WEBVTT

00:00:00.399 --> 00:00:04.720
great um yeah so today we're going to be

00:00:03.320 --> 00:00:07.040
talking a little bit about generation

00:00:04.720 --> 00:00:08.639
algorithms um this will be sort of a

00:00:07.040 --> 00:00:10.160
tour through some of the most common

00:00:08.639 --> 00:00:12.080
methods and we're going to talk a little

00:00:10.160 --> 00:00:13.480
bit about the theory behind them as well

00:00:12.080 --> 00:00:15.080
um if you're looking at the slides on

00:00:13.480 --> 00:00:18.359
the website these might be ever so

00:00:15.080 --> 00:00:20.000
slightly different um but yeah I'll try

00:00:18.359 --> 00:00:21.640
to stop at each section boundary for

00:00:20.000 --> 00:00:23.840
questions also feel free to sort of

00:00:21.640 --> 00:00:25.720
interrupt at any point for

00:00:23.840 --> 00:00:27.720
clarifications so we're starting off

00:00:25.720 --> 00:00:29.560
today with some great news um let's say

00:00:27.720 --> 00:00:31.199
that you have some friend who maybe owns

00:00:29.560 --> 00:00:34.800
a giant tech company and they've gifted

00:00:31.199 --> 00:00:36.480
you this absolutely massive new model M

00:00:34.800 --> 00:00:38.079
um it's a great model it's pre-trained

00:00:36.480 --> 00:00:40.879
with the latest architecture it's

00:00:38.079 --> 00:00:42.920
pre-trained on um trillions of tokens of

00:00:40.879 --> 00:00:44.520
text it's got seven billion parameters

00:00:42.920 --> 00:00:46.399
it looks like a really promising new

00:00:44.520 --> 00:00:48.399
model you know it's the top of all these

00:00:46.399 --> 00:00:50.320
leaderboards um but if you actually take

00:00:48.399 --> 00:00:52.520
your new model M and you sort of open up

00:00:50.320 --> 00:00:53.719
this box and kind of Shake It Out maybe

00:00:52.520 --> 00:00:55.239
from last class you know a little bit

00:00:53.719 --> 00:00:57.000
architecturally what this model might

00:00:55.239 --> 00:00:58.239
look like but if you actually kind of

00:00:57.000 --> 00:01:00.320
take a closer look at it from a

00:00:58.239 --> 00:01:01.719
different angle what you see is that m

00:01:00.320 --> 00:01:04.920
is actually just a conditional

00:01:01.719 --> 00:01:07.200
probability distribution um you put some

00:01:04.920 --> 00:01:09.680
input X into your model and you get some

00:01:07.200 --> 00:01:10.680
probability out for any given sequence

00:01:09.680 --> 00:01:13.360
that you're sort of interested in

00:01:10.680 --> 00:01:14.960
evaluating right um and in particular M

00:01:13.360 --> 00:01:17.560
gives you a probability distribution

00:01:14.960 --> 00:01:19.439
over all tokens in its vocabulary to

00:01:17.560 --> 00:01:21.040
predict like what token you would output

00:01:19.439 --> 00:01:24.840
next right and so this is what this

00:01:21.040 --> 00:01:26.880
equation says um given some input X and

00:01:24.840 --> 00:01:29.520
everything that you've predicted so far

00:01:26.880 --> 00:01:32.399
you get the probability of the next

00:01:29.520 --> 00:01:33.600
token in YJ and if you multiply this out

00:01:32.399 --> 00:01:34.840
over all the probabilities in your

00:01:33.600 --> 00:01:37.159
sequence you can calculate the

00:01:34.840 --> 00:01:41.240
probability of any output y given your

00:01:37.159 --> 00:01:42.640
input X so what this like super fancy

00:01:41.240 --> 00:01:44.119
model that you spend a lot of money to

00:01:42.640 --> 00:01:46.280
train is really just a conditional

00:01:44.119 --> 00:01:47.920
probability distribution um but this

00:01:46.280 --> 00:01:49.600
turns out to be okay because you can use

00:01:47.920 --> 00:01:51.920
a conditional probability distribution

00:01:49.600 --> 00:01:54.399
to do sort of any task that we're really

00:01:51.920 --> 00:01:56.719
interested in in NLP um pretty much any

00:01:54.399 --> 00:01:58.680
task right so by changing what you

00:01:56.719 --> 00:02:01.360
consider your input X and your output y

00:01:58.680 --> 00:02:03.560
to be you can can get outputs from this

00:02:01.360 --> 00:02:06.479
model for things like translation for

00:02:03.560 --> 00:02:08.720
summarization for reasoning Tas um just

00:02:06.479 --> 00:02:10.520
by sort of changing what you consider

00:02:08.720 --> 00:02:12.760
your inputs and outputs in this

00:02:10.520 --> 00:02:14.239
setting but there's sort of both good

00:02:12.760 --> 00:02:15.920
and bad things about your model being a

00:02:14.239 --> 00:02:17.120
probability distribution instead of just

00:02:15.920 --> 00:02:20.599
an oracle that gives you sort of a

00:02:17.120 --> 00:02:22.080
single answer for every input um one

00:02:20.599 --> 00:02:24.480
kind of nice thing about this

00:02:22.080 --> 00:02:26.080
distribution um is that you can get at

00:02:24.480 --> 00:02:27.720
an idea of something like confidence

00:02:26.080 --> 00:02:30.120
right if you give your model the input 2

00:02:27.720 --> 00:02:32.480
plus 2 equals and almost all the

00:02:30.120 --> 00:02:34.200
probability mass is on the token of four

00:02:32.480 --> 00:02:35.760
you can say like the model predicts with

00:02:34.200 --> 00:02:38.319
pretty high confidence that 2 plus 2

00:02:35.760 --> 00:02:39.480
equals four um versus if you give it

00:02:38.319 --> 00:02:40.959
something that's maybe a little more

00:02:39.480 --> 00:02:43.120
open-ended like you ask it to predict

00:02:40.959 --> 00:02:44.640
Graham's favorite color and you see this

00:02:43.120 --> 00:02:47.040
distribution that's sort of a lot

00:02:44.640 --> 00:02:48.440
flatter you know the most likely output

00:02:47.040 --> 00:02:49.720
is green but maybe we don't have a lot

00:02:48.440 --> 00:02:51.560
of confidence that that's the correct

00:02:49.720 --> 00:02:53.040
answer um this is really closely tied

00:02:51.560 --> 00:02:55.200
into the idea of calibration which you

00:02:53.040 --> 00:02:58.879
guys talked about um I guess a couple of

00:02:55.200 --> 00:03:00.640
classes ago now the flip side of this

00:02:58.879 --> 00:03:03.680
though is that you know Noti that for

00:03:00.640 --> 00:03:06.760
this case like 2 plus 2al 4 not all of

00:03:03.680 --> 00:03:08.519
the probability mass is on four um and

00:03:06.760 --> 00:03:09.720
so models that are conditional

00:03:08.519 --> 00:03:11.560
probability distributions can

00:03:09.720 --> 00:03:13.560
hallucinate right um pretty much no

00:03:11.560 --> 00:03:15.799
matter what you do there's going to be

00:03:13.560 --> 00:03:17.680
some nonzero probability to some output

00:03:15.799 --> 00:03:19.920
that's incorrect or

00:03:17.680 --> 00:03:21.239
undesirable um in some cases maybe even

00:03:19.920 --> 00:03:23.760
offensive something that you don't want

00:03:21.239 --> 00:03:25.280
the model to Output um and this is sort

00:03:23.760 --> 00:03:27.840
of an artifact of the way these models

00:03:25.280 --> 00:03:29.280
are trained if there's some great work

00:03:27.840 --> 00:03:31.400
kind of more on the theory side here

00:03:29.280 --> 00:03:32.840
that shows that this is actually true

00:03:31.400 --> 00:03:35.120
even if everything in your input

00:03:32.840 --> 00:03:36.920
training data is sort of correct and

00:03:35.120 --> 00:03:38.439
factual and doesn't have any errors

00:03:36.920 --> 00:03:41.200
you'll still wind up with a situation

00:03:38.439 --> 00:03:44.480
where some nonzero probability mass is

00:03:41.200 --> 00:03:47.000
on some outputs that are undesirable or

00:03:44.480 --> 00:03:50.120
hallucinatory for sort of most inputs

00:03:47.000 --> 00:03:52.159
that you care about evaluating so if we

00:03:50.120 --> 00:03:55.079
have these issues how do we actually get

00:03:52.159 --> 00:03:56.519
a good output out of the model um and to

00:03:55.079 --> 00:03:58.640
do that we're first going to talk about

00:03:56.519 --> 00:04:00.079
some sampling methods um but I want to

00:03:58.640 --> 00:04:01.879
pause here in case there are of any

00:04:00.079 --> 00:04:04.159
questions on this idea of a model is a

00:04:01.879 --> 00:04:04.159
conditional

00:04:05.040 --> 00:04:11.680
distribution great so we can jump right

00:04:07.519 --> 00:04:13.560
in so we have this model right we know

00:04:11.680 --> 00:04:15.959
at each step at each token we might want

00:04:13.560 --> 00:04:17.919
to decode the distribution of likelihood

00:04:15.959 --> 00:04:18.959
over all vocabulary tokens right this

00:04:17.919 --> 00:04:21.680
conditional distribution we've been

00:04:18.959 --> 00:04:24.240
talking about um for the next time step

00:04:21.680 --> 00:04:26.400
and what we want out of this is a good

00:04:24.240 --> 00:04:28.000
output um for some definition of good

00:04:26.400 --> 00:04:30.919
that we can sort of develop as we go

00:04:28.000 --> 00:04:32.479
here so maybe the natural first thing to

00:04:30.919 --> 00:04:34.880
try is we have a probability

00:04:32.479 --> 00:04:36.600
distribution can we just sample from it

00:04:34.880 --> 00:04:39.600
right and this is something called

00:04:36.600 --> 00:04:41.639
ancestral sampling so at each time step

00:04:39.600 --> 00:04:43.560
we're going to draw a token from this

00:04:41.639 --> 00:04:45.039
distribution sort of according to its

00:04:43.560 --> 00:04:47.199
relative probability right so if

00:04:45.039 --> 00:04:48.639
something has twice as much probability

00:04:47.199 --> 00:04:51.280
Mass according to the model we'll draw

00:04:48.639 --> 00:04:54.000
it twice as often um and we can sample

00:04:51.280 --> 00:04:55.560
from this distribution at each time step

00:04:54.000 --> 00:04:58.080
and this is sort of this is sort of a

00:04:55.560 --> 00:05:00.199
nice setup um we get exact samples from

00:04:58.080 --> 00:05:02.639
the model distribution so using the

00:05:00.199 --> 00:05:04.479
setup if you can you imagine like

00:05:02.639 --> 00:05:06.680
drawing an almost infinite number of

00:05:04.479 --> 00:05:08.320
samples like a ridiculously large number

00:05:06.680 --> 00:05:10.160
and you look at their probabilities

00:05:08.320 --> 00:05:11.840
you'd sort of get something from this

00:05:10.160 --> 00:05:13.039
distribution with exactly the

00:05:11.840 --> 00:05:15.720
probability that the real model

00:05:13.039 --> 00:05:17.280
distribution is given you um so this is

00:05:15.720 --> 00:05:19.039
great this gives us an exact sample from

00:05:17.280 --> 00:05:21.400
the model this seems to be exactly what

00:05:19.039 --> 00:05:22.880
we want um but you can guess probably by

00:05:21.400 --> 00:05:24.639
the fact that we're only like 10 minutes

00:05:22.880 --> 00:05:27.000
into class here this is not really the

00:05:24.639 --> 00:05:28.280
end of the story um and there's actually

00:05:27.000 --> 00:05:30.800
a couple of problems with sampling

00:05:28.280 --> 00:05:32.560
directly from our model distribu

00:05:30.800 --> 00:05:35.280
the one that we're really going to focus

00:05:32.560 --> 00:05:37.919
on first here is this idea of a long

00:05:35.280 --> 00:05:41.400
tail so a model like llama and maybe our

00:05:37.919 --> 00:05:43.639
new model M um has 32,000 vocabulary

00:05:41.400 --> 00:05:46.280
tokens and you can imagine maybe out of

00:05:43.639 --> 00:05:48.000
those tokens there might be one or even

00:05:46.280 --> 00:05:49.720
2,000 of those tokens that are sort of a

00:05:48.000 --> 00:05:51.919
reasonable next thing to predict for a

00:05:49.720 --> 00:05:53.479
really open-ended task right but there's

00:05:51.919 --> 00:05:55.440
going to be all kinds of things in that

00:05:53.479 --> 00:05:57.039
distribution um that are maybe like

00:05:55.440 --> 00:05:58.440
punctuation there maybe tokens that

00:05:57.039 --> 00:06:00.280
won't actually lead to the correct

00:05:58.440 --> 00:06:01.840
answer like there's a lot of things in

00:06:00.280 --> 00:06:04.560
this distribution that would be all

00:06:01.840 --> 00:06:06.160
really low likelihood and this is fine

00:06:04.560 --> 00:06:08.759
these things just get low probability

00:06:06.160 --> 00:06:11.039
Mass but the problem is if you give sort

00:06:08.759 --> 00:06:13.639
of a small amount of probability Mass to

00:06:11.039 --> 00:06:16.599
30,000 different things that mass will

00:06:13.639 --> 00:06:19.360
add up pretty quickly um and to see this

00:06:16.599 --> 00:06:20.360
we have sort of this illustration here

00:06:19.360 --> 00:06:21.560
um I don't know if you can see the

00:06:20.360 --> 00:06:23.280
difference between the green and the

00:06:21.560 --> 00:06:25.720
yellow but I've also drawn a little bar

00:06:23.280 --> 00:06:27.800
between them this is a really longtailed

00:06:25.720 --> 00:06:29.720
distribution and the green part of the

00:06:27.800 --> 00:06:31.960
distribution which is a lot of tokens

00:06:29.720 --> 00:06:34.000
with high likelihood has 50% of the

00:06:31.960 --> 00:06:35.560
total probability the Yellow Part which

00:06:34.000 --> 00:06:37.360
is all a lot of things that are all

00:06:35.560 --> 00:06:40.280
individually not super likely is the

00:06:37.360 --> 00:06:41.720
other 50% of the probability and so what

00:06:40.280 --> 00:06:44.360
that means is if you're doing something

00:06:41.720 --> 00:06:46.120
like ancestral sampling 50% of the time

00:06:44.360 --> 00:06:49.160
you'll be sampling something really

00:06:46.120 --> 00:06:51.520
unlikely from this long tail um that

00:06:49.160 --> 00:06:53.759
seems sort of not like what we want

00:06:51.520 --> 00:06:56.080
right um so is there anything we can do

00:06:53.759 --> 00:06:58.080
about this and the obvious for solution

00:06:56.080 --> 00:06:59.400
here is can we just cut off that tail

00:06:58.080 --> 00:07:01.680
like if we know these tokens are not

00:06:59.400 --> 00:07:03.039
super likely can we just ignore them and

00:07:01.680 --> 00:07:05.039
there's a couple of different ways to do

00:07:03.039 --> 00:07:07.919
that um the first of these is something

00:07:05.039 --> 00:07:10.080
called topk sampling where we say okay

00:07:07.919 --> 00:07:12.479
you know maybe we think there are 10

00:07:10.080 --> 00:07:14.000
reasonable like outputs is right maybe

00:07:12.479 --> 00:07:17.280
we'll just sample from the 10 most

00:07:14.000 --> 00:07:19.759
probable tokens um here maybe we say if

00:07:17.280 --> 00:07:21.479
we want to pick top six sampling we'll

00:07:19.759 --> 00:07:23.919
sample from just the six most probable

00:07:21.479 --> 00:07:26.240
tokens and so in this example you can

00:07:23.919 --> 00:07:27.680
see we originally had 10 tokens and

00:07:26.240 --> 00:07:30.560
we're going to sample from just the blue

00:07:27.680 --> 00:07:32.919
ones just the six most likely tokens

00:07:30.560 --> 00:07:34.360
um in this example this distribution is

00:07:32.919 --> 00:07:37.280
pretty flat there's a lot of things that

00:07:34.360 --> 00:07:40.120
are like kind of likely right so that

00:07:37.280 --> 00:07:43.000
those six tokens are only 68% of the

00:07:40.120 --> 00:07:45.360
total probability Mass um if we go like

00:07:43.000 --> 00:07:47.240
one time step further here we might have

00:07:45.360 --> 00:07:49.360
a distribution that's a lot peier most

00:07:47.240 --> 00:07:51.759
of the mass is on just a single token

00:07:49.360 --> 00:07:53.919
and so sampling from just the top six

00:07:51.759 --> 00:07:56.400
tokens actually captures 99% of the

00:07:53.919 --> 00:07:58.360
probability mes maybe we say that seems

00:07:56.400 --> 00:08:01.199
a little excessive right we don't really

00:07:58.360 --> 00:08:03.400
need um maybe all of these tokens that

00:08:01.199 --> 00:08:05.479
are all kind of low probability maybe we

00:08:03.400 --> 00:08:07.000
just want to sort of sample from the top

00:08:05.479 --> 00:08:08.080
half of our distribution or something or

00:08:07.000 --> 00:08:10.840
the top

00:08:08.080 --> 00:08:12.919
90% um so instead of choosing a top

00:08:10.840 --> 00:08:15.560
number of tokens to sample from you

00:08:12.919 --> 00:08:17.400
could choose a top amount of probability

00:08:15.560 --> 00:08:20.000
and this is something called top P or

00:08:17.400 --> 00:08:21.520
nucleus sampling so P here is the amount

00:08:20.000 --> 00:08:24.039
of probability from your distribution

00:08:21.520 --> 00:08:26.639
you want to consider so if you decide

00:08:24.039 --> 00:08:29.280
your p is about like 94% of the

00:08:26.639 --> 00:08:31.639
probability Mass you in this first examp

00:08:29.280 --> 00:08:33.719
example here would choose almost all of

00:08:31.639 --> 00:08:35.440
the tokens you keep adding tokens in

00:08:33.719 --> 00:08:37.159
until you reach an amount of total

00:08:35.440 --> 00:08:39.479
probability that's about

00:08:37.159 --> 00:08:40.880
094 but then when you get to the Second

00:08:39.479 --> 00:08:43.240
Step where you have a couple of really

00:08:40.880 --> 00:08:45.959
highly probable tokens you'd only need a

00:08:43.240 --> 00:08:47.959
couple of tokens to add up to 094 or

00:08:45.959 --> 00:08:50.320
even higher than 0.94 and so you would

00:08:47.959 --> 00:08:52.200
just sample from a smaller set of tokens

00:08:50.320 --> 00:08:54.600
so in top K sampling the total amount of

00:08:52.200 --> 00:08:56.560
probability your sampling from can move

00:08:54.600 --> 00:08:58.120
around in top P sampling the total

00:08:56.560 --> 00:08:59.839
number of tokens you're sampling from

00:08:58.120 --> 00:09:01.959
might change

00:08:59.839 --> 00:09:04.760
um but maybe we sort of don't want to

00:09:01.959 --> 00:09:07.279
impose a strong constraint like we want

00:09:04.760 --> 00:09:09.279
like 94% here maybe just what we really

00:09:07.279 --> 00:09:11.040
care about is saying that we're not

00:09:09.279 --> 00:09:14.000
going to sample anything that's really

00:09:11.040 --> 00:09:16.800
really unlikely right another way of

00:09:14.000 --> 00:09:18.560
doing this is called Epsilon sampling

00:09:16.800 --> 00:09:20.519
where we just sample tokens that have at

00:09:18.560 --> 00:09:22.920
