diff --git "a/CMU Advanced NLP 2024 (13) Debugging and Interpretation/transcript.vtt" "b/CMU Advanced NLP 2024 (13) Debugging and Interpretation/transcript.vtt"
new file mode 100644--- /dev/null
+++ "b/CMU Advanced NLP 2024 (13) Debugging and Interpretation/transcript.vtt"
@@ -0,0 +1,5482 @@
+WEBVTT
+
+00:00:00.919 --> 00:00:05.879
+so in my slides here I'm going to talk
+
+00:00:03.760 --> 00:00:10.040
+about debugging and understanding NLP
+
+00:00:05.879 --> 00:00:12.400
+models and this is how to tell uh when
+
+00:00:10.040 --> 00:00:14.759
+for example both your implementations
+
+00:00:12.400 --> 00:00:17.320
+are wrong and uh for example your
+
+00:00:14.759 --> 00:00:19.000
+underlying assumptions are wrong or your
+
+00:00:17.320 --> 00:00:21.240
+model is failing on particular segments
+
+00:00:19.000 --> 00:00:23.439
+of data or stuff like that so going to
+
+00:00:21.240 --> 00:00:26.160
+go through uh a variety of things that
+
+00:00:23.439 --> 00:00:29.000
+can go wrong with your experiments
+
+00:00:26.160 --> 00:00:31.679
+basically so a typical situation is
+
+00:00:29.000 --> 00:00:33.399
+you've implemented some NLP system you
+
+00:00:31.679 --> 00:00:35.840
+know based on neural networks of course
+
+00:00:33.399 --> 00:00:36.920
+because that's what we use nowadays um
+
+00:00:35.840 --> 00:00:40.000
+and you've looked at the code it
+
+00:00:36.920 --> 00:00:42.000
+basically looks okay um but it has low
+
+00:00:40.000 --> 00:00:44.559
+accuracy or it makes incomprehensible
+
+00:00:42.000 --> 00:00:45.680
+errors and you would like to uh fix
+
+00:00:44.559 --> 00:00:47.440
+these or you'd like to improve the
+
+00:00:45.680 --> 00:00:49.120
+accuracy or something like that and so
+
+00:00:47.440 --> 00:00:52.000
+what do I
+
+00:00:49.120 --> 00:00:53.680
+do and I think there's three dimensions
+
+00:00:52.000 --> 00:00:56.239
+of how you can understand your model and
+
+00:00:53.680 --> 00:00:57.960
+your Model Behavior um the first one is
+
+00:00:56.239 --> 00:01:00.199
+debugging the implementation so it's
+
+00:00:57.960 --> 00:01:03.760
+identifying problems that you have when
+
+00:01:00.199 --> 00:01:05.880
+you uh implemented something uh second
+
+00:01:03.760 --> 00:01:07.759
+thing is actionable evaluation so
+
+00:01:05.880 --> 00:01:09.799
+identifying typical error cases and how
+
+00:01:07.759 --> 00:01:11.840
+you what you can do to fix them and
+
+00:01:09.799 --> 00:01:13.720
+finally uh interpreting predictions or
+
+00:01:11.840 --> 00:01:18.080
+interpreting what's happening inside the
+
+00:01:13.720 --> 00:01:19.920
+model and uh this can maybe give you a
+
+00:01:18.080 --> 00:01:21.520
+deeper idea about what's happening in
+
+00:01:19.920 --> 00:01:22.720
+happening in individual cases and
+
+00:01:21.520 --> 00:01:25.240
+there's a lot of reasons why you might
+
+00:01:22.720 --> 00:01:27.920
+want to do that uh both like to make
+
+00:01:25.240 --> 00:01:30.280
+your models better and also for example
+
+00:01:27.920 --> 00:01:31.840
+if you want to be sure that your ition
+
+00:01:30.280 --> 00:01:34.840
+isn't doing something illegal like
+
+00:01:31.840 --> 00:01:36.439
+discriminating against people uh due to
+
+00:01:34.840 --> 00:01:38.680
+protected attributes or other things
+
+00:01:36.439 --> 00:01:41.399
+like that so um there's a number of
+
+00:01:38.680 --> 00:01:42.920
+reasons why you'd want to do that so I'm
+
+00:01:41.399 --> 00:01:44.399
+going to talk about the first two and
+
+00:01:42.920 --> 00:01:48.840
+Nishant is mainly going to talk about
+
+00:01:44.399 --> 00:01:52.000
+the second one so uh going right into
+
+00:01:48.840 --> 00:01:55.159
+it so in neural network models uh
+
+00:01:52.000 --> 00:01:58.880
+debugging is really important because
+
+00:01:55.159 --> 00:02:00.920
+they're opaque they're unpredictable and
+
+00:01:58.880 --> 00:02:03.119
+uh if you make little mistakes they can
+
+00:02:00.920 --> 00:02:05.439
+cause big problems with your
+
+00:02:03.119 --> 00:02:07.399
+output and another thing is that
+
+00:02:05.439 --> 00:02:09.640
+everything is a hyperparameter including
+
+00:02:07.399 --> 00:02:11.239
+your network size your model variations
+
+00:02:09.640 --> 00:02:14.440
+your batch size your strategy your
+
+00:02:11.239 --> 00:02:18.120
+Optimizer and your learning rate
+
+00:02:14.440 --> 00:02:19.560
+and finally unlike kind of more
+
+00:02:18.120 --> 00:02:21.200
+traditional machine learning methods
+
+00:02:19.560 --> 00:02:23.000
+like logistic progression or support
+
+00:02:21.200 --> 00:02:25.160
+Vector machines or something like that
+
+00:02:23.000 --> 00:02:27.879
+you might that you might have studied in
+
+00:02:25.160 --> 00:02:30.160
+your machine learning class um
+
+00:02:27.879 --> 00:02:32.599
+stochastic optimization has no guarantee
+
+00:02:30.160 --> 00:02:34.239
+about convergence um your loss might go
+
+00:02:32.599 --> 00:02:35.720
+down then it might go up and there might
+
+00:02:34.239 --> 00:02:38.120
+be absolutely nothing wrong with your
+
+00:02:35.720 --> 00:02:40.200
+training or it might be you know a
+
+00:02:38.120 --> 00:02:42.319
+serious problem so that's another issue
+
+00:02:40.200 --> 00:02:45.440
+you need to deal
+
+00:02:42.319 --> 00:02:48.800
+with so first I'd like to go into
+
+00:02:45.440 --> 00:02:51.400
+possible causes of problems with your
+
+00:02:48.800 --> 00:02:53.440
+implementation and I'm going to break
+
+00:02:51.400 --> 00:02:55.040
+them down into a typology and based on
+
+00:02:53.440 --> 00:02:57.040
+what part of the typology you're running
+
+00:02:55.040 --> 00:02:59.200
+into problems with you will need to fix
+
+00:02:57.040 --> 00:03:00.800
+them in different ways so your first
+
+00:02:59.200 --> 00:03:02.599
+goal when you're experiencing the
+
+00:03:00.800 --> 00:03:04.720
+problem is identifying why you're
+
+00:03:02.599 --> 00:03:06.400
+experiencing the problem uh because that
+
+00:03:04.720 --> 00:03:08.760
+will lead you to a
+
+00:03:06.400 --> 00:03:10.440
+solution so for training time problems
+
+00:03:08.760 --> 00:03:12.560
+there's a bunch of uh things that could
+
+00:03:10.440 --> 00:03:14.360
+be wrong uh the first is a lack of model
+
+00:03:12.560 --> 00:03:16.280
+capacity so your model is not able to
+
+00:03:14.360 --> 00:03:18.599
+model the phenomena that you want to
+
+00:03:16.280 --> 00:03:20.000
+model in the first place um you could
+
+00:03:18.599 --> 00:03:22.080
+have a poor training
+
+00:03:20.000 --> 00:03:24.920
+algorithm uh you could just have a bug
+
+00:03:22.080 --> 00:03:27.080
+in your code at training time another
+
+00:03:24.920 --> 00:03:29.319
+thing is uh test time problems and these
+
+00:03:27.080 --> 00:03:30.599
+can include a disconnect between what
+
+00:03:29.319 --> 00:03:33.040
+you're doing at training time and what
+
+00:03:30.599 --> 00:03:35.640
+you're testing at testing time uh
+
+00:03:33.040 --> 00:03:37.959
+failure of search
+
+00:03:35.640 --> 00:03:39.920
+algorithms and another thing you want to
+
+00:03:37.959 --> 00:03:41.360
+deal with is overfitting so you're
+
+00:03:39.920 --> 00:03:44.319
+actually doing well on the training set
+
+00:03:41.360 --> 00:03:48.360
+but you're doing poorly on the test
+
+00:03:44.319 --> 00:03:50.400
+Set uh finally you could have um optimiz
+
+00:03:48.360 --> 00:03:52.640
+a mismatch between the function you're
+
+00:03:50.400 --> 00:03:54.920
+optimizing at evaluation time and uh
+
+00:03:52.640 --> 00:03:56.519
+what you're actually evaluating sorry
+
+00:03:54.920 --> 00:03:58.079
+the fun the function that you're
+
+00:03:56.519 --> 00:04:01.079
+optimizing at training time and what
+
+00:03:58.079 --> 00:04:03.720
+you're actually evaluating at test time
+
+00:04:01.079 --> 00:04:05.280
+and my my best piece of advice for
+
+00:04:03.720 --> 00:04:07.959
+figuring out why things are going wrong
+
+00:04:05.280 --> 00:04:11.040
+is don't uh try to do all of them at
+
+00:04:07.959 --> 00:04:12.560
+once and rather uh start from the top
+
+00:04:11.040 --> 00:04:15.239
+and work it down because the ones at the
+
+00:04:12.560 --> 00:04:17.600
+top are often easier to uh diagnose and
+
+00:04:15.239 --> 00:04:20.680
+the ones at the
+
+00:04:17.600 --> 00:04:23.000
+bottom so looking at how you can debug
+
+00:04:20.680 --> 00:04:25.919
+systems at training time uh there's a
+
+00:04:23.000 --> 00:04:27.360
+number of ways you can do this uh but
+
+00:04:25.919 --> 00:04:30.039
+the most important thing for training
+
+00:04:27.360 --> 00:04:33.479
+time uh issues is looking at the loss
+
+00:04:30.039 --> 00:04:36.759
+function calculated on the training set
+
+00:04:33.479 --> 00:04:38.960
+and what I mean by this is don't look uh
+
+00:04:36.759 --> 00:04:41.240
+we talked about how we can't optimize
+
+00:04:38.960 --> 00:04:45.039
+error or accuracy easily so instead we
+
+00:04:41.240 --> 00:04:47.120
+optimize likelihood um and so you might
+
+00:04:45.039 --> 00:04:49.080
+want to look at accuracy to see whether
+
+00:04:47.120 --> 00:04:50.759
+your model is working well but I would
+
+00:04:49.080 --> 00:04:53.039
+urge you first to look at your
+
+00:04:50.759 --> 00:04:55.080
+likelihood or your loss function on the
+
+00:04:53.039 --> 00:04:57.000
+training set instead of your accuracy on
+
+00:04:55.080 --> 00:04:58.479
+the test set for example to diagnose
+
+00:04:57.000 --> 00:05:00.600
+these variety of
+
+00:04:58.479 --> 00:05:02.919
+problems and the sorts of things you
+
+00:05:00.600 --> 00:05:05.840
+want to look at are um is the loss
+
+00:05:02.919 --> 00:05:10.639
+function going down so is it you know
+
+00:05:05.840 --> 00:05:14.199
+converging into a good place
+
+00:05:10.639 --> 00:05:16.280
+um in general if this is your your
+
+00:05:14.199 --> 00:05:18.600
+loss um the first thing you should know
+
+00:05:16.280 --> 00:05:20.440
+is like what is a good loss uh in most
+
+00:05:18.600 --> 00:05:22.280
+cases a good loss is zero like log
+
+00:05:20.440 --> 00:05:26.280
+likelihood the best loss you can achieve
