ahmedelsayed's picture
commit files to HF hub
2ffb90d
1
00:00:00,399 --> 00:00:04,120
so this time I'm going to be talking
2
00:00:02,080 --> 00:00:05,799
about language modeling uh obviously
3
00:00:04,120 --> 00:00:07,240
language modeling is a big topic and I'm
4
00:00:05,799 --> 00:00:09,880
not going to be able to cover it all in
5
00:00:07,240 --> 00:00:11,320
one class but this is kind of the basics
6
00:00:09,880 --> 00:00:13,080
of uh what does it mean to build a
7
00:00:11,320 --> 00:00:15,320
language model what is a language model
8
00:00:13,080 --> 00:00:18,439
how do we evaluate language models and
9
00:00:15,320 --> 00:00:19,920
other stuff like that and around the end
10
00:00:18,439 --> 00:00:21,320
I'm going to talk a little bit about
11
00:00:19,920 --> 00:00:23,039
efficiently implementing things in
12
00:00:21,320 --> 00:00:25,080
neural networks it's not directly
13
00:00:23,039 --> 00:00:27,760
related to language models but it's very
14
00:00:25,080 --> 00:00:31,200
important to know how to do uh to solve
15
00:00:27,760 --> 00:00:34,200
your assignments so I'll cover both
16
00:00:31,200 --> 00:00:34,200
is
17
00:00:34,239 --> 00:00:38,480
cool okay so the first thing I'd like to
18
00:00:36,760 --> 00:00:41,239
talk about is generative versus
19
00:00:38,480 --> 00:00:43,000
discriminative models and the reason why
20
00:00:41,239 --> 00:00:45,280
is up until now we've been talking about
21
00:00:43,000 --> 00:00:47,559
discriminative models and these are
22
00:00:45,280 --> 00:00:49,640
models uh that are mainly designed to
23
00:00:47,559 --> 00:00:53,800
calculate the probability of a latent
24
00:00:49,640 --> 00:00:56,039
trait uh given the data and so this is
25
00:00:53,800 --> 00:00:58,800
uh P of Y given X where Y is the lat and
26
00:00:56,039 --> 00:01:00,800
trait we want to calculate and X is uh
27
00:00:58,800 --> 00:01:04,760
the input data that we're calculating it
28
00:01:00,800 --> 00:01:07,799
over so just review from last class what
29
00:01:04,760 --> 00:01:10,240
was X from last class from the example
30
00:01:07,799 --> 00:01:10,240
in L
31
00:01:11,360 --> 00:01:15,880
class
32
00:01:13,040 --> 00:01:18,280
anybody yeah some text yeah and then
33
00:01:15,880 --> 00:01:18,280
what was
34
00:01:20,400 --> 00:01:26,119
why it shouldn't be too
35
00:01:23,799 --> 00:01:27,920
hard yeah it was a category or a
36
00:01:26,119 --> 00:01:31,680
sentiment label precisely in the
37
00:01:27,920 --> 00:01:33,399
sentiment analysis tasks so so um a
38
00:01:31,680 --> 00:01:35,560
generative model on the other hand is a
39
00:01:33,399 --> 00:01:38,840
model that calculates the probability of
40
00:01:35,560 --> 00:01:40,880
data itself and is not specifically
41
00:01:38,840 --> 00:01:43,439
conditional and there's a couple of
42
00:01:40,880 --> 00:01:45,439
varieties um this isn't like super
43
00:01:43,439 --> 00:01:48,280
standard terminology I just uh wrote it
44
00:01:45,439 --> 00:01:51,520
myself but here we have a standalone
45
00:01:48,280 --> 00:01:54,360
probability of P of X and we can also
46
00:01:51,520 --> 00:01:58,000
calculate the joint probability P of X
47
00:01:54,360 --> 00:01:58,000
and Y
48
00:01:58,159 --> 00:02:02,880
so probabilistic language models
49
00:02:01,079 --> 00:02:06,640
basically what they do is they calculate
50
00:02:02,880 --> 00:02:08,560
this uh probability usually uh we think
51
00:02:06,640 --> 00:02:10,360
of it as a standalone probability of P
52
00:02:08,560 --> 00:02:11,800
of X where X is something like a
53
00:02:10,360 --> 00:02:15,160
sentence or a
54
00:02:11,800 --> 00:02:16,920
document and it's a generative model
55
00:02:15,160 --> 00:02:19,640
that calculates the probability of
56
00:02:16,920 --> 00:02:22,360
language recently the definition of
57
00:02:19,640 --> 00:02:23,959
language model has expanded a little bit
58
00:02:22,360 --> 00:02:26,160
so now
59
00:02:23,959 --> 00:02:28,640
um people also call things that
60
00:02:26,160 --> 00:02:31,080
calculate the probability of text and
61
00:02:28,640 --> 00:02:35,200
images as like multimodal language
62
00:02:31,080 --> 00:02:38,160
models or uh what are some of the other
63
00:02:35,200 --> 00:02:40,480
ones yeah I think that's the main the
64
00:02:38,160 --> 00:02:42,840
main exception to this rule usually
65
00:02:40,480 --> 00:02:45,080
usually it's calculating either of text
66
00:02:42,840 --> 00:02:47,680
or over text in some multimodal data but
67
00:02:45,080 --> 00:02:47,680
for now we're going to
68
00:02:48,800 --> 00:02:54,200
consider
69
00:02:50,319 --> 00:02:56,440
um then there's kind of two fundamental
70
00:02:54,200 --> 00:02:58,159
operations that we perform with LMS
71
00:02:56,440 --> 00:03:00,519
almost everything else we do with LMS
72
00:02:58,159 --> 00:03:03,640
can be considered like one of these two
73
00:03:00,519 --> 00:03:05,319
types of things the first thing is calc
74
00:03:03,640 --> 00:03:06,440
scoring sentences or calculating the
75
00:03:05,319 --> 00:03:09,599
probability of
76
00:03:06,440 --> 00:03:12,280
sentences and this
77
00:03:09,599 --> 00:03:14,720
is uh for example if we calculate the
78
00:03:12,280 --> 00:03:16,400
probability of Jane went to the store uh
79
00:03:14,720 --> 00:03:19,000
this would have a high probability
80
00:03:16,400 --> 00:03:20,879
ideally um and if we have this kind of
81
00:03:19,000 --> 00:03:23,400
word salid like this this would be given
82
00:03:20,879 --> 00:03:26,080
a low probability uh according to a
83
00:03:23,400 --> 00:03:28,000
English language model if we had a
84
00:03:26,080 --> 00:03:30,000
Chinese language model ideally it would
85
00:03:28,000 --> 00:03:31,319
also probably give low probability first
86
00:03:30,000 --> 00:03:32,879
sentence too because it's a language
87
00:03:31,319 --> 00:03:35,000
model of natural Chinese and not of
88
00:03:32,879 --> 00:03:36,200
natural English so there's also
89
00:03:35,000 --> 00:03:37,360
different types of language models
90
00:03:36,200 --> 00:03:38,400
depending on the type of data you play
91
00:03:37,360 --> 00:03:41,360
in
92
00:03:38,400 --> 00:03:43,599
the another thing I can do is generate
93
00:03:41,360 --> 00:03:45,239
sentences and we'll talk more about the
94
00:03:43,599 --> 00:03:48,280
different methods for generating
95
00:03:45,239 --> 00:03:50,319
sentences but typically they fall into
96
00:03:48,280 --> 00:03:51,799
one of two categories one is sampling
97
00:03:50,319 --> 00:03:53,200
like this where you try to sample a
98
00:03:51,799 --> 00:03:55,480
sentence from the probability
99
00:03:53,200 --> 00:03:57,280
distribution of the language model
100
00:03:55,480 --> 00:03:58,360
possibly with some modifications to the
101
00:03:57,280 --> 00:04:00,760
probability
102
00:03:58,360 --> 00:04:03,079
distribution um the other thing which I
103
00:04:00,760 --> 00:04:04,760
didn't write on the slide is uh finding
104
00:04:03,079 --> 00:04:07,439
the highest scoring sentence according
105
00:04:04,760 --> 00:04:09,760
to the language model um and we do both
106
00:04:07,439 --> 00:04:09,760
of those
107
00:04:10,560 --> 00:04:17,600
S so more concretely how can we apply
108
00:04:15,199 --> 00:04:21,199
these these can be applied to answer
109
00:04:17,600 --> 00:04:23,840
questions so for example um if we have a
110
00:04:21,199 --> 00:04:27,240
multiple choice question we can score
111
00:04:23,840 --> 00:04:30,639
possible multiple choice answers and uh
112
00:04:27,240 --> 00:04:32,880
the way we do this is we calculate
113
00:04:30,639 --> 00:04:35,440
we first
114
00:04:32,880 --> 00:04:38,440
take uh like we have
115
00:04:35,440 --> 00:04:38,440
like
116
00:04:38,560 --> 00:04:43,919
um
117
00:04:40,960 --> 00:04:46,919
where is
118
00:04:43,919 --> 00:04:46,919
CMU
119
00:04:47,560 --> 00:04:51,600
located um
120
00:04:51,960 --> 00:04:59,560
that's and actually maybe promete this
121
00:04:54,560 --> 00:05:01,360
all again to an a here and then we say X
122
00:04:59,560 --> 00:05:05,800
X1 is equal to
123
00:05:01,360 --> 00:05:07,520
this and then we have X2 which is
124
00:05:05,800 --> 00:05:09,720
Q
125
00:05:07,520 --> 00:05:12,479
where is
126
00:05:09,720 --> 00:05:14,120
CMU
127
00:05:12,479 --> 00:05:18,080
located
128
00:05:14,120 --> 00:05:19,720
a um what's something
129
00:05:18,080 --> 00:05:21,960
plausible
130
00:05:19,720 --> 00:05:24,560
uh what was
131
00:05:21,960 --> 00:05:26,319
it okay now now you're going to make it
132
00:05:24,560 --> 00:05:27,960
tricky and make me talk about when we
133
00:05:26,319 --> 00:05:29,960
have multiple right answers and how we
134
00:05:27,960 --> 00:05:31,759
evaluate and stuff let let's ignore that
135
00:05:29,960 --> 00:05:35,080
for now it's say New
136
00:05:31,759 --> 00:05:37,199
York it's not located in New York is
137
00:05:35,080 --> 00:05:40,560
it
138
00:05:37,199 --> 00:05:40,560
okay let's say
139
00:05:40,960 --> 00:05:45,199
Birmingham hopefully there's no CMU
140
00:05:43,199 --> 00:05:47,120
affiliate in Birmingham I think we're
141
00:05:45,199 --> 00:05:49,000
we're pretty so um and then you would
142
00:05:47,120 --> 00:05:53,880
just calculate the probability of X1 and
143
00:05:49,000 --> 00:05:56,440
the probability of X2 X3 X4 Etc and um
144
00:05:53,880 --> 00:06:01,479
then pick the highest saring one and
145
00:05:56,440 --> 00:06:01,479
actually um there's a famous
146
00:06:03,199 --> 00:06:07,440
there's a famous uh leaderboard for
147
00:06:05,840 --> 00:06:08,759
language models that probably a lot of
148
00:06:07,440 --> 00:06:09,759
people know about it's called the open
149
00:06:08,759 --> 00:06:13,120
llm
150
00:06:09,759 --> 00:06:15,639
leaderboard and a lot of these tasks
151
00:06:13,120 --> 00:06:17,319
here basically correspond to doing
152
00:06:15,639 --> 00:06:21,000
something like that like hel swag is
153
00:06:17,319 --> 00:06:22,599
kind of a multiple choice uh is a
154
00:06:21,000 --> 00:06:24,160
multiple choice question answering thing
155
00:06:22,599 --> 00:06:27,880
about common sense where they calculate
156
00:06:24,160 --> 00:06:30,280
it by scoring uh scoring the
157
00:06:27,880 --> 00:06:31,880
outputs so that's a very common way to
158
00:06:30,280 --> 00:06:35,000
use language
159
00:06:31,880 --> 00:06:36,960
models um another thing is generating a
160
00:06:35,000 --> 00:06:40,080
continuation of a question prompt so
161
00:06:36,960 --> 00:06:42,639
basically this is when you uh
162
00:06:40,080 --> 00:06:44,759
sample and so what you would do is you
163
00:06:42,639 --> 00:06:48,440
would prompt the
164
00:06:44,759 --> 00:06:50,560
model with this uh X here and then you
165
00:06:48,440 --> 00:06:53,800
would ask it to generate either the most
166
00:06:50,560 --> 00:06:56,400
likely uh completion or generate um
167
00:06:53,800 --> 00:06:58,960
sample multiple completions to get the
168
00:06:56,400 --> 00:07:00,720
answer so this is very common uh people
169
00:06:58,960 --> 00:07:03,759
are very familiar with this there's lots
170
00:07:00,720 --> 00:07:07,160
of other uh things you can do though so
171
00:07:03,759 --> 00:07:09,400
um you can classify text and there's a
172
00:07:07,160 --> 00:07:12,720
couple ways you can do this uh one way
173
00:07:09,400 --> 00:07:15,960
you can do this is um like let's say we
174
00:07:12,720 --> 00:07:15,960
have a sentiment sentence
175
00:07:16,160 --> 00:07:21,520
here
176
00:07:17,759 --> 00:07:25,440
um you can say uh
177
00:07:21,520 --> 00:07:30,919
this is
178
00:07:25,440 --> 00:07:33,919
gr and then you can say um
179
00:07:30,919 --> 00:07:37,680
star
180
00:07:33,919 --> 00:07:38,879
rating five or something like that and
181
00:07:37,680 --> 00:07:41,400
then you could also have star rating
182
00:07:38,879 --> 00:07:43,680
four star rating three star rating two
183
00:07:41,400 --> 00:07:45,080
star rating one and calculate the
184
00:07:43,680 --> 00:07:46,639
probability of all of these and find
185
00:07:45,080 --> 00:07:50,360
which one has the highest probability so
186
00:07:46,639 --> 00:07:51,800
this is a a common way you can do things
187
00:07:50,360 --> 00:07:54,319
another thing you can do which is kind
188
00:07:51,800 --> 00:07:55,240
of interesting and um there are papers
189
00:07:54,319 --> 00:07:58,319
on this but they're kind of
190
00:07:55,240 --> 00:08:00,800
underexplored is you can do like star
191
00:07:58,319 --> 00:08:04,800
rating
192
00:08:00,800 --> 00:08:04,800
five and then
193
00:08:04,879 --> 00:08:13,280
generate generate the output um and so
194
00:08:10,319 --> 00:08:15,039
that basically says Okay I I want a
195
00:08:13,280 --> 00:08:16,680
positive sentence now I'm going to score
196
00:08:15,039 --> 00:08:19,120
the actual review and see whether that
197
00:08:16,680 --> 00:08:22,319
matches my like conception of a positive
198
00:08:19,120 --> 00:08:24,080
sentence and there's a few uh papers
199
00:08:22,319 --> 00:08:25,680
that do
200
00:08:24,080 --> 00:08:28,240
this
201
00:08:25,680 --> 00:08:31,240
um let
202
00:08:28,240 --> 00:08:31,240
me
203
00:08:34,640 --> 00:08:38,760
this is a kind of older one and then
204
00:08:36,240 --> 00:08:42,080
there's another more recent one by Sean
205
00:08:38,760 --> 00:08:43,839
Min I believe um uh but they demonstrate
206
00:08:42,080 --> 00:08:45,480
how you can do both generative and
207
00:08:43,839 --> 00:08:47,600
discriminative classification in this
208
00:08:45,480 --> 00:08:51,760
way so that's another thing that you can
209
00:08:47,600 --> 00:08:51,760
do uh with language
210
00:08:53,279 --> 00:08:56,839
models and then the other thing you can
211
00:08:55,200 --> 00:08:59,000
do is you can generate the label given a
212
00:08:56,839 --> 00:09:00,680
classification proc so you you say this
213
00:08:59,000 --> 00:09:03,079
is is great star rating and then
214
00:09:00,680 --> 00:09:05,720
generate five
215
00:09:03,079 --> 00:09:09,320
whatever finally um you can do things
216
00:09:05,720 --> 00:09:10,920
like correct a grammar so uh for example
217
00:09:09,320 --> 00:09:12,560
if you score the probability of each
218
00:09:10,920 --> 00:09:14,839
word and you find words that are really
219
00:09:12,560 --> 00:09:17,760
low probability then you can uh replace
220
00:09:14,839 --> 00:09:20,160
them with higher probability words um or
221
00:09:17,760 --> 00:09:21,720
you could ask a model please paraphrase
222
00:09:20,160 --> 00:09:24,000
this output and it will paraphrase it
223
00:09:21,720 --> 00:09:27,640
into something that gives you uh you
224
00:09:24,000 --> 00:09:30,720
know that has better gra so basically
225
00:09:27,640 --> 00:09:33,079
like as I said language models are very
226
00:09:30,720 --> 00:09:34,600
diverse um and they can do a ton of
227
00:09:33,079 --> 00:09:35,680
different things but most of them boil
228
00:09:34,600 --> 00:09:38,440