least some minimum amount of probability

00:09:20.519 --> 00:09:24.720
to them right so maybe we just want

00:09:22.920 --> 00:09:29.519
tokens that have probability of at least

00:09:24.720 --> 00:09:31.240
0.05 here um in this first um example

00:09:29.519 --> 00:09:32.640
everything has at least some reasonable

00:09:31.240 --> 00:09:34.240
amount of probability so we're actually

00:09:32.640 --> 00:09:36.240
going to sample from our full

00:09:34.240 --> 00:09:37.720
distribution and then in the second

00:09:36.240 --> 00:09:39.279
example when we have a lot of things

00:09:37.720 --> 00:09:41.160
that are really unlikely we'll only

00:09:39.279 --> 00:09:43.800
sample from sort of the more likely part

00:09:41.160 --> 00:09:45.240
of the distribution um so all three of

00:09:43.800 --> 00:09:47.000
these methods are sort of different ways

00:09:45.240 --> 00:09:49.399
of trying to cut off the long tail using

00:09:47.000 --> 00:09:51.480
sort of different

00:09:49.399 --> 00:09:53.000
characteristics the tail of the

00:09:51.480 --> 00:09:55.680
distribution though isn't the only thing

00:09:53.000 --> 00:09:58.000
we could choose to modify um we could

00:09:55.680 --> 00:09:59.880
also choose to modify this sort of

00:09:58.000 --> 00:10:02.120
peakiness of the distribution

00:09:59.880 --> 00:10:03.880
so if you look here at the middle of

00:10:02.120 --> 00:10:06.600
these diagrams say this is your original

00:10:03.880 --> 00:10:08.519
distribution over next tokens and maybe

00:10:06.600 --> 00:10:11.040
you want to modify some properties of

00:10:08.519 --> 00:10:12.640
this distribution like you say I want an

00:10:11.040 --> 00:10:14.200
output that's really diverse and

00:10:12.640 --> 00:10:15.680
interesting and open-ended like maybe

00:10:14.200 --> 00:10:17.920
this is something like story generation

00:10:15.680 --> 00:10:20.120
where you want to have sort of a lot of

00:10:17.920 --> 00:10:21.279
maybe surprising things in your output

00:10:20.120 --> 00:10:23.480
you could say I want to sort of

00:10:21.279 --> 00:10:26.440
distribute my probability Mass more over

00:10:23.480 --> 00:10:28.399
the token space and you can do this um

00:10:26.440 --> 00:10:32.720
by sort of flattening this distribution

00:10:28.399 --> 00:10:34.240
like you see on the the right here um

00:10:32.720 --> 00:10:36.800
where now there's sort of more

00:10:34.240 --> 00:10:39.040
probability Mass spread over this um

00:10:36.800 --> 00:10:40.320
like wider set of tokens you could also

00:10:39.040 --> 00:10:42.720
say the opposite right you could say

00:10:40.320 --> 00:10:44.120
maybe I'm doing something like math

00:10:42.720 --> 00:10:45.519
where there shouldn't really be a lot of

00:10:44.120 --> 00:10:47.800
correct answers there should be really

00:10:45.519 --> 00:10:50.399
only one or maybe only like a few

00:10:47.800 --> 00:10:52.320
potential reasonable next answers and so

00:10:50.399 --> 00:10:54.160
you can make your distribution peier or

00:10:52.320 --> 00:10:56.639
sharper so that more of the probability

00:10:54.160 --> 00:11:00.200
mass is on the things at the very top um

00:10:56.639 --> 00:11:02.000
the way you do this is you modify y your

00:11:00.200 --> 00:11:04.320
loges your outputs of the last layer of

00:11:02.000 --> 00:11:06.399
the model before you apply softn so when

00:11:04.320 --> 00:11:08.360
you're predicting you get your outputs

00:11:06.399 --> 00:11:10.040
of the last layer of the model and then

00:11:08.360 --> 00:11:11.560
you apply softmax which turns those

00:11:10.040 --> 00:11:15.240
outputs into a distribution right they

00:11:11.560 --> 00:11:17.399
all sum up the um like Mass over all

00:11:15.240 --> 00:11:18.839
vocabulary tokens sums to one and so

00:11:17.399 --> 00:11:21.920
that is sort of a distribution you could

00:11:18.839 --> 00:11:23.519
sample from if you divide those Logics

00:11:21.920 --> 00:11:26.000
by some number before you apply that

00:11:23.519 --> 00:11:27.880
softmax you can make that distribution

00:11:26.000 --> 00:11:30.760
flatter by using a number greater than

00:11:27.880 --> 00:11:32.440
one or peier by using a number less than

00:11:30.760 --> 00:11:35.079
one and this is this type of parameter

00:11:32.440 --> 00:11:36.839
is called temperature um you can apply

00:11:35.079 --> 00:11:38.480
this with any of the other methods for

00:11:36.839 --> 00:11:40.279
sort of cutting off the long tail but

00:11:38.480 --> 00:11:41.920
what people will often do is just apply

00:11:40.279 --> 00:11:43.639
a temperature and then sample from that

00:11:41.920 --> 00:11:45.320
distribution and that's what we call

00:11:43.639 --> 00:11:48.720
temperature

00:11:45.320 --> 00:11:49.920
sampling so these I think most of you

00:11:48.720 --> 00:11:51.320
might already have been at least a

00:11:49.920 --> 00:11:53.000
little bit familiar with some of these

00:11:51.320 --> 00:11:56.079
methods I want to touch briefly on a

00:11:53.000 --> 00:11:58.160
couple of other ideas for modifying this

00:11:56.079 --> 00:11:59.680
distribution maybe some more complex and

00:11:58.160 --> 00:12:01.839
more recent ideas and the one that I

00:11:59.680 --> 00:12:04.279
want to talk about in more detail is

00:12:01.839 --> 00:12:05.399
something called contrastive decoding so

00:12:04.279 --> 00:12:07.360
the idea here is that we could

00:12:05.399 --> 00:12:10.800
incorporate some extra information at

00:12:07.360 --> 00:12:12.760
decoding time um using some other

00:12:10.800 --> 00:12:15.320
distribution some other data or in this

00:12:12.760 --> 00:12:17.320
case some other model so if you've ever

00:12:15.320 --> 00:12:19.240
played around with a really like

00:12:17.320 --> 00:12:21.800
relatively small language model maybe

00:12:19.240 --> 00:12:23.320
something like gbt2 small um You

00:12:21.800 --> 00:12:26.560
probably noticed you try to give it some

00:12:23.320 --> 00:12:28.240
inputs and maybe it degenerates into

00:12:26.560 --> 00:12:30.160
just repeating the same sequence over

00:12:28.240 --> 00:12:31.720
and over maybe it gives you outputs that

00:12:30.160 --> 00:12:33.399
are just completely incorrect like you

00:12:31.720 --> 00:12:35.320
ask it a factual question and it gets it

00:12:33.399 --> 00:12:37.120
wrong um and you don't see those

00:12:35.320 --> 00:12:39.519
problems if you look at sort of a larger

00:12:37.120 --> 00:12:41.399
model that's trained on more data so the

00:12:39.519 --> 00:12:43.199
question here is can you use what that

00:12:41.399 --> 00:12:46.480
smaller model is getting wrong to make

00:12:43.199 --> 00:12:49.120
your larger model even better um and the

00:12:46.480 --> 00:12:51.360
way we do this is by sort of the

00:12:49.120 --> 00:12:52.880
intuition that if the smaller model

00:12:51.360 --> 00:12:55.079
doesn't have a lot of probability on

00:12:52.880 --> 00:12:57.160
some answer but the the larger model

00:12:55.079 --> 00:12:58.519
does it's likely because that larger

00:12:57.160 --> 00:13:02.279
model has learned something with the

00:12:58.519 --> 00:13:04.000
smaller model didn't know and so here we

00:13:02.279 --> 00:13:06.199
modify the probability distribution

00:13:04.000 --> 00:13:08.199
coming out of the larger model to choose

00:13:06.199 --> 00:13:11.120
outputs that that model thinks are very

00:13:08.199 --> 00:13:12.600
likely and the amateur or the the weaker

00:13:11.120 --> 00:13:15.480
model thinks are not

00:13:12.600 --> 00:13:20.000
likely so in this example here from

00:13:15.480 --> 00:13:22.560
their paper um if you have sort of a

00:13:20.000 --> 00:13:27.199
input like Barack Obama was born in

00:13:22.560 --> 00:13:29.720
Hawaii he was born in L um the smaller

00:13:27.199 --> 00:13:31.360
model would often do something like

00:13:29.720 --> 00:13:35.399
start repeating and actually if you

00:13:31.360 --> 00:13:36.720
sample sort of naively from the um

00:13:35.399 --> 00:13:38.560
larger model you can wind up in these

00:13:36.720 --> 00:13:40.000
situations as well right so if you just

00:13:38.560 --> 00:13:41.959
choose the most likely thing at each

00:13:40.000 --> 00:13:43.399
step you wind up in this Loop where it's

00:13:41.959 --> 00:13:45.560
like he was born in Hawaii he was born

00:13:43.399 --> 00:13:48.199
in Hawaii he was born in Hawaii um and

00:13:45.560 --> 00:13:51.320
this is behavior we generally don't want

00:13:48.199 --> 00:13:52.680
um if you do something like nucleus or

00:13:51.320 --> 00:13:53.720
top PE sampling you can wind up with

00:13:52.680 --> 00:13:55.880
things that are actually completely

00:13:53.720 --> 00:13:58.839
incorrect like he was born in Washington

00:13:55.880 --> 00:14:01.480
DC um but if you use contrastive

00:13:58.839 --> 00:14:04.120
decoding you take the outputs coming out

00:14:01.480 --> 00:14:05.720
of your expert model here and you

00:14:04.120 --> 00:14:07.680
subtract out the probabilities coming

00:14:05.720 --> 00:14:10.160
out of the weaker model and you can wind

00:14:07.680 --> 00:14:11.880
up with things that the higher model the

00:14:10.160 --> 00:14:13.759
stronger model ascribed probability to

00:14:11.880 --> 00:14:15.480
but the weaker model did not likely

00:14:13.759 --> 00:14:16.920
because these are sort of facts that the

00:14:15.480 --> 00:14:18.959
larger model knows that the smaller

00:14:16.920 --> 00:14:20.800
model does not so here we actually get

00:14:18.959 --> 00:14:23.199
the year Barack Obama was born which is

00:14:20.800 --> 00:14:25.800
maybe a fact that the larger model knows

00:14:23.199 --> 00:14:27.639
and the smaller model didn't know um and

00:14:25.800 --> 00:14:29.759
so this is just one of sort of a broad

00:14:27.639 --> 00:14:32.560
class of methods where you use external

00:14:29.759 --> 00:14:35.199
information to improve your decoding by

00:14:32.560 --> 00:14:38.720
modifying this distribution at each

00:14:35.199 --> 00:14:40.720
set um those are sort of a brief tour of

00:14:38.720 --> 00:14:43.920
a couple of different sampling methods

00:14:40.720 --> 00:14:43.920
before we move into search

00:14:44.600 --> 00:14:50.440
yeah

00:14:46.279 --> 00:14:54.880
yeah is it going to improve upon just

00:14:50.440 --> 00:14:57.240
the yeah it generally does um and the

00:14:54.880 --> 00:14:59.800
intuition for why this might be I think

00:14:57.240 --> 00:15:01.680
is that there are sort of these

00:14:59.800 --> 00:15:04.560
degenerate cases like just repeating

00:15:01.680 --> 00:15:06.120
over and over that both the expert and

00:15:04.560 --> 00:15:09.000
the weak model would give relatively

00:15:06.120 --> 00:15:10.880
high probability to um maybe the expert

00:15:09.000 --> 00:15:13.199
model is like slightly less likely to do

00:15:10.880 --> 00:15:14.959
these things but it's still like sort of

00:15:13.199 --> 00:15:16.639
an easy case for the model to learn and

00:15:14.959 --> 00:15:18.120
so both of those models will have high

00:15:16.639 --> 00:15:20.079
probability for those things but the

00:15:18.120 --> 00:15:21.800
things that are genuinely like good

00:15:20.079 --> 00:15:23.880
outputs that only the expert would get

00:15:21.800 --> 00:15:25.519
right those will have low probability

00:15:23.880 --> 00:15:27.600
under the weak model and so you're sort

00:15:25.519 --> 00:15:30.880
of subtracting out all the degenerate

00:15:27.600 --> 00:15:33.759
behaviors and keeping to really good out

00:15:30.880 --> 00:15:35.240
this if you're generating a longer

00:15:33.759 --> 00:15:37.440
sequence with with

00:15:35.240 --> 00:15:40.759
contacing how do you know which steps

00:15:37.440 --> 00:15:45.120
you want to bring out yeah this is a

00:15:40.759 --> 00:15:48.560
great question so for this particular

00:15:45.120 --> 00:15:50.560
case oh yeah sorry so this was if you're

00:15:48.560 --> 00:15:52.279
doing contrastive decoding over a really

00:15:50.560 --> 00:15:54.399
long sequence like when do you choose to

00:15:52.279 --> 00:15:55.800
bring in the expert right and for

00:15:54.399 --> 00:15:58.600
contrastive decoding we're actually

00:15:55.800 --> 00:16:00.759
going to do this at every individual

00:15:58.600 --> 00:16:02.440
time step so we're going to use the

00:16:00.759 --> 00:16:04.800
expert model to decode and we're going

00:16:02.440 --> 00:16:07.000
to bring in the amateur to sort of

00:16:04.800 --> 00:16:09.079
subtract out probabilities at each next

00:16:07.000 --> 00:16:10.399
token prediction um you don't have to do

00:16:09.079 --> 00:16:12.800
that I think that's that's what they do

00:16:10.399 --> 00:16:15.000
in the paper um you could also decide to

00:16:12.800 --> 00:16:16.680
only do this sort of if you have high

00:16:15.000 --> 00:16:19.639
uncertainty or something if you don't

00:16:16.680 --> 00:16:22.639
have a really sharp probability

00:16:19.639 --> 00:16:22.639
distribution

00:16:23.160 --> 00:16:28.160
yeah yeah how weak should the weak

00:16:25.399 --> 00:16:30.199
predictor be um in the in the paper what

00:16:28.160 --> 00:16:31.600
they're look at is actually not a huge

00:16:30.199 --> 00:16:34.560
difference between the two models so you

00:16:31.600 --> 00:16:35.800
can see here this is gpd2 XL and small

00:16:34.560 --> 00:16:37.319
so there's a difference in parameter

00:16:35.800 --> 00:16:39.519
counts and like a bit of a difference in

00:16:37.319 --> 00:16:42.160
data I think here but these are actually

00:16:39.519 --> 00:16:44.959
not like gpd2 XL is certainly not like a

00:16:42.160 --> 00:16:48.399
super strong model now um I think they

00:16:44.959 --> 00:16:50.920
try a couple of different settings and

00:16:48.399 --> 00:16:52.319
the general intuition I think if I'm

00:16:50.920 --> 00:16:54.880
remembering it correctly is that you

00:16:52.319 --> 00:16:56.319
want a model that's not like so close in

00:16:54.880 --> 00:16:58.000
performance to your expert that you're

00:16:56.319 --> 00:16:59.839
basically just subtracting out useful

00:16:58.000 --> 00:17:02.240
things but you also don't want a model

00:16:59.839 --> 00:17:03.519
that's like so degenerate that it is not

00:17:02.240 --> 00:17:04.959
hasn't learned anything useful about

00:17:03.519 --> 00:17:06.839
your task at all so I think it might

00:17:04.959 --> 00:17:09.600
depend on what task you're looking

00:17:06.839 --> 00:17:12.919
at

00:17:09.600 --> 00:17:14.559
yes this is for inference um so actually

00:17:12.919 --> 00:17:17.640
everything we look at today will not

00:17:14.559 --> 00:17:17.640
require aning of the

00:17:19.360 --> 00:17:26.559
model Okay cool so now we're going to

00:17:24.000 --> 00:17:30.039
step into sort of a slightly different

00:17:26.559 --> 00:17:31.280
um set of strategies here which is maybe

00:17:30.039 --> 00:17:33.039
we don't just want something from the

00:17:31.280 --> 00:17:35.160
model distribution or something from a

00:17:33.039 --> 00:17:37.760
modified distribution maybe we actually

00:17:35.160 --> 00:17:39.840
just want the quote unquote best thing

00:17:37.760 --> 00:17:42.960
the single most likely output given our

00:17:39.840 --> 00:17:45.200
input right and here this would be the Y

00:17:42.960 --> 00:17:48.039
hat the single sequence that satisfies

00:17:45.200 --> 00:17:51.919
that has the highest score py given X

00:17:48.039 --> 00:17:54.240
for the X that we gave the model um this

00:17:51.919 --> 00:17:56.000
is this section is called mode seeking

00:17:54.240 --> 00:17:58.039
search because this is the mode of the

00:17:56.000 --> 00:18:00.440
distribution over outputs if you sampled

00:17:58.039 --> 00:18:01.760
a huge huge number of times and you

00:18:00.440 --> 00:18:04.720
looked at the single most likely

00:18:01.760 --> 00:18:06.720
sequence you got it would be this y hat

00:18:04.720 --> 00:18:09.280
and so how do we find this

00:18:06.720 --> 00:18:11.600
thing well one idea is we know the

00:18:09.280 --> 00:18:13.159
distribution at each individual setep

00:18:11.600 --> 00:18:16.000
can we just pick the most likely thing

00:18:13.159 --> 00:18:18.960
from that distribution and so in Greedy

00:18:16.000 --> 00:18:21.080
decoding we take the argmax the single

00:18:18.960 --> 00:18:22.720
highest probability token at each step

00:18:21.080 --> 00:18:24.840
and we continue generating until the

00:18:22.720 --> 00:18:26.600
single highest most the single highest

00:18:24.840 --> 00:18:28.840
probability token is the stop token

00:18:26.600 --> 00:18:31.559
right the end of sequence token

00:18:28.840 --> 00:18:33.400
um for an individual token right if we

00:18:31.559 --> 00:18:35.559
only want a single token output this is

00:18:33.400 --> 00:18:38.320
exactly what we want this is the single

00:18:35.559 --> 00:18:40.400
most likely output um and that's great

00:18:38.320 --> 00:18:44.000
but if we're looking at something that

00:18:40.400 --> 00:18:45.120
is maybe several tokens long are we

00:18:44.000 --> 00:18:47.360
actually going to get the highest

00:18:45.120 --> 00:18:49.720
probability thing and if you kind of

00:18:47.360 --> 00:18:52.159
squint at this you can see that maybe we

00:18:49.720 --> 00:18:54.120
have a problem here where the highest

00:18:52.159 --> 00:18:56.320
probability sequence that you get from

00:18:54.120 --> 00:18:58.039
multiplying across multiple steps

00:18:56.320 --> 00:18:59.559
doesn't necessarily start with the token

00:18:58.039 --> 00:19:01.600
that was highest probability at time

00:18:59.559 --> 00:19:03.200
step one right maybe if you're doing

00:19:01.600 --> 00:19:04.720
something like unconditional generation

00:19:03.200 --> 00:19:06.720
the highest probability token at time

00:19:04.720 --> 00:19:08.360
step one is always the but there could

00:19:06.720 --> 00:19:09.919
be a really probable sentence that just

00:19:08.360 --> 00:19:11.480
doesn't happen to start with the the

00:19:09.919 --> 00:19:12.720
word the' and you would never find it

00:19:11.480 --> 00:19:15.080
