CMU Advanced NLP 2024 (12) Reinforcement Learning/transcript.vtt

WEBVTT

00:00:00.840 --> 00:00:05.920
okay so uh let's get started um today

00:00:04.200 --> 00:00:08.000
I'm going to be talking about learning

00:00:05.920 --> 00:00:09.480
from Human feedback I wrote

00:00:08.000 --> 00:00:12.160
reinforcement learning from Human

00:00:09.480 --> 00:00:14.519
feedback because that's what um you know

00:00:12.160 --> 00:00:15.759
a lot of people talk about nowadays but

00:00:14.519 --> 00:00:18.880
actually there's other methods of

00:00:15.759 --> 00:00:21.840
learning from Human feedback so first

00:00:18.880 --> 00:00:24.760
I'm going to be talking about the ways

00:00:21.840 --> 00:00:27.920
we can get uh human feedback for the

00:00:24.760 --> 00:00:31.039
generations of models and mostly focus

00:00:27.920 --> 00:00:32.960
on generation tasks because is um

00:00:31.039 --> 00:00:35.800
generation tasks are harder than like

00:00:32.960 --> 00:00:38.559
classification tasks that we uh we deal

00:00:35.800 --> 00:00:40.000
with normally so I'll spend a fair

00:00:38.559 --> 00:00:42.239
amount of time talking about how we do

00:00:40.000 --> 00:00:45.760
that and then after I talk about how we

00:00:42.239 --> 00:00:48.360
do that we'll move into um how we

00:00:45.760 --> 00:00:51.160
actually learn from that

00:00:48.360 --> 00:00:53.399
signal so normally what we've done up

00:00:51.160 --> 00:00:56.399
until this point is maximum likelihood

00:00:53.399 --> 00:00:58.199
training uh this is just an overview

00:00:56.399 --> 00:00:59.559
slide so we what we want to do is we

00:00:58.199 --> 00:01:00.760
want to maximize the likelihood of

00:00:59.559 --> 00:01:03.280
predicting the next word and the

00:01:00.760 --> 00:01:05.960
reference given the previous words uh

00:01:03.280 --> 00:01:08.119
which gives us the loss of the output

00:01:05.960 --> 00:01:09.799
given the input uh where you know the

00:01:08.119 --> 00:01:13.960
input can be the prompt the output can

00:01:09.799 --> 00:01:16.080
be the answer to uh the output but

00:01:13.960 --> 00:01:18.360
there's uh lots of problems with

00:01:16.080 --> 00:01:20.439
learning from Maximum likelihood and I'm

00:01:18.360 --> 00:01:22.079
going to give three examples here I

00:01:20.439 --> 00:01:24.159
think all of these are actually real

00:01:22.079 --> 00:01:26.880
problems uh that we need to be worried

00:01:24.159 --> 00:01:30.240
about so the first one is that some

00:01:26.880 --> 00:01:32.439
mistakes are worse than others so um in

00:01:30.240 --> 00:01:33.560
the end we want good outputs and some

00:01:32.439 --> 00:01:36.520
mistaken

00:01:33.560 --> 00:01:38.200
predictions uh can be a bigger problem

00:01:36.520 --> 00:01:42.680
for the output being

00:01:38.200 --> 00:01:46.000
good so to give an example uh let's say

00:01:42.680 --> 00:01:47.600
what we actually wanted from like a

00:01:46.000 --> 00:01:49.320
speech recognition system or a

00:01:47.600 --> 00:01:54.040
translation system or something like

00:01:49.320 --> 00:01:54.040
that is uh please send this package to

00:01:54.280 --> 00:01:58.920
Pittsburgh if I write please send a

00:01:56.880 --> 00:02:01.560
package to Pittsburgh then this is not a

00:01:58.920 --> 00:02:03.560
huge problem

00:02:01.560 --> 00:02:06.479
if I write uh please send this package

00:02:03.560 --> 00:02:07.719
to Tokyo then that might be a big

00:02:06.479 --> 00:02:09.640
problem because the package you wanted

00:02:07.719 --> 00:02:12.760
to come to Pittsburgh goes to Tokyo

00:02:09.640 --> 00:02:13.680
instead and uh you might not want that

00:02:12.760 --> 00:02:16.080
to

00:02:13.680 --> 00:02:18.000
happen you might also have it say

00:02:16.080 --> 00:02:20.400
bleeping send this package to Pittsburgh

00:02:18.000 --> 00:02:22.200
instead of pleas um and that would be a

00:02:20.400 --> 00:02:24.200
problem in a customer service system

00:02:22.200 --> 00:02:28.400
right because your customer would uh

00:02:24.200 --> 00:02:28.400
leave and never come back

00:02:28.840 --> 00:02:32.040
so

00:02:30.360 --> 00:02:33.720
determiner like this is not going to

00:02:32.040 --> 00:02:35.640
cause a huge issue U messing up other

00:02:33.720 --> 00:02:37.519
things is going to cause a larger

00:02:35.640 --> 00:02:39.519
issue but from the point of view of

00:02:37.519 --> 00:02:42.680
Maximum likelihood all of these are just

00:02:39.519 --> 00:02:44.560
tokens and messing up one token is the

00:02:42.680 --> 00:02:47.519
same as messing up another token so

00:02:44.560 --> 00:02:50.040
that's uh you know an

00:02:47.519 --> 00:02:52.080
issue another problem is that the gold

00:02:50.040 --> 00:02:54.640
standard and maximum likelihood

00:02:52.080 --> 00:02:57.480
estimation can be bad it can be like not

00:02:54.640 --> 00:02:59.239
what you want and uh corpa are full of

00:02:57.480 --> 00:03:02.400
outputs that we wouldn't want a language

00:02:59.239 --> 00:03:05.400
model producing so for example uh toxic

00:03:02.400 --> 00:03:07.799
comments on Reddit uh

00:03:05.400 --> 00:03:09.959
disinformation um another thing that a

00:03:07.799 --> 00:03:13.000
lot of people don't think about uh quite

00:03:09.959 --> 00:03:15.640
as much is a lot of the data online is

00:03:13.000 --> 00:03:17.680
uh from is automatically generated

00:03:15.640 --> 00:03:19.720
nowadays for example from machine

00:03:17.680 --> 00:03:24.080
translation a lot of the translations

00:03:19.720 --> 00:03:25.720
online are from uh 2016 Google translate

00:03:24.080 --> 00:03:27.560
uh when Google translate was a lot less

00:03:25.720 --> 00:03:29.120
good than it is now and so you have like

00:03:27.560 --> 00:03:31.760
poor quality translations that were

00:03:29.120 --> 00:03:31.760
automatically

00:03:33.040 --> 00:03:37.959
a final problem is uh something that's

00:03:35.280 --> 00:03:40.360
called exposure bias and exposure bias

00:03:37.959 --> 00:03:44.000
basically what it means is mle training

00:03:40.360 --> 00:03:46.000
doesn't consider um the necessarity the

00:03:44.000 --> 00:03:48.599
necessity for generation and it relies

00:03:46.000 --> 00:03:51.360
on gold standard context so if we go

00:03:48.599 --> 00:03:54.159
back to the mle equation when we're

00:03:51.360 --> 00:03:57.200
calculating mle this y less than T is

00:03:54.159 --> 00:03:59.200
always correct it's always a good output

00:03:57.200 --> 00:04:01.439
and so what the model does is it learns

00:03:59.200 --> 00:04:04.280
to over rely on good

00:04:01.439 --> 00:04:06.079
outputs and one example of a problem

00:04:04.280 --> 00:04:08.360
that this causes is models tend to

00:04:06.079 --> 00:04:10.560
repeat themselves over and over again

00:04:08.360 --> 00:04:12.319
for example um when you use some

00:04:10.560 --> 00:04:15.079
generation algorithms and the reason why

00:04:12.319 --> 00:04:18.519
this happens is because in a gold

00:04:15.079 --> 00:04:22.079
standard output if a word has appeared

00:04:18.519 --> 00:04:25.840
previously that word is more likely to

00:04:22.079 --> 00:04:28.560
happen next so like if you say um like I

00:04:25.840 --> 00:04:29.759
am going um I am going to Pittsburgh

00:04:28.560 --> 00:04:31.880
you're much more likely to say

00:04:29.759 --> 00:04:33.000
Pittsburgh again in the future because

00:04:31.880 --> 00:04:35.720
you're talking about Pittsburgh

00:04:33.000 --> 00:04:37.400
topically as coherent so what you get is

00:04:35.720 --> 00:04:38.639
you get mle trained models saying I'm

00:04:37.400 --> 00:04:40.160
going to Pittsburgh I am going to

00:04:38.639 --> 00:04:41.680
Pittsburgh I am going to Pittsburgh I

00:04:40.160 --> 00:04:45.280
going to Pittsburgh you've probably seen

00:04:41.680 --> 00:04:47.320
this before uh at some point and so um

00:04:45.280 --> 00:04:49.320
exposure bias is basically that the

00:04:47.320 --> 00:04:51.039
model has never been exposed to mistakes

00:04:49.320 --> 00:04:55.240
in the past and so it can't deal with

00:04:51.039 --> 00:04:56.840
them so what this does is um if you have

00:04:55.240 --> 00:04:58.560
an alternative training algorithm you

00:04:56.840 --> 00:05:02.120
can fix this by generating a whole bunch

00:04:58.560 --> 00:05:04.880
of outputs uh down like scoring some of

00:05:02.120 --> 00:05:06.880
them poorly and penalizing the model for

00:05:04.880 --> 00:05:09.960
uh generating po outputs and so that can

00:05:06.880 --> 00:05:09.960
fix these problems as

00:05:10.800 --> 00:05:18.440
well uh any questions about this all

00:05:15.199 --> 00:05:20.800
good Okay cool so now I'd like to get

00:05:18.440 --> 00:05:23.919
into how we measure how good an output

00:05:20.800 --> 00:05:26.360
is and there's different ways of doing

00:05:23.919 --> 00:05:30.319
this um the first one is objective

00:05:26.360 --> 00:05:32.680
assessment so for some uh tasks or for

00:05:30.319 --> 00:05:35.400
many tasks there's kind of objectively a

00:05:32.680 --> 00:05:37.280
correct answer there's also human

00:05:35.400 --> 00:05:40.360
subjective annotations so you can ask

00:05:37.280 --> 00:05:42.919
humans to do annotation for you there's

00:05:40.360 --> 00:05:45.400
machine prediction of human

00:05:42.919 --> 00:05:48.319
preferences and there's also use in

00:05:45.400 --> 00:05:50.840
another system in a downstream

00:05:48.319 --> 00:05:52.960
task so the way objective assessment

00:05:50.840 --> 00:05:54.919
works is you have an annotated correct

00:05:52.960 --> 00:05:57.080
answer in match against this so like if

00:05:54.919 --> 00:06:00.600
you're solving math problems uh

00:05:57.080 --> 00:06:02.560
answering objective questions and and

00:06:00.600 --> 00:06:04.280
you know you can pick any arbitrary

00:06:02.560 --> 00:06:06.840
example you can pick your classification

00:06:04.280 --> 00:06:09.800
example from uh like your text

00:06:06.840 --> 00:06:11.880
classification tasks an even clearer

00:06:09.800 --> 00:06:13.880
example is if you have math problems

00:06:11.880 --> 00:06:15.639
there's kind of objectively one answer

00:06:13.880 --> 00:06:18.080
to any math problem and there's no other

00:06:15.639 --> 00:06:19.680
answer that could be correct so this

00:06:18.080 --> 00:06:21.160
makes your life easy if you're handling

00:06:19.680 --> 00:06:22.560
this type of problem but of course

00:06:21.160 --> 00:06:24.120
there's many other types of problems we

00:06:22.560 --> 00:06:26.039
want to handle that don't have objective

00:06:24.120 --> 00:06:29.039
answers like

00:06:26.039 --> 00:06:31.440
this so let's say we're handling a gener

00:06:29.039 --> 00:06:34.680
a generation task where we don't have an

00:06:31.440 --> 00:06:36.360
objective answer um in this Cas kind of

00:06:34.680 --> 00:06:39.440
one of our gold standards is human

00:06:36.360 --> 00:06:42.360
evaluation so we might have a source

00:06:39.440 --> 00:06:44.919
input like a prompt or an input text for

00:06:42.360 --> 00:06:47.240
machine translation we have one or

00:06:44.919 --> 00:06:49.960
several hypotheses and we ask a human

00:06:47.240 --> 00:06:53.280
annotator to basically give uh a score

00:06:49.960 --> 00:06:55.759
for them or do some sort of other

00:06:53.280 --> 00:06:59.759
annotation and the different varieties

00:06:55.759 --> 00:07:03.080
of annotation that we can give are um

00:06:59.759 --> 00:07:04.599
something called direct assessment so uh

00:07:03.080 --> 00:07:06.599
direct assessment is a term that comes

00:07:04.599 --> 00:07:09.280
from machine translation uh so you might

00:07:06.599 --> 00:07:11.039
not see it used uh lots of other places

00:07:09.280 --> 00:07:13.120
but it's basically just give a score

00:07:11.039 --> 00:07:15.759
directly to how good the output is so

00:07:13.120 --> 00:07:17.199
you can say like if you say please send

00:07:15.759 --> 00:07:18.960
this translation is please send this

00:07:17.199 --> 00:07:21.759
package to Tokyo we give it a score of

00:07:18.960 --> 00:07:24.360
two out of 10 or something like

00:07:21.759 --> 00:07:28.000
this

00:07:24.360 --> 00:07:30.840
so the the question here is like what

00:07:28.000 --> 00:07:32.400
does like let's say I gave a score of

00:07:30.840 --> 00:07:34.520
two out of 10 for please send this

00:07:32.400 --> 00:07:37.680
package to Tokyo what score should I

00:07:34.520 --> 00:07:40.240
give for please send a package to Tokyo

00:07:37.680 --> 00:07:42.360
anyone have any ideas the the correct

00:07:40.240 --> 00:07:46.520
answer is please send this package to

00:07:42.360 --> 00:07:48.000
take out of eight out of 10 yeah but you

00:07:46.520 --> 00:07:50.440
might disagree on that right it's kind

00:07:48.000 --> 00:07:52.159
of like subjective um one of the

00:07:50.440 --> 00:07:54.039
difficulties of direct assessment is

00:07:52.159 --> 00:07:55.520
giving a number like this is pretty

00:07:54.039 --> 00:07:57.800
difficult if you don't have a very clear

00:07:55.520 --> 00:07:59.720
rubric and very skilled annotators and

00:07:57.800 --> 00:08:02.879
it's hard to get consistency between

00:07:59.720 --> 00:08:04.400
people when you do this so the advantage

00:08:02.879 --> 00:08:05.599
is it kind of gives you an idea of how

