CMU Advanced NLP 2024 (14) Ensembling and Mixture of Experts/transcript.vtt

WEBVTT

00:00:00.760 --> 00:00:07.240
he everyone so I'd like to get

00:00:03.279 --> 00:00:09.320
started the first thing is that um I

00:00:07.240 --> 00:00:11.160
heard from the adws people that they

00:00:09.320 --> 00:00:14.440
started the

00:00:11.160 --> 00:00:17.840
process of

00:00:14.440 --> 00:00:19.400
getting things issued on the 26th which

00:00:17.840 --> 00:00:21.480
is three days ago so you should be

00:00:19.400 --> 00:00:23.560
getting it soon uh for reference I

00:00:21.480 --> 00:00:25.599
submitted the form about seven days

00:00:23.560 --> 00:00:28.359
before that so they're moving very

00:00:25.599 --> 00:00:29.599
slowly but I think you should have AWS

00:00:28.359 --> 00:00:31.920
credits by the end of the week if you

00:00:29.599 --> 00:00:35.120
need them to run uh GPU machines or

00:00:31.920 --> 00:00:37.960
stuff like that the moment you get AWS

00:00:35.120 --> 00:00:39.960
credits or maybe even before you get AWS

00:00:37.960 --> 00:00:43.320
credits I might suggest that you try to

00:00:39.960 --> 00:00:46.760
start uh a GPU machine like a P2 machine

00:00:43.320 --> 00:00:49.160
or something like that because um

00:00:46.760 --> 00:00:51.760
sometimes you need to file for a limit

00:00:49.160 --> 00:00:53.640
increase uh to get a P2 machine and that

00:00:51.760 --> 00:00:55.879
also takes a little bit of time so I I

00:00:53.640 --> 00:00:59.160
would suggest that you uh you take a

00:00:55.879 --> 00:01:01.160
look at doing that um so you go to like

00:00:59.160 --> 00:01:02.800
if you're using AWS if you're not using

00:01:01.160 --> 00:01:05.119
AWS it doesn't matter but if you're

00:01:02.800 --> 00:01:08.119
using AWS you can go to launch instance

00:01:05.119 --> 00:01:11.520
and try to launch a p2x large machine um

00:01:08.119 --> 00:01:13.159
or something like that so uh but yeah

00:01:11.520 --> 00:01:14.920
anyway hopefully that will be done soon

00:01:13.159 --> 00:01:16.600
I'm sorry about the delay on this they

00:01:14.920 --> 00:01:21.400
said it would take seven days and it's

00:01:16.600 --> 00:01:24.280
taken almost twice at now so um my

00:01:21.400 --> 00:01:26.439
apologies any other uh things before we

00:01:24.280 --> 00:01:26.439
get

00:01:28.759 --> 00:01:34.520
started um okay I I don't see any so

00:01:31.920 --> 00:01:37.280
I'll go ahead with this um I have

00:01:34.520 --> 00:01:39.240
slightly fewer slides today so I might

00:01:37.280 --> 00:01:40.960
go a little bit off the slides and talk

00:01:39.240 --> 00:01:44.759
about papers and stuff or we might

00:01:40.960 --> 00:01:46.920
finish early uh either way so um but

00:01:44.759 --> 00:01:48.439
what I would like to talk about is um

00:01:46.920 --> 00:01:53.320
combining multiple

00:01:48.439 --> 00:01:55.479
models and this is uh really important

00:01:53.320 --> 00:01:57.520
and useful if you want to get like an

00:01:55.479 --> 00:02:00.719
extra few points of

00:01:57.520 --> 00:02:03.159
accuracy uh for anything basically

00:02:00.719 --> 00:02:04.039
because it's a pretty reliable way to

00:02:03.159 --> 00:02:06.960
get

00:02:04.039 --> 00:02:08.879
improvements um and there's a a bunch of

00:02:06.960 --> 00:02:11.239
different kind of related but different

00:02:08.879 --> 00:02:13.680
topics that I'm going to talk about

00:02:11.239 --> 00:02:15.519
today but anyway the the basic

00:02:13.680 --> 00:02:19.239
background is that we have many models

00:02:15.519 --> 00:02:22.920
uh that exist and the reason why we have

00:02:19.239 --> 00:02:25.840
many models that exist is multiple fold

00:02:22.920 --> 00:02:28.160
number one we could have different model

00:02:25.840 --> 00:02:30.080
architectures um and we could also have

00:02:28.160 --> 00:02:34.440
different initializations of those model

00:02:30.080 --> 00:02:37.879
architectures so um normally you know if

00:02:34.440 --> 00:02:40.319
we do initialization we will initial

00:02:37.879 --> 00:02:42.360
initialize our model architecture like

00:02:40.319 --> 00:02:44.680
let's say we initialize a llama

00:02:42.360 --> 00:02:45.920
architecture uh we start out with random

00:02:44.680 --> 00:02:49.319
7B

00:02:45.920 --> 00:02:52.879
parameters and then we train and we get

00:02:49.319 --> 00:02:53.840
llama 7B for uh our pre-training or

00:02:52.879 --> 00:02:57.280
llama

00:02:53.840 --> 00:02:58.599
27b um we might initialize another model

00:02:57.280 --> 00:03:00.599
this could be you know the same

00:02:58.599 --> 00:03:02.360
architecture different architecture Ure

00:03:00.599 --> 00:03:04.840
train it on the same data or different

00:03:02.360 --> 00:03:07.000
data and get something like mistol

00:03:04.840 --> 00:03:08.599
mistol 7B in this case actually maybe

00:03:07.000 --> 00:03:10.080
these are I should have indicated that

00:03:08.599 --> 00:03:11.680
these are different architectures but

00:03:10.080 --> 00:03:13.879
you know we get a different pre-rain

00:03:11.680 --> 00:03:15.599
model and of course uh we could also

00:03:13.879 --> 00:03:18.640
make it bigger or smaller or whatever

00:03:15.599 --> 00:03:21.720
else and then we get llama 270b over

00:03:18.640 --> 00:03:23.519
here and then after we do that there's a

00:03:21.720 --> 00:03:25.319
lot of fine tuning that goes on

00:03:23.519 --> 00:03:29.360
according to different strategies so we

00:03:25.319 --> 00:03:32.640
have um you know llama 27b instruct uh

00:03:29.360 --> 00:03:37.760
vun 7B uh version

00:03:32.640 --> 00:03:41.000
1.5 um mistol 7B instruct uh news uh

00:03:37.760 --> 00:03:45.239
Hermes 2 mistal 7B or llama 270b

00:03:41.000 --> 00:03:47.239
instruct so we have um a variety of

00:03:45.239 --> 00:03:49.400
architectures a variety of random

00:03:47.239 --> 00:03:51.480
initializations of those architectures a

00:03:49.400 --> 00:03:54.799
variety of pre-train models due to

00:03:51.480 --> 00:03:57.439
pre-training data or base models and

00:03:54.799 --> 00:03:58.920
then a variety of fine dun models um and

00:03:57.439 --> 00:04:01.120
so we have this kind of like branching

00:03:58.920 --> 00:04:02.959
tree basically

00:04:01.120 --> 00:04:04.319
um the reason why this is important is

00:04:02.959 --> 00:04:06.680
because when we're combining multiple

00:04:04.319 --> 00:04:08.400
models together some of the methods are

00:04:06.680 --> 00:04:09.959
applicable to completely different

00:04:08.400 --> 00:04:12.439
models some of the methods are only

00:04:09.959 --> 00:04:15.000
applicable to models that share the same

00:04:12.439 --> 00:04:16.720
architecture and some of them are only

00:04:15.000 --> 00:04:19.199
applicable to models that share the same

00:04:16.720 --> 00:04:20.959
initialization and training trajectory

00:04:19.199 --> 00:04:23.680
and so I'll try to distinguish between

00:04:20.959 --> 00:04:23.680
those as we go

00:04:24.040 --> 00:04:27.919
forward

00:04:25.560 --> 00:04:29.960
cool so the first thing I I'll talk

00:04:27.919 --> 00:04:32.600
about is model ensembling and and

00:04:29.960 --> 00:04:34.320
ensembling is kind of the a very general

00:04:32.600 --> 00:04:37.600
technique that you can use in a lot of

00:04:34.320 --> 00:04:39.360
different uh ways but it has its

00:04:37.600 --> 00:04:43.039
disadvantages as

00:04:39.360 --> 00:04:47.199
well so basically embling is combining

00:04:43.039 --> 00:04:50.320
the predictions from multiple models

00:04:47.199 --> 00:04:52.400
and the easiest way to do this ignore

00:04:50.320 --> 00:04:53.800
the lstm here this is just any sequence

00:04:52.400 --> 00:04:56.320
modeling thing it's because the slides

00:04:53.800 --> 00:05:00.120
are old but like let's say this is a a

00:04:56.320 --> 00:05:03.360
Transformer it is calculating the

00:05:00.120 --> 00:05:05.600
current decoder State and you make a

00:05:03.360 --> 00:05:07.600
prediction um this is calculating a

00:05:05.600 --> 00:05:09.199
current decoder State and make uh

00:05:07.600 --> 00:05:11.560
current decoders sayate in making a

00:05:09.199 --> 00:05:13.039
prediction and based on some combination

00:05:11.560 --> 00:05:17.120
of the two predictions you decide what

00:05:13.039 --> 00:05:17.120
you actually want to Output at the next

00:05:17.680 --> 00:05:23.840
step so why would we want to do this um

00:05:22.080 --> 00:05:25.880
does anyone have any ideas why we want

00:05:23.840 --> 00:05:28.639
to use two models instead of using one

00:05:25.880 --> 00:05:31.639
model or just using the best

00:05:28.639 --> 00:05:31.639
model

00:05:32.319 --> 00:05:36.440
or maybe in what situations we would

00:05:34.520 --> 00:05:39.440
want to do

00:05:36.440 --> 00:05:39.440
this

00:05:45.400 --> 00:05:50.319
yeah and what what's the advantage of

00:05:47.960 --> 00:05:50.319
doing

00:05:51.600 --> 00:05:57.000
that yeah it reduces a bias kind kind of

00:05:54.800 --> 00:05:57.000
yeah

00:05:58.639 --> 00:06:01.639
sure

00:06:28.560 --> 00:06:31.560
m

00:06:35.400 --> 00:06:40.360
yeah so um I I'll repeat all of these I

00:06:38.599 --> 00:06:43.960
think all of these are correct so number

00:06:40.360 --> 00:06:47.479
one um it reduces the bias uh caused by

00:06:43.960 --> 00:06:49.199
a single model uh number two it was it's

00:06:47.479 --> 00:06:52.199
kind of like a beian perspective which

00:06:49.199 --> 00:06:54.000
I'll talk about in a second and then

00:06:52.199 --> 00:06:56.039
number three we have different models

00:06:54.000 --> 00:06:58.520
and models are better at some things and

00:06:56.039 --> 00:07:00.400
worse at other things

00:06:58.520 --> 00:07:02.720
um

00:07:00.400 --> 00:07:05.960
so talking about the better at some

00:07:02.720 --> 00:07:08.319
things and worse at other things um the

00:07:05.960 --> 00:07:10.960
basic idea behind embling is that the

00:07:08.319 --> 00:07:14.240
errors that model m models make tend to

00:07:10.960 --> 00:07:15.840
not be consistent it not tend to not be

00:07:14.240 --> 00:07:21.520
as consistent as when the model is

00:07:15.840 --> 00:07:24.800
getting it correct so we might have um

00:07:21.520 --> 00:07:26.160
we might have one model that says uh

00:07:24.800 --> 00:07:28.199
like let's say we just have really

00:07:26.160 --> 00:07:30.680
really bad models this is kind of a

00:07:28.199 --> 00:07:31.720
really um

00:07:30.680 --> 00:07:35.960
obvious

00:07:31.720 --> 00:07:38.440
example but we have like the dog the dog

00:07:35.960 --> 00:07:42.639
barks and then

00:07:38.440 --> 00:07:46.039
runs and then uh Dives or something like

00:07:42.639 --> 00:07:49.000
that and we have uh one one model that

00:07:46.039 --> 00:07:50.560
just had tons of stuff about diving in

00:07:49.000 --> 00:07:52.120
its training data another model that had

00:07:50.560 --> 00:07:54.240
tons of stuff about running in its

00:07:52.120 --> 00:07:56.560
training data or or marathons or

00:07:54.240 --> 00:08:00.039
something staining data so we'll get

00:07:56.560 --> 00:08:01.800
model one and model one we'll to give

00:08:00.039 --> 00:08:06.240
like a probability of like

00:08:01.800 --> 00:08:08.280
0.3 maybe 0.4 and

00:08:06.240 --> 00:08:10.360
0.05 and then we'll have another one

00:08:08.280 --> 00:08:13.039
over here that's like

00:08:10.360 --> 00:08:17.319
0.32

00:08:13.039 --> 00:08:19.759
0.41 and 0 sorry

00:08:17.319 --> 00:08:23.039
0.05 and

00:08:19.759 --> 00:08:25.759
0.41 or something like this and so when

00:08:23.039 --> 00:08:27.639
you average the two together you tend to

00:08:25.759 --> 00:08:29.240
get the right answer more often because

00:08:27.639 --> 00:08:31.720
kind of the mistakes that they make tend

00:08:29.240 --> 00:08:33.479
to less correlated than the probability

00:08:31.720 --> 00:08:35.880
of getting and of course it's not

00:08:33.479 --> 00:08:38.200
perfect because unbled models are not

00:08:35.880 --> 00:08:39.880
perfect but this is a a general tendency

00:08:38.200 --> 00:08:42.240
that we see a lot in

00:08:39.880 --> 00:08:45.959
models

00:08:42.240 --> 00:08:47.720
um and um it's because of this it kind

00:08:45.959 --> 00:08:52.320
of Smooths over the idiosyncrasies of

00:08:47.720 --> 00:08:54.800
the models you can even um gist Ensemble

00:08:52.320 --> 00:08:57.519
models from different checkpoints and

00:08:54.800 --> 00:08:58.959
that still gives you improvements and so

00:08:57.519 --> 00:09:00.560
when you Ensemble models from different

00:08:58.959 --> 00:09:02.600
checkpoints it's basically just what

00:09:00.560 --> 00:09:05.920
data did they see most recently and that

00:09:02.600 --> 00:09:07.839
also Smooths over you know uh the fact

00:09:05.920 --> 00:09:10.600
that like this model happened to see

00:09:07.839 --> 00:09:13.000
some data more recently and so it's less

00:09:10.600 --> 00:09:16.120
uh you know it's biased towards doing

00:09:13.000 --> 00:09:18.440
that so uh this is a a pretty effective

00:09:16.120 --> 00:09:20.079
method this is one of the few methods

00:09:18.440 --> 00:09:21.959
that I know is going to improve my

00:09:20.079 --> 00:09:25.120
accuracy almost every time like there's

00:09:21.959 --> 00:09:27.880
a bunch of methods that you can apply um

00:09:25.120 --> 00:09:29.680
and I ensembling it's very rare for me

00:09:27.880 --> 00:09:31.959
to Ensemble two models together not get

00:09:29.680 --> 00:09:34.839
a boost in accuracy in some way so it's

00:09:31.959 --> 00:09:34.839
a good thing to

00:09:35.600 --> 00:09:41.040
that there's two main ways to combine

00:09:38.680 --> 00:09:42.560
models together and both of them are

00:09:41.040 --> 00:09:45.800
useful in different

00:09:42.560 --> 00:09:48.079
situations the first one is linear

00:09:45.800 --> 00:09:49.600
interpolation and when you do linear

00:09:48.079 --> 00:09:51.240
interpolation basically what you're

00:09:49.600 --> 00:09:53.720
doing is you're taking the weighted

00:09:51.240 --> 00:09:56.839
average of model

00:09:53.720 --> 00:10:00.360
probabilities and the way that looks

00:09:56.839 --> 00:10:04.040
mathematically is like this um this is a

00:10:00.360 --> 00:10:05.680
probability according to the model M so

00:10:04.040 --> 00:10:08.000
this is just you know the probability of

00:10:05.680 --> 00:10:11.720
the next token according to model M this

00:10:08.000 --> 00:10:13.200
is the probability of selecting model M

00:10:11.720 --> 00:10:18.040
so you talked a little bit about the

00:10:13.200 --> 00:10:19.920
basian approach uh to this and this is

00:10:18.040 --> 00:10:23.519
basically saying what is the probability

00:10:19.920 --> 00:10:26.519
that the parameters of model M

00:10:23.519 --> 00:10:30.320
are the ones that we want to be choosing

00:10:26.519 --> 00:10:32.680
in this at this particular time step and

00:10:30.320 --> 00:10:34.640
then we will we will calculate this and

00:10:32.680 --> 00:10:38.120
so then you take the sum over this and

00:10:34.640 --> 00:10:38.120
this gives you the next

00:10:39.560 --> 00:10:44.800
probability for the second term you can

00:10:42.639 --> 00:10:47.120
do this in two ways the most common way

00:10:44.800 --> 00:10:51.800
to do this is just to have this be a

00:10:47.120 --> 00:10:55.279
constant so you you basically

00:10:51.800 --> 00:10:55.279
Define mixture

00:10:55.920 --> 00:11:01.240
weights uh which are like um

00:11:08.480 --> 00:11:13.480
where the sum of the mixture weights is

00:11:10.760 --> 00:11:16.160
equal to one and this is always between

00:11:13.480 --> 00:11:18.639
zero and one and so if you do this then

00:11:16.160 --> 00:11:21.000
this is just constant and you can uh

00:11:18.639 --> 00:11:23.519
interpolate them together constantly but

00:11:21.000 --> 00:11:25.680
you can also actually explicitly model

00:11:23.519 --> 00:11:27.240
this probability and say oh I'm

00:11:25.680 --> 00:11:30.279
currently in a situation where I really

00:11:27.240 --> 00:11:31.880
think model M will do a good job of uh

00:11:30.279 --> 00:11:33.440
you know predicting the probability so I

00:11:31.880 --> 00:11:36.160
want to put most of my probability on

00:11:33.440 --> 00:11:39.000
model M so you can actually learn this

00:11:36.160 --> 00:11:40.079
dynamically as well um and so if you

00:11:39.000 --> 00:11:44.360
