diff --git "a/CMU Advanced NLP 2024 (14) Ensembling and Mixture of Experts/transcript.srt" "b/CMU Advanced NLP 2024 (14) Ensembling and Mixture of Experts/transcript.srt"
new file mode 100644--- /dev/null
+++ "b/CMU Advanced NLP 2024 (14) Ensembling and Mixture of Experts/transcript.srt"
@@ -0,0 +1,6599 @@
+1
+00:00:00,760 --> 00:00:07,240
+he everyone so I'd like to get
+
+2
+00:00:03,279 --> 00:00:09,320
+started the first thing is that um I
+
+3
+00:00:07,240 --> 00:00:11,160
+heard from the adws people that they
+
+4
+00:00:09,320 --> 00:00:14,440
+started the
+
+5
+00:00:11,160 --> 00:00:17,840
+process of
+
+6
+00:00:14,440 --> 00:00:19,400
+getting things issued on the 26th which
+
+7
+00:00:17,840 --> 00:00:21,480
+is three days ago so you should be
+
+8
+00:00:19,400 --> 00:00:23,560
+getting it soon uh for reference I
+
+9
+00:00:21,480 --> 00:00:25,599
+submitted the form about seven days
+
+10
+00:00:23,560 --> 00:00:28,359
+before that so they're moving very
+
+11
+00:00:25,599 --> 00:00:29,599
+slowly but I think you should have AWS
+
+12
+00:00:28,359 --> 00:00:31,920
+credits by the end of the week if you
+
+13
+00:00:29,599 --> 00:00:35,120
+need them to run uh GPU machines or
+
+14
+00:00:31,920 --> 00:00:37,960
+stuff like that the moment you get AWS
+
+15
+00:00:35,120 --> 00:00:39,960
+credits or maybe even before you get AWS
+
+16
+00:00:37,960 --> 00:00:43,320
+credits I might suggest that you try to
+
+17
+00:00:39,960 --> 00:00:46,760
+start uh a GPU machine like a P2 machine
+
+18
+00:00:43,320 --> 00:00:49,160
+or something like that because um
+
+19
+00:00:46,760 --> 00:00:51,760
+sometimes you need to file for a limit
+
+20
+00:00:49,160 --> 00:00:53,640
+increase uh to get a P2 machine and that
+
+21
+00:00:51,760 --> 00:00:55,879
+also takes a little bit of time so I I
+
+22
+00:00:53,640 --> 00:00:59,160
+would suggest that you uh you take a
+
+23
+00:00:55,879 --> 00:01:01,160
+look at doing that um so you go to like
+
+24
+00:00:59,160 --> 00:01:02,800
+if you're using AWS if you're not using
+
+25
+00:01:01,160 --> 00:01:05,119
+AWS it doesn't matter but if you're
+
+26
+00:01:02,800 --> 00:01:08,119
+using AWS you can go to launch instance
+
+27
+00:01:05,119 --> 00:01:11,520
+and try to launch a p2x large machine um
+
+28
+00:01:08,119 --> 00:01:13,159
+or something like that so uh but yeah
+
+29
+00:01:11,520 --> 00:01:14,920
+anyway hopefully that will be done soon
+
+30
+00:01:13,159 --> 00:01:16,600
+I'm sorry about the delay on this they
+
+31
+00:01:14,920 --> 00:01:21,400
+said it would take seven days and it's
+
+32
+00:01:16,600 --> 00:01:24,280
+taken almost twice at now so um my
+
+33
+00:01:21,400 --> 00:01:26,439
+apologies any other uh things before we
+
+34
+00:01:24,280 --> 00:01:26,439
+get
+
+35
+00:01:28,759 --> 00:01:34,520
+started um okay I I don't see any so
+
+36
+00:01:31,920 --> 00:01:37,280
+I'll go ahead with this um I have
+
+37
+00:01:34,520 --> 00:01:39,240
+slightly fewer slides today so I might
+
+38
+00:01:37,280 --> 00:01:40,960
+go a little bit off the slides and talk
+
+39
+00:01:39,240 --> 00:01:44,759
+about papers and stuff or we might
+
+40
+00:01:40,960 --> 00:01:46,920
+finish early uh either way so um but
+
+41
+00:01:44,759 --> 00:01:48,439
+what I would like to talk about is um
+
+42
+00:01:46,920 --> 00:01:53,320
+combining multiple
+
+43
+00:01:48,439 --> 00:01:55,479
+models and this is uh really important
+
+44
+00:01:53,320 --> 00:01:57,520
+and useful if you want to get like an
+
+45
+00:01:55,479 --> 00:02:00,719
+extra few points of
+
+46
+00:01:57,520 --> 00:02:03,159
+accuracy uh for anything basically
+
+47
+00:02:00,719 --> 00:02:04,039
+because it's a pretty reliable way to
+
+48
+00:02:03,159 --> 00:02:06,960
+get
+
+49
+00:02:04,039 --> 00:02:08,879
+improvements um and there's a a bunch of
+
+50
+00:02:06,960 --> 00:02:11,239
+different kind of related but different
+
+51
+00:02:08,879 --> 00:02:13,680
+topics that I'm going to talk about
+
+52
+00:02:11,239 --> 00:02:15,519
+today but anyway the the basic
+
+53
+00:02:13,680 --> 00:02:19,239
+background is that we have many models
+
+54
+00:02:15,519 --> 00:02:22,920
+uh that exist and the reason why we have
+
+55
+00:02:19,239 --> 00:02:25,840
+many models that exist is multiple fold
+
+56
+00:02:22,920 --> 00:02:28,160
+number one we could have different model
+
+57
+00:02:25,840 --> 00:02:30,080
+architectures um and we could also have
+
+58
+00:02:28,160 --> 00:02:34,440
+different initializations of those model
+
+59
+00:02:30,080 --> 00:02:37,879
+architectures so um normally you know if
+
+60
+00:02:34,440 --> 00:02:40,319
+we do initialization we will initial
+
+61
+00:02:37,879 --> 00:02:42,360
+initialize our model architecture like
+
+62
+00:02:40,319 --> 00:02:44,680
+let's say we initialize a llama
+
+63
+00:02:42,360 --> 00:02:45,920
+architecture uh we start out with random
+
+64
+00:02:44,680 --> 00:02:49,319
+7B
+
+65
+00:02:45,920 --> 00:02:52,879
+parameters and then we train and we get
+
+66
+00:02:49,319 --> 00:02:53,840
+llama 7B for uh our pre-training or
+
+67
+00:02:52,879 --> 00:02:57,280
+llama
+
+68
+00:02:53,840 --> 00:02:58,599
+27b um we might initialize another model
+
+69
+00:02:57,280 --> 00:03:00,599
+this could be you know the same
+
+70
+00:02:58,599 --> 00:03:02,360
+architecture different architecture Ure
+
+71
+00:03:00,599 --> 00:03:04,840
+train it on the same data or different
+
+72
+00:03:02,360 --> 00:03:07,000
+data and get something like mistol
+
+73
+00:03:04,840 --> 00:03:08,599
+mistol 7B in this case actually maybe
+
+74
+00:03:07,000 --> 00:03:10,080
+these are I should have indicated that
+
+75
+00:03:08,599 --> 00:03:11,680
+these are different architectures but
+
+76
+00:03:10,080 --> 00:03:13,879
+you know we get a different pre-rain
+
+77
+00:03:11,680 --> 00:03:15,599
+model and of course uh we could also
+
+78
+00:03:13,879 --> 00:03:18,640
+make it bigger or smaller or whatever
+
+79
+00:03:15,599 --> 00:03:21,720
+else and then we get llama 270b over
+
+80
+00:03:18,640 --> 00:03:23,519
+here and then after we do that there's a
+
+81
+00:03:21,720 --> 00:03:25,319
+lot of fine tuning that goes on
+
+82
+00:03:23,519 --> 00:03:29,360
+according to different strategies so we
+
+83
+00:03:25,319 --> 00:03:32,640
+have um you know llama 27b instruct uh
+
+84
+00:03:29,360 --> 00:03:37,760
+vun 7B uh version
+
+85
+00:03:32,640 --> 00:03:41,000
+1.5 um mistol 7B instruct uh news uh
+
+86
+00:03:37,760 --> 00:03:45,239
+Hermes 2 mistal 7B or llama 270b
+
+87
+00:03:41,000 --> 00:03:47,239
+instruct so we have um a variety of
+
+88
+00:03:45,239 --> 00:03:49,400
+architectures a variety of random
+
+89
+00:03:47,239 --> 00:03:51,480
+initializations of those architectures a
+
+90
+00:03:49,400 --> 00:03:54,799
+variety of pre-train models due to
+
+91
+00:03:51,480 --> 00:03:57,439
+pre-training data or base models and
+
+92
+00:03:54,799 --> 00:03:58,920
+then a variety of fine dun models um and
+
+93
+00:03:57,439 --> 00:04:01,120
+so we have this kind of like branching
+
+94
+00:03:58,920 --> 00:04:02,959
+tree basically
+
+95
+00:04:01,120 --> 00:04:04,319
+um the reason why this is important is
+
+96
+00:04:02,959 --> 00:04:06,680
+because when we're combining multiple
+
+97
+00:04:04,319 --> 00:04:08,400
+models together some of the methods are
+
+98
+00:04:06,680 --> 00:04:09,959
+applicable to completely different
+
+99
+00:04:08,400 --> 00:04:12,439
+models some of the methods are only
+
+100
+00:04:09,959 --> 00:04:15,000
+applicable to models that share the same
+
+101
+00:04:12,439 --> 00:04:16,720
+architecture and some of them are only
+
+102
+00:04:15,000 --> 00:04:19,199
+applicable to models that share the same
+
+103
+00:04:16,720 --> 00:04:20,959
+initialization and training trajectory
+
+104
+00:04:19,199 --> 00:04:23,680
+and so I'll try to distinguish between
+
+105
+00:04:20,959 --> 00:04:23,680
+those as we go
+
+106
+00:04:24,040 --> 00:04:27,919
+forward
+
+107
+00:04:25,560 --> 00:04:29,960
+cool so the first thing I I'll talk
+
+108
+00:04:27,919 --> 00:04:32,600
+about is model ensembling and and
+
+109
+00:04:29,960 --> 00:04:34,320
+ensembling is kind of the a very general
+
+110
+00:04:32,600 --> 00:04:37,600
+technique that you can use in a lot of
+
+111
+00:04:34,320 --> 00:04:39,360
+different uh ways but it has its
+
+112
+00:04:37,600 --> 00:04:43,039
+disadvantages as
+
+113
+00:04:39,360 --> 00:04:47,199
+well so basically embling is combining
+
+114
+00:04:43,039 --> 00:04:50,320
+the predictions from multiple models
+
+115
+00:04:47,199 --> 00:04:52,400
+and the easiest way to do this ignore
+
+116
+00:04:50,320 --> 00:04:53,800
+the lstm here this is just any sequence
+
+117
+00:04:52,400 --> 00:04:56,320
+modeling thing it's because the slides
+
+118
+00:04:53,800 --> 00:05:00,120
+are old but like let's say this is a a
+
+119
+00:04:56,320 --> 00:05:03,360
+Transformer it is calculating the
+
+120
+00:05:00,120 --> 00:05:05,600
+current decoder State and you make a
+
+121
+00:05:03,360 --> 00:05:07,600
+prediction um this is calculating a
+
+122
+00:05:05,600 --> 00:05:09,199
+current decoder State and make uh
+
+123
+00:05:07,600 --> 00:05:11,560
+current decoders sayate in making a
+
+124
+00:05:09,199 --> 00:05:13,039
+prediction and based on some combination
+
+125
+00:05:11,560 --> 00:05:17,120
+of the two predictions you decide what
+
+126
+00:05:13,039 --> 00:05:17,120
+you actually want to Output at the next
+
+127
+00:05:17,680 --> 00:05:23,840
+step so why would we want to do this um
+
+128
+00:05:22,080 --> 00:05:25,880
+does anyone have any ideas why we want
+
+129
+00:05:23,840 --> 00:05:28,639
+to use two models instead of using one
+
+130
+00:05:25,880 --> 00:05:31,639
+model or just using the best
+
+131
+00:05:28,639 --> 00:05:31,639
+model
+
+132
+00:05:32,319 --> 00:05:36,440
+or maybe in what situations we would
+
+133
+00:05:34,520 --> 00:05:39,440
+want to do
+
+134
+00:05:36,440 --> 00:05:39,440
+this
+
+135
+00:05:45,400 --> 00:05:50,319
+yeah and what what's the advantage of
+
+136
+00:05:47,960 --> 00:05:50,319
+doing
+
+137
+00:05:51,600 --> 00:05:57,000
+that yeah it reduces a bias kind kind of
+
+138
+00:05:54,800 --> 00:05:57,000
+yeah
+
+139
+00:05:58,639 --> 00:06:01,639
+sure
+
+140
+00:06:28,560 --> 00:06:31,560
+m
+
+141
+00:06:35,400 --> 00:06:40,360
+yeah so um I I'll repeat all of these I
+
+142
+00:06:38,599 --> 00:06:43,960
+think all of these are correct so number
+
+143
+00:06:40,360 --> 00:06:47,479
+one um it reduces the bias uh caused by
+
+144
+00:06:43,960 --> 00:06:49,199
+a single model uh number two it was it's
+
+145
+00:06:47,479 --> 00:06:52,199
+kind of like a beian perspective which
+
+146
+00:06:49,199 --> 00:06:54,000
+I'll talk about in a second and then
+
+147
+00:06:52,199 --> 00:06:56,039
+number three we have different models
+
+148
+00:06:54,000 --> 00:06:58,520
+and models are better at some things and
+
+149
+00:06:56,039 --> 00:07:00,400
+worse at other things
+
+150
+00:06:58,520 --> 00:07:02,720
+um
+
+151
+00:07:00,400 --> 00:07:05,960
+so talking about the better at some
+
+152
+00:07:02,720 --> 00:07:08,319
+things and worse at other things um the
+
+153
+00:07:05,960 --> 00:07:10,960
+basic idea behind embling is that the
+
+154
+00:07:08,319 --> 00:07:14,240
+errors that model m models make tend to
+
+155
+00:07:10,960 --> 00:07:15,840
+not be consistent it not tend to not be
+
+156
+00:07:14,240 --> 00:07:21,520
+as consistent as when the model is
+
+157
+00:07:15,840 --> 00:07:24,800
+getting it correct so we might have um
+
+158
+00:07:21,520 --> 00:07:26,160
+we might have one model that says uh
+
+159
+00:07:24,800 --> 00:07:28,199
+like let's say we just have really
+
+160
+00:07:26,160 --> 00:07:30,680
+really bad models this is kind of a
+
+161
+00:07:28,199 --> 00:07:31,720
+really um
+
+162
+00:07:30,680 --> 00:07:35,960
+obvious
+
+163
+00:07:31,720 --> 00:07:38,440
+example but we have like the dog the dog
+
+164
+00:07:35,960 --> 00:07:42,639
+barks and then
+
+165
+00:07:38,440 --> 00:07:46,039
+runs and then uh Dives or something like
+
+166
+00:07:42,639 --> 00:07:49,000
+that and we have uh one one model that
+
+167
+00:07:46,039 --> 00:07:50,560
+just had tons of stuff about diving in
+
+168
+00:07:49,000 --> 00:07:52,120
+its training data another model that had
+
+169
+00:07:50,560 --> 00:07:54,240
+tons of stuff about running in its
+
+170
+00:07:52,120 --> 00:07:56,560
+training data or or marathons or
+
+171
+00:07:54,240 --> 00:08:00,039
+something staining data so we'll get
+
+172
+00:07:56,560 --> 00:08:01,800
+model one and model one we'll to give
+
+173
+00:08:00,039 --> 00:08:06,240
+like a probability of like
+
+174
+00:08:01,800 --> 00:08:08,280
+0.3 maybe 0.4 and
+
+175
+00:08:06,240 --> 00:08:10,360
+0.05 and then we'll have another one
+
+176
+00:08:08,280 --> 00:08:13,039
+over here that's like
+
+177
+00:08:10,360 --> 00:08:17,319
+0.32
+
+178
+00:08:13,039 --> 00:08:19,759
+0.41 and 0 sorry
+
+179
+00:08:17,319 --> 00:08:23,039
+0.05 and
+
+180
+00:08:19,759 --> 00:08:25,759
+0.41 or something like this and so when
+
+181
+00:08:23,039 --> 00:08:27,639
+you average the two together you tend to
+
+182
+00:08:25,759 --> 00:08:29,240
+get the right answer more often because
+
+183
+00:08:27,639 --> 00:08:31,720
+kind of the mistakes that they make tend
+
+184
+00:08:29,240 --> 00:08:33,479
+to less correlated than the probability
+
+185
+00:08:31,720 --> 00:08:35,880
+of getting and of course it's not
+
+186
+00:08:33,479 --> 00:08:38,200
+perfect because unbled models are not
+
+187
+00:08:35,880 --> 00:08:39,880
+perfect but this is a a general tendency
+
+188
+00:08:38,200 --> 00:08:42,240
+that we see a lot in
+
+189
+00:08:39,880 --> 00:08:45,959
+models
+
+190
+00:08:42,240 --> 00:08:47,720
+um and um it's because of this it kind
+
+191
+00:08:45,959 --> 00:08:52,320
+of Smooths over the idiosyncrasies of
+
+192
+00:08:47,720 --> 00:08:54,800
+the models you can even um gist Ensemble
+
+193
+00:08:52,320 --> 00:08:57,519
+models from different checkpoints and
+
+194
+00:08:54,800 --> 00:08:58,959
+that still gives you improvements and so
+
+195
+00:08:57,519 --> 00:09:00,560
+when you Ensemble models from different
+
+196
+00:08:58,959 --> 00:09:02,600
+checkpoints it's basically just what
+
+197
+00:09:00,560 --> 00:09:05,920
+data did they see most recently and that
+
+198
+00:09:02,600 --> 00:09:07,839
+also Smooths over you know uh the fact
+
+199
+00:09:05,920 --> 00:09:10,600
+that like this model happened to see
+
+200
+00:09:07,839 --> 00:09:13,000
+some data more recently and so it's less
+
+201
+00:09:10,600 --> 00:09:16,120
+uh you know it's biased towards doing
+
+202
+00:09:13,000 --> 00:09:18,440
+that so uh this is a a pretty effective
+
+203
+00:09:16,120 --> 00:09:20,079
+method this is one of the few methods
+
+204
+00:09:18,440 --> 00:09:21,959
+that I know is going to improve my
+
+205
+00:09:20,079 --> 00:09:25,120
+accuracy almost every time like there's