00:08:04.400 --> 00:08:07.520
good things are overall but the

00:08:05.599 --> 00:08:09.280
disadvantage is it's more difficult to

00:08:07.520 --> 00:08:11.319
annotate and get

00:08:09.280 --> 00:08:13.159
consistency um another thing that I

00:08:11.319 --> 00:08:15.319
should point out is often scores are

00:08:13.159 --> 00:08:18.680
assigned separately based on desirable

00:08:15.319 --> 00:08:20.960
traits so um we don't necessarily just

00:08:18.680 --> 00:08:23.479
say how good is it we say how fluent is

00:08:20.960 --> 00:08:26.120
it like is it fluent uh

00:08:23.479 --> 00:08:28.159
English in Translation there's a concept

00:08:26.120 --> 00:08:30.720
called adequacy which is how well does

00:08:28.159 --> 00:08:34.599
the output reflect the input

00:08:30.720 --> 00:08:36.519
semantics um and if you're assessing

00:08:34.599 --> 00:08:38.440
translation systems actually it's common

00:08:36.519 --> 00:08:40.519
to assess fluency without even looking

00:08:38.440 --> 00:08:43.200
at the input because then you can just

00:08:40.519 --> 00:08:44.880
say how fluent is it but for adequacy

00:08:43.200 --> 00:08:46.320
you definitely need to understand the

00:08:44.880 --> 00:08:49.600
input so you need to be a bilingual

00:08:46.320 --> 00:08:54.680
speaker to be able to assess

00:08:49.600 --> 00:08:57.560
that um factuality um and so factuality

00:08:54.680 --> 00:09:00.160
is tricky um it can either be factuality

00:08:57.560 --> 00:09:03.880
grounded in a particular input text in

00:09:00.160 --> 00:09:05.600
which case um the facts would have to be

00:09:03.880 --> 00:09:07.680
you know things that were said in the

00:09:05.600 --> 00:09:09.399
input or it can be just kind of is the

00:09:07.680 --> 00:09:11.120
statement factual in general in which

00:09:09.399 --> 00:09:13.720
case you need to go online you need to

00:09:11.120 --> 00:09:16.480
search for things and like uh check

00:09:13.720 --> 00:09:18.480
whether the statement is factual or not

00:09:16.480 --> 00:09:20.480
um other things are like coherence does

00:09:18.480 --> 00:09:21.480
the output fit coherently within the

00:09:20.480 --> 00:09:23.680
larger

00:09:21.480 --> 00:09:25.680
discs um and there's many many other

00:09:23.680 --> 00:09:28.120
ones of these this is also task

00:09:25.680 --> 00:09:29.760
dependent so like the things you will

00:09:28.120 --> 00:09:31.000
evaluate for machine transl are

00:09:29.760 --> 00:09:32.880
different than the ones you would do for

00:09:31.000 --> 00:09:35.760
dialog which are different than the ones

00:09:32.880 --> 00:09:38.200
you would do for a general purpose

00:09:35.760 --> 00:09:41.279
chatot uh which is different kind things

00:09:38.200 --> 00:09:44.120
you would do for um summarization for

00:09:41.279 --> 00:09:46.320
example so if you're interested in doing

00:09:44.120 --> 00:09:47.519
something like this uh then I definitely

00:09:46.320 --> 00:09:48.800
encourage you to look at what other

00:09:47.519 --> 00:09:51.399
people have done for the tasks you're

00:09:48.800 --> 00:09:53.079
interested in uh previously and uh find

00:09:51.399 --> 00:09:54.880
out the different types of traits that

00:09:53.079 --> 00:09:58.320
did

00:09:54.880 --> 00:10:00.760
last uh any any questions about this

00:09:58.320 --> 00:10:03.079
also

00:10:00.760 --> 00:10:06.920
okay the next type of feedback is

00:10:03.079 --> 00:10:09.839
preference ratings um and so this is uh

00:10:06.920 --> 00:10:12.600
basically what you do is you have two or

00:10:09.839 --> 00:10:14.240
more outputs from different models or

00:10:12.600 --> 00:10:16.440
different Generations from an individual

00:10:14.240 --> 00:10:18.839
model and you ask a human which one is

00:10:16.440 --> 00:10:22.320
better like is one better than the other

00:10:18.839 --> 00:10:23.839
or are they tied and so in this case um

00:10:22.320 --> 00:10:26.320
you might have please send this package

00:10:23.839 --> 00:10:28.880
to Tokyo please send a package to

00:10:26.320 --> 00:10:31.040
Tokyo we might disagree on how like good

00:10:28.880 --> 00:10:33.959
or bad each of them are but I think most

00:10:31.040 --> 00:10:35.959
people would agree that this one is like

00:10:33.959 --> 00:10:37.480
despite the fact that it got this wrong

00:10:35.959 --> 00:10:40.160
the second one is better than the first

00:10:37.480 --> 00:10:42.240
one so this is a little bit of an easier

00:10:40.160 --> 00:10:45.040
task it's easier to uh get people to

00:10:42.240 --> 00:10:46.839
annotate these things

00:10:45.040 --> 00:10:50.519
consistently however it has the

00:10:46.839 --> 00:10:52.839
disadvantage that you can't really tell

00:10:50.519 --> 00:10:55.360
uh whether systems are really good or

00:10:52.839 --> 00:10:57.200
really bad so let's say you have a bunch

00:10:55.360 --> 00:11:00.279
of really bad systems that you're

00:10:57.200 --> 00:11:01.839
comparing with each other um you might

00:11:00.279 --> 00:11:03.680
find that one is better than the other

00:11:01.839 --> 00:11:06.000
but that still doesn't mean it's ready

00:11:03.680 --> 00:11:07.399
to be deployed or if you have a bunch of

00:11:06.000 --> 00:11:11.040
really good systems they're all

00:11:07.399 --> 00:11:13.000
basically you know very very similar to

00:11:11.040 --> 00:11:14.399
another but one is like slightly more

00:11:13.000 --> 00:11:18.639
fluent than the other you might still

00:11:14.399 --> 00:11:20.680
get a similar result um and so that also

00:11:18.639 --> 00:11:22.760
makes it uh you know a little bit

00:11:20.680 --> 00:11:24.880
difficult to use practically in some

00:11:22.760 --> 00:11:27.040
ways I didn't put it on the slide but

00:11:24.880 --> 00:11:30.680
there's another way you can kind of get

00:11:27.040 --> 00:11:33.920
the best of both worlds um which is a

00:11:30.680 --> 00:11:35.560
side by side assessment and side by-side

00:11:33.920 --> 00:11:38.440
assessment basically what you would do

00:11:35.560 --> 00:11:40.560
is you would say um please send this

00:11:38.440 --> 00:11:43.399
package to Tokyo please send a package

00:11:40.560 --> 00:11:47.279
to Pittsburgh give each of them a direct

00:11:43.399 --> 00:11:48.839
score um but you can use decimal places

00:11:47.279 --> 00:11:51.120
and you can't use the same score for all

00:11:48.839 --> 00:11:55.920
of them and so it's

00:11:51.120 --> 00:11:57.480
like five 500 and 4.99 out of five or

00:11:55.920 --> 00:11:59.519
something like that like you like one

00:11:57.480 --> 00:12:02.639
slightly better than the other or or

00:11:59.519 --> 00:12:04.480
something like that um so there are ways

00:12:02.639 --> 00:12:07.240
to kind of get Best of Both Worlds if

00:12:04.480 --> 00:12:11.720
you're interested in doing

00:12:07.240 --> 00:12:11.720
that um

00:12:14.920 --> 00:12:20.519
so one problem one other problem with

00:12:18.279 --> 00:12:22.519
preference rankings is that there's a

00:12:20.519 --> 00:12:24.440
limited number of things that humans can

00:12:22.519 --> 00:12:28.160
compare before they get really

00:12:24.440 --> 00:12:32.360
overwhelmed so if you say I

00:12:28.160 --> 00:12:35.560
want like I want to

00:12:32.360 --> 00:12:36.920
rate 15 systems or 20 systems with

00:12:35.560 --> 00:12:39.120
respect to how good they are with

00:12:36.920 --> 00:12:40.639
respect to each other it's going to be

00:12:39.120 --> 00:12:43.680
impossible for humans to come up with a

00:12:40.639 --> 00:12:46.959
good preference ranking between them and

00:12:43.680 --> 00:12:49.480
so the typical way around this um which

00:12:46.959 --> 00:12:52.360
is also used in uh things like the

00:12:49.480 --> 00:12:55.440
chatbot Arena by lmis and other things

00:12:52.360 --> 00:12:58.720
like this is to use uh something like an

00:12:55.440 --> 00:13:00.959
ELO or true skill ranking and what these

00:12:58.720 --> 00:13:03.079
are is these are things that were

00:13:00.959 --> 00:13:05.760
created for the ranking of like chess

00:13:03.079 --> 00:13:09.160
players or video game players or other

00:13:05.760 --> 00:13:11.720
things where they like b battle against

00:13:09.160 --> 00:13:13.920
each other in multiple matches uh

00:13:11.720 --> 00:13:16.440
pair-wise and then you put all of the

00:13:13.920 --> 00:13:18.399
wins and losses into these ranking

00:13:16.440 --> 00:13:20.600
algorithms and they give you a score

00:13:18.399 --> 00:13:22.920
about how good like each of the each of

00:13:20.600 --> 00:13:27.079
the players are so if you do something

00:13:22.920 --> 00:13:29.480
like this you can um get basically a

00:13:27.079 --> 00:13:32.120
ranking of systems despite the that you

00:13:29.480 --> 00:13:35.240
only did pairwise assessments so these

00:13:32.120 --> 00:13:35.240
are also a good thing to know

00:13:37.399 --> 00:13:43.839
about a final variety of human feedback

00:13:40.600 --> 00:13:45.320
uh that we create is uh air annotation

00:13:43.839 --> 00:13:47.519
and this can be useful for a number of

00:13:45.320 --> 00:13:49.839
reasons um but basically the way it

00:13:47.519 --> 00:13:53.839
works is you annotate individual errors

00:13:49.839 --> 00:13:55.639
within the outputs and um oh one thing I

00:13:53.839 --> 00:13:58.120
should mention is that um I'm giving a

00:13:55.639 --> 00:14:00.880
lot of examples from machine translation

00:13:58.120 --> 00:14:02.800
um I feel like machine translation has

00:14:00.880 --> 00:14:04.519
been doing evaluation of generated

00:14:02.800 --> 00:14:07.600
outputs for a lot longer than a lot of

00:14:04.519 --> 00:14:09.000
other uh fields of NLP have and

00:14:07.600 --> 00:14:11.800
therefore their methodology is more

00:14:09.000 --> 00:14:13.480
developed than a lot of other fields um

00:14:11.800 --> 00:14:16.199
but a lot of these things can also be

00:14:13.480 --> 00:14:18.079
applied to uh other uh other tasks as

00:14:16.199 --> 00:14:19.079
well but anyway getting back to this

00:14:18.079 --> 00:14:20.680
there's something for machine

00:14:19.079 --> 00:14:23.639
translation called multi-dimensional

00:14:20.680 --> 00:14:26.240
quality metrics and the multidimensional

00:14:23.639 --> 00:14:29.160
quality metrics basically what they do

00:14:26.240 --> 00:14:32.199
is they annotate spans in the output

00:14:29.160 --> 00:14:34.800
where each Span in the output is given a

00:14:32.199 --> 00:14:38.079
severity ranking of the error and it's

00:14:34.800 --> 00:14:40.199
given a type of the error and there's

00:14:38.079 --> 00:14:42.600
about eight different types of Errors

00:14:40.199 --> 00:14:44.839
like this doesn't violate or this

00:14:42.600 --> 00:14:47.399
violates linguistic conventions of using

00:14:44.839 --> 00:14:49.880
the word this instead of uh here by

00:14:47.399 --> 00:14:51.639
using the word uh instead of this here

00:14:49.880 --> 00:14:55.079
and then this is an accuracy error

00:14:51.639 --> 00:14:57.839
because it's not accurately con uh uh

00:14:55.079 --> 00:15:01.720
conveying the output and then this error

00:14:57.839 --> 00:15:04.600
is minor uh this error is Major um and

00:15:01.720 --> 00:15:06.399
then there's also like severe severe

00:15:04.600 --> 00:15:07.440
versus major but minor and major is a

00:15:06.399 --> 00:15:09.680
more important

00:15:07.440 --> 00:15:11.839
distinction um so the advantage of this

00:15:09.680 --> 00:15:14.279
is a couple fold number one it gives you

00:15:11.839 --> 00:15:16.440
more fine grained feedback uh in that

00:15:14.279 --> 00:15:19.199
you can say okay this system has a lot

00:15:16.440 --> 00:15:22.199
of uh accuracy errors this system has a

00:15:19.199 --> 00:15:24.880
lot of linguistic conventions errors um

00:15:22.199 --> 00:15:28.600
it also can be more consistent because

00:15:24.880 --> 00:15:29.839
if you just say to people which output

00:15:28.600 --> 00:15:31.800
is better

00:15:29.839 --> 00:15:34.560
or what is the score of this output

00:15:31.800 --> 00:15:36.360
people have trouble deciding about that

00:15:34.560 --> 00:15:39.560
because it's a more subjective

00:15:36.360 --> 00:15:41.680
evaluation but if I say is this word

00:15:39.560 --> 00:15:43.000
correct it's a little bit easier for

00:15:41.680 --> 00:15:44.759
people to do so you can get more

00:15:43.000 --> 00:15:46.920
consistent annotations

00:15:44.759 --> 00:15:49.720
here the problem with this is this can

00:15:46.920 --> 00:15:50.839
be very time consuming so um you know

00:15:49.720 --> 00:15:52.480
obviously you need to go through and

00:15:50.839 --> 00:15:56.440
annotate every single error if it's for

00:15:52.480 --> 00:15:56.440
a long outputs or something your

00:15:56.959 --> 00:16:03.519
problem so anyway these are just three

00:15:59.800 --> 00:16:05.680
uh ways of collecting human feedback um

00:16:03.519 --> 00:16:08.639
and then there's an alternative which is

00:16:05.680 --> 00:16:10.079
automatic evaluation of outputs and um

00:16:08.639 --> 00:16:14.399
there's a bunch of different ways we can

00:16:10.079 --> 00:16:16.800
do this the basic idea here is we have a

00:16:14.399 --> 00:16:20.199
source um we have a couple

00:16:16.800 --> 00:16:22.800
hypotheses and uh we have an automatic