have

00:11:40.079 --> 00:11:45.920
uh this actually um is rather practical

00:11:44.360 --> 00:11:47.120
and easy to do because what you can do

00:11:45.920 --> 00:11:48.920
is you can just calculate the

00:11:47.120 --> 00:11:51.399
probability according to each model at

00:11:48.920 --> 00:11:53.120
each time step and train this model

00:11:51.399 --> 00:11:55.519
separately without loading these models

00:11:53.120 --> 00:11:59.399
into memory at at the time of training

00:11:55.519 --> 00:12:00.959
those models so uh yeah this is um some

00:11:59.399 --> 00:12:04.800
you can do as

00:12:00.959 --> 00:12:04.800
well any questions about

00:12:06.680 --> 00:12:11.920
this

00:12:08.519 --> 00:12:14.000
Okay cool so the other option is log

00:12:11.920 --> 00:12:15.800
linear interpolation and so linear

00:12:14.000 --> 00:12:18.680
interpolation you're taking a linear

00:12:15.800 --> 00:12:22.040
combination of the probabilities of each

00:12:18.680 --> 00:12:24.959
model log linear interpolation you're

00:12:22.040 --> 00:12:26.079
combining together the log probabilities

00:12:24.959 --> 00:12:29.519
of each

00:12:26.079 --> 00:12:32.639
model and then renormalizing so so that

00:12:29.519 --> 00:12:34.920
you get um that you get an actual

00:12:32.639 --> 00:12:37.760
probabilistic output so basically what

00:12:34.920 --> 00:12:40.720
you do is you have this uh interpolation

00:12:37.760 --> 00:12:44.040
coefficient like I had before but you're

00:12:40.720 --> 00:12:44.040
combining together the log

00:12:44.639 --> 00:12:49.639
probabilities and so here we need to

00:12:47.680 --> 00:12:51.320
take the soft

00:12:49.639 --> 00:12:53.760
Max

00:12:51.320 --> 00:12:55.760
um thinking back here I didn't take the

00:12:53.760 --> 00:12:58.120
softmax does anyone have an idea why I

00:12:55.760 --> 00:13:02.000
didn't take the soft

00:12:58.120 --> 00:13:02.000
Max or why I didn't need

00:13:08.160 --> 00:13:12.199
to why why I need to

00:13:21.600 --> 00:13:27.680
here yeah

00:13:23.639 --> 00:13:30.440
so this probability is gu to be z z and

00:13:27.680 --> 00:13:32.240
one and add up to one this probability

00:13:30.440 --> 00:13:33.760
is also guaranteed to be zero and one

00:13:32.240 --> 00:13:35.680
and add up to one and then when you

00:13:33.760 --> 00:13:37.120
multiply those together uh you can do a

00:13:35.680 --> 00:13:39.160
little bit of math and demonstrate that

00:13:37.120 --> 00:13:41.440
the resulting thing will be between zero

00:13:39.160 --> 00:13:42.839
and one and add up to one that's not the

00:13:41.440 --> 00:13:44.399
case anymore when we start doing things

00:13:42.839 --> 00:13:47.639
in log space because it's just not a

00:13:44.399 --> 00:13:50.160
linear function anyway so um you need to

00:13:47.639 --> 00:13:51.959
renormalize like this luckily this is

00:13:50.160 --> 00:13:54.920
super easy like anything else you do in

00:13:51.959 --> 00:13:56.959
py torch you just add things together

00:13:54.920 --> 00:13:59.320
and take a soft Max and you'll you'll

00:13:56.959 --> 00:14:02.519
get an output but you do need to do

00:13:59.320 --> 00:14:05.279
otherwise you're going to get something

00:14:02.519 --> 00:14:07.279
weird um the interpolation coefficient

00:14:05.279 --> 00:14:09.639
here also can be set to a constant so

00:14:07.279 --> 00:14:12.759
you can you could learn it uh kind of

00:14:09.639 --> 00:14:15.320
dynamically or it could be

00:14:12.759 --> 00:14:17.720
separate cool and these actually have

00:14:15.320 --> 00:14:19.639
different meaning oh sorry go ahead you

00:14:17.720 --> 00:14:23.880
T on

00:14:19.639 --> 00:14:26.759
the Yeah Yeah so basically the

00:14:23.880 --> 00:14:29.880
way the way you would do this is you

00:14:26.759 --> 00:14:32.399
would have either

00:14:29.880 --> 00:14:33.920
the same model you you would either take

00:14:32.399 --> 00:14:35.279
representations from one of these

00:14:33.920 --> 00:14:37.480
language models or you would take

00:14:35.279 --> 00:14:38.440
representations from another model and

00:14:37.480 --> 00:14:41.639
you would

00:14:38.440 --> 00:14:43.959
just have a model that

00:14:41.639 --> 00:14:46.480
predicts uh what this interpolation

00:14:43.959 --> 00:14:48.279
coefficient would be and the

00:14:46.480 --> 00:14:49.720
optimization objective for that

00:14:48.279 --> 00:14:52.759
interpolation coefficient is just

00:14:49.720 --> 00:14:56.120
maximizing the probability

00:14:52.759 --> 00:14:59.600
whatever so this could also be good um

00:14:56.120 --> 00:15:01.839
because this interpolation coefficient

00:14:59.600 --> 00:15:07.160
only like let's say you're interpolating

00:15:01.839 --> 00:15:09.399
two models together it has one degree of

00:15:07.160 --> 00:15:13.320
Freedom at each time step right because

00:15:09.399 --> 00:15:15.320
you're only predicting a probability um

00:15:13.320 --> 00:15:17.839
if you have uh if you have five models

00:15:15.320 --> 00:15:20.240
you have uh you basically would be doing

00:15:17.839 --> 00:15:24.199
a soft match over

00:15:20.240 --> 00:15:25.519
five five outputs and that's a lot fewer

00:15:24.199 --> 00:15:27.600
that's a lot fewer than the whole

00:15:25.519 --> 00:15:29.880
vocabulary right and so this is

00:15:27.600 --> 00:15:31.639
relatively learning a good interpolation

00:15:29.880 --> 00:15:34.160
coefficient is relatively easy compared

00:15:31.639 --> 00:15:35.800
to learning what word to predict next

00:15:34.160 --> 00:15:36.880
and because of this you could actually

00:15:35.800 --> 00:15:39.759
tune

00:15:36.880 --> 00:15:42.880
this um sorry you could tune this

00:15:39.759 --> 00:15:44.600
probability on a very small data set and

00:15:42.880 --> 00:15:46.959
you could even have it be context

00:15:44.600 --> 00:15:48.480
independent so you could just be you

00:15:46.959 --> 00:15:51.399
know

00:15:48.480 --> 00:15:55.880
calculating literally five five

00:15:51.399 --> 00:15:57.399
parameters here um and so because of

00:15:55.880 --> 00:16:00.319
that like let's say you have a special

00:15:57.399 --> 00:16:02.639
domain or a special task where you have

00:16:00.319 --> 00:16:04.920
like 50 training examples or something

00:16:02.639 --> 00:16:07.399
like that or you know 100 training

00:16:04.920 --> 00:16:08.959
examples you can learn this

00:16:07.399 --> 00:16:12.480
interpolation coefficient very

00:16:08.959 --> 00:16:15.880
effectively uh on just a few a very

00:16:12.480 --> 00:16:18.120
small number of training examples um but

00:16:15.880 --> 00:16:20.000
like it could be very useful because

00:16:18.120 --> 00:16:23.920
like let's say you have a special domain

00:16:20.000 --> 00:16:25.639
medical language model that's 1.3

00:16:23.920 --> 00:16:27.759
billion parameters that you trained

00:16:25.639 --> 00:16:29.639
yourself and then you have a 70 billion

00:16:27.759 --> 00:16:31.079
parameter language model

00:16:29.639 --> 00:16:33.680
that's like really good at modeling

00:16:31.079 --> 00:16:35.399
General English um so then you could

00:16:33.680 --> 00:16:39.120
learn the interpolation coefficient

00:16:35.399 --> 00:16:40.600
between those two such that um the large

00:16:39.120 --> 00:16:41.800
general purpose language model will be

00:16:40.600 --> 00:16:43.959
generating all of the kind of

00:16:41.800 --> 00:16:46.360
grammatical stuff but whenever you

00:16:43.959 --> 00:16:48.480
switch over to modeling technical terms

00:16:46.360 --> 00:16:50.040
from the medical domain then it learns

00:16:48.480 --> 00:16:52.480
to upweight the medical language model

00:16:50.040 --> 00:16:54.199
or something so this can be quite uh

00:16:52.480 --> 00:16:57.000
this can be quite effective if you have

00:16:54.199 --> 00:17:00.839
a limited amount of data that you want

00:16:57.000 --> 00:17:00.839
toing thiss

00:17:01.240 --> 00:17:05.600
um any other questions about that

00:17:09.079 --> 00:17:14.880
yeah yeah I'm just gonna talk about that

00:17:11.760 --> 00:17:17.640
next so linear versus log linear you can

00:17:14.880 --> 00:17:20.880
actually think of this in logic um and

00:17:17.640 --> 00:17:23.640
what I mean by that is um linear is kind

00:17:20.880 --> 00:17:26.640
of like a logical or it tries to come up

00:17:23.640 --> 00:17:29.600
with examples where either one of the

00:17:26.640 --> 00:17:31.679
two assigns a high probability so we

00:17:29.600 --> 00:17:36.200
have the example of like bark

00:17:31.679 --> 00:17:36.200
run um bark run

00:17:55.640 --> 00:18:03.840
diet so if we take the average of these

00:18:00.360 --> 00:18:03.840
two in linear

00:18:04.120 --> 00:18:10.240
space this would be

00:18:07.159 --> 00:18:13.679
0.2 this would be

00:18:10.240 --> 00:18:17.240
0.26 and this would

00:18:13.679 --> 00:18:17.240
be um

00:18:17.400 --> 00:18:26.280
0.21 and so a a linear combination

00:18:21.480 --> 00:18:28.600
between the two will find run to be the

00:18:26.280 --> 00:18:30.600
highest scoring one because on the left

00:18:28.600 --> 00:18:32.280
side we have one model that really likes

00:18:30.600 --> 00:18:33.159
this output and we have another model

00:18:32.280 --> 00:18:35.159
that

00:18:33.159 --> 00:18:39.280
doesn't

00:18:35.159 --> 00:18:42.159
um this is this can be good at using

00:18:39.280 --> 00:18:44.440
models that capture uh different traits

00:18:42.159 --> 00:18:47.679
or it can also be useful if like for

00:18:44.440 --> 00:18:49.840
example you have a you have a small

00:18:47.679 --> 00:18:52.320
model that you really that really

00:18:49.840 --> 00:18:53.840
captures like very specific vocabulary

00:18:52.320 --> 00:18:55.520
and you want to upgrate that specific

00:18:53.840 --> 00:18:56.799
vocabulary that gets a really low

00:18:55.520 --> 00:18:57.720
probability according to a general

00:18:56.799 --> 00:19:01.360
purpose

00:18:57.720 --> 00:19:03.200
model um this is also necessary when any

00:19:01.360 --> 00:19:04.520
model can assign zero probabilities so

00:19:03.200 --> 00:19:06.720
if you have like an example of

00:19:04.520 --> 00:19:10.080
vocabulary that isn't included in the

00:19:06.720 --> 00:19:11.159
the like vocabulary of another model or

00:19:10.080 --> 00:19:14.280
you have models with different

00:19:11.159 --> 00:19:17.200
vocabularies it's necessary to do this

00:19:14.280 --> 00:19:19.200
log linear is more like logical and um

00:19:17.200 --> 00:19:22.240
so the interpolated model only likes

00:19:19.200 --> 00:19:23.799
choices where all the models agree and

00:19:22.240 --> 00:19:25.640
this is particularly good when you want

00:19:23.799 --> 00:19:27.440
to restrict possible answers like you

00:19:25.640 --> 00:19:29.280
want to have one model be able to say no

00:19:27.440 --> 00:19:32.080
I really don't like this so never output

00:19:29.280 --> 00:19:34.200
it so um for example if you wanted to

00:19:32.080 --> 00:19:37.360
train a model that you knew was very

00:19:34.200 --> 00:19:38.919
adverse to toxic language and prevent uh

00:19:37.360 --> 00:19:42.600
the model from outputting toxic language

00:19:38.919 --> 00:19:45.200
you could use log linear mod so I I

00:19:42.600 --> 00:19:47.559
can't unfortunately uh calculate logs

00:19:45.200 --> 00:19:50.080
and exponents in my head well enough to

00:19:47.559 --> 00:19:51.600
uh to decide this but I'm sure that a

00:19:50.080 --> 00:19:53.840
linear

00:19:51.600 --> 00:19:56.840
model the linear model would pick the

00:19:53.840 --> 00:19:59.600
first one here and the log linear

00:19:56.840 --> 00:20:01.679
model would pick the second one because

00:19:59.600 --> 00:20:05.640
the second one has a very low score here

00:20:01.679 --> 00:20:08.640
so that would be downrated um

00:20:05.640 --> 00:20:08.640
by

00:20:16.919 --> 00:20:20.640
yeah yeah so

00:20:25.840 --> 00:20:31.000
if yeah and if there's any chance of

00:20:28.760 --> 00:20:34.159
assigning zero probability according to

00:20:31.000 --> 00:20:36.520
a language model then really you can't

00:20:34.159 --> 00:20:38.200
even test that language model on that on

00:20:36.520 --> 00:20:42.120
that test set

00:20:38.200 --> 00:20:43.640
um so the issue becomes like let's say

00:20:42.120 --> 00:20:45.559
you have two models with different

00:20:43.640 --> 00:20:47.080
vocabulary if you have two models with

00:20:45.559 --> 00:20:49.080
different vocabulary it becomes very

00:20:47.080 --> 00:20:50.559
tricky how to reconcile those two but

00:20:49.080 --> 00:20:53.440
you could do linear interpolation

00:20:50.559 --> 00:20:55.200
between them like match the vocab the

00:20:53.440 --> 00:20:57.559
output vocabularies that they do have

00:20:55.200 --> 00:21:00.120
and then just not worry about the fact

00:20:57.559 --> 00:21:02.760
that the vocabularies are dis jointed

00:21:00.120 --> 00:21:05.039
and because one will assign a zero

00:21:02.760 --> 00:21:07.280
probability to those vocabulary items

00:21:05.039 --> 00:21:12.240
but the other one is fine so you can

00:21:07.280 --> 00:21:14.919
just do that but if you're in general it

00:21:12.240 --> 00:21:16.480
will be very tricky to try to get two

00:21:14.919 --> 00:21:18.559
models with different vocabularies to

00:21:16.480 --> 00:21:21.480
play together nicely so I I would

00:21:18.559 --> 00:21:22.919
suggest um thinking about thinking

00:21:21.480 --> 00:21:25.600
seriously about whether you need to do

00:21:22.919 --> 00:21:31.360
that or not before you start out but

00:21:25.600 --> 00:21:31.360
yeah um uh yes there any

00:21:35.559 --> 00:21:40.960
other

00:21:38.039 --> 00:21:43.360
um you could definitely so the question

00:21:40.960 --> 00:21:45.000
is are there any other types of

00:21:43.360 --> 00:21:47.760
interpolation that have other types of

00:21:45.000 --> 00:21:50.159
logical components like exor or nor um

00:21:47.760 --> 00:21:52.840
you could definitely come up with one uh

00:21:50.159 --> 00:21:55.440
I I am struggling a little bit to think

00:21:52.840 --> 00:21:57.520
about when you would want to do that but

00:21:55.440 --> 00:22:02.840
I'm sure

00:21:57.520 --> 00:22:05.840
you is is the inherent that the

00:22:02.840 --> 00:22:05.840
err

00:22:09.120 --> 00:22:14.480
not so what what if the errors are not

00:22:12.640 --> 00:22:15.919
what if the errors are correlated so

00:22:14.480 --> 00:22:18.200
think about what happens if the errors

00:22:15.919 --> 00:22:20.000
are perfectly correlated um which is

00:22:18.200 --> 00:22:25.840
when you're using the same model in two

00:22:20.000 --> 00:22:25.840
parts of the uh like on top so you

00:22:27.000 --> 00:22:30.520
literally uh these

00:22:29.159 --> 00:22:32.679
model one and model two are the same

00:22:30.520 --> 00:22:36.720
model if that's the case nothing happens

00:22:32.679 --> 00:22:39.200
it doesn't get worse um and

00:22:36.720 --> 00:22:43.039
so of course because this is machine

00:22:39.200 --> 00:22:45.080
learning there's no guarantee like you

00:22:43.039 --> 00:22:47.559
know unless we make some assumptions

00:22:45.080 --> 00:22:49.200
about the relationship between like the

00:22:47.559 --> 00:22:52.279
training set and the test set or the

00:22:49.200 --> 00:22:53.760
models errors in the test set um you can

00:22:52.279 --> 00:22:57.039
always do something that will make your

00:22:53.760 --> 00:22:59.240
accuracy worse um like let's say we flip

00:22:57.039 --> 00:23:00.360
the labels of a binary class

00:22:59.240 --> 00:23:03.120
no matter what you do you're going to

00:23:00.360 --> 00:23:06.320
make your accuracy worse but

00:23:03.120 --> 00:23:09.000
um no matter what the normal thing you

00:23:06.320 --> 00:23:10.640
would do is it would make your if it

00:23:09.000 --> 00:23:12.480
would improve accuracy normally it would

00:23:10.640 --> 00:23:14.760
decrease your accuracy but like under

00:23:12.480 --> 00:23:16.080
pretty reasonable assumptions it's

00:23:14.760 --> 00:23:20.400
mostly going to be the case that errors

00:23:16.080 --> 00:23:22.320
are deated to some extent um

00:23:20.400 --> 00:23:25.559
so

00:23:22.320 --> 00:23:30.440
yeah you and because of that ensembly

00:23:25.559 --> 00:23:30.440
usually helps yeah

00:23:36.120 --> 00:23:42.019
um about which one

00:23:38.760 --> 00:23:42.019
[Music]