down to doing one of these two
229
00:09:35,680 --> 00:09:42,079
operations scoring or
230
00:09:38,440 --> 00:09:42,079
generating any questions
231
00:09:42,480 --> 00:09:47,600
s
232
00:09:44,640 --> 00:09:50,000
okay so next I I want to talk about a
233
00:09:47,600 --> 00:09:52,279
specific type of language models uh Auto
234
00:09:50,000 --> 00:09:54,240
regressive language models and auto
235
00:09:52,279 --> 00:09:56,720
regressive language models are language
236
00:09:54,240 --> 00:10:00,240
models that specifically calculate this
237
00:09:56,720 --> 00:10:02,320
probability um in a fashion where you
238
00:10:00,240 --> 00:10:03,680
calculate the probability of one token
239
00:10:02,320 --> 00:10:05,519
and then you calculate the probability
240
00:10:03,680 --> 00:10:07,680
of the next token given the previous
241
00:10:05,519 --> 00:10:10,519
token the probability of the third token
242
00:10:07,680 --> 00:10:13,760
G given the previous two tokens almost
243
00:10:10,519 --> 00:10:18,600
always this happens left to right um or
244
00:10:13,760 --> 00:10:20,519
start to finish um and so this is the
245
00:10:18,600 --> 00:10:25,000
next token here this is a context where
246
00:10:20,519 --> 00:10:28,440
usually um the context is the previous
247
00:10:25,000 --> 00:10:29,640
tokens Can anyone think of a time when
248
00:10:28,440 --> 00:10:32,440
you might want to do
249
00:10:29,640 --> 00:10:37,839
right to left instead of left to
250
00:10:32,440 --> 00:10:40,399
right yeah language that's from right to
251
00:10:37,839 --> 00:10:41,680
yeah that's actually exactly what I what
252
00:10:40,399 --> 00:10:43,079
I was looking for so if you have a
253
00:10:41,680 --> 00:10:46,839
language that's written from right to
254
00:10:43,079 --> 00:10:49,320
left actually uh things like uh Arabic
255
00:10:46,839 --> 00:10:51,360
and Hebrew are written right to left so
256
00:10:49,320 --> 00:10:53,720
um both of those are
257
00:10:51,360 --> 00:10:56,360
chronologically like earlier to later
258
00:10:53,720 --> 00:10:59,399
because you know if if you're thinking
259
00:10:56,360 --> 00:11:01,079
about how people speak um the the first
260
00:10:59,399 --> 00:11:02,440
word that an English speaker speaks is
261
00:11:01,079 --> 00:11:04,000
on the left just because that's the way
262
00:11:02,440 --> 00:11:06,079
you write it but the first word that an
263
00:11:04,000 --> 00:11:09,639
Arabic speaker speaks is on the the
264
00:11:06,079 --> 00:11:12,360
right because chronologically that's uh
265
00:11:09,639 --> 00:11:13,519
that's how it works um there's other
266
00:11:12,360 --> 00:11:16,320
reasons why you might want to do right
267
00:11:13,519 --> 00:11:17,839
to left but uh it's not really that left
268
00:11:16,320 --> 00:11:21,720
to right is important it's that like
269
00:11:17,839 --> 00:11:24,440
start to finish is important in spoken
270
00:11:21,720 --> 00:11:27,880
language so um one thing I should
271
00:11:24,440 --> 00:11:30,240
mention here is that this is just a rule
272
00:11:27,880 --> 00:11:31,560
of probability that if you have multiple
273
00:11:30,240 --> 00:11:33,720
variables and you're calculating the
274
00:11:31,560 --> 00:11:35,760
joint probability of variables the
275
00:11:33,720 --> 00:11:38,000
probability of all of the variables
276
00:11:35,760 --> 00:11:40,240
together is equal to this probability
277
00:11:38,000 --> 00:11:41,920
here so we're not making any
278
00:11:40,240 --> 00:11:44,399
approximations we're not making any
279
00:11:41,920 --> 00:11:46,959
compromises in order to do this but it
280
00:11:44,399 --> 00:11:51,639
all hinges on whether we can predict
281
00:11:46,959 --> 00:11:53,440
this probability um accurately uh
282
00:11:51,639 --> 00:11:56,160
actually another question does anybody
283
00:11:53,440 --> 00:11:57,800
know why we do this decomposition why
284
00:11:56,160 --> 00:12:00,959
don't we just try to predict the
285
00:11:57,800 --> 00:12:00,959
probability of x
286
00:12:02,120 --> 00:12:05,399
directly any
287
00:12:07,680 --> 00:12:12,760
ideas uh of big X sorry uh why don't we
288
00:12:11,079 --> 00:12:17,560
try to calculate the probability of this
289
00:12:12,760 --> 00:12:21,360
is great directly without deated the
290
00:12:17,560 --> 00:12:21,360
IND that
291
00:12:25,519 --> 00:12:31,560
possibility it could be word salid if
292
00:12:27,760 --> 00:12:35,279
you did it in a in a particular way yes
293
00:12:31,560 --> 00:12:35,279
um so that that's a good point
294
00:12:39,519 --> 00:12:47,000
yeah yeah so for example we talked about
295
00:12:43,760 --> 00:12:50,120
um uh we'll talk about
296
00:12:47,000 --> 00:12:51,920
models um or I I mentioned this briefly
297
00:12:50,120 --> 00:12:54,000
last time you can mention it in more
298
00:12:51,920 --> 00:12:55,639
detail this time but this is great we
299
00:12:54,000 --> 00:12:59,880
probably have never seen this before
300
00:12:55,639 --> 00:13:01,399
right so if we predict only things that
301
00:12:59,880 --> 00:13:03,199
we've seen before if we only assign a
302
00:13:01,399 --> 00:13:04,600
non-zero probability to the things we've
303
00:13:03,199 --> 00:13:06,000
seen before there's going to be lots of
304
00:13:04,600 --> 00:13:07,079
sentences that we've never seen before
305
00:13:06,000 --> 00:13:10,000
it makes it
306
00:13:07,079 --> 00:13:13,760
supercars um that that's basically close
307
00:13:10,000 --> 00:13:16,399
to what I wanted to say so um the reason
308
00:13:13,760 --> 00:13:18,040
why we don't typically do it with um
309
00:13:16,399 --> 00:13:21,240
predicting the whole sentence directly
310
00:13:18,040 --> 00:13:22,800
is because if we think about the size of
311
00:13:21,240 --> 00:13:24,959
the classification problem we need to
312
00:13:22,800 --> 00:13:27,880
solve in order to predict the next word
313
00:13:24,959 --> 00:13:30,320
it's a v uh where V is the size of the
314
00:13:27,880 --> 00:13:33,120
vocabulary but the size of the
315
00:13:30,320 --> 00:13:35,399
classification problem that we need to
316
00:13:33,120 --> 00:13:38,040
um we need to solve if we predict
317
00:13:35,399 --> 00:13:40,079
everything directly is V to the N where
318
00:13:38,040 --> 00:13:42,240
n is the length of the sequence and
319
00:13:40,079 --> 00:13:45,240
that's just huge the vocabulary is so
320
00:13:42,240 --> 00:13:48,440
big that it's hard to kind of uh know
321
00:13:45,240 --> 00:13:51,000
how we handle that so basically by doing
322
00:13:48,440 --> 00:13:53,160
this sort of decomposition we decompose
323
00:13:51,000 --> 00:13:56,440
this into uh
324
00:13:53,160 --> 00:13:58,120
n um prediction problems of size V and
325
00:13:56,440 --> 00:13:59,519
that's kind of just a lot more
326
00:13:58,120 --> 00:14:03,079
manageable for from the point of view of
327
00:13:59,519 --> 00:14:06,000
how we train uh know how we train
328
00:14:03,079 --> 00:14:09,399
models um that being said there are
329
00:14:06,000 --> 00:14:11,360
other Alternatives um something very
330
00:14:09,399 --> 00:14:13,920
widely known uh very widely used is
331
00:14:11,360 --> 00:14:16,440
called a MK language model um a mast
332
00:14:13,920 --> 00:14:19,480
language model is something like Bert or
333
00:14:16,440 --> 00:14:21,680
debera or Roberta or all of these models
334
00:14:19,480 --> 00:14:25,000
that you might have heard if you've been
335
00:14:21,680 --> 00:14:28,279
in MLP for more than two years I guess
336
00:14:25,000 --> 00:14:30,680
um and basically what they do is they
337
00:14:28,279 --> 00:14:30,680
predict
338
00:14:32,199 --> 00:14:37,480
uh they like mask out this word and they
339
00:14:34,839 --> 00:14:39,480
predict the middle word so they mask out
340
00:14:37,480 --> 00:14:41,440
is and then try to predict that given
341
00:14:39,480 --> 00:14:45,320
all the other words the problem with
342
00:14:41,440 --> 00:14:48,959
these models is uh twofold number one
343
00:14:45,320 --> 00:14:51,880
they don't actually give you a uh good
344
00:14:48,959 --> 00:14:55,399
probability here uh like a a properly
345
00:14:51,880 --> 00:14:57,800
formed probability here
346
00:14:55,399 --> 00:14:59,160
because this is true only as long as
347
00:14:57,800 --> 00:15:01,920
you're only conditioning on things that
348
00:14:59,160 --> 00:15:03,480
you've previously generated so that
349
00:15:01,920 --> 00:15:04,839
they're not actually true language
350
00:15:03,480 --> 00:15:06,920
models from the point of view of being
351
00:15:04,839 --> 00:15:10,040
able to easily predict the probability
352
00:15:06,920 --> 00:15:11,399
of a sequence um and also it's hard to
353
00:15:10,040 --> 00:15:13,399
generate from them because you need to
354
00:15:11,399 --> 00:15:15,440
generate in some order and mass language
355
00:15:13,399 --> 00:15:17,600
models don't specify economical orders
356
00:15:15,440 --> 00:15:19,120
so they're good for some things like
357
00:15:17,600 --> 00:15:21,720
calculating representations of the
358
00:15:19,120 --> 00:15:22,920
output but they're not useful uh they're
359
00:15:21,720 --> 00:15:25,240
not as useful for
360
00:15:22,920 --> 00:15:26,880
Generation Um there's also energy based
361
00:15:25,240 --> 00:15:28,759
language models which basically create a
362
00:15:26,880 --> 00:15:30,000
scoring function that's not necessarily
363
00:15:28,759 --> 00:15:31,279
left to right or right to left or
364
00:15:30,000 --> 00:15:33,120
anything like that but that's very
365
00:15:31,279 --> 00:15:34,639
Advanced um if you're interested in them
366
00:15:33,120 --> 00:15:36,319
I can talk more about them that we'll
367
00:15:34,639 --> 00:15:38,920
skip
368
00:15:36,319 --> 00:15:41,600
them and um also all of the language
369
00:15:38,920 --> 00:15:45,639
models that you hear about nowadays GPT
370
00:15:41,600 --> 00:15:48,800
uh llama whatever else are all other
371
00:15:45,639 --> 00:15:52,880
models cool so I'm going to go into the
372
00:15:48,800 --> 00:15:52,880
very um any questions about that
373
00:15:57,600 --> 00:16:00,600
yeah
374
00:16:00,680 --> 00:16:04,160
yeah so in Mass language models the
375
00:16:02,680 --> 00:16:06,000
question was in Mass language models
376
00:16:04,160 --> 00:16:08,360
couldn't you just mask out the last
377
00:16:06,000 --> 00:16:10,759
token and predict that sure you could do
378
00:16:08,360 --> 00:16:13,079
that but there it's just not trained
379
00:16:10,759 --> 00:16:14,720
that way so it won't do a very good job
380
00:16:13,079 --> 00:16:16,880
if you always trained it that way it's
381
00:16:14,720 --> 00:16:18,160
an autor regressive language model so
382
00:16:16,880 --> 00:16:22,240
you're you're back to where you were in
383
00:16:18,160 --> 00:16:24,800
the first place um cool so now we I'll
384
00:16:22,240 --> 00:16:26,399
talk about unigram language models and
385
00:16:24,800 --> 00:16:29,319
so the simplest language models are
386
00:16:26,399 --> 00:16:33,560
count-based unigram language models and
387
00:16:29,319 --> 00:16:35,319
the way they work is um basically we
388
00:16:33,560 --> 00:16:38,519
want to calculate this probability
389
00:16:35,319 --> 00:16:41,240
conditioned on all the previous ones and
390
00:16:38,519 --> 00:16:42,360
the way we do this is we just say
391
00:16:41,240 --> 00:16:45,680
actually we're not going to worry about
392
00:16:42,360 --> 00:16:48,759
the order at all and we're just going to
393
00:16:45,680 --> 00:16:52,240
uh predict the probability of the next
394
00:16:48,759 --> 00:16:55,279
word uh independently of all the other
395
00:16:52,240 --> 00:16:57,519
words so if you have something like this
396
00:16:55,279 --> 00:16:59,720
it's actually extremely easy to predict
397
00:16:57,519 --> 00:17:02,480
the probability of this word and the way
398
00:16:59,720 --> 00:17:04,280
you do this is you just count up the
399
00:17:02,480 --> 00:17:08,360
number of times this word appeared in
400
00:17:04,280 --> 00:17:10,480
the training data set and divide by the
401
00:17:08,360 --> 00:17:12,559
uh divide by the total number of words
402
00:17:10,480 --> 00:17:14,240
in the pring data set and now you have a
403
00:17:12,559 --> 00:17:15,959
language model this is like language
404
00:17:14,240 --> 00:17:17,760
model 101 it's the easiest possible
405
00:17:15,959 --> 00:17:19,520
language model you can write in you know
406
00:17:17,760 --> 00:17:21,120
three lines of python
407
00:17:19,520 --> 00:17:25,039
basically
408
00:17:21,120 --> 00:17:28,480
um so it has a few problems uh the first
409
00:17:25,039 --> 00:17:31,120
problem with this language model is um
410
00:17:28,480 --> 00:17:32,960
handling unknown words so what happens
411
00:17:31,120 --> 00:17:38,679
if you have a word that you've never
412
00:17:32,960 --> 00:17:41,000
seen before um in this language model
413
00:17:38,679 --> 00:17:42,240
here what is the probability of any
414
00:17:41,000 --> 00:17:44,720
sequence that has a word that you've
415
00:17:42,240 --> 00:17:47,440
never seen before yeah the probability
416
00:17:44,720 --> 00:17:49,240
of the sequence gets zero so there might
417
00:17:47,440 --> 00:17:51,120
not be such a big problem for generating
418
00:17:49,240 --> 00:17:52,480
things from the language model because
419
00:17:51,120 --> 00:17:54,520
you know maybe it's fine if you only
420
00:17:52,480 --> 00:17:55,960
generate words that you've seen before
421
00:17:54,520 --> 00:17:57,679
uh but it is definitely a problem of
422
00:17:55,960 --> 00:17:59,720
scoring things with the language model
423
00:17:57,679 --> 00:18:02,039
and it's also a problem of uh for
424
00:17:59,720 --> 00:18:04,440
something like translation if you get an
425
00:18:02,039 --> 00:18:05,840
unknown word uh when you're translating
426
00:18:04,440 --> 00:18:07,799
something then you would like to be able
427
00:18:05,840 --> 00:18:11,320
to translate it reasonably but you can't
428
00:18:07,799 --> 00:18:13,799
do that so um that's an issue so how do
429
00:18:11,320 --> 00:18:15,840
we how do we fix this um there's a
430
00:18:13,799 --> 00:18:17,640
couple options the first option is to
431
00:18:15,840 --> 00:18:19,440
segment to characters and subwords and
432
00:18:17,640 --> 00:18:21,720
this is now the preferred option that
433
00:18:19,440 --> 00:18:24,360
most people use nowadays uh just run
434
00:18:21,720 --> 00:18:26,840
sentence piece segment your vocabulary
435
00:18:24,360 --> 00:18:28,400
and you're all set you're you'll now no
436
00:18:26,840 --> 00:18:29,679
longer have any unknown words because
437
00:18:28,400 --> 00:18:30,840
all the unknown words get split into
438
00:18:29,679 --> 00:18:33,559
shorter
439
00:18:30,840 --> 00:18:36,240
units there's also other options that
440
00:18:33,559 --> 00:18:37,919
you can use if you're uh very interested
441
00:18:36,240 --> 00:18:41,280
in or serious about this and want to