+
+206
+00:09:21,959 --> 00:09:27,880
+a bunch of methods that you can apply um
+
+207
+00:09:25,120 --> 00:09:29,680
+and I ensembling it's very rare for me
+
+208
+00:09:27,880 --> 00:09:31,959
+to Ensemble two models together not get
+
+209
+00:09:29,680 --> 00:09:34,839
+a boost in accuracy in some way so it's
+
+210
+00:09:31,959 --> 00:09:34,839
+a good thing to
+
+211
+00:09:35,600 --> 00:09:41,040
+that there's two main ways to combine
+
+212
+00:09:38,680 --> 00:09:42,560
+models together and both of them are
+
+213
+00:09:41,040 --> 00:09:45,800
+useful in different
+
+214
+00:09:42,560 --> 00:09:48,079
+situations the first one is linear
+
+215
+00:09:45,800 --> 00:09:49,600
+interpolation and when you do linear
+
+216
+00:09:48,079 --> 00:09:51,240
+interpolation basically what you're
+
+217
+00:09:49,600 --> 00:09:53,720
+doing is you're taking the weighted
+
+218
+00:09:51,240 --> 00:09:56,839
+average of model
+
+219
+00:09:53,720 --> 00:10:00,360
+probabilities and the way that looks
+
+220
+00:09:56,839 --> 00:10:04,040
+mathematically is like this um this is a
+
+221
+00:10:00,360 --> 00:10:05,680
+probability according to the model M so
+
+222
+00:10:04,040 --> 00:10:08,000
+this is just you know the probability of
+
+223
+00:10:05,680 --> 00:10:11,720
+the next token according to model M this
+
+224
+00:10:08,000 --> 00:10:13,200
+is the probability of selecting model M
+
+225
+00:10:11,720 --> 00:10:18,040
+so you talked a little bit about the
+
+226
+00:10:13,200 --> 00:10:19,920
+basian approach uh to this and this is
+
+227
+00:10:18,040 --> 00:10:23,519
+basically saying what is the probability
+
+228
+00:10:19,920 --> 00:10:26,519
+that the parameters of model M
+
+229
+00:10:23,519 --> 00:10:30,320
+are the ones that we want to be choosing
+
+230
+00:10:26,519 --> 00:10:32,680
+in this at this particular time step and
+
+231
+00:10:30,320 --> 00:10:34,640
+then we will we will calculate this and
+
+232
+00:10:32,680 --> 00:10:38,120
+so then you take the sum over this and
+
+233
+00:10:34,640 --> 00:10:38,120
+this gives you the next
+
+234
+00:10:39,560 --> 00:10:44,800
+probability for the second term you can
+
+235
+00:10:42,639 --> 00:10:47,120
+do this in two ways the most common way
+
+236
+00:10:44,800 --> 00:10:51,800
+to do this is just to have this be a
+
+237
+00:10:47,120 --> 00:10:55,279
+constant so you you basically
+
+238
+00:10:51,800 --> 00:10:55,279
+Define mixture
+
+239
+00:10:55,920 --> 00:11:01,240
+weights uh which are like um
+
+240
+00:11:08,480 --> 00:11:13,480
+where the sum of the mixture weights is
+
+241
+00:11:10,760 --> 00:11:16,160
+equal to one and this is always between
+
+242
+00:11:13,480 --> 00:11:18,639
+zero and one and so if you do this then
+
+243
+00:11:16,160 --> 00:11:21,000
+this is just constant and you can uh
+
+244
+00:11:18,639 --> 00:11:23,519
+interpolate them together constantly but
+
+245
+00:11:21,000 --> 00:11:25,680
+you can also actually explicitly model
+
+246
+00:11:23,519 --> 00:11:27,240
+this probability and say oh I'm
+
+247
+00:11:25,680 --> 00:11:30,279
+currently in a situation where I really
+
+248
+00:11:27,240 --> 00:11:31,880
+think model M will do a good job of uh
+
+249
+00:11:30,279 --> 00:11:33,440
+you know predicting the probability so I
+
+250
+00:11:31,880 --> 00:11:36,160
+want to put most of my probability on
+
+251
+00:11:33,440 --> 00:11:39,000
+model M so you can actually learn this
+
+252
+00:11:36,160 --> 00:11:40,079
+dynamically as well um and so if you
+
+253
+00:11:39,000 --> 00:11:44,360
+have
+
+254
+00:11:40,079 --> 00:11:45,920
+uh this actually um is rather practical
+
+255
+00:11:44,360 --> 00:11:47,120
+and easy to do because what you can do
+
+256
+00:11:45,920 --> 00:11:48,920
+is you can just calculate the
+
+257
+00:11:47,120 --> 00:11:51,399
+probability according to each model at
+
+258
+00:11:48,920 --> 00:11:53,120
+each time step and train this model
+
+259
+00:11:51,399 --> 00:11:55,519
+separately without loading these models
+
+260
+00:11:53,120 --> 00:11:59,399
+into memory at at the time of training
+
+261
+00:11:55,519 --> 00:12:00,959
+those models so uh yeah this is um some
+
+262
+00:11:59,399 --> 00:12:04,800
+you can do as
+
+263
+00:12:00,959 --> 00:12:04,800
+well any questions about
+
+264
+00:12:06,680 --> 00:12:11,920
+this
+
+265
+00:12:08,519 --> 00:12:14,000
+Okay cool so the other option is log
+
+266
+00:12:11,920 --> 00:12:15,800
+linear interpolation and so linear
+
+267
+00:12:14,000 --> 00:12:18,680
+interpolation you're taking a linear
+
+268
+00:12:15,800 --> 00:12:22,040
+combination of the probabilities of each
+
+269
+00:12:18,680 --> 00:12:24,959
+model log linear interpolation you're
+
+270
+00:12:22,040 --> 00:12:26,079
+combining together the log probabilities
+
+271
+00:12:24,959 --> 00:12:29,519
+of each
+
+272
+00:12:26,079 --> 00:12:32,639
+model and then renormalizing so so that
+
+273
+00:12:29,519 --> 00:12:34,920
+you get um that you get an actual
+
+274
+00:12:32,639 --> 00:12:37,760
+probabilistic output so basically what
+
+275
+00:12:34,920 --> 00:12:40,720
+you do is you have this uh interpolation
+
+276
+00:12:37,760 --> 00:12:44,040
+coefficient like I had before but you're
+
+277
+00:12:40,720 --> 00:12:44,040
+combining together the log
+
+278
+00:12:44,639 --> 00:12:49,639
+probabilities and so here we need to
+
+279
+00:12:47,680 --> 00:12:51,320
+take the soft
+
+280
+00:12:49,639 --> 00:12:53,760
+Max
+
+281
+00:12:51,320 --> 00:12:55,760
+um thinking back here I didn't take the
+
+282
+00:12:53,760 --> 00:12:58,120
+softmax does anyone have an idea why I
+
+283
+00:12:55,760 --> 00:13:02,000
+didn't take the soft
+
+284
+00:12:58,120 --> 00:13:02,000
+Max or why I didn't need
+
+285
+00:13:08,160 --> 00:13:12,199
+to why why I need to
+
+286
+00:13:21,600 --> 00:13:27,680
+here yeah
+
+287
+00:13:23,639 --> 00:13:30,440
+so this probability is gu to be z z and
+
+288
+00:13:27,680 --> 00:13:32,240
+one and add up to one this probability
+
+289
+00:13:30,440 --> 00:13:33,760
+is also guaranteed to be zero and one
+
+290
+00:13:32,240 --> 00:13:35,680
+and add up to one and then when you
+
+291
+00:13:33,760 --> 00:13:37,120
+multiply those together uh you can do a
+
+292
+00:13:35,680 --> 00:13:39,160
+little bit of math and demonstrate that
+
+293
+00:13:37,120 --> 00:13:41,440
+the resulting thing will be between zero
+
+294
+00:13:39,160 --> 00:13:42,839
+and one and add up to one that's not the
+
+295
+00:13:41,440 --> 00:13:44,399
+case anymore when we start doing things
+
+296
+00:13:42,839 --> 00:13:47,639
+in log space because it's just not a
+
+297
+00:13:44,399 --> 00:13:50,160
+linear function anyway so um you need to
+
+298
+00:13:47,639 --> 00:13:51,959
+renormalize like this luckily this is
+
+299
+00:13:50,160 --> 00:13:54,920
+super easy like anything else you do in
+
+300
+00:13:51,959 --> 00:13:56,959
+py torch you just add things together
+
+301
+00:13:54,920 --> 00:13:59,320
+and take a soft Max and you'll you'll
+
+302
+00:13:56,959 --> 00:14:02,519
+get an output but you do need to do
+
+303
+00:13:59,320 --> 00:14:05,279
+otherwise you're going to get something
+
+304
+00:14:02,519 --> 00:14:07,279
+weird um the interpolation coefficient
+
+305
+00:14:05,279 --> 00:14:09,639
+here also can be set to a constant so
+
+306
+00:14:07,279 --> 00:14:12,759
+you can you could learn it uh kind of
+
+307
+00:14:09,639 --> 00:14:15,320
+dynamically or it could be
+
+308
+00:14:12,759 --> 00:14:17,720
+separate cool and these actually have
+
+309
+00:14:15,320 --> 00:14:19,639
+different meaning oh sorry go ahead you
+
+310
+00:14:17,720 --> 00:14:23,880
+T on
+
+311
+00:14:19,639 --> 00:14:26,759
+the Yeah Yeah so basically the
+
+312
+00:14:23,880 --> 00:14:29,880
+way the way you would do this is you
+
+313
+00:14:26,759 --> 00:14:32,399
+would have either
+
+314
+00:14:29,880 --> 00:14:33,920
+the same model you you would either take
+
+315
+00:14:32,399 --> 00:14:35,279
+representations from one of these
+
+316
+00:14:33,920 --> 00:14:37,480
+language models or you would take
+
+317
+00:14:35,279 --> 00:14:38,440
+representations from another model and
+
+318
+00:14:37,480 --> 00:14:41,639
+you would
+
+319
+00:14:38,440 --> 00:14:43,959
+just have a model that
+
+320
+00:14:41,639 --> 00:14:46,480
+predicts uh what this interpolation
+
+321
+00:14:43,959 --> 00:14:48,279
+coefficient would be and the
+
+322
+00:14:46,480 --> 00:14:49,720
+optimization objective for that
+
+323
+00:14:48,279 --> 00:14:52,759
+interpolation coefficient is just
+
+324
+00:14:49,720 --> 00:14:56,120
+maximizing the probability
+
+325
+00:14:52,759 --> 00:14:59,600
+whatever so this could also be good um
+
+326
+00:14:56,120 --> 00:15:01,839
+because this interpolation coefficient
+
+327
+00:14:59,600 --> 00:15:07,160
+only like let's say you're interpolating
+
+328
+00:15:01,839 --> 00:15:09,399
+two models together it has one degree of
+
+329
+00:15:07,160 --> 00:15:13,320
+Freedom at each time step right because
+
+330
+00:15:09,399 --> 00:15:15,320
+you're only predicting a probability um
+
+331
+00:15:13,320 --> 00:15:17,839
+if you have uh if you have five models
+
+332
+00:15:15,320 --> 00:15:20,240
+you have uh you basically would be doing
+
+333
+00:15:17,839 --> 00:15:24,199
+a soft match over
+
+334
+00:15:20,240 --> 00:15:25,519
+five five outputs and that's a lot fewer
+
+335
+00:15:24,199 --> 00:15:27,600
+that's a lot fewer than the whole
+
+336
+00:15:25,519 --> 00:15:29,880
+vocabulary right and so this is
+
+337
+00:15:27,600 --> 00:15:31,639
+relatively learning a good interpolation
+
+338
+00:15:29,880 --> 00:15:34,160
+coefficient is relatively easy compared
+
+339
+00:15:31,639 --> 00:15:35,800
+to learning what word to predict next
+
+340
+00:15:34,160 --> 00:15:36,880
+and because of this you could actually
+
+341
+00:15:35,800 --> 00:15:39,759
+tune
+
+342
+00:15:36,880 --> 00:15:42,880
+this um sorry you could tune this
+
+343
+00:15:39,759 --> 00:15:44,600
+probability on a very small data set and
+
+344
+00:15:42,880 --> 00:15:46,959
+you could even have it be context
+
+345
+00:15:44,600 --> 00:15:48,480
+independent so you could just be you
+
+346
+00:15:46,959 --> 00:15:51,399
+know
+
+347
+00:15:48,480 --> 00:15:55,880
+calculating literally five five
+
+348
+00:15:51,399 --> 00:15:57,399
+parameters here um and so because of
+
+349
+00:15:55,880 --> 00:16:00,319
+that like let's say you have a special
+
+350
+00:15:57,399 --> 00:16:02,639
+domain or a special task where you have
+
+351
+00:16:00,319 --> 00:16:04,920
+like 50 training examples or something
+
+352
+00:16:02,639 --> 00:16:07,399
+like that or you know 100 training
+
+353
+00:16:04,920 --> 00:16:08,959
+examples you can learn this
+
+354
+00:16:07,399 --> 00:16:12,480
+interpolation coefficient very
+
+355
+00:16:08,959 --> 00:16:15,880
+effectively uh on just a few a very
+
+356
+00:16:12,480 --> 00:16:18,120
+small number of training examples um but
+
+357
+00:16:15,880 --> 00:16:20,000
+like it could be very useful because
+
+358
+00:16:18,120 --> 00:16:23,920
+like let's say you have a special domain
+
+359
+00:16:20,000 --> 00:16:25,639
+medical language model that's 1.3
+
+360
+00:16:23,920 --> 00:16:27,759
+billion parameters that you trained
+
+361
+00:16:25,639 --> 00:16:29,639
+yourself and then you have a 70 billion
+
+362
+00:16:27,759 --> 00:16:31,079
+parameter language model
+
+363
+00:16:29,639 --> 00:16:33,680
+that's like really good at modeling
+
+364
+00:16:31,079 --> 00:16:35,399
+General English um so then you could
+
+365
+00:16:33,680 --> 00:16:39,120
+learn the interpolation coefficient
+
+366
+00:16:35,399 --> 00:16:40,600
+between those two such that um the large
+
+367
+00:16:39,120 --> 00:16:41,800
+general purpose language model will be
+
+368
+00:16:40,600 --> 00:16:43,959
+generating all of the kind of
+
+369
+00:16:41,800 --> 00:16:46,360
+grammatical stuff but whenever you
+
+370
+00:16:43,959 --> 00:16:48,480
+switch over to modeling technical terms
+
+371
+00:16:46,360 --> 00:16:50,040
+from the medical domain then it learns
+
+372
+00:16:48,480 --> 00:16:52,480
+to upweight the medical language model
+
+373
+00:16:50,040 --> 00:16:54,199
+or something so this can be quite uh
+
+374
+00:16:52,480 --> 00:16:57,000
+this can be quite effective if you have
+
+375
+00:16:54,199 --> 00:17:00,839
+a limited amount of data that you want
+
+376
+00:16:57,000 --> 00:17:00,839
+toing thiss
+
+377
+00:17:01,240 --> 00:17:05,600
+um any other questions about that
+
+378
+00:17:09,079 --> 00:17:14,880
+yeah yeah I'm just gonna talk about that
+
+379
+00:17:11,760 --> 00:17:17,640
+next so linear versus log linear you can
+
+380
+00:17:14,880 --> 00:17:20,880
+actually think of this in logic um and
+
+381
+00:17:17,640 --> 00:17:23,640
+what I mean by that is um linear is kind
+
+382
+00:17:20,880 --> 00:17:26,640
+of like a logical or it tries to come up
+
+383
+00:17:23,640 --> 00:17:29,600
+with examples where either one of the
+
+384
+00:17:26,640 --> 00:17:31,679
+two assigns a high probability so we
+
+385
+00:17:29,600 --> 00:17:36,200
+have the example of like bark
+
+386
+00:17:31,679 --> 00:17:36,200
+run um bark run
+
+387
+00:17:55,640 --> 00:18:03,840
+diet so if we take the average of these
+
+388
+00:18:00,360 --> 00:18:03,840
+two in linear
+
+389
+00:18:04,120 --> 00:18:10,240
+space this would be
+
+390
+00:18:07,159 --> 00:18:13,679
+0.2 this would be
+
+391
+00:18:10,240 --> 00:18:17,240
+0.26 and this would
+
+392
+00:18:13,679 --> 00:18:17,240
+be um
+
+393
+00:18:17,400 --> 00:18:26,280
+0.21 and so a a linear combination
+
+394
+00:18:21,480 --> 00:18:28,600
+between the two will find run to be the
+
+395
+00:18:26,280 --> 00:18:30,600
+highest scoring one because on the left
+
+396
+00:18:28,600 --> 00:18:32,280
+side we have one model that really likes
+
+397
+00:18:30,600 --> 00:18:33,159
+this output and we have another model
+
+398
+00:18:32,280 --> 00:18:35,159
+that
+
+399
+00:18:33,159 --> 00:18:39,280
+doesn't
+
+400
+00:18:35,159 --> 00:18:42,159
+um this is this can be good at using
+
+401
+00:18:39,280 --> 00:18:44,440
+models that capture uh different traits
+
+402
+00:18:42,159 --> 00:18:47,679
+or it can also be useful if like for
+
+403
+00:18:44,440 --> 00:18:49,840
+example you have a you have a small
+
+404
+00:18:47,679 --> 00:18:52,320
+model that you really that really
+
+405
+00:18:49,840 --> 00:18:53,840
+captures like very specific vocabulary
+
+406
+00:18:52,320 --> 00:18:55,520
+and you want to upgrate that specific
+
+407
+00:18:53,840 --> 00:18:56,799
+vocabulary that gets a really low
+
+408
+00:18:55,520 --> 00:18:57,720
+probability according to a general
+
+409
+00:18:56,799 --> 00:19:01,360
+purpose
+
+410
+00:18:57,720 --> 00:19:03,200
+model um this is also necessary when any
+
+411
+00:19:01,360 --> 00:19:04,520
+model can assign zero probabilities so
+
+412
+00:19:03,200 --> 00:19:06,720
+if you have like an example of
+
+413
+00:19:04,520 --> 00:19:10,080
