of degeneracy of beam

1019
00:37:22,040 --> 00:37:25,680
search that this idea that the thing

1020
00:37:24,000 --> 00:37:27,440
that is the absolute highest likelihood

1021
00:37:25,680 --> 00:37:28,599
in your distribution might not actually

1022
00:37:27,440 --> 00:37:31,079
be what you want

1023
00:37:28,599 --> 00:37:33,960
Downstream um this is sort of one of the

1024
00:37:31,079 --> 00:37:35,200
other things that people use to motivate

1025
00:37:33,960 --> 00:37:37,599
why you might want to do something like

1026
00:37:35,200 --> 00:37:39,400
MBR instead um and there's a great paper

1027
00:37:37,599 --> 00:37:41,640
about this problem called the inadequacy

1028
00:37:39,400 --> 00:37:43,680
of the mode because beam search is

1029
00:37:41,640 --> 00:37:45,520
looking for the mode of the

1030
00:37:43,680 --> 00:37:47,880
distribution well one other thing I'd

1031
00:37:45,520 --> 00:37:49,680
like to mention is it also goes together

1032
00:37:47,880 --> 00:37:51,119
with how you train your models because

1033
00:37:49,680 --> 00:37:53,760
most of our models are trained using

1034
00:37:51,119 --> 00:37:57,079
maximum likelihood maximum likelihood

1035
00:37:53,760 --> 00:37:59,040
isn't explicitly maximizing our ability

1036
00:37:57,079 --> 00:38:01,079
to get the best answer it's explicitly

1037
00:37:59,040 --> 00:38:05,720
maximizing our ability to estimate the

1038
00:38:01,079 --> 00:38:10,160
the distribution of answers so if I

1039
00:38:05,720 --> 00:38:13,040
say um if you said like what is what is

1040
00:38:10,160 --> 00:38:15,839
your favorite hobby or something like

1041
00:38:13,040 --> 00:38:17,680
that uh what is your favorite hobby in a

1042
00:38:15,839 --> 00:38:19,280
dialogue system often it'll answer I

1043
00:38:17,680 --> 00:38:22,400
don't know or something like that

1044
00:38:19,280 --> 00:38:24,920
because it like you know that that's

1045
00:38:22,400 --> 00:38:26,599
more likely than answering any specific

1046
00:38:24,920 --> 00:38:29,240
hobby like it's more likely than

1047
00:38:26,599 --> 00:38:32,119
answering basketball bowling you know

1048
00:38:29,240 --> 00:38:35,040
whatever else because you have many many

1049
00:38:32,119 --> 00:38:36,560
different options and so like especially

1050
00:38:35,040 --> 00:38:39,880
if it's something that's a little bit

1051
00:38:36,560 --> 00:38:42,160
more comp complicated it will avoid

1052
00:38:39,880 --> 00:38:44,680
answering that and in particular it ends

1053
00:38:42,160 --> 00:38:47,240
up answering very short things for

1054
00:38:44,680 --> 00:38:49,280
example um or sometimes it ends up

1055
00:38:47,240 --> 00:38:51,160
repeating itself over and over again or

1056
00:38:49,280 --> 00:38:53,240
or things like that so it also goes

1057
00:38:51,160 --> 00:38:57,760
together with like the training of the

1058
00:38:53,240 --> 00:38:59,359
model yeah and this is um one of the

1059
00:38:57,760 --> 00:39:01,079
this is still a problem in modern

1060
00:38:59,359 --> 00:39:02,560
systems so if you actually look at the

1061
00:39:01,079 --> 00:39:03,839
single like if you could enumerate

1062
00:39:02,560 --> 00:39:05,680
everything and see the single most

1063
00:39:03,839 --> 00:39:07,520
likely sequence it's often the empty

1064
00:39:05,680 --> 00:39:10,920
sequence just not opening anything at

1065
00:39:07,520 --> 00:39:12,640
all um and so if that's your true mode

1066
00:39:10,920 --> 00:39:16,119
of the distribution then doing better at

1067
00:39:12,640 --> 00:39:16,119
mode seeking is not always like

1068
00:39:16,599 --> 00:39:19,599
helpful

1069
00:39:25,440 --> 00:39:32,960
yes could this be influenced by the

1070
00:39:28,200 --> 00:39:32,960
confidence problem like um how

1071
00:39:37,560 --> 00:39:41,079
so seems

1072
00:39:49,760 --> 00:39:53,599
bees

1073
00:39:51,010 --> 00:39:57,280
[Music]