442
00:18:37,919 --> 00:18:43,720
handle this like uh as part of a
443
00:18:41,280 --> 00:18:45,960
research project or something like this
444
00:18:43,720 --> 00:18:48,520
and uh the way you can do this is you
445
00:18:45,960 --> 00:18:50,120
can build an unknown word model and an
446
00:18:48,520 --> 00:18:52,200
unknown word model basically what it
447
00:18:50,120 --> 00:18:54,520
does is it uh predicts the probability
448
00:18:52,200 --> 00:18:56,200
of unknown words using characters and
449
00:18:54,520 --> 00:18:59,559
then it models the probability of words
450
00:18:56,200 --> 00:19:01,159
using words and so now you can you have
451
00:18:59,559 --> 00:19:02,559
kind of like a hierarchical model where
452
00:19:01,159 --> 00:19:03,919
you first try to predict words and then
453
00:19:02,559 --> 00:19:06,720
if you can't predict words you predict
454
00:19:03,919 --> 00:19:08,960
unknown words so this isn't us as widely
455
00:19:06,720 --> 00:19:11,520
anymore but it's worth thinking about uh
456
00:19:08,960 --> 00:19:11,520
or knowing
457
00:19:11,840 --> 00:19:20,880
about okay uh so a second detail um a
458
00:19:17,200 --> 00:19:22,799
parameter uh so parameterizing in log
459
00:19:20,880 --> 00:19:25,880
space
460
00:19:22,799 --> 00:19:28,400
so the um multiplication of
461
00:19:25,880 --> 00:19:29,840
probabilities can be reexpressed is the
462
00:19:28,400 --> 00:19:31,840
addition of log
463
00:19:29,840 --> 00:19:34,159
probabilities uh so this is really
464
00:19:31,840 --> 00:19:35,720
important and this is widely used in all
465
00:19:34,159 --> 00:19:37,520
language models whether they're unigram
466
00:19:35,720 --> 00:19:39,640
language models or or neural language
467
00:19:37,520 --> 00:19:41,799
models there's actually a very simple
468
00:19:39,640 --> 00:19:45,440
reason why we why we do it this way does
469
00:19:41,799 --> 00:19:45,440
anybody uh know the
470
00:19:46,440 --> 00:19:52,679
answer what would happen if we
471
00:19:48,280 --> 00:19:56,720
multiplied uh let's say uh 30 30 tokens
472
00:19:52,679 --> 00:20:00,360
worth of probabilities together um
473
00:19:56,720 --> 00:20:02,120
yeah uh yeah too too small um so
474
00:20:00,360 --> 00:20:06,120
basically the problem is numerical
475
00:20:02,120 --> 00:20:07,520
underflow um so modern computers if if
476
00:20:06,120 --> 00:20:08,840
we weren't doing this on a computer and
477
00:20:07,520 --> 00:20:11,240
we were just doing math it wouldn't
478
00:20:08,840 --> 00:20:14,280
matter at all um but because we're doing
479
00:20:11,240 --> 00:20:17,280
it on a computer uh we
480
00:20:14,280 --> 00:20:17,280
have
481
00:20:20,880 --> 00:20:26,000
ours we have our
482
00:20:23,000 --> 00:20:26,000
32bit
483
00:20:27,159 --> 00:20:30,159
float
484
00:20:32,320 --> 00:20:37,720
where we have uh the exponent in the the
485
00:20:35,799 --> 00:20:40,159
fraction over here so the largest the
486
00:20:37,720 --> 00:20:41,960
exponent can get is limited by the
487
00:20:40,159 --> 00:20:45,880
number of exponent bits that we have in
488
00:20:41,960 --> 00:20:48,039
a 32-bit float and um if that's the case
489
00:20:45,880 --> 00:20:52,480
I forget exactly how large it is it's
490
00:20:48,039 --> 00:20:53,440
like yeah something like 30 minus 38 is
491
00:20:52,480 --> 00:20:56,640
that
492
00:20:53,440 --> 00:20:58,520
right yeah but anyway like if the number
493
00:20:56,640 --> 00:21:00,640
gets too small you'll underflow it goes
494
00:20:58,520 --> 00:21:02,400
to zero and you'll get a zero
495
00:21:00,640 --> 00:21:05,720
probability despite the fact that it's
496
00:21:02,400 --> 00:21:07,640
not actually zero so um that's usually
497
00:21:05,720 --> 00:21:09,440
why we do this it's also a little bit
498
00:21:07,640 --> 00:21:12,960
easier for people just to look at like
499
00:21:09,440 --> 00:21:15,200
minus 30 instead of looking to something
500
00:21:12,960 --> 00:21:19,960
something time 10 to the minus 30 or
501
00:21:15,200 --> 00:21:24,520
something so uh that is why we normally
502
00:21:19,960 --> 00:21:27,159
go um another thing that you can note is
503
00:21:24,520 --> 00:21:28,760
uh you can treat each of these in a
504
00:21:27,159 --> 00:21:31,360
unigram model you can treat each of
505
00:21:28,760 --> 00:21:37,039
these as parameters so we talked about
506
00:21:31,360 --> 00:21:39,640
parameters of a model uh like a um like
507
00:21:37,039 --> 00:21:41,120
a bag of words model and we can
508
00:21:39,640 --> 00:21:44,080
similarly treat these unigram
509
00:21:41,120 --> 00:21:47,760
probabilities as parameters so um how
510
00:21:44,080 --> 00:21:47,760
many parameters does a unigram model
511
00:21:48,080 --> 00:21:51,320
have any
512
00:21:57,039 --> 00:22:02,400
ideas
513
00:21:59,600 --> 00:22:04,440
yeah yeah exactly parameters equal to
514
00:22:02,400 --> 00:22:08,120
the size of the vocabulary so this one's
515
00:22:04,440 --> 00:22:10,880
easy and then we can go um we can go to
516
00:22:08,120 --> 00:22:13,880
the slightly less easy ones
517
00:22:10,880 --> 00:22:16,039
there so anyway this is a unigram model
518
00:22:13,880 --> 00:22:17,960
uh it's it's not too hard um you
519
00:22:16,039 --> 00:22:20,480
basically count up and divide and then
520
00:22:17,960 --> 00:22:22,720
you add the the probabilities here you
521
00:22:20,480 --> 00:22:25,440
could easily do it in a short Python
522
00:22:22,720 --> 00:22:28,400
program higher order engram models so
523
00:22:25,440 --> 00:22:31,600
higher order engram models um what these
524
00:22:28,400 --> 00:22:35,520
do is they essentially limit the context
525
00:22:31,600 --> 00:22:40,240
length to a length of N and then they
526
00:22:35,520 --> 00:22:42,600
count and divide so the way it works
527
00:22:40,240 --> 00:22:45,559
here maybe this is a little bit uh
528
00:22:42,600 --> 00:22:47,320
tricky but I can show an example so what
529
00:22:45,559 --> 00:22:49,840
we do is we count up the number of times
530
00:22:47,320 --> 00:22:51,320
we've seen this is an example and then
531
00:22:49,840 --> 00:22:53,480
we divide by the number of times we've
532
00:22:51,320 --> 00:22:55,960
seen this is n and that's the
533
00:22:53,480 --> 00:22:56,960
probability of example given the the
534
00:22:55,960 --> 00:22:58,720
previous
535
00:22:56,960 --> 00:23:00,559
coms
536
00:22:58,720 --> 00:23:02,039
so the problem with this is anytime we
537
00:23:00,559 --> 00:23:03,400
get a sequence that we've never seen
538
00:23:02,039 --> 00:23:04,960
before like we would like to model
539
00:23:03,400 --> 00:23:07,200
longer sequences to make this more
540
00:23:04,960 --> 00:23:08,600
accurate but anytime we've get a uh we
541
00:23:07,200 --> 00:23:10,720
get a sequence that we've never seen
542
00:23:08,600 --> 00:23:12,919
before um it will get a probability of
543
00:23:10,720 --> 00:23:15,919
zero similarly because this count on top
544
00:23:12,919 --> 00:23:19,919
of here will be zero so the way that uh
545
00:23:15,919 --> 00:23:22,640
engram language models work with this uh
546
00:23:19,919 --> 00:23:27,320
handle this is they have fall back to
547
00:23:22,640 --> 00:23:31,840
Shorter uh engram models so um this
548
00:23:27,320 --> 00:23:33,480
model sorry when I say NR uh n is the
549
00:23:31,840 --> 00:23:35,520
length of the context so this is a four
550
00:23:33,480 --> 00:23:37,679
gr model here because the top context is
551
00:23:35,520 --> 00:23:40,520
four so the photogram model would
552
00:23:37,679 --> 00:23:46,640
calculate this and then interpolate it
553
00:23:40,520 --> 00:23:48,640
like this with a um with a trigram model
554
00:23:46,640 --> 00:23:50,400
uh and then the trigram model itself
555
00:23:48,640 --> 00:23:51,720
would interpolate with the Byram model
556
00:23:50,400 --> 00:23:53,440
the Byram model would interpolate with
557
00:23:51,720 --> 00:23:56,880
the unram
558
00:23:53,440 --> 00:23:59,880
model oh this one oh
559
00:23:56,880 --> 00:23:59,880
okay
560
00:24:02,159 --> 00:24:05,440
um one
561
00:24:07,039 --> 00:24:12,320
second could you uh help get it from the
562
00:24:10,000 --> 00:24:12,320
lock
563
00:24:26,799 --> 00:24:29,799
box
564
00:24:43,640 --> 00:24:50,200
um okay sorry
565
00:24:46,880 --> 00:24:53,640
so getting bad
566
00:24:50,200 --> 00:24:56,640
here just
567
00:24:53,640 --> 00:24:56,640
actually
568
00:24:56,760 --> 00:25:02,559
okay uh oh wow that's a lot
569
00:25:02,960 --> 00:25:12,080
better cool okay so
570
00:25:08,279 --> 00:25:14,159
um so this is uh how we deal with the
571
00:25:12,080 --> 00:25:18,799
fact that models can
572
00:25:14,159 --> 00:25:23,919
be um models can be more precise but
573
00:25:18,799 --> 00:25:26,679
more sparse and less precise but less
574
00:25:23,919 --> 00:25:28,720
sparse this is also another concept that
575
00:25:26,679 --> 00:25:31,039
we're going to talk about more later uh
576
00:25:28,720 --> 00:25:33,240
in another class but this is a variety
577
00:25:31,039 --> 00:25:33,240
of
578
00:25:33,679 --> 00:25:38,440
ensembling where we have different
579
00:25:35,960 --> 00:25:40,360
models that are good at different things
580
00:25:38,440 --> 00:25:42,279
and we combine them together so this is
581
00:25:40,360 --> 00:25:44,760
the first instance that you would see of
582
00:25:42,279 --> 00:25:46,159
this there are other instances of this
583
00:25:44,760 --> 00:25:50,320
but the reason why I mentioned that this
584
00:25:46,159 --> 00:25:51,840
is a a variety of ensembling is actually
585
00:25:50,320 --> 00:25:55,520
you're probably not going to be using
586
00:25:51,840 --> 00:25:57,840
engram models super widely unless you
587
00:25:55,520 --> 00:26:00,520
really want to process huge data sets
588
00:25:57,840 --> 00:26:02,399
because that is one advantage of them
589
00:26:00,520 --> 00:26:03,960
but some of these smoothing methods
590
00:26:02,399 --> 00:26:05,720
actually might be interesting even if
591
00:26:03,960 --> 00:26:10,520
you're using other models and ensembling
592
00:26:05,720 --> 00:26:10,520
them together so
593
00:26:10,600 --> 00:26:15,679
the in order to decide this
594
00:26:13,679 --> 00:26:19,559
interpolation coefficient one way we can
595
00:26:15,679 --> 00:26:23,440
do it is just set a fixed um set a fixed
596
00:26:19,559 --> 00:26:26,039
amount of probability that we use for
597
00:26:23,440 --> 00:26:29,000
every um every time so we could say that
598
00:26:26,039 --> 00:26:32,000
we always set this Lambda to 0.8 and
599
00:26:29,000 --> 00:26:34,320
some always set this Lambda 1us Lambda
600
00:26:32,000 --> 00:26:36,559
to 0.2 and interpolate those two
601
00:26:34,320 --> 00:26:39,120
together but actually there's more
602
00:26:36,559 --> 00:26:42,240
sophisticated methods of doing this and
603
00:26:39,120 --> 00:26:44,080
so one way of doing this is uh called
604
00:26:42,240 --> 00:26:47,240
additive
605
00:26:44,080 --> 00:26:50,600
smoothing excuse me and the the way that
606
00:26:47,240 --> 00:26:54,039
additive smoothing works is um basically
607
00:26:50,600 --> 00:26:54,919
we add Alpha to the uh to the top and
608
00:26:54,039 --> 00:26:58,000
the
609
00:26:54,919 --> 00:27:02,159
bottom and the reason why this is slight
610
00:26:58,000 --> 00:27:06,279
different as is as our accounts get
611
00:27:02,159 --> 00:27:10,799
larger we start to approach the true
612
00:27:06,279 --> 00:27:10,799
distribution so just to give an
613
00:27:12,080 --> 00:27:19,480
example let's say we have uh the
614
00:27:17,640 --> 00:27:21,640
box
615
00:27:19,480 --> 00:27:26,279
is
616
00:27:21,640 --> 00:27:26,279
um let's say initially we
617
00:27:26,520 --> 00:27:29,520
have
618
00:27:31,159 --> 00:27:37,600
uh let let's say our Alpha is
619
00:27:33,840 --> 00:27:43,559
one so initially if we have
620
00:27:37,600 --> 00:27:47,320
nothing um if we have no no evidence for
621
00:27:43,559 --> 00:27:47,320
our sorry I I
622
00:27:49,720 --> 00:27:54,960
realize let's say this is
623
00:27:52,640 --> 00:27:56,840
our fallback
624
00:27:54,960 --> 00:27:59,240
distribution um where this is a
625
00:27:56,840 --> 00:28:01,880
probability of Z 0.5 this is a
626
00:27:59,240 --> 00:28:03,360
probability of 0.3 and this is a
627
00:28:01,880 --> 00:28:06,559
probability of
628
00:28:03,360 --> 00:28:09,919
0.2 so now let's talk about our byr
629
00:28:06,559 --> 00:28:13,399
model um and our byr
630
00:28:09,919 --> 00:28:18,000
model has counts which is the
631
00:28:13,399 --> 00:28:18,000
the the box and the
632
00:28:19,039 --> 00:28:24,480
is so if we do something like this then
633
00:28:22,720 --> 00:28:26,720
um initially we have no counts like
634
00:28:24,480 --> 00:28:28,159
let's say we we have no data uh about
635
00:28:26,720 --> 00:28:30,760
this distribution
636
00:28:28,159 --> 00:28:33,200
um our counts would be zero and our
637
00:28:30,760 --> 00:28:35,919
Alpha would be
638
00:28:33,200 --> 00:28:37,840
one and so we would just fall back to
639
00:28:35,919 --> 00:28:40,960
this distribution we just have like one
640
00:28:37,840 --> 00:28:43,320
times uh one times this distribution
641
00:28:40,960 --> 00:28:45,679
let's say we then we have one piece of
642
00:28:43,320 --> 00:28:48,640
evidence and once we have one piece of
643
00:28:45,679 --> 00:28:52,279
evidence now this would be
644
00:28:48,640 --> 00:28:53,960
0.33 um and this would uh be Alpha equal
645
00:28:52,279 --> 00:28:56,399
to 1 so we'd have
646
00:28:53,960 --> 00:28:58,679
0.5 *
647
00:28:56,399 --> 00:29:00,399
0.33
648
00:28:58,679 --> 00:29:04,039
uh and
649
00:29:00,399 --> 00:29:07,720
0.5 time
650
00:29:04,039 --> 00:29:10,840
0.3 uh is the probability of the Box
651
00:29:07,720 --> 00:29:12,840
because um basically we we have one
652
00:29:10,840 --> 00:29:14,720
piece of evidence and we are adding a
653
00:29:12,840 --> 00:29:17,080
count of one to the lower order
654
00:29:14,720 --> 00:29:18,320
distribution then if we increase our
655
00:29:17,080 --> 00:29:24,159
count
656
00:29:18,320 --> 00:29:24,159
here um now we rely more
657
00:29:24,880 --> 00:29:30,960
strongly sorry that that would be wrong
658
00:29:27,720 --> 00:29:32,399
so so now we rely more strongly on the
659
00:29:30,960 --> 00:29:33,880
higher order distribution because we
660
00:29:32,399 --> 00:29:37,039
have more evidence for the higher order
661
00:29:33,880 --> 00:29:39,610
distribution so basically in this case
662
00:29:37,039 --> 00:29:41,240
um the probability
663
00:29:39,610 --> 00:29:44,559
[Music]
664
00:29:41,240 --> 00:29:48,200
of Lambda which I showed
665
00:29:44,559 --> 00:29:52,000
before is equal to the the sum of the
666
00:29:48,200 --> 00:29:54,200
counts plus um the sum of the counts
667
00:29:52,000 --> 00:29:56,480
over the sum of the counts plus
668
00:29:54,200 --> 00:29:58,159
Ali so as the sum of the counts gets
669
00:29:56,480 --> 00:30:00,240