00:16:20.199 --> 00:16:26.000
system that generates outputs uh like

00:16:22.800 --> 00:16:28.279
scores and we optionally have a

00:16:26.000 --> 00:16:30.839
reference output so the reference output

00:16:28.279 --> 00:16:33.519
is a human created gold standard output

00:16:30.839 --> 00:16:35.120
with respect to how good that um uh with

00:16:33.519 --> 00:16:38.240
respect to like what the output should

00:16:35.120 --> 00:16:38.240
be in an ideal

00:16:38.279 --> 00:16:47.079
case and basically the goal of automatic

00:16:43.199 --> 00:16:50.199
evaluation is to

00:16:47.079 --> 00:16:52.839
predict human preferences or to predict

00:16:50.199 --> 00:16:56.240
what the human scores would be um

00:16:52.839 --> 00:16:58.600
because still at this point um we mostly

00:16:56.240 --> 00:16:59.480
view what humans think of the output to

00:16:58.600 --> 00:17:01.680
be

00:16:59.480 --> 00:17:03.280
uh kind of the

00:17:01.680 --> 00:17:06.199
standard

00:17:03.280 --> 00:17:08.439
and this is called a variety of things

00:17:06.199 --> 00:17:10.600
depending on what field you're in um in

00:17:08.439 --> 00:17:12.559
machine translation and summarization

00:17:10.600 --> 00:17:13.520
it's called automatic evaluation also a

00:17:12.559 --> 00:17:16.520
lot in

00:17:13.520 --> 00:17:18.400
dialogue um if you're talking about

00:17:16.520 --> 00:17:21.000
people from reinforcement learning or

00:17:18.400 --> 00:17:24.600
other things um or chat Bots or things

00:17:21.000 --> 00:17:28.240
like that uh a lot of people or uh like

00:17:24.600 --> 00:17:31.280
AGI or whatever um a lot of people call

00:17:28.240 --> 00:17:32.520
it uh word model um because that

00:17:31.280 --> 00:17:34.480
specifically comes from the point of

00:17:32.520 --> 00:17:36.440
view of like learning from this feedback

00:17:34.480 --> 00:17:37.960
but essentially they're the same thing

00:17:36.440 --> 00:17:41.080
uh from my point of view they're trying

00:17:37.960 --> 00:17:42.520
to predict how good an output is and how

00:17:41.080 --> 00:17:44.240
much you should reward the model for

00:17:42.520 --> 00:17:46.559
producing that

00:17:44.240 --> 00:17:48.679
output

00:17:46.559 --> 00:17:50.520
um so there's a bunch of different

00:17:48.679 --> 00:17:51.720
methods to do this I'm not going to

00:17:50.520 --> 00:17:53.799
cover all of them I'm just going to

00:17:51.720 --> 00:17:55.240
cover three paradigms for doing this so

00:17:53.799 --> 00:17:57.880
you know where to look further if you're

00:17:55.240 --> 00:18:00.039
interested in doing these things um the

00:17:57.880 --> 00:18:02.400
first one is embedding based

00:18:00.039 --> 00:18:04.679
evaluation and the way embedding based

00:18:02.400 --> 00:18:06.600
evaluation works is usually it's

00:18:04.679 --> 00:18:11.400
unsupervised calculation based on

00:18:06.600 --> 00:18:14.880
embeding similarity between um

00:18:11.400 --> 00:18:18.080
the output that the model generated and

00:18:14.880 --> 00:18:20.840
a reference output that uh you have

00:18:18.080 --> 00:18:23.400
created so sorry this is very small but

00:18:20.840 --> 00:18:25.559
we have a reference here that says the

00:18:23.400 --> 00:18:27.640
weather is cold today and we have a

00:18:25.559 --> 00:18:30.240
candidate that says it is freezing today

00:18:27.640 --> 00:18:33.000
so this is probably you know like a good

00:18:30.240 --> 00:18:35.480
um a reasonably good

00:18:33.000 --> 00:18:37.640
output and we run this through some

00:18:35.480 --> 00:18:39.120
embedding model uh it was called Bert

00:18:37.640 --> 00:18:40.679
score and so of course you can run it

00:18:39.120 --> 00:18:42.240
through Bert but basically it can be any

00:18:40.679 --> 00:18:43.799
embedding model that gives you embedding

00:18:42.240 --> 00:18:46.200
for each token in the

00:18:43.799 --> 00:18:47.640
sequence and so there are five tokens in

00:18:46.200 --> 00:18:49.720
this sequence four tokens in this

00:18:47.640 --> 00:18:51.960
sequence you get five tokens and then

00:18:49.720 --> 00:18:54.799
four sorry five embeddings and then four

00:18:51.960 --> 00:18:57.400
embeddings you calculate carewise cosine

00:18:54.799 --> 00:18:59.880
similarity between all of them and this

00:18:57.400 --> 00:19:03.480
gives you cosine

00:18:59.880 --> 00:19:06.480
similarity Matrix and then you take the

00:19:03.480 --> 00:19:09.120
ARG Max or you take the maximum

00:19:06.480 --> 00:19:11.280
similarity along either the

00:19:09.120 --> 00:19:15.799
rows or the

00:19:11.280 --> 00:19:19.559
columns and here the rows correspond

00:19:15.799 --> 00:19:22.400
to tokens in the reference and because

00:19:19.559 --> 00:19:24.039
the rows correspond to tokens in the

00:19:22.400 --> 00:19:26.960
reference

00:19:24.039 --> 00:19:28.320
the how well you find something that is

00:19:26.960 --> 00:19:31.679
similar to each of the tokens in the

00:19:28.320 --> 00:19:34.000
reference is like a recall based method

00:19:31.679 --> 00:19:35.919
because it's saying how many tokens in

00:19:34.000 --> 00:19:39.520
the reference have a good match in the

00:19:35.919 --> 00:19:41.120
output and then if you look at the

00:19:39.520 --> 00:19:42.799
columns if you look at the max and the

00:19:41.120 --> 00:19:44.960
columns this is like a precision based

00:19:42.799 --> 00:19:47.000
metric because it's saying how many of

00:19:44.960 --> 00:19:49.360
the things in the output are similar

00:19:47.000 --> 00:19:51.240
have a similar match in the reference so

00:19:49.360 --> 00:19:54.480
basically you can calculate recall and

00:19:51.240 --> 00:19:56.200
precision over all of the tokens and

00:19:54.480 --> 00:20:00.200
then feed this into something that looks

00:19:56.200 --> 00:20:02.400
like fmeasure and you can also use tfidf

00:20:00.200 --> 00:20:06.000
waiting um like what I talked about in

00:20:02.400 --> 00:20:07.799
the rag lecture uh to upweight low

00:20:06.000 --> 00:20:09.520
frequency words because low frequency

00:20:07.799 --> 00:20:11.440
words tend to be more content words and

00:20:09.520 --> 00:20:13.120
going back to my example you know if you

00:20:11.440 --> 00:20:14.280
make a mistake from Pittsburgh to Tokyo

00:20:13.120 --> 00:20:17.880
that's going to be more painful than

00:20:14.280 --> 00:20:21.000
making a mistake from this to um so

00:20:17.880 --> 00:20:22.520
actually if you'll uh if you were paying

00:20:21.000 --> 00:20:25.480
close attention to the rag lecture this

00:20:22.520 --> 00:20:27.360
looks really similar to the co bear um

00:20:25.480 --> 00:20:29.559
the co bear retrieval objective that I

00:20:27.360 --> 00:20:30.960
talked about in the r lecture um I don't

00:20:29.559 --> 00:20:32.840
think it's a coincidence they both came

00:20:30.960 --> 00:20:34.360
out around the same time uh so people

00:20:32.840 --> 00:20:36.360
were thinking about the same thing but

00:20:34.360 --> 00:20:37.600
um this is one method that's pretty

00:20:36.360 --> 00:20:40.200
widely

00:20:37.600 --> 00:20:43.480
use the bird Square code base is also

00:20:40.200 --> 00:20:45.440
really nice and easy to use so um if uh

00:20:43.480 --> 00:20:47.640
you want to try it out feel free to take

00:20:45.440 --> 00:20:47.640
a

00:20:48.159 --> 00:20:53.840
look cool um the next one I'd like to

00:20:51.600 --> 00:20:56.080
talk about is a regression based

00:20:53.840 --> 00:20:58.760
evaluation and the way this works is

00:20:56.080 --> 00:21:02.600
this is usually used in a supervised uh

00:20:58.760 --> 00:21:04.320
setting so uh the way what you have to

00:21:02.600 --> 00:21:07.600
do is you have to calculate a whole

00:21:04.320 --> 00:21:09.799
bunch of like actual human

00:21:07.600 --> 00:21:12.440
judgments and

00:21:09.799 --> 00:21:15.000
usually these judgments can either be

00:21:12.440 --> 00:21:16.960
direct assessment uh where you actually

00:21:15.000 --> 00:21:19.120
have a score or they can be pairwise

00:21:16.960 --> 00:21:20.840
judgments and then if you have direct

00:21:19.120 --> 00:21:23.640
assessment you use a regression based

00:21:20.840 --> 00:21:26.039
loss like uh minimum squared error if

00:21:23.640 --> 00:21:27.520
you have pairwise uh you use a ranking

00:21:26.039 --> 00:21:29.039
based loss that tries to upweight the

00:21:27.520 --> 00:21:31.360
ones that are higher scoring downward

00:21:29.039 --> 00:21:33.200
the ones that are lower scoring one

00:21:31.360 --> 00:21:35.720
typical example of this is Comet which

00:21:33.200 --> 00:21:37.200
is or has been at least for a very long

00:21:35.720 --> 00:21:39.880
time the state-of-the art and machine

00:21:37.200 --> 00:21:41.279
translation evaluation and the reason

00:21:39.880 --> 00:21:43.440
why it works so well is because we have

00:21:41.279 --> 00:21:44.720
a bunch of evaluations for machine

00:21:43.440 --> 00:21:46.080
translation they've been doing

00:21:44.720 --> 00:21:47.600
evaluation and machine translation

00:21:46.080 --> 00:21:50.480
systems for years and you can use that

00:21:47.600 --> 00:21:52.720
as lots of supervised training data so

00:21:50.480 --> 00:21:54.640
basically you just take um these

00:21:52.720 --> 00:21:56.440
evaluation data you have human

00:21:54.640 --> 00:21:59.080
annotations you have the output

00:21:56.440 --> 00:22:00.320
according to a model like comet um you

00:21:59.080 --> 00:22:02.679
calculate the difference between them

00:22:00.320 --> 00:22:05.640
and you update model

00:22:02.679 --> 00:22:07.080
parameters um the problem this is great

00:22:05.640 --> 00:22:08.520
if you have lots of training data the

00:22:07.080 --> 00:22:10.640
problem with this is for a lot of tasks

00:22:08.520 --> 00:22:12.360
we don't have lots of training data so

00:22:10.640 --> 00:22:14.720
um you know training these is a little

00:22:12.360 --> 00:22:14.720
bit less

00:22:15.400 --> 00:22:22.919
feasible and now recently uh what we

00:22:19.600 --> 00:22:25.279
have been moving into is is a QA based

00:22:22.919 --> 00:22:27.120
evaluation which is basically where we

00:22:25.279 --> 00:22:30.760
ask a language model how good the output

00:22:27.120 --> 00:22:32.279
is and so uh gmba is an example one of

00:22:30.760 --> 00:22:34.559
the early examples of this for machine

00:22:32.279 --> 00:22:37.320
translation evaluation uh where they

00:22:34.559 --> 00:22:39.840
basically just ask a g gp4 like score

00:22:37.320 --> 00:22:41.600
the following translation from Source

00:22:39.840 --> 00:22:44.000
language to target language with respect

00:22:41.600 --> 00:22:47.080
to the human reference um on a

00:22:44.000 --> 00:22:49.200
continuous scale from Z to 100 uh where

00:22:47.080 --> 00:22:51.320
the score of zero means no meaning

00:22:49.200 --> 00:22:54.039
preserved and the score of 100 means a

00:22:51.320 --> 00:22:56.880
perfect meaning in grammar uh you feed

00:22:54.039 --> 00:22:58.760
in the source um you feed in the T the

00:22:56.880 --> 00:23:01.000
human reference optionally if you have a

00:22:58.760 --> 00:23:03.320
human reference and then you feed in the

00:23:01.000 --> 00:23:06.760
Target um and you get a

00:23:03.320 --> 00:23:09.919
score and um so this this works pretty

00:23:06.760 --> 00:23:12.720
well this can give you uh better results

00:23:09.919 --> 00:23:15.159
um there's a especially if you have a

00:23:12.720 --> 00:23:16.960
strong language model the problem is

00:23:15.159 --> 00:23:18.279
it's very unpredictable whether this is

00:23:16.960 --> 00:23:20.120
going to work well and it's very

00:23:18.279 --> 00:23:23.039
dependent on the prompt that you're

00:23:20.120 --> 00:23:25.279
using so um right now A lot of people

00:23:23.039 --> 00:23:27.279
are using gp4 without actually

00:23:25.279 --> 00:23:29.039
validating whether it does a good job at

00:23:27.279 --> 00:23:33.080
evaluation and

00:23:29.039 --> 00:23:34.919
and my the results are all across the

00:23:33.080 --> 00:23:36.880
board it can be anywhere from very very

00:23:34.919 --> 00:23:38.640
good to very very bad at evaluating

00:23:36.880 --> 00:23:41.320
particular tasks so I would be at least

00:23:38.640 --> 00:23:43.559
a little bit suspicious of whether gp4

00:23:41.320 --> 00:23:45.679
is doing a good job evaluating for your

00:23:43.559 --> 00:23:49.320
task especially more complex

00:23:45.679 --> 00:23:51.960
tests um I would especially be

00:23:49.320 --> 00:23:54.000
suspicious if you're doing two uh any of

00:23:51.960 --> 00:23:56.760
the two following things number one if

00:23:54.000 --> 00:23:59.880
you're comparing gp4 or any model

00:23:56.760 --> 00:24:02.400
against itself in another model because

00:23:59.880 --> 00:24:05.200
gp4 really likes

00:24:02.400 --> 00:24:06.880
gp4 it really likes its own outputs and

00:24:05.200 --> 00:24:08.120
there are papers uh sorry I don't

00:24:06.880 --> 00:24:09.679
actually have the references here but I

00:24:08.120 --> 00:24:11.200
can follow up if people are interested

00:24:09.679 --> 00:24:13.080
but there are papers that demonstrate

00:24:11.200 --> 00:24:15.799
that gp4 likes it you know its own

00:24:13.080 --> 00:24:19.200
outputs more than others also if you're

00:24:15.799 --> 00:24:22.120