00:23:53.559 --> 00:24:01.240
which let me make sure I didn't mess it

00:23:55.640 --> 00:24:01.240
up on sides okay so in my

00:24:06.960 --> 00:24:13.120
example yeah yeah

00:24:09.640 --> 00:24:13.120
yeah sorry about

00:24:14.360 --> 00:24:19.320
that because this is this is where the

00:24:17.039 --> 00:24:21.840
average is higher and then this is

00:24:19.320 --> 00:24:27.200
one take

00:24:21.840 --> 00:24:29.039
you uh cool any other any other

00:24:27.200 --> 00:24:31.840
questions okay

00:24:29.039 --> 00:24:34.440
okay so

00:24:31.840 --> 00:24:36.320
um another thing I should point out is

00:24:34.440 --> 00:24:39.600
that we don't

00:24:36.320 --> 00:24:41.840
necessarily need to use models only as

00:24:39.600 --> 00:24:44.080
positive evidence so if you're using log

00:24:41.840 --> 00:24:46.039
linear interpolation actually your

00:24:44.080 --> 00:24:49.919
interpolation coefficients do not need

00:24:46.039 --> 00:24:52.520
to be positive they can also be negative

00:24:49.919 --> 00:24:55.360
and you can have uh things where you

00:24:52.520 --> 00:24:57.840
penalize the probabilities given by a

00:24:55.360 --> 00:24:59.679
particular model and this has actually

00:24:57.840 --> 00:25:01.520
been used for a long time it was

00:24:59.679 --> 00:25:04.440
actually used in machine translation

00:25:01.520 --> 00:25:08.840
since like uh 2005 or something like

00:25:04.440 --> 00:25:11.480
this but the basic idea is um that you

00:25:08.840 --> 00:25:13.600
have some models that serve as negative

00:25:11.480 --> 00:25:15.559
evidence so you have kind of a core

00:25:13.600 --> 00:25:17.880
model this might be your really strong

00:25:15.559 --> 00:25:21.520
general purpose language model you have

00:25:17.880 --> 00:25:23.080
a positive uh model which is the model

00:25:21.520 --> 00:25:25.240
that you want to kind of boost up and

00:25:23.080 --> 00:25:27.320
improve and a negative model which you

00:25:25.240 --> 00:25:31.159
want to

00:25:27.320 --> 00:25:33.679
decrease and um one example of this is

00:25:31.159 --> 00:25:36.760
in uh a paper that we did in

00:25:33.679 --> 00:25:40.159
2019 um the core was a machine

00:25:36.760 --> 00:25:42.960
translation model and the negative model

00:25:40.159 --> 00:25:44.880
is an outof domain language model and

00:25:42.960 --> 00:25:46.960
the positive model is an in domain

00:25:44.880 --> 00:25:51.039
language model and so the idea behind

00:25:46.960 --> 00:25:53.880
this is a machine translation model um

00:25:51.039 --> 00:25:55.600
you have to train it on machine

00:25:53.880 --> 00:25:58.320
translation data and machine translation

00:25:55.600 --> 00:26:00.640
data is not very easy to get for

00:25:58.320 --> 00:26:02.360
particular domains for example um you

00:26:00.640 --> 00:26:03.880
might only have machine translation data

00:26:02.360 --> 00:26:06.919
in the news domain and you actually want

00:26:03.880 --> 00:26:09.240
to be uh doing uh translation in the

00:26:06.919 --> 00:26:12.720
medical domain or something so what you

00:26:09.240 --> 00:26:14.640
do is you have your positive model here

00:26:12.720 --> 00:26:17.600
this could be a new this is a machine

00:26:14.640 --> 00:26:19.919
translation model this could be a news

00:26:17.600 --> 00:26:21.320
domain or sorry this could be a medical

00:26:19.919 --> 00:26:22.919
domain language model and this could be

00:26:21.320 --> 00:26:24.360
a news domain language model so you're

00:26:22.919 --> 00:26:25.840
subtracting out the news domain

00:26:24.360 --> 00:26:27.600
probabilities and adding in medical

00:26:25.840 --> 00:26:30.240
domain probabilities move it in that

00:26:27.600 --> 00:26:30.240
direction

00:26:30.440 --> 00:26:36.799
um another example of this is uh

00:26:32.919 --> 00:26:40.000
something called uh D experts um or

00:26:36.799 --> 00:26:43.440
dexperts and the idea here is here you

00:26:40.000 --> 00:26:46.120
have a strong language model as your

00:26:43.440 --> 00:26:48.320
core and then as negative you have a

00:26:46.120 --> 00:26:50.240
weak toxic language model so it was

00:26:48.320 --> 00:26:52.760
trained on lot lots of like bad texts

00:26:50.240 --> 00:26:55.799
that you don't want to be generating and

00:26:52.760 --> 00:26:57.159
the positive is a weak non-toxic

00:26:55.799 --> 00:26:59.279
language model that was trained on lots

00:26:57.159 --> 00:27:03.200
of like inocua

00:26:59.279 --> 00:27:04.399
posts so that would help you detoxify

00:27:03.200 --> 00:27:06.679
the outputs of the

00:27:04.399 --> 00:27:09.799
language so there's lots of examples of

00:27:06.679 --> 00:27:09.799
things like this that you can do

00:27:10.720 --> 00:27:15.880
through

00:27:12.880 --> 00:27:15.880
yeah

00:27:19.320 --> 00:27:25.880
yeah um so the positive in the machine

00:27:22.840 --> 00:27:27.679
translation example this is a so this is

00:27:25.880 --> 00:27:31.760
a machine translation model where the

00:27:27.679 --> 00:27:34.080
input is is like in um English and out

00:27:31.760 --> 00:27:37.880
is in Japanese something like

00:27:34.080 --> 00:27:39.679
that this is only trained on Japanese

00:27:37.880 --> 00:27:42.919
but it's trained on like medical

00:27:39.679 --> 00:27:44.440
Japanese for example Med the domain one

00:27:42.919 --> 00:27:48.480
this is a language model that was

00:27:44.440 --> 00:27:50.600
trained on like news domain um Japanese

00:27:48.480 --> 00:27:54.039
or it could even literally just be

00:27:50.600 --> 00:27:56.360
trained on the side of the machine

00:27:54.039 --> 00:28:00.120
trans um so it's trying to remove out

00:27:56.360 --> 00:28:00.120
the language modeling component from the

00:28:03.720 --> 00:28:06.720
cool

00:28:06.880 --> 00:28:11.480
okay so another thing that I should

00:28:09.880 --> 00:28:14.720
point out I didn't actually put it on

00:28:11.480 --> 00:28:18.399
the slides is um there's a lot of other

00:28:14.720 --> 00:28:19.640
ways to get multiple models and um I

00:28:18.399 --> 00:28:22.600
think a lot of people are probably

00:28:19.640 --> 00:28:23.559
familiar with Dropout um it's a method

00:28:22.600 --> 00:28:27.120
for

00:28:23.559 --> 00:28:29.080
regularizing um it's a method for

00:28:27.120 --> 00:28:31.120
regularizing

00:28:29.080 --> 00:28:33.760
neural networks or deep learning models

00:28:31.120 --> 00:28:37.279
in general and basically the idea is

00:28:33.760 --> 00:28:41.840
every once in a while um during training

00:28:37.279 --> 00:28:45.720
you drop out some portion of the uh like

00:28:41.840 --> 00:28:48.919
nodes in the neural network model and

00:28:45.720 --> 00:28:51.320
you can actually drop

00:28:48.919 --> 00:28:52.640
out and normally what you do is at test

00:28:51.320 --> 00:28:53.919
time then you just don't drop out

00:28:52.640 --> 00:28:56.039
anything and you use the whole neural

00:28:53.919 --> 00:28:59.960
network model but another thing you can

00:28:56.039 --> 00:29:02.559
do is you can drop out a test time drop

00:28:59.960 --> 00:29:04.679
out five times and combine those

00:29:02.559 --> 00:29:06.600
different models together through ensom

00:29:04.679 --> 00:29:10.600
and that's actually something uh that

00:29:06.600 --> 00:29:14.480
people tried in the uh in the Dropout

00:29:10.600 --> 00:29:17.600
paper and this is one way to get

00:29:14.480 --> 00:29:19.640
multiple models uh and actually you can

00:29:17.600 --> 00:29:21.919
demonstrate that this helps the original

00:29:19.640 --> 00:29:24.519
motivation behind Dropout was precisely

00:29:21.919 --> 00:29:26.279
coming from this idea of

00:29:24.519 --> 00:29:29.080
ensembling

00:29:26.279 --> 00:29:31.399
another method

00:29:29.080 --> 00:29:34.799
that has been around for a very long

00:29:31.399 --> 00:29:37.760
time it's another embling method is

00:29:34.799 --> 00:29:41.919
bagging and basically the way bagging

00:29:37.760 --> 00:29:41.919
works is you have a data

00:29:44.000 --> 00:29:50.159
set like this and you just resample the

00:29:47.519 --> 00:29:52.919
data set so you sample all of the output

00:29:50.159 --> 00:29:55.200
with uh replacement and you get another

00:29:52.919 --> 00:29:57.799
data set of equal size and then you

00:29:55.200 --> 00:29:58.559
train on this but you do that like 10

00:29:57.799 --> 00:30:00.120
times

00:29:58.559 --> 00:30:02.679
and you train 10 different models and

00:30:00.120 --> 00:30:04.360
then you emble those models together and

00:30:02.679 --> 00:30:06.000
so this is another way to get multiple

00:30:04.360 --> 00:30:07.519
models and both of these still improve

00:30:06.000 --> 00:30:09.640
your robustness because they basically

00:30:07.519 --> 00:30:11.440
get a different view on the data so they

00:30:09.640 --> 00:30:13.440
smooth over some of the

00:30:11.440 --> 00:30:15.360
idiosyncrasies um and as I mentioned

00:30:13.440 --> 00:30:17.960
before you can also get multiple models

00:30:15.360 --> 00:30:20.120
from different checkpoints and then uh

00:30:17.960 --> 00:30:22.159
put them together and all of these

00:30:20.120 --> 00:30:24.159
methods are pretty related both of them

00:30:22.159 --> 00:30:25.960
basically what they're doing is they're

00:30:24.159 --> 00:30:28.279
taking advantage of the fact that you

00:30:25.960 --> 00:30:29.919
have particular models that saw

00:30:28.279 --> 00:30:32.760
different data or saw data in a

00:30:29.919 --> 00:30:34.120
different order or different nodes saw

00:30:32.760 --> 00:30:35.679
different parts of the data because you

00:30:34.120 --> 00:30:37.799
dropped out some of the nodes when they

00:30:35.679 --> 00:30:41.840
were back propping on particular

00:30:37.799 --> 00:30:44.840
varieties of the data so um even things

00:30:41.840 --> 00:30:46.720
like this can give you models that are

00:30:44.840 --> 00:30:49.760
different enough that to help uh when

00:30:46.720 --> 00:30:49.760
you're onbling or

00:30:52.559 --> 00:30:59.360
combining and then of course um you can

00:30:56.919 --> 00:31:00.799
also

00:30:59.360 --> 00:31:02.480
then of course you can also combine

00:31:00.799 --> 00:31:06.960
together like very different models like

00:31:02.480 --> 00:31:06.960
this and that also works in different

00:31:07.240 --> 00:31:11.159
ways

00:31:09.000 --> 00:31:13.039
cool part of the reason why I wanted to

00:31:11.159 --> 00:31:15.320
mention that Dropout though in

00:31:13.039 --> 00:31:17.120
particular is there's also other

00:31:15.320 --> 00:31:19.240
efficient methods for using multiple

00:31:17.120 --> 00:31:22.000
models so the big problem with

00:31:19.240 --> 00:31:25.399
ensembling is the cost

00:31:22.000 --> 00:31:27.159
and simple ensembling is very expensive

00:31:25.399 --> 00:31:29.240
because it requires you to run multiple

00:31:27.159 --> 00:31:30.519
models at test test time at inference

00:31:29.240 --> 00:31:33.720
time and this is something you don't

00:31:30.519 --> 00:31:35.279
want to be doing if you're you know

00:31:33.720 --> 00:31:38.679
deploying a service or something because

00:31:35.279 --> 00:31:41.080
it like linearly increases your cost by

00:31:38.679 --> 00:31:45.200
um the amount of bottles that you're

00:31:41.080 --> 00:31:47.799
running and it requires both end times

00:31:45.200 --> 00:31:50.120
of computation and end times of memory

00:31:47.799 --> 00:31:51.720
and memory is actually probably the

00:31:50.120 --> 00:31:54.279
worst thing because you need to deploy

00:31:51.720 --> 00:31:58.159
extra GPU machines and other stuff like

00:31:54.279 --> 00:31:59.880
that so um the question is is there any

00:31:58.159 --> 00:32:03.279
way we can get some of the benefits of

00:31:59.880 --> 00:32:06.519
embling without having to create

00:32:03.279 --> 00:32:07.320
multiple models and luckily the answer

00:32:06.519 --> 00:32:09.240
is

00:32:07.320 --> 00:32:11.919
yes

00:32:09.240 --> 00:32:13.960
the method the easiest method for doing

00:32:11.919 --> 00:32:16.600
so is something called parameter

00:32:13.960 --> 00:32:18.399
averaging and basically what you do is

00:32:16.600 --> 00:32:21.960