larger you rely on the higher order
670
00:29:58,159 --> 00:30:01,640
distribution is the sum of the counts is
671
00:30:00,240 --> 00:30:02,760
if the sum of the counts is smaller you
672
00:30:01,640 --> 00:30:04,320
rely more on the lower order
673
00:30:02,760 --> 00:30:06,720
distribution so the more evidence you
674
00:30:04,320 --> 00:30:11,640
have the more you rely on so that's the
675
00:30:06,720 --> 00:30:11,640
basic idea behind these smoothing things
676
00:30:11,679 --> 00:30:16,679
um there's also a number of other
677
00:30:14,519 --> 00:30:18,760
varieties called uh
678
00:30:16,679 --> 00:30:20,799
discounting so uh the discount
679
00:30:18,760 --> 00:30:23,679
hyperparameter basically you subtract
680
00:30:20,799 --> 00:30:26,080
this off um uh you subtract this from
681
00:30:23,679 --> 00:30:27,840
the count so you would subtract like 0.5
682
00:30:26,080 --> 00:30:32,679
from each of the counts that you it's
683
00:30:27,840 --> 00:30:36,279
just empirically this is a better match
684
00:30:32,679 --> 00:30:38,600
for the fact that um natural language
685
00:30:36,279 --> 00:30:40,039
has a very longtailed distribution um
686
00:30:38,600 --> 00:30:41,600
you can kind of do the math and show
687
00:30:40,039 --> 00:30:43,720
that that works and that's actually in
688
00:30:41,600 --> 00:30:46,080
this um in this paper if you're
689
00:30:43,720 --> 00:30:49,880
interested in looking at more details of
690
00:30:46,080 --> 00:30:51,519
that um and then kind of the
691
00:30:49,880 --> 00:30:53,440
stateoftheart in language modeling
692
00:30:51,519 --> 00:30:56,600
before neural language models came out
693
00:30:53,440 --> 00:30:59,919
was this kesser smoothing and what it
694
00:30:56,600 --> 00:31:02,440
does is it discounts but it also
695
00:30:59,919 --> 00:31:04,480
modifies the lower order distribution so
696
00:31:02,440 --> 00:31:07,200
in the lower order distribution you
697
00:31:04,480 --> 00:31:09,039
basically um modify the counts with
698
00:31:07,200 --> 00:31:11,919
respect to how many times that word has
699
00:31:09,039 --> 00:31:13,519
appeared in new contexts with the IDE
700
00:31:11,919 --> 00:31:16,360
idea being that you only use the lower
701
00:31:13,519 --> 00:31:18,880
order distribution when you have uh new
702
00:31:16,360 --> 00:31:21,200
contexts um and so you can kind of Be
703
00:31:18,880 --> 00:31:23,600
Clever
704
00:31:21,200 --> 00:31:25,399
About You Can Be Clever about how you
705
00:31:23,600 --> 00:31:27,639
build this distribution based on the
706
00:31:25,399 --> 00:31:29,360
fact that you're only using it in the
707
00:31:27,639 --> 00:31:31,320
case when this distribution is not very
708
00:31:29,360 --> 00:31:33,960
Rel
709
00:31:31,320 --> 00:31:36,080
so I I would spend a lot more time
710
00:31:33,960 --> 00:31:37,960
teaching this when uh engram models were
711
00:31:36,080 --> 00:31:39,840
kind of the thing uh that people were
712
00:31:37,960 --> 00:31:41,960
using but now I'm going to go over them
713
00:31:39,840 --> 00:31:43,600
very quickly so you know don't worry if
714
00:31:41,960 --> 00:31:46,559
you weren't able to follow all the
715
00:31:43,600 --> 00:31:47,960
details but the basic um the basic thing
716
00:31:46,559 --> 00:31:49,279
take away from this is number one these
717
00:31:47,960 --> 00:31:51,639
are the methods that people use for
718
00:31:49,279 --> 00:31:53,440
engram language models number two if
719
00:31:51,639 --> 00:31:55,720
you're thinking about combining language
720
00:31:53,440 --> 00:31:57,519
models together in some way through you
721
00:31:55,720 --> 00:31:59,279
know ensembling their probability or
722
00:31:57,519 --> 00:32:00,480
something like this this is something
723
00:31:59,279 --> 00:32:02,279
that you should think about a little bit
724
00:32:00,480 --> 00:32:03,679
more carefully because like some
725
00:32:02,279 --> 00:32:05,240
language models might be good in some
726
00:32:03,679 --> 00:32:07,440
context other language models might be
727
00:32:05,240 --> 00:32:09,440
good in other contexts so you would need
728
00:32:07,440 --> 00:32:11,799
to think about that when you're doing um
729
00:32:09,440 --> 00:32:18,200
when you're combining the model
730
00:32:11,799 --> 00:32:18,200
that cool um any any questions about
731
00:32:19,080 --> 00:32:24,840
this Okay
732
00:32:21,159 --> 00:32:27,840
cool so there's a lot of problems that
733
00:32:24,840 --> 00:32:30,760
we have to deal with um when were
734
00:32:27,840 --> 00:32:32,600
creating engram models and that actually
735
00:32:30,760 --> 00:32:35,279
kind of motivated the reason why we
736
00:32:32,600 --> 00:32:36,639
moved to neural language models the
737
00:32:35,279 --> 00:32:38,720
first one is similar to what I talked
738
00:32:36,639 --> 00:32:40,519
about last time with text classification
739
00:32:38,720 --> 00:32:42,600
um that they can't share strength among
740
00:32:40,519 --> 00:32:45,159
similar words like bought and
741
00:32:42,600 --> 00:32:46,919
purchase um another thing is that they
742
00:32:45,159 --> 00:32:49,440
can't easily condition on context with
743
00:32:46,919 --> 00:32:51,240
intervening words so engram models if
744
00:32:49,440 --> 00:32:52,799
you have a rare word in your context
745
00:32:51,240 --> 00:32:54,320
immediately start falling back to the
746
00:32:52,799 --> 00:32:56,799
unigram distribution and they end up
747
00:32:54,320 --> 00:32:58,720
being very bad so uh that was another
748
00:32:56,799 --> 00:33:01,000
issue
749
00:32:58,720 --> 00:33:04,760
and they couldn't handle long distance
750
00:33:01,000 --> 00:33:09,080
um dependencies so if this was beyond
751
00:33:04,760 --> 00:33:10,559
the engram context that they would uh be
752
00:33:09,080 --> 00:33:14,320
handling then you wouldn't be able to
753
00:33:10,559 --> 00:33:15,840
manage this so actually before neural
754
00:33:14,320 --> 00:33:18,000
language models became a really big
755
00:33:15,840 --> 00:33:19,960
thing uh people came up with a bunch of
756
00:33:18,000 --> 00:33:22,760
individual solutions for this in order
757
00:33:19,960 --> 00:33:24,440
to solve the problems but actually it
758
00:33:22,760 --> 00:33:26,679
wasn't that these Solutions didn't work
759
00:33:24,440 --> 00:33:29,159
at all it was just that engineering all
760
00:33:26,679 --> 00:33:30,519
of them together was so hard that nobody
761
00:33:29,159 --> 00:33:32,120
actually ever did that and so they
762
00:33:30,519 --> 00:33:35,120
relied on just engram models out of the
763
00:33:32,120 --> 00:33:37,600
box and that wasn't scalable so it's
764
00:33:35,120 --> 00:33:39,279
kind of a funny example of how like
765
00:33:37,600 --> 00:33:42,000
actually neural networks despite all the
766
00:33:39,279 --> 00:33:43,559
pain that they cause in some areas are a
767
00:33:42,000 --> 00:33:47,120
much better engineering solution to
768
00:33:43,559 --> 00:33:51,279
solve all the issues that previous
769
00:33:47,120 --> 00:33:53,159
method cool um so when they use uh Eng
770
00:33:51,279 --> 00:33:54,799
grab models neural language models
771
00:33:53,159 --> 00:33:56,559
achieve better performance but Eng grab
772
00:33:54,799 --> 00:33:58,440
models are very very fast to estimate
773
00:33:56,559 --> 00:33:59,880
and apply you can even estimate them
774
00:33:58,440 --> 00:34:04,399
completely in
775
00:33:59,880 --> 00:34:07,720
parallel um engram models also I I don't
776
00:34:04,399 --> 00:34:10,399
know if this is necessarily
777
00:34:07,720 --> 00:34:13,200
A a thing that
778
00:34:10,399 --> 00:34:15,079
you a reason to use engram language
779
00:34:13,200 --> 00:34:17,720
models but it is a reason to think a
780
00:34:15,079 --> 00:34:20,320
little bit critically about uh neural
781
00:34:17,720 --> 00:34:22,720
language models which is neural language
782
00:34:20,320 --> 00:34:24,320
models actually can be worse than engram
783
00:34:22,720 --> 00:34:26,679
language models at modeling very low
784
00:34:24,320 --> 00:34:28,480
frequency phenomenas so engram language
785
00:34:26,679 --> 00:34:29,960
model can learn from a single example
786
00:34:28,480 --> 00:34:32,119
they only need a single example of
787
00:34:29,960 --> 00:34:36,879
anything before the probability of that
788
00:34:32,119 --> 00:34:38,639
continuation goes up very high um and uh
789
00:34:36,879 --> 00:34:41,359
but neural language models actually can
790
00:34:38,639 --> 00:34:43,599
forget or not memorize uh appropriately
791
00:34:41,359 --> 00:34:46,280
from single examples so they can be
792
00:34:43,599 --> 00:34:48,040
better at that um there's a toolkit the
793
00:34:46,280 --> 00:34:49,919
standard toolkit for estimating engram
794
00:34:48,040 --> 00:34:54,359
language models is called KLM it's kind
795
00:34:49,919 --> 00:34:57,599
of frighteningly fast um and so people
796
00:34:54,359 --> 00:35:00,400
have been uh saying like I've seen some
797
00:34:57,599 --> 00:35:01,599
jokes which are like job postings that
798
00:35:00,400 --> 00:35:04,040
say people who have been working on
799
00:35:01,599 --> 00:35:05,880
large language models uh for we want
800
00:35:04,040 --> 00:35:07,359
people who have been 10 years of
801
00:35:05,880 --> 00:35:09,240
experience working on large language
802
00:35:07,359 --> 00:35:11,960
models or something like that and a lot
803
00:35:09,240 --> 00:35:13,440
of people are saying wait nobody has 10
804
00:35:11,960 --> 00:35:16,400
years of experience working on large
805
00:35:13,440 --> 00:35:18,160
language models well Kenneth hfield who
806
00:35:16,400 --> 00:35:19,440
created KLM does have 10 years of
807
00:35:18,160 --> 00:35:22,800
experience working on large language
808
00:35:19,440 --> 00:35:24,599
models because he was estimating uh
809
00:35:22,800 --> 00:35:27,720
seven gr
810
00:35:24,599 --> 00:35:30,320
bottles um seven models with a
811
00:35:27,720 --> 00:35:35,040
vocabulary of let's say
812
00:35:30,320 --> 00:35:37,720
100,000 on um you know web text so how
813
00:35:35,040 --> 00:35:41,119
many parameters is at that's more than
814
00:35:37,720 --> 00:35:44,320
any you know large neural language model
815
00:35:41,119 --> 00:35:45,640
that we have nowadays so um they they
816
00:35:44,320 --> 00:35:47,520
have a lot of these parameters are
817
00:35:45,640 --> 00:35:49,400
sparse they're zero counts so obviously
818
00:35:47,520 --> 00:35:52,160
you don't uh you don't memorize all of
819
00:35:49,400 --> 00:35:55,040
them but uh
820
00:35:52,160 --> 00:35:57,800
yeah cool um another thing that maybe I
821
00:35:55,040 --> 00:35:59,359
should mention like so this doesn't
822
00:35:57,800 --> 00:36:01,960
sound completely outdated there was a
823
00:35:59,359 --> 00:36:05,400
really good paper
824
00:36:01,960 --> 00:36:08,400
recently that used the fact that engrams
825
00:36:05,400 --> 00:36:08,400
are
826
00:36:11,079 --> 00:36:17,319
so uses effect that engram models are so
827
00:36:14,280 --> 00:36:18,960
scalable it's this paper um it's called
828
00:36:17,319 --> 00:36:21,079
Data selection for language models via
829
00:36:18,960 --> 00:36:22,359
importance rese sampling and one
830
00:36:21,079 --> 00:36:24,359
interesting thing that they do in this
831
00:36:22,359 --> 00:36:28,920
paper is that they don't
832
00:36:24,359 --> 00:36:31,560
actually um they don't
833
00:36:28,920 --> 00:36:32,800
actually use neural models in any way
834
00:36:31,560 --> 00:36:34,920
despite the fact that they use the
835
00:36:32,800 --> 00:36:36,880
downstream data that they sample in
836
00:36:34,920 --> 00:36:41,319
order to calculate neural models but
837
00:36:36,880 --> 00:36:42,880
they run engram models over um over lots
838
00:36:41,319 --> 00:36:47,359
and lots of data and then they fit a
839
00:36:42,880 --> 00:36:50,000
gaussian distribution to the enr model
840
00:36:47,359 --> 00:36:51,520
counts basically uh in order to select
841
00:36:50,000 --> 00:36:53,040
the data in the reason why they do this
842
00:36:51,520 --> 00:36:55,280
is they want to do this over the entire
843
00:36:53,040 --> 00:36:56,760
web and running a neural model over the
844
00:36:55,280 --> 00:36:58,920
entire web would be too expensive so
845
00:36:56,760 --> 00:37:00,319
they use angr models instead so that's
846
00:36:58,920 --> 00:37:02,359
just an example of something in the
847
00:37:00,319 --> 00:37:04,920
modern context where keeping this in
848
00:37:02,359 --> 00:37:04,920
mind is a good
849
00:37:08,200 --> 00:37:14,000
idea okay I'd like to move to the next
850
00:37:10,960 --> 00:37:15,319
part so a language model evaluation uh
851
00:37:14,000 --> 00:37:17,200
this is important to know I'm not going
852
00:37:15,319 --> 00:37:19,079
to talk about language model evaluation
853
00:37:17,200 --> 00:37:20,599
on other tasks I'm only going to talk
854
00:37:19,079 --> 00:37:23,800
right now about language model
855
00:37:20,599 --> 00:37:26,280
evaluation on the task of language
856
00:37:23,800 --> 00:37:29,079
modeling and there's a number of metrics
857
00:37:26,280 --> 00:37:30,680
that we use for the task of language
858
00:37:29,079 --> 00:37:32,720
modeling evaluating language models on
859
00:37:30,680 --> 00:37:35,560
the task of language modeling the first
860
00:37:32,720 --> 00:37:38,480
one is log likelihood and basically uh
861
00:37:35,560 --> 00:37:40,160
the way we calculate log likelihood is
862
00:37:38,480 --> 00:37:41,640
uh sorry there's an extra parenthesis
863
00:37:40,160 --> 00:37:45,480
here but the way we calculate log
864
00:37:41,640 --> 00:37:47,160
likelihood is we get a test set that
865
00:37:45,480 --> 00:37:50,400
ideally has not been included in our
866
00:37:47,160 --> 00:37:52,520
training data and we take all of the
867
00:37:50,400 --> 00:37:54,200
documents or sentences in the test set
868
00:37:52,520 --> 00:37:57,040
we calculate the log probability of all
869
00:37:54,200 --> 00:37:59,520
of them uh we don't actually use this
870
00:37:57,040 --> 00:38:02,640
super broadly to evaluate models and the
871
00:37:59,520 --> 00:38:04,200
reason why is because this number is
872
00:38:02,640 --> 00:38:05,720
very dependent on the size of the data
873
00:38:04,200 --> 00:38:07,119
set so if you have a larger data set
874
00:38:05,720 --> 00:38:08,720
this number will be larger if you have a
875
00:38:07,119 --> 00:38:10,960
smaller data set this number will be
876
00:38:08,720 --> 00:38:14,040
smaller so the more common thing to do
877
00:38:10,960 --> 00:38:15,839
is per word uh log likelihood and per
878
00:38:14,040 --> 00:38:19,800
word log likelihood is basically
879
00:38:15,839 --> 00:38:22,760
dividing the um dividing the log
880
00:38:19,800 --> 00:38:25,520
probability of the entire corpus with uh
881
00:38:22,760 --> 00:38:28,359
the number of words that you have in the
882
00:38:25,520 --> 00:38:31,000
corpus