explicitly optimizing the outputs using

00:24:19.200 --> 00:24:24.640
rlf um there is something called good

00:24:22.120 --> 00:24:27.120
Hearts law which is basically anytime

00:24:24.640 --> 00:24:29.520
you uh start optimizing towards a metric

00:24:27.120 --> 00:24:32.559
it becomes a bad metric and that also

00:24:29.520 --> 00:24:35.000
happens for gp4 based evaluations so if

00:24:32.559 --> 00:24:37.200
you start optimizing for gp4 based

00:24:35.000 --> 00:24:38.960
evaluations especially for reference

00:24:37.200 --> 00:24:41.679
list metrics that don't use a reference

00:24:38.960 --> 00:24:44.840
output then um you start basically

00:24:41.679 --> 00:24:47.440
exploiting the metric

00:24:44.840 --> 00:24:49.840
um another thing that you can do with QA

00:24:47.440 --> 00:24:53.279
based evaluation is ask about fine grade

00:24:49.840 --> 00:24:54.919
mistakes and so this is a paper by um uh

00:24:53.279 --> 00:24:56.480
Patrick Fernandez who's a student who's

00:24:54.919 --> 00:25:02.080
working with me and basically what we

00:24:56.480 --> 00:25:05.240
did is we asked the model to um not give

00:25:02.080 --> 00:25:07.360
a particular score but actually identify

00:25:05.240 --> 00:25:08.880
the mistakes in the output and when we

00:25:07.360 --> 00:25:10.559
asked it to identify the mistakes in the

00:25:08.880 --> 00:25:13.720
output we found that this gave more

00:25:10.559 --> 00:25:17.320
consistent uh results so kind of

00:25:13.720 --> 00:25:18.840
interestingly we ask humans to identify

00:25:17.320 --> 00:25:21.120
individual mistakes and the output that

00:25:18.840 --> 00:25:24.240
gives humans more consistent results

00:25:21.120 --> 00:25:25.559
it's the same thing for gp4 so um that

00:25:24.240 --> 00:25:27.320
that's another paper you can look at if

00:25:25.559 --> 00:25:29.640
you're

00:25:27.320 --> 00:25:32.679
interested

00:25:29.640 --> 00:25:38.000
cool um so I I mentioned that you could

00:25:32.679 --> 00:25:38.000
or could not uh trust uh yeah sorry go

00:25:44.679 --> 00:25:51.279
ahead uh correct so yeah B basically

00:25:47.360 --> 00:25:53.279
just what you do is you have the source

00:25:51.279 --> 00:25:54.960
um ideally you'll also have a reference

00:25:53.279 --> 00:25:57.840
output that was created by skilled

00:25:54.960 --> 00:25:59.720
humans and then you put in the Target

00:25:57.840 --> 00:26:02.279
you know output basically you have the

00:25:59.720 --> 00:26:08.000
input ideally a reference output created

00:26:02.279 --> 00:26:08.000
by Good by skilled humans and uh like

00:26:15.159 --> 00:26:20.240
hypothesis yeah I

00:26:17.919 --> 00:26:24.559
mean it's a good question and I don't

00:26:20.240 --> 00:26:26.919
know if we actually have a a very clear

00:26:24.559 --> 00:26:31.399
empirical like evidence of why this is

00:26:26.919 --> 00:26:33.320
the case but my hypothesis about this is

00:26:31.399 --> 00:26:36.159
yes we kind of would expect models to be

00:26:33.320 --> 00:26:38.200
more biased towards their own outputs

00:26:36.159 --> 00:26:40.919
and the reason why is because

00:26:38.200 --> 00:26:43.080
essentially you know models

00:26:40.919 --> 00:26:44.279
are within their embeddings they're

00:26:43.080 --> 00:26:45.760
encoding when they're in a high

00:26:44.279 --> 00:26:47.600
probability part of the space and when

00:26:45.760 --> 00:26:50.200
they're in a low probability part of the

00:26:47.600 --> 00:26:51.120
space and like the high probability part

00:26:50.200 --> 00:26:54.600
of the

00:26:51.120 --> 00:26:56.200
space is going to be the high

00:26:54.600 --> 00:26:58.600
probability part of the space is going

00:26:56.200 --> 00:27:02.559
to be associated with good outputs

00:26:58.600 --> 00:27:07.000
because like when

00:27:02.559 --> 00:27:08.600
models are more sure of their outputs

00:27:07.000 --> 00:27:11.960
they're more likely to be

00:27:08.600 --> 00:27:13.520
good just because that indicates that

00:27:11.960 --> 00:27:15.240
like they're closer to the training data

00:27:13.520 --> 00:27:17.760
that it had and other things like that

00:27:15.240 --> 00:27:21.600
so model probabilities are associated

00:27:17.760 --> 00:27:23.760
with outputs uh with uh with good

00:27:21.600 --> 00:27:26.600
outputs but just

00:27:23.760 --> 00:27:29.440
correla separately from

00:27:26.600 --> 00:27:32.120
that I believe a model can identify when

00:27:29.440 --> 00:27:33.320
it's in a high probability segment of

00:27:32.120 --> 00:27:35.799
the space and when it's in a low

00:27:33.320 --> 00:27:39.399
probability segment of the space and

00:27:35.799 --> 00:27:39.399
because of that I expect

00:27:39.519 --> 00:27:45.519
that I like there are segments of the

00:27:43.240 --> 00:27:47.120
embedding space where it's more likely

00:27:45.519 --> 00:27:48.360
to answer yes about something being good

00:27:47.120 --> 00:27:50.960
or not and those are going to be

00:27:48.360 --> 00:27:54.760
associated with high uh like high

00:27:50.960 --> 00:27:56.159
probability outbreaks as well and also

00:27:54.760 --> 00:27:57.760
models are more likely to generate

00:27:56.159 --> 00:28:00.240
outputs that are high probability

00:27:57.760 --> 00:28:02.320
according into their model by definition

00:28:00.240 --> 00:28:03.880
so all three of those effects together

00:28:02.320 --> 00:28:05.640
would basically go into a model being

00:28:03.880 --> 00:28:09.120
bios supports its own outputs compared

00:28:05.640 --> 00:28:11.559
to that puts in another model but um

00:28:09.120 --> 00:28:13.279
yeah this is a very handwavy explanation

00:28:11.559 --> 00:28:15.519
but like putting the two the three

00:28:13.279 --> 00:28:18.600
together models output high probability

00:28:15.519 --> 00:28:20.880
things from their own probability Space

00:28:18.600 --> 00:28:23.440
by definition

00:28:20.880 --> 00:28:25.760
um things that are high probability are

00:28:23.440 --> 00:28:27.519
associated with being good uh just

00:28:25.760 --> 00:28:29.279
because otherwise a model would be

00:28:27.519 --> 00:28:31.840
outputting garbage

00:28:29.279 --> 00:28:33.840
and um the final thing which is more

00:28:31.840 --> 00:28:35.679
tenuous is if the model is in a high

00:28:33.840 --> 00:28:37.919
probability segment of the space it's

00:28:35.679 --> 00:28:39.760
more likely to Output yes according to a

00:28:37.919 --> 00:28:41.480
question of it being good and I I think

00:28:39.760 --> 00:28:44.360
that's probably true but I'm not 100%

00:28:41.480 --> 00:28:44.360
sure about the the

00:28:45.559 --> 00:28:51.039
fin um maybe maybe someone wants to

00:28:49.000 --> 00:28:52.840
examinate examine that as a final

00:28:51.039 --> 00:28:54.200
project it seems like a interesting

00:28:52.840 --> 00:28:57.080
interesting

00:28:54.200 --> 00:29:00.039
question um cool uh were there any other

00:28:57.080 --> 00:29:00.039
questions about these methods

00:29:00.159 --> 00:29:07.120
here um okay so when I say like an

00:29:03.960 --> 00:29:11.080
evaluation metric is good or not what do

00:29:07.120 --> 00:29:13.200
I mean by this being good or not um or a

00:29:11.080 --> 00:29:16.880
reward model or whatever else and

00:29:13.200 --> 00:29:18.440
basically the um the way we typically do

00:29:16.880 --> 00:29:19.840
this is by doing something called meta

00:29:18.440 --> 00:29:22.440
evaluation so it's called meta

00:29:19.840 --> 00:29:25.799
evaluation because it's evaluation of

00:29:22.440 --> 00:29:29.279
evaluation and uh the way we do this is

00:29:25.799 --> 00:29:32.519
we have human uh scores and we have

00:29:29.279 --> 00:29:34.760
automatic scores and we usually

00:29:32.519 --> 00:29:38.640
calculate some sort of correlation

00:29:34.760 --> 00:29:41.000
between the scores so um typical ones

00:29:38.640 --> 00:29:46.440
are rank correlations like Pearson's

00:29:41.000 --> 00:29:48.799
correlation or tendle uh Tow and uh so

00:29:46.440 --> 00:29:51.200
the more Associated the automatic scores

00:29:48.799 --> 00:29:53.960
are with the human scores the higher

00:29:51.200 --> 00:29:55.159
these correlations are going to be um

00:29:53.960 --> 00:29:57.559
there's other things that you can

00:29:55.159 --> 00:30:00.080
calculate so if you're trying to figure

00:29:57.559 --> 00:30:01.640
out whether a model um matches human

00:30:00.080 --> 00:30:04.279
pairwise preferences you can just

00:30:01.640 --> 00:30:06.440
calculate accuracy so I didn't put that

00:30:04.279 --> 00:30:08.080
on um I didn't put that on the slide

00:30:06.440 --> 00:30:10.880
here but you can just calculate accuracy

00:30:08.080 --> 00:30:13.120
of pairwise preferences um you can also

00:30:10.880 --> 00:30:15.360
calculate the absolute error between the

00:30:13.120 --> 00:30:19.320
the judgments if you want to know uh

00:30:15.360 --> 00:30:21.720
whether the absolute error matches so um

00:30:19.320 --> 00:30:24.159
the these are good things to do if you

00:30:21.720 --> 00:30:25.600
want to use an evaluation metric but you

00:30:24.159 --> 00:30:27.200
aren't sure whether it's good or not I

00:30:25.600 --> 00:30:29.640
would check to see whether the authors

00:30:27.200 --> 00:30:32.000
have done this sort of meta evaluation

00:30:29.640 --> 00:30:33.760
if they haven't be a little bit

00:30:32.000 --> 00:30:36.960
suspicious if they have be a little bit

00:30:33.760 --> 00:30:39.799
less suspicious but um

00:30:36.960 --> 00:30:42.960
yeah how do people do this typically uh

00:30:39.799 --> 00:30:45.640
usually they create uh data sets like

00:30:42.960 --> 00:30:49.440
the WM they use data sets like the WMT

00:30:45.640 --> 00:30:53.960
shared tasks um or

00:30:49.440 --> 00:30:57.679
uh uh like some evl um but there's also

00:30:53.960 --> 00:30:59.960
other ways to create um uh there's also

00:30:57.679 --> 00:31:01.639
Lots other data sets but in order to do

00:30:59.960 --> 00:31:05.639
this reliably you need a fairly large

00:31:01.639 --> 00:31:05.639
data set so it's one thing to be aware

00:31:07.080 --> 00:31:10.760
of

00:31:08.720 --> 00:31:14.200
cool

00:31:10.760 --> 00:31:16.360
um then the final thing um all of the

00:31:14.200 --> 00:31:17.919
automatic evaluation methods that I

00:31:16.360 --> 00:31:20.240
talked about now are trying to match

00:31:17.919 --> 00:31:22.679
human preferences but that's not the

00:31:20.240 --> 00:31:24.960
only thing that you necessarily want to

00:31:22.679 --> 00:31:28.440
do the final thing that you might want

00:31:24.960 --> 00:31:30.840
to do is uh use the model outputs in a

00:31:28.440 --> 00:31:34.200
downstream system and see whether they

00:31:30.840 --> 00:31:36.399
are effective for that so there's two

00:31:34.200 --> 00:31:39.080
concepts of intrinsic evaluation and

00:31:36.399 --> 00:31:41.720
extrinsic evaluation so intrinsic

00:31:39.080 --> 00:31:44.159
evaluation um evaluates the quality of

00:31:41.720 --> 00:31:45.720
the output itself and so that would be

00:31:44.159 --> 00:31:48.639
like asking a human directly about how

00:31:45.720 --> 00:31:50.720
good is this output extrinsic evaluation

00:31:48.639 --> 00:31:53.679
is evaluating output quality by its

00:31:50.720 --> 00:31:57.000
utility um and so just to give one

00:31:53.679 --> 00:31:58.360
example um if you can evaluate large

00:31:57.000 --> 00:32:00.200
language model summary

00:31:58.360 --> 00:32:04.200
through question answering

00:32:00.200 --> 00:32:05.880
accuracy um and so you can take the

00:32:04.200 --> 00:32:07.399
output of an llm and feed it through a

00:32:05.880 --> 00:32:09.600
question answering model and see whether

00:32:07.399 --> 00:32:12.399
you're able to answer questions based on

00:32:09.600 --> 00:32:15.799
this and that kind of gives you a better

00:32:12.399 --> 00:32:18.279
idea of whether the summary require uh

00:32:15.799 --> 00:32:20.120
incorporates requisite information but

00:32:18.279 --> 00:32:22.120
if you think about anything an llm can

00:32:20.120 --> 00:32:23.760
be used for usually it's part of a

00:32:22.120 --> 00:32:26.679
bigger system so you can evaluate it as

00:32:23.760 --> 00:32:28.399
a part of that bigger system um the

00:32:26.679 --> 00:32:30.639
problem with this is it's a very

00:32:28.399 --> 00:32:33.960
indirect way of assessing things so like

00:32:30.639 --> 00:32:36.080
let's say your QA model is just bad uh

00:32:33.960 --> 00:32:38.480
how can you disentangle the effect of

00:32:36.080 --> 00:32:41.679
the L summary versus the QA model that's

00:32:38.480 --> 00:32:44.120
not a trivial thing to do so ideally

00:32:41.679 --> 00:32:47.000
like a combination of these two is

00:32:44.120 --> 00:32:47.000
practically the best way

00:32:48.039 --> 00:32:52.200
go cool so

00:32:56.039 --> 00:32:59.960
yeah yeah it wouldn't necessar

00:32:58.360 --> 00:33:05.679
say it's harder to do it might even be

00:32:59.960 --> 00:33:05.679
easier to do um which is like let's

00:33:06.679 --> 00:33:11.720
say Let me let me see if I can come up

00:33:09.360 --> 00:33:11.720
with

00:33:12.639 --> 00:33:17.600
example what let's

00:33:15.000 --> 00:33:19.670
say you

00:33:17.600 --> 00:33:22.979
are trying

00:33:19.670 --> 00:33:22.979
[Music]