1074
00:39:53,599 --> 00:39:59,760
might right I think I I think I see

1075
00:39:57,280 --> 00:40:02,000
what you're saying which is that like

1076
00:39:59,760 --> 00:40:04,200
the the confidence gives you the

1077
00:40:02,000 --> 00:40:06,680
confidence of like a single exact

1078
00:40:04,200 --> 00:40:11,000
sequence right not the like actual sort

1079
00:40:06,680 --> 00:40:13,200
of semantic space of and so yeah if you

1080
00:40:11,000 --> 00:40:14,920
looked at just like the if you look at

1081
00:40:13,200 --> 00:40:17,000
just the probability scores you get the

1082
00:40:14,920 --> 00:40:18,520
probability of an exact string when what

1083
00:40:17,000 --> 00:40:20,119
you really actually care about with

1084
00:40:18,520 --> 00:40:22,319
confidence is the probability of sort of

1085
00:40:20,119 --> 00:40:23,800
like things that mean the same thing

1086
00:40:22,319 --> 00:40:25,359
yeah this is um part of why like

1087
00:40:23,800 --> 00:40:28,359
calibration is really hard for long

1088
00:40:25,359 --> 00:40:28,359
sequences

1089
00:40:30,720 --> 00:40:37,319
great so we're g to touch sort of

1090
00:40:34,359 --> 00:40:39,520
briefly on a couple of other things that

1091
00:40:37,319 --> 00:40:40,920
aren't sort of always explicitly

1092
00:40:39,520 --> 00:40:42,480
described in this framework but that you

1093
00:40:40,920 --> 00:40:45,040
can think of as variance of minimum

1094
00:40:42,480 --> 00:40:46,960
based risk um and if you're interested

1095
00:40:45,040 --> 00:40:49,560
in this analysis um I think as Graham

1096
00:40:46,960 --> 00:40:51,800
mentioned earlier um Alex Z is a first

1097
00:40:49,560 --> 00:40:53,680
year MLT and I wrote a paper about this

1098
00:40:51,800 --> 00:40:57,839
um which you can check out if you're

1099
00:40:53,680 --> 00:41:01,200
interested so the um two that I really

1100
00:40:57,839 --> 00:41:03,800
want to touch on here are other sort of

1101
00:41:01,200 --> 00:41:05,240
inference time things you can consider

1102
00:41:03,800 --> 00:41:07,520
which might look a little bit different

1103
00:41:05,240 --> 00:41:09,480
on the first BL um the first of these is

1104
00:41:07,520 --> 00:41:11,680
output ensembling so say you have

1105
00:41:09,480 --> 00:41:13,240
multiple different models and you get

1106
00:41:11,680 --> 00:41:15,480
outputs from all of them and now you

1107
00:41:13,240 --> 00:41:19,560
need to choose a best output among that

1108
00:41:15,480 --> 00:41:21,599
set um one of the sort of common ways to

1109
00:41:19,560 --> 00:41:24,480
do this is to compare like an embedding

1110
00:41:21,599 --> 00:41:25,920
similarity across models like does model

1111
00:41:24,480 --> 00:41:27,560
one think these two things are really

1112
00:41:25,920 --> 00:41:28,880
similar does model two think these two

1113
00:41:27,560 --> 00:41:32,599
things are really similar and try to

1114
00:41:28,880 --> 00:41:34,680
choose something that the um has really

1115
00:41:32,599 --> 00:41:37,319
high similarity with a lot of other

1116
00:41:34,680 --> 00:41:39,200
outputs um of course now that we've just

1117
00:41:37,319 --> 00:41:41,440
recently been talking about MBR you can

1118
00:41:39,200 --> 00:41:44,920
see that you can probably see that this

1119
00:41:41,440 --> 00:41:46,280
is um the same general formulation just

1120
00:41:44,920 --> 00:41:47,880
rather than summing over a set of

1121
00:41:46,280 --> 00:41:49,520
outputs from a single model now you're

1122
00:41:47,880 --> 00:41:52,160
looking at outputs over a whole set of

1123
00:41:49,520 --> 00:41:54,640
models um so some types of ensembling

1124
00:41:52,160 --> 00:41:57,319
fall into this category of minimum based

1125
00:41:54,640 --> 00:42:00,680
risk methods another thing in this

1126
00:41:57,319 --> 00:42:03,280
category is a um sort of recent decoding

1127
00:42:00,680 --> 00:42:06,079
method called self-consistency and the

1128
00:42:03,280 --> 00:42:08,200
idea here is that you want to do

1129
00:42:06,079 --> 00:42:09,359
something like mathematical reasoning

1130
00:42:08,200 --> 00:42:10,599
and you really care about getting the

1131
00:42:09,359 --> 00:42:12,000
final answer right but you don't

1132
00:42:10,599 --> 00:42:15,000
necessarily care about getting all of

1133
00:42:12,000 --> 00:42:18,079
the the reasoning steps in between right

1134
00:42:15,000 --> 00:42:19,520
so you prompt the model for an answer um

1135
00:42:18,079 --> 00:42:20,800
using something like Chain of Thought

1136
00:42:19,520 --> 00:42:22,680
right you ask it to sort of talk through

1137
00:42:20,800 --> 00:42:26,440
the steps it's going to do and then give

1138
00:42:22,680 --> 00:42:28,599
you a final answer um you sample many

1139
00:42:26,440 --> 00:42:30,400
puts using this and then you completely

1140
00:42:28,599 --> 00:42:32,200
throw away the chains of thought um and

1141
00:42:30,400 --> 00:42:35,359
you just take the answer from each

1142
00:42:32,200 --> 00:42:37,640
output um you have that set of answers

1143
00:42:35,359 --> 00:42:38,960
maybe you have like 20 30 100 answers

1144
00:42:37,640 --> 00:42:40,000
you just return the one that was most

1145
00:42:38,960 --> 00:42:43,720
frequently

1146
00:42:40,000 --> 00:42:46,119
generated um what this is doing is a

1147
00:42:43,720 --> 00:42:48,800
type of MBR where the metric that you

1148
00:42:46,119 --> 00:42:51,160
actually care about is exact match of

1149
00:42:48,800 --> 00:42:51,839
this answer right ignoring the rest of

1150
00:42:51,160 --> 00:42:54,079
the

1151
00:42:51,839 --> 00:42:55,800
generation um and so here we have sort

1152
00:42:54,079 --> 00:42:56,839
of the same intuition that we want an

1153
00:42:55,800 --> 00:42:59,160
output

1154
00:42:56,839 --> 00:43:01,520
that is high probability right we're

1155
00:42:59,160 --> 00:43:03,359
getting it generated a lot but also low

1156
00:43:01,520 --> 00:43:06,079
risk not a lot of the other outputs in

1157
00:43:03,359 --> 00:43:08,440
our in our set disagree with this

1158
00:43:06,079 --> 00:43:10,359
answer so those are a couple of

1159
00:43:08,440 --> 00:43:11,920
different variants of methods where

1160
00:43:10,359 --> 00:43:13,880
we're sort of sampling a wide set of

1161
00:43:11,920 --> 00:43:17,359
sequences and trying to choose the best

1162
00:43:13,880 --> 00:43:20,960
one um MBR is one set is one type of

1163
00:43:17,359 --> 00:43:22,680
sort of sequence set reranking method um

1164
00:43:20,960 --> 00:43:24,760
you could do other things to rerank sets

1165
00:43:22,680 --> 00:43:27,400
as well but this is sort of one

1166
00:43:24,760 --> 00:43:30,359
representative class of these yes uh or

1167
00:43:27,400 --> 00:43:32,280
of the of these methods before we get

1168
00:43:30,359 --> 00:43:35,200
into constrain generation those are sort

1169
00:43:32,280 --> 00:43:37,000
of the three broad categories of

1170
00:43:35,200 --> 00:43:39,480
inference methods we'll discuss which is

1171
00:43:37,000 --> 00:43:41,680
sort of sampling from some distribution

1172
00:43:39,480 --> 00:43:45,040
searching over some space of

1173
00:43:41,680 --> 00:43:47,400
distributions and doing some kind of um

1174
00:43:45,040 --> 00:43:48,559
analysis over a set of samples to choose

1175
00:43:47,400 --> 00:43:51,359
which ones they

1176
00:43:48,559 --> 00:43:52,559
return um does anyone have any questions

1177
00:43:51,359 --> 00:43:55,079
at this

1178
00:43:52,559 --> 00:44:00,680
point

1179
00:43:55,079 --> 00:44:00,680
yeah that a model

1180
00:44:05,800 --> 00:44:12,760
cannot yeah like why is averaging model

1181
00:44:08,359 --> 00:44:16,400
weights not MBR um I think it's not MBR

1182
00:44:12,760 --> 00:44:18,559
because the two um the key thing that I

1183
00:44:16,400 --> 00:44:20,880
think really makes a method MBR is this

1184
00:44:18,559 --> 00:44:22,480
concept of comparing between multiple um

1185
00:44:20,880 --> 00:44:24,880
sort of pseudo

1186
00:44:22,480 --> 00:44:26,839
references um and there you don't have

1187
00:44:24,880 --> 00:44:28,359
the same like you aage model way can you

1188
00:44:26,839 --> 00:44:32,440
wind up with sort of a single output on

1189
00:44:28,359 --> 00:44:34,040
the end that maybe is like using like

1190
00:44:32,440 --> 00:44:35,800
information from these two model

1191
00:44:34,040 --> 00:44:38,240
distributions that you've sort of smush

1192
00:44:35,800 --> 00:44:41,160
together um but it's not the same

1193
00:44:38,240 --> 00:44:44,720
concept of like comparing against pseudo

1194
00:44:41,160 --> 00:44:44,720
references or ranking in a

1195
00:44:48,920 --> 00:44:55,599
set right so now this is sort of a this

1196
00:44:52,720 --> 00:44:57,559
was a wide variety of methods to try to

1197
00:44:55,599 --> 00:44:59,040
find an output that's just sort of good

1198
00:44:57,559 --> 00:45:01,440
right we want an output that that is

1199
00:44:59,040 --> 00:45:03,480
nice out of our model um but now we'd

1200
00:45:01,440 --> 00:45:05,880
like to maybe enclose a few additional

1201
00:45:03,480 --> 00:45:08,280
constraints so say I'm asking our model

1202
00:45:05,880 --> 00:45:10,720
for some Hobbies I could use to stay in

1203
00:45:08,280 --> 00:45:11,920
to stay in shape and no matter what I

1204
00:45:10,720 --> 00:45:14,160
don't want the model to recommend

1205
00:45:11,920 --> 00:45:16,880
climbing like I I just I don't want this

1206
00:45:14,160 --> 00:45:18,400
as an option I've tried it I'm not a fan

1207
00:45:16,880 --> 00:45:21,240
um how do I get the model to stop

1208
00:45:18,400 --> 00:45:22,760
suggesting climbing to me and if you've

1209
00:45:21,240 --> 00:45:24,559
sort of played around with some of the

1210