883
00:38:28,359 --> 00:38:34,599
um it's also common for papers to report
884
00:38:31,000 --> 00:38:36,359
negative log likelihood uh where because
885
00:38:34,599 --> 00:38:37,800
that's used as a loss and there lower is
886
00:38:36,359 --> 00:38:40,440
better so you just need to be careful
887
00:38:37,800 --> 00:38:42,560
about which one is being
888
00:38:40,440 --> 00:38:43,880
reported so this is pretty common I
889
00:38:42,560 --> 00:38:45,400
think most people are are somewhat
890
00:38:43,880 --> 00:38:49,040
familiar with
891
00:38:45,400 --> 00:38:49,800
this another thing that you might see is
892
00:38:49,040 --> 00:38:53,079
uh
893
00:38:49,800 --> 00:38:55,000
entropy and uh specifically this is
894
00:38:53,079 --> 00:38:57,319
often called cross entropy because
895
00:38:55,000 --> 00:38:59,880
you're calculating
896
00:38:57,319 --> 00:39:01,599
the you're estimating the model on a
897
00:38:59,880 --> 00:39:05,079
training data set and then evaluating it
898
00:39:01,599 --> 00:39:08,400
on a separate data set uh so uh on the
899
00:39:05,079 --> 00:39:12,200
test data set and this is calcul often
900
00:39:08,400 --> 00:39:14,640
or usually calculated as log 2 um of the
901
00:39:12,200 --> 00:39:17,119
probability divided by the number of
902
00:39:14,640 --> 00:39:18,760
words or units in the Corpus does anyone
903
00:39:17,119 --> 00:39:23,839
know why this is log
904
00:39:18,760 --> 00:39:23,839
two as opposed to a normal uh
905
00:39:25,440 --> 00:39:31,319
log
906
00:39:28,440 --> 00:39:31,319
anyone yeah
907
00:39:33,119 --> 00:39:38,720
so yeah so it's calculating as bits um
908
00:39:36,760 --> 00:39:43,160
and this is kind of
909
00:39:38,720 --> 00:39:45,240
a um this is kind of a historical thing
910
00:39:43,160 --> 00:39:47,119
and it's not super super important for
911
00:39:45,240 --> 00:39:51,800
language models but it's actually pretty
912
00:39:47,119 --> 00:39:54,599
interesting uh to to think about and so
913
00:39:51,800 --> 00:39:57,480
actually any probabilistic distribution
914
00:39:54,599 --> 00:40:00,040
can also be used for data compression
915
00:39:57,480 --> 00:40:03,319
um and so you know when you're running a
916
00:40:00,040 --> 00:40:05,000
zip file or you're running gzip or bz2
917
00:40:03,319 --> 00:40:07,359
or something like that uh you're
918
00:40:05,000 --> 00:40:09,240
compressing a file into a smaller file
919
00:40:07,359 --> 00:40:12,000
and any language model can also be used
920
00:40:09,240 --> 00:40:15,280
to compress a SM file into a smaller
921
00:40:12,000 --> 00:40:17,119
file um and so the way it does this is
922
00:40:15,280 --> 00:40:19,200
if you have more likely
923
00:40:17,119 --> 00:40:20,960
sequences uh for example more likely
924
00:40:19,200 --> 00:40:25,079
sentences or more likely documents you
925
00:40:20,960 --> 00:40:26,920
can press them into a a shorter uh
926
00:40:25,079 --> 00:40:29,440
output and
927
00:40:26,920 --> 00:40:29,440
kind of
928
00:40:29,640 --> 00:40:33,800
the
929
00:40:31,480 --> 00:40:35,720
ideal I I think it's pretty safe to say
930
00:40:33,800 --> 00:40:37,920
ideal because I think you can't get a
931
00:40:35,720 --> 00:40:42,920
better method for compression than this
932
00:40:37,920 --> 00:40:45,000
uh if I unless I'm uh you know not well
933
00:40:42,920 --> 00:40:46,800
versed enough in information Theory but
934
00:40:45,000 --> 00:40:49,240
I I think this is basically the ideal
935
00:40:46,800 --> 00:40:51,960
method for data compression and the way
936
00:40:49,240 --> 00:40:54,640
it works is um I have a figure up here
937
00:40:51,960 --> 00:40:58,800
but I'd like to recreate it here which
938
00:40:54,640 --> 00:41:02,640
is let's say we have a vocabulary of
939
00:40:58,800 --> 00:41:07,200
a um which has
940
00:41:02,640 --> 00:41:08,800
50% and then we have a vocabulary uh B
941
00:41:07,200 --> 00:41:11,560
which is
942
00:41:08,800 --> 00:41:14,040
33% and a vocabulary
943
00:41:11,560 --> 00:41:18,520
C
944
00:41:14,040 --> 00:41:18,520
uh yeah C which is about
945
00:41:18,640 --> 00:41:25,640
17% and so if you have a single token
946
00:41:22,960 --> 00:41:26,839
sequence um if you have a single token
947
00:41:25,640 --> 00:41:30,880
sequence
948
00:41:26,839 --> 00:41:30,880
what you do is you can
949
00:41:31,319 --> 00:41:38,800
see divide this into zero and one so if
950
00:41:36,400 --> 00:41:40,680
your single token sequence is a you can
951
00:41:38,800 --> 00:41:42,760
just put zero and you'll be done
952
00:41:40,680 --> 00:41:46,800
encoding it if your single token
953
00:41:42,760 --> 00:41:51,920
sequence is B
954
00:41:46,800 --> 00:41:56,520
then um one overlaps with b and c so now
955
00:41:51,920 --> 00:42:00,920
you need to further split this up into
956
00:41:56,520 --> 00:42:00,920
uh o and one and you can see
957
00:42:04,880 --> 00:42:11,440
that let make sure I did that right yeah
958
00:42:08,359 --> 00:42:11,440
you can you can see
959
00:42:15,599 --> 00:42:25,720
that one zero is entirely encompassed by
960
00:42:19,680 --> 00:42:29,200
uh by B so now B is one Z and C uh C is
961
00:42:25,720 --> 00:42:32,359
not L encompassed by that so you would
962
00:42:29,200 --> 00:42:39,240
need to further break this up and say
963
00:42:32,359 --> 00:42:41,880
it's Z one here and now one one
964
00:42:39,240 --> 00:42:45,520
one is encompassed by this so you would
965
00:42:41,880 --> 00:42:48,680
get uh you would get C if it was 111 and
966
00:42:45,520 --> 00:42:51,119
so every every sequence that started
967
00:42:48,680 --> 00:42:53,000
with zero would start out with a every
968
00:42:51,119 --> 00:42:54,960
sequence that started out with one zero
969
00:42:53,000 --> 00:42:57,200
would start with b and every sequence
970
00:42:54,960 --> 00:43:02,079
that started with 11 one1
971
00:42:57,200 --> 00:43:04,920
start um and so then you can look at the
972
00:43:02,079 --> 00:43:06,960
next word and let's say we're using a
973
00:43:04,920 --> 00:43:09,839
unigram model if we're using a unigram
974
00:43:06,960 --> 00:43:12,960
model for the next uh the next token
975
00:43:09,839 --> 00:43:18,200
let's say the next token is C
976
00:43:12,960 --> 00:43:23,640
so now the next token being C we already
977
00:43:18,200 --> 00:43:27,920
have B and now we take we subdivide
978
00:43:23,640 --> 00:43:33,040
B into
979
00:43:27,920 --> 00:43:35,720
a BC ba a BB and BC and then we find the
980
00:43:33,040 --> 00:43:40,720
next binary sequence that is entirely
981
00:43:35,720 --> 00:43:44,000
encompassed by uh BC by this like
982
00:43:40,720 --> 00:43:45,359
interval and so the moment we find a a
983
00:43:44,000 --> 00:43:48,520
binary sequence that's entirely
984
00:43:45,359 --> 00:43:50,599
encompassed by the interval uh then that
985
00:43:48,520 --> 00:43:53,400
is the the sequence that we can use to
986
00:43:50,599 --> 00:43:54,640
represent that SC and so um if you're
987
00:43:53,400 --> 00:43:56,520
interested in this you can look up the
988
00:43:54,640 --> 00:44:00,400
arithmetic coding on on wikip it's
989
00:43:56,520 --> 00:44:02,079
pretty fascinating but basically um here
990
00:44:00,400 --> 00:44:04,040
this is showing the example of the
991
00:44:02,079 --> 00:44:07,160
unigram model where the probabilities
992
00:44:04,040 --> 00:44:10,240
don't change based on the context but
993
00:44:07,160 --> 00:44:13,000
what if we knew that
994
00:44:10,240 --> 00:44:15,599
c had a really high probability of
995
00:44:13,000 --> 00:44:22,160
following B so if that's the case now we
996
00:44:15,599 --> 00:44:24,559
have like a a b c here um like based on
997
00:44:22,160 --> 00:44:25,880
our our byr model or neural language
998
00:44:24,559 --> 00:44:29,319
model or something like that so now this
999
00:44:25,880 --> 00:44:31,240
is interval is much much larger so it's
1000
00:44:29,319 --> 00:44:35,079
much more likely to entirely Encompass a
1001
00:44:31,240 --> 00:44:39,720
shorter string and because of that the
1002
00:44:35,079 --> 00:44:42,440
um the output can be much shorter and so
1003
00:44:39,720 --> 00:44:45,760
if you use this arithmetic encoding um
1004
00:44:42,440 --> 00:44:49,440
over a very long sequence of outputs
1005
00:44:45,760 --> 00:44:52,440
your the length of the sequence that is
1006
00:44:49,440 --> 00:44:56,000
needed to encode this uh this particular
1007
00:44:52,440 --> 00:45:00,359
output is going to be essentially um the
1008
00:44:56,000 --> 00:45:03,319
number of bits according to times the
1009
00:45:00,359 --> 00:45:06,480
times the sequence so this is very
1010
00:45:03,319 --> 00:45:10,000
directly connected to like compression
1011
00:45:06,480 --> 00:45:13,160
and information Theory and stuff like
1012
00:45:10,000 --> 00:45:15,359
that so that that's where entropy comes
1013
00:45:13,160 --> 00:45:17,680
from uh are are there any questions
1014
00:45:15,359 --> 00:45:17,680
about
1015
00:45:19,319 --> 00:45:22,319
this
1016
00:45:24,880 --> 00:45:28,119
yeah
1017
00:45:26,800 --> 00:45:31,880
uh for
1018
00:45:28,119 --> 00:45:34,319
c um so
1019
00:45:31,880 --> 00:45:36,599
111 is
1020
00:45:34,319 --> 00:45:37,920
because let me let me see if I can do
1021
00:45:36,599 --> 00:45:40,559
this
1022
00:45:37,920 --> 00:45:44,240
again
1023
00:45:40,559 --> 00:45:44,240
so I had one
1024
00:45:46,079 --> 00:45:54,520
one so here this interval is
1025
00:45:50,920 --> 00:45:56,839
one this interval is one one this
1026
00:45:54,520 --> 00:46:00,079
interval is 111
1027
00:45:56,839 --> 00:46:03,520
and 111 is the first interval that is
1028
00:46:00,079 --> 00:46:05,520
entirely overlapping with with c um and
1029
00:46:03,520 --> 00:46:08,760
it's not one Z because one one Z is
1030
00:46:05,520 --> 00:46:08,760
overlaping with b and
1031
00:46:09,960 --> 00:46:13,599
c so which
1032
00:46:14,280 --> 00:46:21,720
Cas so which case one
1033
00:46:20,160 --> 00:46:24,800
Z
1034
00:46:21,720 --> 00:46:26,319
one one one
1035
00:46:24,800 --> 00:46:30,800
Z
1036
00:46:26,319 --> 00:46:30,800
when would you use 110 to represent
1037
00:46:32,119 --> 00:46:38,839
something it's a good question I guess
1038
00:46:36,119 --> 00:46:40,599
maybe you wouldn't which seems a little
1039
00:46:38,839 --> 00:46:43,280
bit wasteful
1040
00:46:40,599 --> 00:46:46,160
so let me let me think about that I
1041
00:46:43,280 --> 00:46:49,920
think um it might be the case that you
1042
00:46:46,160 --> 00:46:52,319
just don't use it um
1043
00:46:49,920 --> 00:46:53,559
but yeah I'll try to think about that a
1044
00:46:52,319 --> 00:46:55,920
little bit more because it seems like
1045
00:46:53,559 --> 00:46:59,200
you should use every bet string right so
1046
00:46:55,920 --> 00:47:01,559
um yeah if anybody uh has has the answer
1047
00:46:59,200 --> 00:47:05,160
I'd be happy to hear it otherwise I take
1048
00:47:01,559 --> 00:47:07,079
you cool um so next thing is perplexity
1049
00:47:05,160 --> 00:47:10,640
so this is another one that you see
1050
00:47:07,079 --> 00:47:13,240
commonly and um so perplexity is
1051
00:47:10,640 --> 00:47:16,880
basically two to the ENT uh two to the
1052
00:47:13,240 --> 00:47:20,760
per word entropy or e to the uh negative
1053
00:47:16,880 --> 00:47:24,880
word level log likelihood in log space
1054
00:47:20,760 --> 00:47:28,240
um and so this uh T larger tends to be
1055
00:47:24,880 --> 00:47:32,559
better I'd like to do a little exercise
1056
00:47:28,240 --> 00:47:34,599
to see uh if this works so like let's
1057
00:47:32,559 --> 00:47:39,079
say we have one a dog sees a squirrel it
1058
00:47:34,599 --> 00:47:40,960
will usually um and can anyone guess the
1059
00:47:39,079 --> 00:47:43,480
next word just yell it
1060
00:47:40,960 --> 00:47:46,400
out bar
1061
00:47:43,480 --> 00:47:47,400
okay uh what about that what about
1062
00:47:46,400 --> 00:47:50,400
something
1063
00:47:47,400 --> 00:47:50,400
else
1064
00:47:52,640 --> 00:47:57,520
Chase Run
1065
00:47:54,720 --> 00:48:00,800
Run
1066
00:47:57,520 --> 00:48:00,800
okay John
1067
00:48:01,960 --> 00:48:05,280
John anything
1068
00:48:07,000 --> 00:48:10,400
else any other
1069
00:48:11,280 --> 00:48:16,960
ones so basically what this shows is
1070
00:48:13,640 --> 00:48:16,960
humans are really bad language
1071
00:48:17,160 --> 00:48:24,079
models so uh interestingly every single
1072
00:48:21,520 --> 00:48:26,559
one of the words you predicted here is a
1073
00:48:24,079 --> 00:48:32,240
uh a regular verb
1074
00:48:26,559 --> 00:48:35,200
um but in natural language model gpt2 uh
1075
00:48:32,240 --> 00:48:38,079
the first thing it predicts is B uh
1076
00:48:35,200 --> 00:48:40,440
which is kind of a like the Cula there's
1077
00:48:38,079 --> 00:48:43,400
also start and that will be like start
1078
00:48:40,440 --> 00:48:44,880
running start something um and humans
1079
00:48:43,400 --> 00:48:46,400
actually are really bad at doing this
1080
00:48:44,880 --> 00:48:49,079
are really bad at predicting next words
1081
00:48:46,400 --> 00:48:51,760
we're not trained that way um and so uh
1082
00:48:49,079 --> 00:48:54,319
we end up having these biases but anyway
1083
00:48:51,760 --> 00:48:55,799
um the reason why I did this quiz was
1084
00:48:54,319 --> 00:48:57,280
because that's essentially what
1085
00:48:55,799 --> 00:49:01,160
perplexity
1086
00:48:57,280 --> 00:49:02,680
means um and what what perplexity is is
1087
00:49:01,160 --> 00:49:04,559
it's the number of times you'd have to
1088
00:49:02,680 --> 00:49:07,000
sample from the probability distribution
1089
00:49:04,559 --> 00:49:09,200
before you get the answer right so you
1090
00:49:07,000 --> 00:49:11,160
were a little bit biased here because we
1091
00:49:09,200 --> 00:49:13,359
were doing sampling without replacement
1092
00:49:11,160 --> 00:49:15,480
so like nobody was actually picking a
1093
00:49:13,359 --> 00:49:17,000
word that had already been said but it's
1094
00:49:15,480 --> 00:49:18,319
essentially like if you guessed over and
1095
00:49:17,000 --> 00:49:20,839
over and over again how many times would
1096
00:49:18,319 --> 00:49:22,720
you need until you get it right and so
1097
00:49:20,839 --> 00:49:25,119
here like if the actual answer was start
1098
00:49:22,720 --> 00:49:27,480
the perplexity would be 4.66 so we'd
1099
00:49:25,119 --> 00:49:30,240
expect language model to get it in uh
1100
00:49:27,480 --> 00:49:34,400
four guesses uh between four and five
1101
00:49:30,240 --> 00:49:38,559
guesses and you guys all did six so you
1102
00:49:34,400 --> 00:49:41,599
lose um so uh another important thing to
1103
00:49:38,559 --> 00:49:42,799
mention is evaluation in vocabulary uh
1104
00:49:41,599 --> 00:49:44,880
so for fair
1105
00:49:42,799 --> 00:49:47,319
comparison um make sure that the
1106
00:49:44,880 --> 00:49:49,559
denominator is the same so uh if you're
1107
00:49:47,319 --> 00:49:51,559