using GRE

00:19:12.720 --> 00:19:17.360
decoding so this isn't going to give us

00:19:15.080 --> 00:19:19.799
the highest probability output over a

00:19:17.360 --> 00:19:22.000
sequence that's more than one token one

00:19:19.799 --> 00:19:23.360
can we do anything better to try to find

00:19:22.000 --> 00:19:25.640
this um

00:19:23.360 --> 00:19:27.559
output and here we get into sort of one

00:19:25.640 --> 00:19:29.520
of the most popular decoding methods the

00:19:27.559 --> 00:19:32.600
one that you maybe heard of before which

00:19:29.520 --> 00:19:35.080
is beam search the idea here is that we

00:19:32.600 --> 00:19:36.559
don't want to miss a high probability

00:19:35.080 --> 00:19:38.880
token that's hidden behind a lower

00:19:36.559 --> 00:19:40.200
probability prefix so we want to kind of

00:19:38.880 --> 00:19:42.000
search through a couple of different

00:19:40.200 --> 00:19:43.760
options so that we don't discard

00:19:42.000 --> 00:19:47.120
something too early that might have high

00:19:43.760 --> 00:19:49.360
probability um later on in generation

00:19:47.120 --> 00:19:50.919
and this is a type of bread first search

00:19:49.360 --> 00:19:53.200
so we're going to look at a wide variety

00:19:50.919 --> 00:19:54.600
of options at a given time step we're

00:19:53.200 --> 00:19:55.600
going to pick some set of them to

00:19:54.600 --> 00:19:57.120
continue and then we're going to look at

00:19:55.600 --> 00:19:58.919
a wide variety of options for the next

00:19:57.120 --> 00:19:59.960
time step instead of generating all the

00:19:58.919 --> 00:20:02.200
way through a sequence and then

00:19:59.960 --> 00:20:04.320
generating all the way through another

00:20:02.200 --> 00:20:05.760
sequence um and how this works is we're

00:20:04.320 --> 00:20:07.559
going to pick sort of a number of

00:20:05.760 --> 00:20:09.400
candidates we'd like to explore a beam

00:20:07.559 --> 00:20:11.039
with so in this example we're going to

00:20:09.400 --> 00:20:12.799
pick three and we're going to say all

00:20:11.039 --> 00:20:15.480
right here are maybe three options for

00:20:12.799 --> 00:20:17.640
time step one for if we pick each of

00:20:15.480 --> 00:20:19.760
those three options what would be the

00:20:17.640 --> 00:20:21.799
three most likely things for time step

00:20:19.760 --> 00:20:23.200
two right rather than choosing just the

00:20:21.799 --> 00:20:24.520
single most likely thing in Greedy

00:20:23.200 --> 00:20:26.960
decoding we're going to pick three

00:20:24.520 --> 00:20:29.120
options and so now we have three options

00:20:26.960 --> 00:20:32.559
for time step one three options for time

00:20:29.120 --> 00:20:34.280
step two we now have nine options um

00:20:32.559 --> 00:20:36.320
here right three options and then three

00:20:34.280 --> 00:20:37.679
more for each of these and we don't want

00:20:36.320 --> 00:20:40.159
to continue doing this because this is

00:20:37.679 --> 00:20:41.960
going to sort of combinator explode so

00:20:40.159 --> 00:20:44.080
we need to choose some subset of these

00:20:41.960 --> 00:20:45.880
to continue with and the way we do that

00:20:44.080 --> 00:20:47.799
is we look at the probability over this

00:20:45.880 --> 00:20:49.240
two token sequence and we choose the two

00:20:47.799 --> 00:20:51.520
that have the highest probability

00:20:49.240 --> 00:20:53.400
overall so in this instance we've chosen

00:20:51.520 --> 00:20:55.679
sort of one thing from this first group

00:20:53.400 --> 00:20:57.760
and two things from the second group and

00:20:55.679 --> 00:20:59.760
now we're back down to three hypotheses

00:20:57.760 --> 00:21:02.120
each now two tokens long and we'll

00:20:59.760 --> 00:21:04.000
continue generating to time step three

00:21:02.120 --> 00:21:05.600
we'll get nine options we'll pre it back

00:21:04.000 --> 00:21:07.760
down to three and we'll continue until

00:21:05.600 --> 00:21:09.159
the end of generation where we now have

00:21:07.760 --> 00:21:10.679
three sequences and we'll just pick the

00:21:09.159 --> 00:21:14.000
one that's highest probability out of

00:21:10.679 --> 00:21:15.679
those three to return um this is not

00:21:14.000 --> 00:21:17.360
guaranteed to get you the highest

00:21:15.679 --> 00:21:18.480
probability thing right you still have

00:21:17.360 --> 00:21:20.039
this risk that you could be sort of

00:21:18.480 --> 00:21:22.279
pruning out something that's high

00:21:20.039 --> 00:21:24.159
probability but in general this sort of

00:21:22.279 --> 00:21:26.600
works um much better than greedy

00:21:24.159 --> 00:21:28.520
decoding and this is if you have a

00:21:26.600 --> 00:21:31.120
language model and you're sort of not

00:21:28.520 --> 00:21:32.440
what um decoding method it's using outs

00:21:31.120 --> 00:21:34.200
are pretty good it's either beam search

00:21:32.440 --> 00:21:37.120
or temperature samping right this is

00:21:34.200 --> 00:21:40.039
very effective this is used um pretty

00:21:37.120 --> 00:21:41.760
broadly there are however some issues

00:21:40.039 --> 00:21:43.760
with beam search and one of the biggest

00:21:41.760 --> 00:21:46.159
ones is that when you're doing this

00:21:43.760 --> 00:21:47.679
maximum likelihood sampling you really

00:21:46.159 --> 00:21:50.080
or the sampling to search for something

00:21:47.679 --> 00:21:51.760
that's very high likelihood um you

00:21:50.080 --> 00:21:53.679
really sacrifice a lot of diversity in

00:21:51.760 --> 00:21:55.320
your outputs and in particular you could

00:21:53.679 --> 00:21:57.279
wind up at the end of beam search with

00:21:55.320 --> 00:21:58.919
three different outputs to choose from

00:21:57.279 --> 00:22:00.120
that are all pretty pretty much the same

00:21:58.919 --> 00:22:02.640
like they're slightly different token

00:22:00.120 --> 00:22:04.559
sequences but they look very similar and

00:22:02.640 --> 00:22:07.480
so maybe you want to S get sort of a

00:22:04.559 --> 00:22:08.919
more diverse set um there's a couple of

00:22:07.480 --> 00:22:10.640
different methods in this category I'm

00:22:08.919 --> 00:22:12.679
going to very briefly shout out two of

00:22:10.640 --> 00:22:14.200
them um but the idea here is to sort of

00:22:12.679 --> 00:22:16.440
reintroduce some of the benefits of

00:22:14.200 --> 00:22:19.120
sampling while still doing this kind of

00:22:16.440 --> 00:22:20.919
search for high probability things um

00:22:19.120 --> 00:22:22.600
diverse beam search is one of these

00:22:20.919 --> 00:22:25.520
methods and here the idea is that we

00:22:22.600 --> 00:22:27.279
want to modify that scoring step when we

00:22:25.520 --> 00:22:28.600
choose which three out of our nine beams

00:22:27.279 --> 00:22:30.200
we want to continue

00:22:28.600 --> 00:22:32.000
to avoid choosing things that are really

00:22:30.200 --> 00:22:34.320
really close to each other right so

00:22:32.000 --> 00:22:36.039
maybe our highest probability thing is

00:22:34.320 --> 00:22:37.559
some sequence a and then if we look at

00:22:36.039 --> 00:22:39.520
the other sequences there's one that's

00:22:37.559 --> 00:22:41.279
pretty high probability but very similar

00:22:39.520 --> 00:22:43.600
to that sequence and there's one that's

00:22:41.279 --> 00:22:45.320
like slightly lower probability but very

00:22:43.600 --> 00:22:47.200
different and so maybe we would choose a

00:22:45.320 --> 00:22:49.679
sequence that is a little lower

00:22:47.200 --> 00:22:51.760
probability to maximize diversity in our

00:22:49.679 --> 00:22:53.799
set to try to get like sort of a wider

00:22:51.760 --> 00:22:56.200
range of options to choose from later in

00:22:53.799 --> 00:22:58.200
generation so this modifies the scoring

00:22:56.200 --> 00:23:00.120
to not just take into account likelihood

00:22:58.200 --> 00:23:03.200
but also similarity to other

00:23:00.120 --> 00:23:05.400
KS another option down this path is

00:23:03.200 --> 00:23:07.640
stochastic beam search where we're going

00:23:05.400 --> 00:23:09.279
to keep the scoring the same but rather

00:23:07.640 --> 00:23:11.679
than choosing just the top three most

00:23:09.279 --> 00:23:13.279
likely tokens to expand out each beam

00:23:11.679 --> 00:23:15.200
we're actually going to sample from some

00:23:13.279 --> 00:23:17.000
distribution and you could sample from

00:23:15.200 --> 00:23:18.760
the model distribution directly using

00:23:17.000 --> 00:23:20.200
ancestral sampling or you could use any

00:23:18.760 --> 00:23:22.679
of our sampling methods we talked about

00:23:20.200 --> 00:23:24.200
in the last section to do this and the

00:23:22.679 --> 00:23:25.799
the idea here is sort of similar to

00:23:24.200 --> 00:23:29.279
diverse beam search we want to get sort

00:23:25.799 --> 00:23:31.240
of a wider exploration of our models

00:23:29.279 --> 00:23:33.520
like output space you know we want to

00:23:31.240 --> 00:23:35.360
sort of explore more things instead of

00:23:33.520 --> 00:23:36.760
just seeking winding up with a bunch of

00:23:35.360 --> 00:23:39.679
outputs that look very similar at the

00:23:36.760 --> 00:23:41.120
end of beam search um if folks are

00:23:39.679 --> 00:23:43.679
interested in these I think these are

00:23:41.120 --> 00:23:46.159
both linked on the website um the the

00:23:43.679 --> 00:23:48.679
papers that both of these ideas came

00:23:46.159 --> 00:23:51.480
from

00:23:48.679 --> 00:23:54.400
Yes um for stochastic

00:23:51.480 --> 00:23:57.039
resarch the sampl probability takes into

00:23:54.400 --> 00:23:59.039
account the current part that we already

00:23:57.039 --> 00:24:02.000
travel okay

00:23:59.039 --> 00:24:04.320
yeah exactly so it's this um like

00:24:02.000 --> 00:24:05.640
selection step here but we're instead of

00:24:04.320 --> 00:24:07.760
just doing greedy selection we're going

00:24:05.640 --> 00:24:11.760
to do

00:24:07.760 --> 00:24:17.520
assembling yes my question was on the T

00:24:11.760 --> 00:24:23.200
yeah like you for something super simple

00:24:17.520 --> 00:24:26.520
like if both of them have a high are you

00:24:23.200 --> 00:24:28.120
like yeah so you would if it has a

00:24:26.520 --> 00:24:30.080
really high probability under both

00:24:28.120 --> 00:24:32.880
models it would have a lower probability

00:24:30.080 --> 00:24:35.080
after doing this sort of contrasted

00:24:32.880 --> 00:24:36.600
de right so if the if the smaller

00:24:35.080 --> 00:24:38.799
model's really good at your task this

00:24:36.600 --> 00:24:40.960
might not work very

00:24:38.799 --> 00:24:43.360
well yeah I think in the paper they're

00:24:40.960 --> 00:24:45.320
generally evaluating on these sort of

00:24:43.360 --> 00:24:48.279
like open ended generation task I bet

00:24:45.320 --> 00:24:51.279
this works a lot worse for

00:24:48.279 --> 00:24:51.279
now

00:24:56.760 --> 00:24:59.760
yes

00:25:02.440 --> 00:25:08.120
you yeah this is a great question um and

00:25:05.960 --> 00:25:11.559
so the question is how do we measure

00:25:08.120 --> 00:25:14.120
similar beams um you can sort of Define

00:25:11.559 --> 00:25:15.559
any kind of similarity function you like

00:25:14.120 --> 00:25:17.520
here um anything that you'd use to

00:25:15.559 --> 00:25:20.440
evaluate like how similar something is

00:25:17.520 --> 00:25:22.360
to a gold reference right um I think in

00:25:20.440 --> 00:25:25.039
the original diverse beam search they do

00:25:22.360 --> 00:25:27.760
this by looking at like exact token

00:25:25.039 --> 00:25:30.640
match across the two right like if these

00:25:27.760 --> 00:25:33.880
beams are the same in all but one of the

00:25:30.640 --> 00:25:35.600
tokens or they have like you know 50% of

00:25:33.880 --> 00:25:37.120
the tokens are shared across the beams

00:25:35.600 --> 00:25:38.559
and maybe these are really similar and

00:25:37.120 --> 00:25:40.559
they should try to choose two things

00:25:38.559 --> 00:25:42.600
that are different um but you could swap

00:25:40.559 --> 00:25:46.200
that out for any

00:25:42.600 --> 00:25:49.440
metc yes so

00:25:46.200 --> 00:25:50.960
the there's kind of like a that's Happ

00:25:49.440 --> 00:25:53.360
at

00:25:50.960 --> 00:25:55.000
every for the stochastic be search

00:25:53.360 --> 00:25:57.720
there's like a shering what do you mean

00:25:55.000 --> 00:26:00.520
by a shepher so it says modify the next

00:25:57.720 --> 00:26:03.000
sech selection because they're like um

00:26:00.520 --> 00:26:06.919
it is searching at a different space and

00:26:03.000 --> 00:26:09.679
it's not searching within the same 3D

00:26:06.919 --> 00:26:14.080
SE is it searching in a different space

00:26:09.679 --> 00:26:15.799
yeah so it's um in the same probability

00:26:14.080 --> 00:26:18.399
distribution but it'll see a different

00:26:15.799 --> 00:26:20.840
part of the distribution so when you're

00:26:18.399 --> 00:26:22.640
doing the grey search you'll only ever

00:26:20.840 --> 00:26:24.559
look at the top three tokens in the next

00:26:22.640 --> 00:26:27.120
token distribution because you're just

00:26:24.559 --> 00:26:29.840
selecting like the maximums um but in

00:26:27.120 --> 00:26:31.360
sampling you could you could get the

00:26:29.840 --> 00:26:32.880
same tokens right if they're really high

00:26:31.360 --> 00:26:35.720
likelihood but you could also sample

00:26:32.880 --> 00:26:38.399
something that's further down in the

00:26:35.720 --> 00:26:42.760
distribution yeah as a followup to that

00:26:38.399 --> 00:26:44.880
like into uh our stamping we take into

00:26:42.760 --> 00:26:46.960
account the probability of the prefix

00:26:44.880 --> 00:26:50.679
like the current hypothesis right

00:26:46.960 --> 00:26:51.760
because otherwise it is the same as just

00:26:50.679 --> 00:26:54.279
uh

00:26:51.760 --> 00:26:57.159
in yeah so in the sampling we're taking

00:26:54.279 --> 00:27:00.120
into account the previous the prefix

00:26:57.159 --> 00:27:02.600
yeah so so it we will take into account

00:27:00.120 --> 00:27:06.200
the prefix but this sampling mechanism

00:27:02.600 --> 00:27:08.320
here could be ancestral sampling um the

00:27:06.200 --> 00:27:10.480
only the difference here is that we're

00:27:08.320 --> 00:27:12.600
also doing a sort of search step on top

00:27:10.480 --> 00:27:14.679
of that to choose the maximum likelihood

00:27:12.600 --> 00:27:18.080
things across multiple

00:27:14.679 --> 00:27:20.559
me another important thing um is you

00:27:18.080 --> 00:27:22.279
sample without replacement and so

00:27:20.559 --> 00:27:24.120
normally you sample with replacement and

00:27:22.279 --> 00:27:25.840
you might get exactly the same thing but

00:27:24.120 --> 00:27:28.000
when you're doing stasic beam search you

00:27:25.840 --> 00:27:30.240
sample without replacement so you get

00:27:28.000 --> 00:27:33.279
like three ones according to the

00:27:30.240 --> 00:27:36.080
probability but they're guaranteed to be

00:27:33.279 --> 00:27:37.799
different right so beam search like one

00:27:36.080 --> 00:27:39.559
of the characteristics of beam search is

00:27:37.799 --> 00:27:41.640
you always get three different things

00:27:39.559 --> 00:27:44.240
because you're picking the three top

00:27:41.640 --> 00:27:45.760
when you do sampling uh like stochastic

00:27:44.240 --> 00:27:47.399
Bean shirts you get three different

00:27:45.760 --> 00:27:49.440
things they're not guaranteed to be the

00:27:47.399 --> 00:27:51.760
top they could be distributed according

00:27:49.440 --> 00:27:54.360
to the prob distribution but they're

00:27:51.760 --> 00:27:55.840
guaranteed so um you can take a look at

00:27:54.360 --> 00:27:58.039
the paper for more details of exactly

00:27:55.840 --> 00:28:00.159
how it looks but that that's

00:27:58.039 --> 00:28:03.039
so then is the main difference that

00:28:00.159 --> 00:28:05.120
compared to plus temping that we have n

00:28:03.039 --> 00:28:08.519
options that we're cheing tet instead of

00:28:05.120 --> 00:28:10.320
going with the going with only one and

00:28:08.519 --> 00:28:11.200
you can't yeah you can't simple the same

00:28:10.320 --> 00:28:14.960
thing

00:28:11.200 --> 00:28:16.919
right yeah so just uh repeat recording

00:28:14.960 --> 00:28:19.159
is that n options we're keeping track of

00:28:16.919 --> 00:28:22.240
and they're all going to be unique token

00:28:19.159 --> 00:28:24.240
sequences at least um you can actually

00:28:22.240 --> 00:28:26.200
get the same output sequence from two

00:28:24.240 --> 00:28:28.120
different toen sequences if you tokenize

00:28:26.200 --> 00:28:32.360
slightly differently um but these will

00:28:28.120 --> 00:28:37.840
always be unique tokens

00:28:32.360 --> 00:28:39.279
Le so that was sort of a a why like a a

00:28:37.840 --> 00:28:41.320
set of methods that we've developed to

00:28:39.279 --> 00:28:43.600
try to find the most probable sequence

00:28:41.320 --> 00:28:44.480
out of the model um but in the next

00:28:43.600 --> 00:28:46.039
section here we're going to sort of

00:28:44.480 --> 00:28:50.240
think about whether that's actually what

00:28:46.039 --> 00:28:51.679
we want to do at all um so what is like

00:28:50.240 --> 00:28:54.240
is do we really want the highest

00:28:51.679 --> 00:28:56.880
probability thing um we know that

00:28:54.240 --> 00:28:58.600
outputs with really low probability tend

00:28:56.880 --> 00:29:00.640
to be really like worse than outfits

00:28:58.600 --> 00:29:03.240
with high probability right maybe I'm

00:29:00.640 --> 00:29:05.840
trying to predict like what the next

00:29:03.240 --> 00:29:08.640
sentence should be after the cat saw the

00:29:05.840 --> 00:29:11.240
dog right the cat sat down is way higher

00:29:08.640 --> 00:29:12.559
probability than the cat grew wings and

00:29:11.240 --> 00:29:14.039
at least with the cats I've met that

00:29:12.559 --> 00:29:15.679
sounds pretty that sounds pretty much

00:29:14.039 --> 00:29:19.559
right right like this is a much better

00:29:15.679 --> 00:29:21.720