+vocabulary that isn't included in the
+
+414
+00:19:06,720 --> 00:19:11,159
+the like vocabulary of another model or
+
+415
+00:19:10,080 --> 00:19:14,280
+you have models with different
+
+416
+00:19:11,159 --> 00:19:17,200
+vocabularies it's necessary to do this
+
+417
+00:19:14,280 --> 00:19:19,200
+log linear is more like logical and um
+
+418
+00:19:17,200 --> 00:19:22,240
+so the interpolated model only likes
+
+419
+00:19:19,200 --> 00:19:23,799
+choices where all the models agree and
+
+420
+00:19:22,240 --> 00:19:25,640
+this is particularly good when you want
+
+421
+00:19:23,799 --> 00:19:27,440
+to restrict possible answers like you
+
+422
+00:19:25,640 --> 00:19:29,280
+want to have one model be able to say no
+
+423
+00:19:27,440 --> 00:19:32,080
+I really don't like this so never output
+
+424
+00:19:29,280 --> 00:19:34,200
+it so um for example if you wanted to
+
+425
+00:19:32,080 --> 00:19:37,360
+train a model that you knew was very
+
+426
+00:19:34,200 --> 00:19:38,919
+adverse to toxic language and prevent uh
+
+427
+00:19:37,360 --> 00:19:42,600
+the model from outputting toxic language
+
+428
+00:19:38,919 --> 00:19:45,200
+you could use log linear mod so I I
+
+429
+00:19:42,600 --> 00:19:47,559
+can't unfortunately uh calculate logs
+
+430
+00:19:45,200 --> 00:19:50,080
+and exponents in my head well enough to
+
+431
+00:19:47,559 --> 00:19:51,600
+uh to decide this but I'm sure that a
+
+432
+00:19:50,080 --> 00:19:53,840
+linear
+
+433
+00:19:51,600 --> 00:19:56,840
+model the linear model would pick the
+
+434
+00:19:53,840 --> 00:19:59,600
+first one here and the log linear
+
+435
+00:19:56,840 --> 00:20:01,679
+model would pick the second one because
+
+436
+00:19:59,600 --> 00:20:05,640
+the second one has a very low score here
+
+437
+00:20:01,679 --> 00:20:08,640
+so that would be downrated um
+
+438
+00:20:05,640 --> 00:20:08,640
+by
+
+439
+00:20:16,919 --> 00:20:20,640
+yeah yeah so
+
+440
+00:20:25,840 --> 00:20:31,000
+if yeah and if there's any chance of
+
+441
+00:20:28,760 --> 00:20:34,159
+assigning zero probability according to
+
+442
+00:20:31,000 --> 00:20:36,520
+a language model then really you can't
+
+443
+00:20:34,159 --> 00:20:38,200
+even test that language model on that on
+
+444
+00:20:36,520 --> 00:20:42,120
+that test set
+
+445
+00:20:38,200 --> 00:20:43,640
+um so the issue becomes like let's say
+
+446
+00:20:42,120 --> 00:20:45,559
+you have two models with different
+
+447
+00:20:43,640 --> 00:20:47,080
+vocabulary if you have two models with
+
+448
+00:20:45,559 --> 00:20:49,080
+different vocabulary it becomes very
+
+449
+00:20:47,080 --> 00:20:50,559
+tricky how to reconcile those two but
+
+450
+00:20:49,080 --> 00:20:53,440
+you could do linear interpolation
+
+451
+00:20:50,559 --> 00:20:55,200
+between them like match the vocab the
+
+452
+00:20:53,440 --> 00:20:57,559
+output vocabularies that they do have
+
+453
+00:20:55,200 --> 00:21:00,120
+and then just not worry about the fact
+
+454
+00:20:57,559 --> 00:21:02,760
+that the vocabularies are dis jointed
+
+455
+00:21:00,120 --> 00:21:05,039
+and because one will assign a zero
+
+456
+00:21:02,760 --> 00:21:07,280
+probability to those vocabulary items
+
+457
+00:21:05,039 --> 00:21:12,240
+but the other one is fine so you can
+
+458
+00:21:07,280 --> 00:21:14,919
+just do that but if you're in general it
+
+459
+00:21:12,240 --> 00:21:16,480
+will be very tricky to try to get two
+
+460
+00:21:14,919 --> 00:21:18,559
+models with different vocabularies to
+
+461
+00:21:16,480 --> 00:21:21,480
+play together nicely so I I would
+
+462
+00:21:18,559 --> 00:21:22,919
+suggest um thinking about thinking
+
+463
+00:21:21,480 --> 00:21:25,600
+seriously about whether you need to do
+
+464
+00:21:22,919 --> 00:21:31,360
+that or not before you start out but
+
+465
+00:21:25,600 --> 00:21:31,360
+yeah um uh yes there any
+
+466
+00:21:35,559 --> 00:21:40,960
+other
+
+467
+00:21:38,039 --> 00:21:43,360
+um you could definitely so the question
+
+468
+00:21:40,960 --> 00:21:45,000
+is are there any other types of
+
+469
+00:21:43,360 --> 00:21:47,760
+interpolation that have other types of
+
+470
+00:21:45,000 --> 00:21:50,159
+logical components like exor or nor um
+
+471
+00:21:47,760 --> 00:21:52,840
+you could definitely come up with one uh
+
+472
+00:21:50,159 --> 00:21:55,440
+I I am struggling a little bit to think
+
+473
+00:21:52,840 --> 00:21:57,520
+about when you would want to do that but
+
+474
+00:21:55,440 --> 00:22:02,840
+I'm sure
+
+475
+00:21:57,520 --> 00:22:05,840
+you is is the inherent that the
+
+476
+00:22:02,840 --> 00:22:05,840
+err
+
+477
+00:22:09,120 --> 00:22:14,480
+not so what what if the errors are not
+
+478
+00:22:12,640 --> 00:22:15,919
+what if the errors are correlated so
+
+479
+00:22:14,480 --> 00:22:18,200
+think about what happens if the errors
+
+480
+00:22:15,919 --> 00:22:20,000
+are perfectly correlated um which is
+
+481
+00:22:18,200 --> 00:22:25,840
+when you're using the same model in two
+
+482
+00:22:20,000 --> 00:22:25,840
+parts of the uh like on top so you
+
+483
+00:22:27,000 --> 00:22:30,520
+literally uh these
+
+484
+00:22:29,159 --> 00:22:32,679
+model one and model two are the same
+
+485
+00:22:30,520 --> 00:22:36,720
+model if that's the case nothing happens
+
+486
+00:22:32,679 --> 00:22:39,200
+it doesn't get worse um and
+
+487
+00:22:36,720 --> 00:22:43,039
+so of course because this is machine
+
+488
+00:22:39,200 --> 00:22:45,080
+learning there's no guarantee like you
+
+489
+00:22:43,039 --> 00:22:47,559
+know unless we make some assumptions
+
+490
+00:22:45,080 --> 00:22:49,200
+about the relationship between like the
+
+491
+00:22:47,559 --> 00:22:52,279
+training set and the test set or the
+
+492
+00:22:49,200 --> 00:22:53,760
+models errors in the test set um you can
+
+493
+00:22:52,279 --> 00:22:57,039
+always do something that will make your
+
+494
+00:22:53,760 --> 00:22:59,240
+accuracy worse um like let's say we flip
+
+495
+00:22:57,039 --> 00:23:00,360
+the labels of a binary class
+
+496
+00:22:59,240 --> 00:23:03,120
+no matter what you do you're going to
+
+497
+00:23:00,360 --> 00:23:06,320
+make your accuracy worse but
+
+498
+00:23:03,120 --> 00:23:09,000
+um no matter what the normal thing you
+
+499
+00:23:06,320 --> 00:23:10,640
+would do is it would make your if it
+
+500
+00:23:09,000 --> 00:23:12,480
+would improve accuracy normally it would
+
+501
+00:23:10,640 --> 00:23:14,760
+decrease your accuracy but like under
+
+502
+00:23:12,480 --> 00:23:16,080
+pretty reasonable assumptions it's
+
+503
+00:23:14,760 --> 00:23:20,400
+mostly going to be the case that errors
+
+504
+00:23:16,080 --> 00:23:22,320
+are deated to some extent um
+
+505
+00:23:20,400 --> 00:23:25,559
+so
+
+506
+00:23:22,320 --> 00:23:30,440
+yeah you and because of that ensembly
+
+507
+00:23:25,559 --> 00:23:30,440
+usually helps yeah
+
+508
+00:23:36,120 --> 00:23:42,019
+um about which one
+
+509
+00:23:38,760 --> 00:23:42,019
+[Music]
+
+510
+00:23:53,559 --> 00:24:01,240
+which let me make sure I didn't mess it
+
+511
+00:23:55,640 --> 00:24:01,240
+up on sides okay so in my
+
+512
+00:24:06,960 --> 00:24:13,120
+example yeah yeah
+
+513
+00:24:09,640 --> 00:24:13,120
+yeah sorry about
+
+514
+00:24:14,360 --> 00:24:19,320
+that because this is this is where the
+
+515
+00:24:17,039 --> 00:24:21,840
+average is higher and then this is
+
+516
+00:24:19,320 --> 00:24:27,200
+one take
+
+517
+00:24:21,840 --> 00:24:29,039
+you uh cool any other any other
+
+518
+00:24:27,200 --> 00:24:31,840
+questions okay
+
+519
+00:24:29,039 --> 00:24:34,440
+okay so
+
+520
+00:24:31,840 --> 00:24:36,320
+um another thing I should point out is
+
+521
+00:24:34,440 --> 00:24:39,600
+that we don't
+
+522
+00:24:36,320 --> 00:24:41,840
+necessarily need to use models only as
+
+523
+00:24:39,600 --> 00:24:44,080
+positive evidence so if you're using log
+
+524
+00:24:41,840 --> 00:24:46,039
+linear interpolation actually your
+
+525
+00:24:44,080 --> 00:24:49,919
+interpolation coefficients do not need
+
+526
+00:24:46,039 --> 00:24:52,520
+to be positive they can also be negative
+
+527
+00:24:49,919 --> 00:24:55,360
+and you can have uh things where you
+
+528
+00:24:52,520 --> 00:24:57,840
+penalize the probabilities given by a
+
+529
+00:24:55,360 --> 00:24:59,679
+particular model and this has actually
+
+530
+00:24:57,840 --> 00:25:01,520
+been used for a long time it was
+
+531
+00:24:59,679 --> 00:25:04,440
+actually used in machine translation
+
+532
+00:25:01,520 --> 00:25:08,840
+since like uh 2005 or something like
+
+533
+00:25:04,440 --> 00:25:11,480
+this but the basic idea is um that you
+
+534
+00:25:08,840 --> 00:25:13,600
+have some models that serve as negative
+
+535
+00:25:11,480 --> 00:25:15,559
+evidence so you have kind of a core
+
+536
+00:25:13,600 --> 00:25:17,880
+model this might be your really strong
+
+537
+00:25:15,559 --> 00:25:21,520
+general purpose language model you have
+
+538
+00:25:17,880 --> 00:25:23,080
+a positive uh model which is the model
+
+539
+00:25:21,520 --> 00:25:25,240
+that you want to kind of boost up and
+
+540
+00:25:23,080 --> 00:25:27,320
+improve and a negative model which you
+
+541
+00:25:25,240 --> 00:25:31,159
+want to
+
+542
+00:25:27,320 --> 00:25:33,679
+decrease and um one example of this is
+
+543
+00:25:31,159 --> 00:25:36,760
+in uh a paper that we did in
+
+544
+00:25:33,679 --> 00:25:40,159
+2019 um the core was a machine
+
+545
+00:25:36,760 --> 00:25:42,960
+translation model and the negative model
+
+546
+00:25:40,159 --> 00:25:44,880
+is an outof domain language model and
+
+547
+00:25:42,960 --> 00:25:46,960
+the positive model is an in domain
+
+548
+00:25:44,880 --> 00:25:51,039
+language model and so the idea behind
+
+549
+00:25:46,960 --> 00:25:53,880
+this is a machine translation model um
+
+550
+00:25:51,039 --> 00:25:55,600
+you have to train it on machine
+
+551
+00:25:53,880 --> 00:25:58,320
+translation data and machine translation
+
+552
+00:25:55,600 --> 00:26:00,640
+data is not very easy to get for
+
+553
+00:25:58,320 --> 00:26:02,360
+particular domains for example um you
+
+554
+00:26:00,640 --> 00:26:03,880
+might only have machine translation data
+
+555
+00:26:02,360 --> 00:26:06,919
+in the news domain and you actually want
+
+556
+00:26:03,880 --> 00:26:09,240
+to be uh doing uh translation in the
+
+557
+00:26:06,919 --> 00:26:12,720
+medical domain or something so what you
+
+558
+00:26:09,240 --> 00:26:14,640
+do is you have your positive model here
+
+559
+00:26:12,720 --> 00:26:17,600
+this could be a new this is a machine
+
+560
+00:26:14,640 --> 00:26:19,919
+translation model this could be a news
+
+561
+00:26:17,600 --> 00:26:21,320
+domain or sorry this could be a medical
+
+562
+00:26:19,919 --> 00:26:22,919
+domain language model and this could be
+
+563
+00:26:21,320 --> 00:26:24,360
+a news domain language model so you're
+
+564
+00:26:22,919 --> 00:26:25,840
+subtracting out the news domain
+
+565
+00:26:24,360 --> 00:26:27,600
+probabilities and adding in medical
+
+566
+00:26:25,840 --> 00:26:30,240
+domain probabilities move it in that
+
+567
+00:26:27,600 --> 00:26:30,240
+direction
+
+568
+00:26:30,440 --> 00:26:36,799
+um another example of this is uh
+
+569
+00:26:32,919 --> 00:26:40,000
+something called uh D experts um or
+
+570
+00:26:36,799 --> 00:26:43,440
+dexperts and the idea here is here you
+
+571
+00:26:40,000 --> 00:26:46,120
+have a strong language model as your
+
+572
+00:26:43,440 --> 00:26:48,320
+core and then as negative you have a
+
+573
+00:26:46,120 --> 00:26:50,240
+weak toxic language model so it was
+
+574
+00:26:48,320 --> 00:26:52,760
+trained on lot lots of like bad texts
+
+575
+00:26:50,240 --> 00:26:55,799
+that you don't want to be generating and
+
+576
+00:26:52,760 --> 00:26:57,159
+the positive is a weak non-toxic
+
+577
+00:26:55,799 --> 00:26:59,279
+language model that was trained on lots
+
+578
+00:26:57,159 --> 00:27:03,200
+of like inocua
+
+579
+00:26:59,279 --> 00:27:04,399
+posts so that would help you detoxify
+
+580
+00:27:03,200 --> 00:27:06,679
+the outputs of the
+
+581
+00:27:04,399 --> 00:27:09,799
+language so there's lots of examples of
+
+582
+00:27:06,679 --> 00:27:09,799
+things like this that you can do
+
+583
+00:27:10,720 --> 00:27:15,880
+through
+
+584
+00:27:12,880 --> 00:27:15,880
+yeah
+
+585
+00:27:19,320 --> 00:27:25,880
+yeah um so the positive in the machine
+
+586
+00:27:22,840 --> 00:27:27,679
+translation example this is a so this is
+
+587
+00:27:25,880 --> 00:27:31,760
+a machine translation model where the
+
+588
+00:27:27,679 --> 00:27:34,080
+input is is like in um English and out
+
+589
+00:27:31,760 --> 00:27:37,880
+is in Japanese something like
+
+590
+00:27:34,080 --> 00:27:39,679
+that this is only trained on Japanese
+
+591
+00:27:37,880 --> 00:27:42,919
+but it's trained on like medical
+
+592
+00:27:39,679 --> 00:27:44,440
+Japanese for example Med the domain one
+
+593
+00:27:42,919 --> 00:27:48,480
+this is a language model that was
+
+594
+00:27:44,440 --> 00:27:50,600
+trained on like news domain um Japanese
+
+595
+00:27:48,480 --> 00:27:54,039
+or it could even literally just be
+
+596
+00:27:50,600 --> 00:27:56,360
+trained on the side of the machine
+
+597
+00:27:54,039 --> 00:28:00,120
+trans um so it's trying to remove out
+
+598
+00:27:56,360 --> 00:28:00,120
+the language modeling component from the
+
+599
+00:28:03,720 --> 00:28:06,720
+cool
+
+600
+00:28:06,880 --> 00:28:11,480
+okay so another thing that I should
+
+601
+00:28:09,880 --> 00:28:14,720
+point out I didn't actually put it on
+
+602
+00:28:11,480 --> 00:28:18,399
+the slides is um there's a lot of other
+
+603
+00:28:14,720 --> 00:28:19,640
+ways to get multiple models and um I
+
+604
+00:28:18,399 --> 00:28:22,600
+think a lot of people are probably
+
+605
+00:28:19,640 --> 00:28:23,559
+familiar with Dropout um it's a method
+
+606
+00:28:22,600 --> 00:28:27,120
+for
+
+607
+00:28:23,559 --> 00:28:29,080
+regularizing um it's a method for
+
+608
+00:28:27,120 --> 00:28:31,120
+regularizing
+
+609
+00:28:29,080 --> 00:28:33,760
+neural networks or deep learning models
+
+610
+00:28:31,120 --> 00:28:37,279
+in general and basically the idea is
+
+611
+00:28:33,760 --> 00:28:41,840
+every once in a while um during training
+
+612
+00:28:37,279 --> 00:28:45,720
+you drop out some portion of the uh like
+
+613
+00:28:41,840 --> 00:28:48,919
+nodes in the neural network model and
+
+614
+00:28:45,720 --> 00:28:51,320
+you can actually drop
+
+615
+00:28:48,919 --> 00:28:52,640
+out and normally what you do is at test
+
+616
+00:28:51,320 --> 00:28:53,919
+time then you just don't drop out
+
+617
+00:28:52,640 --> 00:28:56,039
+anything and you use the whole neural
+
+618
+00:28:53,919 --> 00:28:59,960
+network model but another thing you can
+
+619
+00:28:56,039 --> 00:29:02,559
+do is you can drop out a test time drop
+
+620
+00:28:59,960 --> 00:29:04,679
+out five times and combine those