00:45:22,760 --> 00:45:26,200
more recent llms you'd say maybe this is

1211
00:45:24,559 --> 00:45:27,480
easy right you just tell the model the

1212
00:45:26,200 --> 00:45:30,160
instruction that you don't want to talk

1213
00:45:27,480 --> 00:45:31,640
about climbing and having talked to Bard

1214
00:45:30,160 --> 00:45:33,640
recently I can tell you unfortunately

1215
00:45:31,640 --> 00:45:34,800
that it's not that easy so I tell the

1216
00:45:33,640 --> 00:45:36,599
model I don't want to talk about

1217
00:45:34,800 --> 00:45:38,000
climbing it does okay for a little bit

1218
00:45:36,599 --> 00:45:40,920
and then it's like all right but maybe

1219
00:45:38,000 --> 00:45:42,359
you want to try rap climbing um and so

1220
00:45:40,920 --> 00:45:44,559
we could continue trying to instruction

1221
00:45:42,359 --> 00:45:46,200
to our model but maybe there's sort of a

1222
00:45:44,559 --> 00:45:49,079
way to impose this constraint on the

1223
00:45:46,200 --> 00:45:50,680
decoding side instead and so I'd say all

1224
00:45:49,079 --> 00:45:52,960
right I'm going to do something dramatic

1225
00:45:50,680 --> 00:45:54,440
right I know I can manipulate the

1226
00:45:52,960 --> 00:45:56,200
probability distribution I'm just going

1227
00:45:54,440 --> 00:45:57,920
to set the probability of climbing to be

1228
00:45:56,200 --> 00:46:00,440
zero I don't want to see this token like

1229
00:45:57,920 --> 00:46:02,640
I'm I'm completely over it um and this

1230
00:46:00,440 --> 00:46:04,839
is sort of nice in some sense because

1231
00:46:02,640 --> 00:46:06,720
this is pretty easy to do um remember

1232
00:46:04,839 --> 00:46:08,440
we're doing a soft Max over the outputs

1233
00:46:06,720 --> 00:46:10,599
to choose this probability distribution

1234
00:46:08,440 --> 00:46:12,400
and so if we add a huge negative number

1235
00:46:10,599 --> 00:46:14,160
to the logic for climbing before we do

1236
00:46:12,400 --> 00:46:15,520
this softmax its probability is going to

1237
00:46:14,160 --> 00:46:18,640
be basically zero and we're never going

1238
00:46:15,520 --> 00:46:20,240
to see it as an output um but this

1239
00:46:18,640 --> 00:46:22,480
doesn't seem like a perfect solution

1240
00:46:20,240 --> 00:46:24,400
right because you know what if the model

1241
00:46:22,480 --> 00:46:26,160
recommends bouldering to me do I have to

1242
00:46:24,400 --> 00:46:28,599
write like a sort of a list of every

1243
00:46:26,160 --> 00:46:30,599
possible climbing synonym in the world

1244
00:46:28,599 --> 00:46:32,079
um what if there's sort of an allowable

1245
00:46:30,599 --> 00:46:33,920
way to use this token like I want the

1246
00:46:32,079 --> 00:46:35,319
model to suggest hiking because climbing

1247
00:46:33,920 --> 00:46:37,480
up a mountain to see a good view is

1248
00:46:35,319 --> 00:46:38,720
relaxing but that's a use of the word

1249
00:46:37,480 --> 00:46:41,400
climbing and we just said that we can't

1250
00:46:38,720 --> 00:46:43,520
use the word climbing um or what if we

1251
00:46:41,400 --> 00:46:45,480
sort of generate other related terms

1252
00:46:43,520 --> 00:46:47,520
before we get to the restricted term

1253
00:46:45,480 --> 00:46:49,359
like the model starts suggesting maybe

1254
00:46:47,520 --> 00:46:51,480
you can work out by going to an indoor

1255
00:46:49,359 --> 00:46:52,920
rock blank and then what are we going to

1256
00:46:51,480 --> 00:46:54,800
say there's not we can't say rock

1257
00:46:52,920 --> 00:46:57,079
climbing so maybe the model suggests

1258
00:46:54,800 --> 00:46:58,640
rock climbing is rock collecting is a

1259
00:46:57,079 --> 00:47:01,400
hobby to stay in shape and that doesn't

1260
00:46:58,640 --> 00:47:03,480
sound good either um you could continue

1261
00:47:01,400 --> 00:47:05,640
like sort of engineering more and more

1262
00:47:03,480 --> 00:47:06,599
complicated rules here but maybe we

1263
00:47:05,640 --> 00:47:08,760
could do something that's a little

1264
00:47:06,599 --> 00:47:10,559
simpler so what if I just sample a bunch

1265
00:47:08,760 --> 00:47:11,920
of outputs from the model and then I

1266
00:47:10,559 --> 00:47:14,359
check if they're about climbing and I

1267
00:47:11,920 --> 00:47:16,280
get rid of them if they are right um

1268
00:47:14,359 --> 00:47:18,200
this is sort of the advantage that it's

1269
00:47:16,280 --> 00:47:19,599
pretty easy to check after the fact if

1270
00:47:18,200 --> 00:47:22,480
the sequence has satisfied this

1271
00:47:19,599 --> 00:47:24,400
constraint you know we could train some

1272
00:47:22,480 --> 00:47:26,200
smaller model to guess if the topic of a

1273
00:47:24,400 --> 00:47:27,960
sentence is about climbing could check

1274
00:47:26,200 --> 00:47:30,040
for keywords we could have a friend

1275
00:47:27,960 --> 00:47:31,359
who's willing to see this content like

1276
00:47:30,040 --> 00:47:33,040
filter through it and throw everything

1277
00:47:31,359 --> 00:47:36,480
out that's not about climing that is

1278
00:47:33,040 --> 00:47:38,280
about climbing but if this model um

1279
00:47:36,480 --> 00:47:40,119
ascribes really high likelihood to this

1280
00:47:38,280 --> 00:47:42,559
like if this model was trained on you

1281
00:47:40,119 --> 00:47:44,760
know data from CS PhD students this

1282
00:47:42,559 --> 00:47:46,240
could be an extremely high likelihood

1283
00:47:44,760 --> 00:47:48,319
suggestion and so we might need to

1284
00:47:46,240 --> 00:47:49,839
regenerate hundreds or thousands of

1285
00:47:48,319 --> 00:47:52,559
sequences to find something that's not

1286
00:47:49,839 --> 00:47:55,240
about climing um and that feels a little

1287
00:47:52,559 --> 00:47:56,920
bit inefficient right so is there

1288
00:47:55,240 --> 00:47:59,040
something that we can do that's a little

1289
00:47:56,920 --> 00:48:01,599
bit better than that well really we'd

1290
00:47:59,040 --> 00:48:03,200
like to guess at some point during our

1291
00:48:01,599 --> 00:48:05,200
generation if the sequence is going to

1292
00:48:03,200 --> 00:48:08,000
be about climbing and maybe like

1293
00:48:05,200 --> 00:48:10,640
recalibrate or you know we could even

1294
00:48:08,000 --> 00:48:12,079
restart or sort of shape Our Generations

1295
00:48:10,640 --> 00:48:14,520
so that we don't wind up with a sequence

1296
00:48:12,079 --> 00:48:16,319
that's about climbing in the first place

1297
00:48:14,520 --> 00:48:19,359
um one of the methods that we'll discuss

1298
00:48:16,319 --> 00:48:20,920
to do this is a method called fudge um

1299
00:48:19,359 --> 00:48:22,800
and unfortunately in their paper they

1300
00:48:20,920 --> 00:48:24,240
don't have the same anti-climbing bias I

1301
00:48:22,800 --> 00:48:27,000
do so this example is actually about

1302
00:48:24,240 --> 00:48:29,000
formality instead um the idea here is

1303
00:48:27,000 --> 00:48:32,079
that we want a sequence output of the

1304
00:48:29,000 --> 00:48:34,079
model that is sort of satisfies this

1305
00:48:32,079 --> 00:48:36,079
constraint of being formal and the way

1306
00:48:34,079 --> 00:48:39,960
we're going to do this is at each step

1307
00:48:36,079 --> 00:48:41,640
of prediction we get the outputs of what

1308
00:48:39,960 --> 00:48:44,160
the model predicts is the next token

1309
00:48:41,640 --> 00:48:47,319
right this sort of distribution here in

1310
00:48:44,160 --> 00:48:49,760
blue and we also have some second

1311
00:48:47,319 --> 00:48:52,079
distribution which says given sort of

1312
00:48:49,760 --> 00:48:54,480
what we have so far How likely is this

1313
00:48:52,079 --> 00:48:56,920
to be a formal sentence at the end right

1314
00:48:54,480 --> 00:48:58,880
does a sentence that starts do you want

1315
00:48:56,920 --> 00:49:01,200
have a high likelihood of being formal

1316
00:48:58,880 --> 00:49:04,559
versus a sentence that starts do you

1317
00:49:01,200 --> 00:49:07,200
prefer and so this sort of guess at what

1318
00:49:04,559 --> 00:49:09,520
will be formal at the end of the um

1319
00:49:07,200 --> 00:49:10,960
generation will put High likelihood on

1320
00:49:09,520 --> 00:49:13,599
things that result in really formal

1321
00:49:10,960 --> 00:49:15,880
sentences like do you prefer or do you

1322
00:49:13,599 --> 00:49:17,200
thus whereas the original model might

1323
00:49:15,880 --> 00:49:19,440
have higher likelihood on things that

1324
00:49:17,200 --> 00:49:22,559
are maybe more commonly said like do you

1325
00:49:19,440 --> 00:49:24,319
want um so we combine these two

1326
00:49:22,559 --> 00:49:26,280
distributions you can just multiply them

1327
00:49:24,319 --> 00:49:29,079
together and then we sample from this

1328
00:49:26,280 --> 00:49:30,520
modified distribution which now has some

1329
00:49:29,079 --> 00:49:32,359
sort of high weight on things that the

1330
00:49:30,520 --> 00:49:33,559
model thinks are likely but also takes

1331
00:49:32,359 --> 00:49:35,960
into account the likelihood of

1332
00:49:33,559 --> 00:49:38,240
satisfying a constraint um this is

1333
00:49:35,960 --> 00:49:40,640
another sort of method of modifying or

1334
00:49:38,240 --> 00:49:42,520
sampling distribution um with some

1335
00:49:40,640 --> 00:49:44,520
external information here and so there's

1336
00:49:42,520 --> 00:49:47,440
results and sequences that wind up being

1337
00:49:44,520 --> 00:49:48,799
sort of more likely to be formal without

1338
00:49:47,440 --> 00:49:50,280
having to sample a whole bunch of

1339
00:49:48,799 --> 00:49:52,880
sentences and reject the ones that we

1340
00:49:50,280 --> 00:49:54,720
think don't satisfy this constraint so

1341
00:49:52,880 --> 00:49:57,119
how do we get sort of a guess of what

1342
00:49:54,720 --> 00:49:58,839