00:33:24.639 --> 00:33:29.760
to let's say you're trying to

00:33:30.559 --> 00:33:33.559
guess

00:33:39.000 --> 00:33:45.399
whether let's say you're trying to guess

00:33:42.399 --> 00:33:46.559
whether a someone will be hired at a

00:33:45.399 --> 00:33:52.039
company or

00:33:46.559 --> 00:33:53.880
not based on an llm generated summary of

00:33:52.039 --> 00:33:58.880
their qualifications for a position or

00:33:53.880 --> 00:34:01.799
something like that um and

00:33:58.880 --> 00:34:03.080
you what actually maybe this is not a

00:34:01.799 --> 00:34:04.720
great example because whether you should

00:34:03.080 --> 00:34:06.960
be doing this ethically is a little bit

00:34:04.720 --> 00:34:08.159
unclear but let's say you were doing

00:34:06.960 --> 00:34:09.560
let's say you were doing something like

00:34:08.159 --> 00:34:11.520
that just because it's one example I can

00:34:09.560 --> 00:34:14.320
think of right now whether they will get

00:34:11.520 --> 00:34:16.320
hired or not is um is clear because you

00:34:14.320 --> 00:34:19.399
have a objective answer right whether

00:34:16.320 --> 00:34:21.480
they were hired or not um or maybe maybe

00:34:19.399 --> 00:34:23.800
another example would be like let's say

00:34:21.480 --> 00:34:26.320
um let's say you want to predict the

00:34:23.800 --> 00:34:29.599
diagnosis in a medical application based

00:34:26.320 --> 00:34:32.960
on an llm generated some of somebody's

00:34:29.599 --> 00:34:35.919
uh you know LM generated summary of

00:34:32.960 --> 00:34:38.480
somebody's you know past medical history

00:34:35.919 --> 00:34:40.839
and all this stuff and here you want the

00:34:38.480 --> 00:34:43.440
llm generated summary you definitely

00:34:40.839 --> 00:34:44.879
want the summary because the summary is

00:34:43.440 --> 00:34:47.560
going to be viewed by a doctor who will

00:34:44.879 --> 00:34:49.359
make the final decision but you also

00:34:47.560 --> 00:34:50.760
have information about the diagnoses of

00:34:49.359 --> 00:34:52.399
all the people in your medical system

00:34:50.760 --> 00:34:54.560
later because you know they went through

00:34:52.399 --> 00:34:56.480
your medical system for years and you

00:34:54.560 --> 00:34:58.200
know later like through lots of tests

00:34:56.480 --> 00:35:00.800
and stuff uh whether how they were

00:34:58.200 --> 00:35:02.320
diagnosed so you generate an LM based

00:35:00.800 --> 00:35:05.000
summary and then you predict the

00:35:02.320 --> 00:35:06.599
diagnosis from the summary so there the

00:35:05.000 --> 00:35:08.040
evaluation of the diagnosis is very

00:35:06.599 --> 00:35:11.480
clear because you kind of have a gold

00:35:08.040 --> 00:35:12.599
standard answer um but the EV intrinsic

00:35:11.480 --> 00:35:14.839
evaluation of whether it's a good

00:35:12.599 --> 00:35:16.839
summary or not is not as clear because

00:35:14.839 --> 00:35:19.400
you'd have pass do whether it's good and

00:35:16.839 --> 00:35:21.079
understandable summary so the extrinsic

00:35:19.400 --> 00:35:24.920
evaluation might be easier because it's

00:35:21.079 --> 00:35:26.480
clearer um so there are cases like that

00:35:24.920 --> 00:35:30.720
um the problem is you would have to have

00:35:26.480 --> 00:35:33.800
that data in order to do that um yeah do

00:35:30.720 --> 00:35:38.240
like evaluation yeah I was just

00:35:33.800 --> 00:35:40.800
wondering typically the

00:35:38.240 --> 00:35:42.880
like like how do you accomodate the

00:35:40.800 --> 00:35:47.160
diversity oh yeah that's a great that's

00:35:42.880 --> 00:35:50.240
a great question um so how do you how do

00:35:47.160 --> 00:35:50.240
you get these scores

00:35:50.720 --> 00:35:55.800
here there's a number of different

00:35:53.200 --> 00:35:59.160
things in the WMT shared tasks what they

00:35:55.800 --> 00:36:00.280
did is they did

00:35:59.160 --> 00:36:03.200
the first thing they do is they

00:36:00.280 --> 00:36:06.319
normalize by annotator and what they do

00:36:03.200 --> 00:36:10.400
is they basically take the zcore or Z

00:36:06.319 --> 00:36:12.240
score of the um of the human annotator's

00:36:10.400 --> 00:36:14.880
actual scores because some people are

00:36:12.240 --> 00:36:16.400
more harsh than other people and so what

00:36:14.880 --> 00:36:20.680
that means is you basically normalize to

00:36:16.400 --> 00:36:22.119
have zero mean in unit variance um and

00:36:20.680 --> 00:36:24.119
then after they've normalized to zero

00:36:22.119 --> 00:36:29.560
mean and unit variance then I think they

00:36:24.119 --> 00:36:29.560
average together different humans so um

00:36:30.160 --> 00:36:36.520
then for how do you deal with the fact

00:36:33.680 --> 00:36:38.040
that humans disagree on things and I

00:36:36.520 --> 00:36:39.480
think it's pretty varied I don't know if

00:36:38.040 --> 00:36:42.160
there's any gold standard way of doing

00:36:39.480 --> 00:36:43.839
it but sometimes you just average

00:36:42.160 --> 00:36:46.359
sometimes you throw away examples where

00:36:43.839 --> 00:36:47.960
humans disagree a lot um because like

00:36:46.359 --> 00:36:50.200
you can't get the humans to agree how

00:36:47.960 --> 00:36:53.319
could you expect how could you expect a

00:36:50.200 --> 00:36:55.119
machine to do well um so I think it it's

00:36:53.319 --> 00:36:59.200
a little bit test

00:36:55.119 --> 00:37:01.560
defending yeah so for

00:36:59.200 --> 00:37:04.560
generation inin

00:37:01.560 --> 00:37:06.280
andin yeah so for code generation that's

00:37:04.560 --> 00:37:08.200
I I I love this example because I've

00:37:06.280 --> 00:37:09.960
worked on code generation a lot of

00:37:08.200 --> 00:37:12.680
people only think about extrinsic

00:37:09.960 --> 00:37:14.400
evaluation of code Generation Um or I

00:37:12.680 --> 00:37:16.160
don't know if it's extrinsic but only

00:37:14.400 --> 00:37:19.160
think about execution based evaluation

00:37:16.160 --> 00:37:20.520
of code generation which is like you

00:37:19.160 --> 00:37:22.400
execute the code you see whether it

00:37:20.520 --> 00:37:25.040
passs unit tests and other things like

00:37:22.400 --> 00:37:26.839
this but in reality actually there's a

00:37:25.040 --> 00:37:28.599
lot of other important things for code

00:37:26.839 --> 00:37:30.560
like readability and other stuff like

00:37:28.599 --> 00:37:32.160
that and you should be evaluating those

00:37:30.560 --> 00:37:34.920
things but I think a lot of people like

00:37:32.160 --> 00:37:36.520
kind of ignore that so um there there

00:37:34.920 --> 00:37:38.880
are a few Pap that do that but most of

00:37:36.520 --> 00:37:41.000
the time people just execute the Cod

00:37:38.880 --> 00:37:45.520
process

00:37:41.000 --> 00:37:47.760
un cool okay um so yeah moving on to the

00:37:45.520 --> 00:37:51.160
learning part so now I'd like to talk

00:37:47.760 --> 00:37:55.280
about uh learning and the first thing

00:37:51.160 --> 00:37:59.480
I'll cover is error and risk and so

00:37:55.280 --> 00:38:02.280
basically um the way we calculate air is

00:37:59.480 --> 00:38:03.119
we generate an output and we calculate

00:38:02.280 --> 00:38:07.680
its

00:38:03.119 --> 00:38:09.480
Badness um and so generating the output

00:38:07.680 --> 00:38:13.160
could be argmax it could be sampling it

00:38:09.480 --> 00:38:15.800
could be anything else like that um and

00:38:13.160 --> 00:38:18.640
we calculate its Badness uh which is one

00:38:15.800 --> 00:38:21.040
minus in which could be like how bad is

00:38:18.640 --> 00:38:22.720
the output uh if you're you have a

00:38:21.040 --> 00:38:24.760
Badness measure or it could be one minus

00:38:22.720 --> 00:38:28.400
the evaluation Square to calculate its

00:38:24.760 --> 00:38:30.160
Badness and this is defined as error

00:38:28.400 --> 00:38:31.440
and generally what you want to do is you

00:38:30.160 --> 00:38:33.520
want to minimize

00:38:31.440 --> 00:38:36.800
error

00:38:33.520 --> 00:38:39.400
um because in the end you're going to be

00:38:36.800 --> 00:38:42.359
deploying A system that just outputs you

00:38:39.400 --> 00:38:46.079
know one thing and uh you're going to

00:38:42.359 --> 00:38:49.800
want that to be as good a thing as

00:38:46.079 --> 00:38:53.000
possible um but the problem with this is

00:38:49.800 --> 00:38:56.400
there's no easy way to actually optimize

00:38:53.000 --> 00:38:59.079
this value in especially in a text

00:38:56.400 --> 00:39:01.800
generation sty setting but even in the

00:38:59.079 --> 00:39:06.839
classification setting we can't easily

00:39:01.800 --> 00:39:06.839
maximize err because um if you look at

00:39:09.040 --> 00:39:14.200
the if you look at the surface of air uh

00:39:12.760 --> 00:39:15.960
at some point you're going to have a

00:39:14.200 --> 00:39:18.319
non-differentiable part when you take

00:39:15.960 --> 00:39:21.119
the argmax and or when you do sampling

00:39:18.319 --> 00:39:23.319
or anything like that so um you're not

00:39:21.119 --> 00:39:27.119
going to be able to do gradient based

00:39:23.319 --> 00:39:29.200
optimization so what we do normally is

00:39:27.119 --> 00:39:33.400
um

00:39:29.200 --> 00:39:37.000
we instead calculate something uh called

00:39:33.400 --> 00:39:38.560
risk and what risk looks like is uh we

00:39:37.000 --> 00:39:40.599
talked a little bit about minimum based

00:39:38.560 --> 00:39:43.520
risk for decoding but this is for uh

00:39:40.599 --> 00:39:46.160
training time and what it looks like is

00:39:43.520 --> 00:39:49.040
it's essentially the expected err of the

00:39:46.160 --> 00:39:52.359
output and the expected err of the

00:39:49.040 --> 00:39:54.760
output um includes a probability in the

00:39:52.359 --> 00:39:58.240
objective function here and that

00:39:54.760 --> 00:40:01.079
probability uh is differential basically

00:39:58.240 --> 00:40:02.319
so we can um uh we can easily do

00:40:01.079 --> 00:40:05.720
gradient based

00:40:02.319 --> 00:40:09.119
optimization through it um the problem

00:40:05.720 --> 00:40:12.200
with this is It's differentiable but for

00:40:09.119 --> 00:40:17.160
text generation for example the sum is

00:40:12.200 --> 00:40:20.319
intractable because we have a combinator

00:40:17.160 --> 00:40:23.880
large number of potential outputs um

00:40:20.319 --> 00:40:25.520
because you know if this is we've talked

00:40:23.880 --> 00:40:28.720
about this before but if this is like

00:40:25.520 --> 00:40:30.680
link you know 50 and we have a 30,000

00:40:28.720 --> 00:40:32.839
vocabul that's 30,000 to the 50

00:40:30.680 --> 00:40:34.599
possibilities we can't take a su over

00:40:32.839 --> 00:40:36.359
that many

00:40:34.599 --> 00:40:38.400
possibilities

00:40:36.359 --> 00:40:42.680
um

00:40:38.400 --> 00:40:45.839
so minimum R risk training uh tries to

00:40:42.680 --> 00:40:48.440
minimize risk reinforcement learning

00:40:45.839 --> 00:40:50.040
also many of the models especially

00:40:48.440 --> 00:40:53.599
policy gradient models are trying to

00:40:50.040 --> 00:40:55.240
minimize risk as well so um but the

00:40:53.599 --> 00:40:58.040
reason why I wanted to talk about risk

00:40:55.240 --> 00:41:00.440
first is because this is very simple to

00:40:58.040 --> 00:41:01.640
get to from the uh the point of view of

00:41:00.440 --> 00:41:06.560
like all the things that we've studied

00:41:01.640 --> 00:41:06.560
so so I think it's talking about

00:41:06.760 --> 00:41:11.800
that

00:41:08.319 --> 00:41:15.520
um one other thing that I should mention

00:41:11.800 --> 00:41:18.400
about is

00:41:15.520 --> 00:41:23.079
um or no sorry I'll I'll talk about that

00:41:18.400 --> 00:41:26.880
later so when we want to optimize risk

00:41:23.079 --> 00:41:30.560
um what we do is we sample in order to

00:41:26.880 --> 00:41:35.520
make this trct so a very simple way to

00:41:30.560 --> 00:41:37.640
minimize risk is instead of um instead

00:41:35.520 --> 00:41:39.359
of summing over all of the possible

00:41:37.640 --> 00:41:42.760
outputs we sum over a small number of

00:41:39.359 --> 00:41:46.079
possible outputs and we upgrade uh and

00:41:42.760 --> 00:41:47.359
we uh sorry normalize uh to make this

00:41:46.079 --> 00:41:51.200
all add up to

00:41:47.359 --> 00:41:52.839
one and so this normalizer here is

00:41:51.200 --> 00:41:55.319
basically the sum over all of the

00:41:52.839 --> 00:41:58.599
probabilities that we have uh on the top

00:41:55.319 --> 00:42:02.119
part here and and these samples can be

00:41:58.599 --> 00:42:05.480
created either using sampling or n best

00:42:02.119 --> 00:42:07.040
search we don't need to have from the

00:42:05.480 --> 00:42:11.040
point of view of doing this sort of

00:42:07.040 --> 00:42:13.960
minimum risk training the kind of

00:42:11.040 --> 00:42:16.880
correct way of doing this is sampling

00:42:13.960 --> 00:42:19.880
using ancestral sampling uh like we

00:42:16.880 --> 00:42:23.079
talked about before and um in minimizing

00:42:19.880 --> 00:42:25.839
the output based on the the samples but

00:42:23.079 --> 00:42:28.480
the problem with that is um as many of

00:42:25.839 --> 00:42:31.440
you also might have seen when you were

00:42:28.480 --> 00:42:33.599
sampling from your language model uh

00:42:31.440 --> 00:42:35.160
from assignment one if you sample with

00:42:33.599 --> 00:42:38.040
temperature one it gives you a lot of

00:42:35.160 --> 00:42:40.720
like not very good outlets right and so

00:42:38.040 --> 00:42:43.400
if you're sampling with temperature one

00:42:40.720 --> 00:42:45.000
um you'll be exploring a a very large

00:42:43.400 --> 00:42:47.880
part of the space that actually isn't

00:42:45.000 --> 00:42:49.720
very good and so because of this uh some

00:42:47.880 --> 00:42:51.480
other Alternatives that you can use is

00:42:49.720 --> 00:42:53.400
you can just do endb search to find the

00:42:51.480 --> 00:42:55.280
best outputs or you can sample with a

00:42:53.400 --> 00:42:58.079
temperature that's not one or something

00:42:55.280 --> 00:43:00.240
like that and basically create uh you

00:42:58.079 --> 00:43:02.520
know a list of possible hypotheses and

00:43:00.240 --> 00:43:04.079
then normalize other B so that's another

00:43:02.520 --> 00:43:06.240
option and very often not using

00:43:04.079 --> 00:43:11.200
temperature one is a better

00:43:06.240 --> 00:43:15.280
way um if you're sampling with not

00:43:11.200 --> 00:43:18.640
temperature one and you are um

00:43:15.280 --> 00:43:20.920
potentially getting multiple outputs you

00:43:18.640 --> 00:43:23.400
should try to D duplicate or sample

00:43:20.920 --> 00:43:25.480
without replacement because if you get

00:43:23.400 --> 00:43:27.559
multiple outputs here it messes up your

00:43:25.480 --> 00:43:30.680
equations if you basically uh have the

00:43:27.559 --> 00:43:30.680
same one in there multiple

00:43:32.160 --> 00:43:37.800
times cool so so this is a really simple

00:43:35.880 --> 00:43:40.079
example of how you can do minimal risk

00:43:37.800 --> 00:43:42.119
training but now I want to get into uh

00:43:40.079 --> 00:43:44.640
like reinforcement learning which is the

00:43:42.119 --> 00:43:48.119
framing that most um

00:43:44.640 --> 00:43:50.760
modern Works about this Paulo uh one

00:43:48.119 --> 00:43:52.559
thing I should mention is there are

00:43:50.760 --> 00:43:55.240
actually other alternatives to learning

00:43:52.559 --> 00:43:57.359
from uh human feedback including like

00:43:55.240 --> 00:43:59.359
margin loss margin based losses and

00:43:57.359 --> 00:44:00.960
other stuff like that but most people

00:43:59.359 --> 00:44:03.440
nowadays use reinforcement learning so