+
+621
+00:29:02,559 --> 00:29:06,600
+different models together through ensom
+
+622
+00:29:04,679 --> 00:29:10,600
+and that's actually something uh that
+
+623
+00:29:06,600 --> 00:29:14,480
+people tried in the uh in the Dropout
+
+624
+00:29:10,600 --> 00:29:17,600
+paper and this is one way to get
+
+625
+00:29:14,480 --> 00:29:19,640
+multiple models uh and actually you can
+
+626
+00:29:17,600 --> 00:29:21,919
+demonstrate that this helps the original
+
+627
+00:29:19,640 --> 00:29:24,519
+motivation behind Dropout was precisely
+
+628
+00:29:21,919 --> 00:29:26,279
+coming from this idea of
+
+629
+00:29:24,519 --> 00:29:29,080
+ensembling
+
+630
+00:29:26,279 --> 00:29:31,399
+another method
+
+631
+00:29:29,080 --> 00:29:34,799
+that has been around for a very long
+
+632
+00:29:31,399 --> 00:29:37,760
+time it's another embling method is
+
+633
+00:29:34,799 --> 00:29:41,919
+bagging and basically the way bagging
+
+634
+00:29:37,760 --> 00:29:41,919
+works is you have a data
+
+635
+00:29:44,000 --> 00:29:50,159
+set like this and you just resample the
+
+636
+00:29:47,519 --> 00:29:52,919
+data set so you sample all of the output
+
+637
+00:29:50,159 --> 00:29:55,200
+with uh replacement and you get another
+
+638
+00:29:52,919 --> 00:29:57,799
+data set of equal size and then you
+
+639
+00:29:55,200 --> 00:29:58,559
+train on this but you do that like 10
+
+640
+00:29:57,799 --> 00:30:00,120
+times
+
+641
+00:29:58,559 --> 00:30:02,679
+and you train 10 different models and
+
+642
+00:30:00,120 --> 00:30:04,360
+then you emble those models together and
+
+643
+00:30:02,679 --> 00:30:06,000
+so this is another way to get multiple
+
+644
+00:30:04,360 --> 00:30:07,519
+models and both of these still improve
+
+645
+00:30:06,000 --> 00:30:09,640
+your robustness because they basically
+
+646
+00:30:07,519 --> 00:30:11,440
+get a different view on the data so they
+
+647
+00:30:09,640 --> 00:30:13,440
+smooth over some of the
+
+648
+00:30:11,440 --> 00:30:15,360
+idiosyncrasies um and as I mentioned
+
+649
+00:30:13,440 --> 00:30:17,960
+before you can also get multiple models
+
+650
+00:30:15,360 --> 00:30:20,120
+from different checkpoints and then uh
+
+651
+00:30:17,960 --> 00:30:22,159
+put them together and all of these
+
+652
+00:30:20,120 --> 00:30:24,159
+methods are pretty related both of them
+
+653
+00:30:22,159 --> 00:30:25,960
+basically what they're doing is they're
+
+654
+00:30:24,159 --> 00:30:28,279
+taking advantage of the fact that you
+
+655
+00:30:25,960 --> 00:30:29,919
+have particular models that saw
+
+656
+00:30:28,279 --> 00:30:32,760
+different data or saw data in a
+
+657
+00:30:29,919 --> 00:30:34,120
+different order or different nodes saw
+
+658
+00:30:32,760 --> 00:30:35,679
+different parts of the data because you
+
+659
+00:30:34,120 --> 00:30:37,799
+dropped out some of the nodes when they
+
+660
+00:30:35,679 --> 00:30:41,840
+were back propping on particular
+
+661
+00:30:37,799 --> 00:30:44,840
+varieties of the data so um even things
+
+662
+00:30:41,840 --> 00:30:46,720
+like this can give you models that are
+
+663
+00:30:44,840 --> 00:30:49,760
+different enough that to help uh when
+
+664
+00:30:46,720 --> 00:30:49,760
+you're onbling or
+
+665
+00:30:52,559 --> 00:30:59,360
+combining and then of course um you can
+
+666
+00:30:56,919 --> 00:31:00,799
+also
+
+667
+00:30:59,360 --> 00:31:02,480
+then of course you can also combine
+
+668
+00:31:00,799 --> 00:31:06,960
+together like very different models like
+
+669
+00:31:02,480 --> 00:31:06,960
+this and that also works in different
+
+670
+00:31:07,240 --> 00:31:11,159
+ways
+
+671
+00:31:09,000 --> 00:31:13,039
+cool part of the reason why I wanted to
+
+672
+00:31:11,159 --> 00:31:15,320
+mention that Dropout though in
+
+673
+00:31:13,039 --> 00:31:17,120
+particular is there's also other
+
+674
+00:31:15,320 --> 00:31:19,240
+efficient methods for using multiple
+
+675
+00:31:17,120 --> 00:31:22,000
+models so the big problem with
+
+676
+00:31:19,240 --> 00:31:25,399
+ensembling is the cost
+
+677
+00:31:22,000 --> 00:31:27,159
+and simple ensembling is very expensive
+
+678
+00:31:25,399 --> 00:31:29,240
+because it requires you to run multiple
+
+679
+00:31:27,159 --> 00:31:30,519
+models at test test time at inference
+
+680
+00:31:29,240 --> 00:31:33,720
+time and this is something you don't
+
+681
+00:31:30,519 --> 00:31:35,279
+want to be doing if you're you know
+
+682
+00:31:33,720 --> 00:31:38,679
+deploying a service or something because
+
+683
+00:31:35,279 --> 00:31:41,080
+it like linearly increases your cost by
+
+684
+00:31:38,679 --> 00:31:45,200
+um the amount of bottles that you're
+
+685
+00:31:41,080 --> 00:31:47,799
+running and it requires both end times
+
+686
+00:31:45,200 --> 00:31:50,120
+of computation and end times of memory
+
+687
+00:31:47,799 --> 00:31:51,720
+and memory is actually probably the
+
+688
+00:31:50,120 --> 00:31:54,279
+worst thing because you need to deploy
+
+689
+00:31:51,720 --> 00:31:58,159
+extra GPU machines and other stuff like
+
+690
+00:31:54,279 --> 00:31:59,880
+that so um the question is is there any
+
+691
+00:31:58,159 --> 00:32:03,279
+way we can get some of the benefits of
+
+692
+00:31:59,880 --> 00:32:06,519
+embling without having to create
+
+693
+00:32:03,279 --> 00:32:07,320
+multiple models and luckily the answer
+
+694
+00:32:06,519 --> 00:32:09,240
+is
+
+695
+00:32:07,320 --> 00:32:11,919
+yes
+
+696
+00:32:09,240 --> 00:32:13,960
+the method the easiest method for doing
+
+697
+00:32:11,919 --> 00:32:16,600
+so is something called parameter
+
+698
+00:32:13,960 --> 00:32:18,399
+averaging and basically what you do is
+
+699
+00:32:16,600 --> 00:32:21,960
+you just average the parameters of
+
+700
+00:32:18,399 --> 00:32:26,039
+multiple models together um this only
+
+701
+00:32:21,960 --> 00:32:29,200
+works under certain conditions so does
+
+702
+00:32:26,039 --> 00:32:31,120
+anyone um does anyone know what these
+
+703
+00:32:29,200 --> 00:32:33,320
+conditions might be there's a few
+
+704
+00:32:31,120 --> 00:32:35,919
+obvious ones and maybe a few slightly
+
+705
+00:32:33,320 --> 00:32:35,919
+less obvious
+
+706
+00:32:36,039 --> 00:32:40,799
+ones so like first question do you think
+
+707
+00:32:38,799 --> 00:32:41,919
+you could combine together do you think
+
+708
+00:32:40,799 --> 00:32:45,880
+you could average together the
+
+709
+00:32:41,919 --> 00:32:45,880
+parameters of llama 7B and Lama
+
+710
+00:32:46,440 --> 00:32:52,639
+70b
+
+711
+00:32:48,480 --> 00:32:52,639
+no the answer is no but why
+
+712
+00:32:54,480 --> 00:32:58,440
+not I mean what does that even mean in
+
+713
+00:32:56,760 --> 00:33:00,480
+the first place right like they have
+
+714
+00:32:58,440 --> 00:33:02,799
+totally different numbers of parameters
+
+715
+00:33:00,480 --> 00:33:05,840
+uh you wouldn't be able to find a one
+
+716
+00:33:02,799 --> 00:33:07,840
+toone association between 7B and like 7
+
+717
+00:33:05,840 --> 00:33:12,320
+billion parameters and 70 billion
+
+718
+00:33:07,840 --> 00:33:16,880
+parameters um what about averaging
+
+719
+00:33:12,320 --> 00:33:19,399
+together uh let's let's say llama 7B and
+
+720
+00:33:16,880 --> 00:33:19,399
+mistol
+
+721
+00:33:23,080 --> 00:33:29,760
+7bs yes no y I'm guessing that like for
+
+722
+00:33:27,440 --> 00:33:29,760
+the
+
+723
+00:33:33,760 --> 00:33:38,120
+yeah for different architectures the um
+
+724
+00:33:36,760 --> 00:33:41,799
+the parameters could mean different
+
+725
+00:33:38,120 --> 00:33:44,159
+things and even if the architecture is
+
+726
+00:33:41,799 --> 00:33:45,880
+exactly the same even if your random
+
+727
+00:33:44,159 --> 00:33:49,880
+initialization is different then that
+
+728
+00:33:45,880 --> 00:33:52,360
+would be a disastrous because basically
+
+729
+00:33:49,880 --> 00:33:54,760
+in neural networks there's no inherent
+
+730
+00:33:52,360 --> 00:33:58,559
+meaning to like parameter number one
+
+731
+00:33:54,760 --> 00:34:01,919
+right um and there's the idea of permut
+
+732
+00:33:58,559 --> 00:34:06,679
+Inari which is
+
+733
+00:34:01,919 --> 00:34:07,639
+um you could like randomly Swap all of
+
+734
+00:34:06,679 --> 00:34:10,280
+the
+
+735
+00:34:07,639 --> 00:34:12,079
+dimensions uh between within a neural
+
+736
+00:34:10,280 --> 00:34:14,760
+network and get exactly the same
+
+737
+00:34:12,079 --> 00:34:17,919
+function
+
+738
+00:34:14,760 --> 00:34:22,560
+uh as long as kind
+
+739
+00:34:17,919 --> 00:34:24,839
+of in layer number one you swap and then
+
+740
+00:34:22,560 --> 00:34:30,359
+also take the inputs in the next layer
+
+741
+00:34:24,839 --> 00:34:30,359
+also so um you know you know as long
+
+742
+00:34:30,960 --> 00:34:36,399
+as if you have a weight Matrix that
+
+743
+00:34:33,679 --> 00:34:40,800
+results in the um in the outputs being
+
+744
+00:34:36,399 --> 00:34:49,639
+ordered like 1 two three four
+
+745
+00:34:40,800 --> 00:34:54,159
+five one or 2 1 3 five four as long as
+
+746
+00:34:49,639 --> 00:34:55,720
+you also swap the input direct input
+
+747
+00:34:54,159 --> 00:34:58,400
+dimensions of this weight Matrix you get
+
+748
+00:34:55,720 --> 00:35:01,520
+exactly the same because they
+
+749
+00:34:58,400 --> 00:35:04,200
+linear combinations of the parameters
+
+750
+00:35:01,520 --> 00:35:06,480
+together so neural networks have this
+
+751
+00:35:04,200 --> 00:35:08,599
+feature of permutation and variance so
+
+752
+00:35:06,480 --> 00:35:11,800
+models that were trained from like
+
+753
+00:35:08,599 --> 00:35:13,280
+different uh different initializations
+
+754
+00:35:11,800 --> 00:35:15,040
+won't be able to be combined together in
+
+755
+00:35:13,280 --> 00:35:18,320
+this
+
+756
+00:35:15,040 --> 00:35:20,079
+way um but the good luck the good thing
+
+757
+00:35:18,320 --> 00:35:21,359
+is actually we have a whole bunch of
+
+758
+00:35:20,079 --> 00:35:25,320
+models that come from the same
+
+759
+00:35:21,359 --> 00:35:26,720
+pre-trained model right uh so we we have
+
+760
+00:35:25,320 --> 00:35:28,640
+this initialization here this
+
+761
+00:35:26,720 --> 00:35:31,280
+initialization was used to train Lama
+
+762
+00:35:28,640 --> 00:35:32,920
+27b but now we have like hundreds
+
+763
+00:35:31,280 --> 00:35:34,440
+hundreds of models that are DED from
+
+764
+00:35:32,920 --> 00:35:37,400
+Lama 2 we have hundreds of models that
+
+765
+00:35:34,440 --> 00:35:39,599
+are DED from mixol and there all of the
+
+766
+00:35:37,400 --> 00:35:40,920
+dimensions actually mean the same thing
+
+767
+00:35:39,599 --> 00:35:43,280
+because they're derived from the same
+
+768
+00:35:40,920 --> 00:35:46,680
+parameters in the first place so those
+
+769
+00:35:43,280 --> 00:35:48,119
+ones we can average together and um
+
+770
+00:35:46,680 --> 00:35:50,359
+there's basically two ways that we can
+
+771
+00:35:48,119 --> 00:35:53,520
+do this uh one is by averaging together
+
+772
+00:35:50,359 --> 00:35:55,240
+multiple checkpoints during training so
+
+773
+00:35:53,520 --> 00:35:57,960
+originally this was the big thing that
+
+774
+00:35:55,240 --> 00:36:00,359
+people did uh like you would train model
+
+775
+00:35:57,960 --> 00:36:02,119
+from scratch for a really long time but
+
+776
+00:36:00,359 --> 00:36:03,920
+then you would take the final five
+
+777
+00:36:02,119 --> 00:36:07,520
+checkpoints and you would just average
+
+778
+00:36:03,920 --> 00:36:09,280
+them together and this helps reduce some
+
+779
+00:36:07,520 --> 00:36:11,040
+of the noise that you get from
+
+780
+00:36:09,280 --> 00:36:13,839
+stochastic gradient descent and can
+
+781
+00:36:11,040 --> 00:36:15,520
+improve your overall accuracy if you're
+
+782
+00:36:13,839 --> 00:36:17,280
+fine-tuning any models this is something
+
+783
+00:36:15,520 --> 00:36:18,680
+you can do also uh because you're
+
+784
+00:36:17,280 --> 00:36:19,800
+probably going to be saving checkpoints
+
+785
+00:36:18,680 --> 00:36:21,160
+you can just take the best five
+
+786
+00:36:19,800 --> 00:36:23,079
+checkpoints and average them together
+
+787
+00:36:21,160 --> 00:36:27,280
+and that actually can improve your
+
+788
+00:36:23,079 --> 00:36:28,160
+accuracy quite a bit um another thing is
+
+789
+00:36:27,280 --> 00:36:31,520
+find
+
+790
+00:36:28,160 --> 00:36:32,880
+uh tuned model merging soine tune um in
+
+791
+00:36:31,520 --> 00:36:35,000
+several ways and then merge them
+
+792
+00:36:32,880 --> 00:36:39,079
+together and so for example we might
+
+793
+00:36:35,000 --> 00:36:41,240
+take Lama 27b instruct and um vuna 7B
+
+794
+00:36:39,079 --> 00:36:44,760
+1.5 and merg them together with some
+
+795
+00:36:41,240 --> 00:36:47,599
+weights and uh we could you
+
+796
+00:36:44,760 --> 00:36:50,319
+know smooth over their idos synchr dises
+
+797
+00:36:47,599 --> 00:36:52,520
+and get better results
+
+798
+00:36:50,319 --> 00:36:56,280
+too
+
+799
+00:36:52,520 --> 00:36:56,280
+cool uh any questions
+
+800
+00:36:56,520 --> 00:36:59,520
+here
+
+801
+00:37:00,920 --> 00:37:03,119
+oh
+
+802
+00:37:04,680 --> 00:37:11,920
+yeah want to so I just
+
+803
+00:37:09,680 --> 00:37:14,079
+came
+
+804
+00:37:11,920 --> 00:37:19,040
+non I
+
+805
+00:37:14,079 --> 00:37:19,040
+use like those different chain and
+
+806
+00:37:19,640 --> 00:37:23,319
+just
+
+807
+00:37:21,160 --> 00:37:26,640
+I pretty
+
+808
+00:37:23,319 --> 00:37:29,520
+efficient because on the same model you
+
+809
+00:37:26,640 --> 00:37:29,520
+get
+
+810
+00:37:35,640 --> 00:37:40,839
+yeah so would this would this parameter
+
+811
+00:37:38,000 --> 00:37:46,119
+averaging be a good method for U making
+
+812
+00:37:40,839 --> 00:37:49,839
+a model less toxic for example the
+
+813
+00:37:46,119 --> 00:37:53,200
+answer is a little bit trickier there I
+
+814
+00:37:49,839 --> 00:37:56,119
+guess because um I I feel like this is
+
+815
+00:37:53,200 --> 00:37:58,160
+good for mixing two models together so
+
+816
+00:37:56,119 --> 00:38:01,400
+if you're mixing your
+
+817
+00:37:58,160 --> 00:38:03,359
+like non-toxicity tuned model or your
+
+818
+00:38:01,400 --> 00:38:06,079
+safety tuned model with the original
+
+819
+00:38:03,359 --> 00:38:07,520
+base model that was not uh safety tuned
+
+820
+00:38:06,079 --> 00:38:08,800
+or something like that then you might
+
+821
+00:38:07,520 --> 00:38:11,240
+get something in the middle so you might
+
+822
+00:38:08,800 --> 00:38:13,319
+get something that's less safe than the
+
+823
+00:38:11,240 --> 00:38:18,720
+uh like the model that was tuned to not
+
+824
+00:38:13,319 --> 00:38:21,400
+be toxic so it might be uh yeah I'm not
+
+825
+00:38:18,720 --> 00:38:23,920
+sure but like let's say you let's say
+
+826
+00:38:21,400 --> 00:38:26,240
+you have a model that somebody
+
+827
+00:38:23,920 --> 00:38:28,640
+else did like a really good job
+
+828
+00:38:26,240 --> 00:38:31,359
+instruction tuning for you
+
+829
+00:38:28,640 --> 00:38:33,640
+um and anytime you start using safety