+
+00:05:22.280 --> 00:05:28.639
+is zero so you have zero down here um
+
+00:05:26.280 --> 00:05:31.639
+something
+
+00:05:28.639 --> 00:05:31.639
+like
+
+00:05:31.919 --> 00:05:36.680
+this is uh essentially a good loss
+
+00:05:38.080 --> 00:05:43.120
+function something like that uh
+
+00:05:41.360 --> 00:05:45.120
+especially if this is a relatively High
+
+00:05:43.120 --> 00:05:47.759
+number is usually a bad loss
+
+00:05:45.120 --> 00:05:50.319
+function
+
+00:05:47.759 --> 00:05:52.680
+um something like that on your training
+
+00:05:50.319 --> 00:05:54.240
+set is a very bad loss function uh
+
+00:05:52.680 --> 00:05:55.840
+something something is going seriously
+
+00:05:54.240 --> 00:05:57.960
+wrong if you see this on your Dev set
+
+00:05:55.840 --> 00:05:59.800
+that could be or your test set that
+
+00:05:57.960 --> 00:06:01.199
+could be uh overfitting but but if
+
+00:05:59.800 --> 00:06:03.440
+you're seeing that on your training set
+
+00:06:01.199 --> 00:06:05.759
+that's usually symptomatic of a problem
+
+00:06:03.440 --> 00:06:09.160
+so uh these are uh things that you
+
+00:06:05.759 --> 00:06:10.960
+should be uh knowing um is it going down
+
+00:06:09.160 --> 00:06:13.520
+basically to zero if you run training
+
+00:06:10.960 --> 00:06:16.000
+long enough um for many epochs over your
+
+00:06:13.520 --> 00:06:17.479
+training data so if it's not going down
+
+00:06:16.000 --> 00:06:20.599
+to zero and it's sticking up here then
+
+00:06:17.479 --> 00:06:20.599
+that's also an
+
+00:06:21.120 --> 00:06:25.759
+issue and um if it's not going down to
+
+00:06:23.840 --> 00:06:27.919
+close to zero on whatever training set
+
+00:06:25.759 --> 00:06:30.199
+you're training on um let's say you make
+
+00:06:27.919 --> 00:06:31.840
+your training set extremely small
+
+00:06:30.199 --> 00:06:33.319
+uh at least in that case it should go
+
+00:06:31.840 --> 00:06:34.960
+down to zero otherwise you might have a
+
+00:06:33.319 --> 00:06:37.199
+serious problem in your
+
+00:06:34.960 --> 00:06:39.240
+implementation so these are good things
+
+00:06:37.199 --> 00:06:41.960
+to check first when you're training a
+
+00:06:39.240 --> 00:06:45.199
+model um and there's a number of reasons
+
+00:06:41.960 --> 00:06:47.759
+why this might not be helping or why
+
+00:06:45.199 --> 00:06:50.880
+this might not be happening so um your
+
+00:06:47.759 --> 00:06:53.120
+Mo model might be too weak and so in
+
+00:06:50.880 --> 00:06:55.440
+general larger models tend to perform
+
+00:06:53.120 --> 00:06:58.000
+better uh especially if you're using a
+
+00:06:55.440 --> 00:06:59.800
+pre-trained model and um this is just an
+
+00:06:58.000 --> 00:07:03.800
+example from the T5 paper where they
+
+00:06:59.800 --> 00:07:06.680
+scale up the T5 model um from a
+
+00:07:03.800 --> 00:07:09.319
+relatively small model to what at the
+
+00:07:06.680 --> 00:07:12.199
+time was a very large model of 11
+
+00:07:09.319 --> 00:07:14.360
+billion parameters now this is you know
+
+00:07:12.199 --> 00:07:17.479
+a moderately sized model or maybe even
+
+00:07:14.360 --> 00:07:20.879
+small model by some standards but anyway
+
+00:07:17.479 --> 00:07:23.800
+you can see that it uh in continues to
+
+00:07:20.879 --> 00:07:26.479
+increase one really interesting
+
+00:07:23.800 --> 00:07:30.080
+phenomenon is uh that actually larger
+
+00:07:26.479 --> 00:07:33.879
+models can learn faster or at least with
+
+00:07:30.080 --> 00:07:36.680
+fewer steps than uh smaller
+
+00:07:33.879 --> 00:07:40.199
+models and so this
+
+00:07:36.680 --> 00:07:42.240
+is an interesting example this paper uh
+
+00:07:40.199 --> 00:07:43.919
+on neural scaling was it's a very
+
+00:07:42.240 --> 00:07:48.000
+influential paper but basically what
+
+00:07:43.919 --> 00:07:51.000
+they show is the darker purple ones are
+
+00:07:48.000 --> 00:07:54.599
+smaller models the yellow ones are
+
+00:07:51.000 --> 00:07:57.159
+bigger models and what you can see here
+
+00:07:54.599 --> 00:07:59.639
+is the purple model and on the left side
+
+00:07:57.159 --> 00:08:02.120
+they have the number of tokens processed
+
+00:07:59.639 --> 00:08:05.759
+the right side they have the number of
+
+00:08:02.120 --> 00:08:08.159
+uh compute or the amount of compute um
+
+00:08:05.759 --> 00:08:10.080
+and so what you can see is if you just
+
+00:08:08.159 --> 00:08:12.240
+look at the number of tokens processed
+
+00:08:10.080 --> 00:08:14.280
+the larger the model the faster it
+
+00:08:12.240 --> 00:08:17.720
+converges which
+
+00:08:14.280 --> 00:08:21.400
+is maybe a little bit surprising maybe a
+
+00:08:17.720 --> 00:08:22.680
+little bit you or maybe uh like some
+
+00:08:21.400 --> 00:08:24.879
+people have the intuition that this
+
+00:08:22.680 --> 00:08:26.440
+should be the case but when I first saw
+
+00:08:24.879 --> 00:08:27.759
+this I found it a little bit surprising
+
+00:08:26.440 --> 00:08:29.000
+because I thought it would be so large
+
+00:08:27.759 --> 00:08:29.960
+and noisy that the model would have
+
+00:08:29.000 --> 00:08:32.320
+trouble fit
+
+00:08:29.960 --> 00:08:34.200
+you know fitting the data as quickly but
+
+00:08:32.320 --> 00:08:36.200
+there's actually a good reason for this
+
+00:08:34.200 --> 00:08:37.240
+does anyone have a guess about why this
+
+00:08:36.200 --> 00:08:39.719
+is
+
+00:08:37.240 --> 00:08:41.240
+thee we've talked a little bit about the
+
+00:08:39.719 --> 00:08:44.120
+underlying phenomena for this in
+
+00:08:41.240 --> 00:08:48.360
+previous classes so you might be able to
+
+00:08:44.120 --> 00:08:48.360
+think back to some of the things you
+
+00:08:50.480 --> 00:08:56.040
+yeah yeah so um just to repeat there's a
+
+00:08:54.160 --> 00:08:57.720
+lot of different parameters so it can
+
+00:08:56.040 --> 00:08:59.880
+try to converge along a lot of different
+
+00:08:57.720 --> 00:09:01.920
+dimensions so if we think back to the
+
+00:08:59.880 --> 00:09:04.079
+like model pruning class and other stuff
+
+00:09:01.920 --> 00:09:06.640
+like that um part of the reason why we
+
+00:09:04.079 --> 00:09:08.000
+can prune large models so efficiently is
+
+00:09:06.640 --> 00:09:10.200
+because only like a small number of the
+
+00:09:08.000 --> 00:09:12.440
+parameters are actually useful and so if
+
+00:09:10.200 --> 00:09:15.120
+you start out with a much larger model
+
+00:09:12.440 --> 00:09:17.720
+it's more likely to have useful subsets
+
+00:09:15.120 --> 00:09:20.320
+of the parameters basically um which is
+
+00:09:17.720 --> 00:09:21.560
+called the lottery ticket hypothesis uh
+
+00:09:20.320 --> 00:09:23.839
+there there's a famous paper called the
+
+00:09:21.560 --> 00:09:27.560
+lottery ticket hypothesis examines this
+
+00:09:23.839 --> 00:09:29.680
+phenomenon so um one one interesting
+
+00:09:27.560 --> 00:09:32.160
+thing is you can see that even if you
+
+00:09:29.680 --> 00:09:35.640
+scale up the compute even if you measure
+
+00:09:32.160 --> 00:09:37.640
+based on compute the uh larger models
+
+00:09:35.640 --> 00:09:38.959
+eventually surpass the smaller models in
+
+00:09:37.640 --> 00:09:41.920
+terms of how efficient they are at
+
+00:09:38.959 --> 00:09:44.680
+modeling the data and that's just
+
+00:09:41.920 --> 00:09:46.760
+because models tend to learn well for a
+
+00:09:44.680 --> 00:09:49.560
+while and then they basically reach
+
+00:09:46.760 --> 00:09:51.760
+their capacity and stop learning well or
+
+00:09:49.560 --> 00:09:53.680
+they start learning very slowly and once
+
+00:09:51.760 --> 00:09:57.120
+you get to that point the larger models
+
+00:09:53.680 --> 00:09:58.800
+work better so there's a kind of
+
+00:09:57.120 --> 00:10:00.640
+counterintuitive thing that if you want
+
+00:09:58.800 --> 00:10:04.160
+to train faster you actually can train a
+
+00:10:00.640 --> 00:10:06.839
+larger model and uh that will that will
+
+00:10:04.160 --> 00:10:08.000
+uh get you to a good solution at some
+
+00:10:06.839 --> 00:10:09.640
+point that will get you to a good
+
+00:10:08.000 --> 00:10:11.120
+solution faster than a smaller model
+
+00:10:09.640 --> 00:10:15.200
+would you know of course you need memory
+
+00:10:11.120 --> 00:10:15.200
+and stuff but why are looking
+
+00:10:20.040 --> 00:10:26.920
+at so this is test loss training loss
+
+00:10:22.760 --> 00:10:30.680
+also looks like this um I think on
+
+00:10:26.920 --> 00:10:34.360
+this particular
+
+00:10:30.680 --> 00:10:37.519
+on this particular paper they never
+
+00:10:34.360 --> 00:10:39.399
+repeated data and if you never repeat
+
+00:10:37.519 --> 00:10:42.560
+data actually your training loss looks
+
+00:10:39.399 --> 00:10:44.680
+very similar to your test loss because
+
+00:10:42.560 --> 00:10:46.079
+it if you like actually if you can
+
+00:10:44.680 --> 00:10:48.760
+assume your training data set and your
+
+00:10:46.079 --> 00:10:50.880
+test data set are um uh identically
+
+00:10:48.760 --> 00:10:52.279
+distributed your training loss of new
+
+00:10:50.880 --> 00:10:54.600
+training data should be exactly the same
+
+00:10:52.279 --> 00:10:55.959
+as your test loss so I think that's
+
+00:10:54.600 --> 00:10:57.760
+basically why they were justified in
+
+00:10:55.959 --> 00:11:01.000
+doing that good but they probably did
+
+00:10:57.760 --> 00:11:03.639
+test loss to like I swashed the concern
+
+00:11:01.000 --> 00:11:05.839
+that this was overfitting comp or
+
+00:11:03.639 --> 00:11:09.200
+something but good
+
+00:11:05.839 --> 00:11:11.279
+question um cool so the these are are
+
+00:11:09.200 --> 00:11:13.000
+good things to know um so basically if
+
+00:11:11.279 --> 00:11:14.839
+you see your model doing something like
+
+00:11:13.000 --> 00:11:16.279
+this um plateauing out maybe your
+
+00:11:14.839 --> 00:11:18.680
+model's too small and you need to tr a
+
+00:11:16.279 --> 00:11:20.920
+big
+
+00:11:18.680 --> 00:11:22.200
+basically another uh piece of trouble
+
+00:11:20.920 --> 00:11:26.800
+that you can have is trouble with
+
+00:11:22.200 --> 00:11:29.519