you just average the parameters of

00:32:18.399 --> 00:32:26.039
multiple models together um this only

00:32:21.960 --> 00:32:29.200
works under certain conditions so does

00:32:26.039 --> 00:32:31.120
anyone um does anyone know what these

00:32:29.200 --> 00:32:33.320
conditions might be there's a few

00:32:31.120 --> 00:32:35.919
obvious ones and maybe a few slightly

00:32:33.320 --> 00:32:35.919
less obvious

00:32:36.039 --> 00:32:40.799
ones so like first question do you think

00:32:38.799 --> 00:32:41.919
you could combine together do you think

00:32:40.799 --> 00:32:45.880
you could average together the

00:32:41.919 --> 00:32:45.880
parameters of llama 7B and Lama

00:32:46.440 --> 00:32:52.639
70b

00:32:48.480 --> 00:32:52.639
no the answer is no but why

00:32:54.480 --> 00:32:58.440
not I mean what does that even mean in

00:32:56.760 --> 00:33:00.480
the first place right like they have

00:32:58.440 --> 00:33:02.799
totally different numbers of parameters

00:33:00.480 --> 00:33:05.840
uh you wouldn't be able to find a one

00:33:02.799 --> 00:33:07.840
toone association between 7B and like 7

00:33:05.840 --> 00:33:12.320
billion parameters and 70 billion

00:33:07.840 --> 00:33:16.880
parameters um what about averaging

00:33:12.320 --> 00:33:19.399
together uh let's let's say llama 7B and

00:33:16.880 --> 00:33:19.399
mistol

00:33:23.080 --> 00:33:29.760
7bs yes no y I'm guessing that like for

00:33:27.440 --> 00:33:29.760
the

00:33:33.760 --> 00:33:38.120
yeah for different architectures the um

00:33:36.760 --> 00:33:41.799
the parameters could mean different

00:33:38.120 --> 00:33:44.159
things and even if the architecture is

00:33:41.799 --> 00:33:45.880
exactly the same even if your random

00:33:44.159 --> 00:33:49.880
initialization is different then that

00:33:45.880 --> 00:33:52.360
would be a disastrous because basically

00:33:49.880 --> 00:33:54.760
in neural networks there's no inherent

00:33:52.360 --> 00:33:58.559
meaning to like parameter number one

00:33:54.760 --> 00:34:01.919
right um and there's the idea of permut

00:33:58.559 --> 00:34:06.679
Inari which is

00:34:01.919 --> 00:34:07.639
um you could like randomly Swap all of

00:34:06.679 --> 00:34:10.280
the

00:34:07.639 --> 00:34:12.079
dimensions uh between within a neural

00:34:10.280 --> 00:34:14.760
network and get exactly the same

00:34:12.079 --> 00:34:17.919
function

00:34:14.760 --> 00:34:22.560
uh as long as kind

00:34:17.919 --> 00:34:24.839
of in layer number one you swap and then

00:34:22.560 --> 00:34:30.359
also take the inputs in the next layer

00:34:24.839 --> 00:34:30.359
also so um you know you know as long

00:34:30.960 --> 00:34:36.399
as if you have a weight Matrix that

00:34:33.679 --> 00:34:40.800
results in the um in the outputs being

00:34:36.399 --> 00:34:49.639
ordered like 1 two three four

00:34:40.800 --> 00:34:54.159
five one or 2 1 3 five four as long as

00:34:49.639 --> 00:34:55.720
you also swap the input direct input

00:34:54.159 --> 00:34:58.400
dimensions of this weight Matrix you get

00:34:55.720 --> 00:35:01.520
exactly the same because they

00:34:58.400 --> 00:35:04.200
linear combinations of the parameters

00:35:01.520 --> 00:35:06.480
together so neural networks have this

00:35:04.200 --> 00:35:08.599
feature of permutation and variance so

00:35:06.480 --> 00:35:11.800
models that were trained from like

00:35:08.599 --> 00:35:13.280
different uh different initializations

00:35:11.800 --> 00:35:15.040
won't be able to be combined together in

00:35:13.280 --> 00:35:18.320
this

00:35:15.040 --> 00:35:20.079
way um but the good luck the good thing

00:35:18.320 --> 00:35:21.359
is actually we have a whole bunch of

00:35:20.079 --> 00:35:25.320
models that come from the same

00:35:21.359 --> 00:35:26.720
pre-trained model right uh so we we have

00:35:25.320 --> 00:35:28.640
this initialization here this

00:35:26.720 --> 00:35:31.280
initialization was used to train Lama

00:35:28.640 --> 00:35:32.920
27b but now we have like hundreds

00:35:31.280 --> 00:35:34.440
hundreds of models that are DED from

00:35:32.920 --> 00:35:37.400
Lama 2 we have hundreds of models that

00:35:34.440 --> 00:35:39.599
are DED from mixol and there all of the

00:35:37.400 --> 00:35:40.920
dimensions actually mean the same thing

00:35:39.599 --> 00:35:43.280
because they're derived from the same

00:35:40.920 --> 00:35:46.680
parameters in the first place so those

00:35:43.280 --> 00:35:48.119
ones we can average together and um

00:35:46.680 --> 00:35:50.359
there's basically two ways that we can

00:35:48.119 --> 00:35:53.520
do this uh one is by averaging together

00:35:50.359 --> 00:35:55.240
multiple checkpoints during training so

00:35:53.520 --> 00:35:57.960
originally this was the big thing that

00:35:55.240 --> 00:36:00.359
people did uh like you would train model

00:35:57.960 --> 00:36:02.119
from scratch for a really long time but

00:36:00.359 --> 00:36:03.920
then you would take the final five

00:36:02.119 --> 00:36:07.520
checkpoints and you would just average

00:36:03.920 --> 00:36:09.280
them together and this helps reduce some

00:36:07.520 --> 00:36:11.040
of the noise that you get from

00:36:09.280 --> 00:36:13.839
stochastic gradient descent and can

00:36:11.040 --> 00:36:15.520
improve your overall accuracy if you're

00:36:13.839 --> 00:36:17.280
fine-tuning any models this is something

00:36:15.520 --> 00:36:18.680
you can do also uh because you're

00:36:17.280 --> 00:36:19.800
probably going to be saving checkpoints

00:36:18.680 --> 00:36:21.160
you can just take the best five

00:36:19.800 --> 00:36:23.079
checkpoints and average them together

00:36:21.160 --> 00:36:27.280
and that actually can improve your

00:36:23.079 --> 00:36:28.160
accuracy quite a bit um another thing is

00:36:27.280 --> 00:36:31.520
find

00:36:28.160 --> 00:36:32.880
uh tuned model merging soine tune um in

00:36:31.520 --> 00:36:35.000
several ways and then merge them

00:36:32.880 --> 00:36:39.079
together and so for example we might

00:36:35.000 --> 00:36:41.240
take Lama 27b instruct and um vuna 7B

00:36:39.079 --> 00:36:44.760
1.5 and merg them together with some

00:36:41.240 --> 00:36:47.599
weights and uh we could you

00:36:44.760 --> 00:36:50.319
know smooth over their idos synchr dises

00:36:47.599 --> 00:36:52.520
and get better results

00:36:50.319 --> 00:36:56.280
too

00:36:52.520 --> 00:36:56.280
cool uh any questions

00:36:56.520 --> 00:36:59.520
here

00:37:00.920 --> 00:37:03.119
oh

00:37:04.680 --> 00:37:11.920
yeah want to so I just

00:37:09.680 --> 00:37:14.079
came

00:37:11.920 --> 00:37:19.040
non I

00:37:14.079 --> 00:37:19.040
use like those different chain and

00:37:19.640 --> 00:37:23.319
just

00:37:21.160 --> 00:37:26.640
I pretty

00:37:23.319 --> 00:37:29.520
efficient because on the same model you

00:37:26.640 --> 00:37:29.520
get

00:37:35.640 --> 00:37:40.839
yeah so would this would this parameter

00:37:38.000 --> 00:37:46.119
averaging be a good method for U making

00:37:40.839 --> 00:37:49.839
a model less toxic for example the

00:37:46.119 --> 00:37:53.200
answer is a little bit trickier there I

00:37:49.839 --> 00:37:56.119
guess because um I I feel like this is

00:37:53.200 --> 00:37:58.160
good for mixing two models together so

00:37:56.119 --> 00:38:01.400
if you're mixing your

00:37:58.160 --> 00:38:03.359
like non-toxicity tuned model or your

00:38:01.400 --> 00:38:06.079
safety tuned model with the original

00:38:03.359 --> 00:38:07.520
base model that was not uh safety tuned

00:38:06.079 --> 00:38:08.800
or something like that then you might

00:38:07.520 --> 00:38:11.240
get something in the middle so you might

00:38:08.800 --> 00:38:13.319
get something that's less safe than the

00:38:11.240 --> 00:38:18.720
uh like the model that was tuned to not

00:38:13.319 --> 00:38:21.400
be toxic so it might be uh yeah I'm not

00:38:18.720 --> 00:38:23.920
sure but like let's say you let's say

00:38:21.400 --> 00:38:26.240
you have a model that somebody

00:38:23.920 --> 00:38:28.640
else did like a really good job

00:38:26.240 --> 00:38:31.359
instruction tuning for you

00:38:28.640 --> 00:38:33.640
um and anytime you start using safety

00:38:31.359 --> 00:38:35.560
tuning on it you like hurt the

00:38:33.640 --> 00:38:38.680
instruction tuning like the model gets

00:38:35.560 --> 00:38:40.560
worse I could see a world where you take

00:38:38.680 --> 00:38:43.000
the base model the same base model you

00:38:40.560 --> 00:38:45.280
take llama 27b you train like a less

00:38:43.000 --> 00:38:47.480
toxic version of llama 27d and then do

00:38:45.280 --> 00:38:51.319
parameter averaging with the like well

00:38:47.480 --> 00:38:53.160
instruction tuned model um that might

00:38:51.319 --> 00:38:55.359
work that might make something that's

00:38:53.160 --> 00:38:57.560
more safe and like not much worse

00:38:55.359 --> 00:39:01.440
instruction to so there's definitely I

00:38:57.560 --> 00:39:01.440
think creative things that you can do

00:39:01.520 --> 00:39:08.400
that um maybe I'll go directly into the

00:39:04.960 --> 00:39:11.480
methods um

00:39:08.400 --> 00:39:13.240
so uh there's a few uh recent papers on

00:39:11.480 --> 00:39:16.000
this like this method has been around

00:39:13.240 --> 00:39:17.880
for a long time since at least 1996 but

00:39:16.000 --> 00:39:20.880
uh recently people have examined it a

00:39:17.880 --> 00:39:24.800
lot in the context of uh kind of modern

00:39:20.880 --> 00:39:27.400
networks and uh this paper model soup uh

00:39:24.800 --> 00:39:29.000
examines two strategies the first one is

00:39:27.400 --> 00:39:31.400
uniform averaging where you just average

00:39:29.000 --> 00:39:33.560
all the parameters together uh like as

00:39:31.400 --> 00:39:35.480
you would expect but they also have a

00:39:33.560 --> 00:39:38.319
greedy averaging method and basically

00:39:35.480 --> 00:39:40.240
what they do here is they add one model

00:39:38.319 --> 00:39:42.119
and check if the whole like averaged

00:39:40.240 --> 00:39:43.680
model improves and then only if the

00:39:42.119 --> 00:39:45.760
whole averaged model improves do they

00:39:43.680 --> 00:39:49.040
keep that model otherwise they throw it

00:39:45.760 --> 00:39:52.960
out and then they um they don't uh use

00:39:49.040 --> 00:39:54.520
it so what they demonstrate uh this is a

00:39:52.960 --> 00:39:57.560
little bit small but basically the

00:39:54.520 --> 00:40:00.520
purple star here is uh when the use

00:39:57.560 --> 00:40:02.480
greedy averaging and then the blue

00:40:00.520 --> 00:40:05.119
circle here is when they use the uniform

00:40:02.480 --> 00:40:08.280
averaging and then green is all of the

00:40:05.119 --> 00:40:09.960
models that they they put into this

00:40:08.280 --> 00:40:12.560
average

00:40:09.960 --> 00:40:16.680
and what they found

00:40:12.560 --> 00:40:18.480
is this is average uh accuracy on image

00:40:16.680 --> 00:40:22.400
net which is the thing that they they

00:40:18.480 --> 00:40:25.160
used in deciding which models to merge

00:40:22.400 --> 00:40:26.920
in greedily and then this is on

00:40:25.160 --> 00:40:28.640
distribution shifts so this is on other

00:40:26.920 --> 00:40:31.119
data sets other than the ones they use

00:40:28.640 --> 00:40:33.040
specifically for training and what you

00:40:31.119 --> 00:40:34.720
can see is the greedy averaging method

00:40:33.040 --> 00:40:38.720
does

00:40:34.720 --> 00:40:40.839
better um than the best single model on

00:40:38.720 --> 00:40:42.319
the data set that they used to decide

00:40:40.839 --> 00:40:44.800
that greedy

00:40:42.319 --> 00:40:46.560
average the uniform average actually

00:40:44.800 --> 00:40:48.359
does worse than the best model so you

00:40:46.560 --> 00:40:50.960
would actually be better off for image

00:40:48.359 --> 00:40:52.960
net accuracy to just use the best model

00:40:50.960 --> 00:40:56.000
but it's more robust so on the

00:40:52.960 --> 00:40:57.319
distribution shift like data set it

00:40:56.000 --> 00:41:00.000
actually does better than any of them

00:40:57.319 --> 00:41:02.280