+
+830
+00:38:31,359 --> 00:38:35,560
+tuning on it you like hurt the
+
+831
+00:38:33,640 --> 00:38:38,680
+instruction tuning like the model gets
+
+832
+00:38:35,560 --> 00:38:40,560
+worse I could see a world where you take
+
+833
+00:38:38,680 --> 00:38:43,000
+the base model the same base model you
+
+834
+00:38:40,560 --> 00:38:45,280
+take llama 27b you train like a less
+
+835
+00:38:43,000 --> 00:38:47,480
+toxic version of llama 27d and then do
+
+836
+00:38:45,280 --> 00:38:51,319
+parameter averaging with the like well
+
+837
+00:38:47,480 --> 00:38:53,160
+instruction tuned model um that might
+
+838
+00:38:51,319 --> 00:38:55,359
+work that might make something that's
+
+839
+00:38:53,160 --> 00:38:57,560
+more safe and like not much worse
+
+840
+00:38:55,359 --> 00:39:01,440
+instruction to so there's definitely I
+
+841
+00:38:57,560 --> 00:39:01,440
+think creative things that you can do
+
+842
+00:39:01,520 --> 00:39:08,400
+that um maybe I'll go directly into the
+
+843
+00:39:04,960 --> 00:39:11,480
+methods um
+
+844
+00:39:08,400 --> 00:39:13,240
+so uh there's a few uh recent papers on
+
+845
+00:39:11,480 --> 00:39:16,000
+this like this method has been around
+
+846
+00:39:13,240 --> 00:39:17,880
+for a long time since at least 1996 but
+
+847
+00:39:16,000 --> 00:39:20,880
+uh recently people have examined it a
+
+848
+00:39:17,880 --> 00:39:24,800
+lot in the context of uh kind of modern
+
+849
+00:39:20,880 --> 00:39:27,400
+networks and uh this paper model soup uh
+
+850
+00:39:24,800 --> 00:39:29,000
+examines two strategies the first one is
+
+851
+00:39:27,400 --> 00:39:31,400
+uniform averaging where you just average
+
+852
+00:39:29,000 --> 00:39:33,560
+all the parameters together uh like as
+
+853
+00:39:31,400 --> 00:39:35,480
+you would expect but they also have a
+
+854
+00:39:33,560 --> 00:39:38,319
+greedy averaging method and basically
+
+855
+00:39:35,480 --> 00:39:40,240
+what they do here is they add one model
+
+856
+00:39:38,319 --> 00:39:42,119
+and check if the whole like averaged
+
+857
+00:39:40,240 --> 00:39:43,680
+model improves and then only if the
+
+858
+00:39:42,119 --> 00:39:45,760
+whole averaged model improves do they
+
+859
+00:39:43,680 --> 00:39:49,040
+keep that model otherwise they throw it
+
+860
+00:39:45,760 --> 00:39:52,960
+out and then they um they don't uh use
+
+861
+00:39:49,040 --> 00:39:54,520
+it so what they demonstrate uh this is a
+
+862
+00:39:52,960 --> 00:39:57,560
+little bit small but basically the
+
+863
+00:39:54,520 --> 00:40:00,520
+purple star here is uh when the use
+
+864
+00:39:57,560 --> 00:40:02,480
+greedy averaging and then the blue
+
+865
+00:40:00,520 --> 00:40:05,119
+circle here is when they use the uniform
+
+866
+00:40:02,480 --> 00:40:08,280
+averaging and then green is all of the
+
+867
+00:40:05,119 --> 00:40:09,960
+models that they they put into this
+
+868
+00:40:08,280 --> 00:40:12,560
+average
+
+869
+00:40:09,960 --> 00:40:16,680
+and what they found
+
+870
+00:40:12,560 --> 00:40:18,480
+is this is average uh accuracy on image
+
+871
+00:40:16,680 --> 00:40:22,400
+net which is the thing that they they
+
+872
+00:40:18,480 --> 00:40:25,160
+used in deciding which models to merge
+
+873
+00:40:22,400 --> 00:40:26,920
+in greedily and then this is on
+
+874
+00:40:25,160 --> 00:40:28,640
+distribution shifts so this is on other
+
+875
+00:40:26,920 --> 00:40:31,119
+data sets other than the ones they use
+
+876
+00:40:28,640 --> 00:40:33,040
+specifically for training and what you
+
+877
+00:40:31,119 --> 00:40:34,720
+can see is the greedy averaging method
+
+878
+00:40:33,040 --> 00:40:38,720
+does
+
+879
+00:40:34,720 --> 00:40:40,839
+better um than the best single model on
+
+880
+00:40:38,720 --> 00:40:42,319
+the data set that they used to decide
+
+881
+00:40:40,839 --> 00:40:44,800
+that greedy
+
+882
+00:40:42,319 --> 00:40:46,560
+average the uniform average actually
+
+883
+00:40:44,800 --> 00:40:48,359
+does worse than the best model so you
+
+884
+00:40:46,560 --> 00:40:50,960
+would actually be better off for image
+
+885
+00:40:48,359 --> 00:40:52,960
+net accuracy to just use the best model
+
+886
+00:40:50,960 --> 00:40:56,000
+but it's more robust so on the
+
+887
+00:40:52,960 --> 00:40:57,319
+distribution shift like data set it
+
+888
+00:40:56,000 --> 00:41:00,000
+actually does better than any of them
+
+889
+00:40:57,319 --> 00:41:02,280
+models so um you can see that there's
+
+890
+00:41:00,000 --> 00:41:04,720
+kind of trade-offs between choosing
+
+891
+00:41:02,280 --> 00:41:06,480
+those
+
+892
+00:41:04,720 --> 00:41:09,319
+essentially
+
+893
+00:41:06,480 --> 00:41:12,040
+um whoops that's a that's a typo that
+
+894
+00:41:09,319 --> 00:41:15,760
+should be ensembling but um they also
+
+895
+00:41:12,040 --> 00:41:18,440
+demonstrate that um averaging is
+
+896
+00:41:15,760 --> 00:41:22,720
+correlated with ensembling so this is
+
+897
+00:41:18,440 --> 00:41:25,200
+the um image accuracy of the parameter
+
+898
+00:41:22,720 --> 00:41:27,000
+average model this is image not accuracy
+
+899
+00:41:25,200 --> 00:41:30,200
+of the Ensemble so this is actually I
+
+900
+00:41:27,000 --> 00:41:33,720
+think really interesting figure um what
+
+901
+00:41:30,200 --> 00:41:36,440
+it shows is that there's a pretty strong
+
+902
+00:41:33,720 --> 00:41:38,760
+correlation between the two averaging is
+
+903
+00:41:36,440 --> 00:41:41,400
+almost never better than ensembling the
+
+904
+00:41:38,760 --> 00:41:44,800
+two together but it's faster of course
+
+905
+00:41:41,400 --> 00:41:48,119
+so it's better because it's faster and
+
+906
+00:41:44,800 --> 00:41:50,000
+there are situations where the Ensemble
+
+907
+00:41:48,119 --> 00:41:51,680
+is much better than the average model so
+
+908
+00:41:50,000 --> 00:41:55,720
+like the average model hurts the
+
+909
+00:41:51,680 --> 00:41:58,560
+averaging hurts um onbling does not hurt
+
+910
+00:41:55,720 --> 00:42:01,319
+so what this shows you is parameter
+
+911
+00:41:58,560 --> 00:42:03,119
+averaging is is safe and it nearly
+
+912
+00:42:01,319 --> 00:42:04,359
+approximates model on samping most of
+
+913
+00:42:03,119 --> 00:42:06,720
+the time but there are cases where it
+
+914
+00:42:04,359 --> 00:42:08,119
+doesn't so you do need to be a little
+
+915
+00:42:06,720 --> 00:42:11,720
+bit careful and it might hurt your
+
+916
+00:42:08,119 --> 00:42:11,720
+accuracy in some cases
+
+917
+00:42:16,680 --> 00:42:21,520
+yeah oh yeah sorry very good point yes
+
+918
+00:42:19,280 --> 00:42:21,520
+it's
+
+919
+00:42:22,319 --> 00:42:29,119
+paralel yeah
+
+920
+00:42:26,119 --> 00:42:29,119
+this
+
+921
+00:42:36,480 --> 00:42:41,520
+um how do you know
+
+922
+00:42:39,400 --> 00:42:45,720
+it's
+
+923
+00:42:41,520 --> 00:42:48,280
+particular yeah so notably all of these
+
+924
+00:42:45,720 --> 00:42:48,280
+are
+
+925
+00:42:48,800 --> 00:42:52,240
+initialized it's been a little while
+
+926
+00:42:50,800 --> 00:42:54,079
+since I read this but I know all of
+
+927
+00:42:52,240 --> 00:42:56,520
+these were initialized from a model that
+
+928
+00:42:54,079 --> 00:42:58,160
+was already pretty good on image that
+
+929
+00:42:56,520 --> 00:43:01,760
+and then they were tuned in different
+
+930
+00:42:58,160 --> 00:43:03,800
+ways I guess and so this I think this
+
+931
+00:43:01,760 --> 00:43:05,319
+might be initialized with a model that
+
+932
+00:43:03,800 --> 00:43:09,160
+was trained on a different data set or
+
+933
+00:43:05,319 --> 00:43:10,160
+something like that um and so they are
+
+934
+00:43:09,160 --> 00:43:12,480
+all starting from the same
+
+935
+00:43:10,160 --> 00:43:14,480
+initialization so parameter U
+
+936
+00:43:12,480 --> 00:43:16,599
+permutation inv variance is not an issue
+
+937
+00:43:14,480 --> 00:43:19,200
+there because they're starting from the
+
+938
+00:43:16,599 --> 00:43:23,480
+pre um but despite the fact that it's
+
+939
+00:43:19,200 --> 00:43:26,520
+not a problem there are there are cases
+
+940
+00:43:23,480 --> 00:43:29,119
+where like averaging is detrimental
+
+941
+00:43:26,520 --> 00:43:29,119
+compared to
+
+942
+00:43:32,839 --> 00:43:37,559
+um okay so
+
+943
+00:43:42,800 --> 00:43:45,800
+yeah
+
+944
+00:43:51,720 --> 00:43:54,720
+yep
+
+945
+00:43:56,040 --> 00:43:59,040
+y
+
+946
+00:44:07,079 --> 00:44:10,079
+okay
+
+947
+00:44:26,040 --> 00:44:29,040
+y
+
+948
+00:44:46,319 --> 00:44:52,520
+yeah so that's a great question um I'll
+
+949
+00:44:48,240 --> 00:44:54,920
+just repeat it which is um the these
+
+950
+00:44:52,520 --> 00:44:57,520
+experiments were done on CNN's or image
+
+951
+00:44:54,920 --> 00:44:59,280
+net like uh CNN based image that
+
+952
+00:44:57,520 --> 00:45:01,119
+classifiers is there something different
+
+953
+00:44:59,280 --> 00:45:04,040
+than Transformers particularly because
+
+954
+00:45:01,119 --> 00:45:06,240
+Transformer representations tend to be
+
+955
+00:45:04,040 --> 00:45:09,000
+uh like very concentrated in particular
+
+956
+00:45:06,240 --> 00:45:11,359
+parts of the space that's an excellent
+
+957
+00:45:09,000 --> 00:45:14,040
+question um what I do know is a lot of
+
+958
+00:45:11,359 --> 00:45:15,319
+people do merge together Transformer
+
+959
+00:45:14,040 --> 00:45:18,319
+models in fact if you look at the
+
+960
+00:45:15,319 --> 00:45:20,079
+hugging face leaderboard there's like
+
+961
+00:45:18,319 --> 00:45:22,240
+something and something merg together
+
+962
+00:45:20,079 --> 00:45:24,200
+like all over the leader board and it
+
+963
+00:45:22,240 --> 00:45:25,960
+does tend to improve accuracy so I I
+
+964
+00:45:24,200 --> 00:45:27,480
+know it is definitely effective for
+
+965
+00:45:25,960 --> 00:45:28,559
+Transformers
+
+966
+00:45:27,480 --> 00:45:32,040
+however Are
+
+967
+00:45:28,559 --> 00:45:34,640
+there specific model like parameter
+
+968
+00:45:32,040 --> 00:45:37,040
+averaging or model merging methods that
+
+969
+00:45:34,640 --> 00:45:38,599
+could improve accuracy by taking
+
+970
+00:45:37,040 --> 00:45:40,680
+advantage of the fact that Transformers
+
+971
+00:45:38,599 --> 00:45:42,480
+behaving a c certain way I think that's
+
+972
+00:45:40,680 --> 00:45:44,920
+totally possible and you know it would
+
+973
+00:45:42,480 --> 00:45:48,800
+be an interesting research Direction um
+
+974
+00:45:44,920 --> 00:45:51,680
+I'm not familiar enough with that
+
+975
+00:45:48,800 --> 00:45:53,359
+particular part myself to say oh I have
+
+976
+00:45:51,680 --> 00:45:55,160
+this great idea that you should work on
+
+977
+00:45:53,359 --> 00:45:55,920
+but I think if you're interested in it
+
+978
+00:45:55,160 --> 00:45:58,160
+you
+
+979
+00:45:55,920 --> 00:46:00,280
+definitely
+
+980
+00:45:58,160 --> 00:46:05,240
+cool anything
+
+981
+00:46:00,280 --> 00:46:08,920
+El okay so there's also the idea of uh
+
+982
+00:46:05,240 --> 00:46:12,440
+task vectors and um basically task
+
+983
+00:46:08,920 --> 00:46:15,280
+vectors here we are just merging
+
+984
+00:46:12,440 --> 00:46:17,280
+together two models by taking the
+
+985
+00:46:15,280 --> 00:46:18,280
+parameters of the models and averaging
+
+986
+00:46:17,280 --> 00:46:22,079
+them
+
+987
+00:46:18,280 --> 00:46:24,480
+together task vectors and other related
+
+988
+00:46:22,079 --> 00:46:26,040
+works specifically take advantage of the
+
+989
+00:46:24,480 --> 00:46:27,640
+fact that we're looking at different
+
+990
+00:46:26,040 --> 00:46:29,160
+fine-tuned models
+
+991
+00:46:27,640 --> 00:46:31,480
+and so these are models where we have a
+
+992
+00:46:29,160 --> 00:46:33,920
+base model and we know that uh that we
+
+993
+00:46:31,480 --> 00:46:35,760
+fine-tuned from this base model and the
+
+994
+00:46:33,920 --> 00:46:38,480
+basic idea is that we have our base
+
+995
+00:46:35,760 --> 00:46:40,319
+model here and the task Vector is the
+
+996
+00:46:38,480 --> 00:46:43,280
+difference between the base models
+
+997
+00:46:40,319 --> 00:46:45,559
+Vector uh parameters and the uh fine
+
+998
+00:46:43,280 --> 00:46:49,480
+tune models parameters so that's what
+
+999
+00:46:45,559 --> 00:46:52,720
+they Define as a task Vector um what
+
+1000
+00:46:49,480 --> 00:46:56,000
+does this allow us to do this allows us
+
+1001
+00:46:52,720 --> 00:46:58,040
+to do a number of interesting things um
+
+1002
+00:46:56,000 --> 00:47:02,359
+the first one
+
+1003
+00:46:58,040 --> 00:47:05,119
+is that we can actually subtract out uh
+
+1004
+00:47:02,359 --> 00:47:08,960
+quote unquote tasks that we don't want
+
+1005
+00:47:05,119 --> 00:47:11,559
+so like let's say we had a model that
+
+1006
+00:47:08,960 --> 00:47:13,440
+was trained on lots of toxic text or we
+
+1007
+00:47:11,559 --> 00:47:15,760
+had a model that was trained on lots of
+
+1008
+00:47:13,440 --> 00:47:18,760
+private text or something like that we
+
+1009
+00:47:15,760 --> 00:47:22,040
+could actually subtract out the task
+
+1010
+00:47:18,760 --> 00:47:24,240
+Vector from this and basically attempt
+
+1011
+00:47:22,040 --> 00:47:27,480
+to remove the model's ability to uh do
+
+1012
+00:47:24,240 --> 00:47:31,240
+that sort of things um you can also
+
+1013
+00:47:27,480 --> 00:47:36,040
+take two task vectors and combine them
+
+1014
+00:47:31,240 --> 00:47:39,280
+together and uh like get the model uh
+
+1015
+00:47:36,040 --> 00:47:42,200
+from the combination of the two um this
+
+1016
+00:47:39,280 --> 00:47:44,280
+isn't exactly the same as averaging the
+
+1017
+00:47:42,200 --> 00:47:45,440
+parameters because if you average the
+
+1018
+00:47:44,280 --> 00:47:47,400
+parameters you would probably get
+
+1019
+00:47:45,440 --> 00:47:49,160
+something in the middle right here but
+
+1020
+00:47:47,400 --> 00:47:50,440
+if you average the two vectors or add
+
+1021
+00:47:49,160 --> 00:47:52,040
+the two vectors together you would get
+
+1022
+00:47:50,440 --> 00:47:53,760
+something over here actually sorry if
+
+1023
+00:47:52,040 --> 00:47:56,520
+you average the vectors maybe it's the
+
+1024
+00:47:53,760 --> 00:47:58,119
+same so you could like add together the
+
+1025
+00:47:56,520 --> 00:47:59,480
+two vectors and and that would be
+
+1026
+00:47:58,119 --> 00:48:01,640
+something different than taking the
+
+1027
+00:47:59,480 --> 00:48:05,280
+average so it gives you a little bit
+
+1028
+00:48:01,640 --> 00:48:07,720
+more flexibility about things to do
+
+1029
+00:48:05,280 --> 00:48:09,599
+um and another thing this allows you to
+
+1030
+00:48:07,720 --> 00:48:12,920
+do is this allows you to try to resolve
+
+1031
+00:48:09,599 --> 00:48:15,400
+conflicts between um vectors of
+
+1032
+00:48:12,920 --> 00:48:19,720
+different tasks and so this is an
+
+1033
+00:48:15,400 --> 00:48:22,480
+illustration of of this method here
+
+1034
+00:48:19,720 --> 00:48:25,680
+and this has three tasks basically it
+
+1035
+00:48:22,480 --> 00:48:27,720
+has model one model two model three and
+
+1036
+00:48:25,680 --> 00:48:29,920