+optimization and basically um you should
+
+00:11:26.800 --> 00:11:31.600
+check your Optimizer um usually people
+
+00:11:29.519 --> 00:11:35.639
+are using atom variants nowadays like
+
+00:11:31.600 --> 00:11:37.839
+atom or atom W so just use that um
+
+00:11:35.639 --> 00:11:39.639
+learning rate uh so make sure that the
+
+00:11:37.839 --> 00:11:41.160
+learning rate you're using is standard
+
+00:11:39.639 --> 00:11:43.399
+for kind of the model size that you're
+
+00:11:41.160 --> 00:11:44.920
+using and the best way to do this is uh
+
+00:11:43.399 --> 00:11:46.000
+look at previous papers and see what
+
+00:11:44.920 --> 00:11:50.160
+they're
+
+00:11:46.000 --> 00:11:51.680
+using um initialization most people
+
+00:11:50.160 --> 00:11:53.440
+nowadays will not be training from
+
+00:11:51.680 --> 00:11:55.440
+scratch but if you are training from
+
+00:11:53.440 --> 00:11:58.040
+scratch how you initialize your model is
+
+00:11:55.440 --> 00:11:59.399
+really important and normally the way
+
+00:11:58.040 --> 00:12:03.320
+you do this is you do this with some
+
+00:11:59.399 --> 00:12:05.079
+sort of uniform random noise and uh
+
+00:12:03.320 --> 00:12:06.959
+specifically you can pick the uniform
+
+00:12:05.079 --> 00:12:08.800
+random noise in intelligent ways based
+
+00:12:06.959 --> 00:12:12.240
+on the the data size which I'll talk
+
+00:12:08.800 --> 00:12:13.920
+about in a second um also mini batching
+
+00:12:12.240 --> 00:12:15.639
+um are you using sufficiently large
+
+00:12:13.920 --> 00:12:17.480
+batches of data if you're using small
+
+00:12:15.639 --> 00:12:18.720
+batches of data you might have too much
+
+00:12:17.480 --> 00:12:21.279
+noise in your training and it might
+
+00:12:18.720 --> 00:12:23.839
+diverge so uh these are things you need
+
+00:12:21.279 --> 00:12:23.839
+think about as
+
+00:12:25.279 --> 00:12:30.560
+well
+
+00:12:27.519 --> 00:12:35.000
+cool um so these are training time
+
+00:12:30.560 --> 00:12:37.320
+things um the next thing is debugging at
+
+00:12:35.000 --> 00:12:37.320
+test
+
+00:12:38.160 --> 00:12:43.839
+time and this is particularly important
+
+00:12:41.240 --> 00:12:47.320
+if you're doing any sort
+
+00:12:43.839 --> 00:12:48.880
+of like I guess a lot of this has kind
+
+00:12:47.320 --> 00:12:51.360
+of been commoditized and it's
+
+00:12:48.880 --> 00:12:52.560
+implemented in hugging face and stuff
+
+00:12:51.360 --> 00:12:55.120
+like that and as long as you're using
+
+00:12:52.560 --> 00:12:57.279
+the standard implementations you're less
+
+00:12:55.120 --> 00:12:59.000
+likely to run into these bugs but if you
+
+00:12:57.279 --> 00:13:00.519
+are implementing anything on your own
+
+00:12:59.000 --> 00:13:03.040
+this is actually really tricky and you
+
+00:13:00.519 --> 00:13:07.880
+can easily make mistakes so uh it's
+
+00:13:03.040 --> 00:13:08.959
+important to to know about it so um what
+
+00:13:07.880 --> 00:13:10.680
+one of the reasons why you can have
+
+00:13:08.959 --> 00:13:12.240
+training and test disconnects especially
+
+00:13:10.680 --> 00:13:14.399
+if you're doing something like text
+
+00:13:12.240 --> 00:13:15.959
+generation is that usually your loss
+
+00:13:14.399 --> 00:13:17.720
+calculation and prodiction functions
+
+00:13:15.959 --> 00:13:20.480
+will be implemented in different
+
+00:13:17.720 --> 00:13:23.360
+functions and like anything in software
+
+00:13:20.480 --> 00:13:25.440
+engineering Um this can be a source of
+
+00:13:23.360 --> 00:13:26.760
+bugs duplicated sour code can be a
+
+00:13:25.440 --> 00:13:28.440
+source of bugs because you might
+
+00:13:26.760 --> 00:13:30.199
+Implement one thing in one place in one
+
+00:13:28.440 --> 00:13:33.000
+way another thing in another place in
+
+00:13:30.199 --> 00:13:35.560
+another way so this is no exception to
+
+00:13:33.000 --> 00:13:37.399
+that um it's especially true for
+
+00:13:35.560 --> 00:13:39.000
+structured prediction models so anything
+
+00:13:37.399 --> 00:13:40.399
+where you're not just making a single
+
+00:13:39.000 --> 00:13:42.079
+prediction but you're making multiple
+
+00:13:40.399 --> 00:13:43.839
+predictions in a row so you need to be a
+
+00:13:42.079 --> 00:13:46.959
+little bit careful about
+
+00:13:43.839 --> 00:13:49.880
+that um another thing that you need to
+
+00:13:46.959 --> 00:13:51.079
+be pay attention about is often uh
+
+00:13:49.880 --> 00:13:52.680
+especially if you're doing your own
+
+00:13:51.079 --> 00:13:55.880
+implementation loss calculation it's
+
+00:13:52.680 --> 00:13:59.800
+mini batched and generation is not or in
+
+00:13:55.880 --> 00:14:02.199
+highly optimized versions of um of
+
+00:13:59.800 --> 00:14:03.880
+inference you might be doing inference
+
+00:14:02.199 --> 00:14:05.360
+with Dynamic batching and stuff like
+
+00:14:03.880 --> 00:14:06.720
+that and it might become complicated you
+
+00:14:05.360 --> 00:14:09.800
+might make
+
+00:14:06.720 --> 00:14:12.160
+mistakes um so how do
+
+00:14:09.800 --> 00:14:15.839
+we make sure that we're not making any
+
+00:14:12.160 --> 00:14:18.560
+mistakes here um there's a really simple
+
+00:14:15.839 --> 00:14:21.199
+way to debug any sort of mini batched
+
+00:14:18.560 --> 00:14:24.199
+loss calculation because normally when
+
+00:14:21.199 --> 00:14:27.000
+we mini batch loss calculations we're
+
+00:14:24.199 --> 00:14:31.079
+simultaneously calculating uh the loss
+
+00:14:27.000 --> 00:14:35.600
+for like uh four four or eight or
+
+00:14:31.079 --> 00:14:37.560
+whatever sequences at a time and so you
+
+00:14:35.600 --> 00:14:40.279
+can calculate the loss with a large
+
+00:14:37.560 --> 00:14:42.000
+batch size like 32 and then calculate
+
+00:14:40.279 --> 00:14:44.920
+the loss for each uh sentence
+
+00:14:42.000 --> 00:14:47.720
+individually and sum them together and
+
+00:14:44.920 --> 00:14:49.480
+these uh value should be the same and
+
+00:14:47.720 --> 00:14:52.160
+this can help make sure that you don't
+
+00:14:49.480 --> 00:14:55.120
+have any you know issues with your
+
+00:14:52.160 --> 00:14:57.959
+padding or your masking or other things
+
+00:14:55.120 --> 00:14:59.800
+like this um so this is particularly
+
+00:14:57.959 --> 00:15:01.959
+important if you're not just using out
+
+00:14:59.800 --> 00:15:04.240
+of the box things so you have a slightly
+
+00:15:01.959 --> 00:15:06.240
+unusually structured model with like
+
+00:15:04.240 --> 00:15:08.880
+hierarchical encoding or anything like
+
+00:15:06.240 --> 00:15:11.680
+that you need to be really careful about
+
+00:15:08.880 --> 00:15:15.440
+that um you can even create unit tests
+
+00:15:11.680 --> 00:15:17.399
+that test this so like um in machine
+
+00:15:15.440 --> 00:15:18.959
+learning code we don't write unit test
+
+00:15:17.399 --> 00:15:20.160
+or especially neural network based
+
+00:15:18.959 --> 00:15:22.440
+machine learning code we don't write
+
+00:15:20.160 --> 00:15:24.160
+unit tests that often because it's kind
+
+00:15:22.440 --> 00:15:26.279
+of hard to do there's lots of Randomness
+
+00:15:24.160 --> 00:15:27.959
+and other stuff like that um but this is
+
+00:15:26.279 --> 00:15:30.959
+one thing that you can easily test and
+
+00:15:27.959 --> 00:15:30.959
+and make sure that you don't hear the
+
+00:15:32.440 --> 00:15:39.319
+mistakes um any sort of uh generation
+
+00:15:36.480 --> 00:15:43.199
+algorithm uh so when you're generating
+
+00:15:39.319 --> 00:15:44.639
+or decoding um you can make sure that
+
+00:15:43.199 --> 00:15:47.639
+your decoding code is getting the same
+
+00:15:44.639 --> 00:15:50.040
+score is when you calculate the loss and
+
+00:15:47.639 --> 00:15:52.959
+an easy way to do this is you call the
+
+00:15:50.040 --> 00:15:54.759
+decoding function to generate an output
+
+00:15:52.959 --> 00:15:57.399
+and normally when you're doing any sort
+
+00:15:54.759 --> 00:15:59.480
+of search or sampling or something like
+
+00:15:57.399 --> 00:16:02.120
+that during the search or sampling
+
+00:15:59.480 --> 00:16:05.000
+you're calculating the logits or the log
+
+00:16:02.120 --> 00:16:07.399
+probabilities of each step that you
+
+00:16:05.000 --> 00:16:09.120
+sample so you keep track of that during
+
+00:16:07.399 --> 00:16:12.279
+your sampling
+
+00:16:09.120 --> 00:16:14.319
+algorithm and then after that you call
+
+00:16:12.279 --> 00:16:16.800
+the loss function on the generated
+
+00:16:14.319 --> 00:16:18.639
+output and you calculate the loss
+
+00:16:16.800 --> 00:16:20.360
+according to the loss function and the
+
+00:16:18.639 --> 00:16:22.240
+score of these two things should be the
+
+00:16:20.360 --> 00:16:26.440
+same uh
+
+00:16:22.240 --> 00:16:26.440
+so um you know you do your
+
+00:16:27.920 --> 00:16:35.279
+generate and that gives you an
+
+00:16:32.000 --> 00:16:35.279
+output in
+
+00:16:35.600 --> 00:16:42.360
+score and then you do um
+
+00:16:39.319 --> 00:16:45.839
+loss on the
+
+00:16:42.360 --> 00:16:49.040
+output and that gives you the score
+
+00:16:45.839 --> 00:16:53.079
+two and then you just compare these two
+
+00:16:49.040 --> 00:16:56.360
+things together and this can uh in in my
+
+00:16:53.079 --> 00:17:01.120
+experience this has allowed me to find
+
+00:16:56.360 --> 00:17:03.240
+the majority of the bugs in um these two
+
+00:17:01.120 --> 00:17:04.679
+things um have allowed me to find the
+
+00:17:03.240 --> 00:17:06.600
+majority of the bugs whenever I was
+
+00:17:04.679 --> 00:17:09.199
+doing any sort of like complex thing
+
+00:17:06.600 --> 00:17:11.880
+with respect to generation or models and
+
+00:17:09.199 --> 00:17:13.360
+stuff like that so um it's a very common
+
+00:17:11.880 --> 00:17:15.439
+place for bugs even if you're pretty
+
+00:17:13.360 --> 00:17:17.280
+familiar with models so I I would highly
+
+00:17:15.439 --> 00:17:19.760
+recommend
+
+00:17:17.280 --> 00:17:21.319
+that um this is particularly bad when
+
+00:17:19.760 --> 00:17:25.559
+you're doing something like a search
+
+00:17:21.319 --> 00:17:28.400
+algorithm like beam search um and
+