output than the cat gr wings but if you

00:29:19.559 --> 00:29:24.159
look at just the outputs with relatively

00:29:21.720 --> 00:29:25.960
high probability it's sort of less clear

00:29:24.159 --> 00:29:27.880
that this defines an exact ranking

00:29:25.960 --> 00:29:30.559
between those outputs right

00:29:27.880 --> 00:29:32.600
is the cat sat down necessarily better

00:29:30.559 --> 00:29:34.519
than the cat ran away these both seem

00:29:32.600 --> 00:29:35.720
like pretty reasonable outputs to me

00:29:34.519 --> 00:29:40.200
even though one of them is slightly

00:29:35.720 --> 00:29:42.799
higher probability and so we do we

00:29:40.200 --> 00:29:45.240
really like necessarily need to recover

00:29:42.799 --> 00:29:47.200
the cat that down um and this gets a

00:29:45.240 --> 00:29:49.399
little a little more complicated still

00:29:47.200 --> 00:29:51.120
if we look at sort of a range of outputs

00:29:49.399 --> 00:29:53.120
so say there's sort of six outputs that

00:29:51.120 --> 00:29:55.240
our model could give us um and here

00:29:53.120 --> 00:29:57.559
we're looking at sort of full sequences

00:29:55.240 --> 00:30:00.120
not individual tokens just for clarity

00:29:57.559 --> 00:30:02.640
so maybe our outputs in order of

00:30:00.120 --> 00:30:05.840
probability are the cat sat down it ran

00:30:02.640 --> 00:30:08.240
away it sprinted off it got out of there

00:30:05.840 --> 00:30:09.720
it's very small and it grew Wings right

00:30:08.240 --> 00:30:11.440
so we're definitely sure that the cat

00:30:09.720 --> 00:30:13.159
sat down is a better output than the cat

00:30:11.440 --> 00:30:15.360
grew wings and if we're doing a mod

00:30:13.159 --> 00:30:17.600
seeking search we would find that as our

00:30:15.360 --> 00:30:19.440
most likely thing if we're if we you

00:30:17.600 --> 00:30:21.440
know do a good job searching and we'd

00:30:19.440 --> 00:30:23.519
return that as our output but if you

00:30:21.440 --> 00:30:25.919
look at the rest of this distribution

00:30:23.519 --> 00:30:27.880
you see that there's actually a whole

00:30:25.919 --> 00:30:29.240
set of outputs after that all say

00:30:27.880 --> 00:30:31.720
something that kind of means the cat

00:30:29.240 --> 00:30:33.480
left the area right it's just that this

00:30:31.720 --> 00:30:35.200
probability is split over these three

00:30:33.480 --> 00:30:37.080
different generations and if you

00:30:35.200 --> 00:30:39.120
actually add up the probability mass of

00:30:37.080 --> 00:30:40.880
all three of these sequences this is

00:30:39.120 --> 00:30:42.919
double the probability mass of the cat

00:30:40.880 --> 00:30:44.360
sat down but because none of these

00:30:42.919 --> 00:30:45.960
individual sequences is higher

00:30:44.360 --> 00:30:47.399
probability if you're doing mode seeking

00:30:45.960 --> 00:30:50.640
search you wouldn't you wouldn't be able

00:30:47.399 --> 00:30:52.480
to see this effect right so do we really

00:30:50.640 --> 00:30:53.760
want to return the cat sat down or do we

00:30:52.480 --> 00:30:55.200
want to return something that means the

00:30:53.760 --> 00:30:57.559
cat left the

00:30:55.200 --> 00:30:59.200
area the question then is like if it's

00:30:57.559 --> 00:31:03.120
not probability that makes an output

00:30:59.200 --> 00:31:04.679
good what is it so we have this one

00:31:03.120 --> 00:31:06.039
output that's really high probability

00:31:04.679 --> 00:31:09.000
but it's very different from everything

00:31:06.039 --> 00:31:10.720
else in our set and then we have a

00:31:09.000 --> 00:31:13.200
couple of outputs that are all pretty

00:31:10.720 --> 00:31:15.080
high probability and similar to a bunch

00:31:13.200 --> 00:31:17.840
of other relatively high probability

00:31:15.080 --> 00:31:19.720
things so maybe it's sort of less risky

00:31:17.840 --> 00:31:21.399
to return one of these right are thing

00:31:19.720 --> 00:31:23.200
that's higher probability but different

00:31:21.399 --> 00:31:24.600
than everything else could be different

00:31:23.200 --> 00:31:26.840
because it's way better or it could be

00:31:24.600 --> 00:31:29.000
different because it's way worse um

00:31:26.840 --> 00:31:31.120
another way to think about this is you

00:31:29.000 --> 00:31:32.600
know maybe if you and your friends were

00:31:31.120 --> 00:31:34.200
cheating on a test which you shouldn't

00:31:32.600 --> 00:31:35.480
do but if you were going to do it and

00:31:34.200 --> 00:31:37.519
all of your friends sent you their

00:31:35.480 --> 00:31:39.240
answers um maybe one of your friends has

00:31:37.519 --> 00:31:40.960
a slightly higher score in the class

00:31:39.240 --> 00:31:42.519
than everyone else but they said the

00:31:40.960 --> 00:31:44.480
answer was answer a and everyone else

00:31:42.519 --> 00:31:45.799
said the answer was B right you still

00:31:44.480 --> 00:31:48.480
might go with the answer that everyone

00:31:45.799 --> 00:31:50.679
else said because like what there's it

00:31:48.480 --> 00:31:52.679
sort of feels less risky like maybe

00:31:50.679 --> 00:31:54.440
everyone else got the answer get that

00:31:52.679 --> 00:31:55.880
answer and so your one friend could be

00:31:54.440 --> 00:31:56.919
right when everyone else is wrong or

00:31:55.880 --> 00:31:59.679
they could have made a mistake that no

00:31:56.919 --> 00:32:01.240
one El else is making so this is sort of

00:31:59.679 --> 00:32:03.519
the same concept right we want an output

00:32:01.240 --> 00:32:06.320
that's relatively high probability but

00:32:03.519 --> 00:32:09.399
also relatively low

00:32:06.320 --> 00:32:11.320
risk and so here maybe if we were using

00:32:09.399 --> 00:32:13.679
this criteria we'd return the cat ran

00:32:11.320 --> 00:32:14.720
away as our sort of as our sort of

00:32:13.679 --> 00:32:16.720
single

00:32:14.720 --> 00:32:19.440
output so how do you find something

00:32:16.720 --> 00:32:21.000
that's high probability and low risk

00:32:19.440 --> 00:32:22.480
there's sort of two questions here right

00:32:21.000 --> 00:32:24.399
we have to figure out how to estimate

00:32:22.480 --> 00:32:26.120
probability and if we're looking at a

00:32:24.399 --> 00:32:28.519
set of outputs like the six we saw

00:32:26.120 --> 00:32:29.880
before maybe we can just do this by

00:32:28.519 --> 00:32:31.720
counting right we could sample

00:32:29.880 --> 00:32:34.000
everything from the model and get exact

00:32:31.720 --> 00:32:35.200
probability or we could take a sample

00:32:34.000 --> 00:32:38.080
from the model and just look at

00:32:35.200 --> 00:32:40.200
probabilities in that set and from there

00:32:38.080 --> 00:32:41.840
from that sample um sort of one

00:32:40.200 --> 00:32:43.559
reasonable thing to do is just count

00:32:41.840 --> 00:32:45.320
frequency right if something's in our

00:32:43.559 --> 00:32:47.919
sample twice as often we just say it's

00:32:45.320 --> 00:32:49.799
twice as frequent or it's twice as

00:32:47.919 --> 00:32:52.880
probable um this is something called

00:32:49.799 --> 00:32:54.440
Monte Carlos sampling if you do this um

00:32:52.880 --> 00:32:56.039
enough times like if you sample an

00:32:54.440 --> 00:32:58.279
infinite set this is would give you

00:32:56.039 --> 00:33:00.880
exactly the model distri distribution um

00:32:58.279 --> 00:33:02.840
but for the sort of reasonable size sets

00:33:00.880 --> 00:33:04.200
we're working with maybe like a 100

00:33:02.840 --> 00:33:06.320
samples this gives us a sort of

00:33:04.200 --> 00:33:09.440
reasonable approximation for what we for

00:33:06.320 --> 00:33:10.840
what we need to do here at least so

00:33:09.440 --> 00:33:12.000
we're just going to take a sample to get

00:33:10.840 --> 00:33:13.440
probability and we're just going to

00:33:12.000 --> 00:33:15.519
count things in that sample to see how

00:33:13.440 --> 00:33:17.320
likely things are that doesn't seem too

00:33:15.519 --> 00:33:20.080
bad how do we estimate

00:33:17.320 --> 00:33:21.679
risk the idea here is that we have a

00:33:20.080 --> 00:33:24.080
bunch of other things in this set of

00:33:21.679 --> 00:33:26.080
outputs and we can treat those as sort

00:33:24.080 --> 00:33:27.880
of like pseudo references right we can

00:33:26.080 --> 00:33:29.840
evaluate agreement between the thing

00:33:27.880 --> 00:33:31.519
we're looking at and each of those other

00:33:29.840 --> 00:33:33.480
references and this is sort of the same

00:33:31.519 --> 00:33:35.519
idea of calculating similarity in

00:33:33.480 --> 00:33:37.159
diverse beam search we're going to use

00:33:35.519 --> 00:33:39.639
some kind of metric to compare how

00:33:37.159 --> 00:33:41.279
similar these things are um this metric

00:33:39.639 --> 00:33:43.080
could be anything you use Downstream it

00:33:41.279 --> 00:33:44.840
could be like an engram overlap metric

00:33:43.080 --> 00:33:48.600
like Rouge or blue or it could also be

00:33:44.840 --> 00:33:51.120
something um neural or semantic like um

00:33:48.600 --> 00:33:54.799
something like BT score or Bart

00:33:51.120 --> 00:33:56.600
score and so this concept um is a type

00:33:54.799 --> 00:33:57.919
of decoding called minimum based risk

00:33:56.600 --> 00:33:59.600
decoding

00:33:57.919 --> 00:34:01.840
and what this equation captures is

00:33:59.600 --> 00:34:03.919
exactly the intuition that we were um

00:34:01.840 --> 00:34:06.600
sort of talking about just a slide ago

00:34:03.919 --> 00:34:08.159
where we're going to choose something

00:34:06.600 --> 00:34:09.919
that is low risk which means it's

00:34:08.159 --> 00:34:11.960
similar to a lot of other things in this

00:34:09.919 --> 00:34:12.800
set of outputs we've sampled and we're

00:34:11.960 --> 00:34:14.800
going to choose something that's

00:34:12.800 --> 00:34:17.560
relatively high probability which means

00:34:14.800 --> 00:34:19.159
that sort of when we sum up over this if

00:34:17.560 --> 00:34:21.399
something occurs in our set a bunch of

00:34:19.159 --> 00:34:23.320
times it's going to have pretty strong

00:34:21.399 --> 00:34:25.800
weight in picking which um of these

00:34:23.320 --> 00:34:27.000
outputs are similar right if sort of

00:34:25.800 --> 00:34:28.399
there's one thing in the set that

00:34:27.000 --> 00:34:29.919
appears a bunch of times it's going to

00:34:28.399 --> 00:34:32.040
have a strong influence on which thing

00:34:29.919 --> 00:34:34.119
we pick and so that kind of captures

00:34:32.040 --> 00:34:38.520
high probability in this

00:34:34.119 --> 00:34:41.119
setting so to see how this works we can

00:34:38.520 --> 00:34:44.639
look at an example um in

00:34:41.119 --> 00:34:47.399
summarization so we choose some Metric

00:34:44.639 --> 00:34:49.639
maybe we choose um Rouge which is an

00:34:47.399 --> 00:34:51.399
engram overlap metric for summarization

00:34:49.639 --> 00:34:52.879
and we say we're going to sample 100

00:34:51.399 --> 00:34:55.960
things and we're going to use this

00:34:52.879 --> 00:35:00.359
equation to choose the one that has the

00:34:55.960 --> 00:35:03.960
sort of lower EST risk according to MBR

00:35:00.359 --> 00:35:06.480
um so if we do that and we look at this

00:35:03.960 --> 00:35:07.560
sort of table of results here um you can

00:35:06.480 --> 00:35:09.680
see that this

00:35:07.560 --> 00:35:11.320
outperforms the other sampling methods

00:35:09.680 --> 00:35:13.720
that we've looked at before so greedy

00:35:11.320 --> 00:35:15.640
decoding here is just sampling the

00:35:13.720 --> 00:35:18.760
single most likely thing in each step

00:35:15.640 --> 00:35:21.800
beam search here is the BS with five or

00:35:18.760 --> 00:35:24.359
10 beams and DBS is the diverse beam

00:35:21.800 --> 00:35:27.040
search we were talking about um if we

00:35:24.359 --> 00:35:29.440
use minimum based risk and we use grou

00:35:27.040 --> 00:35:31.240
is the sort of determiner of similarity

00:35:29.440 --> 00:35:32.680
we do way better across all of our

00:35:31.240 --> 00:35:33.960
metrics but we especially do really good

00:35:32.680 --> 00:35:36.680
at Rouge because that's sort of the

00:35:33.960 --> 00:35:38.119
metric that we've been using to evaluate

00:35:36.680 --> 00:35:40.240
and then if we swap this out for other

00:35:38.119 --> 00:35:43.599
metrics you still see an performance

00:35:40.240 --> 00:35:46.440
improvement over these um search methods

00:35:43.599 --> 00:35:48.119
here um what's the sort of catch here

00:35:46.440 --> 00:35:49.920
the catch here is that MBR requires you

00:35:48.119 --> 00:35:51.599
to sample a hundred things and so this

00:35:49.920 --> 00:35:54.760
is a lot more expensive it's a lot

00:35:51.599 --> 00:35:54.760
slower at infin

00:35:54.800 --> 00:35:58.800
time um yes

00:36:04.200 --> 00:36:10.040
yes a great question why does the beam

00:36:07.000 --> 00:36:14.000
search with more beams perform worse um

00:36:10.040 --> 00:36:16.720
this is a well a relatively welln

00:36:14.000 --> 00:36:19.359
phenomena called the cursive beam search

00:36:16.720 --> 00:36:21.640
which is we actually lost your M so you

00:36:19.359 --> 00:36:24.599
mic and we can speak okay yeah so this

00:36:21.640 --> 00:36:26.079
is called the cursive beam search um and

00:36:24.599 --> 00:36:27.760
the idea here is that beam search is

00:36:26.079 --> 00:36:29.359
like an approxim search right so if you

00:36:27.760 --> 00:36:31.200
add more beams you should be doing

00:36:29.359 --> 00:36:33.319
better and better at finding the maximum

00:36:31.200 --> 00:36:34.800
likelihood thing and generally you are

00:36:33.319 --> 00:36:37.160
you get something that is higher

00:36:34.800 --> 00:36:39.160
probability but as you add more beams

00:36:37.160 --> 00:36:42.319
you also often get something that does

00:36:39.160 --> 00:36:42.319
worse on your Downstream

00:36:44.160 --> 00:36:47.560
metrics back up

00:36:54.240 --> 00:36:58.680
there is that back online

00:36:59.119 --> 00:37:06.520
yeah is that back is that any louder no

00:37:03.520 --> 00:37:06.520
it

00:37:07.000 --> 00:37:12.640
question oh there we go is that better

00:37:09.599 --> 00:37:13.760
great um yeah so why why does this

00:37:12.640 --> 00:37:16.040
happen right why do you get something

00:37:13.760 --> 00:37:18.560
that's higher likelihood but um lower

00:37:16.040 --> 00:37:22.040
performance Downstream um and this is

00:37:18.560 --> 00:37:24.000
like another sort of degeneracy of beam

00:37:22.040 --> 00:37:25.680
search that this idea that the thing

00:37:24.000 --> 00:37:27.440
that is the absolute highest likelihood

00:37:25.680 --> 00:37:28.599
in your distribution might not actually

00:37:27.440 --> 00:37:31.079
be what you want

00:37:28.599 --> 00:37:33.960
Downstream um this is sort of one of the

00:37:31.079 --> 00:37:35.200
other things that people use to motivate

00:37:33.960 --> 00:37:37.599
why you might want to do something like

00:37:35.200 --> 00:37:39.400
MBR instead um and there's a great paper

00:37:37.599 --> 00:37:41.640
about this problem called the inadequacy

00:37:39.400 --> 00:37:43.680
of the mode because beam search is

00:37:41.640 --> 00:37:45.520
looking for the mode of the

00:37:43.680 --> 00:37:47.880
distribution well one other thing I'd

00:37:45.520 --> 00:37:49.680
like to mention is it also goes together

00:37:47.880 --> 00:37:51.119
with how you train your models because

00:37:49.680 --> 00:37:53.760
most of our models are trained using

00:37:51.119 --> 00:37:57.079
maximum likelihood maximum likelihood

00:37:53.760 --> 00:37:59.040
isn't explicitly maximizing our ability

00:37:57.079 --> 00:38:01.079
to get the best answer it's explicitly

00:37:59.040 --> 00:38:05.720
maximizing our ability to estimate the

00:38:01.079 --> 00:38:10.160
the distribution of answers so if I

00:38:05.720 --> 00:38:13.040
say um if you said like what is what is

00:38:10.160 --> 00:38:15.839
your favorite hobby or something like

00:38:13.040 --> 00:38:17.680
that uh what is your favorite hobby in a

00:38:15.839 --> 00:38:19.280
dialogue system often it'll answer I

00:38:17.680 --> 00:38:22.400
don't know or something like that

00:38:19.280 --> 00:38:24.920
because it like you know that that's

00:38:22.400 --> 00:38:26.599
more likely than answering any specific

00:38:24.920 --> 00:38:29.240
hobby like it's more likely than

00:38:26.599 --> 00:38:32.119
answering basketball bowling you know

00:38:29.240 --> 00:38:35.040
whatever else because you have many many

00:38:32.119 --> 00:38:36.560
different options and so like especially

00:38:35.040 --> 00:38:39.880
if it's something that's a little bit

00:38:36.560 --> 00:38:42.160
more comp complicated it will avoid

00:38:39.880 --> 00:38:44.680
answering that and in particular it ends

00:38:42.160 --> 00:38:47.240
up answering very short things for

00:38:44.680 --> 00:38:49.280
example um or sometimes it ends up

00:38:47.240 --> 00:38:51.160
repeating itself over and over again or

00:38:49.280 --> 00:38:53.240
or things like that so it also goes

00:38:51.160 --> 00:38:57.760
together with like the training of the

00:38:53.240 --> 00:38:59.359
model yeah and this is um one of the

00:38:57.760 --> 00:39:01.079
this is still a problem in modern

00:38:59.359 --> 00:39:02.560
systems so if you actually look at the

00:39:01.079 --> 00:39:03.839
single like if you could enumerate

00:39:02.560 --> 00:39:05.680
everything and see the single most

00:39:03.839 --> 00:39:07.520
likely sequence it's often the empty

00:39:05.680 --> 00:39:10.920
sequence just not opening anything at

00:39:07.520 --> 00:39:12.640
all um and so if that's your true mode

00:39:10.920 --> 00:39:16.119
of the distribution then doing better at

00:39:12.640 --> 00:39:16.119
mode seeking is not always like

00:39:16.599 --> 00:39:19.599
helpful

00:39:25.440 --> 00:39:32.960
yes could this be influenced by the

00:39:28.200 --> 00:39:32.960
confidence problem like um how

00:39:37.560 --> 00:39:41.079
so seems

00:39:49.760 --> 00:39:53.599
bees

00:39:51.010 --> 00:39:57.280
[Music]