00:44:00.960 --> 00:44:06.359
I'm only going to cover that

00:44:03.440 --> 00:44:08.440
here so what is reinforcement learning

00:44:06.359 --> 00:44:11.000
um learning reinforcement learning is

00:44:08.440 --> 00:44:14.559
learning where we have an environment uh

00:44:11.000 --> 00:44:16.079
x uh ability to make actions a and get a

00:44:14.559 --> 00:44:20.160
delayed reward

00:44:16.079 --> 00:44:21.880
R and um there's a really nice example

00:44:20.160 --> 00:44:24.400
uh if you're not familiar with the

00:44:21.880 --> 00:44:27.480
basics of policy gradient by Andre

00:44:24.400 --> 00:44:28.800
karpathy which I linked in the um in the

00:44:27.480 --> 00:44:29.680
recommended reading so you can take a

00:44:28.800 --> 00:44:34.680
look at

00:44:29.680 --> 00:44:37.240
that um but in that example gives an

00:44:34.680 --> 00:44:39.440
example of pong uh where you're playing

00:44:37.240 --> 00:44:42.640
the game pong where X is your observed

00:44:39.440 --> 00:44:45.640
image a is up or down and R is the wind

00:44:42.640 --> 00:44:47.480
loss at the end of the game uh does

00:44:45.640 --> 00:44:50.559
anyone have an idea about uh what this

00:44:47.480 --> 00:44:52.119
looks like for any arbitrary NLP task

00:44:50.559 --> 00:44:56.520
that we might want to do reinforcement

00:44:52.119 --> 00:44:59.040
learning for so what what is X what is a

00:44:56.520 --> 00:44:59.040
and what is

00:45:00.040 --> 00:45:04.680
are pick your favorite uh your favorite

00:45:06.920 --> 00:45:09.920
Trask

00:45:10.960 --> 00:45:18.400
anybody

00:45:12.520 --> 00:45:18.400
yeah be or what what's X first

00:45:19.680 --> 00:45:28.720
yeah you have generate okay is the

00:45:24.440 --> 00:45:29.720
next be like the Buton like whether or

00:45:28.720 --> 00:45:32.520
not

00:45:29.720 --> 00:45:35.240
you okay yeah I I think this is very

00:45:32.520 --> 00:45:37.119
close just to repeat it it's like X is

00:45:35.240 --> 00:45:39.599
what you've generated so far a is the

00:45:37.119 --> 00:45:41.559
next token and R is the button that the

00:45:39.599 --> 00:45:45.400
user clicks about whether it's good or

00:45:41.559 --> 00:45:46.920
not um I think that's reasonably good

00:45:45.400 --> 00:45:48.760
although I don't know if we'd expect

00:45:46.920 --> 00:45:52.960
them to click the button every token we

00:45:48.760 --> 00:45:54.880
generate right so um it might be that X

00:45:52.960 --> 00:45:57.880
is the conversational history up till

00:45:54.880 --> 00:46:02.319
this point um a

00:45:57.880 --> 00:46:04.280
a could be a next token generation and

00:46:02.319 --> 00:46:06.520
then R is a reward we get in an

00:46:04.280 --> 00:46:08.280
arbitrary time point it might not be

00:46:06.520 --> 00:46:09.960
like immediately after generating the

00:46:08.280 --> 00:46:12.040
next token but it might be later and

00:46:09.960 --> 00:46:13.480
that's actually really really important

00:46:12.040 --> 00:46:15.040
from the point of view of reinforcement

00:46:13.480 --> 00:46:19.599
learning and I'll I'll talk about that

00:46:15.040 --> 00:46:23.040
in a second um anyone have an idea from

00:46:19.599 --> 00:46:24.960
I don't know uh code generation or

00:46:23.040 --> 00:46:28.119
translation or some other

00:46:24.960 --> 00:46:31.160
things C generation maybe s is a

00:46:28.119 --> 00:46:33.040
compiler or like the gra scpt and then

00:46:31.160 --> 00:46:37.000
the

00:46:33.040 --> 00:46:42.520
is the actual code that right and reward

00:46:37.000 --> 00:46:44.839
is yep um so X could be the compiler

00:46:42.520 --> 00:46:47.559
it's probably the compiler and all of

00:46:44.839 --> 00:46:50.200
the surrounding code context like what

00:46:47.559 --> 00:46:52.520
what is the natural language output and

00:46:50.200 --> 00:46:53.960
it's also um you know what is the

00:46:52.520 --> 00:46:57.280
project that you're you're working on

00:46:53.960 --> 00:47:00.079
and stuff like that um a i think

00:46:57.280 --> 00:47:02.800
typically we would treat each token in

00:47:00.079 --> 00:47:04.160
the code to be an action um and then R

00:47:02.800 --> 00:47:06.599
would be the reward after a long

00:47:04.160 --> 00:47:08.640
sequence of actions um and it could be

00:47:06.599 --> 00:47:11.119
the reward from the compiler it could be

00:47:08.640 --> 00:47:13.160
the reward from a code readability model

00:47:11.119 --> 00:47:15.720
it could be the reward from a speed

00:47:13.160 --> 00:47:17.079
execution speed and stuff like that so

00:47:15.720 --> 00:47:18.839
like one of the interesting things about

00:47:17.079 --> 00:47:22.640
R is you can be really creative about

00:47:18.839 --> 00:47:25.400
how you form R um which is not easy to

00:47:22.640 --> 00:47:27.319
do uh if you're just doing maximum

00:47:25.400 --> 00:47:29.240
likelihood also so you can come up with

00:47:27.319 --> 00:47:32.920
a r that really matches with like what

00:47:29.240 --> 00:47:36.559
you want um what you want in an output

00:47:32.920 --> 00:47:40.079
so why reinforcement learning in NLP um

00:47:36.559 --> 00:47:42.599
and I think there's basically three um

00:47:40.079 --> 00:47:44.240
three answers the first one is you have

00:47:42.599 --> 00:47:49.000
a typical reinforcement learning

00:47:44.240 --> 00:47:51.119
scenario um where you have a dialogue

00:47:49.000 --> 00:47:52.720
where you get lots of responses and then

00:47:51.119 --> 00:47:54.559
you get a reward at the end so the

00:47:52.720 --> 00:47:57.359
thumbs up and thumbs down from humans is

00:47:54.559 --> 00:47:59.839
a very typical example of

00:47:57.359 --> 00:48:02.800
uh reinforcement learning because you

00:47:59.839 --> 00:48:05.000
get a delayed reward uh at some point in

00:48:02.800 --> 00:48:07.599
the dialogue when a human presses up or

00:48:05.000 --> 00:48:09.280
down um another like actually more

00:48:07.599 --> 00:48:11.680
technical scenario where reinforcement

00:48:09.280 --> 00:48:14.960
learning has been used um for a long

00:48:11.680 --> 00:48:17.400
time is call centers so we've had

00:48:14.960 --> 00:48:20.680
dialogue systems for call centers and

00:48:17.400 --> 00:48:23.160
then if you complete a ticket purchase

00:48:20.680 --> 00:48:24.839
um or you complete resolve a ticket

00:48:23.160 --> 00:48:27.480
without ever having to go to a human

00:48:24.839 --> 00:48:30.800
operator you get a really big reward

00:48:27.480 --> 00:48:33.640
if you have to go to the human operator

00:48:30.800 --> 00:48:36.400
you get maybe a smaller reward and if

00:48:33.640 --> 00:48:39.200
the person yells at you and hangs up

00:48:36.400 --> 00:48:41.640
then you get a really negative reward so

00:48:39.200 --> 00:48:43.040
um this is kind of the typical example

00:48:41.640 --> 00:48:45.599
reinforcement learning has been used for

00:48:43.040 --> 00:48:48.520
a long time there another example is if

00:48:45.599 --> 00:48:53.280
you have like latent variables uh chains

00:48:48.520 --> 00:48:55.799
of thought where um you decide the

00:48:53.280 --> 00:48:58.839
latent variable and then get a reward um

00:48:55.799 --> 00:49:02.799
you get a reward based Bas on how those

00:48:58.839 --> 00:49:03.920
latent variables affect the output so um

00:49:02.799 --> 00:49:07.200
this

00:49:03.920 --> 00:49:09.799
is uh this is another example

00:49:07.200 --> 00:49:12.599
because the Chain of Thought itself

00:49:09.799 --> 00:49:13.880
might not actually be good you might

00:49:12.599 --> 00:49:15.839
have a bad Chain of Thought and still

00:49:13.880 --> 00:49:17.760
get the correct answer so you don't

00:49:15.839 --> 00:49:19.640
actually know for sure that a chain of

00:49:17.760 --> 00:49:22.359
thought that was automatically generated

00:49:19.640 --> 00:49:24.799
is good or not but um that so that kind

00:49:22.359 --> 00:49:27.000
of makes it a reinforcement learning

00:49:24.799 --> 00:49:29.520
problem and another thing is you might

00:49:27.000 --> 00:49:32.520
have a sequence level evaluation metric

00:49:29.520 --> 00:49:34.240
um so that you can't optimize the

00:49:32.520 --> 00:49:36.839
evaluation metric without uh first

00:49:34.240 --> 00:49:38.480
generating the whole like sequence so

00:49:36.839 --> 00:49:40.880
that would be any of the evaluation

00:49:38.480 --> 00:49:42.400
metrics that I talked about before so um

00:49:40.880 --> 00:49:44.720
these are three scenarios where you can

00:49:42.400 --> 00:49:47.079
use reinforcement

00:49:44.720 --> 00:49:50.000
planning so

00:49:47.079 --> 00:49:51.400
um I'm going to make a few steps through

00:49:50.000 --> 00:49:54.640
but like let's start again with our

00:49:51.400 --> 00:49:57.359
supervised mle loss and uh that's just

00:49:54.640 --> 00:50:01.799
the log probability here um in the

00:49:57.359 --> 00:50:04.160
context of reinforcement learning this

00:50:01.799 --> 00:50:07.079
is also called imitation

00:50:04.160 --> 00:50:08.880
learning because um essentially you're

00:50:07.079 --> 00:50:12.680
learning how to perform actions by

00:50:08.880 --> 00:50:14.559
imitating a teacher um and imitation

00:50:12.680 --> 00:50:15.960
learning is not just supervised mle

00:50:14.559 --> 00:50:18.440
there's also other varieties of

00:50:15.960 --> 00:50:21.440
imitation learning but um this is one

00:50:18.440 --> 00:50:21.440
variety of imitation

00:50:22.520 --> 00:50:27.640
learning the next thing I'd like to talk

00:50:24.599 --> 00:50:30.079
about is self-training and basically

00:50:27.640 --> 00:50:31.760
self-training the idea is that you

00:50:30.079 --> 00:50:33.720
sample or argmax according to the

00:50:31.760 --> 00:50:36.119
current model so you have your current

00:50:33.720 --> 00:50:38.000
model and you get a sample from it and

00:50:36.119 --> 00:50:41.520
then you use the sample or samples to

00:50:38.000 --> 00:50:43.680
maximize likelihood so um basically

00:50:41.520 --> 00:50:47.520
instead of doing maximum likelihood with

00:50:43.680 --> 00:50:49.520
respect to the a gold standard output

00:50:47.520 --> 00:50:51.280
you're doing it with respect to your own

00:50:49.520 --> 00:50:55.280
output

00:50:51.280 --> 00:50:55.280
so does this seem like a good

00:50:55.640 --> 00:51:03.880
idea I see a few people shaking heads um

00:51:00.480 --> 00:51:03.880
any ideas why this is not a good

00:51:04.680 --> 00:51:07.680
idea

00:51:15.040 --> 00:51:20.599
yeah yeah exactly so if you don't have

00:51:17.720 --> 00:51:23.760
any access to any notion well it's good

00:51:20.599 --> 00:51:27.480
um this will be optimizing towards good

00:51:23.760 --> 00:51:28.839
outputs and bad outputs right so um your

00:51:27.480 --> 00:51:30.200
model might be outputting bad outputs

00:51:28.839 --> 00:51:32.839
and you're just reinforcing the errors

00:51:30.200 --> 00:51:35.160
set the model R already nonetheless like

00:51:32.839 --> 00:51:37.799
self trining actually improves your

00:51:35.160 --> 00:51:39.680
accuracy somewhat in some cases like for

00:51:37.799 --> 00:51:43.040
example if your accuracy is if your

00:51:39.680 --> 00:51:45.520
model is Right more often than not um

00:51:43.040 --> 00:51:49.119
basically optimizing towards the more

00:51:45.520 --> 00:51:51.720
often the not right outputs can actually

00:51:49.119 --> 00:51:53.640
um due to the implicit regularization

00:51:51.720 --> 00:51:55.000
that models have and early stopping and

00:51:53.640 --> 00:51:56.559
other things like that it can actually

00:51:55.000 --> 00:51:59.280
move you in the right direction and

00:51:56.559 --> 00:52:01.559
improve accuracy

00:51:59.280 --> 00:52:05.000
um

00:52:01.559 --> 00:52:06.640
so there are alternatives to this that

00:52:05.000 --> 00:52:09.520
further improve accuracy so like for

00:52:06.640 --> 00:52:12.720
example if you have multiple models and

00:52:09.520 --> 00:52:16.200
um you only generate sentences where the

00:52:12.720 --> 00:52:17.760
models agree then this can improve your

00:52:16.200 --> 00:52:20.000
uh overall accuracy

00:52:17.760 --> 00:52:24.240
further um this is called code training

00:52:20.000 --> 00:52:27.799
it was actually uh created by uh uh

00:52:24.240 --> 00:52:30.160
people at at CMU as well and another

00:52:27.799 --> 00:52:32.280
successful alternative uh is adding

00:52:30.160 --> 00:52:34.920
noise to the input to match the noise

00:52:32.280 --> 00:52:38.760
that you find in the output so if you uh

00:52:34.920 --> 00:52:40.720
add like word uh word-based Dropout or

00:52:38.760 --> 00:52:44.000
other things like that this can also

00:52:40.720 --> 00:52:47.400
help uh accommodate these things but

00:52:44.000 --> 00:52:48.920
anyway um so self trining is is useful

00:52:47.400 --> 00:52:50.480
but there are better Alternatives if you

00:52:48.920 --> 00:52:54.079
can get a reward

00:52:50.480 --> 00:52:55.559
function so um the simplest variety of

00:52:54.079 --> 00:52:56.960
this is something called policy gradient

00:52:55.559 --> 00:52:59.720
or reinforce

00:52:56.960 --> 00:53:02.319
um or more specifically reinforce and

00:52:59.720 --> 00:53:06.280
basically what this does is this adds a

00:53:02.319 --> 00:53:08.359
term that scales the loss by the reward

00:53:06.280 --> 00:53:12.400
so if you can get a reward for each

00:53:08.359 --> 00:53:15.680
output basically this

00:53:12.400 --> 00:53:18.119
um you uh instead of doing self trining

00:53:15.680 --> 00:53:21.760
entirely by itself you multiply it by a

00:53:18.119 --> 00:53:23.119
reward and this allows you to increase

00:53:21.760 --> 00:53:24.640
the likelihood of things that get a high

00:53:23.119 --> 00:53:28.440
reward decrease the likelihood of things

00:53:24.640 --> 00:53:28.440
that get a low reward

00:53:29.680 --> 00:53:34.960
so uh a brief quiz here under what

00:53:32.440 --> 00:53:37.599
conditions is this equal equivalent to

00:53:34.960 --> 00:53:41.480
ml or essentially equivalent to maximum

00:53:37.599 --> 00:53:43.079
leg uh estimation and so like in order

00:53:41.480 --> 00:53:45.480
to make this quiz easier I'll go back to

00:53:43.079 --> 00:53:47.720
maximum likelihood estimation so it

00:53:45.480 --> 00:53:50.359
looked a bit like this um you calculated

00:53:47.720 --> 00:53:53.440
the log probability of the true output

00:53:50.359 --> 00:53:55.440
and now let me go uh to

00:53:53.440 --> 00:53:56.960
here any

00:53:55.440 --> 00:54:00.119
ideas

00:53:56.960 --> 00:54:05.040
yeah when your reward equals to

00:54:00.119 --> 00:54:05.040
one some sometimes in zero other times

00:54:07.760 --> 00:54:10.960
what any

00:54:12.760 --> 00:54:17.520
ideas what when when does your reward

00:54:15.280 --> 00:54:19.640
need to be equal to one in order to make

00:54:17.520 --> 00:54:23.400
this

00:54:19.640 --> 00:54:23.400
equation equivalent this

00:54:24.960 --> 00:54:31.680
equation yeah when Y and Y hat are the

00:54:27.319 --> 00:54:36.119
same so um basically

00:54:31.680 --> 00:54:38.880
this objective is equivalent to the mle

00:54:36.119 --> 00:54:43.160
objective when you're using a zero1

00:54:38.880 --> 00:54:44.480
loss um where or you're using an

00:54:43.160 --> 00:54:46.359
evaluation function that gives you a

00:54:44.480 --> 00:54:50.920
score of one when it's exact match and

00:54:46.359 --> 00:54:51.720
zero when it's not exact match so um but

00:54:50.920 --> 00:54:54.480
that

00:54:51.720 --> 00:54:56.440
also demonstrates that this can be more

00:54:54.480 --> 00:54:58.400
flexible because you can have other

00:54:56.440 --> 00:55:00.160
rewards that are not just one and zero

00:54:58.400 --> 00:55:02.599
for exact match but you can use things

00:55:00.160 --> 00:55:05.359