will be formal at the end of Generation

1343
00:49:57,119 --> 00:50:01,319
Um this is where the name fudge comes

1344
00:49:58,839 --> 00:50:03,319
from the fud stands for future

1345
00:50:01,319 --> 00:50:06,640
discriminator and so what they do is

1346
00:50:03,319 --> 00:50:08,920
they train a model on prefixes to guess

1347
00:50:06,640 --> 00:50:10,400
whether that sequence will be formal um

1348
00:50:08,920 --> 00:50:12,040
you can do this if you have a bunch of

1349
00:50:10,400 --> 00:50:15,319
data that's sort of sorted into formal

1350
00:50:12,040 --> 00:50:17,720
and not formal right every um sort of

1351
00:50:15,319 --> 00:50:20,119
prefix of a sentence in the formal

1352
00:50:17,720 --> 00:50:21,480
category is a training example right you

1353
00:50:20,119 --> 00:50:23,720
know a sentence that starts do you

1354
00:50:21,480 --> 00:50:27,599
prefer you can shop off each token to

1355
00:50:23,720 --> 00:50:29,920
get sort of a um set of sequ of prefixes

1356
00:50:27,599 --> 00:50:31,160
to sequences that have the label formal

1357
00:50:29,920 --> 00:50:33,559
and you can do the same thing to your

1358
00:50:31,160 --> 00:50:34,920
informal set and train a discriminator

1359
00:50:33,559 --> 00:50:36,559
to choose between them to say like

1360
00:50:34,920 --> 00:50:38,400
what's the probability the sentence but

1361
00:50:36,559 --> 00:50:41,160
will belong to the formal set when we

1362
00:50:38,400 --> 00:50:43,319
finish and so this idea of sort of

1363
00:50:41,160 --> 00:50:44,359
trying to guess at a given decoding step

1364
00:50:43,319 --> 00:50:49,480
if we're going to wind up with our

1365
00:50:44,359 --> 00:50:50,799
constraints satisfied at the end um is a

1366
00:50:49,480 --> 00:50:53,000
sort of key way to do constraint

1367
00:50:50,799 --> 00:50:56,000
decoding um and one that we'll return to

1368
00:50:53,000 --> 00:50:58,280
in just a couple slides here

1369
00:50:56,000 --> 00:51:00,440
I want to talk touch on something

1370
00:50:58,280 --> 00:51:03,079
slightly different which is that maybe

1371
00:51:00,440 --> 00:51:04,599
one of the constraints we care about is

1372
00:51:03,079 --> 00:51:07,319
something a little more nebulous like we

1373
00:51:04,599 --> 00:51:09,160
want to match human preference um the

1374
00:51:07,319 --> 00:51:12,079
way that we usually accomplish this

1375
00:51:09,160 --> 00:51:14,920
constraint is a little bit different

1376
00:51:12,079 --> 00:51:16,040
right um this we' usually do through

1377
00:51:14,920 --> 00:51:18,839
like reinforcement learning through

1378
00:51:16,040 --> 00:51:21,559
human feedback um and so we take sort of

1379
00:51:18,839 --> 00:51:24,960
our original model distribution and we

1380
00:51:21,559 --> 00:51:27,960
take a sort of really like tight like

1381
00:51:24,960 --> 00:51:30,200
distrib tion of evidence that says like

1382
00:51:27,960 --> 00:51:31,680
um this model says that this sequence is

1383
00:51:30,200 --> 00:51:33,960
really high reward this sequence is

1384
00:51:31,680 --> 00:51:35,640
really low reward and we try to sort of

1385
00:51:33,960 --> 00:51:38,200
combine them somehow through training so

1386
00:51:35,640 --> 00:51:41,240
we get a new model that is um quote

1387
00:51:38,200 --> 00:51:43,240
unquote aligned and that it has like a

1388
00:51:41,240 --> 00:51:45,280
higher likelihood of giving us things

1389
00:51:43,240 --> 00:51:48,640
that have really high reward according

1390
00:51:45,280 --> 00:51:51,319
to our reward distribution um you can

1391
00:51:48,640 --> 00:51:53,599
view this though as a type of basian

1392
00:51:51,319 --> 00:51:55,119
inference and so what this means is the

1393
00:51:53,599 --> 00:51:57,440
distribution that we really want to get

1394
00:51:55,119 --> 00:51:59,880
at the end is a distribution that

1395
00:51:57,440 --> 00:52:03,160
combines our original models

1396
00:51:59,880 --> 00:52:05,680
distribution and some idea of like How

1397
00:52:03,160 --> 00:52:08,480
likely we are to satisfy the reward

1398
00:52:05,680 --> 00:52:10,720
right um this we do through

1399
00:52:08,480 --> 00:52:12,359
reinforcement learning but if we sort of

1400
00:52:10,720 --> 00:52:14,480
know what these two distributions look

1401
00:52:12,359 --> 00:52:16,119
like we've we've just been talking about

1402
00:52:14,480 --> 00:52:17,680
a lot of methods that modify the

1403
00:52:16,119 --> 00:52:20,119
original models distribution with

1404
00:52:17,680 --> 00:52:21,880
external information it seems like maybe

1405
00:52:20,119 --> 00:52:24,760
we could just add that external

1406
00:52:21,880 --> 00:52:26,200
information in at decoding time to get

1407
00:52:24,760 --> 00:52:29,040
some of the same

1408
00:52:26,200 --> 00:52:31,040
effects um and it turns out you can do

1409
00:52:29,040 --> 00:52:32,799
exactly this so this is a paper from

1410
00:52:31,040 --> 00:52:36,680
last year called reward augmented

1411
00:52:32,799 --> 00:52:39,079
decoding and the idea here is sort of um

1412
00:52:36,680 --> 00:52:41,839
in the same conceptual class as fudge

1413
00:52:39,079 --> 00:52:44,079
but instead of um predicting whether

1414
00:52:41,839 --> 00:52:46,079
we're likely to satisfy the constraint

1415
00:52:44,079 --> 00:52:47,599
we're predicting how much reward we

1416
00:52:46,079 --> 00:52:49,880
think that sequence will have at the end

1417
00:52:47,599 --> 00:52:52,599
of generation so we take our original

1418
00:52:49,880 --> 00:52:54,839
model without doing any rhf and we get

1419
00:52:52,599 --> 00:52:58,160
the output we get the predictions for

1420
00:52:54,839 --> 00:52:59,400
the next token and then we use a model

1421
00:52:58,160 --> 00:53:02,359
that's been trained to predict the

1422
00:52:59,400 --> 00:53:05,040
likely reward given some prefix like a

1423
00:53:02,359 --> 00:53:06,720
future discriminator and we calculate

1424
00:53:05,040 --> 00:53:08,200
the likely reward if we pick each of

1425
00:53:06,720 --> 00:53:09,799
those tokens and then we use the

1426
00:53:08,200 --> 00:53:12,319
combination of those two distributions

1427
00:53:09,799 --> 00:53:13,720
to choose what to decode next um and

1428
00:53:12,319 --> 00:53:16,000
this sort of gives you some of the

1429
00:53:13,720 --> 00:53:18,440
benefits of rlf without actually having

1430
00:53:16,000 --> 00:53:21,200
to do reinforcement learning so it's a

1431
00:53:18,440 --> 00:53:23,160
way of treating like aligning to human

1432
00:53:21,200 --> 00:53:26,839
feedback as just another constraint that

1433
00:53:23,160 --> 00:53:30,400
you can impose at decoding point

1434
00:53:26,839 --> 00:53:32,319
so those were sort of a a subset of the

1435
00:53:30,400 --> 00:53:34,280
um constrains decoding strategies that

1436
00:53:32,319 --> 00:53:35,799
people use um before we get into the

1437
00:53:34,280 --> 00:53:38,400
human and the loop stack are there any

1438
00:53:35,799 --> 00:53:38,400
questions on

1439
00:53:39,040 --> 00:53:43,599
this yes for

1440
00:53:44,960 --> 00:53:48,319
the do you have

1441
00:53:52,799 --> 00:53:57,440
to right so for the discrimin do you

1442
00:53:55,640 --> 00:54:00,000
need to train one for every constraint

1443
00:53:57,440 --> 00:54:01,440
and you do yeah so you need to have some

1444
00:54:00,000 --> 00:54:02,920
set of data that satisfies your

1445
00:54:01,440 --> 00:54:05,319
constraint and some set of data that

1446
00:54:02,920 --> 00:54:08,200
doesn't before you can enforce a new

1447
00:54:05,319 --> 00:54:10,200
constraint in an alternative might be

1448
00:54:08,200 --> 00:54:12,040
like in the paper that's what they did

1449
00:54:10,200 --> 00:54:16,400
but an alternative might be just to

1450
00:54:12,040 --> 00:54:18,359
train a discriminator to determine

1451
00:54:16,400 --> 00:54:20,880
whether any constraint was violated so

1452
00:54:18,359 --> 00:54:23,359
if you have 100 constraints you could do

1453
00:54:20,880 --> 00:54:25,599
a binary prier about whether any

1454
00:54:23,359 --> 00:54:26,880
constraint is violated and then

1455
00:54:25,599 --> 00:54:29,040
also

1456
00:54:26,880 --> 00:54:30,559
sufficient but if you wanted to add a

1457
00:54:29,040 --> 00:54:34,079
new constraint you'd still have to

1458
00:54:30,559 --> 00:54:34,079
retrain or you have to retrain

1459
00:54:35,160 --> 00:54:41,319
or the the reason that this is sort of

1460
00:54:38,119 --> 00:54:43,119
relatively reasonable to do is that this

1461
00:54:41,319 --> 00:54:45,240
determination of if a constraint is

1462
00:54:43,119 --> 00:54:46,960
likely to be violated is sort of a a

1463
00:54:45,240 --> 00:54:48,520
lighter weight or an easier task to

1464
00:54:46,960 --> 00:54:50,520
learn you can use a relatively small

1465
00:54:48,520 --> 00:54:52,079
model for this versus like your big

1466
00:54:50,520 --> 00:54:53,680
model just that has to be able to

1467
00:54:52,079 --> 00:54:55,920
predict the next token for any sequence

1468
00:54:53,680 --> 00:54:58,400
anymore yeah another another like

1469
00:54:55,920 --> 00:55:00,760
interesting thing is if you think about

1470
00:54:58,400 --> 00:55:01,520
it normally you're predicting with your

1471
00:55:00,760 --> 00:55:04,119
big

1472
00:55:01,520 --> 00:55:06,359
softmax like this over all of your

1473
00:55:04,119 --> 00:55:09,680
vocabulary you can even use the same

1474
00:55:06,359 --> 00:55:11,920
representations here to predict with a

1475
00:55:09,680 --> 00:55:13,359
binary classifier uh whether the

1476
00:55:11,920 --> 00:55:14,559
constraint is violated let's say you

1477
00:55:13,359 --> 00:55:17,240
have 100

1478
00:55:14,559 --> 00:55:19,240
constraints this is still a vector of

1479