+each of them has vectors and you'll see
+
+1037
+00:48:27,720 --> 00:48:32,880
+that in some cases these vectors
+
+1038
+00:48:29,920 --> 00:48:34,599
+conflict so we have like pink going up
+
+1039
+00:48:32,880 --> 00:48:36,079
+we have yellow and purple going down we
+
+1040
+00:48:34,599 --> 00:48:37,800
+have yellow going up we have pink and
+
+1041
+00:48:36,079 --> 00:48:40,720
+purple going down etc
+
+1042
+00:48:37,800 --> 00:48:43,040
+etc and what this does is this
+
+1043
+00:48:40,720 --> 00:48:45,960
+identifies the vectors that are uh
+
+1044
+00:48:43,040 --> 00:48:48,040
+pointing the most strongly in particular
+
+1045
+00:48:45,960 --> 00:48:50,440
+directions and then it resolves
+
+1046
+00:48:48,040 --> 00:48:52,240
+conflicts between them and comes up with
+
+1047
+00:48:50,440 --> 00:48:54,559
+a vector that tries to move in a
+
+1048
+00:48:52,240 --> 00:48:55,920
+direction that improves all of the tasks
+
+1049
+00:48:54,559 --> 00:48:59,319
+at the same time and they demonstrate
+
+1050
+00:48:55,920 --> 00:49:01,480
+that this is better method for um kind
+
+1051
+00:48:59,319 --> 00:49:04,599
+of improving the ability to do all of
+
+1052
+00:49:01,480 --> 00:49:09,599
+the tasks compared to just averaging
+
+1053
+00:49:04,599 --> 00:49:09,599
+things together so yeah first
+
+1054
+00:49:11,920 --> 00:49:15,559
+exle like it just
+
+1055
+00:49:16,880 --> 00:49:23,640
+add yeah so this is
+
+1056
+00:49:20,680 --> 00:49:25,760
+um yeah you could move it more in that
+
+1057
+00:49:23,640 --> 00:49:27,319
+direction it there's obviously no
+
+1058
+00:49:25,760 --> 00:49:29,720
+guarantee that it would make it better
+
+1059
+00:49:27,319 --> 00:49:32,319
+but it might make it more extreme at
+
+1060
+00:49:29,720 --> 00:49:35,760
+least so uh
+
+1061
+00:49:32,319 --> 00:49:35,760
+yeah any other
+
+1062
+00:49:36,680 --> 00:49:39,960
+questions all
+
+1063
+00:49:55,640 --> 00:49:58,640
+yes
+
+1064
+00:50:25,640 --> 00:50:28,640
+one
+
+1065
+00:50:32,319 --> 00:50:37,240
+yeah yeah so this is a a great question
+
+1066
+00:50:35,599 --> 00:50:38,760
+um I can explain a little bit I'm not
+
+1067
+00:50:37,240 --> 00:50:40,760
+going to talk about Metal learning
+
+1068
+00:50:38,760 --> 00:50:42,680
+extensively in this class but just to
+
+1069
+00:50:40,760 --> 00:50:46,040
+give a very quick primer for people who
+
+1070
+00:50:42,680 --> 00:50:46,040
+don't know about it
+
+1071
+00:50:55,640 --> 00:50:58,640
+um
+
+1072
+00:51:00,359 --> 00:51:06,040
+this is an example of a paper on metal
+
+1073
+00:51:03,319 --> 00:51:09,559
+learning for low resource machine
+
+1074
+00:51:06,040 --> 00:51:12,680
+translation um I you can take a look at
+
+1075
+00:51:09,559 --> 00:51:16,200
+this paper um or not take a look at this
+
+1076
+00:51:12,680 --> 00:51:17,760
+paper um uh but the reason why I wanted
+
+1077
+00:51:16,200 --> 00:51:20,799
+to look at this paper is because it has
+
+1078
+00:51:17,760 --> 00:51:25,160
+a good um uh it has a good illustration
+
+1079
+00:51:20,799 --> 00:51:27,200
+of what metal learning is and basically
+
+1080
+00:51:25,160 --> 00:51:29,160
+um if we
+
+1081
+00:51:27,200 --> 00:51:33,839
+are doing transfer learning from a
+
+1082
+00:51:29,160 --> 00:51:35,880
+single task what we do is we have like a
+
+1083
+00:51:33,839 --> 00:51:37,960
+Spanish English machine translation
+
+1084
+00:51:35,880 --> 00:51:41,839
+system and then we fine-tune it to try
+
+1085
+00:51:37,960 --> 00:51:45,280
+to hit like to try to be a good Romanian
+
+1086
+00:51:41,839 --> 00:51:48,680
+uh English or latan English system if
+
+1087
+00:51:45,280 --> 00:51:50,400
+we're doing multitask learning um or
+
+1088
+00:51:48,680 --> 00:51:53,079
+which also could be equivalent to like
+
+1089
+00:51:50,400 --> 00:51:55,680
+instruction tuning for example we have
+
+1090
+00:51:53,079 --> 00:51:57,680
+uh French uh Spanish and Portuguese we
+
+1091
+00:51:55,680 --> 00:52:03,319
+train on all the then we
+
+1092
+00:51:57,680 --> 00:52:06,520
+fine-tune to uh to be a good Romanian uh
+
+1093
+00:52:03,319 --> 00:52:09,240
+translator latan trans uh
+
+1094
+00:52:06,520 --> 00:52:10,760
+translator whereas metal learning what
+
+1095
+00:52:09,240 --> 00:52:12,119
+it's trying to do is it's trying to
+
+1096
+00:52:10,760 --> 00:52:14,680
+learn a good
+
+1097
+00:52:12,119 --> 00:52:17,480
+initialization that makes it easy to
+
+1098
+00:52:14,680 --> 00:52:21,280
+fine-tune to try to come up with a model
+
+1099
+00:52:17,480 --> 00:52:23,839
+that is good uh for fine-tuning into new
+
+1100
+00:52:21,280 --> 00:52:29,040
+tasks
+
+1101
+00:52:23,839 --> 00:52:32,200
+um the way you do this is basically um
+
+1102
+00:52:29,040 --> 00:52:36,599
+you have two
+
+1103
+00:52:32,200 --> 00:52:39,400
+steps um of gradient descent and so you
+
+1104
+00:52:36,599 --> 00:52:42,400
+have a first step where you uh train the
+
+1105
+00:52:39,400 --> 00:52:42,400
+model
+
+1106
+00:52:42,599 --> 00:52:50,160
+um where you have an update on like data
+
+1107
+00:52:47,119 --> 00:52:50,160
+from French for
+
+1108
+00:52:55,440 --> 00:53:02,400
+example
+
+1109
+00:52:57,920 --> 00:53:02,400
+and then you have another
+
+1110
+00:53:04,640 --> 00:53:10,599
+update um where you train on like black
+
+1111
+00:53:07,880 --> 00:53:10,599
+or something like
+
+1112
+00:53:12,559 --> 00:53:17,040
+this and this is a very informal very
+
+1113
+00:53:15,599 --> 00:53:18,200
+informal description there's a lot of
+
+1114
+00:53:17,040 --> 00:53:19,599
+stuff we could talk about here I could
+
+1115
+00:53:18,200 --> 00:53:22,119
+have a whole class on this but we're not
+
+1116
+00:53:19,599 --> 00:53:27,200
+going to um I don't have one planned at
+
+1117
+00:53:22,119 --> 00:53:28,559
+the moment um and so you uh you up once
+
+1118
+00:53:27,200 --> 00:53:30,319
+and then you update again and you
+
+1119
+00:53:28,559 --> 00:53:33,400
+differentiate through this update
+
+1120
+00:53:30,319 --> 00:53:35,160
+process uh so that this becomes like
+
+1121
+00:53:33,400 --> 00:53:37,440
+essentially a good initialization for
+
+1122
+00:53:35,160 --> 00:53:40,640
+training on other languages or for other
+
+1123
+00:53:37,440 --> 00:53:43,000
+tasks or things like that
+
+1124
+00:53:40,640 --> 00:53:44,920
+um now going back to the original
+
+1125
+00:53:43,000 --> 00:53:46,240
+question the original question is is
+
+1126
+00:53:44,920 --> 00:53:50,000
+there a connection between metal
+
+1127
+00:53:46,240 --> 00:53:50,000
+learning in these uh task
+
+1128
+00:53:54,720 --> 00:53:58,440
+vectors I'm not
+
+1129
+00:53:59,079 --> 00:54:03,720
+100% sure about that because I think
+
+1130
+00:54:01,760 --> 00:54:06,599
+these test backs are generally created
+
+1131
+00:54:03,720 --> 00:54:08,480
+post Haw and so they're not like there's
+
+1132
+00:54:06,599 --> 00:54:12,680
+no explicit learning step to try to make
+
+1133
+00:54:08,480 --> 00:54:14,440
+them uh you know generalize well um one
+
+1134
+00:54:12,680 --> 00:54:15,960
+one thing that maybe might be
+
+1135
+00:54:14,440 --> 00:54:18,559
+interesting to people this is a paper
+
+1136
+00:54:15,960 --> 00:54:23,040
+that we like literally just put on
+
+1137
+00:54:18,559 --> 00:54:23,040
+archive about last week
+
+1138
+00:54:25,359 --> 00:54:28,359
+um
+
+1139
+00:54:34,520 --> 00:54:39,880
+and we didn't actually use metal
+
+1140
+00:54:36,400 --> 00:54:41,960
+learning in this uh in this paper um
+
+1141
+00:54:39,880 --> 00:54:44,520
+just because metal learning actually is
+
+1142
+00:54:41,960 --> 00:54:46,160
+hard to implement uh because you need to
+
+1143
+00:54:44,520 --> 00:54:48,680
+do this kind of double differentiation
+
+1144
+00:54:46,160 --> 00:54:50,720
+and can become very very expensive for
+
+1145
+00:54:48,680 --> 00:54:52,839
+large models but we did something a
+
+1146
+00:54:50,720 --> 00:54:55,920
+little bit motivated by
+
+1147
+00:54:52,839 --> 00:54:59,680
+um uh by metal learning and what we did
+
+1148
+00:54:55,920 --> 00:55:01,280
+is we took a pre-trained LM and normally
+
+1149
+00:54:59,680 --> 00:55:04,359
+what you do is something like continued
+
+1150
+00:55:01,280 --> 00:55:06,799
+pre-training on new documents to learn
+
+1151
+00:55:04,359 --> 00:55:10,160
+knowledge from the new documents or
+
+1152
+00:55:06,799 --> 00:55:12,200
+maybe um instruction tuning including
+
+1153
+00:55:10,160 --> 00:55:15,960
+instruction tuning on data on documents
+
+1154
+00:55:12,200 --> 00:55:17,520
+about the kind of uh data that you would
+
+1155
+00:55:15,960 --> 00:55:18,880
+want to be answering questions about so
+
+1156
+00:55:17,520 --> 00:55:20,640
+like let's say you're trying to train a
+
+1157
+00:55:18,880 --> 00:55:23,000
+medical language model you might train
+
+1158
+00:55:20,640 --> 00:55:26,680
+on lots of medical documents but what we
+
+1159
+00:55:23,000 --> 00:55:29,839
+did here is we had a step where we train
+
+1160
+00:55:26,680 --> 00:55:33,720
+in advance to
+
+1161
+00:55:29,839 --> 00:55:38,079
+get on question answer Pairs and
+
+1162
+00:55:33,720 --> 00:55:40,400
+documents from another domain and then
+
+1163
+00:55:38,079 --> 00:55:43,359
+we have a step after that where we train
+
+1164
+00:55:40,400 --> 00:55:46,400
+on documents from the domain we want to
+
+1165
+00:55:43,359 --> 00:55:48,400
+answer on so like we might train on
+
+1166
+00:55:46,400 --> 00:55:51,079
+Wikipedia question answer Pairs and
+
+1167
+00:55:48,400 --> 00:55:52,559
+Wikipedia documents and then in the
+
+1168
+00:55:51,079 --> 00:55:54,079
+second step we would train on medical
+
+1169
+00:55:52,559 --> 00:55:56,680
+documents and we demonstrate that
+
+1170
+00:55:54,079 --> 00:55:58,880
+basically this allows the model to do a
+
+1171
+00:55:56,680 --> 00:56:00,880
+better job of question answering over
+
+1172
+00:55:58,880 --> 00:56:03,640
+these uh documents that we find tune on
+
+1173
+00:56:00,880 --> 00:56:05,000
+over here and so kind of going back to
+
+1174
+00:56:03,640 --> 00:56:06,760
+the metal learning paper that I talked
+
+1175
+00:56:05,000 --> 00:56:08,359
+about before the metal learning paper
+
+1176
+00:56:06,760 --> 00:56:10,640
+tries to get the parameters in a good
+
+1177
+00:56:08,359 --> 00:56:12,559
+space so that after you find ton on
+
+1178
+00:56:10,640 --> 00:56:15,520
+another data set you do a good job of
+
+1179
+00:56:12,559 --> 00:56:17,799
+that in this paper our motivation is
+
+1180
+00:56:15,520 --> 00:56:20,359
+that the model kind of learns that when
+
+1181
+00:56:17,799 --> 00:56:22,039
+you train on documents you should be
+
+1182
+00:56:20,359 --> 00:56:24,079
+able to answer questions about those
+
+1183
+00:56:22,039 --> 00:56:25,480
+documents and so when you get a new set
+
+1184
+00:56:24,079 --> 00:56:27,200
+of documents it's kind of in a good part
+
+1185
+00:56:25,480 --> 00:56:31,079
+of the parameter space to make that easy
+
+1186
+00:56:27,200 --> 00:56:33,520
+to do so um if that if metal learning is
+
+1187
+00:56:31,079 --> 00:56:34,640
+interesting um there are tutorials on
+
+1188
+00:56:33,520 --> 00:56:37,119
+metal learning that I could probably
+
+1189
+00:56:34,640 --> 00:56:39,599
+share and then um if you're interested
+
+1190
+00:56:37,119 --> 00:56:42,599
+in kind of like learning Knowledge from
+
+1191
+00:56:39,599 --> 00:56:45,039
+uh learning Knowledge
+
+1192
+00:56:42,599 --> 00:56:46,079
+from continued pre-training or something
+
+1193
+00:56:45,039 --> 00:56:47,400
+like that you could take a look at this
+
+1194
+00:56:46,079 --> 00:56:49,920
+right there as
+
+1195
+00:56:47,400 --> 00:56:54,480
+well uh
+
+1196
+00:56:49,920 --> 00:56:54,480
+cool any questions about that
+
+1197
+00:56:55,240 --> 00:57:00,880
+or
+
+1198
+00:56:57,599 --> 00:57:02,480
+okay cool I I'll jump on this so anyway
+
+1199
+00:57:00,880 --> 00:57:05,520
+um I talked about several methods for
+
+1200
+00:57:02,480 --> 00:57:07,520
+merging models together um there's a
+
+1201
+00:57:05,520 --> 00:57:09,440
+popular toolkit called merge kit that
+
+1202
+00:57:07,520 --> 00:57:10,960
+makes it relatively easy to do this it
+
+1203
+00:57:09,440 --> 00:57:13,280
+implements a lot of the models that I
+
+1204
+00:57:10,960 --> 00:57:17,160
+talked about here including uh the
+
+1205
+00:57:13,280 --> 00:57:19,880
+linear methods um uh the task arithmetic
+
+1206
+00:57:17,160 --> 00:57:23,079
+method and ties uh so I talked about
+
+1207
+00:57:19,880 --> 00:57:25,480
+these there is kind of like a expansion
+
+1208
+00:57:23,079 --> 00:57:27,240
+on this so if you want to merge together
+
+1209
+00:57:25,480 --> 00:57:28,760
+models it's Rel easy to do from a
+
+1210
+00:57:27,240 --> 00:57:30,760
+software standpoint as so so you can
+
+1211
+00:57:28,760 --> 00:57:35,119
+take a look at
+
+1212
+00:57:30,760 --> 00:57:38,000
+that um another really simple thing uh
+
+1213
+00:57:35,119 --> 00:57:39,880
+is uh distilling ensembles and so we
+
+1214
+00:57:38,000 --> 00:57:43,039
+already talked about distillation the
+
+1215
+00:57:39,880 --> 00:57:45,599
+idea is simple um
+
+1216
+00:57:43,039 --> 00:57:47,680
+you so parameter averaging only really
+
+1217
+00:57:45,599 --> 00:57:49,200
+works for models within the same run uh
+
+1218
+00:57:47,680 --> 00:57:51,760
+same model architecture same
+
+1219
+00:57:49,200 --> 00:57:54,280
+initialization so knowledge distillation
+
+1220
+00:57:51,760 --> 00:57:55,559
+uh trains a model to copy The Ensemble
+
+1221
+00:57:54,280 --> 00:57:57,359
+and so it tries to match the
+
+1222
+00:57:55,559 --> 00:57:59,119
+distribution over the predicted words
+
+1223
+00:57:57,359 --> 00:58:00,760
+for an
+
+1224
+00:57:59,119 --> 00:58:05,319
+on
+
+1225
+00:58:00,760 --> 00:58:07,799
+um and so this allows the model to make
+
+1226
+00:58:05,319 --> 00:58:09,079
+the same you know good predictions as
+
+1227
+00:58:07,799 --> 00:58:11,079
+The Ensemble make the same bad
+
+1228
+00:58:09,079 --> 00:58:12,799
+predictions as Ensemble it just allows
+
+1229
+00:58:11,079 --> 00:58:14,799
+you to learn more efficiently just like
+
+1230
+00:58:12,799 --> 00:58:16,680
+distillation does in general and they
+
+1231
+00:58:14,799 --> 00:58:18,960
+actually model distillation the original
+
+1232
+00:58:16,680 --> 00:58:22,240
+motivation for it when Jeff Hinton
+
+1233
+00:58:18,960 --> 00:58:24,599
+proposed it in 2015 in in this paper was
+
+1234
+00:58:22,240 --> 00:58:25,680
+to copy an ensemble now we use it for a
+
+1235
+00:58:24,599 --> 00:58:27,039
+lot of other things like in the
+
+1236
+00:58:25,680 --> 00:58:31,160
+distillation
+
+1237
+00:58:27,039 --> 00:58:31,160
+like weed the class but was the
+
+1238
+00:58:34,119 --> 00:58:39,599
+original
+
+1239
+00:58:35,760 --> 00:58:42,640
+um next I'll move on to sparse mixture
+
+1240
+00:58:39,599 --> 00:58:44,960
+of experts models and this is really
+
+1241
+00:58:42,640 --> 00:58:47,599
+important uh this is used in a lot of