+00:17:25.559 --> 00:17:30.400
+so beam search uh as you know from the
+
+00:17:28.400 --> 00:17:34.200
+generation class instead of picking one
+
+00:17:30.400 --> 00:17:37.080
+high probability uh you know word in
+
+00:17:34.200 --> 00:17:40.160
+your next step you maintain several
+
+00:17:37.080 --> 00:17:41.960
+paths and one way that you can fix this
+
+00:17:40.160 --> 00:17:44.320
+is as you make search better the model
+
+00:17:41.960 --> 00:17:45.760
+score should get better so the log
+
+00:17:44.320 --> 00:17:48.240
+likelihood of the output should get
+
+00:17:45.760 --> 00:17:50.280
+better almost all of the time so you can
+
+00:17:48.240 --> 00:17:51.840
+search with varying beam sizes and make
+
+00:17:50.280 --> 00:17:55.280
+sure that you get a better overall model
+
+00:17:51.840 --> 00:17:57.559
+score at the end so um and you can even
+
+00:17:55.280 --> 00:17:59.320
+create a unit test testing this as well
+
+00:17:57.559 --> 00:18:01.000
+I don't think that that many people will
+
+00:17:59.320 --> 00:18:02.480
+be reimplementing beam search so you
+
+00:18:01.000 --> 00:18:04.120
+might not need to worry about that too
+
+00:18:02.480 --> 00:18:05.679
+much but in case you are doing anything
+
+00:18:04.120 --> 00:18:08.159
+with respect to search algorithms it's a
+
+00:18:05.679 --> 00:18:08.159
+good thing to
+
+00:18:08.880 --> 00:18:15.159
+know
+
+00:18:10.480 --> 00:18:15.159
+cool um any questions about these two so
+
+00:18:16.919 --> 00:18:24.159
+far no okay um so the second the next
+
+00:18:22.600 --> 00:18:25.400
+thing I want to talk about this is
+
+00:18:24.159 --> 00:18:27.840
+something that people think about a
+
+00:18:25.400 --> 00:18:29.400
+little bit less uh but it's actually
+
+00:18:27.840 --> 00:18:31.280
+something really important to know
+
+00:18:29.400 --> 00:18:34.280
+because it will affect you it will
+
+00:18:31.280 --> 00:18:35.799
+affect everybody uh to some extent it
+
+00:18:34.280 --> 00:18:40.760
+will affect you to a greater or lesser
+
+00:18:35.799 --> 00:18:41.520
+extent depending on um what uh type of
+
+00:18:40.760 --> 00:18:44.480
+you
+
+00:18:41.520 --> 00:18:46.799
+know system you're building but it will
+
+00:18:44.480 --> 00:18:48.760
+definitely affect everybody and that's
+
+00:18:46.799 --> 00:18:50.960
+the mismatch between the the function
+
+00:18:48.760 --> 00:18:53.440
+that you're optimizing at training time
+
+00:18:50.960 --> 00:18:55.240
+and the evaluation metric that you're
+
+00:18:53.440 --> 00:18:58.000
+evaluating and
+
+00:18:55.240 --> 00:18:59.679
+so uh like as I said in the
+
+00:18:58.000 --> 00:19:01.679
+reinforcement learning class it's very
+
+00:18:59.679 --> 00:19:03.640
+common to optimize for maximum
+
+00:19:01.679 --> 00:19:06.039
+likelihood for training uh but there's
+
+00:19:03.640 --> 00:19:07.840
+all kinds of problems with this you know
+
+00:19:06.039 --> 00:19:09.640
+um with respect to the mistake it not
+
+00:19:07.840 --> 00:19:11.640
+being sensitive to mistakes it not being
+
+00:19:09.640 --> 00:19:14.799
+sensitive to your generation
+
+00:19:11.640 --> 00:19:16.520
+algorithm um but even though your
+
+00:19:14.799 --> 00:19:19.880
+likelihood is getting better accuracy
+
+00:19:16.520 --> 00:19:22.799
+can get worse and this is a super simple
+
+00:19:19.880 --> 00:19:25.080
+example with uh image classification on
+
+00:19:22.799 --> 00:19:27.919
+mest and I I ran this experiment with
+
+00:19:25.080 --> 00:19:30.880
+like 10 lines of pytorch code or
+
+00:19:27.919 --> 00:19:36.840
+something like this uh maybe more like
+
+00:19:30.880 --> 00:19:40.080
+40 lines of P um and so here um on the
+
+00:19:36.840 --> 00:19:43.120
+left side we have the loss on the
+
+00:19:40.080 --> 00:19:46.600
+training set and the test set or the dev
+
+00:19:43.120 --> 00:19:48.559
+set and here we have accuracy on the
+
+00:19:46.600 --> 00:19:50.799
+training set in the test
+
+00:19:48.559 --> 00:19:55.000
+set
+
+00:19:50.799 --> 00:19:56.159
+and so oops I showed you the answer so I
+
+00:19:55.000 --> 00:19:58.799
+was going to do a quiz but I
+
+00:19:56.159 --> 00:20:00.559
+accidentally showed you the answer um
+
+00:19:58.799 --> 00:20:04.440
+but the problem here is basically
+
+00:20:00.559 --> 00:20:06.320
+because um the the loss you're
+
+00:20:04.440 --> 00:20:09.400
+calculating the likelihood of the
+
+00:20:06.320 --> 00:20:11.120
+correct answer and the likelihood of the
+
+00:20:09.400 --> 00:20:12.440
+correct answer is the probability of
+
+00:20:11.120 --> 00:20:15.000
+getting the correct
+
+00:20:12.440 --> 00:20:17.240
+answer the accuracy is the number of
+
+00:20:15.000 --> 00:20:20.280
+times you're getting the correct answer
+
+00:20:17.240 --> 00:20:23.799
+so as you train a model to get more and
+
+00:20:20.280 --> 00:20:25.440
+more confident it gets better it gets
+
+00:20:23.799 --> 00:20:27.840
+better and better at getting more
+
+00:20:25.440 --> 00:20:30.039
+answers correct but it also gets more
+
+00:20:27.840 --> 00:20:33.360
+and more confident in its answers and so
+
+00:20:30.039 --> 00:20:36.200
+if the you know there's any example that
+
+00:20:33.360 --> 00:20:37.840
+it's really bad at um it might get very
+
+00:20:36.200 --> 00:20:42.320
+confident in
+
+00:20:37.840 --> 00:20:44.760
+that answer that bad answer and the log
+
+00:20:42.320 --> 00:20:47.320
+likelihood of that answer will go up or
+
+00:20:44.760 --> 00:20:49.679
+sorry the log likelihood will go down so
+
+00:20:47.320 --> 00:20:54.360
+the negative log likelihood will go up
+
+00:20:49.679 --> 00:20:56.720
+is the loss so basically
+
+00:20:54.360 --> 00:20:59.559
+um the
+
+00:20:56.720 --> 00:21:01.039
+uh the loss that you're calculating and
+
+00:20:59.559 --> 00:21:03.840
+the thing that you care about in the end
+
+00:21:01.039 --> 00:21:07.120
+accuracy can be decorrelated
+
+00:21:03.840 --> 00:21:09.520
+um so there's also an interesting
+
+00:21:07.120 --> 00:21:12.080
+example um in text generation and this
+
+00:21:09.520 --> 00:21:14.000
+is part of the reason why uh we have all
+
+00:21:12.080 --> 00:21:15.880
+these other text generation algorithms
+
+00:21:14.000 --> 00:21:20.080
+like nucleus samp playing or topk samp
+
+00:21:15.880 --> 00:21:23.039
+playing or other things like this is um
+
+00:21:20.080 --> 00:21:25.080
+actually in a maximum likelihood trained
+
+00:21:23.039 --> 00:21:27.799
+model better
+
+00:21:25.080 --> 00:21:29.559
+search uh in in other words finding a
+
+00:21:27.799 --> 00:21:32.159
+better model scope
+
+00:21:29.559 --> 00:21:36.120
+doesn't necessarily give you a better
+
+00:21:32.159 --> 00:21:37.840
+generation result and this is an example
+
+00:21:36.120 --> 00:21:39.080
+uh from machine translation from a
+
+00:21:37.840 --> 00:21:41.880
+really long time
+
+00:21:39.080 --> 00:21:44.000
+ago uh but you know it still persists
+
+00:21:41.880 --> 00:21:47.520
+today which is they did beam search with
+
+00:21:44.000 --> 00:21:53.600
+a larger and larger beam
+
+00:21:47.520 --> 00:21:56.640
+and the be the best Beam for finding um
+
+00:21:53.600 --> 00:21:59.640
+the best scoring output basically was
+
+00:21:56.640 --> 00:22:01.600
+four and then the accuracy goes down and
+
+00:21:59.640 --> 00:22:05.559
+down and down as they find a better
+
+00:22:01.600 --> 00:22:07.200
+output and does anyone remember when we
+
+00:22:05.559 --> 00:22:09.679
+talked about the generation class where
+
+00:22:07.200 --> 00:22:09.679
+this comes
+
+00:22:10.120 --> 00:22:15.000
+from I don't know how explicitly we said
+
+00:22:12.960 --> 00:22:18.600
+we mentioned it in the generation class
+
+00:22:15.000 --> 00:22:20.360
+but basically the problem is um maximum
+
+00:22:18.600 --> 00:22:22.559
+likelihood train models like shorter
+
+00:22:20.360 --> 00:22:25.240
+outputs generally because if as we make
+
+00:22:22.559 --> 00:22:27.760
+the output longer uh the probability of
+
+00:22:25.240 --> 00:22:29.679
+the longer outputs goes down so as you
+
+00:22:27.760 --> 00:22:32.039
+improve the beam it will start
+
+00:22:29.679 --> 00:22:34.799
+generating shorter and shorter outputs
+
+00:22:32.039 --> 00:22:36.480
+and because of that the score goes down
+
+00:22:34.799 --> 00:22:39.039
+because blue score doesn't like outputs
+
+00:22:36.480 --> 00:22:41.520
+that are too short essentially so there
+
+00:22:39.039 --> 00:22:44.039
+are um there are hex around this for
+
+00:22:41.520 --> 00:22:46.200
+beam search where essentially what you
+
+00:22:44.039 --> 00:22:48.559
+do is you uh take the average log
+
+00:22:46.200 --> 00:22:51.159
+likelihood of each token instead of the
+
+00:22:48.559 --> 00:22:52.760
+overall log likelihood of each token um
+
+00:22:51.159 --> 00:22:54.679
+and that improves a little bit but still
+
+00:22:52.760 --> 00:22:59.720
+you can see as you search more the the
+
+00:22:54.679 --> 00:23:01.440
+accuracy goes down so um so that's the
+
+00:22:59.720 --> 00:23:04.039
+the general idea
+
+00:23:01.440 --> 00:23:08.760
+here there's a bunch of ways you can fix
+
+00:23:04.039 --> 00:23:10.600
+this um the most principled way is to
+
+00:23:08.760 --> 00:23:12.760
+use a method like reinforcement learning
+
+00:23:10.600 --> 00:23:14.120
+or something uh some sort of you know
+
+00:23:12.760 --> 00:23:15.520
+structured training algorithm that
+
+00:23:14.120 --> 00:23:17.159
+allows you to train your models so that
+
+00:23:15.520 --> 00:23:20.159
+you don't get these bad
+
+00:23:17.159 --> 00:23:22.159
+outputs um another way that's much
+
+00:23:20.159 --> 00:23:25.640
+easier is to do early stopping with the
+
+00:23:22.159 --> 00:23:30.480
+evaluation metric as opposed to um early
+
+00:23:25.640 --> 00:23:32.840
+stopping with the loss and by doing this
+
+00:23:30.480 --> 00:23:34.520
+you would stop here so you would stop
+
+00:23:32.840 --> 00:23:37.159
+where you get the highest evaluation
+
+00:23:34.520 --> 00:23:42.600
+metric uh that you care about instead of
+
+00:23:37.159 --> 00:23:44.400
+stopping here uh so that's um that's one