that give you partial credit you can use

00:55:02.599 --> 00:55:06.880
things that uplate multiple potential uh

00:55:05.359 --> 00:55:08.880
potentially correct outputs and other

00:55:06.880 --> 00:55:13.400
things like

00:55:08.880 --> 00:55:17.160
that so one problem with these methods

00:55:13.400 --> 00:55:21.799
is um how do we know which action led to

00:55:17.160 --> 00:55:24.720
the reward so the best scenario is after

00:55:21.799 --> 00:55:26.359
each action you get a reward so after

00:55:24.720 --> 00:55:28.960
each token that you generated you get

00:55:26.359 --> 00:55:31.240
get a thumbs up or thumbs down uh from

00:55:28.960 --> 00:55:34.280
the user about whether they like that

00:55:31.240 --> 00:55:36.000
token or not um and how much happier

00:55:34.280 --> 00:55:37.720
they are after you generated that token

00:55:36.000 --> 00:55:42.400
than they were before you generated that

00:55:37.720 --> 00:55:44.200
token um the problem with this is that

00:55:42.400 --> 00:55:45.799
that's completely infeasible right like

00:55:44.200 --> 00:55:47.039
every time after you use chat GPD you're

00:55:45.799 --> 00:55:50.480
not going to press thumbs up and thumbs

00:55:47.039 --> 00:55:52.559
down after each token so um in reality

00:55:50.480 --> 00:55:55.559
what we get is usually we get it at the

00:55:52.559 --> 00:55:57.000
end of uh roll out of many many

00:55:55.559 --> 00:55:58.640
different actions and we're not sure

00:55:57.000 --> 00:55:59.720
which action is responsible for giving

00:55:58.640 --> 00:56:02.559
us the

00:55:59.720 --> 00:56:05.440
reward and

00:56:02.559 --> 00:56:08.000
so there's a few typical ways of dealing

00:56:05.440 --> 00:56:09.640
with this um the most typical way of

00:56:08.000 --> 00:56:13.359
dealing with this right now is just not

00:56:09.640 --> 00:56:15.440
dealing with it um and just hoping that

00:56:13.359 --> 00:56:17.200
your optimization algorithm internally

00:56:15.440 --> 00:56:21.480
will be able to do credit

00:56:17.200 --> 00:56:24.520
assignment um and so what that entails

00:56:21.480 --> 00:56:27.319
is essentially you um give an equal

00:56:24.520 --> 00:56:29.880
reward for each token in the output

00:56:27.319 --> 00:56:32.480
other ways that you can deal with it are

00:56:29.880 --> 00:56:35.640
um you can assign decaying rewards from

00:56:32.480 --> 00:56:37.559
future events so like let's say let's

00:56:35.640 --> 00:56:41.839
say you're talking about a chat bot for

00:56:37.559 --> 00:56:44.119
example maybe this is the the most uh

00:56:41.839 --> 00:56:46.599
kind of intuitive way of thinking about

00:56:44.119 --> 00:56:50.400
it but you you have a chat bot you have

00:56:46.599 --> 00:56:52.599
like 20 chat turns and you have the user

00:56:50.400 --> 00:56:55.640
give a thumbs up or a thumbs down on the

00:56:52.599 --> 00:56:58.920
20th chat turn there you would assign a

00:56:55.640 --> 00:57:01.440
reward of um like let's say it gave a

00:56:58.920 --> 00:57:03.640
thumbs up there you would re assign a

00:57:01.440 --> 00:57:06.559
reward of one for the previous chat turn

00:57:03.640 --> 00:57:09.839
a reward of like 0.5 for the second to

00:57:06.559 --> 00:57:11.720
previous chat term a reward of 0.25 for

00:57:09.839 --> 00:57:14.319
the third to previous chat term to

00:57:11.720 --> 00:57:16.160
basically say yeah like the user is

00:57:14.319 --> 00:57:18.240
feeling good at the moment they gave the

00:57:16.160 --> 00:57:20.359
thumbs up and that's probably more

00:57:18.240 --> 00:57:23.400
likely due to the things that happened

00:57:20.359 --> 00:57:23.400
recently so

00:57:23.559 --> 00:57:28.119
yeah we have a

00:57:26.680 --> 00:57:32.280
like not

00:57:28.119 --> 00:57:34.160
learning so the reward model can be any

00:57:32.280 --> 00:57:35.839
of the methods that I talked about

00:57:34.160 --> 00:57:37.480
before so it can be human feedback

00:57:35.839 --> 00:57:39.000
directly like a thumbs up or a thumbs

00:57:37.480 --> 00:57:42.200
down it could also be from a reward

00:57:39.000 --> 00:57:44.599
model uh that was pre-trained you could

00:57:42.200 --> 00:57:47.680
also theoretically learn the reward

00:57:44.599 --> 00:57:52.720
model simultaneously but you'd have to

00:57:47.680 --> 00:57:55.200
simultaneously with the model itself um

00:57:52.720 --> 00:57:57.280
so yeah I'm going to talk a little bit

00:57:55.200 --> 00:58:00.359
about DP which kind of does that a

00:57:57.280 --> 00:58:01.720
little bit but um I I would basically

00:58:00.359 --> 00:58:03.160
say that wherever you're getting your

00:58:01.720 --> 00:58:06.280
reward is probably from one of the

00:58:03.160 --> 00:58:06.280
things I talked about earlier

00:58:06.359 --> 00:58:14.960
today cool any other

00:58:09.319 --> 00:58:17.720
questions okay um so that's the basic

00:58:14.960 --> 00:58:20.640
the basic idea the very simplest thing

00:58:17.720 --> 00:58:23.359
that you can do is you can just sample

00:58:20.640 --> 00:58:26.079
um optimize the subjective function this

00:58:23.359 --> 00:58:28.359
is dead easy you it's not hard to imp

00:58:26.079 --> 00:58:30.799
imp it all as long as you have some

00:58:28.359 --> 00:58:32.760
source of reward signal um but the

00:58:30.799 --> 00:58:35.559
problem is uh reinforcement learning can

00:58:32.760 --> 00:58:38.599
be very unstable and it's hard to get it

00:58:35.559 --> 00:58:40.160
to uh you know work properly if you uh

00:58:38.599 --> 00:58:42.400
don't do some additional tricks so I'd

00:58:40.160 --> 00:58:45.720
like to talk about this

00:58:42.400 --> 00:58:45.720
next oh yeah

00:58:48.880 --> 00:58:51.880
sir

00:58:55.039 --> 00:58:58.039
yeah

00:59:03.280 --> 00:59:08.960
yeah the typical the typical way is you

00:59:05.440 --> 00:59:12.960
just have an exponential decay um so you

00:59:08.960 --> 00:59:16.200
you multiply each time by what 0.5 0. or

00:59:12.960 --> 00:59:19.400
something like that

00:59:16.200 --> 00:59:19.400
um from

00:59:20.319 --> 00:59:27.720
A6 um cool okay

00:59:25.039 --> 00:59:30.720
so

00:59:27.720 --> 00:59:33.319
and that's one option and sorry just to

00:59:30.720 --> 00:59:35.760
clarify the most common option nowadays

00:59:33.319 --> 00:59:37.920
um at least from the point of view of

00:59:35.760 --> 00:59:39.839
models is not to Decay it at all and

00:59:37.920 --> 00:59:43.880
just assign the same amount for each

00:59:39.839 --> 00:59:45.319
token um I'm not actually 100% sure what

00:59:43.880 --> 00:59:47.319
people are doing with respect to like

00:59:45.319 --> 00:59:49.280
long chat things I think probably

00:59:47.319 --> 00:59:51.720
they're only assigning it to the current

00:59:49.280 --> 00:59:54.240
like utterance and then not optimizing

00:59:51.720 --> 00:59:57.240
the previous utterances so like if they

00:59:54.240 --> 00:59:59.039
get a thumbs up or thumbs down signal um

00:59:57.240 --> 01:00:00.720
then they they would assign an

00:59:59.039 --> 01:00:02.440
equivalent reward for all of the tokens

01:00:00.720 --> 01:00:04.640
and the current utterance and zero

01:00:02.440 --> 01:00:06.119
reward for the previous ones but I'm not

01:00:04.640 --> 01:00:08.480
100% sure about that there might be

01:00:06.119 --> 01:00:11.200
other methods that people are

01:00:08.480 --> 01:00:13.960
using um

01:00:11.200 --> 01:00:16.680
cool so uh stabilizing reinforcement

01:00:13.960 --> 01:00:18.520
learning so um stabilizing reinforcement

01:00:16.680 --> 01:00:21.839
learning there's a lot of reasons why

01:00:18.520 --> 01:00:23.880
it's unstable um the first reason is

01:00:21.839 --> 01:00:27.200
you're sampling an individual output and

01:00:23.880 --> 01:00:30.160
calculating the um uh calculating based

01:00:27.200 --> 01:00:32.039
on the S individual sampled output and

01:00:30.160 --> 01:00:33.440
then there's an Infinity of other

01:00:32.039 --> 01:00:36.480
outputs that you could be optimizing

01:00:33.440 --> 01:00:39.119
over for mle this is not a problem

01:00:36.480 --> 01:00:41.319
because for mle you're always

01:00:39.119 --> 01:00:45.359
contrasting the gold standard output to

01:00:41.319 --> 01:00:46.599
all of the other outputs in the space um

01:00:45.359 --> 01:00:48.280
and you're saying I want to upweight the

01:00:46.599 --> 01:00:51.200
gold standard output and down we all of

01:00:48.280 --> 01:00:53.039
the other ones but for reinforcement

01:00:51.200 --> 01:00:54.760
learning you only have a single sampled

01:00:53.039 --> 01:00:57.520
output that output might be wrong and

01:00:54.760 --> 01:00:59.359
that's a source of inst ility this is

01:00:57.520 --> 01:01:02.079
particularly a problem when using bigger

01:00:59.359 --> 01:01:05.960
output spaces like all of the in the

01:01:02.079 --> 01:01:07.920
vocabul another problem is uh anytime

01:01:05.960 --> 01:01:11.599
you start using negative

01:01:07.920 --> 01:01:15.160
rewards um because if you start using

01:01:11.599 --> 01:01:17.559
negative rewards those rewards will be

01:01:15.160 --> 01:01:19.520
downweighting the probability of a

01:01:17.559 --> 01:01:20.680
particular output sequence and that

01:01:19.520 --> 01:01:22.440
might be a good idea maybe you're

01:01:20.680 --> 01:01:24.319
getting a toxic output or something like

01:01:22.440 --> 01:01:25.960
that and you want to down it but at the

01:01:24.319 --> 01:01:28.280
same time in addition to that toxic

01:01:25.960 --> 01:01:30.000
output there's like you know a

01:01:28.280 --> 01:01:31.599
combinatorial number of completely

01:01:30.000 --> 01:01:33.880
nonsense outputs that aren't even

01:01:31.599 --> 01:01:36.599
English and so basically you can start

01:01:33.880 --> 01:01:38.920
diverge from the N starting start to

01:01:36.599 --> 01:01:40.799
diverge from the natural like language

01:01:38.920 --> 01:01:44.720
modeling distribution that you have

01:01:40.799 --> 01:01:49.079
before so this is a big uh a big

01:01:44.720 --> 01:01:51.880
problem so a number of uh strategies can

01:01:49.079 --> 01:01:53.880
be used to stabilize the first one is

01:01:51.880 --> 01:01:55.480
this is completely obvious right now and

01:01:53.880 --> 01:01:57.240
nobody in their right mind would avoid

01:01:55.480 --> 01:02:00.119
doing this but the first one is

01:01:57.240 --> 01:02:02.839
pre-training with mle and so you start

01:02:00.119 --> 01:02:04.920
with a pre-trained model um and then

01:02:02.839 --> 01:02:09.359
switch over to RL after you finished

01:02:04.920 --> 01:02:11.520
pre-training the model um and so

01:02:09.359 --> 01:02:13.279
this makes a lot of sense if you're

01:02:11.520 --> 01:02:14.960
training a language model which I assume

01:02:13.279 --> 01:02:17.039
that almost everybody in this class is

01:02:14.960 --> 01:02:20.279
going to be doing but it does only work

01:02:17.039 --> 01:02:22.720
in scenarios where you can run mle and

01:02:20.279 --> 01:02:24.359
so it doesn't work if you're predicting

01:02:22.720 --> 01:02:27.240
like latent variables that aren't

01:02:24.359 --> 01:02:28.760
included in the original space

01:02:27.240 --> 01:02:31.960
um it

01:02:28.760 --> 01:02:34.279
also doesn't work in a setting where

01:02:31.960 --> 01:02:36.640
like you want to learn a

01:02:34.279 --> 01:02:40.799
chatbot you want to learn a chatbot for

01:02:36.640 --> 01:02:44.200
customer service for a

01:02:40.799 --> 01:02:48.039
company that

01:02:44.200 --> 01:02:49.960
has like for example a product catalog

01:02:48.039 --> 01:02:53.559
that the language model has never seen

01:02:49.960 --> 01:02:56.000
before and so if the language model has

01:02:53.559 --> 01:02:57.359
no information about the product catalog

01:02:56.000 --> 01:02:59.920
whatsoever you don't provide it through

01:02:57.359 --> 01:03:02.440
rag or something like that it's going to

01:02:59.920 --> 01:03:04.039
have to explore infinitely or not

01:03:02.440 --> 01:03:05.599
infinitely but it's going to have to

01:03:04.039 --> 01:03:08.359
explore too large of a space and you're

01:03:05.599 --> 01:03:10.000
never going to converge with um with

01:03:08.359 --> 01:03:12.359
your language modeling objectives so you

01:03:10.000 --> 01:03:15.000
need to basically be able to create at

01:03:12.359 --> 01:03:16.079
least some supervised training data to

01:03:15.000 --> 01:03:19.279
train with

01:03:16.079 --> 01:03:20.720
mle um but assuming you can do that I'm

01:03:19.279 --> 01:03:22.920
assuming that almost everybody is going

01:03:20.720 --> 01:03:26.400
to do some sort of pre-training with

01:03:22.920 --> 01:03:27.880
ML um The Next Step that people use uh

01:03:26.400 --> 01:03:30.520
in reinforcement learning that's really

01:03:27.880 --> 01:03:34.319
important to stabilize is regularization

01:03:30.520 --> 01:03:35.880
to an existing model and you have an

01:03:34.319 --> 01:03:39.039
existing model and you want to prevent

01:03:35.880 --> 01:03:40.559
it from getting too far away and the

01:03:39.039 --> 01:03:42.279
reason why you want to do this is like

01:03:40.559 --> 01:03:45.720
let's say you start assigning a negative

01:03:42.279 --> 01:03:47.440
reward to toxic utterances for example

01:03:45.720 --> 01:03:49.200
if your model stops being a language

01:03:47.440 --> 01:03:51.920
model whatsoever that's a bad idea so

01:03:49.200 --> 01:03:53.400
you want to keep it as a language model

01:03:51.920 --> 01:03:55.599
keep it close enough to still being a

01:03:53.400 --> 01:03:57.559
competent language model while you know

01:03:55.599 --> 01:03:59.599
like removing the toxic

01:03:57.559 --> 01:04:03.039
utterances so there's a number of

01:03:59.599 --> 01:04:05.680
methods that people use to do this um uh

01:04:03.039 --> 01:04:08.359
the most prominent ones are kale

01:04:05.680 --> 01:04:10.279
regularization uh well so the the first

01:04:08.359 --> 01:04:13.119
most prominent one is K regularization

01:04:10.279 --> 01:04:15.839
and the way this works is basically in

01:04:13.119 --> 01:04:19.400
addition you add you have two

01:04:15.839 --> 01:04:22.279
terms the first term is a term that

01:04:19.400 --> 01:04:25.760
improves your reward so you have your

01:04:22.279 --> 01:04:28.039
old model where your old model is

01:04:25.760 --> 01:04:31.279
creating a

01:04:28.039 --> 01:04:32.440
probability uh it has a probability here

01:04:31.279 --> 01:04:34.960
and then you have the probability

01:04:32.440 --> 01:04:38.160
assigned by your new model and then you

01:04:34.960 --> 01:04:41.200
have your reward signal here and so this

01:04:38.160 --> 01:04:43.599
is basically improving the log odds or

01:04:41.200 --> 01:04:46.960
improving the odds of getting a good

01:04:43.599 --> 01:04:49.720
reward for high reward

01:04:46.960 --> 01:04:52.920
sequences separately from this you have

01:04:49.720 --> 01:04:55.920
this K regularization term and this K

01:04:52.920 --> 01:04:58.119
regularization term is keeping the

01:04:55.920 --> 01:05:00.279
scores of or it's keeping the

01:04:58.119 --> 01:05:02.400
probability distribution of your new

01:05:00.279 --> 01:05:03.960
model similar to the probability

01:05:02.400 --> 01:05:09.200
distribution of your old

01:05:03.960 --> 01:05:11.359
model and this beta parameter basically

01:05:09.200 --> 01:05:15.240
you can increase it or decrease it based

01:05:11.359 --> 01:05:18.400
on how similar you want to keep the um

01:05:15.240 --> 01:05:18.400
how similar you want to keep the

01:05:20.720 --> 01:05:24.640
model another method that people use is

01:05:23.160 --> 01:05:29.279
something called proximal policy

01:05:24.640 --> 01:05:30.920