models so um you can see that there's

00:41:00.000 --> 00:41:04.720
kind of trade-offs between choosing

00:41:02.280 --> 00:41:06.480
those

00:41:04.720 --> 00:41:09.319
essentially

00:41:06.480 --> 00:41:12.040
um whoops that's a that's a typo that

00:41:09.319 --> 00:41:15.760
should be ensembling but um they also

00:41:12.040 --> 00:41:18.440
demonstrate that um averaging is

00:41:15.760 --> 00:41:22.720
correlated with ensembling so this is

00:41:18.440 --> 00:41:25.200
the um image accuracy of the parameter

00:41:22.720 --> 00:41:27.000
average model this is image not accuracy

00:41:25.200 --> 00:41:30.200
of the Ensemble so this is actually I

00:41:27.000 --> 00:41:33.720
think really interesting figure um what

00:41:30.200 --> 00:41:36.440
it shows is that there's a pretty strong

00:41:33.720 --> 00:41:38.760
correlation between the two averaging is

00:41:36.440 --> 00:41:41.400
almost never better than ensembling the

00:41:38.760 --> 00:41:44.800
two together but it's faster of course

00:41:41.400 --> 00:41:48.119
so it's better because it's faster and

00:41:44.800 --> 00:41:50.000
there are situations where the Ensemble

00:41:48.119 --> 00:41:51.680
is much better than the average model so

00:41:50.000 --> 00:41:55.720
like the average model hurts the

00:41:51.680 --> 00:41:58.560
averaging hurts um onbling does not hurt

00:41:55.720 --> 00:42:01.319
so what this shows you is parameter

00:41:58.560 --> 00:42:03.119
averaging is is safe and it nearly

00:42:01.319 --> 00:42:04.359
approximates model on samping most of

00:42:03.119 --> 00:42:06.720
the time but there are cases where it

00:42:04.359 --> 00:42:08.119
doesn't so you do need to be a little

00:42:06.720 --> 00:42:11.720
bit careful and it might hurt your

00:42:08.119 --> 00:42:11.720
accuracy in some cases

00:42:16.680 --> 00:42:21.520
yeah oh yeah sorry very good point yes

00:42:19.280 --> 00:42:21.520
it's

00:42:22.319 --> 00:42:29.119
paralel yeah

00:42:26.119 --> 00:42:29.119
this

00:42:36.480 --> 00:42:41.520
um how do you know

00:42:39.400 --> 00:42:45.720
it's

00:42:41.520 --> 00:42:48.280
particular yeah so notably all of these

00:42:45.720 --> 00:42:48.280
are

00:42:48.800 --> 00:42:52.240
initialized it's been a little while

00:42:50.800 --> 00:42:54.079
since I read this but I know all of

00:42:52.240 --> 00:42:56.520
these were initialized from a model that

00:42:54.079 --> 00:42:58.160
was already pretty good on image that

00:42:56.520 --> 00:43:01.760
and then they were tuned in different

00:42:58.160 --> 00:43:03.800
ways I guess and so this I think this

00:43:01.760 --> 00:43:05.319
might be initialized with a model that

00:43:03.800 --> 00:43:09.160
was trained on a different data set or

00:43:05.319 --> 00:43:10.160
something like that um and so they are

00:43:09.160 --> 00:43:12.480
all starting from the same

00:43:10.160 --> 00:43:14.480
initialization so parameter U

00:43:12.480 --> 00:43:16.599
permutation inv variance is not an issue

00:43:14.480 --> 00:43:19.200
there because they're starting from the

00:43:16.599 --> 00:43:23.480
pre um but despite the fact that it's

00:43:19.200 --> 00:43:26.520
not a problem there are there are cases

00:43:23.480 --> 00:43:29.119
where like averaging is detrimental

00:43:26.520 --> 00:43:29.119
compared to

00:43:32.839 --> 00:43:37.559
um okay so

00:43:42.800 --> 00:43:45.800
yeah

00:43:51.720 --> 00:43:54.720
yep

00:43:56.040 --> 00:43:59.040
y

00:44:07.079 --> 00:44:10.079
okay

00:44:26.040 --> 00:44:29.040
y

00:44:46.319 --> 00:44:52.520
yeah so that's a great question um I'll

00:44:48.240 --> 00:44:54.920
just repeat it which is um the these

00:44:52.520 --> 00:44:57.520
experiments were done on CNN's or image

00:44:54.920 --> 00:44:59.280
net like uh CNN based image that

00:44:57.520 --> 00:45:01.119
classifiers is there something different

00:44:59.280 --> 00:45:04.040
than Transformers particularly because

00:45:01.119 --> 00:45:06.240
Transformer representations tend to be

00:45:04.040 --> 00:45:09.000
uh like very concentrated in particular

00:45:06.240 --> 00:45:11.359
parts of the space that's an excellent

00:45:09.000 --> 00:45:14.040
question um what I do know is a lot of

00:45:11.359 --> 00:45:15.319
people do merge together Transformer

00:45:14.040 --> 00:45:18.319
models in fact if you look at the

00:45:15.319 --> 00:45:20.079
hugging face leaderboard there's like

00:45:18.319 --> 00:45:22.240
something and something merg together

00:45:20.079 --> 00:45:24.200
like all over the leader board and it

00:45:22.240 --> 00:45:25.960
does tend to improve accuracy so I I

00:45:24.200 --> 00:45:27.480
know it is definitely effective for

00:45:25.960 --> 00:45:28.559
Transformers

00:45:27.480 --> 00:45:32.040
however Are

00:45:28.559 --> 00:45:34.640
there specific model like parameter

00:45:32.040 --> 00:45:37.040
averaging or model merging methods that

00:45:34.640 --> 00:45:38.599
could improve accuracy by taking

00:45:37.040 --> 00:45:40.680
advantage of the fact that Transformers

00:45:38.599 --> 00:45:42.480
behaving a c certain way I think that's

00:45:40.680 --> 00:45:44.920
totally possible and you know it would

00:45:42.480 --> 00:45:48.800
be an interesting research Direction um

00:45:44.920 --> 00:45:51.680
I'm not familiar enough with that

00:45:48.800 --> 00:45:53.359
particular part myself to say oh I have

00:45:51.680 --> 00:45:55.160
this great idea that you should work on

00:45:53.359 --> 00:45:55.920
but I think if you're interested in it

00:45:55.160 --> 00:45:58.160
you

00:45:55.920 --> 00:46:00.280
definitely

00:45:58.160 --> 00:46:05.240
cool anything

00:46:00.280 --> 00:46:08.920
El okay so there's also the idea of uh

00:46:05.240 --> 00:46:12.440
task vectors and um basically task

00:46:08.920 --> 00:46:15.280
vectors here we are just merging

00:46:12.440 --> 00:46:17.280
together two models by taking the

00:46:15.280 --> 00:46:18.280
parameters of the models and averaging

00:46:17.280 --> 00:46:22.079
them

00:46:18.280 --> 00:46:24.480
together task vectors and other related

00:46:22.079 --> 00:46:26.040
works specifically take advantage of the

00:46:24.480 --> 00:46:27.640
fact that we're looking at different

00:46:26.040 --> 00:46:29.160
fine-tuned models

00:46:27.640 --> 00:46:31.480
and so these are models where we have a

00:46:29.160 --> 00:46:33.920
base model and we know that uh that we

00:46:31.480 --> 00:46:35.760
fine-tuned from this base model and the

00:46:33.920 --> 00:46:38.480
basic idea is that we have our base

00:46:35.760 --> 00:46:40.319
model here and the task Vector is the

00:46:38.480 --> 00:46:43.280
difference between the base models

00:46:40.319 --> 00:46:45.559
Vector uh parameters and the uh fine

00:46:43.280 --> 00:46:49.480
tune models parameters so that's what

00:46:45.559 --> 00:46:52.720
they Define as a task Vector um what

00:46:49.480 --> 00:46:56.000
does this allow us to do this allows us

00:46:52.720 --> 00:46:58.040
to do a number of interesting things um

00:46:56.000 --> 00:47:02.359
the first one

00:46:58.040 --> 00:47:05.119
is that we can actually subtract out uh

00:47:02.359 --> 00:47:08.960
quote unquote tasks that we don't want

00:47:05.119 --> 00:47:11.559
so like let's say we had a model that

00:47:08.960 --> 00:47:13.440
was trained on lots of toxic text or we

00:47:11.559 --> 00:47:15.760
had a model that was trained on lots of

00:47:13.440 --> 00:47:18.760
private text or something like that we

00:47:15.760 --> 00:47:22.040
could actually subtract out the task

00:47:18.760 --> 00:47:24.240
Vector from this and basically attempt

00:47:22.040 --> 00:47:27.480
to remove the model's ability to uh do

00:47:24.240 --> 00:47:31.240
that sort of things um you can also

00:47:27.480 --> 00:47:36.040
take two task vectors and combine them

00:47:31.240 --> 00:47:39.280
together and uh like get the model uh

00:47:36.040 --> 00:47:42.200
from the combination of the two um this

00:47:39.280 --> 00:47:44.280
isn't exactly the same as averaging the

00:47:42.200 --> 00:47:45.440
parameters because if you average the

00:47:44.280 --> 00:47:47.400
parameters you would probably get

00:47:45.440 --> 00:47:49.160
something in the middle right here but

00:47:47.400 --> 00:47:50.440
if you average the two vectors or add

00:47:49.160 --> 00:47:52.040
the two vectors together you would get

00:47:50.440 --> 00:47:53.760
something over here actually sorry if

00:47:52.040 --> 00:47:56.520
you average the vectors maybe it's the

00:47:53.760 --> 00:47:58.119
same so you could like add together the

00:47:56.520 --> 00:47:59.480
two vectors and and that would be

00:47:58.119 --> 00:48:01.640
something different than taking the

00:47:59.480 --> 00:48:05.280
average so it gives you a little bit

00:48:01.640 --> 00:48:07.720
more flexibility about things to do

00:48:05.280 --> 00:48:09.599
um and another thing this allows you to

00:48:07.720 --> 00:48:12.920
do is this allows you to try to resolve

00:48:09.599 --> 00:48:15.400
conflicts between um vectors of

00:48:12.920 --> 00:48:19.720
different tasks and so this is an

00:48:15.400 --> 00:48:22.480
illustration of of this method here

00:48:19.720 --> 00:48:25.680
and this has three tasks basically it

00:48:22.480 --> 00:48:27.720
has model one model two model three and

00:48:25.680 --> 00:48:29.920
each of them has vectors and you'll see

00:48:27.720 --> 00:48:32.880
that in some cases these vectors

00:48:29.920 --> 00:48:34.599
conflict so we have like pink going up

00:48:32.880 --> 00:48:36.079
we have yellow and purple going down we

00:48:34.599 --> 00:48:37.800
have yellow going up we have pink and

00:48:36.079 --> 00:48:40.720
purple going down etc

00:48:37.800 --> 00:48:43.040
etc and what this does is this

00:48:40.720 --> 00:48:45.960
identifies the vectors that are uh

00:48:43.040 --> 00:48:48.040
pointing the most strongly in particular

00:48:45.960 --> 00:48:50.440
directions and then it resolves

00:48:48.040 --> 00:48:52.240
conflicts between them and comes up with

00:48:50.440 --> 00:48:54.559
a vector that tries to move in a

00:48:52.240 --> 00:48:55.920
direction that improves all of the tasks

00:48:54.559 --> 00:48:59.319
at the same time and they demonstrate

00:48:55.920 --> 00:49:01.480
that this is better method for um kind

00:48:59.319 --> 00:49:04.599
of improving the ability to do all of

00:49:01.480 --> 00:49:09.599
the tasks compared to just averaging

00:49:04.599 --> 00:49:09.599
things together so yeah first

00:49:11.920 --> 00:49:15.559
exle like it just

00:49:16.880 --> 00:49:23.640
add yeah so this is

00:49:20.680 --> 00:49:25.760
um yeah you could move it more in that

00:49:23.640 --> 00:49:27.319
direction it there's obviously no

00:49:25.760 --> 00:49:29.720
guarantee that it would make it better

00:49:27.319 --> 00:49:32.319
but it might make it more extreme at

00:49:29.720 --> 00:49:35.760
least so uh

00:49:32.319 --> 00:49:35.760
yeah any other

00:49:36.680 --> 00:49:39.960
questions all

00:49:55.640 --> 00:49:58.640
yes

00:50:25.640 --> 00:50:28.640
one

00:50:32.319 --> 00:50:37.240
yeah yeah so this is a a great question

00:50:35.599 --> 00:50:38.760
um I can explain a little bit I'm not

00:50:37.240 --> 00:50:40.760
going to talk about Metal learning

00:50:38.760 --> 00:50:42.680
extensively in this class but just to

00:50:40.760 --> 00:50:46.040
give a very quick primer for people who

00:50:42.680 --> 00:50:46.040
don't know about it

00:50:55.640 --> 00:50:58.640
um

00:51:00.359 --> 00:51:06.040
this is an example of a paper on metal

00:51:03.319 --> 00:51:09.559
learning for low resource machine

00:51:06.040 --> 00:51:12.680
translation um I you can take a look at

00:51:09.559 --> 00:51:16.200
this paper um or not take a look at this

00:51:12.680 --> 00:51:17.760
paper um uh but the reason why I wanted

00:51:16.200 --> 00:51:20.799
to look at this paper is because it has

00:51:17.760 --> 00:51:25.160
a good um uh it has a good illustration

00:51:20.799 --> 00:51:27.200
of what metal learning is and basically

00:51:25.160 --> 00:51:29.160
um if we

00:51:27.200 --> 00:51:33.839
are doing transfer learning from a

00:51:29.160 --> 00:51:35.880