+
+00:23:42.600 --> 00:23:46.600
+way you can fix this
+
+00:23:44.400 --> 00:23:49.760
+problem does anyone have an idea about
+
+00:23:46.600 --> 00:23:49.760
+why this might be a bad
+
+00:23:49.840 --> 00:23:57.159
+idea why might it be a bad idea to stop
+
+00:23:52.480 --> 00:23:57.159
+here instead of stopping here for
+
+00:23:57.440 --> 00:24:00.440
+example
+
+00:24:05.320 --> 00:24:10.200
+yeah it's kind of overfitting it's
+
+00:24:07.760 --> 00:24:13.640
+overfitting in a particular way um but
+
+00:24:10.200 --> 00:24:16.000
+remember here this is still the accuracy
+
+00:24:13.640 --> 00:24:18.400
+on the dev set so we're not overfitting
+
+00:24:16.000 --> 00:24:20.080
+so much that the dev accuracy is going
+
+00:24:18.400 --> 00:24:24.279
+down that would be a different variety
+
+00:24:20.080 --> 00:24:27.360
+of overfitting but any any
+
+00:24:24.279 --> 00:24:29.799
+ideas go for it we don't want to be too
+
+00:24:27.360 --> 00:24:31.600
+confident yeah exactly we don't want it
+
+00:24:29.799 --> 00:24:32.880
+to be too confident in its wrong answers
+
+00:24:31.600 --> 00:24:35.279
+and we talked about
+
+00:24:32.880 --> 00:24:38.000
+calibration um where calibration is
+
+00:24:35.279 --> 00:24:40.039
+basically like how accurate are the
+
+00:24:38.000 --> 00:24:41.480
+probability estimates so this model over
+
+00:24:40.039 --> 00:24:43.600
+here is going to be really poorly
+
+00:24:41.480 --> 00:24:45.159
+calibrated it's going to be very
+
+00:24:43.600 --> 00:24:46.240
+confident regardless of whether it's
+
+00:24:45.159 --> 00:24:49.440
+correct or not and that could be a
+
+00:24:46.240 --> 00:24:50.840
+problem in dopram uh dopram tasks
+
+00:24:49.440 --> 00:24:52.130
+there's also another thing that I I
+
+00:24:50.840 --> 00:24:55.189
+forgot to put on
+
+00:24:52.130 --> 00:24:55.189
+[Music]
+
+00:24:57.320 --> 00:25:00.320
+um
+
+00:25:02.919 --> 00:25:08.120
+that I forgot to put on the slides but
+
+00:25:04.520 --> 00:25:10.720
+it's a um an interesting phenomenon that
+
+00:25:08.120 --> 00:25:12.720
+actually um kind of a lot of people in
+
+00:25:10.720 --> 00:25:16.360
+interpretability are interested in it's
+
+00:25:12.720 --> 00:25:18.120
+this uh generalization gring
+
+00:25:16.360 --> 00:25:19.640
+generalization Beyond overfitting on
+
+00:25:18.120 --> 00:25:21.120
+small algorithmic data sets and
+
+00:25:19.640 --> 00:25:27.360
+basically what they
+
+00:25:21.120 --> 00:25:29.720
+show is um you can be training for a
+
+00:25:27.360 --> 00:25:31.320
+very very long time
+
+00:25:29.720 --> 00:25:34.279
+um
+
+00:25:31.320 --> 00:25:35.919
+and uh like reducing the loss reducing
+
+00:25:34.279 --> 00:25:40.399
+the loss reducing the loss and reducing
+
+00:25:35.919 --> 00:25:42.480
+the loss and it's only after a very long
+
+00:25:40.399 --> 00:25:43.840
+time does your Model start generalizing
+
+00:25:42.480 --> 00:25:48.240
+well and getting good
+
+00:25:43.840 --> 00:25:49.799
+accuracy um the this paper the types of
+
+00:25:48.240 --> 00:25:52.120
+data sets it's talking about are data
+
+00:25:49.799 --> 00:25:55.520
+sets where you need to get many things
+
+00:25:52.120 --> 00:25:58.640
+in a row correct before you get the
+
+00:25:55.520 --> 00:26:00.880
+final answer correct so basically you
+
+00:25:58.640 --> 00:26:02.320
+need to get like 20 steps in a row or 50
+
+00:26:00.880 --> 00:26:06.200
+steps in a row correct before you get
+
+00:26:02.320 --> 00:26:10.679
+the final answer correct and um
+
+00:26:06.200 --> 00:26:13.000
+basically the reason why this happens is
+
+00:26:10.679 --> 00:26:15.720
+because this accuracy will keep going up
+
+00:26:13.000 --> 00:26:17.760
+but you only get the accuracy of each
+
+00:26:15.720 --> 00:26:20.520
+individual decision will keep going up
+
+00:26:17.760 --> 00:26:22.880
+but you only get marked like
+
+00:26:20.520 --> 00:26:25.440
+correct uh
+
+00:26:22.880 --> 00:26:29.799
+after you get like all 50 in a row
+
+00:26:25.440 --> 00:26:31.200
+correct so um it this difference can be
+
+00:26:29.799 --> 00:26:33.039
+even more Stark when you're talking
+
+00:26:31.200 --> 00:26:35.399
+about things that require like 50 steps
+
+00:26:33.039 --> 00:26:37.399
+of reasoning or like multiple steps of
+
+00:26:35.399 --> 00:26:39.559
+reasoning but like 50 token Generations
+
+00:26:37.399 --> 00:26:42.679
+correct before you get them right so um
+
+00:26:39.559 --> 00:26:42.679
+that's another thing to be aware
+
+00:26:43.000 --> 00:26:49.240
+of cool um so now I want to switch gears
+
+00:26:46.960 --> 00:26:51.919
+a little bit to actionable evaluation
+
+00:26:49.240 --> 00:26:54.240
+and how you can um evaluate your models
+
+00:26:51.919 --> 00:26:56.640
+in a way that makes it easy to find uh
+
+00:26:54.240 --> 00:26:58.600
+next steps to be
+
+00:26:56.640 --> 00:27:00.159
+improving uh are there any questions
+
+00:26:58.600 --> 00:27:02.600
+about the debugging part before we get
+
+00:27:00.159 --> 00:27:02.600
+into this
+
+00:27:03.360 --> 00:27:10.120
+part okay I'll
+
+00:27:05.880 --> 00:27:12.840
+go so um my first suggestion with
+
+00:27:10.120 --> 00:27:15.559
+respect to how you can actually you know
+
+00:27:12.840 --> 00:27:17.440
+improve systems is make sure that you're
+
+00:27:15.559 --> 00:27:21.039
+looking at the data that you're
+
+00:27:17.440 --> 00:27:22.679
+using and um both bugs and new research
+
+00:27:21.039 --> 00:27:24.080
+directions can be found by looking at
+
+00:27:22.679 --> 00:27:27.159
+your model
+
+00:27:24.080 --> 00:27:31.640
+outputs um
+
+00:27:27.159 --> 00:27:33.279
+so to give one example um of a very
+
+00:27:31.640 --> 00:27:36.200
+common mistake that you can make when
+
+00:27:33.279 --> 00:27:40.159
+you're creating a a generation algorithm
+
+00:27:36.200 --> 00:27:41.600
+it's these sort of off by one erors um
+
+00:27:40.159 --> 00:27:43.919
+so like let's say you implemented a
+
+00:27:41.600 --> 00:27:46.039
+translation system and it's generating
+
+00:27:43.919 --> 00:27:49.440
+outputs like went to the store yesterday
+
+00:27:46.039 --> 00:27:51.080
+bought a dog um you can immediately look
+
+00:27:49.440 --> 00:27:53.440
+at this and say hey this doesn't look
+
+00:27:51.080 --> 00:27:58.360
+like natural English what's going uh
+
+00:27:53.440 --> 00:28:00.000
+what's going on and the the problem here
+
+00:27:58.360 --> 00:28:04.600
+is
+
+00:28:00.000 --> 00:28:04.600
+you're um you're doing something
+
+00:28:05.159 --> 00:28:12.720
+like output uh
+
+00:28:09.240 --> 00:28:14.600
+one uh and you have a slice of like one
+
+00:28:12.720 --> 00:28:17.399
+instead of zero here or something like
+
+00:28:14.600 --> 00:28:18.640
+this and so this is a really silly error
+
+00:28:17.399 --> 00:28:21.000
+that you might just make a mistake on
+
+00:28:18.640 --> 00:28:23.679
+python on your you know pre-processing
+
+00:28:21.000 --> 00:28:26.200
+or postprocessing or something like this
+
+00:28:23.679 --> 00:28:28.399
+um but the problem is like if you look
+
+00:28:26.200 --> 00:28:30.600
+at your blue score based evaluation or
+
+00:28:28.399 --> 00:28:32.840
+something like that you'll have like
+
+00:28:30.600 --> 00:28:34.760
+you'll be one point worse or two points
+
+00:28:32.840 --> 00:28:36.720
+worse or something like that and you'll
+
+00:28:34.760 --> 00:28:38.600
+be like Oh I'm I'm two points worse why
+
+00:28:36.720 --> 00:28:40.600
+am I two point wor two points worse in
+
+00:28:38.600 --> 00:28:43.760
+the state of the art and it turns out it
+
+00:28:40.600 --> 00:28:45.279
+was a really like silly thing like this
+
+00:28:43.760 --> 00:28:46.519
+and immediately you'll see this if you
+
+00:28:45.279 --> 00:28:47.960
+look at your data but if you're doing
+
+00:28:46.519 --> 00:28:49.600
+all your experiments and just looking at
+
+00:28:47.960 --> 00:28:51.519
+the numbers it's really hard to tell you
+
+00:28:49.600 --> 00:28:53.720
+know why this is
+
+00:28:51.519 --> 00:28:58.720
+happening
+
+00:28:53.720 --> 00:29:02.360
+um another thing is uh if you
+
+00:28:58.720 --> 00:29:04.799
+have a good eye and can like just look
+
+00:29:02.360 --> 00:29:07.799
+through the data points
+
+00:29:04.799 --> 00:29:09.640
+um we as humans are pretty good uh
+
+00:29:07.799 --> 00:29:14.200
+pattern recognizers and especially you
+
+00:29:09.640 --> 00:29:16.360
+know CMU students uh you're uh very good
+
+00:29:14.200 --> 00:29:18.519
+and quick at picking up on things so if
+
+00:29:16.360 --> 00:29:20.600
+you look at the data and pour through
+
+00:29:18.519 --> 00:29:22.880
+things you can uh probably pick up
+
+00:29:20.600 --> 00:29:24.880
+patterns about why things are failing
+
+00:29:22.880 --> 00:29:27.720
+and so um you know you might look and
+
+00:29:24.880 --> 00:29:29.919
+see that uh compared to some other model
+
+00:29:27.720 --> 00:29:31.679
+your model is really bad at answering
+
+00:29:29.919 --> 00:29:33.679
+questions about people or something like
+
+00:29:31.679 --> 00:29:36.480
+that and then you figure out you'll need
+
+00:29:33.679 --> 00:29:38.320
+a better model of uh people or your rag
+
+00:29:36.480 --> 00:29:40.519
+systems uh that you're building for
+
+00:29:38.320 --> 00:29:42.880
+assignment two is maybe failing on all
+
+00:29:40.519 --> 00:29:45.559
+the research related questions so you
+
+00:29:42.880 --> 00:29:47.080
+need to come up with the research uh
+
+00:29:45.559 --> 00:29:48.320
+like scrape more research data or
+
+00:29:47.080 --> 00:29:50.080
+something like
+
+00:29:48.320 --> 00:29:53.840
+that
+
+00:29:50.080 --> 00:29:55.760
+um so there are methods to do this more
+
+00:29:53.840 --> 00:29:58.039
+systematically and this is something I
+
+00:29:55.760 --> 00:29:59.720
+picked up when I was doing an internship
+
+00:29:58.039 --> 00:30:04.080
+at Google and it really stuck with me
+
+00:29:59.720 --> 00:30:09.080
+for you know 14 uh 14 years now I guess
+
+00:30:04.080 --> 00:30:10.960
+13 years um so uh a very simple way to
+
+00:30:09.080 --> 00:30:12.600