00:55:17,240 --> 00:55:21,520
size 100 compared to your vector of size

1480
00:55:19,240 --> 00:55:26,240
32,000 that you're using for llama right

1481
00:55:21,520 --> 00:55:28,280
so it's not like this adds the training

1482
00:55:26,240 --> 00:55:32,799
would cost some time but it adds very

1483
00:55:28,280 --> 00:55:32,799
little like inference time I guess

1484
00:55:33,440 --> 00:55:38,960
basically the rock

1485
00:55:35,880 --> 00:55:41,400
sound so when you do the constraint you

1486
00:55:38,960 --> 00:55:43,160
use like a more General

1487
00:55:41,400 --> 00:55:44,680
like do

1488
00:55:43,160 --> 00:55:48,160
notest

1489
00:55:44,680 --> 00:55:50,799
or I guess like in that constraint for

1490
00:55:48,160 --> 00:55:50,799
you can add

1491
00:55:52,559 --> 00:55:57,000
like, is there

1492
00:55:57,880 --> 00:56:00,720
like is there a way to generalize your

1493
00:55:59,400 --> 00:56:04,760
constraint would be like don't talk

1494
00:56:00,720 --> 00:56:07,039
about this whole set of hobes um you

1495
00:56:04,760 --> 00:56:08,960
could do that by training a

1496
00:56:07,039 --> 00:56:10,400
discriminator um by training one

1497
00:56:08,960 --> 00:56:12,359
discriminator that considers all of

1498
00:56:10,400 --> 00:56:15,119
those or by training like a hundred

1499
00:56:12,359 --> 00:56:17,559
different discriminators and then um

1500
00:56:15,119 --> 00:56:19,520
sort of taking like the maximum score

1501
00:56:17,559 --> 00:56:21,240
from any of them right like you want to

1502
00:56:19,520 --> 00:56:23,240
you want to be able to exclude all of

1503
00:56:21,240 --> 00:56:27,799
these things so you consider if any of

1504
00:56:23,240 --> 00:56:30,720
them are violated yeah and for um reward

1505
00:56:27,799 --> 00:56:32,839
augmented recoding how do we sort of

1506
00:56:30,720 --> 00:56:36,039
like frame that reward model or is that

1507
00:56:32,839 --> 00:56:38,400
just come from the previously done rhf

1508
00:56:36,039 --> 00:56:41,079
data that the store from there and then

1509
00:56:38,400 --> 00:56:44,119
you sort of like FR another

1510
00:56:41,079 --> 00:56:47,880
discriminator but this one

1511
00:56:44,119 --> 00:56:50,799
now I I fully understand yeah so how do

1512
00:56:47,880 --> 00:56:52,920
we get the the reward model here this is

1513
00:56:50,799 --> 00:56:55,280
we can use the same data that we' use

1514
00:56:52,920 --> 00:56:58,000
for rhf but we need a slightly different

1515
00:56:55,280 --> 00:57:01,119
model so for rhf we'll train a reward

1516
00:56:58,000 --> 00:57:02,599
model over full sequences right and here

1517
00:57:01,119 --> 00:57:05,280
we need to do the same trick where we

1518
00:57:02,599 --> 00:57:07,280
sort of look at just prefixes and try to

1519
00:57:05,280 --> 00:57:09,640
guess the reward Downstream but if we

1520
00:57:07,280 --> 00:57:12,440
have already have preference data then

1521
00:57:09,640 --> 00:57:15,119
we have some um like we have a data

1522
00:57:12,440 --> 00:57:16,720
source to do this with I think if I'm

1523
00:57:15,119 --> 00:57:19,240
remembering correctly they also had a

1524
00:57:16,720 --> 00:57:20,920
couple more sort of tricks for data

1525
00:57:19,240 --> 00:57:22,640
augmentation to get this to work this is

1526
00:57:20,920 --> 00:57:25,720
sort of like a non-trivial thing to

1527
00:57:22,640 --> 00:57:28,039
figure out um because like reward is

1528
00:57:25,720 --> 00:57:30,200
generally a secret bual

1529
00:57:28,039 --> 00:57:32,280
attribute and also if you don't know

1530
00:57:30,200 --> 00:57:34,160
very much about rhf we're going to cover

1531
00:57:32,280 --> 00:57:36,400
that the future class so don't worry if

1532
00:57:34,160 --> 00:57:37,880
this is a yeah sorry to Jump Ahead a

1533
00:57:36,400 --> 00:57:39,880
little no no

1534
00:57:37,880 --> 00:57:43,640
wores

1535
00:57:39,880 --> 00:57:47,240
yeah application this like why would we

1536
00:57:43,640 --> 00:57:49,640
doing this to ensure it could be like

1537
00:57:47,240 --> 00:57:52,839
our llm would want to highlight certain

1538
00:57:49,640 --> 00:57:53,799
qualities like we want our evence to be

1539
00:57:52,839 --> 00:57:55,960
more

1540
00:57:53,799 --> 00:57:57,839
empathetic is there

1541
00:57:55,960 --> 00:57:59,440
something yeah like what are the real

1542
00:57:57,839 --> 00:58:01,280
world applications like could we use

1543
00:57:59,440 --> 00:58:03,680
this to make L more empathetic or

1544
00:58:01,280 --> 00:58:06,359
something yeah any any real attribute

1545
00:58:03,680 --> 00:58:08,000
that you can sort of collect like

1546
00:58:06,359 --> 00:58:09,839
positive and negative data for you could

1547
00:58:08,000 --> 00:58:12,200
do this kind of constraints for I think

1548
00:58:09,839 --> 00:58:15,119
the the ones you see most commonly are

1549
00:58:12,200 --> 00:58:16,480
the human preference and then like

1550
00:58:15,119 --> 00:58:18,839
negative constraints like you don't want

1551
00:58:16,480 --> 00:58:20,000
your model to generate offensive content

1552
00:58:18,839 --> 00:58:21,839
and if you can build like a good

1553
00:58:20,000 --> 00:58:23,319
discriminator for is a sentence going in

1554
00:58:21,839 --> 00:58:26,160
a really offensive Direction you can

1555
00:58:23,319 --> 00:58:28,440
kind of stop it from gener

1556
00:58:26,160 --> 00:58:30,480
yeah would it be a good idea if you

1557
00:58:28,440 --> 00:58:33,760
generate a bunch of cons and ask the

1558
00:58:30,480 --> 00:58:35,480
model itself whether it violates the

1559
00:58:33,760 --> 00:58:37,319
yeah you could do that for sure could

1560
00:58:35,480 --> 00:58:38,920
you ask like could you generate a bunch

1561
00:58:37,319 --> 00:58:42,440
of samples and ask the model if it

1562
00:58:38,920 --> 00:58:44,720
violates the constraint um this is also

1563
00:58:42,440 --> 00:58:47,119
a type of sort of sample and then rerank

1564
00:58:44,720 --> 00:58:52,319
strategy um but yeah this would be sort

1565
00:58:47,119 --> 00:58:54,000
of a more um clever like less

1566
00:58:52,319 --> 00:58:55,559
heavyweight version of this checking if

1567
00:58:54,000 --> 00:58:57,319
it's about climate means right you'd

1568
00:58:55,559 --> 00:58:58,520
like ask the model if it violated the

1569
00:58:57,319 --> 00:59:00,160
constraint and if it's a good enough

1570
00:58:58,520 --> 00:59:02,480
model it could probably do that pretty

1571
00:59:00,160 --> 00:59:05,160
well I suppose in that case you don't

1572
00:59:02,480 --> 00:59:08,160
have to thing anything yeah yeah and

1573
00:59:05,160 --> 00:59:10,359
this is sort of a general like the

1574
00:59:08,160 --> 00:59:12,240
generating text that like satisfies a

1575
00:59:10,359 --> 00:59:14,079
constraint is harder than checking if a

1576
00:59:12,240 --> 00:59:16,280
text satisfies a constraint so even if

1577
00:59:14,079 --> 00:59:17,880
the model isn't good about like not

1578
00:59:16,280 --> 00:59:19,440
generating text about climbing when you

1579
00:59:17,880 --> 00:59:20,520
tell it to it might be able to tell if

1580
00:59:19,440 --> 00:59:23,640
text is

1581
00:59:20,520 --> 00:59:26,640
about yeah yeah so how do

1582
00:59:23,640 --> 00:59:26,640
you

1583
00:59:28,400 --> 00:59:32,359
have different

1584
00:59:32,920 --> 00:59:36,319
different you have

1585
00:59:36,599 --> 00:59:42,119
to yeah like how do you collect the data

1586
00:59:38,839 --> 00:59:45,720
to train this discriminator um generally

1587
00:59:42,119 --> 00:59:47,160
you're going to see like you'll look to

1588
00:59:45,720 --> 00:59:48,720
see if there are data sets that already

1589
00:59:47,160 --> 00:59:50,160
captured this attribute or you could

1590
00:59:48,720 --> 00:59:51,599
sort of write her istics to try to

1591
00:59:50,160 --> 00:59:53,839
recover it if it's an attribute that not

1592
00:59:51,599 --> 00:59:55,480
a lot of other people care about like

1593
00:59:53,839 --> 00:59:58,280
you could write your puristic to check

1594
00:59:55,480 --> 01:00:00,160
if text is about climbing for instance

1595
00:59:58,280 --> 01:00:02,359
um and then try to recover what noisy

1596
01:00:00,160 --> 01:00:04,200
samples of data that is or is not about

1597
01:00:02,359 --> 01:00:05,559
climbing maybe you could scrape a

1598
01:00:04,200 --> 01:00:07,000
climbing forum and then scrape like a

1599
01:00:05,559 --> 01:00:09,079
hiking forum and use the difference

1600
01:00:07,000 --> 01:00:10,319
between them um but for a lot of tests

1601
01:00:09,079 --> 01:00:11,760
there's actually pretty good data sets

1602
01:00:10,319 --> 01:00:14,400
already out there for this so there's

1603
01:00:11,760 --> 01:00:17,480
like in there's a lot of style transfer

1604
01:00:14,400 --> 01:00:20,200
tasks that are like go from informal to

1605
01:00:17,480 --> 01:00:22,240
formal or go from this to that or like

1606
01:00:20,200 --> 01:00:24,039
make this text in an iic contamin and

1607
01:00:22,240 --> 01:00:26,559
you can find like data from those

1608
01:00:24,039 --> 01:00:26,559
sources

1609
01:00:26,799 --> 01:00:31,599
we never like talked about F yet but I'm

1610
01:00:29,520 --> 01:00:34,520
really curious with like the word a

1611
01:00:31,599 --> 01:00:38,039
beting whether it would perform better

1612
01:00:34,520 --> 01:00:39,079
than like fineing on RF like certainly

1613
01:00:38,039 --> 01:00:42,720
more

1614
01:00:39,079 --> 01:00:45,039
efficient but I I was I think this is a

1615
01:00:42,720 --> 01:00:49,760
comparison they make in their paper but

1616
01:00:45,039 --> 01:00:52,520
I don't remember their pun on yeah um in

1617
01:00:49,760 --> 01:00:55,280
general there's this sort of a like you

1618
01:00:52,520 --> 01:00:57,039