single task what we do is we have like a

00:51:33.839 --> 00:51:37.960
Spanish English machine translation

00:51:35.880 --> 00:51:41.839
system and then we fine-tune it to try

00:51:37.960 --> 00:51:45.280
to hit like to try to be a good Romanian

00:51:41.839 --> 00:51:48.680
uh English or latan English system if

00:51:45.280 --> 00:51:50.400
we're doing multitask learning um or

00:51:48.680 --> 00:51:53.079
which also could be equivalent to like

00:51:50.400 --> 00:51:55.680
instruction tuning for example we have

00:51:53.079 --> 00:51:57.680
uh French uh Spanish and Portuguese we

00:51:55.680 --> 00:52:03.319
train on all the then we

00:51:57.680 --> 00:52:06.520
fine-tune to uh to be a good Romanian uh

00:52:03.319 --> 00:52:09.240
translator latan trans uh

00:52:06.520 --> 00:52:10.760
translator whereas metal learning what

00:52:09.240 --> 00:52:12.119
it's trying to do is it's trying to

00:52:10.760 --> 00:52:14.680
learn a good

00:52:12.119 --> 00:52:17.480
initialization that makes it easy to

00:52:14.680 --> 00:52:21.280
fine-tune to try to come up with a model

00:52:17.480 --> 00:52:23.839
that is good uh for fine-tuning into new

00:52:21.280 --> 00:52:29.040
tasks

00:52:23.839 --> 00:52:32.200
um the way you do this is basically um

00:52:29.040 --> 00:52:36.599
you have two

00:52:32.200 --> 00:52:39.400
steps um of gradient descent and so you

00:52:36.599 --> 00:52:42.400
have a first step where you uh train the

00:52:39.400 --> 00:52:42.400
model

00:52:42.599 --> 00:52:50.160
um where you have an update on like data

00:52:47.119 --> 00:52:50.160
from French for

00:52:55.440 --> 00:53:02.400
example

00:52:57.920 --> 00:53:02.400
and then you have another

00:53:04.640 --> 00:53:10.599
update um where you train on like black

00:53:07.880 --> 00:53:10.599
or something like

00:53:12.559 --> 00:53:17.040
this and this is a very informal very

00:53:15.599 --> 00:53:18.200
informal description there's a lot of

00:53:17.040 --> 00:53:19.599
stuff we could talk about here I could

00:53:18.200 --> 00:53:22.119
have a whole class on this but we're not

00:53:19.599 --> 00:53:27.200
going to um I don't have one planned at

00:53:22.119 --> 00:53:28.559
the moment um and so you uh you up once

00:53:27.200 --> 00:53:30.319
and then you update again and you

00:53:28.559 --> 00:53:33.400
differentiate through this update

00:53:30.319 --> 00:53:35.160
process uh so that this becomes like

00:53:33.400 --> 00:53:37.440
essentially a good initialization for

00:53:35.160 --> 00:53:40.640
training on other languages or for other

00:53:37.440 --> 00:53:43.000
tasks or things like that

00:53:40.640 --> 00:53:44.920
um now going back to the original

00:53:43.000 --> 00:53:46.240
question the original question is is

00:53:44.920 --> 00:53:50.000
there a connection between metal

00:53:46.240 --> 00:53:50.000
learning in these uh task

00:53:54.720 --> 00:53:58.440
vectors I'm not

00:53:59.079 --> 00:54:03.720
100% sure about that because I think

00:54:01.760 --> 00:54:06.599
these test backs are generally created

00:54:03.720 --> 00:54:08.480
post Haw and so they're not like there's

00:54:06.599 --> 00:54:12.680
no explicit learning step to try to make

00:54:08.480 --> 00:54:14.440
them uh you know generalize well um one

00:54:12.680 --> 00:54:15.960
one thing that maybe might be

00:54:14.440 --> 00:54:18.559
interesting to people this is a paper

00:54:15.960 --> 00:54:23.040
that we like literally just put on

00:54:18.559 --> 00:54:23.040
archive about last week

00:54:25.359 --> 00:54:28.359
um

00:54:34.520 --> 00:54:39.880
and we didn't actually use metal

00:54:36.400 --> 00:54:41.960
learning in this uh in this paper um

00:54:39.880 --> 00:54:44.520
just because metal learning actually is

00:54:41.960 --> 00:54:46.160
hard to implement uh because you need to

00:54:44.520 --> 00:54:48.680
do this kind of double differentiation

00:54:46.160 --> 00:54:50.720
and can become very very expensive for

00:54:48.680 --> 00:54:52.839
large models but we did something a

00:54:50.720 --> 00:54:55.920
little bit motivated by

00:54:52.839 --> 00:54:59.680
um uh by metal learning and what we did

00:54:55.920 --> 00:55:01.280
is we took a pre-trained LM and normally

00:54:59.680 --> 00:55:04.359
what you do is something like continued

00:55:01.280 --> 00:55:06.799
pre-training on new documents to learn

00:55:04.359 --> 00:55:10.160
knowledge from the new documents or

00:55:06.799 --> 00:55:12.200
maybe um instruction tuning including

00:55:10.160 --> 00:55:15.960
instruction tuning on data on documents

00:55:12.200 --> 00:55:17.520
about the kind of uh data that you would

00:55:15.960 --> 00:55:18.880
want to be answering questions about so

00:55:17.520 --> 00:55:20.640
like let's say you're trying to train a

00:55:18.880 --> 00:55:23.000
medical language model you might train

00:55:20.640 --> 00:55:26.680
on lots of medical documents but what we

00:55:23.000 --> 00:55:29.839
did here is we had a step where we train

00:55:26.680 --> 00:55:33.720
in advance to

00:55:29.839 --> 00:55:38.079
get on question answer Pairs and

00:55:33.720 --> 00:55:40.400
documents from another domain and then

00:55:38.079 --> 00:55:43.359
we have a step after that where we train

00:55:40.400 --> 00:55:46.400
on documents from the domain we want to

00:55:43.359 --> 00:55:48.400
answer on so like we might train on

00:55:46.400 --> 00:55:51.079
Wikipedia question answer Pairs and

00:55:48.400 --> 00:55:52.559
Wikipedia documents and then in the

00:55:51.079 --> 00:55:54.079
second step we would train on medical

00:55:52.559 --> 00:55:56.680
documents and we demonstrate that

00:55:54.079 --> 00:55:58.880
basically this allows the model to do a

00:55:56.680 --> 00:56:00.880
better job of question answering over

00:55:58.880 --> 00:56:03.640
these uh documents that we find tune on

00:56:00.880 --> 00:56:05.000
over here and so kind of going back to

00:56:03.640 --> 00:56:06.760
the metal learning paper that I talked

00:56:05.000 --> 00:56:08.359
about before the metal learning paper

00:56:06.760 --> 00:56:10.640
tries to get the parameters in a good

00:56:08.359 --> 00:56:12.559
space so that after you find ton on

00:56:10.640 --> 00:56:15.520
another data set you do a good job of

00:56:12.559 --> 00:56:17.799
that in this paper our motivation is

00:56:15.520 --> 00:56:20.359
that the model kind of learns that when

00:56:17.799 --> 00:56:22.039
you train on documents you should be

00:56:20.359 --> 00:56:24.079
able to answer questions about those

00:56:22.039 --> 00:56:25.480
documents and so when you get a new set

00:56:24.079 --> 00:56:27.200
of documents it's kind of in a good part

00:56:25.480 --> 00:56:31.079
of the parameter space to make that easy

00:56:27.200 --> 00:56:33.520
to do so um if that if metal learning is

00:56:31.079 --> 00:56:34.640
interesting um there are tutorials on

00:56:33.520 --> 00:56:37.119
metal learning that I could probably

00:56:34.640 --> 00:56:39.599
share and then um if you're interested

00:56:37.119 --> 00:56:42.599
in kind of like learning Knowledge from

00:56:39.599 --> 00:56:45.039
uh learning Knowledge

00:56:42.599 --> 00:56:46.079
from continued pre-training or something

00:56:45.039 --> 00:56:47.400
like that you could take a look at this

00:56:46.079 --> 00:56:49.920
right there as

00:56:47.400 --> 00:56:54.480
well uh

00:56:49.920 --> 00:56:54.480
cool any questions about that

00:56:55.240 --> 00:57:00.880
or

00:56:57.599 --> 00:57:02.480
okay cool I I'll jump on this so anyway

00:57:00.880 --> 00:57:05.520
um I talked about several methods for

00:57:02.480 --> 00:57:07.520
merging models together um there's a

00:57:05.520 --> 00:57:09.440
popular toolkit called merge kit that

00:57:07.520 --> 00:57:10.960
makes it relatively easy to do this it

00:57:09.440 --> 00:57:13.280
implements a lot of the models that I

00:57:10.960 --> 00:57:17.160
talked about here including uh the

00:57:13.280 --> 00:57:19.880
linear methods um uh the task arithmetic

00:57:17.160 --> 00:57:23.079
method and ties uh so I talked about

00:57:19.880 --> 00:57:25.480
these there is kind of like a expansion

00:57:23.079 --> 00:57:27.240
on this so if you want to merge together

00:57:25.480 --> 00:57:28.760
models it's Rel easy to do from a

00:57:27.240 --> 00:57:30.760
software standpoint as so so you can

00:57:28.760 --> 00:57:35.119
take a look at

00:57:30.760 --> 00:57:38.000
that um another really simple thing uh

00:57:35.119 --> 00:57:39.880
is uh distilling ensembles and so we

00:57:38.000 --> 00:57:43.039
already talked about distillation the

00:57:39.880 --> 00:57:45.599
idea is simple um

00:57:43.039 --> 00:57:47.680
you so parameter averaging only really

00:57:45.599 --> 00:57:49.200
works for models within the same run uh

00:57:47.680 --> 00:57:51.760
same model architecture same

00:57:49.200 --> 00:57:54.280
initialization so knowledge distillation

00:57:51.760 --> 00:57:55.559
uh trains a model to copy The Ensemble

00:57:54.280 --> 00:57:57.359
and so it tries to match the

00:57:55.559 --> 00:57:59.119
distribution over the predicted words

00:57:57.359 --> 00:58:00.760
for an

00:57:59.119 --> 00:58:05.319
on

00:58:00.760 --> 00:58:07.799
um and so this allows the model to make

00:58:05.319 --> 00:58:09.079
the same you know good predictions as

00:58:07.799 --> 00:58:11.079
The Ensemble make the same bad

00:58:09.079 --> 00:58:12.799
predictions as Ensemble it just allows

00:58:11.079 --> 00:58:14.799
you to learn more efficiently just like

00:58:12.799 --> 00:58:16.680
distillation does in general and they

00:58:14.799 --> 00:58:18.960
actually model distillation the original

00:58:16.680 --> 00:58:22.240
motivation for it when Jeff Hinton

00:58:18.960 --> 00:58:24.599
proposed it in 2015 in in this paper was

00:58:22.240 --> 00:58:25.680
to copy an ensemble now we use it for a

00:58:24.599 --> 00:58:27.039
lot of other things like in the

00:58:25.680 --> 00:58:31.160
distillation

00:58:27.039 --> 00:58:31.160
like weed the class but was the

00:58:34.119 --> 00:58:39.599
original

00:58:35.760 --> 00:58:42.640
um next I'll move on to sparse mixture

00:58:39.599 --> 00:58:44.960
of experts models and this is really

00:58:42.640 --> 00:58:47.599
important uh this is used in a lot of

00:58:44.960 --> 00:58:51.319
modern models it's allegedly used in GPD

00:58:47.599 --> 00:58:53.160
4 um and it is uh definitely used in

00:58:51.319 --> 00:58:55.280
mixl uh which is kind of one of the

00:58:53.160 --> 00:58:58.039
state-ofthe-art open models so I think

00:58:55.280 --> 00:58:58.039
it's a good thing to know

00:58:59.880 --> 00:59:05.720
um what these do is they take advantage

00:59:02.680 --> 00:59:08.160
of sparse computation so if you think

00:59:05.720 --> 00:59:09.359
about what happens when you do a scalar

00:59:08.160 --> 00:59:12.760
tensor

00:59:09.359 --> 00:59:14.720
multiply where the scaler is zero and

00:59:12.760 --> 00:59:17.160
basically the result of the entire

00:59:14.720 --> 00:59:19.680
resulting tensor is guaranteed to be

00:59:17.160 --> 00:59:21.440
zero and so you don't even need to do

00:59:19.680 --> 00:59:25.440
the computation you don't need to even

00:59:21.440 --> 00:59:27.520
bother um and so this manifests itself

00:59:25.440 --> 00:59:30.240
in a bunch of different places in modern

00:59:27.520 --> 00:59:35.000
models um the first one could be single

00:59:30.240 --> 00:59:38.400
rows in a matrix multiply so um if you

00:59:35.000 --> 00:59:40.480
have a big Matrix multiply like

00:59:38.400 --> 00:59:44.240
this

00:59:40.480 --> 00:59:47.880
um or Matrix Vector multiply like this

00:59:44.240 --> 00:59:50.200
um and some of the rows are zero then uh

00:59:47.880 --> 00:59:54.559
that that's one place where it

00:59:50.200 --> 00:59:58.200
happens um you can also uh do this

00:59:54.559 --> 01:00:00.119
between zero and in not just rows but

00:59:58.200 --> 01:00:02.200
also larger

01:00:00.119 --> 01:00:05.799
tensors um and you can even do it in

01:00:02.200 --> 01:00:07.599
whole models in an ensemble so um the

01:00:05.799 --> 01:00:10.799
first one this can be optimized