+do this more systematically than just
+
+00:30:10.960 --> 00:30:16.200
+browsing through things is to randomly
+
+00:30:12.600 --> 00:30:19.000
+sample a 100 outputs and look at a 100
+
+00:30:16.200 --> 00:30:21.840
+erors and try to group them into some
+
+00:30:19.000 --> 00:30:23.799
+sort of typology and say oh uh this kind
+
+00:30:21.840 --> 00:30:27.799
+of air is particularly
+
+00:30:23.799 --> 00:30:31.279
+frequent and this is just one example of
+
+00:30:27.799 --> 00:30:33.120
+a typology that was defined by V at all
+
+00:30:31.279 --> 00:30:37.320
+um where they tried to take machine
+
+00:30:33.120 --> 00:30:39.480
+translation errors and group them into
+
+00:30:37.320 --> 00:30:43.440
+uh various varieties like correct words
+
+00:30:39.480 --> 00:30:46.640
+filler words local uh local range long
+
+00:30:43.440 --> 00:30:48.440
+range um uh sorry word word level word
+
+00:30:46.640 --> 00:30:50.440
+ordering erors local range long range
+
+00:30:48.440 --> 00:30:54.279
+phrase level local range long range and
+
+00:30:50.440 --> 00:30:55.679
+stuff like this um you can definitely
+
+00:30:54.279 --> 00:30:58.399
+look at previous work and see the
+
+00:30:55.679 --> 00:31:00.559
+typologies of errors that they used but
+
+00:30:58.399 --> 00:31:02.440
+the problem is like systems get better
+
+00:31:00.559 --> 00:31:04.240
+and actually I don't think this is a
+
+00:31:02.440 --> 00:31:06.760
+super relevant typology for machine
+
+00:31:04.240 --> 00:31:10.120
+translation anymore uh because machine
+
+00:31:06.760 --> 00:31:12.159
+translation systems like they don't make
+
+00:31:10.120 --> 00:31:14.639
+a whole lot of local range Word level
+
+00:31:12.159 --> 00:31:16.159
+errors anymore and rather we might want
+
+00:31:14.639 --> 00:31:18.279
+to know more fine grain like are they
+
+00:31:16.159 --> 00:31:21.720
+making mistakes on named entities or
+
+00:31:18.279 --> 00:31:24.720
+other things like that so actually
+
+00:31:21.720 --> 00:31:24.720
+we
+
+00:31:26.919 --> 00:31:29.919
+um
+
+00:31:30.519 --> 00:31:36.279
+did a re a more recent thing it's I
+
+00:31:34.279 --> 00:31:39.159
+guess four years ago now um but it was
+
+00:31:36.279 --> 00:31:42.720
+when uh people first started saying that
+
+00:31:39.159 --> 00:31:46.200
+machine translation systems are about as
+
+00:31:42.720 --> 00:31:50.720
+good as humans at doing a
+
+00:31:46.200 --> 00:31:50.720
+translation and when we did this we
+
+00:31:52.480 --> 00:31:58.440
+compared we compared machine translation
+
+00:31:55.200 --> 00:31:59.960
+systems to humans and we tried to find
+
+00:31:58.440 --> 00:32:02.240
+you know different types of things and
+
+00:31:59.960 --> 00:32:03.919
+we were inspired by V but we recreated
+
+00:32:02.240 --> 00:32:06.159
+our typology based on the things that we
+
+00:32:03.919 --> 00:32:10.279
+thought were you know the most important
+
+00:32:06.159 --> 00:32:13.399
+types of errors in like 2020 instead of
+
+00:32:10.279 --> 00:32:16.799
+2006 so this is really helpful the
+
+00:32:13.399 --> 00:32:19.039
+reason why it's really helpful is if you
+
+00:32:16.799 --> 00:32:20.440
+can do this even for a small sample of
+
+00:32:19.039 --> 00:32:23.440
+the outputs that you're looking at and
+
+00:32:20.440 --> 00:32:25.279
+identify the most like prominent types
+
+00:32:23.440 --> 00:32:27.440
+of eras that you're facing it often
+
+00:32:25.279 --> 00:32:29.360
+leads you to the most successful ways of
+
+00:32:27.440 --> 00:32:31.519
+improving the accuracy of your systems
+
+00:32:29.360 --> 00:32:33.120
+because you might if you don't do this
+
+00:32:31.519 --> 00:32:35.000
+you might be focusing on an air type
+
+00:32:33.120 --> 00:32:38.000
+that's not actually an error it's kind
+
+00:32:35.000 --> 00:32:39.200
+of like if you learned in uh programming
+
+00:32:38.000 --> 00:32:40.799
+you know software engineering or
+
+00:32:39.200 --> 00:32:42.639
+something like that you should never
+
+00:32:40.799 --> 00:32:46.360
+optimize your code until you run a
+
+00:32:42.639 --> 00:32:47.799
+profiler um because actually your code
+
+00:32:46.360 --> 00:32:50.320
+might be slow in a place that you never
+
+00:32:47.799 --> 00:32:52.720
+expected and so it's kind of the same
+
+00:32:50.320 --> 00:32:56.600
+principle here right so don't optimize
+
+00:32:52.720 --> 00:32:58.720
+your systems errors in a place uh where
+
+00:32:56.600 --> 00:33:03.240
+like actually it's not having in years
+
+00:32:58.720 --> 00:33:06.440
+so um that's a general principle
+
+00:33:03.240 --> 00:33:09.440
+here uh cool another thing you can do is
+
+00:33:06.440 --> 00:33:11.760
+quantitative analysis so um if you can
+
+00:33:09.440 --> 00:33:13.880
+think of the phenomenon that you choose
+
+00:33:11.760 --> 00:33:17.480
+to focus on um is that phenomenon
+
+00:33:13.880 --> 00:33:19.159
+getting better so if you focused on uh
+
+00:33:17.480 --> 00:33:22.240
+something that should improve the
+
+00:33:19.159 --> 00:33:23.760
+quality of low frequency words uh you
+
+00:33:22.240 --> 00:33:26.200
+can check if the accuracy on low
+
+00:33:23.760 --> 00:33:27.399
+frequency words is increasing if you
+
+00:33:26.200 --> 00:33:29.600
+focused on something that should be
+
+00:33:27.399 --> 00:33:32.120
+improving the syntax in a low resource
+
+00:33:29.600 --> 00:33:36.080
+language you can measure um whether it's
+
+00:33:32.120 --> 00:33:37.360
+doing better on word ordering or uh long
+
+00:33:36.080 --> 00:33:41.840
+distance
+
+00:33:37.360 --> 00:33:44.360
+dependencies um if you focused on
+
+00:33:41.840 --> 00:33:46.039
+improving a search algorithm for you
+
+00:33:44.360 --> 00:33:47.519
+know generation or something like that
+
+00:33:46.039 --> 00:33:49.880
+are the number of search errors that
+
+00:33:47.519 --> 00:33:53.120
+you're encountering being reduced so
+
+00:33:49.880 --> 00:33:56.320
+depending on what you planned on uh you
+
+00:33:53.120 --> 00:33:57.919
+know improving it's often a good idea to
+
+00:33:56.320 --> 00:33:59.480
+measure more directly whether it's
+
+00:33:57.919 --> 00:34:00.559
+improving the the thing that you think
+
+00:33:59.480 --> 00:34:04.880
+it should
+
+00:34:00.559 --> 00:34:06.000
+improve um one example of um so I I
+
+00:34:04.880 --> 00:34:09.240
+basically
+
+00:34:06.000 --> 00:34:11.240
+created since my experience doing this
+
+00:34:09.240 --> 00:34:15.159
+manually uh when I I was on an
+
+00:34:11.240 --> 00:34:18.280
+internship at Google um I've
+
+00:34:15.159 --> 00:34:20.639
+gradually improved my methodology for
+
+00:34:18.280 --> 00:34:20.639
+doing
+
+00:34:21.679 --> 00:34:26.320
+this and um and worked on automating
+
+00:34:24.879 --> 00:34:30.599
+things and
+
+00:34:26.320 --> 00:34:33.839
+so the first thing I had was a super
+
+00:34:30.599 --> 00:34:35.560
+hacky uh hacky script that basically
+
+00:34:33.839 --> 00:34:37.720
+writes out HTML
+
+00:34:35.560 --> 00:34:39.320
+files um and then I I had something
+
+00:34:37.720 --> 00:34:42.320
+called explainer board where we had a
+
+00:34:39.320 --> 00:34:44.879
+leader board and uh recently one of the
+
+00:34:42.320 --> 00:34:47.800
+things I've worked on is uh this uh
+
+00:34:44.879 --> 00:34:53.200
+together with um Alex Alex Cabrera who's
+
+00:34:47.800 --> 00:34:56.760
+a student here um is this toolkit called
+
+00:34:53.200 --> 00:34:59.640
+Zeno and um this is just an example from
+
+00:34:56.760 --> 00:34:59.640
+machine translation
+
+00:35:03.440 --> 00:35:09.200
+it's being a little bit
+
+00:35:06.599 --> 00:35:11.079
+slow um but basically what it does is it
+
+00:35:09.200 --> 00:35:14.920
+allows you to look at the data on the
+
+00:35:11.079 --> 00:35:18.000
+right side um and so these are just
+
+00:35:14.920 --> 00:35:19.680
+examples um but you can go in and do
+
+00:35:18.000 --> 00:35:22.760
+things like say Okay I want to look at
+
+00:35:19.680 --> 00:35:24.640
+all machine translation examples
+
+00:35:22.760 --> 00:35:28.040
+from
+
+00:35:24.640 --> 00:35:30.920
+uh housea and so it shows you the ones
+
+00:35:28.040 --> 00:35:32.960
+from housea I want to look
+
+00:35:30.920 --> 00:35:36.240
+at all
+
+00:35:32.960 --> 00:35:38.880
+examples let me clear that off I want to
+
+00:35:36.240 --> 00:35:40.800
+look at all examples where the accuracy
+
+00:35:38.880 --> 00:35:43.440
+is
+
+00:35:40.800 --> 00:35:45.280
+low um and so now I can look at all the
+
+00:35:43.440 --> 00:35:49.640
+examples where the accuracy is low and I
+
+00:35:45.280 --> 00:35:52.640
+I can go in and uh uh examine them so uh
+
+00:35:49.640 --> 00:35:54.880
+you can also go in and build charts like
+
+00:35:52.640 --> 00:35:58.280
+this so like what is the overall
+
+00:35:54.880 --> 00:36:02.200
+performance um what is is the
+
+00:35:58.280 --> 00:36:05.960
+performance what is the performance
+
+00:36:02.200 --> 00:36:07.520
+um on different scripts so you can see
+
+00:36:05.960 --> 00:36:10.880
+which model which model is doing better
+
+00:36:07.520 --> 00:36:13.960
+at scripts and stuff like that so um or
+
+00:36:10.880 --> 00:36:16.000
+you can put things side by side and say
+
+00:36:13.960 --> 00:36:20.720
+okay I want to find all the examples
+
+00:36:16.000 --> 00:36:21.800
+where uh chat GPT is doing much worse
+
+00:36:20.720 --> 00:36:25.280
+than GPT
+
+00:36:21.800 --> 00:36:28.240
+4 uh or like GPT 3.5 is doing much worse
+
+00:36:25.280 --> 00:36:29.680
+than gp4 and here we can see that oh in
+
+00:36:28.240 --> 00:36:31.520
+this case it's generating something in
+
+00:36:29.680 --> 00:36:34.079
+the wrong script or something like that
+
+00:36:31.520 --> 00:36:37.839
+so um there's also tooling that you can
+
+00:36:34.079 --> 00:36:40.480
+use to make this easier as
+
+00:36:37.839 --> 00:36:43.520
+well and the way uh the way you use this
+
+00:36:40.480 --> 00:36:46.079
+is you basically
+
+00:36:43.520 --> 00:36:48.000
+um uh create a pandas data frame with
+
+00:36:46.079 --> 00:36:49.680
+all of your data in it and you upload
+