+
+1242
+00:58:44,960 --> 00:58:51,319
+modern models it's allegedly used in GPD
+
+1243
+00:58:47,599 --> 00:58:53,160
+4 um and it is uh definitely used in
+
+1244
+00:58:51,319 --> 00:58:55,280
+mixl uh which is kind of one of the
+
+1245
+00:58:53,160 --> 00:58:58,039
+state-ofthe-art open models so I think
+
+1246
+00:58:55,280 --> 00:58:58,039
+it's a good thing to know
+
+1247
+00:58:59,880 --> 00:59:05,720
+um what these do is they take advantage
+
+1248
+00:59:02,680 --> 00:59:08,160
+of sparse computation so if you think
+
+1249
+00:59:05,720 --> 00:59:09,359
+about what happens when you do a scalar
+
+1250
+00:59:08,160 --> 00:59:12,760
+tensor
+
+1251
+00:59:09,359 --> 00:59:14,720
+multiply where the scaler is zero and
+
+1252
+00:59:12,760 --> 00:59:17,160
+basically the result of the entire
+
+1253
+00:59:14,720 --> 00:59:19,680
+resulting tensor is guaranteed to be
+
+1254
+00:59:17,160 --> 00:59:21,440
+zero and so you don't even need to do
+
+1255
+00:59:19,680 --> 00:59:25,440
+the computation you don't need to even
+
+1256
+00:59:21,440 --> 00:59:27,520
+bother um and so this manifests itself
+
+1257
+00:59:25,440 --> 00:59:30,240
+in a bunch of different places in modern
+
+1258
+00:59:27,520 --> 00:59:35,000
+models um the first one could be single
+
+1259
+00:59:30,240 --> 00:59:38,400
+rows in a matrix multiply so um if you
+
+1260
+00:59:35,000 --> 00:59:40,480
+have a big Matrix multiply like
+
+1261
+00:59:38,400 --> 00:59:44,240
+this
+
+1262
+00:59:40,480 --> 00:59:47,880
+um or Matrix Vector multiply like this
+
+1263
+00:59:44,240 --> 00:59:50,200
+um and some of the rows are zero then uh
+
+1264
+00:59:47,880 --> 00:59:54,559
+that that's one place where it
+
+1265
+00:59:50,200 --> 00:59:58,200
+happens um you can also uh do this
+
+1266
+00:59:54,559 --> 01:00:00,119
+between zero and in not just rows but
+
+1267
+00:59:58,200 --> 01:00:02,200
+also larger
+
+1268
+01:00:00,119 --> 01:00:05,799
+tensors um and you can even do it in
+
+1269
+01:00:02,200 --> 01:00:07,599
+whole models in an ensemble so um the
+
+1270
+01:00:05,799 --> 01:00:10,799
+first one this can be optimized
+
+1271
+01:00:07,599 --> 01:00:13,880
+automatically by GPU um the second one
+
+1272
+01:00:10,799 --> 01:00:15,400
+this often occurs in uh sparse mixture
+
+1273
+01:00:13,880 --> 01:00:18,000
+of experts
+
+1274
+01:00:15,400 --> 01:00:19,400
+models and the final one uh basically
+
+1275
+01:00:18,000 --> 01:00:21,880
+you just don't need to even use the
+
+1276
+01:00:19,400 --> 01:00:24,119
+model in emble so if you somehow
+
+1277
+01:00:21,880 --> 01:00:25,640
+optimize an ensemble and it turns out
+
+1278
+01:00:24,119 --> 01:00:27,599
+that the probability of one of the
+
+1279
+01:00:25,640 --> 01:00:29,680
+models is zero you just can throw it out
+
+1280
+01:00:27,599 --> 01:00:33,640
+and not use it at
+
+1281
+01:00:29,680 --> 01:00:36,839
+all so um GPU level sparsity
+
+1282
+01:00:33,640 --> 01:00:39,839
+support uh Nvidia gpus support a bunch
+
+1283
+01:00:36,839 --> 01:00:42,559
+of different types of sparsity and uh
+
+1284
+01:00:39,839 --> 01:00:44,599
+the people the wonderful people at
+
+1285
+01:00:42,559 --> 01:00:48,280
+Nvidia have worked hard to make the
+
+1286
+01:00:44,599 --> 01:00:51,319
+support uh work to some extent anyway
+
+1287
+01:00:48,280 --> 01:00:53,119
+and uh there's a library called cpar and
+
+1288
+01:00:51,319 --> 01:00:56,119
+this is used in pytorch and all these
+
+1289
+01:00:53,119 --> 01:00:58,280
+other things as well and just to give
+
+1290
+01:00:56,119 --> 01:01:01,240
+example a vector Matrix multiply with a
+
+1291
+01:00:58,280 --> 01:01:03,240
+sparse Vector um such as one that comes
+
+1292
+01:01:01,240 --> 01:01:06,160
+from a relu activation basically what
+
+1293
+01:01:03,240 --> 01:01:09,319
+happens is let's say you only have three
+
+1294
+01:01:06,160 --> 01:01:11,799
+uh parts of this Vector that are active
+
+1295
+01:01:09,319 --> 01:01:15,240
+um you actually just don't need to cop
+
+1296
+01:01:11,799 --> 01:01:18,200
+uh calculate any of the columns here so
+
+1297
+01:01:15,240 --> 01:01:19,720
+that makes your life relatively
+
+1298
+01:01:18,200 --> 01:01:22,880
+easy
+
+1299
+01:01:19,720 --> 01:01:24,480
+um but the specific thing that I wanted
+
+1300
+01:01:22,880 --> 01:01:26,640
+to talk about is a sparsely gated
+
+1301
+01:01:24,480 --> 01:01:29,799
+mixture of experts layer because this is
+
+1302
+01:01:26,640 --> 01:01:33,960
+uh what is used in mixol and probably uh
+
+1303
+01:01:29,799 --> 01:01:38,200
+the GPT models as well and what you do
+
+1304
+01:01:33,960 --> 01:01:41,760
+is you have a feed forward Network and
+
+1305
+01:01:38,200 --> 01:01:41,760
+normally a feed forward Network in a
+
+1306
+01:01:43,640 --> 01:01:52,119
+Transformer is this like really wide
+
+1307
+01:01:49,319 --> 01:01:57,240
+thing this huge wide feed forward
+
+1308
+01:01:52,119 --> 01:01:59,359
+Network um that you use to extract a
+
+1309
+01:01:57,240 --> 01:02:00,520
+whole bunch of features at each layer
+
+1310
+01:01:59,359 --> 01:02:02,640
+and that's where a lot of the
+
+1311
+01:02:00,520 --> 01:02:05,799
+computation and Transformer
+
+1312
+01:02:02,640 --> 01:02:10,079
+happens um and what sparsely gated
+
+1313
+01:02:05,799 --> 01:02:13,079
+mixture of uh experts layers do is they
+
+1314
+01:02:10,079 --> 01:02:15,640
+first have this gating Network here
+
+1315
+01:02:13,079 --> 01:02:17,880
+where it calculates uh mixture
+
+1316
+01:02:15,640 --> 01:02:21,119
+probability but the mixture probability
+
+1317
+01:02:17,880 --> 01:02:23,039
+is zero and for many or most of the
+
+1318
+01:02:21,119 --> 01:02:26,880
+parts of this feed forward
+
+1319
+01:02:23,039 --> 01:02:28,760
+Network and so for the ones where it's
+
+1320
+01:02:26,880 --> 01:02:31,319
+zero you just don't calculate
+
+1321
+01:02:28,760 --> 01:02:34,319
+it um and then when you mix them
+
+1322
+01:02:31,319 --> 01:02:37,359
+together you use the mixture rates and
+
+1323
+01:02:34,319 --> 01:02:39,520
+this is actually really simple um it's
+
+1324
+01:02:37,359 --> 01:02:42,400
+like several lines of pytorch code maybe
+
+1325
+01:02:39,520 --> 01:02:45,319
+like seven or eight lines of P torch
+
+1326
+01:02:42,400 --> 01:02:48,720
+code but the basic uh idea here is you
+
+1327
+01:02:45,319 --> 01:02:50,599
+have um this gating function where you
+
+1328
+01:02:48,720 --> 01:02:52,799
+calculate the gating function based on
+
+1329
+01:02:50,599 --> 01:02:53,640
+the input and then you have this keep
+
+1330
+01:02:52,799 --> 01:02:56,720
+top
+
+1331
+01:02:53,640 --> 01:02:58,319
+K uh operation and then you take the
+
+1332
+01:02:56,720 --> 01:03:02,559
+soft Max over
+
+1333
+01:02:58,319 --> 01:03:04,359
+this and the keep top K operation is if
+
+1334
+01:03:02,559 --> 01:03:06,160
+the value is within the top K you just
+
+1335
+01:03:04,359 --> 01:03:07,319
+keep it and if it's not in the top K you
+
+1336
+01:03:06,160 --> 01:03:11,960
+don't keep
+
+1337
+01:03:07,319 --> 01:03:13,119
+it so that that's all basically but what
+
+1338
+01:03:11,960 --> 01:03:14,760
+what's great about this is then you
+
+1339
+01:03:13,119 --> 01:03:17,799
+don't have to calculate like many of
+
+1340
+01:03:14,760 --> 01:03:20,119
+them and so for example um uh if you
+
+1341
+01:03:17,799 --> 01:03:22,640
+keep the top two out of eight you reduce
+
+1342
+01:03:20,119 --> 01:03:26,760
+your calcul uh your computation by four
+
+1343
+01:03:22,640 --> 01:03:30,000
+times for this part so
+
+1344
+01:03:26,760 --> 01:03:33,000
+um any any questions
+
+1345
+01:03:30,000 --> 01:03:33,000
+here
+
+1346
+01:03:54,720 --> 01:03:57,720
+yeah
+
+1347
+01:04:03,160 --> 01:04:07,039
+um sorry what what exactly do you mean
+
+1348
+01:04:05,559 --> 01:04:09,400
+by easy to paralyze are you talking
+
+1349
+01:04:07,039 --> 01:04:12,400
+about like a GPU can calculate lots of
+
+1350
+01:04:09,400 --> 01:04:15,680
+things at the same time yeah so I think
+
+1351
+01:04:12,400 --> 01:04:17,720
+if you have a very small model um you're
+
+1352
+01:04:15,680 --> 01:04:21,680
+actually not going to get as much from
+
+1353
+01:04:17,720 --> 01:04:25,079
+this uh because you're not you're
+
+1354
+01:04:21,680 --> 01:04:26,359
+essentially not bound by computation uh
+
+1355
+01:04:25,079 --> 01:04:27,880
+like you're bound more by memory
+
+1356
+01:04:26,359 --> 01:04:29,079
+movement and the GPU and other stuff
+
+1357
+01:04:27,880 --> 01:04:30,520
+like that but once you start getting up
+
+1358
+01:04:29,079 --> 01:04:32,920
+to the bigger models you actually are
+
+1359
+01:04:30,520 --> 01:04:34,640
+bound by computation so reducing your
+
+1360
+01:04:32,920 --> 01:04:37,039
+computation by four actually is a big
+
+1361
+01:04:34,640 --> 01:04:42,559
+one so it's a really really good
+
+1362
+01:04:37,039 --> 01:04:42,559
+question um any any other questions
+
+1363
+01:04:44,039 --> 01:04:50,520
+yeah so so this will
+
+1364
+01:04:48,240 --> 01:04:53,160
+um probably
+
+1365
+01:04:50,520 --> 01:04:56,039
+be
+
+1366
+01:04:53,160 --> 01:04:59,279
+just oh sorry I I don't have this here
+
+1367
+01:04:56,039 --> 01:05:01,760
+but this will be a often a linear layer
+
+1368
+01:04:59,279 --> 01:05:01,760
+followed by a
+
+1369
+01:05:03,039 --> 01:05:08,000
+seance um or or actually no it doesn't
+
+1370
+01:05:06,359 --> 01:05:10,520
+even need to be followed by softb it
+
+1371
+01:05:08,000 --> 01:05:10,520
+could just be a
+
+1372
+01:05:12,520 --> 01:05:17,920
+linear and I think actually I didn't put
+
+1373
+01:05:14,960 --> 01:05:19,680
+it on this slide but I have the in the
+
+1374
+01:05:17,920 --> 01:05:21,359
+references on the website I have the
+
+1375
+01:05:19,680 --> 01:05:22,760
+actual implementation in mix roll you
+
+1376
+01:05:21,359 --> 01:05:25,279
+can go in and look at it it's really
+
+1377
+01:05:22,760 --> 01:05:27,160
+simple um one thing I didn't put on here
+
+1378
+01:05:25,279 --> 01:05:31,000
+um which actually uh relates to the
+
+1379
+01:05:27,160 --> 01:05:32,920
+question before is Hardware wise this
+
+1380
+01:05:31,000 --> 01:05:34,799
+implementation is tricky if you do
+
+1381
+01:05:32,920 --> 01:05:37,599
+batching um and the reason why It's
+
+1382
+01:05:34,799 --> 01:05:39,480
+Tricky if you do batching is because um
+
+1383
+01:05:37,599 --> 01:05:43,000
+different experts will be active for
+
+1384
+01:05:39,480 --> 01:05:45,240
+different like parts of the batch so if
+
+1385
+01:05:43,000 --> 01:05:48,559
+you do that you need to do some tricky
+
+1386
+01:05:45,240 --> 01:05:48,559
+stuff uh there's
+
+1387
+01:05:54,640 --> 01:05:57,640
+this
+
+1388
+01:06:03,240 --> 01:06:12,039
+like so much of AI research nowadays uh
+
+1389
+01:06:08,200 --> 01:06:12,039
+the best resource for this is social
+
+1390
+01:06:13,680 --> 01:06:20,000
+media so this is uh there's a kind of
+
+1391
+01:06:16,880 --> 01:06:23,240
+interesting discussion of
+
+1392
+01:06:20,000 --> 01:06:25,359
+this um if you search for like gpk Fast
+
+1393
+01:06:23,240 --> 01:06:28,400
+mixed r on Twitter it it talks about
+
+1394
+01:06:25,359 --> 01:06:30,200
+this but basically there's a bunch of uh
+
+1395
+01:06:28,400 --> 01:06:32,680
+little little things you need to pay
+
+1396
+01:06:30,200 --> 01:06:34,760
+attention to um and ways that you can do
+
+1397
+01:06:32,680 --> 01:06:36,960
+tricks to make this work fast on GPU
+
+1398
+01:06:34,760 --> 01:06:40,000
+which also kind of uh addresses the
+
+1399
+01:06:36,960 --> 01:06:42,359
+concern so you can look for Horus H's
+
+1400
+01:06:40,000 --> 01:06:44,200
+discussion
+
+1401
+01:06:42,359 --> 01:06:46,680
+this
+
+1402
+01:06:44,200 --> 01:06:49,000
+cool
+
+1403
+01:06:46,680 --> 01:06:50,799
+um so the final thing I'd like to talk
+
+1404
+01:06:49,000 --> 01:06:52,480
+about in the last 10 minutes is pipeline
+
+1405
+01:06:50,799 --> 01:06:55,359
+systems
+
+1406
+01:06:52,480 --> 01:06:57,039
+um and pipeline systems are systems
+
+1407
+01:06:55,359 --> 01:07:00,279
+where we
+
+1408
+01:06:57,039 --> 01:07:02,319
+have models that basically the output of
+
+1409
+01:07:00,279 --> 01:07:05,319
+one model becomes the input of another
+
+1410
+01:07:02,319 --> 01:07:05,319
+model
+
+1411
+01:07:05,599 --> 01:07:10,359
+and to give an example of this a
+
+1412
+01:07:08,200 --> 01:07:13,480
+cascaded system is basically a system
+
+1413
+01:07:10,359 --> 01:07:15,119
+like this where you uh take the output
+
+1414
+01:07:13,480 --> 01:07:16,960
+of one system and then you feed it into
+
+1415
+01:07:15,119 --> 01:07:19,640
+the input of another system so a very
+
+1416
+01:07:16,960 --> 01:07:22,880
+stereotypical example of This is speech
+
+1417
+01:07:19,640 --> 01:07:25,559
+translation um where you run speech and
+
+1418
+01:07:22,880 --> 01:07:27,720
+then you uh do speech recognition into
+
+1419
+01:07:25,559 --> 01:07:29,319
+text and then text you do machine
+
+1420
+01:07:27,720 --> 01:07:32,160
+translation into another
+
+1421
+01:07:29,319 --> 01:07:33,920
+language
+
+1422
+01:07:32,160 --> 01:07:36,440
+and
+
+1423
+01:07:33,920 --> 01:07:39,039
+um one of the frustrating things about
+
+1424
+01:07:36,440 --> 01:07:43,000
+speech translation is these systems are
+
+1425
+01:07:39,039 --> 01:07:45,799
+stubbornly better uh for a long time
+
+1426
+01:07:43,000 --> 01:07:47,680
+than many systems that try to do end to
+
+1427
+01:07:45,799 --> 01:07:49,960
+end like speech to text in another
+
+1428
+01:07:47,680 --> 01:07:52,160
+language there's a couple reasons for
+
+1429
+01:07:49,960 --> 01:07:54,440
+this does anyone have an idea why what
+
+1430
+01:07:52,160 --> 01:07:57,039
+one of those reasons might
+
+1431
+01:07:54,440 --> 01:07:58,839
+be
+
+1432
+01:07:57,039 --> 01:08:01,559
+yeah the
+
+1433
+01:07:58,839 --> 01:08:05,279
+data
+
+1434
+01:08:01,559 --> 01:08:08,680
+anying exactly so data data availability
+
+1435
+01:08:05,279 --> 01:08:10,920
+is way better for speech to text in the
+
+1436
+01:08:08,680 --> 01:08:13,319
+same language and text to text in
+
+1437
+01:08:10,920 --> 01:08:15,720
+another language than it is for uh
+
+1438
+01:08:13,319 --> 01:08:17,759
+Speech to te text in another language
+
+1439
+01:08:15,720 --> 01:08:19,319
+because there just aren't large data
+
+1440
+01:08:17,759 --> 01:08:21,679
+sets that have speech and text in many
+
+1441
+01:08:19,319 --> 01:08:25,719
+languages so there's a bunch of tricks
+
+1442
+01:08:21,679 --> 01:08:31,759
+that you can do uh to you know fix this
+
+1443
+01:08:25,719 --> 01:08:34,239
+but still it it's uh you know uh tricky
+
+1444
+01:08:31,759 --> 01:08:36,120
+and there's a couple other reasons
+
+1445
+01:08:34,239 --> 01:08:38,159
+another reason is like actually speech
+
+1446
+01:08:36,120 --> 01:08:39,319
+to text in the same language is just a
+
+1447
+01:08:38,159 --> 01:08:42,520
+much more
+
+1448
+01:08:39,319 --> 01:08:45,359