01:00:07.599 --> 01:00:13.880
automatically by GPU um the second one

01:00:10.799 --> 01:00:15.400
this often occurs in uh sparse mixture

01:00:13.880 --> 01:00:18.000
of experts

01:00:15.400 --> 01:00:19.400
models and the final one uh basically

01:00:18.000 --> 01:00:21.880
you just don't need to even use the

01:00:19.400 --> 01:00:24.119
model in emble so if you somehow

01:00:21.880 --> 01:00:25.640
optimize an ensemble and it turns out

01:00:24.119 --> 01:00:27.599
that the probability of one of the

01:00:25.640 --> 01:00:29.680
models is zero you just can throw it out

01:00:27.599 --> 01:00:33.640
and not use it at

01:00:29.680 --> 01:00:36.839
all so um GPU level sparsity

01:00:33.640 --> 01:00:39.839
support uh Nvidia gpus support a bunch

01:00:36.839 --> 01:00:42.559
of different types of sparsity and uh

01:00:39.839 --> 01:00:44.599
the people the wonderful people at

01:00:42.559 --> 01:00:48.280
Nvidia have worked hard to make the

01:00:44.599 --> 01:00:51.319
support uh work to some extent anyway

01:00:48.280 --> 01:00:53.119
and uh there's a library called cpar and

01:00:51.319 --> 01:00:56.119
this is used in pytorch and all these

01:00:53.119 --> 01:00:58.280
other things as well and just to give

01:00:56.119 --> 01:01:01.240
example a vector Matrix multiply with a

01:00:58.280 --> 01:01:03.240
sparse Vector um such as one that comes

01:01:01.240 --> 01:01:06.160
from a relu activation basically what

01:01:03.240 --> 01:01:09.319
happens is let's say you only have three

01:01:06.160 --> 01:01:11.799
uh parts of this Vector that are active

01:01:09.319 --> 01:01:15.240
um you actually just don't need to cop

01:01:11.799 --> 01:01:18.200
uh calculate any of the columns here so

01:01:15.240 --> 01:01:19.720
that makes your life relatively

01:01:18.200 --> 01:01:22.880
easy

01:01:19.720 --> 01:01:24.480
um but the specific thing that I wanted

01:01:22.880 --> 01:01:26.640
to talk about is a sparsely gated

01:01:24.480 --> 01:01:29.799
mixture of experts layer because this is

01:01:26.640 --> 01:01:33.960
uh what is used in mixol and probably uh

01:01:29.799 --> 01:01:38.200
the GPT models as well and what you do

01:01:33.960 --> 01:01:41.760
is you have a feed forward Network and

01:01:38.200 --> 01:01:41.760
normally a feed forward Network in a

01:01:43.640 --> 01:01:52.119
Transformer is this like really wide

01:01:49.319 --> 01:01:57.240
thing this huge wide feed forward

01:01:52.119 --> 01:01:59.359
Network um that you use to extract a

01:01:57.240 --> 01:02:00.520
whole bunch of features at each layer

01:01:59.359 --> 01:02:02.640
and that's where a lot of the

01:02:00.520 --> 01:02:05.799
computation and Transformer

01:02:02.640 --> 01:02:10.079
happens um and what sparsely gated

01:02:05.799 --> 01:02:13.079
mixture of uh experts layers do is they

01:02:10.079 --> 01:02:15.640
first have this gating Network here

01:02:13.079 --> 01:02:17.880
where it calculates uh mixture

01:02:15.640 --> 01:02:21.119
probability but the mixture probability

01:02:17.880 --> 01:02:23.039
is zero and for many or most of the

01:02:21.119 --> 01:02:26.880
parts of this feed forward

01:02:23.039 --> 01:02:28.760
Network and so for the ones where it's

01:02:26.880 --> 01:02:31.319
zero you just don't calculate

01:02:28.760 --> 01:02:34.319
it um and then when you mix them

01:02:31.319 --> 01:02:37.359
together you use the mixture rates and

01:02:34.319 --> 01:02:39.520
this is actually really simple um it's

01:02:37.359 --> 01:02:42.400
like several lines of pytorch code maybe

01:02:39.520 --> 01:02:45.319
like seven or eight lines of P torch

01:02:42.400 --> 01:02:48.720
code but the basic uh idea here is you

01:02:45.319 --> 01:02:50.599
have um this gating function where you

01:02:48.720 --> 01:02:52.799
calculate the gating function based on

01:02:50.599 --> 01:02:53.640
the input and then you have this keep

01:02:52.799 --> 01:02:56.720
top

01:02:53.640 --> 01:02:58.319
K uh operation and then you take the

01:02:56.720 --> 01:03:02.559
soft Max over

01:02:58.319 --> 01:03:04.359
this and the keep top K operation is if

01:03:02.559 --> 01:03:06.160
the value is within the top K you just

01:03:04.359 --> 01:03:07.319
keep it and if it's not in the top K you

01:03:06.160 --> 01:03:11.960
don't keep

01:03:07.319 --> 01:03:13.119
it so that that's all basically but what

01:03:11.960 --> 01:03:14.760
what's great about this is then you

01:03:13.119 --> 01:03:17.799
don't have to calculate like many of

01:03:14.760 --> 01:03:20.119
them and so for example um uh if you

01:03:17.799 --> 01:03:22.640
keep the top two out of eight you reduce

01:03:20.119 --> 01:03:26.760
your calcul uh your computation by four

01:03:22.640 --> 01:03:30.000
times for this part so

01:03:26.760 --> 01:03:33.000
um any any questions

01:03:30.000 --> 01:03:33.000
here

01:03:54.720 --> 01:03:57.720
yeah

01:04:03.160 --> 01:04:07.039
um sorry what what exactly do you mean

01:04:05.559 --> 01:04:09.400
by easy to paralyze are you talking

01:04:07.039 --> 01:04:12.400
about like a GPU can calculate lots of

01:04:09.400 --> 01:04:15.680
things at the same time yeah so I think

01:04:12.400 --> 01:04:17.720
if you have a very small model um you're

01:04:15.680 --> 01:04:21.680
actually not going to get as much from

01:04:17.720 --> 01:04:25.079
this uh because you're not you're

01:04:21.680 --> 01:04:26.359
essentially not bound by computation uh

01:04:25.079 --> 01:04:27.880
like you're bound more by memory

01:04:26.359 --> 01:04:29.079
movement and the GPU and other stuff

01:04:27.880 --> 01:04:30.520
like that but once you start getting up

01:04:29.079 --> 01:04:32.920
to the bigger models you actually are

01:04:30.520 --> 01:04:34.640
bound by computation so reducing your

01:04:32.920 --> 01:04:37.039
computation by four actually is a big

01:04:34.640 --> 01:04:42.559
one so it's a really really good

01:04:37.039 --> 01:04:42.559
question um any any other questions

01:04:44.039 --> 01:04:50.520
yeah so so this will

01:04:48.240 --> 01:04:53.160
um probably

01:04:50.520 --> 01:04:56.039
be

01:04:53.160 --> 01:04:59.279
just oh sorry I I don't have this here

01:04:56.039 --> 01:05:01.760
but this will be a often a linear layer

01:04:59.279 --> 01:05:01.760
followed by a

01:05:03.039 --> 01:05:08.000
seance um or or actually no it doesn't

01:05:06.359 --> 01:05:10.520
even need to be followed by softb it

01:05:08.000 --> 01:05:10.520
could just be a

01:05:12.520 --> 01:05:17.920
linear and I think actually I didn't put

01:05:14.960 --> 01:05:19.680
it on this slide but I have the in the

01:05:17.920 --> 01:05:21.359
references on the website I have the

01:05:19.680 --> 01:05:22.760
actual implementation in mix roll you

01:05:21.359 --> 01:05:25.279
can go in and look at it it's really

01:05:22.760 --> 01:05:27.160
simple um one thing I didn't put on here

01:05:25.279 --> 01:05:31.000
um which actually uh relates to the

01:05:27.160 --> 01:05:32.920
question before is Hardware wise this

01:05:31.000 --> 01:05:34.799
implementation is tricky if you do

01:05:32.920 --> 01:05:37.599
batching um and the reason why It's

01:05:34.799 --> 01:05:39.480
Tricky if you do batching is because um

01:05:37.599 --> 01:05:43.000
different experts will be active for

01:05:39.480 --> 01:05:45.240
different like parts of the batch so if

01:05:43.000 --> 01:05:48.559
you do that you need to do some tricky

01:05:45.240 --> 01:05:48.559
stuff uh there's

01:05:54.640 --> 01:05:57.640
this

01:06:03.240 --> 01:06:12.039
like so much of AI research nowadays uh

01:06:08.200 --> 01:06:12.039
the best resource for this is social

01:06:13.680 --> 01:06:20.000
media so this is uh there's a kind of

01:06:16.880 --> 01:06:23.240
interesting discussion of

01:06:20.000 --> 01:06:25.359
this um if you search for like gpk Fast

01:06:23.240 --> 01:06:28.400
mixed r on Twitter it it talks about

01:06:25.359 --> 01:06:30.200
this but basically there's a bunch of uh

01:06:28.400 --> 01:06:32.680
little little things you need to pay

01:06:30.200 --> 01:06:34.760
attention to um and ways that you can do

01:06:32.680 --> 01:06:36.960
tricks to make this work fast on GPU

01:06:34.760 --> 01:06:40.000
which also kind of uh addresses the

01:06:36.960 --> 01:06:42.359
concern so you can look for Horus H's

01:06:40.000 --> 01:06:44.200
discussion

01:06:42.359 --> 01:06:46.680
this

01:06:44.200 --> 01:06:49.000
cool

01:06:46.680 --> 01:06:50.799
um so the final thing I'd like to talk

01:06:49.000 --> 01:06:52.480
about in the last 10 minutes is pipeline

01:06:50.799 --> 01:06:55.359
systems

01:06:52.480 --> 01:06:57.039
um and pipeline systems are systems

01:06:55.359 --> 01:07:00.279
where we

01:06:57.039 --> 01:07:02.319
have models that basically the output of

01:07:00.279 --> 01:07:05.319
one model becomes the input of another

01:07:02.319 --> 01:07:05.319
model

01:07:05.599 --> 01:07:10.359
and to give an example of this a

01:07:08.200 --> 01:07:13.480
cascaded system is basically a system

01:07:10.359 --> 01:07:15.119
like this where you uh take the output

01:07:13.480 --> 01:07:16.960
of one system and then you feed it into

01:07:15.119 --> 01:07:19.640
the input of another system so a very

01:07:16.960 --> 01:07:22.880
stereotypical example of This is speech

01:07:19.640 --> 01:07:25.559
translation um where you run speech and

01:07:22.880 --> 01:07:27.720
then you uh do speech recognition into

01:07:25.559 --> 01:07:29.319
text and then text you do machine

01:07:27.720 --> 01:07:32.160
translation into another

01:07:29.319 --> 01:07:33.920
language

01:07:32.160 --> 01:07:36.440
and

01:07:33.920 --> 01:07:39.039
um one of the frustrating things about

01:07:36.440 --> 01:07:43.000
speech translation is these systems are

01:07:39.039 --> 01:07:45.799
stubbornly better uh for a long time

01:07:43.000 --> 01:07:47.680
than many systems that try to do end to

01:07:45.799 --> 01:07:49.960
end like speech to text in another

01:07:47.680 --> 01:07:52.160
language there's a couple reasons for

01:07:49.960 --> 01:07:54.440
this does anyone have an idea why what

01:07:52.160 --> 01:07:57.039
one of those reasons might

01:07:54.440 --> 01:07:58.839
be

01:07:57.039 --> 01:08:01.559
yeah the

01:07:58.839 --> 01:08:05.279
data

01:08:01.559 --> 01:08:08.680
anying exactly so data data availability

01:08:05.279 --> 01:08:10.920
is way better for speech to text in the

01:08:08.680 --> 01:08:13.319
same language and text to text in

01:08:10.920 --> 01:08:15.720
another language than it is for uh

01:08:13.319 --> 01:08:17.759
Speech to te text in another language

01:08:15.720 --> 01:08:19.319
because there just aren't large data

01:08:17.759 --> 01:08:21.679
sets that have speech and text in many

01:08:19.319 --> 01:08:25.719
languages so there's a bunch of tricks

01:08:21.679 --> 01:08:31.759
that you can do uh to you know fix this

01:08:25.719 --> 01:08:34.239
but still it it's uh you know uh tricky

01:08:31.759 --> 01:08:36.120
and there's a couple other reasons

01:08:34.239 --> 01:08:38.159
another reason is like actually speech

01:08:36.120 --> 01:08:39.319
to text in the same language is just a

01:08:38.159 --> 01:08:42.520
much more

01:08:39.319 --> 01:08:45.359
straightforward task um and so it's a

01:08:42.520 --> 01:08:47.839
bit easier to learn another thing is

01:08:45.359 --> 01:08:50.839
interpretability and the reason why

01:08:47.839 --> 01:08:52.120
interpretability is important is

01:08:50.839 --> 01:08:54.920
basically

01:08:52.120 --> 01:08:56.640
like if I'm talking to you in a

01:08:54.920 --> 01:08:58.000
different language like you speak a

01:08:56.640 --> 01:09:00.319
different language I'm talking to you

01:08:58.000 --> 01:09:02.679
through a speech translation system I

01:09:00.319 --> 01:09:05.799
actually want to know if the speech

01:09:02.679 --> 01:09:07.600
recognition worked because I know if the

01:09:05.799 --> 01:09:08.920
speech recognition didn't work then I'll

01:09:07.600 --> 01:09:10.440
I'm pretty sure that the translation

01:09:08.920 --> 01:09:11.920
didn't work either right and I can

01:09:10.440 --> 01:09:14.880
verify the speech recognition but I