00:39:53.599 --> 00:39:59.760
might right I think I I think I see

00:39:57.280 --> 00:40:02.000
what you're saying which is that like

00:39:59.760 --> 00:40:04.200
the the confidence gives you the

00:40:02.000 --> 00:40:06.680
confidence of like a single exact

00:40:04.200 --> 00:40:11.000
sequence right not the like actual sort

00:40:06.680 --> 00:40:13.200
of semantic space of and so yeah if you

00:40:11.000 --> 00:40:14.920
looked at just like the if you look at

00:40:13.200 --> 00:40:17.000
just the probability scores you get the

00:40:14.920 --> 00:40:18.520
probability of an exact string when what

00:40:17.000 --> 00:40:20.119
you really actually care about with

00:40:18.520 --> 00:40:22.319
confidence is the probability of sort of

00:40:20.119 --> 00:40:23.800
like things that mean the same thing

00:40:22.319 --> 00:40:25.359
yeah this is um part of why like

00:40:23.800 --> 00:40:28.359
calibration is really hard for long

00:40:25.359 --> 00:40:28.359
sequences

00:40:30.720 --> 00:40:37.319
great so we're g to touch sort of

00:40:34.359 --> 00:40:39.520
briefly on a couple of other things that

00:40:37.319 --> 00:40:40.920
aren't sort of always explicitly

00:40:39.520 --> 00:40:42.480
described in this framework but that you

00:40:40.920 --> 00:40:45.040
can think of as variance of minimum

00:40:42.480 --> 00:40:46.960
based risk um and if you're interested

00:40:45.040 --> 00:40:49.560
in this analysis um I think as Graham

00:40:46.960 --> 00:40:51.800
mentioned earlier um Alex Z is a first

00:40:49.560 --> 00:40:53.680
year MLT and I wrote a paper about this

00:40:51.800 --> 00:40:57.839
um which you can check out if you're

00:40:53.680 --> 00:41:01.200
interested so the um two that I really

00:40:57.839 --> 00:41:03.800
want to touch on here are other sort of

00:41:01.200 --> 00:41:05.240
inference time things you can consider

00:41:03.800 --> 00:41:07.520
which might look a little bit different

00:41:05.240 --> 00:41:09.480
on the first BL um the first of these is

00:41:07.520 --> 00:41:11.680
output ensembling so say you have

00:41:09.480 --> 00:41:13.240
multiple different models and you get

00:41:11.680 --> 00:41:15.480
outputs from all of them and now you

00:41:13.240 --> 00:41:19.560
need to choose a best output among that

00:41:15.480 --> 00:41:21.599
set um one of the sort of common ways to

00:41:19.560 --> 00:41:24.480
do this is to compare like an embedding

00:41:21.599 --> 00:41:25.920
similarity across models like does model

00:41:24.480 --> 00:41:27.560
one think these two things are really

00:41:25.920 --> 00:41:28.880
similar does model two think these two

00:41:27.560 --> 00:41:32.599
things are really similar and try to

00:41:28.880 --> 00:41:34.680
choose something that the um has really

00:41:32.599 --> 00:41:37.319
high similarity with a lot of other

00:41:34.680 --> 00:41:39.200
outputs um of course now that we've just

00:41:37.319 --> 00:41:41.440
recently been talking about MBR you can

00:41:39.200 --> 00:41:44.920
see that you can probably see that this

00:41:41.440 --> 00:41:46.280
is um the same general formulation just

00:41:44.920 --> 00:41:47.880
rather than summing over a set of

00:41:46.280 --> 00:41:49.520
outputs from a single model now you're

00:41:47.880 --> 00:41:52.160
looking at outputs over a whole set of

00:41:49.520 --> 00:41:54.640
models um so some types of ensembling

00:41:52.160 --> 00:41:57.319
fall into this category of minimum based

00:41:54.640 --> 00:42:00.680
risk methods another thing in this

00:41:57.319 --> 00:42:03.280
category is a um sort of recent decoding

00:42:00.680 --> 00:42:06.079
method called self-consistency and the

00:42:03.280 --> 00:42:08.200
idea here is that you want to do

00:42:06.079 --> 00:42:09.359
something like mathematical reasoning

00:42:08.200 --> 00:42:10.599
and you really care about getting the

00:42:09.359 --> 00:42:12.000
final answer right but you don't

00:42:10.599 --> 00:42:15.000
necessarily care about getting all of

00:42:12.000 --> 00:42:18.079
the the reasoning steps in between right

00:42:15.000 --> 00:42:19.520
so you prompt the model for an answer um

00:42:18.079 --> 00:42:20.800
using something like Chain of Thought

00:42:19.520 --> 00:42:22.680
right you ask it to sort of talk through

00:42:20.800 --> 00:42:26.440
the steps it's going to do and then give

00:42:22.680 --> 00:42:28.599
you a final answer um you sample many

00:42:26.440 --> 00:42:30.400
puts using this and then you completely

00:42:28.599 --> 00:42:32.200
throw away the chains of thought um and

00:42:30.400 --> 00:42:35.359
you just take the answer from each

00:42:32.200 --> 00:42:37.640
output um you have that set of answers

00:42:35.359 --> 00:42:38.960
maybe you have like 20 30 100 answers

00:42:37.640 --> 00:42:40.000
you just return the one that was most

00:42:38.960 --> 00:42:43.720
frequently

00:42:40.000 --> 00:42:46.119
generated um what this is doing is a

00:42:43.720 --> 00:42:48.800
type of MBR where the metric that you

00:42:46.119 --> 00:42:51.160
actually care about is exact match of

00:42:48.800 --> 00:42:51.839
this answer right ignoring the rest of

00:42:51.160 --> 00:42:54.079
the

00:42:51.839 --> 00:42:55.800
generation um and so here we have sort

00:42:54.079 --> 00:42:56.839
of the same intuition that we want an

00:42:55.800 --> 00:42:59.160
output

00:42:56.839 --> 00:43:01.520
that is high probability right we're

00:42:59.160 --> 00:43:03.359
getting it generated a lot but also low

00:43:01.520 --> 00:43:06.079
risk not a lot of the other outputs in

00:43:03.359 --> 00:43:08.440
our in our set disagree with this

00:43:06.079 --> 00:43:10.359
answer so those are a couple of

00:43:08.440 --> 00:43:11.920
different variants of methods where

00:43:10.359 --> 00:43:13.880
we're sort of sampling a wide set of

00:43:11.920 --> 00:43:17.359
sequences and trying to choose the best

00:43:13.880 --> 00:43:20.960
one um MBR is one set is one type of

00:43:17.359 --> 00:43:22.680
sort of sequence set reranking method um

00:43:20.960 --> 00:43:24.760
you could do other things to rerank sets

00:43:22.680 --> 00:43:27.400
as well but this is sort of one

00:43:24.760 --> 00:43:30.359
representative class of these yes uh or

00:43:27.400 --> 00:43:32.280
of the of these methods before we get

00:43:30.359 --> 00:43:35.200
into constrain generation those are sort

00:43:32.280 --> 00:43:37.000
of the three broad categories of

00:43:35.200 --> 00:43:39.480
inference methods we'll discuss which is

00:43:37.000 --> 00:43:41.680
sort of sampling from some distribution

00:43:39.480 --> 00:43:45.040
searching over some space of

00:43:41.680 --> 00:43:47.400
distributions and doing some kind of um

00:43:45.040 --> 00:43:48.559
analysis over a set of samples to choose

00:43:47.400 --> 00:43:51.359
which ones they

00:43:48.559 --> 00:43:52.559
return um does anyone have any questions

00:43:51.359 --> 00:43:55.079
at this

00:43:52.559 --> 00:44:00.680
point

00:43:55.079 --> 00:44:00.680
yeah that a model

00:44:05.800 --> 00:44:12.760
cannot yeah like why is averaging model

00:44:08.359 --> 00:44:16.400
weights not MBR um I think it's not MBR

00:44:12.760 --> 00:44:18.559
because the two um the key thing that I

00:44:16.400 --> 00:44:20.880
think really makes a method MBR is this

00:44:18.559 --> 00:44:22.480
concept of comparing between multiple um

00:44:20.880 --> 00:44:24.880
sort of pseudo

00:44:22.480 --> 00:44:26.839
references um and there you don't have

00:44:24.880 --> 00:44:28.359
the same like you aage model way can you

00:44:26.839 --> 00:44:32.440
wind up with sort of a single output on

00:44:28.359 --> 00:44:34.040
the end that maybe is like using like

00:44:32.440 --> 00:44:35.800
information from these two model

00:44:34.040 --> 00:44:38.240
distributions that you've sort of smush

00:44:35.800 --> 00:44:41.160
together um but it's not the same

00:44:38.240 --> 00:44:44.720
concept of like comparing against pseudo

00:44:41.160 --> 00:44:44.720
references or ranking in a

00:44:48.920 --> 00:44:55.599
set right so now this is sort of a this

00:44:52.720 --> 00:44:57.559
was a wide variety of methods to try to

00:44:55.599 --> 00:44:59.040
find an output that's just sort of good

00:44:57.559 --> 00:45:01.440
right we want an output that that is

00:44:59.040 --> 00:45:03.480
nice out of our model um but now we'd

00:45:01.440 --> 00:45:05.880
like to maybe enclose a few additional

00:45:03.480 --> 00:45:08.280
constraints so say I'm asking our model

00:45:05.880 --> 00:45:10.720
for some Hobbies I could use to stay in

00:45:08.280 --> 00:45:11.920
to stay in shape and no matter what I

00:45:10.720 --> 00:45:14.160
don't want the model to recommend

00:45:11.920 --> 00:45:16.880
climbing like I I just I don't want this

00:45:14.160 --> 00:45:18.400
as an option I've tried it I'm not a fan

00:45:16.880 --> 00:45:21.240
um how do I get the model to stop

00:45:18.400 --> 00:45:22.760
suggesting climbing to me and if you've

00:45:21.240 --> 00:45:24.559
sort of played around with some of the

00:45:22.760 --> 00:45:26.200
more recent llms you'd say maybe this is

00:45:24.559 --> 00:45:27.480
easy right you just tell the model the

00:45:26.200 --> 00:45:30.160
instruction that you don't want to talk

00:45:27.480 --> 00:45:31.640
about climbing and having talked to Bard

00:45:30.160 --> 00:45:33.640
recently I can tell you unfortunately

00:45:31.640 --> 00:45:34.800
that it's not that easy so I tell the

00:45:33.640 --> 00:45:36.599
model I don't want to talk about

00:45:34.800 --> 00:45:38.000
climbing it does okay for a little bit

00:45:36.599 --> 00:45:40.920
and then it's like all right but maybe

00:45:38.000 --> 00:45:42.359
you want to try rap climbing um and so

00:45:40.920 --> 00:45:44.559
we could continue trying to instruction

00:45:42.359 --> 00:45:46.200
to our model but maybe there's sort of a

00:45:44.559 --> 00:45:49.079
way to impose this constraint on the

00:45:46.200 --> 00:45:50.680
decoding side instead and so I'd say all

00:45:49.079 --> 00:45:52.960
right I'm going to do something dramatic

00:45:50.680 --> 00:45:54.440
right I know I can manipulate the

00:45:52.960 --> 00:45:56.200
probability distribution I'm just going

00:45:54.440 --> 00:45:57.920
to set the probability of climbing to be

00:45:56.200 --> 00:46:00.440
zero I don't want to see this token like

00:45:57.920 --> 00:46:02.640
I'm I'm completely over it um and this

00:46:00.440 --> 00:46:04.839
is sort of nice in some sense because

00:46:02.640 --> 00:46:06.720
this is pretty easy to do um remember

00:46:04.839 --> 00:46:08.440
we're doing a soft Max over the outputs

00:46:06.720 --> 00:46:10.599
to choose this probability distribution

00:46:08.440 --> 00:46:12.400
and so if we add a huge negative number

00:46:10.599 --> 00:46:14.160
to the logic for climbing before we do

00:46:12.400 --> 00:46:15.520
this softmax its probability is going to

00:46:14.160 --> 00:46:18.640
be basically zero and we're never going

00:46:15.520 --> 00:46:20.240
to see it as an output um but this

00:46:18.640 --> 00:46:22.480
doesn't seem like a perfect solution

00:46:20.240 --> 00:46:24.400
right because you know what if the model

00:46:22.480 --> 00:46:26.160
recommends bouldering to me do I have to

00:46:24.400 --> 00:46:28.599
write like a sort of a list of every

00:46:26.160 --> 00:46:30.599
possible climbing synonym in the world

00:46:28.599 --> 00:46:32.079
um what if there's sort of an allowable

00:46:30.599 --> 00:46:33.920
way to use this token like I want the

00:46:32.079 --> 00:46:35.319
model to suggest hiking because climbing

00:46:33.920 --> 00:46:37.480
up a mountain to see a good view is

00:46:35.319 --> 00:46:38.720
relaxing but that's a use of the word

00:46:37.480 --> 00:46:41.400
climbing and we just said that we can't

00:46:38.720 --> 00:46:43.520
use the word climbing um or what if we

00:46:41.400 --> 00:46:45.480
sort of generate other related terms

00:46:43.520 --> 00:46:47.520
before we get to the restricted term

00:46:45.480 --> 00:46:49.359
like the model starts suggesting maybe

00:46:47.520 --> 00:46:51.480
you can work out by going to an indoor

00:46:49.359 --> 00:46:52.920
rock blank and then what are we going to

00:46:51.480 --> 00:46:54.800
say there's not we can't say rock

00:46:52.920 --> 00:46:57.079
climbing so maybe the model suggests

00:46:54.800 --> 00:46:58.640
rock climbing is rock collecting is a

00:46:57.079 --> 00:47:01.400
hobby to stay in shape and that doesn't

00:46:58.640 --> 00:47:03.480
sound good either um you could continue

00:47:01.400 --> 00:47:05.640
like sort of engineering more and more

00:47:03.480 --> 00:47:06.599
complicated rules here but maybe we

00:47:05.640 --> 00:47:08.760
could do something that's a little

00:47:06.599 --> 00:47:10.559
simpler so what if I just sample a bunch

00:47:08.760 --> 00:47:11.920
of outputs from the model and then I

00:47:10.559 --> 00:47:14.359
check if they're about climbing and I

00:47:11.920 --> 00:47:16.280
get rid of them if they are right um

00:47:14.359 --> 00:47:18.200
this is sort of the advantage that it's

00:47:16.280 --> 00:47:19.599
pretty easy to check after the fact if

00:47:18.200 --> 00:47:22.480
the sequence has satisfied this

00:47:19.599 --> 00:47:24.400
constraint you know we could train some

00:47:22.480 --> 00:47:26.200
smaller model to guess if the topic of a

00:47:24.400 --> 00:47:27.960
sentence is about climbing could check

00:47:26.200 --> 00:47:30.040
for keywords we could have a friend

00:47:27.960 --> 00:47:31.359
who's willing to see this content like

00:47:30.040 --> 00:47:33.040
filter through it and throw everything

00:47:31.359 --> 00:47:36.480
out that's not about climing that is

00:47:33.040 --> 00:47:38.280
about climbing but if this model um

00:47:36.480 --> 00:47:40.119
ascribes really high likelihood to this

00:47:38.280 --> 00:47:42.559
like if this model was trained on you

00:47:40.119 --> 00:47:44.760
know data from CS PhD students this

00:47:42.559 --> 00:47:46.240
could be an extremely high likelihood

00:47:44.760 --> 00:47:48.319
suggestion and so we might need to

00:47:46.240 --> 00:47:49.839
regenerate hundreds or thousands of

00:47:48.319 --> 00:47:52.559
sequences to find something that's not

00:47:49.839 --> 00:47:55.240
about climing um and that feels a little

00:47:52.559 --> 00:47:56.920
bit inefficient right so is there

00:47:55.240 --> 00:47:59.040
something that we can do that's a little

00:47:56.920 --> 00:48:01.599
bit better than that well really we'd

00:47:59.040 --> 00:48:03.200
like to guess at some point during our

00:48:01.599 --> 00:48:05.200
generation if the sequence is going to

00:48:03.200 --> 00:48:08.000
be about climbing and maybe like

00:48:05.200 --> 00:48:10.640
recalibrate or you know we could even

00:48:08.000 --> 00:48:12.079
restart or sort of shape Our Generations

00:48:10.640 --> 00:48:14.520
so that we don't wind up with a sequence

00:48:12.079 --> 00:48:16.319
that's about climbing in the first place

00:48:14.520 --> 00:48:19.359
um one of the methods that we'll discuss

00:48:16.319 --> 00:48:20.920
to do this is a method called fudge um

00:48:19.359 --> 00:48:22.800
and unfortunately in their paper they

00:48:20.920 --> 00:48:24.240
don't have the same anti-climbing bias I

00:48:22.800 --> 00:48:27.000
do so this example is actually about

00:48:24.240 --> 00:48:29.000
formality instead um the idea here is

00:48:27.000 --> 00:48:32.079
that we want a sequence output of the

00:48:29.000 --> 00:48:34.079
model that is sort of satisfies this

00:48:32.079 --> 00:48:36.079
constraint of being formal and the way

00:48:34.079 --> 00:48:39.960
we're going to do this is at each step

00:48:36.079 --> 00:48:41.640
of prediction we get the outputs of what

00:48:39.960 --> 00:48:44.160
the model predicts is the next token