+straightforward task um and so it's a
+
+1449
+01:08:42,520 --> 01:08:47,839
+bit easier to learn another thing is
+
+1450
+01:08:45,359 --> 01:08:50,839
+interpretability and the reason why
+
+1451
+01:08:47,839 --> 01:08:52,120
+interpretability is important is
+
+1452
+01:08:50,839 --> 01:08:54,920
+basically
+
+1453
+01:08:52,120 --> 01:08:56,640
+like if I'm talking to you in a
+
+1454
+01:08:54,920 --> 01:08:58,000
+different language like you speak a
+
+1455
+01:08:56,640 --> 01:09:00,319
+different language I'm talking to you
+
+1456
+01:08:58,000 --> 01:09:02,679
+through a speech translation system I
+
+1457
+01:09:00,319 --> 01:09:05,799
+actually want to know if the speech
+
+1458
+01:09:02,679 --> 01:09:07,600
+recognition worked because I know if the
+
+1459
+01:09:05,799 --> 01:09:08,920
+speech recognition didn't work then I'll
+
+1460
+01:09:07,600 --> 01:09:10,440
+I'm pretty sure that the translation
+
+1461
+01:09:08,920 --> 01:09:11,920
+didn't work either right and I can
+
+1462
+01:09:10,440 --> 01:09:14,880
+verify the speech recognition but I
+
+1463
+01:09:11,920 --> 01:09:16,199
+can't verify the transation so um
+
+1464
+01:09:14,880 --> 01:09:18,279
+there's other reasons why you might want
+
+1465
+01:09:16,199 --> 01:09:20,239
+a Cascade system other than just like
+
+1466
+01:09:18,279 --> 01:09:22,440
+accuracy or or other things like that
+
+1467
+01:09:20,239 --> 01:09:25,880
+but this is a thing we definitely
+
+1468
+01:09:22,440 --> 01:09:29,120
+do um there's another idea of stacking
+
+1469
+01:09:25,880 --> 01:09:32,560
+and stacking is um very similar to cast
+
+1470
+01:09:29,120 --> 01:09:34,560
+skating but it allows you to take two
+
+1471
+01:09:32,560 --> 01:09:37,120
+different models for the same task but
+
+1472
+01:09:34,560 --> 01:09:39,400
+with predictions in different ways so
+
+1473
+01:09:37,120 --> 01:09:41,120
+just taking another um
+
+1474
+01:09:39,400 --> 01:09:43,600
+example
+
+1475
+01:09:41,120 --> 01:09:45,040
+uh actually maybe maybe ignore the
+
+1476
+01:09:43,600 --> 01:09:47,159
+example I have here but we could just
+
+1477
+01:09:45,040 --> 01:09:50,679
+take the example of speech uh
+
+1478
+01:09:47,159 --> 01:09:53,000
+translation um the speech translation
+
+1479
+01:09:50,679 --> 01:09:55,760
+model uh we would first do speech
+
+1480
+01:09:53,000 --> 01:09:57,520
+recognition into like let's say English
+
+1481
+01:09:55,760 --> 01:09:59,640
+and then we would do translation and the
+
+1482
+01:09:57,520 --> 01:10:03,840
+input to the translation model would be
+
+1483
+01:09:59,640 --> 01:10:05,560
+speech in English um text in English and
+
+1484
+01:10:03,840 --> 01:10:07,320
+we would generate the output in Japanese
+
+1485
+01:10:05,560 --> 01:10:10,080
+so it would take both the speech and the
+
+1486
+01:10:07,320 --> 01:10:12,920
+text uh when it was doing translation
+
+1487
+01:10:10,080 --> 01:10:14,840
+and that would allow it to number one
+
+1488
+01:10:12,920 --> 01:10:17,719
+basically get a second opinion about
+
+1489
+01:10:14,840 --> 01:10:21,080
+whether the transcription was correct
+
+1490
+01:10:17,719 --> 01:10:23,800
+but also like let's say there was
+
+1491
+01:10:21,080 --> 01:10:26,440
+some unique information that only
+
+1492
+01:10:23,800 --> 01:10:29,480
+appeared in the
+
+1493
+01:10:26,440 --> 01:10:31,679
+um uh that only appeared in the speech
+
+1494
+01:10:29,480 --> 01:10:34,840
+so just to give an example I read the
+
+1495
+01:10:31,679 --> 01:10:37,040
+book I read the book are both
+
+1496
+01:10:34,840 --> 01:10:38,640
+transcribed exactly the same way and
+
+1497
+01:10:37,040 --> 01:10:41,679
+they're different translations obviously
+
+1498
+01:10:38,640 --> 01:10:42,920
+because one is uh you know present or
+
+1499
+01:10:41,679 --> 01:10:45,560
+present tense and the other is past
+
+1500
+01:10:42,920 --> 01:10:47,239
+tense so there are examples where uh
+
+1501
+01:10:45,560 --> 01:10:51,600
+adding a cascaded system would lose
+
+1502
+01:10:47,239 --> 01:10:51,600
+information and a stacked system would
+
+1503
+01:10:53,400 --> 01:10:57,679
+not another thing is of refinement I
+
+1504
+01:10:56,440 --> 01:10:59,480
+think this is actually really
+
+1505
+01:10:57,679 --> 01:11:01,000
+interesting because large language
+
+1506
+01:10:59,480 --> 01:11:03,920
+models have opened up a whole bunch of
+
+1507
+01:11:01,000 --> 01:11:05,640
+possibilities for us in this space um
+
+1508
+01:11:03,920 --> 01:11:07,760
+this is like cascading and stacking but
+
+1509
+01:11:05,640 --> 01:11:09,640
+it it can be done multiple times and it
+
+1510
+01:11:07,760 --> 01:11:12,960
+can be done multiple times with the same
+
+1511
+01:11:09,640 --> 01:11:15,040
+model so um we have an input we feed it
+
+1512
+01:11:12,960 --> 01:11:17,320
+into the model we get an output and then
+
+1513
+01:11:15,040 --> 01:11:19,360
+we feed the output back in and gradually
+
+1514
+01:11:17,320 --> 01:11:23,080
+refine it and make it better and
+
+1515
+01:11:19,360 --> 01:11:24,760
+better and the first time this was done
+
+1516
+01:11:23,080 --> 01:11:27,440
+in neural networks was through something
+
+1517
+01:11:24,760 --> 01:11:29,679
+called Del ation networks and basically
+
+1518
+01:11:27,440 --> 01:11:32,360
+deliberation networks what they do is
+
+1519
+01:11:29,679 --> 01:11:33,760
+they uh take in an output and then they
+
+1520
+01:11:32,360 --> 01:11:34,920
+just gradually refine it to make it
+
+1521
+01:11:33,760 --> 01:11:37,280
+better and better they used a
+
+1522
+01:11:34,920 --> 01:11:39,159
+reinforcement learning algorithm to do
+
+1523
+01:11:37,280 --> 01:11:41,159
+this where you generated the output and
+
+1524
+01:11:39,159 --> 01:11:43,600
+then um improved
+
+1525
+01:11:41,159 --> 01:11:46,719
+it another thing that's really popular
+
+1526
+01:11:43,600 --> 01:11:48,280
+nowadays is uh diffusion models and I
+
+1527
+01:11:46,719 --> 01:11:50,400
+haven't quite decided whether I'll have
+
+1528
+01:11:48,280 --> 01:11:51,880
+time to cover diffusion models in depth
+
+1529
+01:11:50,400 --> 01:11:54,880
+but basically the way a diffusion model
+
+1530
+01:11:51,880 --> 01:11:55,880
+works is very similar you start out with
+
+1531
+01:11:54,880 --> 01:11:57,239
+nothing
+
+1532
+01:11:55,880 --> 01:11:59,840
+and then you gradually make it better
+
+1533
+01:11:57,239 --> 01:12:01,360
+and better um the key difference between
+
+1534
+01:11:59,840 --> 01:12:03,520
+deliberation networks and diffusion
+
+1535
+01:12:01,360 --> 01:12:05,520
+models is diffusion models um you can
+
+1536
+01:12:03,520 --> 01:12:08,600
+train from scratch by basically noising
+
+1537
+01:12:05,520 --> 01:12:10,600
+the input uh applying noise to the input
+
+1538
+01:12:08,600 --> 01:12:12,880
+um in training very efficiently and
+
+1539
+01:12:10,600 --> 01:12:15,639
+these are very widely used
+
+1540
+01:12:12,880 --> 01:12:18,199
+in image generation they're not super
+
+1541
+01:12:15,639 --> 01:12:20,120
+widely used in text just because regular
+
+1542
+01:12:18,199 --> 01:12:22,840
+autor regressive models are so good for
+
+1543
+01:12:20,120 --> 01:12:24,159
+text um but there are a few efforts to
+
+1544
+01:12:22,840 --> 01:12:26,880
+do
+
+1545
+01:12:24,159 --> 01:12:30,920
+that and then a final one is self-
+
+1546
+01:12:26,880 --> 01:12:35,120
+refine and the idea behind self- refine
+
+1547
+01:12:30,920 --> 01:12:39,400
+is you um actually maybe I can open the
+
+1548
+01:12:35,120 --> 01:12:39,400
+paper because the paper has a good
+
+1549
+01:12:54,120 --> 01:12:58,239
+figure
+
+1550
+01:12:56,280 --> 01:13:02,679
+actually I thought it had a good
+
+1551
+01:12:58,239 --> 01:13:05,600
+figure um yeah so maybe this is a figure
+
+1552
+01:13:02,679 --> 01:13:08,639
+um so basically uh what you do is you
+
+1553
+01:13:05,600 --> 01:13:10,639
+feed in the input you generate an output
+
+1554
+01:13:08,639 --> 01:13:12,679
+and then you ask the model to give you
+
+1555
+01:13:10,639 --> 01:13:15,520
+feedback on the output and say yes this
+
+1556
+01:13:12,679 --> 01:13:16,760
+output is good or um like let's say
+
+1557
+01:13:15,520 --> 01:13:19,679
+you're doing code generation it could
+
+1558
+01:13:16,760 --> 01:13:21,920
+say no this output has an error in it um
+
+1559
+01:13:19,679 --> 01:13:24,719
+this is a problem with your output and
+
+1560
+01:13:21,920 --> 01:13:27,840
+then you feed in both the output and the
+
+1561
+01:13:24,719 --> 01:13:29,480
+feedback back uh and ask the model to
+
+1562
+01:13:27,840 --> 01:13:32,239
+refine its output and you do this over
+
+1563
+01:13:29,480 --> 01:13:35,280
+and over again and this allows you to uh
+
+1564
+01:13:32,239 --> 01:13:36,840
+improve the output and uh this is has
+
+1565
+01:13:35,280 --> 01:13:39,600
+ended up being pretty effective in a
+
+1566
+01:13:36,840 --> 01:13:41,159
+pretty wide number of tasks one caveat
+
+1567
+01:13:39,600 --> 01:13:44,040
+about this is your model has to be
+
+1568
+01:13:41,159 --> 01:13:47,000
+really good for this to work so um only
+
+1569
+01:13:44,040 --> 01:13:49,239
+models kind of on the level of GPT 4 not
+
+1570
+01:13:47,000 --> 01:13:52,000
+on the level of GPT 3.5 have the ability
+
+1571
+01:13:49,239 --> 01:13:54,040
+to do this pretty consistently so it is
+
+1572
+01:13:52,000 --> 01:13:57,040
+something you need to be aware
+
+1573
+01:13:54,040 --> 01:13:57,040
+of
+
+1574
+01:13:59,760 --> 01:14:03,600
+cool yep that's all I I had for today
+
+1575
+01:14:02,400 --> 01:14:06,600
+I'm happy
+
+1576
+01:14:03,600 --> 01:14:06,600
+to
+
+1577
+01:14:07,159 --> 01:14:10,159
+take
+
+1578
+01:14:20,600 --> 01:14:27,320
+yep yep that this is a great question so
+
+1579
+01:14:23,920 --> 01:14:28,840
+if sta has the potential to address
+
+1580
+01:14:27,320 --> 01:14:32,120
+information loss why would we ever
+
+1581
+01:14:28,840 --> 01:14:33,840
+choose a Cascade model I think basically
+
+1582
+01:14:32,120 --> 01:14:37,440
+there's potentially two reasons one
+
+1583
+01:14:33,840 --> 01:14:39,199
+reason is um data availability so in
+
+1584
+01:14:37,440 --> 01:14:42,639
+order to train a stacked model you
+
+1585
+01:14:39,199 --> 01:14:43,430
+obviously need the outputs I guess you
+
+1586
+01:14:42,639 --> 01:14:44,639
+could
+
+1587
+01:14:43,430 --> 01:14:48,440
+[Music]
+
+1588
+01:14:44,639 --> 01:14:50,880
+um yeah I guess you could run
+
+1589
+01:14:48,440 --> 01:14:53,199
+the and generate outputs for every
+
+1590
+01:14:50,880 --> 01:14:54,840
+training example you have um but you
+
+1591
+01:14:53,199 --> 01:14:55,840
+would need to do that so you would need
+
+1592
+01:14:54,840 --> 01:14:58,639
+to to
+
+1593
+01:14:55,840 --> 01:14:59,920
+run speech recognition for every example
+
+1594
+01:14:58,639 --> 01:15:02,760
+and you also
+
+1595
+01:14:59,920 --> 01:15:05,199
+couldn't you couldn't use any examples
+
+1596
+01:15:02,760 --> 01:15:07,600
+where you don't have the original input
+
+1597
+01:15:05,199 --> 01:15:10,320
+so you couldn't use text to text
+
+1598
+01:15:07,600 --> 01:15:12,239
+examples unless you like synthesize
+
+1599
+01:15:10,320 --> 01:15:14,159
+speech from text for machine translation
+
+1600
+01:15:12,239 --> 01:15:15,840
+for example so makes it a little bit
+
+1601
+01:15:14,159 --> 01:15:17,360
+more tricky due to the data requirements
+
+1602
+01:15:15,840 --> 01:15:19,239
+but that's not
+
+1603
+01:15:17,360 --> 01:15:22,560
+insurmountable the second reason is
+
+1604
+01:15:19,239 --> 01:15:24,400
+complexity and efficiency so you know
+
+1605
+01:15:22,560 --> 01:15:27,920
+you do have to come up with a model that
+
+1606
+01:15:24,400 --> 01:15:29,520
+takes in speed and text and run set and
+
+1607
+01:15:27,920 --> 01:15:30,920
+it might be easier just to hook together
+
+1608
+01:15:29,520 --> 01:15:34,719
+a speech recognitional with a
+
+1609
+01:15:30,920 --> 01:15:37,920
+translation so but like I think overall
+
+1610
+01:15:34,719 --> 01:15:39,639
+I I like these methods I I think these
+
+1611
+01:15:37,920 --> 01:15:41,159
+are good methods to use if you're if
+
+1612
+01:15:39,639 --> 01:15:42,480
+you're thinking about using a Cascade
+
+1613
+01:15:41,159 --> 01:15:44,199
+system you should definitely consider
+
+1614
+01:15:42,480 --> 01:15:47,199
+using a stack system in
+
+1615
+01:15:44,199 --> 01:15:47,199
+sense
+
+1616
+01:15:52,080 --> 01:15:56,960
+yeah yeah can you measure the
+
+1617
+01:15:55,159 --> 01:15:59,400
+contribution of each component to an
+
+1618
+01:15:56,960 --> 01:16:00,639
+ensemble um the very very easy way to do
+
+1619
+01:15:59,400 --> 01:16:02,199
+that is look at the interpolation
+
+1620
+01:16:00,639 --> 01:16:05,360
+coefficients if you train the
+
+1621
+01:16:02,199 --> 01:16:06,800
+interpolation coefficients um otherwise
+
+1622
+01:16:05,360 --> 01:16:08,920
+I guess it depends on what you mean by
+
+1623
+01:16:06,800 --> 01:16:10,480
+each contribution but I you know looking
+
+1624
+01:16:08,920 --> 01:16:12,280
+at the interpolation coefficients is a
+
+1625
+01:16:10,480 --> 01:16:16,320
+pretty good way to do
+
+1626
+01:16:12,280 --> 01:16:16,320
+it also just how much did the
+
+1627
+01:16:21,480 --> 01:16:27,400
+accuracy is iterative refinement the
+
+1628
+01:16:24,159 --> 01:16:30,199
+same idea as boosting in traditional
+
+1629
+01:16:27,400 --> 01:16:30,199
+like machine Learning
+
+1630
+01:16:30,320 --> 01:16:34,920
+Systems I think it's a little bit
+
+1631
+01:16:32,920 --> 01:16:36,520
+different um because iterative
+
+1632
+01:16:34,920 --> 01:16:38,920
+refinement what I'm talking about here
+
+1633
+01:16:36,520 --> 01:16:41,120
+it's usually taking in the output like
+
+1634
+01:16:38,920 --> 01:16:43,320
+rather complex output of a system and
+
+1635
+01:16:41,120 --> 01:16:44,920
+modifying it so you're not just
+
+1636
+01:16:43,320 --> 01:16:47,080
+modifying the
+
+1637
+01:16:44,920 --> 01:16:49,880
+probabilities of like a single
+
+1638
+01:16:47,080 --> 01:16:53,080
+classifier you're modifying the actual
+
+1639
+01:16:49,880 --> 01:16:55,960
+outputs that were generated then from
+
+1640
+01:16:53,080 --> 01:16:59,560
+the point of view of a boosting
+
+1641
+01:16:55,960 --> 01:17:02,560
+model over a single categorical output
+
+1642
+01:16:59,560 --> 01:17:04,520
+it might actually be similar or the same
+
+1643
+01:17:02,560 --> 01:17:06,480
+but this is more like uh you you
+
+1644
+01:17:04,520 --> 01:17:08,159
+generated a textual output and then you
+
+1645
+01:17:06,480 --> 01:17:10,400
+feed in the textual output to the other
+
+1646
+01:17:08,159 --> 01:17:12,120
+model and refine like generated a new
+
+1647
+01:17:10,400 --> 01:17:14,239
+textual output so I feel like it's a lot
+
+1648
+01:17:12,120 --> 01:17:18,639
+more
+
+1649
+01:17:14,239 --> 01:17:18,639
+complex cool okay thank thanks a lot
+
+1650
+01:17:18,840 --> 01:17:21,840
+everyone
\ No newline at end of file