01:09:11.920 --> 01:09:16.199
can't verify the transation so um

01:09:14.880 --> 01:09:18.279
there's other reasons why you might want

01:09:16.199 --> 01:09:20.239
a Cascade system other than just like

01:09:18.279 --> 01:09:22.440
accuracy or or other things like that

01:09:20.239 --> 01:09:25.880
but this is a thing we definitely

01:09:22.440 --> 01:09:29.120
do um there's another idea of stacking

01:09:25.880 --> 01:09:32.560
and stacking is um very similar to cast

01:09:29.120 --> 01:09:34.560
skating but it allows you to take two

01:09:32.560 --> 01:09:37.120
different models for the same task but

01:09:34.560 --> 01:09:39.400
with predictions in different ways so

01:09:37.120 --> 01:09:41.120
just taking another um

01:09:39.400 --> 01:09:43.600
example

01:09:41.120 --> 01:09:45.040
uh actually maybe maybe ignore the

01:09:43.600 --> 01:09:47.159
example I have here but we could just

01:09:45.040 --> 01:09:50.679
take the example of speech uh

01:09:47.159 --> 01:09:53.000
translation um the speech translation

01:09:50.679 --> 01:09:55.760
model uh we would first do speech

01:09:53.000 --> 01:09:57.520
recognition into like let's say English

01:09:55.760 --> 01:09:59.640
and then we would do translation and the

01:09:57.520 --> 01:10:03.840
input to the translation model would be

01:09:59.640 --> 01:10:05.560
speech in English um text in English and

01:10:03.840 --> 01:10:07.320
we would generate the output in Japanese

01:10:05.560 --> 01:10:10.080
so it would take both the speech and the

01:10:07.320 --> 01:10:12.920
text uh when it was doing translation

01:10:10.080 --> 01:10:14.840
and that would allow it to number one

01:10:12.920 --> 01:10:17.719
basically get a second opinion about

01:10:14.840 --> 01:10:21.080
whether the transcription was correct

01:10:17.719 --> 01:10:23.800
but also like let's say there was

01:10:21.080 --> 01:10:26.440
some unique information that only

01:10:23.800 --> 01:10:29.480
appeared in the

01:10:26.440 --> 01:10:31.679
um uh that only appeared in the speech

01:10:29.480 --> 01:10:34.840
so just to give an example I read the

01:10:31.679 --> 01:10:37.040
book I read the book are both

01:10:34.840 --> 01:10:38.640
transcribed exactly the same way and

01:10:37.040 --> 01:10:41.679
they're different translations obviously

01:10:38.640 --> 01:10:42.920
because one is uh you know present or

01:10:41.679 --> 01:10:45.560
present tense and the other is past

01:10:42.920 --> 01:10:47.239
tense so there are examples where uh

01:10:45.560 --> 01:10:51.600
adding a cascaded system would lose

01:10:47.239 --> 01:10:51.600
information and a stacked system would

01:10:53.400 --> 01:10:57.679
not another thing is of refinement I

01:10:56.440 --> 01:10:59.480
think this is actually really

01:10:57.679 --> 01:11:01.000
interesting because large language

01:10:59.480 --> 01:11:03.920
models have opened up a whole bunch of

01:11:01.000 --> 01:11:05.640
possibilities for us in this space um

01:11:03.920 --> 01:11:07.760
this is like cascading and stacking but

01:11:05.640 --> 01:11:09.640
it it can be done multiple times and it

01:11:07.760 --> 01:11:12.960
can be done multiple times with the same

01:11:09.640 --> 01:11:15.040
model so um we have an input we feed it

01:11:12.960 --> 01:11:17.320
into the model we get an output and then

01:11:15.040 --> 01:11:19.360
we feed the output back in and gradually

01:11:17.320 --> 01:11:23.080
refine it and make it better and

01:11:19.360 --> 01:11:24.760
better and the first time this was done

01:11:23.080 --> 01:11:27.440
in neural networks was through something

01:11:24.760 --> 01:11:29.679
called Del ation networks and basically

01:11:27.440 --> 01:11:32.360
deliberation networks what they do is

01:11:29.679 --> 01:11:33.760
they uh take in an output and then they

01:11:32.360 --> 01:11:34.920
just gradually refine it to make it

01:11:33.760 --> 01:11:37.280
better and better they used a

01:11:34.920 --> 01:11:39.159
reinforcement learning algorithm to do

01:11:37.280 --> 01:11:41.159
this where you generated the output and

01:11:39.159 --> 01:11:43.600
then um improved

01:11:41.159 --> 01:11:46.719
it another thing that's really popular

01:11:43.600 --> 01:11:48.280
nowadays is uh diffusion models and I

01:11:46.719 --> 01:11:50.400
haven't quite decided whether I'll have

01:11:48.280 --> 01:11:51.880
time to cover diffusion models in depth

01:11:50.400 --> 01:11:54.880
but basically the way a diffusion model

01:11:51.880 --> 01:11:55.880
works is very similar you start out with

01:11:54.880 --> 01:11:57.239
nothing

01:11:55.880 --> 01:11:59.840
and then you gradually make it better

01:11:57.239 --> 01:12:01.360
and better um the key difference between

01:11:59.840 --> 01:12:03.520
deliberation networks and diffusion

01:12:01.360 --> 01:12:05.520
models is diffusion models um you can

01:12:03.520 --> 01:12:08.600
train from scratch by basically noising

01:12:05.520 --> 01:12:10.600
the input uh applying noise to the input

01:12:08.600 --> 01:12:12.880
um in training very efficiently and

01:12:10.600 --> 01:12:15.639
these are very widely used

01:12:12.880 --> 01:12:18.199
in image generation they're not super

01:12:15.639 --> 01:12:20.120
widely used in text just because regular

01:12:18.199 --> 01:12:22.840
autor regressive models are so good for

01:12:20.120 --> 01:12:24.159
text um but there are a few efforts to

01:12:22.840 --> 01:12:26.880
do

01:12:24.159 --> 01:12:30.920
that and then a final one is self-

01:12:26.880 --> 01:12:35.120
refine and the idea behind self- refine

01:12:30.920 --> 01:12:39.400
is you um actually maybe I can open the

01:12:35.120 --> 01:12:39.400
paper because the paper has a good

01:12:54.120 --> 01:12:58.239
figure

01:12:56.280 --> 01:13:02.679
actually I thought it had a good

01:12:58.239 --> 01:13:05.600
figure um yeah so maybe this is a figure

01:13:02.679 --> 01:13:08.639
um so basically uh what you do is you

01:13:05.600 --> 01:13:10.639
feed in the input you generate an output

01:13:08.639 --> 01:13:12.679
and then you ask the model to give you

01:13:10.639 --> 01:13:15.520
feedback on the output and say yes this

01:13:12.679 --> 01:13:16.760
output is good or um like let's say

01:13:15.520 --> 01:13:19.679
you're doing code generation it could

01:13:16.760 --> 01:13:21.920
say no this output has an error in it um

01:13:19.679 --> 01:13:24.719
this is a problem with your output and

01:13:21.920 --> 01:13:27.840
then you feed in both the output and the

01:13:24.719 --> 01:13:29.480
feedback back uh and ask the model to

01:13:27.840 --> 01:13:32.239
refine its output and you do this over

01:13:29.480 --> 01:13:35.280
and over again and this allows you to uh

01:13:32.239 --> 01:13:36.840
improve the output and uh this is has

01:13:35.280 --> 01:13:39.600
ended up being pretty effective in a

01:13:36.840 --> 01:13:41.159
pretty wide number of tasks one caveat

01:13:39.600 --> 01:13:44.040
about this is your model has to be

01:13:41.159 --> 01:13:47.000
really good for this to work so um only

01:13:44.040 --> 01:13:49.239
models kind of on the level of GPT 4 not

01:13:47.000 --> 01:13:52.000
on the level of GPT 3.5 have the ability

01:13:49.239 --> 01:13:54.040
to do this pretty consistently so it is

01:13:52.000 --> 01:13:57.040
something you need to be aware

01:13:54.040 --> 01:13:57.040
of

01:13:59.760 --> 01:14:03.600
cool yep that's all I I had for today

01:14:02.400 --> 01:14:06.600
I'm happy

01:14:03.600 --> 01:14:06.600
to

01:14:07.159 --> 01:14:10.159
take

01:14:20.600 --> 01:14:27.320
yep yep that this is a great question so

01:14:23.920 --> 01:14:28.840
if sta has the potential to address

01:14:27.320 --> 01:14:32.120
information loss why would we ever

01:14:28.840 --> 01:14:33.840
choose a Cascade model I think basically

01:14:32.120 --> 01:14:37.440
there's potentially two reasons one

01:14:33.840 --> 01:14:39.199
reason is um data availability so in

01:14:37.440 --> 01:14:42.639
order to train a stacked model you

01:14:39.199 --> 01:14:43.430
obviously need the outputs I guess you

01:14:42.639 --> 01:14:44.639
could

01:14:43.430 --> 01:14:48.440
[Music]

01:14:44.639 --> 01:14:50.880
um yeah I guess you could run

01:14:48.440 --> 01:14:53.199
the and generate outputs for every

01:14:50.880 --> 01:14:54.840
training example you have um but you

01:14:53.199 --> 01:14:55.840
would need to do that so you would need

01:14:54.840 --> 01:14:58.639
to to

01:14:55.840 --> 01:14:59.920
run speech recognition for every example

01:14:58.639 --> 01:15:02.760
and you also

01:14:59.920 --> 01:15:05.199
couldn't you couldn't use any examples

01:15:02.760 --> 01:15:07.600
where you don't have the original input

01:15:05.199 --> 01:15:10.320
so you couldn't use text to text

01:15:07.600 --> 01:15:12.239
examples unless you like synthesize

01:15:10.320 --> 01:15:14.159
speech from text for machine translation

01:15:12.239 --> 01:15:15.840
for example so makes it a little bit

01:15:14.159 --> 01:15:17.360
more tricky due to the data requirements

01:15:15.840 --> 01:15:19.239
but that's not

01:15:17.360 --> 01:15:22.560
insurmountable the second reason is

01:15:19.239 --> 01:15:24.400
complexity and efficiency so you know

01:15:22.560 --> 01:15:27.920
you do have to come up with a model that

01:15:24.400 --> 01:15:29.520
takes in speed and text and run set and

01:15:27.920 --> 01:15:30.920
it might be easier just to hook together

01:15:29.520 --> 01:15:34.719
a speech recognitional with a

01:15:30.920 --> 01:15:37.920
translation so but like I think overall

01:15:34.719 --> 01:15:39.639
I I like these methods I I think these

01:15:37.920 --> 01:15:41.159
are good methods to use if you're if

01:15:39.639 --> 01:15:42.480
you're thinking about using a Cascade

01:15:41.159 --> 01:15:44.199
system you should definitely consider

01:15:42.480 --> 01:15:47.199
using a stack system in

01:15:44.199 --> 01:15:47.199
sense

01:15:52.080 --> 01:15:56.960
yeah yeah can you measure the

01:15:55.159 --> 01:15:59.400
contribution of each component to an

01:15:56.960 --> 01:16:00.639
ensemble um the very very easy way to do

01:15:59.400 --> 01:16:02.199
that is look at the interpolation

01:16:00.639 --> 01:16:05.360
coefficients if you train the

01:16:02.199 --> 01:16:06.800
interpolation coefficients um otherwise

01:16:05.360 --> 01:16:08.920
I guess it depends on what you mean by

01:16:06.800 --> 01:16:10.480
each contribution but I you know looking

01:16:08.920 --> 01:16:12.280
at the interpolation coefficients is a

01:16:10.480 --> 01:16:16.320
pretty good way to do

01:16:12.280 --> 01:16:16.320
it also just how much did the

01:16:21.480 --> 01:16:27.400
accuracy is iterative refinement the

01:16:24.159 --> 01:16:30.199
same idea as boosting in traditional

01:16:27.400 --> 01:16:30.199
like machine Learning

01:16:30.320 --> 01:16:34.920
Systems I think it's a little bit

01:16:32.920 --> 01:16:36.520
different um because iterative

01:16:34.920 --> 01:16:38.920
refinement what I'm talking about here

01:16:36.520 --> 01:16:41.120
it's usually taking in the output like

01:16:38.920 --> 01:16:43.320
rather complex output of a system and

01:16:41.120 --> 01:16:44.920
modifying it so you're not just

01:16:43.320 --> 01:16:47.080
modifying the

01:16:44.920 --> 01:16:49.880
probabilities of like a single

01:16:47.080 --> 01:16:53.080
classifier you're modifying the actual

01:16:49.880 --> 01:16:55.960
outputs that were generated then from

01:16:53.080 --> 01:16:59.560
the point of view of a boosting

01:16:55.960 --> 01:17:02.560
model over a single categorical output

01:16:59.560 --> 01:17:04.520
it might actually be similar or the same

01:17:02.560 --> 01:17:06.480
but this is more like uh you you

01:17:04.520 --> 01:17:08.159
generated a textual output and then you

01:17:06.480 --> 01:17:10.400
feed in the textual output to the other

01:17:08.159 --> 01:17:12.120
model and refine like generated a new

01:17:10.400 --> 01:17:14.239
textual output so I feel like it's a lot

01:17:12.120 --> 01:17:18.639
more

01:17:14.239 --> 01:17:18.639
complex cool okay thank thanks a lot

01:17:18.840 --> 01:17:21.840
everyone