00:48:41.640 --> 00:48:47.319
right this sort of distribution here in

00:48:44.160 --> 00:48:49.760
blue and we also have some second

00:48:47.319 --> 00:48:52.079
distribution which says given sort of

00:48:49.760 --> 00:48:54.480
what we have so far How likely is this

00:48:52.079 --> 00:48:56.920
to be a formal sentence at the end right

00:48:54.480 --> 00:48:58.880
does a sentence that starts do you want

00:48:56.920 --> 00:49:01.200
have a high likelihood of being formal

00:48:58.880 --> 00:49:04.559
versus a sentence that starts do you

00:49:01.200 --> 00:49:07.200
prefer and so this sort of guess at what

00:49:04.559 --> 00:49:09.520
will be formal at the end of the um

00:49:07.200 --> 00:49:10.960
generation will put High likelihood on

00:49:09.520 --> 00:49:13.599
things that result in really formal

00:49:10.960 --> 00:49:15.880
sentences like do you prefer or do you

00:49:13.599 --> 00:49:17.200
thus whereas the original model might

00:49:15.880 --> 00:49:19.440
have higher likelihood on things that

00:49:17.200 --> 00:49:22.559
are maybe more commonly said like do you

00:49:19.440 --> 00:49:24.319
want um so we combine these two

00:49:22.559 --> 00:49:26.280
distributions you can just multiply them

00:49:24.319 --> 00:49:29.079
together and then we sample from this

00:49:26.280 --> 00:49:30.520
modified distribution which now has some

00:49:29.079 --> 00:49:32.359
sort of high weight on things that the

00:49:30.520 --> 00:49:33.559
model thinks are likely but also takes

00:49:32.359 --> 00:49:35.960
into account the likelihood of

00:49:33.559 --> 00:49:38.240
satisfying a constraint um this is

00:49:35.960 --> 00:49:40.640
another sort of method of modifying or

00:49:38.240 --> 00:49:42.520
sampling distribution um with some

00:49:40.640 --> 00:49:44.520
external information here and so there's

00:49:42.520 --> 00:49:47.440
results and sequences that wind up being

00:49:44.520 --> 00:49:48.799
sort of more likely to be formal without

00:49:47.440 --> 00:49:50.280
having to sample a whole bunch of

00:49:48.799 --> 00:49:52.880
sentences and reject the ones that we

00:49:50.280 --> 00:49:54.720
think don't satisfy this constraint so

00:49:52.880 --> 00:49:57.119
how do we get sort of a guess of what

00:49:54.720 --> 00:49:58.839
will be formal at the end of Generation

00:49:57.119 --> 00:50:01.319
Um this is where the name fudge comes

00:49:58.839 --> 00:50:03.319
from the fud stands for future

00:50:01.319 --> 00:50:06.640
discriminator and so what they do is

00:50:03.319 --> 00:50:08.920
they train a model on prefixes to guess

00:50:06.640 --> 00:50:10.400
whether that sequence will be formal um

00:50:08.920 --> 00:50:12.040
you can do this if you have a bunch of

00:50:10.400 --> 00:50:15.319
data that's sort of sorted into formal

00:50:12.040 --> 00:50:17.720
and not formal right every um sort of

00:50:15.319 --> 00:50:20.119
prefix of a sentence in the formal

00:50:17.720 --> 00:50:21.480
category is a training example right you

00:50:20.119 --> 00:50:23.720
know a sentence that starts do you

00:50:21.480 --> 00:50:27.599
prefer you can shop off each token to

00:50:23.720 --> 00:50:29.920
get sort of a um set of sequ of prefixes

00:50:27.599 --> 00:50:31.160
to sequences that have the label formal

00:50:29.920 --> 00:50:33.559
and you can do the same thing to your

00:50:31.160 --> 00:50:34.920
informal set and train a discriminator

00:50:33.559 --> 00:50:36.559
to choose between them to say like

00:50:34.920 --> 00:50:38.400
what's the probability the sentence but

00:50:36.559 --> 00:50:41.160
will belong to the formal set when we

00:50:38.400 --> 00:50:43.319
finish and so this idea of sort of

00:50:41.160 --> 00:50:44.359
trying to guess at a given decoding step

00:50:43.319 --> 00:50:49.480
if we're going to wind up with our

00:50:44.359 --> 00:50:50.799
constraints satisfied at the end um is a

00:50:49.480 --> 00:50:53.000
sort of key way to do constraint

00:50:50.799 --> 00:50:56.000
decoding um and one that we'll return to

00:50:53.000 --> 00:50:58.280
in just a couple slides here

00:50:56.000 --> 00:51:00.440
I want to talk touch on something

00:50:58.280 --> 00:51:03.079
slightly different which is that maybe

00:51:00.440 --> 00:51:04.599
one of the constraints we care about is

00:51:03.079 --> 00:51:07.319
something a little more nebulous like we

00:51:04.599 --> 00:51:09.160
want to match human preference um the

00:51:07.319 --> 00:51:12.079
way that we usually accomplish this

00:51:09.160 --> 00:51:14.920
constraint is a little bit different

00:51:12.079 --> 00:51:16.040
right um this we' usually do through

00:51:14.920 --> 00:51:18.839
like reinforcement learning through

00:51:16.040 --> 00:51:21.559
human feedback um and so we take sort of

00:51:18.839 --> 00:51:24.960
our original model distribution and we

00:51:21.559 --> 00:51:27.960
take a sort of really like tight like

00:51:24.960 --> 00:51:30.200
distrib tion of evidence that says like

00:51:27.960 --> 00:51:31.680
um this model says that this sequence is

00:51:30.200 --> 00:51:33.960
really high reward this sequence is

00:51:31.680 --> 00:51:35.640
really low reward and we try to sort of

00:51:33.960 --> 00:51:38.200
combine them somehow through training so

00:51:35.640 --> 00:51:41.240
we get a new model that is um quote

00:51:38.200 --> 00:51:43.240
unquote aligned and that it has like a

00:51:41.240 --> 00:51:45.280
higher likelihood of giving us things

00:51:43.240 --> 00:51:48.640
that have really high reward according

00:51:45.280 --> 00:51:51.319
to our reward distribution um you can

00:51:48.640 --> 00:51:53.599
view this though as a type of basian

00:51:51.319 --> 00:51:55.119
inference and so what this means is the

00:51:53.599 --> 00:51:57.440
distribution that we really want to get

00:51:55.119 --> 00:51:59.880
at the end is a distribution that

00:51:57.440 --> 00:52:03.160
combines our original models

00:51:59.880 --> 00:52:05.680
distribution and some idea of like How

00:52:03.160 --> 00:52:08.480
likely we are to satisfy the reward

00:52:05.680 --> 00:52:10.720
right um this we do through

00:52:08.480 --> 00:52:12.359
reinforcement learning but if we sort of

00:52:10.720 --> 00:52:14.480
know what these two distributions look

00:52:12.359 --> 00:52:16.119
like we've we've just been talking about

00:52:14.480 --> 00:52:17.680
a lot of methods that modify the

00:52:16.119 --> 00:52:20.119
original models distribution with

00:52:17.680 --> 00:52:21.880
external information it seems like maybe

00:52:20.119 --> 00:52:24.760
we could just add that external

00:52:21.880 --> 00:52:26.200
information in at decoding time to get

00:52:24.760 --> 00:52:29.040
some of the same

00:52:26.200 --> 00:52:31.040
effects um and it turns out you can do

00:52:29.040 --> 00:52:32.799
exactly this so this is a paper from

00:52:31.040 --> 00:52:36.680
last year called reward augmented

00:52:32.799 --> 00:52:39.079
decoding and the idea here is sort of um

00:52:36.680 --> 00:52:41.839
in the same conceptual class as fudge

00:52:39.079 --> 00:52:44.079
but instead of um predicting whether

00:52:41.839 --> 00:52:46.079
we're likely to satisfy the constraint

00:52:44.079 --> 00:52:47.599
we're predicting how much reward we

00:52:46.079 --> 00:52:49.880
think that sequence will have at the end

00:52:47.599 --> 00:52:52.599
of generation so we take our original

00:52:49.880 --> 00:52:54.839
model without doing any rhf and we get

00:52:52.599 --> 00:52:58.160
the output we get the predictions for

00:52:54.839 --> 00:52:59.400
the next token and then we use a model

00:52:58.160 --> 00:53:02.359
that's been trained to predict the

00:52:59.400 --> 00:53:05.040
likely reward given some prefix like a

00:53:02.359 --> 00:53:06.720
future discriminator and we calculate

00:53:05.040 --> 00:53:08.200
the likely reward if we pick each of

00:53:06.720 --> 00:53:09.799
those tokens and then we use the

00:53:08.200 --> 00:53:12.319
combination of those two distributions

00:53:09.799 --> 00:53:13.720
to choose what to decode next um and

00:53:12.319 --> 00:53:16.000
this sort of gives you some of the

00:53:13.720 --> 00:53:18.440
benefits of rlf without actually having

00:53:16.000 --> 00:53:21.200
to do reinforcement learning so it's a

00:53:18.440 --> 00:53:23.160
way of treating like aligning to human

00:53:21.200 --> 00:53:26.839
feedback as just another constraint that

00:53:23.160 --> 00:53:30.400
you can impose at decoding point

00:53:26.839 --> 00:53:32.319
so those were sort of a a subset of the

00:53:30.400 --> 00:53:34.280
um constrains decoding strategies that

00:53:32.319 --> 00:53:35.799
people use um before we get into the

00:53:34.280 --> 00:53:38.400
human and the loop stack are there any

00:53:35.799 --> 00:53:38.400
questions on

00:53:39.040 --> 00:53:43.599
this yes for

00:53:44.960 --> 00:53:48.319
the do you have

00:53:52.799 --> 00:53:57.440
to right so for the discrimin do you

00:53:55.640 --> 00:54:00.000
need to train one for every constraint

00:53:57.440 --> 00:54:01.440
and you do yeah so you need to have some

00:54:00.000 --> 00:54:02.920
set of data that satisfies your

00:54:01.440 --> 00:54:05.319
constraint and some set of data that

00:54:02.920 --> 00:54:08.200
doesn't before you can enforce a new

00:54:05.319 --> 00:54:10.200
constraint in an alternative might be

00:54:08.200 --> 00:54:12.040
like in the paper that's what they did

00:54:10.200 --> 00:54:16.400
but an alternative might be just to

00:54:12.040 --> 00:54:18.359
train a discriminator to determine

00:54:16.400 --> 00:54:20.880
whether any constraint was violated so

00:54:18.359 --> 00:54:23.359
if you have 100 constraints you could do

00:54:20.880 --> 00:54:25.599
a binary prier about whether any

00:54:23.359 --> 00:54:26.880
constraint is violated and then

00:54:25.599 --> 00:54:29.040
also

00:54:26.880 --> 00:54:30.559
sufficient but if you wanted to add a

00:54:29.040 --> 00:54:34.079
new constraint you'd still have to

00:54:30.559 --> 00:54:34.079
retrain or you have to retrain

00:54:35.160 --> 00:54:41.319
or the the reason that this is sort of

00:54:38.119 --> 00:54:43.119
relatively reasonable to do is that this

00:54:41.319 --> 00:54:45.240
determination of if a constraint is

00:54:43.119 --> 00:54:46.960
likely to be violated is sort of a a

00:54:45.240 --> 00:54:48.520
lighter weight or an easier task to

00:54:46.960 --> 00:54:50.520
learn you can use a relatively small

00:54:48.520 --> 00:54:52.079
model for this versus like your big

00:54:50.520 --> 00:54:53.680
model just that has to be able to

00:54:52.079 --> 00:54:55.920
predict the next token for any sequence

00:54:53.680 --> 00:54:58.400
anymore yeah another another like

00:54:55.920 --> 00:55:00.760
interesting thing is if you think about

00:54:58.400 --> 00:55:01.520
it normally you're predicting with your

00:55:00.760 --> 00:55:04.119
big

00:55:01.520 --> 00:55:06.359
softmax like this over all of your

00:55:04.119 --> 00:55:09.680
vocabulary you can even use the same

00:55:06.359 --> 00:55:11.920
representations here to predict with a

00:55:09.680 --> 00:55:13.359
binary classifier uh whether the

00:55:11.920 --> 00:55:14.559
constraint is violated let's say you

00:55:13.359 --> 00:55:17.240
have 100

00:55:14.559 --> 00:55:19.240
constraints this is still a vector of

00:55:17.240 --> 00:55:21.520
size 100 compared to your vector of size

00:55:19.240 --> 00:55:26.240
32,000 that you're using for llama right

00:55:21.520 --> 00:55:28.280
so it's not like this adds the training

00:55:26.240 --> 00:55:32.799
would cost some time but it adds very

00:55:28.280 --> 00:55:32.799
little like inference time I guess

00:55:33.440 --> 00:55:38.960
basically the rock

00:55:35.880 --> 00:55:41.400
sound so when you do the constraint you

00:55:38.960 --> 00:55:43.160
use like a more General

00:55:41.400 --> 00:55:44.680
like do

00:55:43.160 --> 00:55:48.160
notest

00:55:44.680 --> 00:55:50.799
or I guess like in that constraint for

00:55:48.160 --> 00:55:50.799
you can add

00:55:52.559 --> 00:55:57.000
like, is there

00:55:57.880 --> 00:56:00.720
like is there a way to generalize your

00:55:59.400 --> 00:56:04.760
constraint would be like don't talk

00:56:00.720 --> 00:56:07.039
about this whole set of hobes um you

00:56:04.760 --> 00:56:08.960
could do that by training a

00:56:07.039 --> 00:56:10.400
discriminator um by training one

00:56:08.960 --> 00:56:12.359
discriminator that considers all of

00:56:10.400 --> 00:56:15.119
those or by training like a hundred

00:56:12.359 --> 00:56:17.559
different discriminators and then um

00:56:15.119 --> 00:56:19.520
sort of taking like the maximum score

00:56:17.559 --> 00:56:21.240
from any of them right like you want to

00:56:19.520 --> 00:56:23.240
you want to be able to exclude all of

00:56:21.240 --> 00:56:27.799
these things so you consider if any of

00:56:23.240 --> 00:56:30.720
them are violated yeah and for um reward

00:56:27.799 --> 00:56:32.839
augmented recoding how do we sort of

00:56:30.720 --> 00:56:36.039
like frame that reward model or is that

00:56:32.839 --> 00:56:38.400
just come from the previously done rhf

00:56:36.039 --> 00:56:41.079
data that the store from there and then

00:56:38.400 --> 00:56:44.119
you sort of like FR another

00:56:41.079 --> 00:56:47.880
discriminator but this one

00:56:44.119 --> 00:56:50.799
now I I fully understand yeah so how do

00:56:47.880 --> 00:56:52.920
we get the the reward model here this is

00:56:50.799 --> 00:56:55.280
we can use the same data that we' use

00:56:52.920 --> 00:56:58.000
for rhf but we need a slightly different

00:56:55.280 --> 00:57:01.119
model so for rhf we'll train a reward

00:56:58.000 --> 00:57:02.599
model over full sequences right and here

00:57:01.119 --> 00:57:05.280
we need to do the same trick where we

00:57:02.599 --> 00:57:07.280
sort of look at just prefixes and try to

00:57:05.280 --> 00:57:09.640
guess the reward Downstream but if we

00:57:07.280 --> 00:57:12.440
have already have preference data then

00:57:09.640 --> 00:57:15.119
we have some um like we have a data

00:57:12.440 --> 00:57:16.720
source to do this with I think if I'm

00:57:15.119 --> 00:57:19.240
remembering correctly they also had a

00:57:16.720 --> 00:57:20.920
couple more sort of tricks for data

00:57:19.240 --> 00:57:22.640
augmentation to get this to work this is

00:57:20.920 --> 00:57:25.720
sort of like a non-trivial thing to

00:57:22.640 --> 00:57:28.039
figure out um because like reward is

00:57:25.720 --> 00:57:30.200
generally a secret bual

00:57:28.039 --> 00:57:32.280
attribute and also if you don't know

00:57:30.200 --> 00:57:34.160
very much about rhf we're going to cover

00:57:32.280 --> 00:57:36.400
that the future class so don't worry if

00:57:34.160 --> 00:57:37.880
this is a yeah sorry to Jump Ahead a

00:57:36.400 --> 00:57:39.880
little no no

00:57:37.880 --> 00:57:43.640
wores

00:57:39.880 --> 00:57:47.240
yeah application this like why would we

00:57:43.640 --> 00:57:49.640
doing this to ensure it could be like

00:57:47.240 --> 00:57:52.839
our llm would want to highlight certain

00:57:49.640 --> 00:57:53.799
qualities like we want our evence to be

00:57:52.839 --> 00:57:55.960
more

00:57:53.799 --> 00:57:57.839
empathetic is there

00:57:55.960 --> 00:57:59.440
something yeah like what are the real

00:57:57.839 --> 00:58:01.280
world applications like could we use

00:57:59.440 --> 00:58:03.680
this to make L more empathetic or

00:58:01.280 --> 00:58:06.359
something yeah any any real attribute

00:58:03.680 --> 00:58:08.000
that you can sort of collect like

00:58:06.359 --> 00:58:09.839
positive and negative data for you could

00:58:08.000 --> 00:58:12.200
do this kind of constraints for I think

00:58:09.839 --> 00:58:15.119
the the ones you see most commonly are

00:58:12.200 --> 00:58:16.480