can pay a onetime kind of heavy cost to

1619
01:00:55,280 --> 01:00:58,880
fine-tune or you can pay costs at

1620
01:00:57,039 --> 01:01:01,160
inference time every time to make sort

1621
01:00:58,880 --> 01:01:03,880
of a to make your model better in any of

1622
01:01:01,160 --> 01:01:06,160
these ways and depending on how much

1623
01:01:03,880 --> 01:01:09,119
inference you're playing do like one or

1624
01:01:06,160 --> 01:01:09,119
the other of these could be

1625
01:01:11,240 --> 01:01:16,400
better

1626
01:01:12,839 --> 01:01:19,200
great so now we're going to talk about

1627
01:01:16,400 --> 01:01:21,160
sort of methods for introducing human

1628
01:01:19,200 --> 01:01:22,680
interaction into the decoding process

1629
01:01:21,160 --> 01:01:25,240
and everything we've looked at so far

1630
01:01:22,680 --> 01:01:26,920
has been very sort of black booss kind

1631
01:01:25,240 --> 01:01:28,920
of hands off right like you give the

1632
01:01:26,920 --> 01:01:30,640
model M some input maybe we do some kind

1633
01:01:28,920 --> 01:01:33,640
of manipulation on the decoding side you

1634
01:01:30,640 --> 01:01:37,160
get one output back right um but in a

1635
01:01:33,640 --> 01:01:38,920
lot of situations where maybe you have

1636
01:01:37,160 --> 01:01:40,960
some high-risk application and you need

1637
01:01:38,920 --> 01:01:42,640
somebody to be consistently monitoring

1638
01:01:40,960 --> 01:01:43,799
and maybe intervening or you're doing

1639
01:01:42,640 --> 01:01:46,359
something where you want to do some kind

1640
01:01:43,799 --> 01:01:47,960
of human AI collaboration um and you

1641
01:01:46,359 --> 01:01:49,160
want to be able to go back and forth or

1642
01:01:47,960 --> 01:01:50,960
you want to have a conversation with the

1643
01:01:49,160 --> 01:01:53,480
model what you're actually doing is sort

1644
01:01:50,960 --> 01:01:54,960
of a series of decodings with human

1645
01:01:53,480 --> 01:01:56,319
intervention in between

1646
01:01:54,960 --> 01:01:58,640
um and I'm going to talk about a couple

1647
01:01:56,319 --> 01:02:00,760
of these strategies briefly I think if

1648
01:01:58,640 --> 01:02:02,200
you've used sort of a modern llm you're

1649
01:02:00,760 --> 01:02:04,440
probably familiar with at least a few of

1650
01:02:02,200 --> 01:02:06,720
them already um we'll sort of put names

1651
01:02:04,440 --> 01:02:08,359
to each of them and the set of examples

1652
01:02:06,720 --> 01:02:10,880
that we're running with here are from a

1653
01:02:08,359 --> 01:02:13,880
paper called wordcraft which is about um

1654
01:02:10,880 --> 01:02:15,480
story generation with llm assistants but

1655
01:02:13,880 --> 01:02:17,559
these can also be applied sort of more

1656
01:02:15,480 --> 01:02:20,319
generally to any kind of task where

1657
01:02:17,559 --> 01:02:23,799
you'd want to go back and forth with a

1658
01:02:20,319 --> 01:02:25,319
model um the sort of easiest or maybe

1659
01:02:23,799 --> 01:02:27,599
simplest place to start here is just

1660
01:02:25,319 --> 01:02:29,760
with interleaving text right you can

1661
01:02:27,599 --> 01:02:31,400
choose when the model starts and stops

1662
01:02:29,760 --> 01:02:33,720
decoding and you can choose when a human

1663
01:02:31,400 --> 01:02:34,920
is writing text in between and you can

1664
01:02:33,720 --> 01:02:36,680
condition your model in sort of a

1665
01:02:34,920 --> 01:02:39,240
mixture of human and model generated

1666
01:02:36,680 --> 01:02:41,279
text to choose what to continue next um

1667
01:02:39,240 --> 01:02:43,680
you can also do something like have the

1668
01:02:41,279 --> 01:02:45,319
model generate a set of text edit that

1669
01:02:43,680 --> 01:02:47,119
text in some way maybe the human is

1670
01:02:45,319 --> 01:02:48,640
imposing some really subtle constraint

1671
01:02:47,119 --> 01:02:50,559
like I want it to sound like my writing

1672
01:02:48,640 --> 01:02:52,200
style we don't have a discriminator for

1673
01:02:50,559 --> 01:02:54,119
this but the human can sort of modify

1674
01:02:52,200 --> 01:02:55,680
the text and then continue generating

1675
01:02:54,119 --> 01:02:57,160
from that point and that will influence

1676
01:02:55,680 --> 01:03:01,160
the style of the text that continues

1677
01:02:57,160 --> 01:03:03,240
being generative um a this case here is

1678
01:03:01,160 --> 01:03:04,720
sort of a you're writing a story

1679
01:03:03,240 --> 01:03:06,520
together and so you're going back and

1680
01:03:04,720 --> 01:03:07,799
forth and editing the text like that but

1681
01:03:06,520 --> 01:03:10,319
you can also think of any kind of

1682
01:03:07,799 --> 01:03:11,920
conversation with a model as the same

1683
01:03:10,319 --> 01:03:15,319
kind of interleaving of text right the

1684
01:03:11,920 --> 01:03:17,000
model gives some um text you provide

1685
01:03:15,319 --> 01:03:18,599
some text you go back and forth on like

1686
01:03:17,000 --> 01:03:20,480
who's providing the text that conditions

1687
01:03:18,599 --> 01:03:23,039
the

1688
01:03:20,480 --> 01:03:24,880
model you also might want to do things

1689
01:03:23,039 --> 01:03:26,760
like more fine brain replace

1690
01:03:24,880 --> 01:03:28,559
so here the person has highlighted some

1691
01:03:26,760 --> 01:03:31,640
text and said like make this more

1692
01:03:28,559 --> 01:03:33,960
descriptive or shorten this to two words

1693
01:03:31,640 --> 01:03:36,079
or maybe you want some additional

1694
01:03:33,960 --> 01:03:38,520
constraint like can this be happier can

1695
01:03:36,079 --> 01:03:40,960
this be sad like change the ending or

1696
01:03:38,520 --> 01:03:43,760
something um you can accomplish this in

1697
01:03:40,960 --> 01:03:45,799
a variety of ways um here this is done

1698
01:03:43,760 --> 01:03:47,680
through input manipulation so you prompt

1699
01:03:45,799 --> 01:03:50,359
your model differently with different

1700
01:03:47,680 --> 01:03:52,200
constraints you can also do this with an

1701
01:03:50,359 --> 01:03:54,440
actual modeling change like if you want

1702
01:03:52,200 --> 01:03:56,119
some kind of infilling model um

1703
01:03:54,440 --> 01:03:57,720
particularly for things like code this

1704
01:03:56,119 --> 01:04:01,119
can be helpful so you want context from

1705
01:03:57,720 --> 01:04:02,440
left and right sides um or you can do

1706
01:04:01,119 --> 01:04:03,799
this with the decoding changes that we

1707
01:04:02,440 --> 01:04:05,960
talked about in the previous section

1708
01:04:03,799 --> 01:04:07,799
right you could add a discriminator for

1709
01:04:05,960 --> 01:04:09,680
descriptiveness of text or you could do

1710
01:04:07,799 --> 01:04:11,680
some kind of sampling ranking method to

1711
01:04:09,680 --> 01:04:13,880
recover a more descriptive

1712
01:04:11,680 --> 01:04:16,640
output another thing that's very common

1713
01:04:13,880 --> 01:04:17,960
in this space is sampling and reranking

1714
01:04:16,640 --> 01:04:20,839
methods where the human is the one

1715
01:04:17,960 --> 01:04:23,640
choosing what to return right so in

1716
01:04:20,839 --> 01:04:25,960
wordcraft you see a set of choices and

1717
01:04:23,640 --> 01:04:28,200
you can choose text to insert but more

1718
01:04:25,960 --> 01:04:30,720
commonly in something like um chat gbt

1719
01:04:28,200 --> 01:04:33,160
or Bard you see this little option to

1720
01:04:30,720 --> 01:04:34,880
regenerate text right you as the human

1721
01:04:33,160 --> 01:04:36,160
can reject the text and say like no I

1722
01:04:34,880 --> 01:04:38,680
don't like this give me a different

1723
01:04:36,160 --> 01:04:41,359
output and this is also sort of a way of

1724
01:04:38,680 --> 01:04:44,079
controlling decoding um just by doing it

1725
01:04:41,359 --> 01:04:46,319
on on a human rather in an algorithmic

1726
01:04:44,079 --> 01:04:49,279
level of course you don't necessarily

1727
01:04:46,319 --> 01:04:51,200
need a human in here and so um some

1728
01:04:49,279 --> 01:04:52,960
recent work has looked at functionally

1729
01:04:51,200 --> 01:04:55,799
using models to make these decisions

1730
01:04:52,960 --> 01:04:57,480
instead um this is a a a prompting paper

1731
01:04:55,799 --> 01:05:00,359
called free of thought which was sort of

1732
01:04:57,480 --> 01:05:02,279
very popular on Twitter last summer um

1733
01:05:00,359 --> 01:05:06,119
and the idea here is that you're going

1734
01:05:02,279 --> 01:05:08,480
to generate um several smaller sequences

1735
01:05:06,119 --> 01:05:11,200
um like a couple of sentences a

1736
01:05:08,480 --> 01:05:13,160
reasoning step or a thought in the paper

1737
01:05:11,200 --> 01:05:14,839
and you're going to use a model to

1738
01:05:13,160 --> 01:05:16,839
choose which ones to continue and you

1739
01:05:14,839 --> 01:05:19,000
can do different sort of constraints

1740
01:05:16,839 --> 01:05:21,960
here like I want to sort of rank this

1741
01:05:19,000 --> 01:05:25,079
set of three or maybe I want to predict

1742
01:05:21,960 --> 01:05:26,839
if any in this set is wrong like is this

1743
01:05:25,079 --> 01:05:29,400
a good reasoning step and if the model

1744
01:05:26,839 --> 01:05:32,240
says no you no longer continue that but

1745
01:05:29,400 --> 01:05:33,559
the idea here is through prompting

1746
01:05:32,240 --> 01:05:35,640
really achieving something that's sort

1747
01:05:33,559 --> 01:05:38,960
of if you squint at it looks a lot like

1748
01:05:35,640 --> 01:05:41,279
beam search right instead of doing a um

1749
01:05:38,960 --> 01:05:43,160
like token level thing and making a

1750
01:05:41,279 --> 01:05:45,079
decision based on likelihood you're

1751
01:05:43,160 --> 01:05:47,880
generating sort of several sentences out