+00:36:48.000 --> 00:36:52.400
+the pandas data frame with any metadata
+
+00:36:49.680 --> 00:36:54.280
+you want to use and you can uh use and I
+
+00:36:52.400 --> 00:36:56.520
+think VJ will be having a recitation on
+
+00:36:54.280 --> 00:37:02.560
+this if you're interested in taking a
+
+00:36:56.520 --> 00:37:04.680
+look cool um so that is the my part and
+
+00:37:02.560 --> 00:37:07.760
+then we'll be doing Nishant next while
+
+00:37:04.680 --> 00:37:09.480
+Nishant comes up to set up are there any
+
+00:37:07.760 --> 00:37:10.520
+questions about the thing that I talked
+
+00:37:09.480 --> 00:37:14.079
+about
+
+00:37:10.520 --> 00:37:14.079
+here yeah
+
+00:37:14.359 --> 00:37:18.200
+so that when I
+
+00:37:26.200 --> 00:37:30.079
+regular um
+
+00:37:28.160 --> 00:37:32.560
+is that does that make a difference in
+
+00:37:30.079 --> 00:37:35.400
+terms of like what we're expecting when
+
+00:37:32.560 --> 00:37:38.800
+we're evaluating the model
+
+00:37:35.400 --> 00:37:41.720
+model yeah so just to repeat the
+
+00:37:38.800 --> 00:37:43.680
+question it's a a great question so if
+
+00:37:41.720 --> 00:37:49.440
+you apply
+
+00:37:43.680 --> 00:37:49.440
+regularization um will that change the
+
+00:37:49.640 --> 00:37:54.079
+overall expectation for the model loss
+
+00:37:52.040 --> 00:37:55.680
+so I was saying loss should converge to
+
+00:37:54.079 --> 00:37:57.200
+zero once you start applying
+
+00:37:55.680 --> 00:37:59.079
+regularization or weight Decay or
+
+00:37:57.200 --> 00:38:02.640
+something like that it definitely might
+
+00:37:59.079 --> 00:38:04.520
+not converge to Z um and the reason why
+
+00:38:02.640 --> 00:38:06.520
+is because once you start applying
+
+00:38:04.520 --> 00:38:09.319
+regularization there is no zero loss
+
+00:38:06.520 --> 00:38:11.480
+solion um because in order to reduce the
+
+00:38:09.319 --> 00:38:14.960
+loss you need to make move things away
+
+00:38:11.480 --> 00:38:16.359
+move weights away from zero um but when
+
+00:38:14.960 --> 00:38:19.560
+you move weights away from zero the
+
+00:38:16.359 --> 00:38:22.200
+regularization L becomes n negative so
+
+00:38:19.560 --> 00:38:24.599
+one thing you can do however is measure
+
+00:38:22.200 --> 00:38:26.880
+the losses separately so measure the
+
+00:38:24.599 --> 00:38:27.960
+regularization component of the loss and
+
+00:38:26.880 --> 00:38:29.760
+the um
+
+00:38:27.960 --> 00:38:31.920
+the log like we had component with the
+
+00:38:29.760 --> 00:38:33.560
+loss and with any reasonable
+
+00:38:31.920 --> 00:38:35.280
+regularization and a reasonably
+
+00:38:33.560 --> 00:38:38.000
+parameterized model I do think the loss
+
+00:38:35.280 --> 00:38:39.760
+should be getting closer to Zer like the
+
+00:38:38.000 --> 00:38:41.920
+actual likely should be getting closer
+
+00:38:39.760 --> 00:38:41.920
+to
+
+00:38:42.200 --> 00:38:46.520
+zero uh you were using an extremely
+
+00:38:44.480 --> 00:38:49.240
+small model in the mini L signed though
+
+00:38:46.520 --> 00:38:53.680
+so that might make it more
+
+00:38:49.240 --> 00:38:56.440
+difficult yeah and any other
+
+00:38:53.680 --> 00:38:59.440
+things okay if not
+
+00:38:56.440 --> 00:38:59.440
+I'll
+
+00:39:13.720 --> 00:39:19.160
+all right can everyone hear
+
+00:39:15.319 --> 00:39:21.440
+me sweet okay move this it looks like
+
+00:39:19.160 --> 00:39:24.200
+I'm talking to someone instead of
+
+00:39:21.440 --> 00:39:24.200
+between both of
+
+00:39:26.359 --> 00:39:29.359
+you
+
+00:39:33.319 --> 00:39:37.680
+all right so hi everyone um I'm going to
+
+00:39:35.720 --> 00:39:39.400
+talk about model interpretability for
+
+00:39:37.680 --> 00:39:41.680
+for those who don't know me I'm one of
+
+00:39:39.400 --> 00:39:44.359
+your Tas I'm a first year PhD student
+
+00:39:41.680 --> 00:39:47.359
+working with Mona diab on model
+
+00:39:44.359 --> 00:39:47.359
+interpretability
+
+00:39:48.800 --> 00:39:55.400
+um where what do I
+
+00:39:51.839 --> 00:39:59.119
+click your your mouse should be there
+
+00:39:55.400 --> 00:40:01.599
+yeah just
+
+00:39:59.119 --> 00:40:04.160
+cool okay um
+
+00:40:01.599 --> 00:40:06.079
+so what I want you to take away if you
+
+00:40:04.160 --> 00:40:08.359
+if you fall asleep this is too boring
+
+00:40:06.079 --> 00:40:09.839
+here are sort of the two main takeaways
+
+00:40:08.359 --> 00:40:12.040
+one I want to convince you that model
+
+00:40:09.839 --> 00:40:14.720
+interpretability is important to study
+
+00:40:12.040 --> 00:40:16.720
+and two I want I want you to find this
+
+00:40:14.720 --> 00:40:18.880
+interesting um and something you want to
+
+00:40:16.720 --> 00:40:20.079
+explore more there's a bunch of details
+
+00:40:18.880 --> 00:40:21.800
+here this is going to be kind of a
+
+00:40:20.079 --> 00:40:24.599
+whirlwind tour you're not going to get
+
+00:40:21.800 --> 00:40:27.440
+super deep into anything um so hopefully
+
+00:40:24.599 --> 00:40:28.839
+this acts as a starting point um then
+
+00:40:27.440 --> 00:40:33.800
+than anything
+
+00:40:28.839 --> 00:40:37.040
+else so interpretability in AI um the
+
+00:40:33.800 --> 00:40:38.480
+the definition is it's the study of
+
+00:40:37.040 --> 00:40:40.440
+understanding the decisions that AI
+
+00:40:38.480 --> 00:40:42.640
+systems make and putting them into
+
+00:40:40.440 --> 00:40:44.280
+easily human understandable terms this
+
+00:40:42.640 --> 00:40:47.640
+can mean a lot of different things and
+
+00:40:44.280 --> 00:40:49.280
+this is often really hard um and the why
+
+00:40:47.640 --> 00:40:51.319
+is to use that understanding to
+
+00:40:49.280 --> 00:40:54.040
+iteratively better Design Systems that
+
+00:40:51.319 --> 00:40:56.240
+are better they're more more performant
+
+00:40:54.040 --> 00:40:59.240
+but also those that are more human
+
+00:40:56.240 --> 00:40:59.240
+understandable
+
+00:41:00.119 --> 00:41:06.599
+um so interpretability is this big blah
+
+00:41:03.720 --> 00:41:08.440
+but there's a bunch of other uh spheres
+
+00:41:06.599 --> 00:41:11.920
+that intersect with it this is a super
+
+00:41:08.440 --> 00:41:14.920
+incomplete list uh so bear with me the
+
+00:41:11.920 --> 00:41:16.560
+causality and data integrate with this
+
+00:41:14.920 --> 00:41:19.000
+there's aspects that are interpretable
+
+00:41:16.560 --> 00:41:20.480
+there's aspects that matter here um
+
+00:41:19.000 --> 00:41:22.400
+explainable AI is another thing that
+
+00:41:20.480 --> 00:41:24.440
+you've probably heard this sits firmly
+
+00:41:22.400 --> 00:41:27.800
+in the interpretability blob and
+
+00:41:24.440 --> 00:41:30.520
+connects with ideas and causality and uh
+
+00:41:27.800 --> 00:41:32.680
+in data too um model interpretability
+
+00:41:30.520 --> 00:41:34.200
+sits on this kind of other side of
+
+00:41:32.680 --> 00:41:37.680
+things it intersects a little bit with
+
+00:41:34.200 --> 00:41:40.000
+causality and explainable AI but uh is a
+
+00:41:37.680 --> 00:41:42.280
+little bit separate for it um and from
+
+00:41:40.000 --> 00:41:43.880
+it and mechanistic interpretability
+
+00:41:42.280 --> 00:41:45.400
+which which you've probably heard of
+
+00:41:43.880 --> 00:41:47.680
+it's gotten a lot of Buzz recently kind
+
+00:41:45.400 --> 00:41:48.880
+of sits inside of model interpretability
+
+00:41:47.680 --> 00:41:51.680
+it's a special case of model
+
+00:41:48.880 --> 00:41:53.160
+interpretability I hope the mech people
+
+00:41:51.680 --> 00:41:56.640
+agree with me
+
+00:41:53.160 --> 00:41:58.040
+but um so yeah so historically we've
+
+00:41:56.640 --> 00:42:00.880
+been dealing with really really really
+
+00:41:58.040 --> 00:42:03.680
+small models you had Bas Nets this is a
+
+00:42:00.880 --> 00:42:07.560
+this is very small model um if all these
+
+00:42:03.680 --> 00:42:10.000
+are binary variables this is uh eight
+
+00:42:07.560 --> 00:42:12.680
+total parameters and only four of which
+
+00:42:10.000 --> 00:42:14.880
+are independent uh we also used to work
+
+00:42:12.680 --> 00:42:18.160
+with linear regression a lot and in the
+
+00:42:14.880 --> 00:42:20.680
+first case that's a nice line can be two
+
+00:42:18.160 --> 00:42:23.240
+parameters the multivariate case again
+
+00:42:20.680 --> 00:42:25.880
+that's a a small number of parameters
+
+00:42:23.240 --> 00:42:27.880
+we've moved to of more things we've
+
+00:42:25.880 --> 00:42:30.400
+moved to
+
+00:42:27.880 --> 00:42:32.160
+MLPs that have larger weight matrices
+
+00:42:30.400 --> 00:42:33.920
+but all these are kind of digestible and
+
+00:42:32.160 --> 00:42:37.200
+interpretable so the interpretability
+
+00:42:33.920 --> 00:42:40.160
+world was sort of uh not super concerned
+
+00:42:37.200 --> 00:42:41.280
+with large ginormous things but we're
+
+00:42:40.160 --> 00:42:44.800
+not there
+
+00:42:41.280 --> 00:42:47.000
+anymore uh this is a language model this
+
+00:42:44.800 --> 00:42:50.839
+is part of still part of a language
+
+00:42:47.000 --> 00:42:51.960
+model now it's getting more and more and
+
+00:42:50.839 --> 00:42:55.119
+more
+
+00:42:51.960 --> 00:42:57.920
+hairing and this is just not
+
+00:42:55.119 --> 00:43:00.520
+interpretable um I mentioned
+
+00:42:57.920 --> 00:43:03.280
+on on the first day of class that I hate
+
+00:43:00.520 --> 00:43:05.240
+when we update parameters of models also
+
+00:43:03.280 --> 00:43:07.720
+hate when models are this big and this
+
+00:43:05.240 --> 00:43:10.000
+is a six layer Transformer this is way
+
+00:43:07.720 --> 00:43:15.920
+smaller than basically anything that we
+
+00:43:10.000 --> 00:43:18.040
+have um and this makes things very very
+
+00:43:15.920 --> 00:43:20.920
+uninterpretable um so we'll talk about
+
+00:43:18.040 --> 00:43:22.880
+one one way that people sort of uh five
+
+00:43:20.920 --> 00:43:24.599
+years ago started addressing this
+
+00:43:22.880 --> 00:43:25.680
+problem and this is and this is the idea
+
+00:43:24.599 --> 00:43:28.000
+of
+
+00:43:25.680 --> 00:43:30.880
+probing so how do we make sense of a
+
+00:43:28.000 --> 00:43:35.160