the human preference and then like

00:58:15.119 --> 00:58:18.839
negative constraints like you don't want

00:58:16.480 --> 00:58:20.000
your model to generate offensive content

00:58:18.839 --> 00:58:21.839
and if you can build like a good

00:58:20.000 --> 00:58:23.319
discriminator for is a sentence going in

00:58:21.839 --> 00:58:26.160
a really offensive Direction you can

00:58:23.319 --> 00:58:28.440
kind of stop it from gener

00:58:26.160 --> 00:58:30.480
yeah would it be a good idea if you

00:58:28.440 --> 00:58:33.760
generate a bunch of cons and ask the

00:58:30.480 --> 00:58:35.480
model itself whether it violates the

00:58:33.760 --> 00:58:37.319
yeah you could do that for sure could

00:58:35.480 --> 00:58:38.920
you ask like could you generate a bunch

00:58:37.319 --> 00:58:42.440
of samples and ask the model if it

00:58:38.920 --> 00:58:44.720
violates the constraint um this is also

00:58:42.440 --> 00:58:47.119
a type of sort of sample and then rerank

00:58:44.720 --> 00:58:52.319
strategy um but yeah this would be sort

00:58:47.119 --> 00:58:54.000
of a more um clever like less

00:58:52.319 --> 00:58:55.559
heavyweight version of this checking if

00:58:54.000 --> 00:58:57.319
it's about climate means right you'd

00:58:55.559 --> 00:58:58.520
like ask the model if it violated the

00:58:57.319 --> 00:59:00.160
constraint and if it's a good enough

00:58:58.520 --> 00:59:02.480
model it could probably do that pretty

00:59:00.160 --> 00:59:05.160
well I suppose in that case you don't

00:59:02.480 --> 00:59:08.160
have to thing anything yeah yeah and

00:59:05.160 --> 00:59:10.359
this is sort of a general like the

00:59:08.160 --> 00:59:12.240
generating text that like satisfies a

00:59:10.359 --> 00:59:14.079
constraint is harder than checking if a

00:59:12.240 --> 00:59:16.280
text satisfies a constraint so even if

00:59:14.079 --> 00:59:17.880
the model isn't good about like not

00:59:16.280 --> 00:59:19.440
generating text about climbing when you

00:59:17.880 --> 00:59:20.520
tell it to it might be able to tell if

00:59:19.440 --> 00:59:23.640
text is

00:59:20.520 --> 00:59:26.640
about yeah yeah so how do

00:59:23.640 --> 00:59:26.640
you

00:59:28.400 --> 00:59:32.359
have different

00:59:32.920 --> 00:59:36.319
different you have

00:59:36.599 --> 00:59:42.119
to yeah like how do you collect the data

00:59:38.839 --> 00:59:45.720
to train this discriminator um generally

00:59:42.119 --> 00:59:47.160
you're going to see like you'll look to

00:59:45.720 --> 00:59:48.720
see if there are data sets that already

00:59:47.160 --> 00:59:50.160
captured this attribute or you could

00:59:48.720 --> 00:59:51.599
sort of write her istics to try to

00:59:50.160 --> 00:59:53.839
recover it if it's an attribute that not

00:59:51.599 --> 00:59:55.480
a lot of other people care about like

00:59:53.839 --> 00:59:58.280
you could write your puristic to check

00:59:55.480 --> 01:00:00.160
if text is about climbing for instance

00:59:58.280 --> 01:00:02.359
um and then try to recover what noisy

01:00:00.160 --> 01:00:04.200
samples of data that is or is not about

01:00:02.359 --> 01:00:05.559
climbing maybe you could scrape a

01:00:04.200 --> 01:00:07.000
climbing forum and then scrape like a

01:00:05.559 --> 01:00:09.079
hiking forum and use the difference

01:00:07.000 --> 01:00:10.319
between them um but for a lot of tests

01:00:09.079 --> 01:00:11.760
there's actually pretty good data sets

01:00:10.319 --> 01:00:14.400
already out there for this so there's

01:00:11.760 --> 01:00:17.480
like in there's a lot of style transfer

01:00:14.400 --> 01:00:20.200
tasks that are like go from informal to

01:00:17.480 --> 01:00:22.240
formal or go from this to that or like

01:00:20.200 --> 01:00:24.039
make this text in an iic contamin and

01:00:22.240 --> 01:00:26.559
you can find like data from those

01:00:24.039 --> 01:00:26.559
sources

01:00:26.799 --> 01:00:31.599
we never like talked about F yet but I'm

01:00:29.520 --> 01:00:34.520
really curious with like the word a

01:00:31.599 --> 01:00:38.039
beting whether it would perform better

01:00:34.520 --> 01:00:39.079
than like fineing on RF like certainly

01:00:38.039 --> 01:00:42.720
more

01:00:39.079 --> 01:00:45.039
efficient but I I was I think this is a

01:00:42.720 --> 01:00:49.760
comparison they make in their paper but

01:00:45.039 --> 01:00:52.520
I don't remember their pun on yeah um in

01:00:49.760 --> 01:00:55.280
general there's this sort of a like you

01:00:52.520 --> 01:00:57.039
can pay a onetime kind of heavy cost to

01:00:55.280 --> 01:00:58.880
fine-tune or you can pay costs at

01:00:57.039 --> 01:01:01.160
inference time every time to make sort

01:00:58.880 --> 01:01:03.880
of a to make your model better in any of

01:01:01.160 --> 01:01:06.160
these ways and depending on how much

01:01:03.880 --> 01:01:09.119
inference you're playing do like one or

01:01:06.160 --> 01:01:09.119
the other of these could be

01:01:11.240 --> 01:01:16.400
better

01:01:12.839 --> 01:01:19.200
great so now we're going to talk about

01:01:16.400 --> 01:01:21.160
sort of methods for introducing human

01:01:19.200 --> 01:01:22.680
interaction into the decoding process

01:01:21.160 --> 01:01:25.240
and everything we've looked at so far

01:01:22.680 --> 01:01:26.920
has been very sort of black booss kind

01:01:25.240 --> 01:01:28.920
of hands off right like you give the

01:01:26.920 --> 01:01:30.640
model M some input maybe we do some kind

01:01:28.920 --> 01:01:33.640
of manipulation on the decoding side you

01:01:30.640 --> 01:01:37.160
get one output back right um but in a

01:01:33.640 --> 01:01:38.920
lot of situations where maybe you have

01:01:37.160 --> 01:01:40.960
some high-risk application and you need

01:01:38.920 --> 01:01:42.640
somebody to be consistently monitoring

01:01:40.960 --> 01:01:43.799
and maybe intervening or you're doing

01:01:42.640 --> 01:01:46.359
something where you want to do some kind

01:01:43.799 --> 01:01:47.960
of human AI collaboration um and you

01:01:46.359 --> 01:01:49.160
want to be able to go back and forth or

01:01:47.960 --> 01:01:50.960
you want to have a conversation with the

01:01:49.160 --> 01:01:53.480
model what you're actually doing is sort

01:01:50.960 --> 01:01:54.960
of a series of decodings with human

01:01:53.480 --> 01:01:56.319
intervention in between

01:01:54.960 --> 01:01:58.640
um and I'm going to talk about a couple

01:01:56.319 --> 01:02:00.760
of these strategies briefly I think if

01:01:58.640 --> 01:02:02.200
you've used sort of a modern llm you're

01:02:00.760 --> 01:02:04.440
probably familiar with at least a few of

01:02:02.200 --> 01:02:06.720
them already um we'll sort of put names

01:02:04.440 --> 01:02:08.359
to each of them and the set of examples

01:02:06.720 --> 01:02:10.880
that we're running with here are from a

01:02:08.359 --> 01:02:13.880
paper called wordcraft which is about um

01:02:10.880 --> 01:02:15.480
story generation with llm assistants but

01:02:13.880 --> 01:02:17.559
these can also be applied sort of more

01:02:15.480 --> 01:02:20.319
generally to any kind of task where

01:02:17.559 --> 01:02:23.799
you'd want to go back and forth with a

01:02:20.319 --> 01:02:25.319
model um the sort of easiest or maybe

01:02:23.799 --> 01:02:27.599
simplest place to start here is just

01:02:25.319 --> 01:02:29.760
with interleaving text right you can

01:02:27.599 --> 01:02:31.400
choose when the model starts and stops

01:02:29.760 --> 01:02:33.720
decoding and you can choose when a human

01:02:31.400 --> 01:02:34.920
is writing text in between and you can

01:02:33.720 --> 01:02:36.680
condition your model in sort of a

01:02:34.920 --> 01:02:39.240
mixture of human and model generated

01:02:36.680 --> 01:02:41.279
text to choose what to continue next um

01:02:39.240 --> 01:02:43.680
you can also do something like have the

01:02:41.279 --> 01:02:45.319
model generate a set of text edit that

01:02:43.680 --> 01:02:47.119
text in some way maybe the human is

01:02:45.319 --> 01:02:48.640
imposing some really subtle constraint

01:02:47.119 --> 01:02:50.559
like I want it to sound like my writing

01:02:48.640 --> 01:02:52.200
style we don't have a discriminator for

01:02:50.559 --> 01:02:54.119
this but the human can sort of modify

01:02:52.200 --> 01:02:55.680
the text and then continue generating

01:02:54.119 --> 01:02:57.160
from that point and that will influence

01:02:55.680 --> 01:03:01.160
the style of the text that continues

01:02:57.160 --> 01:03:03.240
being generative um a this case here is

01:03:01.160 --> 01:03:04.720
sort of a you're writing a story

01:03:03.240 --> 01:03:06.520
together and so you're going back and

01:03:04.720 --> 01:03:07.799
forth and editing the text like that but

01:03:06.520 --> 01:03:10.319
you can also think of any kind of

01:03:07.799 --> 01:03:11.920
conversation with a model as the same

01:03:10.319 --> 01:03:15.319
kind of interleaving of text right the

01:03:11.920 --> 01:03:17.000
model gives some um text you provide

01:03:15.319 --> 01:03:18.599
some text you go back and forth on like

01:03:17.000 --> 01:03:20.480
who's providing the text that conditions

01:03:18.599 --> 01:03:23.039
the

01:03:20.480 --> 01:03:24.880
model you also might want to do things

01:03:23.039 --> 01:03:26.760
like more fine brain replace

01:03:24.880 --> 01:03:28.559
so here the person has highlighted some

01:03:26.760 --> 01:03:31.640
text and said like make this more

01:03:28.559 --> 01:03:33.960
descriptive or shorten this to two words

01:03:31.640 --> 01:03:36.079
or maybe you want some additional

01:03:33.960 --> 01:03:38.520
constraint like can this be happier can

01:03:36.079 --> 01:03:40.960
this be sad like change the ending or

01:03:38.520 --> 01:03:43.760
something um you can accomplish this in

01:03:40.960 --> 01:03:45.799
a variety of ways um here this is done

01:03:43.760 --> 01:03:47.680
through input manipulation so you prompt

01:03:45.799 --> 01:03:50.359
your model differently with different

01:03:47.680 --> 01:03:52.200
constraints you can also do this with an

01:03:50.359 --> 01:03:54.440
actual modeling change like if you want

01:03:52.200 --> 01:03:56.119
some kind of infilling model um

01:03:54.440 --> 01:03:57.720
particularly for things like code this

01:03:56.119 --> 01:04:01.119
can be helpful so you want context from

01:03:57.720 --> 01:04:02.440
left and right sides um or you can do

01:04:01.119 --> 01:04:03.799
this with the decoding changes that we

01:04:02.440 --> 01:04:05.960
talked about in the previous section

01:04:03.799 --> 01:04:07.799
right you could add a discriminator for

01:04:05.960 --> 01:04:09.680
descriptiveness of text or you could do

01:04:07.799 --> 01:04:11.680
some kind of sampling ranking method to

01:04:09.680 --> 01:04:13.880
recover a more descriptive

01:04:11.680 --> 01:04:16.640
output another thing that's very common

01:04:13.880 --> 01:04:17.960
in this space is sampling and reranking

01:04:16.640 --> 01:04:20.839
methods where the human is the one

01:04:17.960 --> 01:04:23.640
choosing what to return right so in

01:04:20.839 --> 01:04:25.960
wordcraft you see a set of choices and

01:04:23.640 --> 01:04:28.200
you can choose text to insert but more

01:04:25.960 --> 01:04:30.720
commonly in something like um chat gbt

01:04:28.200 --> 01:04:33.160
or Bard you see this little option to

01:04:30.720 --> 01:04:34.880
regenerate text right you as the human

01:04:33.160 --> 01:04:36.160
can reject the text and say like no I

01:04:34.880 --> 01:04:38.680
don't like this give me a different

01:04:36.160 --> 01:04:41.359
output and this is also sort of a way of

01:04:38.680 --> 01:04:44.079
controlling decoding um just by doing it

01:04:41.359 --> 01:04:46.319
on on a human rather in an algorithmic

01:04:44.079 --> 01:04:49.279
level of course you don't necessarily

01:04:46.319 --> 01:04:51.200
need a human in here and so um some

01:04:49.279 --> 01:04:52.960
recent work has looked at functionally

01:04:51.200 --> 01:04:55.799
using models to make these decisions

01:04:52.960 --> 01:04:57.480
instead um this is a a a prompting paper

01:04:55.799 --> 01:05:00.359
called free of thought which was sort of

01:04:57.480 --> 01:05:02.279
very popular on Twitter last summer um

01:05:00.359 --> 01:05:06.119
and the idea here is that you're going

01:05:02.279 --> 01:05:08.480
to generate um several smaller sequences

01:05:06.119 --> 01:05:11.200
um like a couple of sentences a

01:05:08.480 --> 01:05:13.160
reasoning step or a thought in the paper

01:05:11.200 --> 01:05:14.839
and you're going to use a model to

01:05:13.160 --> 01:05:16.839
choose which ones to continue and you

01:05:14.839 --> 01:05:19.000
can do different sort of constraints

01:05:16.839 --> 01:05:21.960
here like I want to sort of rank this

01:05:19.000 --> 01:05:25.079
set of three or maybe I want to predict

01:05:21.960 --> 01:05:26.839
if any in this set is wrong like is this

01:05:25.079 --> 01:05:29.400
a good reasoning step and if the model

01:05:26.839 --> 01:05:32.240
says no you no longer continue that but

01:05:29.400 --> 01:05:33.559
the idea here is through prompting

01:05:32.240 --> 01:05:35.640
really achieving something that's sort

01:05:33.559 --> 01:05:38.960
of if you squint at it looks a lot like

01:05:35.640 --> 01:05:41.279
beam search right instead of doing a um

01:05:38.960 --> 01:05:43.160
like token level thing and making a

01:05:41.279 --> 01:05:45.079
decision based on likelihood you're

01:05:43.160 --> 01:05:47.880
generating sort of several sentences out

01:05:45.079 --> 01:05:50.599
a time and making a decision based on

01:05:47.880 --> 01:05:52.359
this models feedback right this signal

01:05:50.599 --> 01:05:53.799
from an external source which here is a

01:05:52.359 --> 01:05:55.279
model but could also be a human if

01:05:53.799 --> 01:05:57.920
you're willing willing to sort of wait

01:05:55.279 --> 01:06:01.559
around for them to make the decision and

01:05:57.920 --> 01:06:03.839
so this is a way of sort of giving

01:06:01.559 --> 01:06:06.640
feedback on a broader level than single

01:06:03.839 --> 01:06:09.079
tokens um to guide a decoding process to

01:06:06.640 --> 01:06:09.079
a final

01:06:09.839 --> 01:06:15.079
outut so the last couple of things we'll

01:06:12.760 --> 01:06:17.520
talk about here are sort of practical

01:06:15.079 --> 01:06:19.839
considerations speed choosing decoding

01:06:17.520 --> 01:06:22.599
methods um but I can take any questions

01:06:19.839 --> 01:06:22.599
before that

01:06:23.000 --> 01:06:26.000
to

01:06:26.760 --> 01:06:32.920
great so how do you make this fast and

01:06:30.359 --> 01:06:34.920
in particular if you've ever tried to

01:06:32.920 --> 01:06:36.920
sort of Benchmark performance of a model

01:06:34.920 --> 01:06:38.720
what you realize pretty quickly is that

01:06:36.920 --> 01:06:40.720
the vast majority of time is actually

01:06:38.720 --> 01:06:43.440
spent in decoding you have to generate

01:06:40.720 --> 01:06:45.319
one token at a time you have to sort of

01:06:43.440 --> 01:06:46.920
pass that back through the model to get

01:06:45.319 --> 01:06:51.279
conditioning to generate the next token

01:06:46.920 --> 01:06:53.599
and so this is um generally fairly slow

01:06:51.279 --> 01:06:54.839
um this is sort of a a major impediment

01:06:53.599 --> 01:06:56.359
if you're d to do something like a

01:06:54.839 --> 01:06:57.839
streaming application where you want or

01:06:56.359 --> 01:06:59.559
a chat application where you don't want

01:06:57.839 --> 01:07:03.599
the person to be waiting around for an

01:06:59.559 --> 01:07:06.799
answer um one way to do this is a method

01:07:03.599 --> 01:07:09.160
called Spectra of decoding and this is a

01:07:06.799 --> 01:07:12.599
method where you're using a smaller