calculating the perplexity make sure
1108
00:49:49,559 --> 00:49:53,359
that you're dividing by the same number
1109
00:49:51,559 --> 00:49:55,799
uh every time you're dividing by words
1110
00:49:53,359 --> 00:49:58,520
if it's uh the other paper or whatever
1111
00:49:55,799 --> 00:50:00,680
is dividing by words or like let's say
1112
00:49:58,520 --> 00:50:02,160
you're comparing llama to gp2 they have
1113
00:50:00,680 --> 00:50:04,880
different tokenizers so they'll have
1114
00:50:02,160 --> 00:50:07,040
different numbers of tokens so comparing
1115
00:50:04,880 --> 00:50:10,880
uh with different denominators is not uh
1116
00:50:07,040 --> 00:50:12,440
not fair um if you're allowing unknown
1117
00:50:10,880 --> 00:50:14,559
words or characters so if you allow the
1118
00:50:12,440 --> 00:50:17,640
model to not predict
1119
00:50:14,559 --> 00:50:19,119
any token then you need to be fair about
1120
00:50:17,640 --> 00:50:22,040
that
1121
00:50:19,119 --> 00:50:25,160
too um so I'd like to go into a few
1122
00:50:22,040 --> 00:50:27,960
Alternatives these are very similar to
1123
00:50:25,160 --> 00:50:29,400
the Network classifiers and bag of words
1124
00:50:27,960 --> 00:50:30,680
classifiers that I talked about before
1125
00:50:29,400 --> 00:50:32,480
so I'm going to go through them rather
1126
00:50:30,680 --> 00:50:35,480
quickly because I think you should get
1127
00:50:32,480 --> 00:50:38,119
the basic idea but basically the
1128
00:50:35,480 --> 00:50:40,000
alternative is uh featued models so we
1129
00:50:38,119 --> 00:50:42,559
calculate features of to account based
1130
00:50:40,000 --> 00:50:44,599
models as featued models so we calculate
1131
00:50:42,559 --> 00:50:46,880
features of the context and based on the
1132
00:50:44,599 --> 00:50:48,280
features calculate probabilities
1133
00:50:46,880 --> 00:50:50,480
optimize the feature weights using
1134
00:50:48,280 --> 00:50:53,839
gradient descent uh
1135
00:50:50,480 --> 00:50:56,119
Etc and so for example if we have uh
1136
00:50:53,839 --> 00:50:58,880
input giving a
1137
00:50:56,119 --> 00:51:02,960
uh we calculate features so um we might
1138
00:50:58,880 --> 00:51:05,400
look up uh the word identity of the two
1139
00:51:02,960 --> 00:51:08,240
previous words look up the word identity
1140
00:51:05,400 --> 00:51:11,000
of the word uh directly previous add a
1141
00:51:08,240 --> 00:51:13,480
bias add them all together get scores
1142
00:51:11,000 --> 00:51:14,960
and calculate probabilities where each
1143
00:51:13,480 --> 00:51:16,920
Vector is the size of the output
1144
00:51:14,960 --> 00:51:19,680
vocabulary and feature weights are
1145
00:51:16,920 --> 00:51:21,799
optimized using SGD so this is basically
1146
00:51:19,680 --> 00:51:24,240
a bag of words classifier but it's a
1147
00:51:21,799 --> 00:51:27,200
multiclass bag of words classifier over
1148
00:51:24,240 --> 00:51:28,960
the next token so it's very similar to
1149
00:51:27,200 --> 00:51:30,839
our classification task before except
1150
00:51:28,960 --> 00:51:33,160
now instead of having two classes we
1151
00:51:30,839 --> 00:51:36,280
have you know 10,000 classes or 100,000
1152
00:51:33,160 --> 00:51:38,480
classes oh yeah sorry very quick aside
1153
00:51:36,280 --> 00:51:40,280
um these were actually invented by Rony
1154
00:51:38,480 --> 00:51:41,440
Rosenfeld who's the head of the machine
1155
00:51:40,280 --> 00:51:45,119
learning department at the end the
1156
00:51:41,440 --> 00:51:47,799
machine learning Department uh so um 27
1157
00:51:45,119 --> 00:51:50,760
years ago I guess so he has even more
1158
00:51:47,799 --> 00:51:52,680
experience large language modeling than
1159
00:51:50,760 --> 00:51:55,880
um
1160
00:51:52,680 --> 00:51:58,599
cool so um the one difference with a bag
1161
00:51:55,880 --> 00:52:02,119
of words classifier is
1162
00:51:58,599 --> 00:52:05,480
um we we have
1163
00:52:02,119 --> 00:52:07,640
biases um and we have the probability
1164
00:52:05,480 --> 00:52:09,400
Vector given the previous word but
1165
00:52:07,640 --> 00:52:11,720
instead of using a bag of words this
1166
00:52:09,400 --> 00:52:15,440
actually is using uh How likely is it
1167
00:52:11,720 --> 00:52:16,960
giving given two words previous so uh
1168
00:52:15,440 --> 00:52:18,040
the feature design would be a little bit
1169
00:52:16,960 --> 00:52:19,119
different and that would give you a
1170
00:52:18,040 --> 00:52:22,920
total
1171
00:52:19,119 --> 00:52:24,359
score um as a reminder uh last time we
1172
00:52:22,920 --> 00:52:26,440
did a training algorithm where we
1173
00:52:24,359 --> 00:52:27,480
calculated gradients loss function with
1174
00:52:26,440 --> 00:52:29,960
respect to the
1175
00:52:27,480 --> 00:52:32,319
parameters and uh we can use the chain
1176
00:52:29,960 --> 00:52:33,839
Rule and back propagation and updates to
1177
00:52:32,319 --> 00:52:36,400
move in the direction that increases
1178
00:52:33,839 --> 00:52:39,040
enough so nothing extremely different
1179
00:52:36,400 --> 00:52:42,640
from what we had for our
1180
00:52:39,040 --> 00:52:44,240
B um similarly this solves some problems
1181
00:52:42,640 --> 00:52:47,240
so this didn't solve the problem of
1182
00:52:44,240 --> 00:52:49,119
sharing strength among similar words it
1183
00:52:47,240 --> 00:52:50,839
did solve the problem of conditioning on
1184
00:52:49,119 --> 00:52:52,839
context with intervening words because
1185
00:52:50,839 --> 00:52:56,920
now we can condition directly on Doctor
1186
00:52:52,839 --> 00:52:59,680
without having to um combine with
1187
00:52:56,920 --> 00:53:01,200
gitrid um and it doesn't necessarily
1188
00:52:59,680 --> 00:53:03,480
handle longdistance dependencies because
1189
00:53:01,200 --> 00:53:05,240
we're still limited in our context with
1190
00:53:03,480 --> 00:53:09,079
the model I just
1191
00:53:05,240 --> 00:53:11,920
described so um if we so sorry back to
1192
00:53:09,079 --> 00:53:13,480
neural networks is what I should say um
1193
00:53:11,920 --> 00:53:15,160
so if we have a feedforward neural
1194
00:53:13,480 --> 00:53:18,480
network language model the way this
1195
00:53:15,160 --> 00:53:20,400
could work is instead of looking up
1196
00:53:18,480 --> 00:53:23,079
discrete features uh like we had in a
1197
00:53:20,400 --> 00:53:25,960
bag of words model uh we would look up
1198
00:53:23,079 --> 00:53:27,400
dents embeddings and so we concatenate
1199
00:53:25,960 --> 00:53:29,359
together these dense
1200
00:53:27,400 --> 00:53:32,319
embeddings and based on the dense
1201
00:53:29,359 --> 00:53:34,599
embeddings uh we do some sort of uh
1202
00:53:32,319 --> 00:53:36,079
intermediate layer transforms to extract
1203
00:53:34,599 --> 00:53:37,200
features like we did for our neural
1204
00:53:36,079 --> 00:53:39,359
network based
1205
00:53:37,200 --> 00:53:41,520
classifier um we multiply this by
1206
00:53:39,359 --> 00:53:43,559
weights uh we have a bias and we
1207
00:53:41,520 --> 00:53:46,559
calculate
1208
00:53:43,559 --> 00:53:49,200
scores and uh then we take a soft Max to
1209
00:53:46,559 --> 00:53:49,200
do
1210
00:53:50,400 --> 00:53:55,799
classification so um this can calculate
1211
00:53:53,359 --> 00:53:58,000
combination features uh like we we also
1212
00:53:55,799 --> 00:54:02,280
used in our uh neural network based
1213
00:53:58,000 --> 00:54:04,119
classifiers so um this could uh give us
1214
00:54:02,280 --> 00:54:05,760
a positive number for example if the
1215
00:54:04,119 --> 00:54:07,760
previous word is a determiner and the
1216
00:54:05,760 --> 00:54:10,440
second previous word is a verb so that
1217
00:54:07,760 --> 00:54:14,520
would be like uh in giving and then that
1218
00:54:10,440 --> 00:54:14,520
would allow us upway to that particular
1219
00:54:15,000 --> 00:54:19,559
examples um so this allows us to share
1220
00:54:17,640 --> 00:54:21,640
strength in various places in our model
1221
00:54:19,559 --> 00:54:23,520
which was also You Know instrumental in
1222
00:54:21,640 --> 00:54:25,599
making our our neural network
1223
00:54:23,520 --> 00:54:28,000
classifiers work for similar work and
1224
00:54:25,599 --> 00:54:30,119
stuff and so these would be word
1225
00:54:28,000 --> 00:54:32,160
embeddings so similar words get similar
1226
00:54:30,119 --> 00:54:35,079
embeddings another really important
1227
00:54:32,160 --> 00:54:38,480
thing is uh similar output words also
1228
00:54:35,079 --> 00:54:41,839
get similar rows in The softmax Matrix
1229
00:54:38,480 --> 00:54:44,440
and so here remember if you remember
1230
00:54:41,839 --> 00:54:48,240
from last class this was a big Matrix
1231
00:54:44,440 --> 00:54:50,400
where the size of the Matrix was the
1232
00:54:48,240 --> 00:54:53,319
number of vocabulary items times the
1233
00:54:50,400 --> 00:54:55,920
size of a word embedding this is also a
1234
00:54:53,319 --> 00:54:58,319
matrix where this is
1235
00:54:55,920 --> 00:55:02,200
the number of vocabulary items times the
1236
00:54:58,319 --> 00:55:04,160
size of a context embedding gr and so
1237
00:55:02,200 --> 00:55:06,160
these will also be similar because words
1238
00:55:04,160 --> 00:55:08,280
that appear in similar contexts will
1239
00:55:06,160 --> 00:55:11,920
also you know want similar embeddings so
1240
00:55:08,280 --> 00:55:15,119
they get uploaded in at the same
1241
00:55:11,920 --> 00:55:17,119
time and similar hidden States will have
1242
00:55:15,119 --> 00:55:19,799
similar context so ideally like if you
1243
00:55:17,119 --> 00:55:20,920
have giving a or delivering a or
1244
00:55:19,799 --> 00:55:22,680
something like that those would be
1245
00:55:20,920 --> 00:55:27,000
similar contexts so they would get
1246
00:55:22,680 --> 00:55:27,000
similar purple embeddings out out of the
1247
00:55:28,440 --> 00:55:31,599
so one trick that's widely used in
1248
00:55:30,200 --> 00:55:34,960
language model that further takes
1249
00:55:31,599 --> 00:55:38,799
advantage of this is uh tying
1250
00:55:34,960 --> 00:55:44,160
embeddings so here what this does is
1251
00:55:38,799 --> 00:55:48,280
sharing parameters between this um
1252
00:55:44,160 --> 00:55:49,920
lookup Matrix here and this uh Matrix
1253
00:55:48,280 --> 00:55:51,119
over here that we use for calculating
1254
00:55:49,920 --> 00:55:56,200
the
1255
00:55:51,119 --> 00:55:58,839
softmax and um the reason why this is
1256
00:55:56,200 --> 00:56:00,559
useful is twofold number one it gives
1257
00:55:58,839 --> 00:56:02,079
you essentially more training data to
1258
00:56:00,559 --> 00:56:04,440
learn these embeddings because instead
1259
00:56:02,079 --> 00:56:05,799
of learning the embeddings whenever a
1260
00:56:04,440 --> 00:56:08,520
word is in
1261
00:56:05,799 --> 00:56:10,599
context separately from learning the
1262
00:56:08,520 --> 00:56:13,520
embeddings whenever a word is predicted
1263
00:56:10,599 --> 00:56:15,480
you learn the the same embedding Matrix
1264
00:56:13,520 --> 00:56:19,319
whenever the word is in the context or
1265
00:56:15,480 --> 00:56:21,520
whatever it's predicted and so um that
1266
00:56:19,319 --> 00:56:24,119
makes it more accurate to learn these uh
1267
00:56:21,520 --> 00:56:26,960
embeddings well another thing is the
1268
00:56:24,119 --> 00:56:31,119
embedding mat can actually be very large
1269
00:56:26,960 --> 00:56:34,920
so like let's say we have aab of
1270
00:56:31,119 --> 00:56:37,520
10 100,000 and we have an embedding a
1271
00:56:34,920 --> 00:56:40,799
word embedding size of like 512 or
1272
00:56:37,520 --> 00:56:45,319
something like that
1273
00:56:40,799 --> 00:56:45,319
that's um 51 million
1274
00:56:46,839 --> 00:56:52,440
parameters um and this doesn't sound
1275
00:56:49,559 --> 00:56:55,520
like a lot of parameters at first but it
1276
00:56:52,440 --> 00:56:57,880
actually is a lot to learn when um
1277
00:56:55,520 --> 00:57:01,000
these get updated relatively
1278
00:56:57,880 --> 00:57:03,400
infrequently uh because
1279
00:57:01,000 --> 00:57:06,079
um these get updated relatively
1280
00:57:03,400 --> 00:57:07,960
infrequently because they only are
1281
00:57:06,079 --> 00:57:09,559
updated whenever that word or token
1282
00:57:07,960 --> 00:57:12,319
actually appears in your training data
1283
00:57:09,559 --> 00:57:14,119
so um this can be a good thing for
1284
00:57:12,319 --> 00:57:16,319
parameter savings parameter efficiency
1285
00:57:14,119 --> 00:57:16,319
as
1286
00:57:16,440 --> 00:57:22,520
well um so this uh solves most of the
1287
00:57:19,599 --> 00:57:24,319
problems here um but it doesn't solve
1288
00:57:22,520 --> 00:57:26,839
the problem of longdistance dependencies
1289
00:57:24,319 --> 00:57:29,839
because still limited by the overall
1290
00:57:26,839 --> 00:57:31,359
length of uh the context that we're
1291
00:57:29,839 --> 00:57:32,520
concatenating together here sure we
1292
00:57:31,359 --> 00:57:35,760
could make that longer but that would
1293
00:57:32,520 --> 00:57:37,200
make our model larger and um and bring
1294
00:57:35,760 --> 00:57:39,720
various
1295
00:57:37,200 --> 00:57:42,520
issues and so what I'm going to talk
1296
00:57:39,720 --> 00:57:44,599
about in on thur day is how we solve
1297
00:57:42,520 --> 00:57:47,559
this problem of modeling long contexts
1298
00:57:44,599 --> 00:57:49,720
so how do we um build recurrent neural
1299
00:57:47,559 --> 00:57:52,559
networks uh how do we build
1300
00:57:49,720 --> 00:57:54,960
convolutional uh convolutional networks
1301
00:57:52,559 --> 00:57:57,520
or how do we build attention based
1302
00:57:54,960 --> 00:58:00,720
Transformer models and these are all
1303
00:57:57,520 --> 00:58:02,119
options that are used um Transformers
1304
00:58:00,720 --> 00:58:04,359
are kind of
1305
00:58:02,119 --> 00:58:06,039
the the main thing that people use
1306
00:58:04,359 --> 00:58:08,400
nowadays but there's a lot of versions
1307
00:58:06,039 --> 00:58:11,880
of Transformers that borrow ideas from
1308
00:58:08,400 --> 00:58:14,960
recurrent uh and convolutional models
1309
00:58:11,880 --> 00:58:17,359
um recently a lot of long context models
1310
00:58:14,960 --> 00:58:19,440
us use ideas from recurrent networks and
1311
00:58:17,359 --> 00:58:22,160
a lot of for example speech models or
1312
00:58:19,440 --> 00:58:24,160
things like or image models use ideas
1313
00:58:22,160 --> 00:58:25,920
from convolutional networks so I think
1314
00:58:24,160 --> 00:58:28,760
learning all but at the same time is a
1315
00:58:25,920 --> 00:58:32,160
good idea in comparing
1316
00:58:28,760 --> 00:58:34,319
them cool uh any any questions about
1317
00:58:32,160 --> 00:58:35,799
this part I went through this kind of
1318
00:58:34,319 --> 00:58:37,319
quickly because it's pretty similar to
1319
00:58:35,799 --> 00:58:40,079
the the classification stuff that we
1320
00:58:37,319 --> 00:58:42,680
covered last time but uh any any things