1752
01:05:45,079 --> 01:05:50,599
a time and making a decision based on

1753
01:05:47,880 --> 01:05:52,359
this models feedback right this signal

1754
01:05:50,599 --> 01:05:53,799
from an external source which here is a

1755
01:05:52,359 --> 01:05:55,279
model but could also be a human if

1756
01:05:53,799 --> 01:05:57,920
you're willing willing to sort of wait

1757
01:05:55,279 --> 01:06:01,559
around for them to make the decision and

1758
01:05:57,920 --> 01:06:03,839
so this is a way of sort of giving

1759
01:06:01,559 --> 01:06:06,640
feedback on a broader level than single

1760
01:06:03,839 --> 01:06:09,079
tokens um to guide a decoding process to

1761
01:06:06,640 --> 01:06:09,079
a final

1762
01:06:09,839 --> 01:06:15,079
outut so the last couple of things we'll

1763
01:06:12,760 --> 01:06:17,520
talk about here are sort of practical

1764
01:06:15,079 --> 01:06:19,839
considerations speed choosing decoding

1765
01:06:17,520 --> 01:06:22,599
methods um but I can take any questions

1766
01:06:19,839 --> 01:06:22,599
before that

1767
01:06:23,000 --> 01:06:26,000
to

1768
01:06:26,760 --> 01:06:32,920
great so how do you make this fast and

1769
01:06:30,359 --> 01:06:34,920
in particular if you've ever tried to

1770
01:06:32,920 --> 01:06:36,920
sort of Benchmark performance of a model

1771
01:06:34,920 --> 01:06:38,720
what you realize pretty quickly is that

1772
01:06:36,920 --> 01:06:40,720
the vast majority of time is actually

1773
01:06:38,720 --> 01:06:43,440
spent in decoding you have to generate

1774
01:06:40,720 --> 01:06:45,319
one token at a time you have to sort of

1775
01:06:43,440 --> 01:06:46,920
pass that back through the model to get

1776
01:06:45,319 --> 01:06:51,279
conditioning to generate the next token

1777
01:06:46,920 --> 01:06:53,599
and so this is um generally fairly slow

1778
01:06:51,279 --> 01:06:54,839
um this is sort of a a major impediment

1779
01:06:53,599 --> 01:06:56,359
if you're d to do something like a

1780
01:06:54,839 --> 01:06:57,839
streaming application where you want or

1781
01:06:56,359 --> 01:06:59,559
a chat application where you don't want

1782
01:06:57,839 --> 01:07:03,599
the person to be waiting around for an

1783
01:06:59,559 --> 01:07:06,799
answer um one way to do this is a method

1784
01:07:03,599 --> 01:07:09,160
called Spectra of decoding and this is a

1785
01:07:06,799 --> 01:07:12,599
method where you're using a smaller

1786
01:07:09,160 --> 01:07:14,039
model um not as like we're in contrast

1787
01:07:12,599 --> 01:07:16,240
of decoding right we're using a smaller

1788
01:07:14,039 --> 01:07:17,559
model to decide what not to generate but

1789
01:07:16,240 --> 01:07:20,119
here we're using a smaller model to

1790
01:07:17,559 --> 01:07:21,880
decide be what to generate um and the

1791
01:07:20,119 --> 01:07:24,960
idea here is that most of these tokens

1792
01:07:21,880 --> 01:07:26,480
are maybe not super hard to side it's

1793
01:07:24,960 --> 01:07:27,400
just that occasionally the bigger model

1794
01:07:26,480 --> 01:07:30,240
might want to go in a different

1795
01:07:27,400 --> 01:07:32,920
direction so these green tokens here are

1796
01:07:30,240 --> 01:07:35,160
generated by a smaller model our amateur

1797
01:07:32,920 --> 01:07:37,079
model here and the larger model acts

1798
01:07:35,160 --> 01:07:39,960
largely as a verifier and what it does

1799
01:07:37,079 --> 01:07:43,000
is it checks if the output so far is

1800
01:07:39,960 --> 01:07:44,920
going in a an a Direction that's sort of

1801
01:07:43,000 --> 01:07:46,400
in distribution for the big model like

1802
01:07:44,920 --> 01:07:49,240
something that's within the realm of

1803
01:07:46,400 --> 01:07:50,720
what it might SLE and to there's sort of

1804
01:07:49,240 --> 01:07:52,400
an involved discussion in this paper of

1805
01:07:50,720 --> 01:07:55,200
how you determine if something is in

1806
01:07:52,400 --> 01:07:58,000
distribution um so here the smaller

1807
01:07:55,200 --> 01:08:00,240
models generates like five or six tokens

1808
01:07:58,000 --> 01:08:02,559
that the larger model says okay this

1809
01:08:00,240 --> 01:08:03,680
looks great until it hits a token that

1810
01:08:02,559 --> 01:08:06,079
the larger model would not have

1811
01:08:03,680 --> 01:08:07,920
generated in that circumstance and then

1812
01:08:06,079 --> 01:08:10,279
the larger model rejects that token and

1813
01:08:07,920 --> 01:08:13,000
generates a different token instead so

1814
01:08:10,279 --> 01:08:15,440
you can see here each of these red and

1815
01:08:13,000 --> 01:08:17,600
then blue sections is where the larger

1816
01:08:15,440 --> 01:08:19,400
model has rejected something and has to

1817
01:08:17,600 --> 01:08:21,920
actually autor regressively decode a

1818
01:08:19,400 --> 01:08:24,199
single token by contrast if you were

1819
01:08:21,920 --> 01:08:27,359
doing regular decoding at each

1820
01:08:24,199 --> 01:08:28,799
individual token in this sequence the um

1821
01:08:27,359 --> 01:08:31,640
larger model would have had to make the

1822
01:08:28,799 --> 01:08:35,359
fall forward pass to decoda token so

1823
01:08:31,640 --> 01:08:37,359
here rather than de doing maybe what

1824
01:08:35,359 --> 01:08:39,239
probably like 20ish decoding steps to

1825
01:08:37,359 --> 01:08:41,560
get this full sequence the larger model

1826
01:08:39,239 --> 01:08:43,040
has done about eight decoring steps and

1827
01:08:41,560 --> 01:08:47,560
everything else is able to sort of

1828
01:08:43,040 --> 01:08:49,759
verify a block of tokens at once um this

1829
01:08:47,560 --> 01:08:51,400
sort of idea of like using a smaller

1830
01:08:49,759 --> 01:08:54,120
model as an approximation is pretty

1831
01:08:51,400 --> 01:08:55,839
powerful um and there's some great um

1832
01:08:54,120 --> 01:08:58,159
followup work cons specul decoding and

1833
01:08:55,839 --> 01:08:59,000
sort of ways to do this faster or with

1834
01:08:58,159 --> 01:09:01,520
stronger

1835
01:08:59,000 --> 01:09:04,839
guarantees um but this General concept

1836
01:09:01,520 --> 01:09:06,920
is I would bet probably how models like

1837
01:09:04,839 --> 01:09:09,080
um part of how models like chat GPT or

1838
01:09:06,920 --> 01:09:11,159
Bard are sort of generating text so

1839
01:09:09,080 --> 01:09:13,120
quickly um there's another element here

1840
01:09:11,159 --> 01:09:16,159
which is like the model architecture

1841
01:09:13,120 --> 01:09:17,679
being sparse but I think that um if you

1842
01:09:16,159 --> 01:09:19,920
folks talk about mixture of experts we

1843
01:09:17,679 --> 01:09:22,880
might get into that

1844
01:09:19,920 --> 01:09:26,080
later um how do you do this kind of fast

1845
01:09:22,880 --> 01:09:27,679
inference um libraries like BLM will

1846
01:09:26,080 --> 01:09:29,440
Implement things I think Implement

1847
01:09:27,679 --> 01:09:32,199
speculative decoding and Implement sort

1848
01:09:29,440 --> 01:09:34,400
of Hardware level tricks like choosing

1849
01:09:32,199 --> 01:09:37,799
which attention um weights to Cash wear

1850
01:09:34,400 --> 01:09:39,199
to do faster inflence um there's also

1851
01:09:37,799 --> 01:09:40,799
great libraries for doing things like

1852
01:09:39,199 --> 01:09:42,679
constraint decoding so things like

1853
01:09:40,799 --> 01:09:45,520
outlines will let you set constraints

1854
01:09:42,679 --> 01:09:46,960
like I want my outputs to all be Json

1855
01:09:45,520 --> 01:09:48,640
and it will impose additional

1856
01:09:46,960 --> 01:09:50,839
constraints during decoding to ensure

1857
01:09:48,640 --> 01:09:52,279
that that happens and then pretty much

1858
01:09:50,839 --> 01:09:53,960
anything in these first couple of

1859
01:09:52,279 --> 01:09:56,560
sections we talked about um like

1860
01:09:53,960 --> 01:09:58,440
sampling mode seeking search and

1861
01:09:56,560 --> 01:10:00,400
sometimes MBR will also be implemented

1862
01:09:58,440 --> 01:10:05,080
in pretty much any Library you use for

1863
01:10:00,400 --> 01:10:07,679
models like huggingface Fair seek or

1864
01:10:05,080 --> 01:10:10,000
Jacks so to kind of take a step back

1865
01:10:07,679 --> 01:10:12,520
here is when you get to the end of class

1866
01:10:10,000 --> 01:10:15,640
um there's really two broad categories

1867
01:10:12,520 --> 01:10:17,679
of methods that we talked about today um

1868
01:10:15,640 --> 01:10:20,360
given our initial distribution from the

1869
01:10:17,679 --> 01:10:22,600
model for a next token given our our

1870
01:10:20,360 --> 01:10:24,920
input we can do two kind of different

1871
01:10:22,600 --> 01:10:26,400
things we can each individual decoding

1872
01:10:24,920 --> 01:10:28,360
step choose some kind of function to

1873
01:10:26,400 --> 01:10:30,280
manipulate this distribution and this

1874
01:10:28,360 --> 01:10:32,280
could be something like short like

1875
01:10:30,280 --> 01:10:33,960
cutting off the long tail like modifying

1876
01:10:32,280 --> 01:10:36,239
the temperature or adding external

1877
01:10:33,960 --> 01:10:38,400
information from another model or from a

1878
01:10:36,239 --> 01:10:41,480
discriminator model

1879
01:10:38,400 --> 01:10:43,159
right or we can over a larger part of

1880
01:10:41,480 --> 01:10:45,120
the decoding process choose some

1881
01:10:43,159 --> 01:10:47,120
function to choose between sequences and

1882
01:10:45,120 --> 01:10:49,199
this could be like choosing between next

1883
01:10:47,120 --> 01:10:51,679
tokens in beam search when we pruning

1884
01:10:49,199 --> 01:10:53,120
beams this could be choosing from Full

1885
01:10:51,679 --> 01:10:56,760
sequences when we're doing something

1886
01:10:53,120 --> 01:10:58,040