01:07:09.160 --> 01:07:14.039
model um not as like we're in contrast

01:07:12.599 --> 01:07:16.240
of decoding right we're using a smaller

01:07:14.039 --> 01:07:17.559
model to decide what not to generate but

01:07:16.240 --> 01:07:20.119
here we're using a smaller model to

01:07:17.559 --> 01:07:21.880
decide be what to generate um and the

01:07:20.119 --> 01:07:24.960
idea here is that most of these tokens

01:07:21.880 --> 01:07:26.480
are maybe not super hard to side it's

01:07:24.960 --> 01:07:27.400
just that occasionally the bigger model

01:07:26.480 --> 01:07:30.240
might want to go in a different

01:07:27.400 --> 01:07:32.920
direction so these green tokens here are

01:07:30.240 --> 01:07:35.160
generated by a smaller model our amateur

01:07:32.920 --> 01:07:37.079
model here and the larger model acts

01:07:35.160 --> 01:07:39.960
largely as a verifier and what it does

01:07:37.079 --> 01:07:43.000
is it checks if the output so far is

01:07:39.960 --> 01:07:44.920
going in a an a Direction that's sort of

01:07:43.000 --> 01:07:46.400
in distribution for the big model like

01:07:44.920 --> 01:07:49.240
something that's within the realm of

01:07:46.400 --> 01:07:50.720
what it might SLE and to there's sort of

01:07:49.240 --> 01:07:52.400
an involved discussion in this paper of

01:07:50.720 --> 01:07:55.200
how you determine if something is in

01:07:52.400 --> 01:07:58.000
distribution um so here the smaller

01:07:55.200 --> 01:08:00.240
models generates like five or six tokens

01:07:58.000 --> 01:08:02.559
that the larger model says okay this

01:08:00.240 --> 01:08:03.680
looks great until it hits a token that

01:08:02.559 --> 01:08:06.079
the larger model would not have

01:08:03.680 --> 01:08:07.920
generated in that circumstance and then

01:08:06.079 --> 01:08:10.279
the larger model rejects that token and

01:08:07.920 --> 01:08:13.000
generates a different token instead so

01:08:10.279 --> 01:08:15.440
you can see here each of these red and

01:08:13.000 --> 01:08:17.600
then blue sections is where the larger

01:08:15.440 --> 01:08:19.400
model has rejected something and has to

01:08:17.600 --> 01:08:21.920
actually autor regressively decode a

01:08:19.400 --> 01:08:24.199
single token by contrast if you were

01:08:21.920 --> 01:08:27.359
doing regular decoding at each

01:08:24.199 --> 01:08:28.799
individual token in this sequence the um

01:08:27.359 --> 01:08:31.640
larger model would have had to make the

01:08:28.799 --> 01:08:35.359
fall forward pass to decoda token so

01:08:31.640 --> 01:08:37.359
here rather than de doing maybe what

01:08:35.359 --> 01:08:39.239
probably like 20ish decoding steps to

01:08:37.359 --> 01:08:41.560
get this full sequence the larger model

01:08:39.239 --> 01:08:43.040
has done about eight decoring steps and

01:08:41.560 --> 01:08:47.560
everything else is able to sort of

01:08:43.040 --> 01:08:49.759
verify a block of tokens at once um this

01:08:47.560 --> 01:08:51.400
sort of idea of like using a smaller

01:08:49.759 --> 01:08:54.120
model as an approximation is pretty

01:08:51.400 --> 01:08:55.839
powerful um and there's some great um

01:08:54.120 --> 01:08:58.159
followup work cons specul decoding and

01:08:55.839 --> 01:08:59.000
sort of ways to do this faster or with

01:08:58.159 --> 01:09:01.520
stronger

01:08:59.000 --> 01:09:04.839
guarantees um but this General concept

01:09:01.520 --> 01:09:06.920
is I would bet probably how models like

01:09:04.839 --> 01:09:09.080
um part of how models like chat GPT or

01:09:06.920 --> 01:09:11.159
Bard are sort of generating text so

01:09:09.080 --> 01:09:13.120
quickly um there's another element here

01:09:11.159 --> 01:09:16.159
which is like the model architecture

01:09:13.120 --> 01:09:17.679
being sparse but I think that um if you

01:09:16.159 --> 01:09:19.920
folks talk about mixture of experts we

01:09:17.679 --> 01:09:22.880
might get into that

01:09:19.920 --> 01:09:26.080
later um how do you do this kind of fast

01:09:22.880 --> 01:09:27.679
inference um libraries like BLM will

01:09:26.080 --> 01:09:29.440
Implement things I think Implement

01:09:27.679 --> 01:09:32.199
speculative decoding and Implement sort

01:09:29.440 --> 01:09:34.400
of Hardware level tricks like choosing

01:09:32.199 --> 01:09:37.799
which attention um weights to Cash wear

01:09:34.400 --> 01:09:39.199
to do faster inflence um there's also

01:09:37.799 --> 01:09:40.799
great libraries for doing things like

01:09:39.199 --> 01:09:42.679
constraint decoding so things like

01:09:40.799 --> 01:09:45.520
outlines will let you set constraints

01:09:42.679 --> 01:09:46.960
like I want my outputs to all be Json

01:09:45.520 --> 01:09:48.640
and it will impose additional

01:09:46.960 --> 01:09:50.839
constraints during decoding to ensure

01:09:48.640 --> 01:09:52.279
that that happens and then pretty much

01:09:50.839 --> 01:09:53.960
anything in these first couple of

01:09:52.279 --> 01:09:56.560
sections we talked about um like

01:09:53.960 --> 01:09:58.440
sampling mode seeking search and

01:09:56.560 --> 01:10:00.400
sometimes MBR will also be implemented

01:09:58.440 --> 01:10:05.080
in pretty much any Library you use for

01:10:00.400 --> 01:10:07.679
models like huggingface Fair seek or

01:10:05.080 --> 01:10:10.000
Jacks so to kind of take a step back

01:10:07.679 --> 01:10:12.520
here is when you get to the end of class

01:10:10.000 --> 01:10:15.640
um there's really two broad categories

01:10:12.520 --> 01:10:17.679
of methods that we talked about today um

01:10:15.640 --> 01:10:20.360
given our initial distribution from the

01:10:17.679 --> 01:10:22.600
model for a next token given our our

01:10:20.360 --> 01:10:24.920
input we can do two kind of different

01:10:22.600 --> 01:10:26.400
things we can each individual decoding

01:10:24.920 --> 01:10:28.360
step choose some kind of function to

01:10:26.400 --> 01:10:30.280
manipulate this distribution and this

01:10:28.360 --> 01:10:32.280
could be something like short like

01:10:30.280 --> 01:10:33.960
cutting off the long tail like modifying

01:10:32.280 --> 01:10:36.239
the temperature or adding external

01:10:33.960 --> 01:10:38.400
information from another model or from a

01:10:36.239 --> 01:10:41.480
discriminator model

01:10:38.400 --> 01:10:43.159
right or we can over a larger part of

01:10:41.480 --> 01:10:45.120
the decoding process choose some

01:10:43.159 --> 01:10:47.120
function to choose between sequences and

01:10:45.120 --> 01:10:49.199
this could be like choosing between next

01:10:47.120 --> 01:10:51.679
tokens in beam search when we pruning

01:10:49.199 --> 01:10:53.120
beams this could be choosing from Full

01:10:51.679 --> 01:10:56.760
sequences when we're doing something

01:10:53.120 --> 01:10:58.040
like MB r or sample and rerank methods

01:10:56.760 --> 01:11:00.239
um and you can do these two things in

01:10:58.040 --> 01:11:01.440
parallel right you can choose like a

01:11:00.239 --> 01:11:03.159
different function to manipulate the

01:11:01.440 --> 01:11:04.760
next token distribution and then some

01:11:03.159 --> 01:11:06.199
sort of like broader thing to choose

01:11:04.760 --> 01:11:08.280
what you do with the full sequences you

01:11:06.199 --> 01:11:09.920
get out of that distribution um but

01:11:08.280 --> 01:11:12.040
there are sort of these two broad

01:11:09.920 --> 01:11:14.880
categories of

01:11:12.040 --> 01:11:17.440
decoding so what should you take away

01:11:14.880 --> 01:11:19.400
from this um I think a couple of things

01:11:17.440 --> 01:11:21.000
you decoding methods can be really

01:11:19.400 --> 01:11:23.040
powerful to control features of your

01:11:21.000 --> 01:11:25.040
output if you want to impose particular

01:11:23.040 --> 01:11:26.679
constraints if you want to factor in

01:11:25.040 --> 01:11:27.960
reward function or factor in a data

01:11:26.679 --> 01:11:31.800
source that you maybe didn't have at

01:11:27.960 --> 01:11:34.239
training time um and to some extent you

01:11:31.800 --> 01:11:36.120
can do a more expensive decoding method

01:11:34.239 --> 01:11:37.520
to compensate for a worse model or to

01:11:36.120 --> 01:11:39.080
compensate for a model that hasn't been

01:11:37.520 --> 01:11:42.480
trained to do exactly the thing you want

01:11:39.080 --> 01:11:44.800
it to do um of course you can't you know

01:11:42.480 --> 01:11:47.679
use this to make gpt2 small as good as

01:11:44.800 --> 01:11:49.840
gp4 but you can sort of for some points

01:11:47.679 --> 01:11:51.679
in the middle spend more um computed

01:11:49.840 --> 01:11:53.159
inference time to pay for not spending

01:11:51.679 --> 01:11:55.639
as much computed training time and

01:11:53.159 --> 01:11:57.440
particularly if you don't have access to

01:11:55.639 --> 01:11:59.400
the kind of giant gpus you might need to

01:11:57.440 --> 01:12:01.840
continue fine-tuning your model this can

01:11:59.400 --> 01:12:05.679
be a really a really powerful

01:12:01.840 --> 01:12:07.800
alternative um yeah so say like you're

01:12:05.679 --> 01:12:12.560
building like something in production

01:12:07.800 --> 01:12:15.920
right people usually do um sort of like

01:12:12.560 --> 01:12:18.760
that you know inance before cling to see

01:12:15.920 --> 01:12:21.840
if it's G to work at do

01:12:18.760 --> 01:12:25.080
that like try to see like if you have a

01:12:21.840 --> 01:12:26.800
model that you can do some kind of

01:12:25.080 --> 01:12:29.199
expensive decoding method for to get

01:12:26.800 --> 01:12:31.120
good outputs is it then worth try

01:12:29.199 --> 01:12:34.000
training that model right um there's

01:12:31.120 --> 01:12:36.560
some great recent work on like training

01:12:34.000 --> 01:12:39.400
models to produce the same kind of

01:12:36.560 --> 01:12:40.760
outputs you get out of MVR without um

01:12:39.400 --> 01:12:43.239
actually doing a really expensive

01:12:40.760 --> 01:12:45.600
inference Stu so at some level like yeah

01:12:43.239 --> 01:12:48.120
you can decide like this model is good

01:12:45.600 --> 01:12:49.920
enough with its expensive method we can

01:12:48.120 --> 01:12:50.920
try to make it cheaper by spending more

01:12:49.920 --> 01:12:53.960
money on

01:12:50.920 --> 01:12:55.520
funing um but that's not it's not like

01:12:53.960 --> 01:12:57.320
necessarily guaranteed that that's will

01:12:55.520 --> 01:13:00.679
be the case

01:12:57.320 --> 01:13:03.040
Okay um the methods that we looked at

01:13:00.679 --> 01:13:06.199
have these sort of trade-offs in quality

01:13:03.040 --> 01:13:07.960
in diversity and in inference speed so

01:13:06.199 --> 01:13:10.320
sampling from your model directly is

01:13:07.960 --> 01:13:13.120
pretty fast to do you get really diverse

01:13:10.320 --> 01:13:14.960
outputs but it tends to be lower quality

01:13:13.120 --> 01:13:16.320
um whereas more restricted sampling

01:13:14.960 --> 01:13:18.520
these sort of mode seeking search

01:13:16.320 --> 01:13:20.639
methods tend to be higher quality but

01:13:18.520 --> 01:13:21.880
you get less less diverse outputs and

01:13:20.639 --> 01:13:23.560
that's why we have these methods like

01:13:21.880 --> 01:13:26.719
diverse and stochastic resarch to

01:13:23.560 --> 01:13:28.760
counter this a bit um and then methods

01:13:26.719 --> 01:13:30.400
like MBR or other sample and rerank

01:13:28.760 --> 01:13:32.679
methods tend to be very high quality

01:13:30.400 --> 01:13:34.280
outputs but you pay for this with much

01:13:32.679 --> 01:13:36.520
slower inference

01:13:34.280 --> 01:13:38.679
time um but if I can kind of convince

01:13:36.520 --> 01:13:41.560
you of anything today I think it would

01:13:38.679 --> 01:13:43.600
be this which is that these the decoding

01:13:41.560 --> 01:13:45.600
method you choose for your model has a

01:13:43.600 --> 01:13:47.960
really strong impact on performance

01:13:45.600 --> 01:13:49.520
Downstream um you can get radically

01:13:47.960 --> 01:13:51.239
different results out of the same model

01:13:49.520 --> 01:13:52.639
without doing any additional training

01:13:51.239 --> 01:13:55.120
just by choosing the different decoding

01:13:52.639 --> 01:13:57.880
method that you might want to try and so

01:13:55.120 --> 01:13:59.679
when you sort of let your libraries pick

01:13:57.880 --> 01:14:01.159
a quote unquote like sensible default

01:13:59.679 --> 01:14:03.760
you can leave a lot of performance on

01:14:01.159 --> 01:14:06.480
the train on the table so I encourage

01:14:03.760 --> 01:14:08.199
you folks that if if you're um deploying

01:14:06.480 --> 01:14:09.760
models in production or if you're doing

01:14:08.199 --> 01:14:10.840
research or you know maybe look at your

01:14:09.760 --> 01:14:13.280
outputs and your model has some

01:14:10.840 --> 01:14:15.320
undesirable behaviors to consider if the

01:14:13.280 --> 01:14:17.800
decoding method you're using is imposing

01:14:15.320 --> 01:14:20.000
some kind of Intuition or some kind of

01:14:17.800 --> 01:14:21.840
inductive bias and if you can alter that

01:14:20.000 --> 01:14:24.239
to get some of these behaviors without

01:14:21.840 --> 01:14:26.320
resorting to additional training

01:14:24.239 --> 01:14:28.719
um and that's sort of the end I can take

01:14:26.320 --> 01:14:28.719
any other

01:14:34.320 --> 01:14:38.719
questions okay um yeah I guess we don't

01:14:37.199 --> 01:14:41.360
have any questions we can take questions

01:14:38.719 --> 01:14:45.560
up here um one one thing I'd like to

01:14:41.360 --> 01:14:47.679
point out also is that um I I love the

01:14:45.560 --> 01:14:50.760
final thing that Amanda said here

01:14:47.679 --> 01:14:54.199
another thing is that my impression from

01:14:50.760 --> 01:14:56.400
dealing with things is that it's a lot

01:14:54.199 --> 01:14:58.159
easier to predict the effect of

01:14:56.400 --> 01:14:59.920
inference time decoding time

01:14:58.159 --> 01:15:01.120
manipulations than it is to predict the

01:14:59.920 --> 01:15:04.239
effect of

01:15:01.120 --> 01:15:07.480
like um fine-tuning or something like

01:15:04.239 --> 01:15:11.040
this like just to give a an

01:15:07.480 --> 01:15:12.480
example beam search with the maximum

01:15:11.040 --> 01:15:15.199
likelihood trained model tends to

01:15:12.480 --> 01:15:16.719
generate things that are shorter um

01:15:15.199 --> 01:15:18.040
whereas greedy decoding tends to

01:15:16.719 --> 01:15:19.639
generate things that are longer and

01:15:18.040 --> 01:15:22.000
repeat more often and stuff like that

01:15:19.639 --> 01:15:25.920
and if you try a few methods like this

01:15:22.000 --> 01:15:28.920
you'll quickly find these kind of qus of

01:15:25.920 --> 01:15:31.320
each of the methods and so by forming a

01:15:28.920 --> 01:15:32.719
good intuition of this you will also

01:15:31.320 --> 01:15:34.000
know how to fix these problems when you

01:15:32.719 --> 01:15:35.600
see them it's like oh my model's

01:15:34.000 --> 01:15:37.320
repeating itself a lot maybe I shouldn't

01:15:35.600 --> 01:15:38.679
be using grey search I should be

01:15:37.320 --> 01:15:41.199
switching over to something else or

01:15:38.679 --> 01:15:43.320
something like that so um this is a good

01:15:41.199 --> 01:15:45.880
thing to know and play around with yeah

01:15:43.320 --> 01:15:47.239
and I think pretty underutilized too um

01:15:45.880 --> 01:15:48.880
a lot of folks will not think about a

01:15:47.239 --> 01:15:50.920
decoding method to fix their problem

01:15:48.880 --> 01:15:52.280
even if like your model might actually

01:15:50.920 --> 01:15:53.760
be perfectly fine under a different

01:15:52.280 --> 01:15:56.000
decoding strategy

01:15:53.760 --> 01:15:58.320
great okay thanks a lot everyone you can

01:15:56.000 --> 01:15:58.320
uh

01:16:02.280 --> 01:16:05.280
finish