1321
00:58:40,079 --> 00:58:42,680
that people want to
1322
00:58:43,880 --> 00:58:49,039
ask okay so next I'm going to talk about
1323
00:58:46,839 --> 00:58:51,559
a few other desiderata of language
1324
00:58:49,039 --> 00:58:53,039
models so the next one is really really
1325
00:58:51,559 --> 00:58:55,640
important it's a concept I want
1326
00:58:53,039 --> 00:58:57,640
everybody to know I actually
1327
00:58:55,640 --> 00:58:59,520
taught this informally up until this
1328
00:58:57,640 --> 00:59:02,039
class but now I I actually made slides
1329
00:58:59,520 --> 00:59:05,079
for it starting this time which is
1330
00:59:02,039 --> 00:59:07,240
calibration so the idea of calibration
1331
00:59:05,079 --> 00:59:10,200
is that the model quote unquote knows
1332
00:59:07,240 --> 00:59:14,559
when it knows or the the fact that it is
1333
00:59:10,200 --> 00:59:17,480
able to provide a a good answer um uh
1334
00:59:14,559 --> 00:59:21,640
provide a good confidence in its answer
1335
00:59:17,480 --> 00:59:23,640
and more formally this can be specified
1336
00:59:21,640 --> 00:59:25,240
as
1337
00:59:23,640 --> 00:59:27,799
the
1338
00:59:25,240 --> 00:59:29,200
feature that the model probability of
1339
00:59:27,799 --> 00:59:33,119
the answer matches the actual
1340
00:59:29,200 --> 00:59:37,319
probability of getting it right um and
1341
00:59:33,119 --> 00:59:37,319
so what this means
1342
00:59:41,960 --> 00:59:47,480
is the
1343
00:59:44,240 --> 00:59:51,839
probability of the
1344
00:59:47,480 --> 00:59:51,839
answer um is
1345
00:59:52,720 --> 00:59:59,880
correct given the fact that
1346
00:59:56,319 --> 00:59:59,880
the model
1347
01:00:00,160 --> 01:00:07,440
probability is equal to
1348
01:00:03,640 --> 01:00:07,440
P is equal to
1349
01:00:08,559 --> 01:00:12,760
ke
1350
01:00:10,480 --> 01:00:15,319
so I know this is a little bit hard to
1351
01:00:12,760 --> 01:00:18,240
parse I it always took me like a few
1352
01:00:15,319 --> 01:00:21,720
seconds to parse before I uh like when I
1353
01:00:18,240 --> 01:00:25,160
looked at it but basically if the model
1354
01:00:21,720 --> 01:00:26,920
if the model says the probability of it
1355
01:00:25,160 --> 01:00:29,440
being correct is
1356
01:00:26,920 --> 01:00:33,559
0.7 then the probability that the answer
1357
01:00:29,440 --> 01:00:35,960
is correct is actually 0.7 so um you
1358
01:00:33,559 --> 01:00:41,520
know if it says uh the probability is
1359
01:00:35,960 --> 01:00:41,520
0.7 100 times then it will be right 70
1360
01:00:43,640 --> 01:00:52,160
times and so the way we formalize this
1361
01:00:48,039 --> 01:00:55,200
um is is by this uh it was proposed by
1362
01:00:52,160 --> 01:00:57,760
this seminal paper by gu it all in
1363
01:00:55,200 --> 01:01:00,319
2017
1364
01:00:57,760 --> 01:01:03,319
and
1365
01:01:00,319 --> 01:01:05,520
unfortunately this data itself is hard
1366
01:01:03,319 --> 01:01:08,119
to collect
1367
01:01:05,520 --> 01:01:11,200
because the model probability is always
1368
01:01:08,119 --> 01:01:13,359
different right and so if the model
1369
01:01:11,200 --> 01:01:15,359
probability is like if the model
1370
01:01:13,359 --> 01:01:20,480
probability was actually 0.7 that'd be
1371
01:01:15,359 --> 01:01:22,000
nice but actually it's 0.793 to 6 8 5
1372
01:01:20,480 --> 01:01:24,599
and you never get another example where
1373
01:01:22,000 --> 01:01:26,319
the probability is exactly the same so
1374
01:01:24,599 --> 01:01:28,280
what we do instead is we divide the
1375
01:01:26,319 --> 01:01:30,240
model probabilities into buckets so we
1376
01:01:28,280 --> 01:01:32,880
say the model probability is between 0
1377
01:01:30,240 --> 01:01:36,599
and 0.1 we say the model probability is
1378
01:01:32,880 --> 01:01:40,319
between 0.1 and 0.2 0.2 and 0.3 so we
1379
01:01:36,599 --> 01:01:44,599
create buckets like this like these and
1380
01:01:40,319 --> 01:01:46,520
then we looked at the model confidence
1381
01:01:44,599 --> 01:01:52,839
the average model confidence within that
1382
01:01:46,520 --> 01:01:55,000
bucket so maybe uh between 0.1 and 0 uh
1383
01:01:52,839 --> 01:01:58,000
between 0 and 0.1 the model confidence
1384
01:01:55,000 --> 01:02:00,920
on average is 0 055 or something like
1385
01:01:58,000 --> 01:02:02,640
that so that would be this T here and
1386
01:02:00,920 --> 01:02:05,079
then the accuracy is how often did it
1387
01:02:02,640 --> 01:02:06,680
actually get a correct and this can be
1388
01:02:05,079 --> 01:02:09,720
plotted in this thing called a
1389
01:02:06,680 --> 01:02:15,039
reliability diagram and the reliability
1390
01:02:09,720 --> 01:02:17,599
diagram basically um the the
1391
01:02:15,039 --> 01:02:20,359
outputs uh
1392
01:02:17,599 --> 01:02:26,359
here so this is
1393
01:02:20,359 --> 01:02:26,359
um the this is the model
1394
01:02:27,520 --> 01:02:34,119
yeah I think the red is the model
1395
01:02:30,760 --> 01:02:36,400
um expected probability and then the
1396
01:02:34,119 --> 01:02:40,559
blue uh the blue is the actual
1397
01:02:36,400 --> 01:02:43,240
probability and then um
1398
01:02:40,559 --> 01:02:45,160
the difference between the expected and
1399
01:02:43,240 --> 01:02:47,160
the actual probability is kind of like
1400
01:02:45,160 --> 01:02:48,359
the penalty there is how how poorly
1401
01:02:47,160 --> 01:02:52,000
calibrated
1402
01:02:48,359 --> 01:02:55,880
the and one really important thing to
1403
01:02:52,000 --> 01:02:58,440
know is that calibration in accuracy are
1404
01:02:55,880 --> 01:03:00,599
not necessarily they don't go hand inand
1405
01:02:58,440 --> 01:03:02,359
uh they do to some extent but they don't
1406
01:03:00,599 --> 01:03:06,440
uh they don't necessarily go hand in
1407
01:03:02,359 --> 01:03:06,440
hand and
1408
01:03:07,200 --> 01:03:14,319
the example on the left is a a bad model
1409
01:03:11,200 --> 01:03:16,279
but a well calibrated so its accuracy is
1410
01:03:14,319 --> 01:03:18,720
uh its error is
1411
01:03:16,279 --> 01:03:20,000
44.9% um but it's well calibrated as you
1412
01:03:18,720 --> 01:03:21,440
can see like when it says it knows the
1413
01:03:20,000 --> 01:03:23,880
answer it knows the answer when it
1414
01:03:21,440 --> 01:03:27,799
doesn't answer does this model on the
1415
01:03:23,880 --> 01:03:30,000
other hand has better erir and um but
1416
01:03:27,799 --> 01:03:31,880
worse calibration so the reason why is
1417
01:03:30,000 --> 01:03:36,680
the model is very very confident all the
1418
01:03:31,880 --> 01:03:39,640
time and usually what happens is um
1419
01:03:36,680 --> 01:03:41,200
models that overfit to the data
1420
01:03:39,640 --> 01:03:43,359
especially when you do early stopping on
1421
01:03:41,200 --> 01:03:44,760
something like accuracy uh when you stop
1422
01:03:43,359 --> 01:03:47,279
the training on something like accuracy
1423
01:03:44,760 --> 01:03:49,960
will become very overconfident and uh
1424
01:03:47,279 --> 01:03:52,599
give confidence estimates um that are in
1425
01:03:49,960 --> 01:03:54,000
cor like this so this is important to
1426
01:03:52,599 --> 01:03:56,079
know and the reason why it's important
1427
01:03:54,000 --> 01:03:58,000
to know is actually because you know
1428
01:03:56,079 --> 01:04:00,960
models are very good at making up things
1429
01:03:58,000 --> 01:04:02,359
that aren't actually correct nowadays um
1430
01:04:00,960 --> 01:04:04,920
and but if you have a really well
1431
01:04:02,359 --> 01:04:07,760
calibrated model you could at least say
1432
01:04:04,920 --> 01:04:09,920
with what confidence you have this
1433
01:04:07,760 --> 01:04:12,760
working so how do you calculate the
1434
01:04:09,920 --> 01:04:14,160
probability of an answer so H yeah sorry
1435
01:04:12,760 --> 01:04:17,599
uh yes
1436
01:04:14,160 --> 01:04:17,599
yes yeah please
1437
01:04:17,799 --> 01:04:26,559
go the probability of percent or
1438
01:04:23,200 --> 01:04:28,039
percent um usually this would be for a
1439
01:04:26,559 --> 01:04:29,599
generated output because you want to
1440
01:04:28,039 --> 01:04:32,559
know the the probability that the
1441
01:04:29,599 --> 01:04:32,559
generated output is
1442
01:04:53,160 --> 01:04:56,160
cor
1443
01:05:01,079 --> 01:05:06,319
great that's what I'm about to talk
1444
01:05:03,000 --> 01:05:07,839
about so perfect perfect question um so
1445
01:05:06,319 --> 01:05:10,160
how do we calculate the answer
1446
01:05:07,839 --> 01:05:13,279
probability or um how do we calculate
1447
01:05:10,160 --> 01:05:15,039
the confidence in an answer um we're
1448
01:05:13,279 --> 01:05:18,319
actually going to go into more detail
1449
01:05:15,039 --> 01:05:20,760
about this um in a a later class but the
1450
01:05:18,319 --> 01:05:23,200
first thing is probability of the answer
1451
01:05:20,760 --> 01:05:25,799
and this is easy when there's a single
1452
01:05:23,200 --> 01:05:29,079
answer um like if there's only one
1453
01:05:25,799 --> 01:05:31,839
correct answer and you want your model
1454
01:05:29,079 --> 01:05:34,160
to be solving math problems and you want
1455
01:05:31,839 --> 01:05:38,319
it to return only the answer and nothing
1456
01:05:34,160 --> 01:05:40,760
else if it returns anything else like it
1457
01:05:38,319 --> 01:05:44,920
won't work then you can just use the
1458
01:05:40,760 --> 01:05:47,119
probability of the answer but what
1459
01:05:44,920 --> 01:05:49,559
if
1460
01:05:47,119 --> 01:05:52,000
um what if there are multiple acceptable
1461
01:05:49,559 --> 01:05:54,680
answers um and maybe a perfect example
1462
01:05:52,000 --> 01:06:02,240
of that is like where is CMU located
1463
01:05:54,680 --> 01:06:04,400
or um uh where where are we right now um
1464
01:06:02,240 --> 01:06:06,960
if the answer is where are we right
1465
01:06:04,400 --> 01:06:08,880
now um could be
1466
01:06:06,960 --> 01:06:12,880
Pittsburgh could be
1467
01:06:08,880 --> 01:06:12,880
CMU could be carnegy
1468
01:06:16,200 --> 01:06:24,440
melon could be other other things like
1469
01:06:18,760 --> 01:06:26,760
this right um and so another way that
1470
01:06:24,440 --> 01:06:28,319
you can calculate the confidence is
1471
01:06:26,760 --> 01:06:31,240
calculating the probability of the
1472
01:06:28,319 --> 01:06:33,680
answer plus uh you know paraphrases of
1473
01:06:31,240 --> 01:06:35,799
the answer or other uh other things like
1474
01:06:33,680 --> 01:06:37,680
this and so then you would just sum the
1475
01:06:35,799 --> 01:06:38,839
probability over all the qu like
1476
01:06:37,680 --> 01:06:41,680
acceptable
1477
01:06:38,839 --> 01:06:45,359
answers
1478
01:06:41,680 --> 01:06:47,680
um another thing that you can do is um
1479
01:06:45,359 --> 01:06:49,279
sample multiple outputs and count the
1480
01:06:47,680 --> 01:06:51,000
number of times you get a particular
1481
01:06:49,279 --> 01:06:54,440
answer this doesn't solve the problem of
1482
01:06:51,000 --> 01:06:58,119
paraphrasing ex paraphrases existing but
1483
01:06:54,440 --> 01:06:59,880
it does solve the problem of uh it does
1484
01:06:58,119 --> 01:07:01,480
solve two problems sometimes there are
1485
01:06:59,880 --> 01:07:05,240
language models where you can't get
1486
01:07:01,480 --> 01:07:06,640
probabilities out of them um this is not
1487
01:07:05,240 --> 01:07:08,680
so much of a problem anymore with the
1488
01:07:06,640 --> 01:07:11,240
GPT models because they're reintroducing
1489
01:07:08,680 --> 01:07:12,440
the ability to get probabilities but um
1490
01:07:11,240 --> 01:07:13,720
there are some models where you can just
1491
01:07:12,440 --> 01:07:16,279
sample from them and you can't get
1492
01:07:13,720 --> 01:07:18,680
probabilities out but also more
1493
01:07:16,279 --> 01:07:21,039
importantly um sometimes when you're
1494
01:07:18,680 --> 01:07:23,000
using things like uh Chain of Thought
1495
01:07:21,039 --> 01:07:26,520
reasoning which I'll talk about in more
1496
01:07:23,000 --> 01:07:29,839
detail but basically it's like um please
1497
01:07:26,520 --> 01:07:31,480
solve this math problem and explain
1498
01:07:29,839 --> 01:07:33,480
explain your solution and then if it
1499
01:07:31,480 --> 01:07:35,119
will do that it will generate you know a
1500
01:07:33,480 --> 01:07:36,279
really long explanation of how it got to
1501
01:07:35,119 --> 01:07:40,119
the solution and then it will give you
1502
01:07:36,279 --> 01:07:41,640
the answer at the very end and so then
1503
01:07:40,119 --> 01:07:44,960
you can't calculate the probability of
1504
01:07:41,640 --> 01:07:47,720
the actual like answer itself because
1505
01:07:44,960 --> 01:07:49,359
there's this long reasoning chain in
1506
01:07:47,720 --> 01:07:51,960
between and you have like all these
1507
01:07:49,359 --> 01:07:53,559
other all that other text there but what
1508
01:07:51,960 --> 01:07:55,480
you can do is you can sample those
1509
01:07:53,559 --> 01:07:56,920
reasoning chains 100 times and then see
1510
01:07:55,480 --> 01:07:59,599
how many times you got a particular
1511
01:07:56,920 --> 01:08:02,960
answer and that's actually a pretty um a
1512
01:07:59,599 --> 01:08:06,079
Prett pretty reasonable way of uh
1513
01:08:02,960 --> 01:08:09,000
getting a have
1514
01:08:06,079 --> 01:08:11,200
yet this is my favorite one I I love how
1515
01:08:09,000 --> 01:08:12,880
we can do this now it's just absolutely
1516
01:08:11,200 --> 01:08:16,480
ridiculous but you could ask the model
1517
01:08:12,880 --> 01:08:20,279
how confident it is and um it sometimes
1518
01:08:16,480 --> 01:08:22,359
gives you a reasonable uh a reasonable
1519
01:08:20,279 --> 01:08:24,600
answer um there's a really nice
1520
01:08:22,359 --> 01:08:26,400
comparison of different methods uh in
1521
01:08:24,600 --> 01:08:29,679
this paper which is also on on the
1522
01:08:26,400 --> 01:08:31,960
website and basically long story short
1523
01:08:29,679 --> 01:08:34,000
the conclusion from this paper is the
1524
01:08:31,960 --> 01:08:35,640
sampling multiple outputs one is the
1525
01:08:34,000 --> 01:08:36,839
best way to do it if you can't directly
1526
01:08:35,640 --> 01:08:39,520
calculate
1527
01:08:36,839 --> 01:08:41,359
probabilities um another thing that I'd
1528
01:08:39,520 --> 01:08:42,600
like people to pay very close attention
1529
01:08:41,359 --> 01:08:45,040
to is in the
1530
01:08:42,600 --> 01:08:46,480
Generation Um in the generation class
1531
01:08:45,040 --> 01:08:49,600
we're going to be talking about minimum
1532
01:08:46,480 --> 01:08:52,600
based risk which is a Criterion for
1533
01:08:49,600 --> 01:08:54,719
deciding how risky an output is and it's
1534
01:08:52,600 --> 01:08:56,199
actually a really good uh confidence
1535
01:08:54,719 --> 01:08:58,000
metric as well but I'm going to leave
1536
01:08:56,199 --> 01:08:59,440
that till when we discuss it more detail
1537
01:08:58,000 --> 01:09:02,759
with
1538