optimization or or Po and this is a

01:05:29.279 --> 01:05:33.920
method that is based on

01:05:30.920 --> 01:05:38.160
clipping uh the

01:05:33.920 --> 01:05:40.920
outputs and we Define uh this ratio

01:05:38.160 --> 01:05:43.880
here so this ratio is equivalent to this

01:05:40.920 --> 01:05:46.160
here so it's basically um kind of the

01:05:43.880 --> 01:05:47.839
amount that you're learning or the

01:05:46.160 --> 01:05:51.720
amount that the new model up weights

01:05:47.839 --> 01:05:54.039
High reward sequences and so here we

01:05:51.720 --> 01:05:58.200
have the same thing that we had

01:05:54.039 --> 01:06:01.200
above so it it looks like this but over

01:05:58.200 --> 01:06:03.720
here we have a clipped version of this

01:06:01.200 --> 01:06:07.000
where essentially what we do is we

01:06:03.720 --> 01:06:07.000
clip this

01:06:21.119 --> 01:06:27.880
ratio this ratio to be within uh a

01:06:24.720 --> 01:06:32.160
certain range of the original ratio and

01:06:27.880 --> 01:06:37.880
what this is doing is this is

01:06:32.160 --> 01:06:41.400
essentially forcing the model to um not

01:06:37.880 --> 01:06:44.000
reward large jumps in the space um

01:06:41.400 --> 01:06:47.559
because if you take the

01:06:44.000 --> 01:06:49.160
minimum and actually I'm I'm sorry I

01:06:47.559 --> 01:06:50.720
just realized I I might have done

01:06:49.160 --> 01:06:52.520
something confusing here because this is

01:06:50.720 --> 01:06:53.960
actually higher as better so this isn't

01:06:52.520 --> 01:06:56.079
really a loss function this is something

01:06:53.960 --> 01:06:57.680
you're attempting to maximize so

01:06:56.079 --> 01:06:59.839
in contrast to all of the other things I

01:06:57.680 --> 01:07:01.680
was talking about before um this is

01:06:59.839 --> 01:07:04.400
something where higher is better instead

01:07:01.680 --> 01:07:07.599
of lower is better but anyway basically

01:07:04.400 --> 01:07:09.599
by taking the minimum of this you're

01:07:07.599 --> 01:07:11.960
encouraging the model

01:07:09.599 --> 01:07:16.279
to

01:07:11.960 --> 01:07:18.559
uh keep examining the space where you

01:07:16.279 --> 01:07:20.799
don't diverge much from the original

01:07:18.559 --> 01:07:22.920
model and if the space where the

01:07:20.799 --> 01:07:25.240
original model was in is better than the

01:07:22.920 --> 01:07:27.440
new space that your model has moved into

01:07:25.240 --> 01:07:30.920
you move back towards the original model

01:07:27.440 --> 01:07:33.000
so basically like if you had um if you

01:07:30.920 --> 01:07:34.960
learned a model if you started learning

01:07:33.000 --> 01:07:37.960
a model that looked like it was

01:07:34.960 --> 01:07:40.279
optimizing uh your your reward but then

01:07:37.960 --> 01:07:43.119
suddenly the model went off the rails

01:07:40.279 --> 01:07:45.000
and um it starts generating completely

01:07:43.119 --> 01:07:47.319
nonsense outputs that get really bad

01:07:45.000 --> 01:07:49.119
reward this will push it back towards

01:07:47.319 --> 01:07:50.920
the original policy and that's the basic

01:07:49.119 --> 01:07:54.279
idea behind

01:07:50.920 --> 01:07:57.640
P um in terms of what I see people using

01:07:54.279 --> 01:07:59.799
um po was like really really popular for

01:07:57.640 --> 01:08:01.880
a while but I've started to see people

01:07:59.799 --> 01:08:04.799
use alternative strategies that use K

01:08:01.880 --> 01:08:06.880
regularization so I don't I don't think

01:08:04.799 --> 01:08:08.520
either one of them is like particularly

01:08:06.880 --> 01:08:10.039
more popular than any of the others and

01:08:08.520 --> 01:08:13.720
this one's a little bit simpler

01:08:10.039 --> 01:08:13.720
conceptually so I like the the

01:08:14.880 --> 01:08:19.279
one cool um any questions about

01:08:20.359 --> 01:08:26.759
this okay um and actually one thing I

01:08:24.640 --> 01:08:29.679
should mention is um all of these things

01:08:26.759 --> 01:08:32.120
are implemented uh in you know whatever

01:08:29.679 --> 01:08:33.759
libraries you use like hugging face TRL

01:08:32.120 --> 01:08:35.679
Transformer reinforcement learning as an

01:08:33.759 --> 01:08:37.040
example Library all of these methods are

01:08:35.679 --> 01:08:38.400
implemented there so if you actually

01:08:37.040 --> 01:08:40.600
want to use these in practice that's

01:08:38.400 --> 01:08:40.600
good

01:08:40.839 --> 01:08:46.359
place the next thing is adding a

01:08:42.920 --> 01:08:48.679
Baseline and so the basic idea is that

01:08:46.359 --> 01:08:52.199
you have ex expectations about your

01:08:48.679 --> 01:08:54.640
reward for a particular sentence and um

01:08:52.199 --> 01:08:56.560
like let's say we wanted to uh translate

01:08:54.640 --> 01:08:58.400
a sentence and we have uh something like

01:08:56.560 --> 01:09:01.279
this is an easy sentence and buffalo

01:08:58.400 --> 01:09:02.920
buffalo buffalo which is a harder

01:09:01.279 --> 01:09:07.799
sentence to

01:09:02.920 --> 01:09:09.679
translate and so we have a reward um if

01:09:07.799 --> 01:09:11.759
if you're not familiar with this example

01:09:09.679 --> 01:09:13.480
you can search on Wikipedia for buffalo

01:09:11.759 --> 01:09:16.759
buffalo buffalo and you'll you'll find

01:09:13.480 --> 01:09:19.520
out what I'm talking about um but uh

01:09:16.759 --> 01:09:21.440
there's a reward uh and let's say you

01:09:19.520 --> 01:09:24.359
got a reward of 0.8 for the first one

01:09:21.440 --> 01:09:29.679
and a reward of 0.3 for the second

01:09:24.359 --> 01:09:31.679
one but the problem is if um the first

01:09:29.679 --> 01:09:33.640
one actually is really easy and the

01:09:31.679 --> 01:09:36.120
second one is really hard getting a

01:09:33.640 --> 01:09:37.799
reward of 0.8 for the second one for

01:09:36.120 --> 01:09:40.080
like a translation or something is

01:09:37.799 --> 01:09:41.120
actually bad right and a reward of 0.3

01:09:40.080 --> 01:09:45.239
is good because you're moving in the

01:09:41.120 --> 01:09:49.359
right direction and so you basically um

01:09:45.239 --> 01:09:52.239
you have uh the Baseline uh minus reward

01:09:49.359 --> 01:09:54.960
or sorry reward minus Baseline and this

01:09:52.239 --> 01:09:56.520
would give you a negative value for this

01:09:54.960 --> 01:09:59.320
first one a positive value for the

01:09:56.520 --> 01:10:01.360
second one and so the basic idea is can

01:09:59.320 --> 01:10:04.400
we predict a priori how difficult this

01:10:01.360 --> 01:10:05.440
example is and then uh adjust our reward

01:10:04.400 --> 01:10:08.360
based on

01:10:05.440 --> 01:10:10.960
that and

01:10:08.360 --> 01:10:13.679
so that's the basic idea you just have

01:10:10.960 --> 01:10:15.560
kind of like a baseline model um you

01:10:13.679 --> 01:10:19.320
have a baseline model that predicts this

01:10:15.560 --> 01:10:19.320
and uh you adjust uh

01:10:19.760 --> 01:10:25.000
appropriately um there's two major ways

01:10:22.719 --> 01:10:27.600
you can do this the first one um the

01:10:25.000 --> 01:10:29.800
Baseline doesn't need to be anything um

01:10:27.600 --> 01:10:32.960
the only hope is that it decreases the

01:10:29.800 --> 01:10:35.960
variance in your reward uh and makes

01:10:32.960 --> 01:10:38.239
learning more stable um there's two

01:10:35.960 --> 01:10:40.159
options that I see done pretty widely

01:10:38.239 --> 01:10:43.000
the first one is predicting the final

01:10:40.159 --> 01:10:47.360
reward um predicting the final reward

01:10:43.000 --> 01:10:50.960
using a model that doesn't look at

01:10:47.360 --> 01:10:53.400
all at the answer that you provided it

01:10:50.960 --> 01:10:55.880
only looks at the input or it only looks

01:10:53.400 --> 01:10:58.840
at the intermediate States of uh you

01:10:55.880 --> 01:11:00.480
know a model or something and so at the

01:10:58.840 --> 01:11:03.280
sentence level you can have one Baseline

01:11:00.480 --> 01:11:04.719
per sentence um you can also do it at

01:11:03.280 --> 01:11:10.560
each decoder

01:11:04.719 --> 01:11:11.640
State and this is uh basically you can

01:11:10.560 --> 01:11:13.040
do this anytime you're doing

01:11:11.640 --> 01:11:15.199
reinforcement learning by just training

01:11:13.040 --> 01:11:18.199
a regression model that does this for

01:11:15.199 --> 01:11:19.679
you based on the rewards you get the

01:11:18.199 --> 01:11:21.040
important thing is the Baseline is not

01:11:19.679 --> 01:11:22.640
allowed to use any of your actual

01:11:21.040 --> 01:11:25.679
predictions because once you start using

01:11:22.640 --> 01:11:26.640
the predictions then um your uh it's not

01:11:25.679 --> 01:11:28.679
a

01:11:26.640 --> 01:11:30.840
baseline another option which is

01:11:28.679 --> 01:11:33.440
relatively easy to implement but can

01:11:30.840 --> 01:11:36.320
still be effective is you calculate the

01:11:33.440 --> 01:11:38.719
mean of the rewards in a batch and so if

01:11:36.320 --> 01:11:40.880
you have a big batch of data and your

01:11:38.719 --> 01:11:44.440
average reward in the batch is like

01:11:40.880 --> 01:11:46.480
0.4 uh then you just subtract that 0.4

01:11:44.440 --> 01:11:50.080
uh and calculate your reward based on

01:11:46.480 --> 01:11:50.080
that so that's another option that can

01:11:51.800 --> 01:11:57.800
use

01:11:53.639 --> 01:12:00.000
um a kind of extreme example of this uh

01:11:57.800 --> 01:12:01.199
of creating a baseline is contrasting

01:12:00.000 --> 01:12:03.639
pairwise

01:12:01.199 --> 01:12:05.880
examples um or

01:12:03.639 --> 01:12:08.280
contrasting different outputs for the

01:12:05.880 --> 01:12:12.040
same input

01:12:08.280 --> 01:12:13.920
and you can easily learn uh directly

01:12:12.040 --> 01:12:16.239
from pairwise Human

01:12:13.920 --> 01:12:18.199
preferences uh which can provide more

01:12:16.239 --> 01:12:20.760
stability because you know one is better

01:12:18.199 --> 01:12:23.880
than the other so you essentially can be

01:12:20.760 --> 01:12:26.199
sure that uh you're upweighting a better

01:12:23.880 --> 01:12:29.560
one and down weting a worse one

01:12:26.199 --> 01:12:31.400
um this is the idea behind DPO which is

01:12:29.560 --> 01:12:33.719
a recently pretty popular model but

01:12:31.400 --> 01:12:36.800
there's also other previous methods that

01:12:33.719 --> 01:12:40.199
did similar things and the way DPO works

01:12:36.800 --> 01:12:45.040
is it basically calculates this ratio of

01:12:40.199 --> 01:12:49.280
uh the probability of the new uh the new

01:12:45.040 --> 01:12:51.639
model to the old model but it UPS this

01:12:49.280 --> 01:12:53.639
probability for a good output and it

01:12:51.639 --> 01:12:56.280
downweights this probability for a bad

01:12:53.639 --> 01:12:57.679
output and so

01:12:56.280 --> 01:13:00.120
here we have our better outputs over

01:12:57.679 --> 01:13:02.040
here here we have our worse outputs and

01:13:00.120 --> 01:13:03.600
you just it's basically learning to

01:13:02.040 --> 01:13:05.639
upate the probability and downweight

01:13:03.600 --> 01:13:09.320
probability

01:13:05.639 --> 01:13:09.320
accordingly so

01:13:09.360 --> 01:13:15.040
um you can notice that DPO is very

01:13:12.280 --> 01:13:18.040
similar to PO um and that it's learning

01:13:15.040 --> 01:13:19.679
uh it's using these ratios but the

01:13:18.040 --> 01:13:21.520
disadvantage of this is you obviously

01:13:19.679 --> 01:13:23.120
require pairwise judgments and you can't

01:13:21.520 --> 01:13:26.120
learn a model if you don't have these

01:13:23.120 --> 01:13:28.080
pawise judgments so

01:13:26.120 --> 01:13:30.760
the

01:13:28.080 --> 01:13:33.159
beta yeah so the beta term is is

01:13:30.760 --> 01:13:35.840
basically a normalization term it's a

01:13:33.159 --> 01:13:39.960
hyper parameter um

01:13:35.840 --> 01:13:41.840
for DPO sorry I read the paper right

01:13:39.960 --> 01:13:43.639
when it came out and I don't remember if

01:13:41.840 --> 01:13:45.600
it's a direct derivation from the K

01:13:43.639 --> 01:13:47.960
Divergence term or not but I think it

01:13:45.600 --> 01:13:49.800
might be um I'd have to go back and look

01:13:47.960 --> 01:13:50.480
at the look at the paper but basically

01:13:49.800 --> 01:13:53.600
the

01:13:50.480 --> 01:13:56.760
more the larger this is the larger

01:13:53.600 --> 01:13:59.320
gradient steps you'll be

01:13:56.760 --> 01:14:00.639
it also um like you'll notice there

01:13:59.320 --> 01:14:03.400
sorry I didn't mention this but you'll

01:14:00.639 --> 01:14:06.120
notice there's a sigmoid term here so

01:14:03.400 --> 01:14:09.000
the the

01:14:06.120 --> 01:14:10.080
beta the larger you increase the beta

01:14:09.000 --> 01:14:13.239
the

01:14:10.080 --> 01:14:16.600
more small differences in these

01:14:13.239 --> 01:14:18.719
values like it basically like stretches

01:14:16.600 --> 01:14:22.280
or shrinks the sigmoid with respect to

01:14:18.719 --> 01:14:24.120
how beak the it is so it will um it will

01:14:22.280 --> 01:14:25.800
affect how much like small differences

01:14:24.120 --> 01:14:27.960
in this will affect

01:14:25.800 --> 01:14:30.120
but I I think this was derived from the

01:14:27.960 --> 01:14:31.760
K regularization term that we had

01:14:30.120 --> 01:14:34.400
previously in

01:14:31.760 --> 01:14:35.800
um in this slide here but I have to go

01:14:34.400 --> 01:14:40.520
back and double check unless somebody

01:14:35.800 --> 01:14:43.239
knows it is okay good yeah

01:14:40.520 --> 01:14:45.000
so I don't want to say wrong things but

01:14:43.239 --> 01:14:48.239
I also don't want

01:14:45.000 --> 01:14:50.920
to okay cool um and so then increasing

01:14:48.239 --> 01:14:55.080
batch size

01:14:50.920 --> 01:14:57.360
um because each uh another thing is um

01:14:55.080 --> 01:14:58.440
kind of NE necessarily reinforcement

01:14:57.360 --> 01:14:59.920
learning is going to have higher

01:14:58.440 --> 01:15:01.400
variance and maximum likelihood

01:14:59.920 --> 01:15:04.199
estimation just because we're doing samp

01:15:01.400 --> 01:15:07.840
playing and other things like this and

01:15:04.199 --> 01:15:09.440
um so one very simple thing you can do

01:15:07.840 --> 01:15:11.280
is just increase the number of examples

01:15:09.440 --> 01:15:13.679
or rollouts that you do before an update

01:15:11.280 --> 01:15:15.800
to stabilize and so I I would definitely

01:15:13.679 --> 01:15:17.480
suggest that if you're seeing any

01:15:15.800 --> 01:15:18.679
stability after doing all of the tricks

01:15:17.480 --> 01:15:20.400
that I mentioned before that you

01:15:18.679 --> 01:15:23.040
increase your batch size and often that

01:15:20.400 --> 01:15:25.480
can just resolve your problems

01:15:23.040 --> 01:15:28.760
um another uh

01:15:25.480 --> 01:15:30.560
thing that people often do is um save

01:15:28.760 --> 01:15:32.040
many many previous rollouts because

01:15:30.560 --> 01:15:34.199
generally doing rollouts is more

01:15:32.040 --> 01:15:37.840
expensive doing rollouts and collecting

01:15:34.199 --> 01:15:39.560
rewards is more expensive and so um you

01:15:37.840 --> 01:15:42.360
can save the roll outs that you have

01:15:39.560 --> 01:15:43.840
done before and uh keep them around so

01:15:42.360 --> 01:15:46.600
you can update parameters with larger

01:15:43.840 --> 01:15:50.800
batches in a more efficient

01:15:46.600 --> 01:15:53.120
way cool so that's all I have uh I just

01:15:50.800 --> 01:15:54.400
realized we're exactly at time so uh I

01:15:53.120 --> 01:15:56.440
should finish up here but I'll be happy

01:15:54.400 --> 01:15:59.440
to take any

01:15:56.440 --> 01:15:59.440
for

01:16:01.679 --> 01:16:04.679
thanks