like MB r or sample and rerank methods

1887
01:10:56,760 --> 01:11:00,239
um and you can do these two things in

1888
01:10:58,040 --> 01:11:01,440
parallel right you can choose like a

1889
01:11:00,239 --> 01:11:03,159
different function to manipulate the

1890
01:11:01,440 --> 01:11:04,760
next token distribution and then some

1891
01:11:03,159 --> 01:11:06,199
sort of like broader thing to choose

1892
01:11:04,760 --> 01:11:08,280
what you do with the full sequences you

1893
01:11:06,199 --> 01:11:09,920
get out of that distribution um but

1894
01:11:08,280 --> 01:11:12,040
there are sort of these two broad

1895
01:11:09,920 --> 01:11:14,880
categories of

1896
01:11:12,040 --> 01:11:17,440
decoding so what should you take away

1897
01:11:14,880 --> 01:11:19,400
from this um I think a couple of things

1898
01:11:17,440 --> 01:11:21,000
you decoding methods can be really

1899
01:11:19,400 --> 01:11:23,040
powerful to control features of your

1900
01:11:21,000 --> 01:11:25,040
output if you want to impose particular

1901
01:11:23,040 --> 01:11:26,679
constraints if you want to factor in

1902
01:11:25,040 --> 01:11:27,960
reward function or factor in a data

1903
01:11:26,679 --> 01:11:31,800
source that you maybe didn't have at

1904
01:11:27,960 --> 01:11:34,239
training time um and to some extent you

1905
01:11:31,800 --> 01:11:36,120
can do a more expensive decoding method

1906
01:11:34,239 --> 01:11:37,520
to compensate for a worse model or to

1907
01:11:36,120 --> 01:11:39,080
compensate for a model that hasn't been

1908
01:11:37,520 --> 01:11:42,480
trained to do exactly the thing you want

1909
01:11:39,080 --> 01:11:44,800
it to do um of course you can't you know

1910
01:11:42,480 --> 01:11:47,679
use this to make gpt2 small as good as

1911
01:11:44,800 --> 01:11:49,840
gp4 but you can sort of for some points

1912
01:11:47,679 --> 01:11:51,679
in the middle spend more um computed

1913
01:11:49,840 --> 01:11:53,159
inference time to pay for not spending

1914
01:11:51,679 --> 01:11:55,639
as much computed training time and

1915
01:11:53,159 --> 01:11:57,440
particularly if you don't have access to

1916
01:11:55,639 --> 01:11:59,400
the kind of giant gpus you might need to

1917
01:11:57,440 --> 01:12:01,840
continue fine-tuning your model this can

1918
01:11:59,400 --> 01:12:05,679
be a really a really powerful

1919
01:12:01,840 --> 01:12:07,800
alternative um yeah so say like you're

1920
01:12:05,679 --> 01:12:12,560
building like something in production

1921
01:12:07,800 --> 01:12:15,920
right people usually do um sort of like

1922
01:12:12,560 --> 01:12:18,760
that you know inance before cling to see

1923
01:12:15,920 --> 01:12:21,840
if it's G to work at do

1924
01:12:18,760 --> 01:12:25,080
that like try to see like if you have a

1925
01:12:21,840 --> 01:12:26,800
model that you can do some kind of

1926
01:12:25,080 --> 01:12:29,199
expensive decoding method for to get

1927
01:12:26,800 --> 01:12:31,120
good outputs is it then worth try

1928
01:12:29,199 --> 01:12:34,000
training that model right um there's

1929
01:12:31,120 --> 01:12:36,560
some great recent work on like training

1930
01:12:34,000 --> 01:12:39,400
models to produce the same kind of

1931
01:12:36,560 --> 01:12:40,760
outputs you get out of MVR without um

1932
01:12:39,400 --> 01:12:43,239
actually doing a really expensive

1933
01:12:40,760 --> 01:12:45,600
inference Stu so at some level like yeah

1934
01:12:43,239 --> 01:12:48,120
you can decide like this model is good

1935
01:12:45,600 --> 01:12:49,920
enough with its expensive method we can

1936
01:12:48,120 --> 01:12:50,920
try to make it cheaper by spending more

1937
01:12:49,920 --> 01:12:53,960
money on

1938
01:12:50,920 --> 01:12:55,520
funing um but that's not it's not like

1939
01:12:53,960 --> 01:12:57,320
necessarily guaranteed that that's will

1940
01:12:55,520 --> 01:13:00,679
be the case

1941
01:12:57,320 --> 01:13:03,040
Okay um the methods that we looked at

1942
01:13:00,679 --> 01:13:06,199
have these sort of trade-offs in quality

1943
01:13:03,040 --> 01:13:07,960
in diversity and in inference speed so

1944
01:13:06,199 --> 01:13:10,320
sampling from your model directly is

1945
01:13:07,960 --> 01:13:13,120
pretty fast to do you get really diverse

1946
01:13:10,320 --> 01:13:14,960
outputs but it tends to be lower quality

1947
01:13:13,120 --> 01:13:16,320
um whereas more restricted sampling

1948
01:13:14,960 --> 01:13:18,520
these sort of mode seeking search

1949
01:13:16,320 --> 01:13:20,639
methods tend to be higher quality but

1950
01:13:18,520 --> 01:13:21,880
you get less less diverse outputs and

1951
01:13:20,639 --> 01:13:23,560
that's why we have these methods like

1952
01:13:21,880 --> 01:13:26,719
diverse and stochastic resarch to

1953
01:13:23,560 --> 01:13:28,760
counter this a bit um and then methods

1954
01:13:26,719 --> 01:13:30,400
like MBR or other sample and rerank

1955
01:13:28,760 --> 01:13:32,679
methods tend to be very high quality

1956
01:13:30,400 --> 01:13:34,280
outputs but you pay for this with much

1957
01:13:32,679 --> 01:13:36,520
slower inference

1958
01:13:34,280 --> 01:13:38,679
time um but if I can kind of convince

1959
01:13:36,520 --> 01:13:41,560
you of anything today I think it would

1960
01:13:38,679 --> 01:13:43,600
be this which is that these the decoding

1961
01:13:41,560 --> 01:13:45,600
method you choose for your model has a

1962
01:13:43,600 --> 01:13:47,960
really strong impact on performance

1963
01:13:45,600 --> 01:13:49,520
Downstream um you can get radically

1964
01:13:47,960 --> 01:13:51,239
different results out of the same model

1965
01:13:49,520 --> 01:13:52,639
without doing any additional training

1966
01:13:51,239 --> 01:13:55,120
just by choosing the different decoding

1967
01:13:52,639 --> 01:13:57,880
method that you might want to try and so

1968
01:13:55,120 --> 01:13:59,679
when you sort of let your libraries pick

1969
01:13:57,880 --> 01:14:01,159
a quote unquote like sensible default

1970
01:13:59,679 --> 01:14:03,760
you can leave a lot of performance on

1971
01:14:01,159 --> 01:14:06,480
the train on the table so I encourage

1972
01:14:03,760 --> 01:14:08,199
you folks that if if you're um deploying

1973
01:14:06,480 --> 01:14:09,760
models in production or if you're doing

1974
01:14:08,199 --> 01:14:10,840
research or you know maybe look at your

1975
01:14:09,760 --> 01:14:13,280
outputs and your model has some

1976
01:14:10,840 --> 01:14:15,320
undesirable behaviors to consider if the

1977
01:14:13,280 --> 01:14:17,800
decoding method you're using is imposing

1978
01:14:15,320 --> 01:14:20,000
some kind of Intuition or some kind of

1979
01:14:17,800 --> 01:14:21,840
inductive bias and if you can alter that

1980
01:14:20,000 --> 01:14:24,239
to get some of these behaviors without

1981
01:14:21,840 --> 01:14:26,320
resorting to additional training

1982
01:14:24,239 --> 01:14:28,719
um and that's sort of the end I can take

1983
01:14:26,320 --> 01:14:28,719
any other

1984
01:14:34,320 --> 01:14:38,719
questions okay um yeah I guess we don't

1985
01:14:37,199 --> 01:14:41,360
have any questions we can take questions

1986
01:14:38,719 --> 01:14:45,560
up here um one one thing I'd like to

1987
01:14:41,360 --> 01:14:47,679
point out also is that um I I love the

1988
01:14:45,560 --> 01:14:50,760
final thing that Amanda said here

1989
01:14:47,679 --> 01:14:54,199
another thing is that my impression from

1990
01:14:50,760 --> 01:14:56,400
dealing with things is that it's a lot

1991
01:14:54,199 --> 01:14:58,159
easier to predict the effect of

1992
01:14:56,400 --> 01:14:59,920
inference time decoding time

1993
01:14:58,159 --> 01:15:01,120
manipulations than it is to predict the

1994
01:14:59,920 --> 01:15:04,239
effect of

1995
01:15:01,120 --> 01:15:07,480
like um fine-tuning or something like

1996
01:15:04,239 --> 01:15:11,040
this like just to give a an

1997
01:15:07,480 --> 01:15:12,480
example beam search with the maximum

1998
01:15:11,040 --> 01:15:15,199
likelihood trained model tends to

1999
01:15:12,480 --> 01:15:16,719
generate things that are shorter um

2000
01:15:15,199 --> 01:15:18,040
whereas greedy decoding tends to

2001
01:15:16,719 --> 01:15:19,639
generate things that are longer and

2002
01:15:18,040 --> 01:15:22,000
repeat more often and stuff like that

2003
01:15:19,639 --> 01:15:25,920
and if you try a few methods like this

2004
01:15:22,000 --> 01:15:28,920
you'll quickly find these kind of qus of

2005
01:15:25,920 --> 01:15:31,320
each of the methods and so by forming a

2006
01:15:28,920 --> 01:15:32,719
good intuition of this you will also

2007
01:15:31,320 --> 01:15:34,000
know how to fix these problems when you

2008
01:15:32,719 --> 01:15:35,600
see them it's like oh my model's

2009
01:15:34,000 --> 01:15:37,320
repeating itself a lot maybe I shouldn't

2010
01:15:35,600 --> 01:15:38,679
be using grey search I should be

2011
01:15:37,320 --> 01:15:41,199
switching over to something else or

2012
01:15:38,679 --> 01:15:43,320
something like that so um this is a good

2013
01:15:41,199 --> 01:15:45,880
thing to know and play around with yeah

2014
01:15:43,320 --> 01:15:47,239
and I think pretty underutilized too um

2015
01:15:45,880 --> 01:15:48,880
a lot of folks will not think about a

2016
01:15:47,239 --> 01:15:50,920
decoding method to fix their problem

2017
01:15:48,880 --> 01:15:52,280
even if like your model might actually

2018
01:15:50,920 --> 01:15:53,760
be perfectly fine under a different

2019
01:15:52,280 --> 01:15:56,000
decoding strategy

2020
01:15:53,760 --> 01:15:58,320
great okay thanks a lot everyone you can

2021
01:15:56,000 --> 01:15:58,320
uh

2022
01:16:02,280 --> 01:16:05,280
finish