01:08:59,440 --> 01:09:05,359
it um any any questions
1539
01:09:02,759 --> 01:09:08,440
here okay
1540
01:09:05,359 --> 01:09:10,480
cool um so the other Criterion uh this
1541
01:09:08,440 --> 01:09:12,520
is just yet another Criterion that we
1542
01:09:10,480 --> 01:09:15,239
would like language models to be good at
1543
01:09:12,520 --> 01:09:17,600
um its efficiency and so basically the
1544
01:09:15,239 --> 01:09:21,920
model is easy to run on limited Hardware
1545
01:09:17,600 --> 01:09:25,400
by some you know uh metric of easy and
1546
01:09:21,920 --> 01:09:29,319
some metrics that we like to talk about
1547
01:09:25,400 --> 01:09:32,400
our parameter account so often you will
1548
01:09:29,319 --> 01:09:34,239
see oh this is the best model under
1549
01:09:32,400 --> 01:09:35,520
three billion parameters or this is the
1550
01:09:34,239 --> 01:09:37,960
best model under seven billion
1551
01:09:35,520 --> 01:09:39,600
parameters or um we trained a model with
1552
01:09:37,960 --> 01:09:42,159
one trillion parameters or something
1553
01:09:39,600 --> 01:09:44,719
like that you know
1554
01:09:42,159 --> 01:09:46,839
uh the thing is parameter count doesn't
1555
01:09:44,719 --> 01:09:49,640
really mean that much um from the point
1556
01:09:46,839 --> 01:09:52,839
of view of like ease of using the model
1557
01:09:49,640 --> 01:09:54,400
um unless you also think about other uh
1558
01:09:52,839 --> 01:09:56,480
you know deser
1559
01:09:54,400 --> 01:09:58,840
like just to give one example this is a
1560
01:09:56,480 --> 01:10:00,880
parameter count um let's say you have a
1561
01:09:58,840 --> 01:10:02,960
parameter count of 7 billion is that 7
1562
01:10:00,880 --> 01:10:05,719
billion parameters at 32-bit Precision
1563
01:10:02,960 --> 01:10:07,800
or is that 7 billion parameters at 4bit
1564
01:10:05,719 --> 01:10:09,400
Precision um will make a huge difference
1565
01:10:07,800 --> 01:10:12,960
in your memory footprint your speed
1566
01:10:09,400 --> 01:10:14,920
other things like that um so some of the
1567
01:10:12,960 --> 01:10:18,040
things that are more direct with respect
1568
01:10:14,920 --> 01:10:19,800
to efficiency are memory usage um and
1569
01:10:18,040 --> 01:10:22,440
there's two varieties of memory usage
1570
01:10:19,800 --> 01:10:24,280
one is model uh model only memory usage
1571
01:10:22,440 --> 01:10:27,120
so when you load loaded the model into
1572
01:10:24,280 --> 01:10:29,120
memory uh how much space does it take
1573
01:10:27,120 --> 01:10:31,159
and also Peak memory consumption when
1574
01:10:29,120 --> 01:10:33,159
you run have run the model over a
1575
01:10:31,159 --> 01:10:35,920
sequence of a certain length how much is
1576
01:10:33,159 --> 01:10:40,040
it going to P so that's another
1577
01:10:35,920 --> 01:10:43,000
thing another thing is latency um and
1578
01:10:40,040 --> 01:10:46,440
with respect to latency this can be
1579
01:10:43,000 --> 01:10:49,440
either how long does it take to start
1580
01:10:46,440 --> 01:10:52,080
outputting the first token um and how
1581
01:10:49,440 --> 01:10:54,840
long does it take to uh finish
1582
01:10:52,080 --> 01:10:59,480
outputting uh a generation of a certain
1583
01:10:54,840 --> 01:11:01,199
length and the first will have more to
1584
01:10:59,480 --> 01:11:04,960
do with how long does it take to encode
1585
01:11:01,199 --> 01:11:06,480
a sequence um which is usually faster
1586
01:11:04,960 --> 01:11:09,080
than how long does it take to generate a
1587
01:11:06,480 --> 01:11:11,360
sequence so this will have to do with
1588
01:11:09,080 --> 01:11:13,000
like encoding time this will require
1589
01:11:11,360 --> 01:11:15,880
encoding time of course but it will also
1590
01:11:13,000 --> 01:11:15,880
require generation
1591
01:11:16,280 --> 01:11:21,840
time also throughput so you know how
1592
01:11:19,239 --> 01:11:23,679
much um how many sentences can you
1593
01:11:21,840 --> 01:11:25,400
process in a certain amount of time so
1594
01:11:23,679 --> 01:11:26,480
of these are kind of desad that you you
1595
01:11:25,400 --> 01:11:29,000
would
1596
01:11:26,480 --> 01:11:30,280
say um we're going to be talking about
1597
01:11:29,000 --> 01:11:31,920
this more in the distillation and
1598
01:11:30,280 --> 01:11:33,199
compression and generation algorithms
1599
01:11:31,920 --> 01:11:35,640
classes so I won't go into a whole lot
1600
01:11:33,199 --> 01:11:36,840
of detail about this but um it's just
1601
01:11:35,640 --> 01:11:39,960
another thing that we want to be
1602
01:11:36,840 --> 01:11:43,560
thinking about in addition to
1603
01:11:39,960 --> 01:11:45,360
complexity um but since I'm I'm on the
1604
01:11:43,560 --> 01:11:47,800
topic of efficiency I would like to talk
1605
01:11:45,360 --> 01:11:49,480
just a little bit about it um in terms
1606
01:11:47,800 --> 01:11:51,000
of especially things that will be useful
1607
01:11:49,480 --> 01:11:53,600
for implementing your first
1608
01:11:51,000 --> 01:11:55,840
assignment and uh one thing that every
1609
01:11:53,600 --> 01:11:58,639
body should know about um if you've done
1610
01:11:55,840 --> 01:11:59,920
any like deep learning with pytorch or
1611
01:11:58,639 --> 01:12:02,639
something like this you already know
1612
01:11:59,920 --> 01:12:05,880
about this probably but uh I think it's
1613
01:12:02,639 --> 01:12:08,760
worth mentioning but basically mini
1614
01:12:05,880 --> 01:12:12,120
batching or batching uh is uh very
1615
01:12:08,760 --> 01:12:15,320
useful and the basic idea behind it is
1616
01:12:12,120 --> 01:12:17,560
that on Modern Hardware if you do many
1617
01:12:15,320 --> 01:12:20,520
of the same operations at once it's much
1618
01:12:17,560 --> 01:12:24,320
faster than doing um
1619
01:12:20,520 --> 01:12:25,480
like uh operations executively and
1620
01:12:24,320 --> 01:12:27,280
that's especially the case if you're
1621
01:12:25,480 --> 01:12:30,520
programming in an extremely slow
1622
01:12:27,280 --> 01:12:33,239
programming language like python um I
1623
01:12:30,520 --> 01:12:37,239
love python but it's slow I mean like
1624
01:12:33,239 --> 01:12:38,719
there's no argument about that um and so
1625
01:12:37,239 --> 01:12:40,520
what mini batching does is it combines
1626
01:12:38,719 --> 01:12:43,600
together smaller operations into one big
1627
01:12:40,520 --> 01:12:47,480
one and the basic idea uh for example if
1628
01:12:43,600 --> 01:12:51,679
we want to calculate our um our linear
1629
01:12:47,480 --> 01:12:56,560
layer with a t uh nonlinearity after it
1630
01:12:51,679 --> 01:12:59,760
we will take several inputs X1 X2 X3
1631
01:12:56,560 --> 01:13:02,040
concatenate them together and do a
1632
01:12:59,760 --> 01:13:04,600
Matrix Matrix multiply instead of doing
1633
01:13:02,040 --> 01:13:07,960
three Vector Matrix
1634
01:13:04,600 --> 01:13:09,239
multiplies and so what we do is we take
1635
01:13:07,960 --> 01:13:11,280
a whole bunch of examples we take like
1636
01:13:09,239 --> 01:13:13,840
64 examples or something like that and
1637
01:13:11,280 --> 01:13:18,000
we combine them together and calculate
1638
01:13:13,840 --> 01:13:21,280
out thingsit one thing to know is that
1639
01:13:18,000 --> 01:13:22,560
if you're working with sentences there's
1640
01:13:21,280 --> 01:13:24,719
different ways you can calculate the
1641
01:13:22,560 --> 01:13:27,360
size of your mini
1642
01:13:24,719 --> 01:13:28,880
normally nowadays the thing that people
1643
01:13:27,360 --> 01:13:30,400
do and the thing that I recommend is to
1644
01:13:28,880 --> 01:13:31,679
calculate the size of your mini batches
1645
01:13:30,400 --> 01:13:33,639
based on the number of tokens in the
1646
01:13:31,679 --> 01:13:35,840
mini batch it used to be that you would
1647
01:13:33,639 --> 01:13:39,719
do it based on the number of sequences
1648
01:13:35,840 --> 01:13:43,800
but the the problem is um one like 50
1649
01:13:39,719 --> 01:13:47,120
sequences of length like 100 is much
1650
01:13:43,800 --> 01:13:49,480
more memory intensive than uh 50
1651
01:13:47,120 --> 01:13:51,960
sequences of Link five and so you get
1652
01:13:49,480 --> 01:13:53,920
these vastly varying these mini batches
1653
01:13:51,960 --> 01:13:57,000
of vastly varying size and that's both
1654
01:13:53,920 --> 01:13:59,800
bad for you know memory overflows and
1655
01:13:57,000 --> 01:14:01,639
bad for um and bad for learning
1656
01:13:59,800 --> 01:14:04,280
stability so I I definitely recommend
1657
01:14:01,639 --> 01:14:06,880
doing it based on the number of
1658
01:14:04,280 --> 01:14:09,080
comps uh another thing is gpus versus
1659
01:14:06,880 --> 01:14:12,400
CPUs so
1660
01:14:09,080 --> 01:14:14,600
um uh CPUs one way you can think of it
1661
01:14:12,400 --> 01:14:17,320
is a CPUs kind of like a motorcycle it's
1662
01:14:14,600 --> 01:14:19,600
very fast at picking up and doing a
1663
01:14:17,320 --> 01:14:23,960
bunch of uh things very quickly
1664
01:14:19,600 --> 01:14:26,600
accelerating uh into starting new uh new
1665
01:14:23,960 --> 01:14:28,760
tasks a GPU is more like an airplane
1666
01:14:26,600 --> 01:14:30,719
which uh you wait forever in line in
1667
01:14:28,760 --> 01:14:33,360
security and
1668
01:14:30,719 --> 01:14:34,800
then and then uh it takes a long time to
1669
01:14:33,360 --> 01:14:40,400
get off the ground and start working but
1670
01:14:34,800 --> 01:14:43,679
once it does it's extremely fast um and
1671
01:14:40,400 --> 01:14:45,360
so if we do a simple example of how long
1672
01:14:43,679 --> 01:14:47,600
does it take to do a Matrix Matrix
1673
01:14:45,360 --> 01:14:49,040
multiply I calculated this a really long
1674
01:14:47,600 --> 01:14:51,280
time ago it's probably horribly out of
1675
01:14:49,040 --> 01:14:55,120
date now but the same general principle
1676
01:14:51,280 --> 01:14:56,560
stands which is if we have have um the
1677
01:14:55,120 --> 01:14:58,480
number of seconds that it takes to do a
1678
01:14:56,560 --> 01:15:02,080
Matrix Matrix multiply doing one of size
1679
01:14:58,480 --> 01:15:03,920
16 is actually faster on CPU because uh
1680
01:15:02,080 --> 01:15:07,760
the overhead it takes to get started is
1681
01:15:03,920 --> 01:15:10,880
very low but if you um once you start
1682
01:15:07,760 --> 01:15:13,360
getting up to size like 128 by 128
1683
01:15:10,880 --> 01:15:15,800
Matrix multiplies then doing it on GPU
1684
01:15:13,360 --> 01:15:17,320
is faster and then um it's you know a
1685
01:15:15,800 --> 01:15:19,679
100 times faster once you start getting
1686
01:15:17,320 --> 01:15:21,600
up to very large matrices so um if
1687
01:15:19,679 --> 01:15:24,000
you're dealing with very large networks
1688
01:15:21,600 --> 01:15:26,800
handling a GPU is good
1689
01:15:24,000 --> 01:15:30,159
um and this is the the speed up
1690
01:15:26,800 --> 01:15:31,440
percentage um one thing I should mention
1691
01:15:30,159 --> 01:15:34,239
is
1692
01:15:31,440 --> 01:15:36,440
um compute with respect to like doing
1693
01:15:34,239 --> 01:15:39,800
the assignments for this class if you
1694
01:15:36,440 --> 01:15:43,199
have a relatively recent Mac you're kind
1695
01:15:39,800 --> 01:15:44,760
of in luck because actually the gpus on
1696
01:15:43,199 --> 01:15:47,239
the Mac are pretty fast and they're well
1697
01:15:44,760 --> 01:15:48,960
integrated with um they're well
1698
01:15:47,239 --> 01:15:52,080
integrated with pipor and other things
1699
01:15:48,960 --> 01:15:53,440
like that so decently sized models maybe
1700
01:15:52,080 --> 01:15:54,840
up to the size that you would need to
1701
01:15:53,440 --> 01:15:57,840
run for assignment one or even
1702
01:15:54,840 --> 01:16:00,880
assignment two might uh just run on your
1703
01:15:57,840 --> 01:16:03,639
uh laptop computer um if you don't have
1704
01:16:00,880 --> 01:16:05,280
a GPU uh that you have immediately
1705
01:16:03,639 --> 01:16:06,760
accessible to you I we're going to
1706
01:16:05,280 --> 01:16:08,400
recommend that you use collab where you
1707
01:16:06,760 --> 01:16:10,120
can get a GPU uh for the first
1708
01:16:08,400 --> 01:16:12,440
assignments and then we'll have plug
1709
01:16:10,120 --> 01:16:15,159
reddits that you can use otherwise but
1710
01:16:12,440 --> 01:16:16,800
um GPU is usually like something that
1711
01:16:15,159 --> 01:16:18,440
you can get on the cloud or one that you
1712
01:16:16,800 --> 01:16:21,080
have on your Mac or one that you have on
1713
01:16:18,440 --> 01:16:24,600
your gaming computer or something like
1714
01:16:21,080 --> 01:16:26,040
that um there's a few speed tricks that
1715
01:16:24,600 --> 01:16:30,000
you should know for efficient GPU
1716
01:16:26,040 --> 01:16:32,480
operations so um one mistake that people
1717
01:16:30,000 --> 01:16:35,880
make when creating models is they repeat
1718
01:16:32,480 --> 01:16:38,080
operations over and over again and um
1719
01:16:35,880 --> 01:16:40,600
you don't want to be doing this so like
1720
01:16:38,080 --> 01:16:43,239
for example um this is multiplying a
1721
01:16:40,600 --> 01:16:45,320
matrix by a constant multiple times and
1722
01:16:43,239 --> 01:16:46,880
if you're just using out of thee box pie
1723
01:16:45,320 --> 01:16:49,280
torch this would be really bad because
1724
01:16:46,880 --> 01:16:50,400
you'd be repeating the operation uh when
1725
01:16:49,280 --> 01:16:52,679
it's not
1726
01:16:50,400 --> 01:16:54,480
necessary um you can also reduce the
1727
01:16:52,679 --> 01:16:57,360
number of operations that you need to
1728
01:16:54,480 --> 01:17:00,320
use so uh use Matrix Matrix multiplies
1729
01:16:57,360 --> 01:17:03,080
instead of Matrix Vector
1730
01:17:00,320 --> 01:17:07,920
multiplies and another thing is uh
1731
01:17:03,080 --> 01:17:10,719
reducing CPU GPU data movement and um so
1732
01:17:07,920 --> 01:17:12,360
when you do try to move memory um when
1733
01:17:10,719 --> 01:17:17,080
you do try to move memory try to do it
1734
01:17:12,360 --> 01:17:20,040
as early as possible and as uh and as
1735
01:17:17,080 --> 01:17:22,199
few times as possible and the reason why
1736
01:17:20,040 --> 01:17:24,199
you want to move things early or start
1737
01:17:22,199 --> 01:17:25,920
operations early is many GPU operations
1738
01:17:24,199 --> 01:17:27,159
are asynchronous so you can start the
1739
01:17:25,920 --> 01:17:28,800
operation and it will run in the
1740
01:17:27,159 --> 01:17:33,120
background while other things are
1741
01:17:28,800 --> 01:17:36,080
processing so um it's a good idea to try
1742
01:17:33,120 --> 01:17:39,840
to um to optimize and you can also use
1743
01:17:36,080 --> 01:17:42,360
your python profiler or um envidia GPU
1744
01:17:39,840 --> 01:17:43,679
profilers to try to optimize these
1745
01:17:42,360 --> 01:17:46,520
things as
1746
01:17:43,679 --> 01:17:49,840
well cool that's all I have uh we're
1747
01:17:46,520 --> 01:17:49,840
right at time