+giant model this is one way so we take
+
+00:43:30.880 --> 00:43:38.200
+our giant model we cut the top off
+
+00:43:35.160 --> 00:43:40.520
+basically um and now we have this thing
+
+00:43:38.200 --> 00:43:42.119
+we stick a probe which actually in a lot
+
+00:43:40.520 --> 00:43:44.559
+of cases looks very similar to a
+
+00:43:42.119 --> 00:43:47.280
+language modeling head uh usually it's a
+
+00:43:44.559 --> 00:43:51.640
+small two layer or one layer
+
+00:43:47.280 --> 00:43:54.319
+MLP um and we basically treat the model
+
+00:43:51.640 --> 00:43:56.760
+as something that uh that exists and we
+
+00:43:54.319 --> 00:44:00.240
+only really care about the output of of
+
+00:43:56.760 --> 00:44:03.240
+the model so more specifically what is a
+
+00:44:00.240 --> 00:44:05.720
+probe it's a classifier this this green
+
+00:44:03.240 --> 00:44:07.680
+thing here uh that is specifically
+
+00:44:05.720 --> 00:44:09.200
+trained to predict some specific
+
+00:44:07.680 --> 00:44:11.480
+property from the pre-trained models
+
+00:44:09.200 --> 00:44:16.440
+representations
+
+00:44:11.480 --> 00:44:18.480
+alone so um in 2019 Ian Tenny and folks
+
+00:44:16.440 --> 00:44:21.319
+introduced Edge probing so this is a
+
+00:44:18.480 --> 00:44:23.240
+general method um it works to probe
+
+00:44:21.319 --> 00:44:27.559
+different types of information out of a
+
+00:44:23.240 --> 00:44:29.960
+model so this bottom part here uh yeah
+
+00:44:27.559 --> 00:44:33.160
+this bottom part here it you pass it in
+
+00:44:29.960 --> 00:44:36.520
+a sequence you pass it into a model this
+
+00:44:33.160 --> 00:44:38.839
+is Burt in their experiments often uh
+
+00:44:36.520 --> 00:44:40.960
+and that outputs a set of contextual
+
+00:44:38.839 --> 00:44:44.359
+vectors these contextual vectors can be
+
+00:44:40.960 --> 00:44:45.920
+at any layer um often it's near the
+
+00:44:44.359 --> 00:44:49.280
+often it's near the top but we'll talk
+
+00:44:45.920 --> 00:44:51.079
+about uh the the fact that this can work
+
+00:44:49.280 --> 00:44:53.359
+kind of across layers and different
+
+00:44:51.079 --> 00:44:55.599
+layers and code different information
+
+00:44:53.359 --> 00:44:58.960
+and on top of this you have this MLP
+
+00:44:55.599 --> 00:45:02.480
+that you train to Output a prediction
+
+00:44:58.960 --> 00:45:05.599
+your model is always always fixed um in
+
+00:45:02.480 --> 00:45:08.079
+these cases so you can do things like
+
+00:45:05.599 --> 00:45:09.880
+part of speech tagging where each
+
+00:45:08.079 --> 00:45:12.400
+specific word you try to determine what
+
+00:45:09.880 --> 00:45:16.640
+its part of speech is and in that case
+
+00:45:12.400 --> 00:45:18.000
+this these S1 and S2 spans here uh only
+
+00:45:16.640 --> 00:45:19.440
+one of them is active because you're
+
+00:45:18.000 --> 00:45:21.440
+predicting for every single
+
+00:45:19.440 --> 00:45:23.240
+contextualized Vector you're predicting
+
+00:45:21.440 --> 00:45:25.359
+whether that thing is a noun or a verb
+
+00:45:23.240 --> 00:45:27.440
+or something like this you can have
+
+00:45:25.359 --> 00:45:29.599
+other sorts of tasks too like in ailment
+
+00:45:27.440 --> 00:45:32.520
+where you have two sequences and two
+
+00:45:29.599 --> 00:45:35.079
+spans um and you use the embeddings for
+
+00:45:32.520 --> 00:45:37.359
+those spans um for like sentence one and
+
+00:45:35.079 --> 00:45:39.319
+sentence two you pull them together in
+
+00:45:37.359 --> 00:45:43.359
+some way and then you pass them to this
+
+00:45:39.319 --> 00:45:47.480
+MLP and you see whether the MLP can uh
+
+00:45:43.359 --> 00:45:49.680
+solve that test so they did this uh in
+
+00:45:47.480 --> 00:45:52.559
+another paper uh Bert rediscovers the
+
+00:45:49.680 --> 00:45:54.280
+NLP Pipeline and this there's a lot
+
+00:45:52.559 --> 00:45:57.079
+going on in this figure the the only
+
+00:45:54.280 --> 00:45:59.599
+major thing here to take away um is
+
+00:45:57.079 --> 00:46:02.720
+these numbers that are in this like pink
+
+00:45:59.599 --> 00:46:05.359
+purple color um so these are a bunch of
+
+00:46:02.720 --> 00:46:07.960
+different uh properties supp part of
+
+00:46:05.359 --> 00:46:11.319
+speech uh and and a bunch of other
+
+00:46:07.960 --> 00:46:13.520
+things um and what they basically find
+
+00:46:11.319 --> 00:46:15.640
+is that at earlier layers in the model
+
+00:46:13.520 --> 00:46:18.760
+the things that are closer to the Token
+
+00:46:15.640 --> 00:46:21.480
+level representation are more um
+
+00:46:18.760 --> 00:46:23.400
+extractable using a probe and the things
+
+00:46:21.480 --> 00:46:26.440
+that require more contextualized
+
+00:46:23.400 --> 00:46:29.440
+information um is extractable later from
+
+00:46:26.440 --> 00:46:32.359
+later layers in the model and so here's
+
+00:46:29.440 --> 00:46:34.599
+sort of a brief uh description of what
+
+00:46:32.359 --> 00:46:37.599
+these tasks are so the ones on the
+
+00:46:34.599 --> 00:46:40.040
+bottom are more semantic more
+
+00:46:37.599 --> 00:46:42.040
+contextualized like uh semantic prot
+
+00:46:40.040 --> 00:46:43.880
+roles and relation relation
+
+00:46:42.040 --> 00:46:45.839
+classification and then the first few
+
+00:46:43.880 --> 00:46:48.200
+are more you know chunking in part of
+
+00:46:45.839 --> 00:46:51.880
+speech tagging and um dependency
+
+00:46:48.200 --> 00:46:51.880
+labeling in these in these sorts of
+
+00:46:52.040 --> 00:46:57.200
+tests um so there's a bunch of issues
+
+00:46:54.480 --> 00:46:59.520
+with probing um and there aren't many
+
+00:46:57.200 --> 00:47:03.559
+probing papers now as there were many
+
+00:46:59.520 --> 00:47:05.960
+years ago um and so if your probe let's
+
+00:47:03.559 --> 00:47:07.960
+say your probe
+
+00:47:05.960 --> 00:47:09.920
+works it's possible that the
+
+00:47:07.960 --> 00:47:12.200
+representation actually encodes that
+
+00:47:09.920 --> 00:47:14.520
+information it's also possible that it
+
+00:47:12.200 --> 00:47:16.359
+doesn't and the probe solved the task by
+
+00:47:14.520 --> 00:47:18.119
+itself uh keep in mind that you're
+
+00:47:16.359 --> 00:47:20.640
+learning this probe you're training this
+
+00:47:18.119 --> 00:47:22.720
+Probe on labeled data uh let's say your
+
+00:47:20.640 --> 00:47:24.599
+probe doesn't work does that tell you
+
+00:47:22.720 --> 00:47:27.119
+anything Maybe not maybe the
+
+00:47:24.599 --> 00:47:30.280
+representation lacks the information or
+
+00:47:27.119 --> 00:47:31.800
+maybe your probe doesn't doesn't
+
+00:47:30.280 --> 00:47:33.800
+actually isn't actually able to
+
+00:47:31.800 --> 00:47:35.240
+disentangle that information from your
+
+00:47:33.800 --> 00:47:36.720
+representation maybe the probe is not
+
+00:47:35.240 --> 00:47:38.359
+the right function class maybe you
+
+00:47:36.720 --> 00:47:40.839
+poorly trained your probe there's
+
+00:47:38.359 --> 00:47:42.280
+hyperparameters for your probe so often
+
+00:47:40.839 --> 00:47:43.000
+times your probe doesn't give you that
+
+00:47:42.280 --> 00:47:46.119
+much
+
+00:47:43.000 --> 00:47:49.040
+information there's more problems too so
+
+00:47:46.119 --> 00:47:50.800
+often we want to probe tasks themselves
+
+00:47:49.040 --> 00:47:53.240
+and that requires a lot of supervised
+
+00:47:50.800 --> 00:47:55.880
+data um but we can't collect a lot of
+
+00:47:53.240 --> 00:47:58.440
+supervised data so we collect some of it
+
+00:47:55.880 --> 00:48:00.040
+and then that instead produces this
+
+00:47:58.440 --> 00:48:02.480
+convenient sample that we have that's a
+
+00:48:00.040 --> 00:48:04.119
+data set that is a convenient sample of
+
+00:48:02.480 --> 00:48:07.000
+your task so really what you're probing
+
+00:48:04.119 --> 00:48:10.040
+is the data set and so with all these
+
+00:48:07.000 --> 00:48:11.800
+limitations it's it's fallen out of
+
+00:48:10.040 --> 00:48:13.599
+favor a little bit it's still very very
+
+00:48:11.800 --> 00:48:16.400
+useful but it's fallen out of favor as
+
+00:48:13.599 --> 00:48:20.000
+like a core model interpretability
+
+00:48:16.400 --> 00:48:22.160
+idea um also probes designed in this way
+
+00:48:20.000 --> 00:48:26.079
+are correlated they're correlative not
+
+00:48:22.160 --> 00:48:27.880
+really causitive so your your underlying
+
+00:48:26.079 --> 00:48:29.640
+model is trained in a specific way all
+
+00:48:27.880 --> 00:48:31.359
+of that information is disentangled and
+
+00:48:29.640 --> 00:48:32.920
+kind of thrown away and you're only
+
+00:48:31.359 --> 00:48:34.599
+looking at the output representation and
+
+00:48:32.920 --> 00:48:36.559
+you're saying is my output
+
+00:48:34.599 --> 00:48:39.200
+representation correlated to the thing
+
+00:48:36.559 --> 00:48:42.400
+that I'm training this probe for there's
+
+00:48:39.200 --> 00:48:44.960
+no notion of intervening on this lat and
+
+00:48:42.400 --> 00:48:46.559
+space there's no notion of of causation
+
+00:48:44.960 --> 00:48:49.119
+really so you're just seeing whether
+
+00:48:46.559 --> 00:48:52.559
+your representation is correlated with
+
+00:48:49.119 --> 00:48:54.480
+your property that you're probing for um
+
+00:48:52.559 --> 00:48:56.200
+and with these limitations the
+
+00:48:54.480 --> 00:48:58.720
+community's moved a little bit away from
+
+00:48:56.200 --> 00:48:58.720
+this area
+
+00:48:59.040 --> 00:49:02.200
+uh there's a bunch of other probing
+
+00:49:00.240 --> 00:49:04.920
+works so a bunch of people aim to solve
+
+00:49:02.200 --> 00:49:06.000
+a bunch of these problems um and uh for
+
+00:49:04.920 --> 00:49:09.200
+the sake of time I'm not going to go
+
+00:49:06.000 --> 00:49:12.599
+into all of these but uh I'd encourage
+
+00:49:09.200 --> 00:49:14.000
+you to look into these they for for some
+
+00:49:12.599 --> 00:49:17.319
+of these problems they're able to
+
+00:49:14.000 --> 00:49:19.520
+control for um control for like the
+
+00:49:17.319 --> 00:49:22.200
+complexity of the of the probe and
+
+00:49:19.520 --> 00:49:24.359
+things like this um but even despite
+
+00:49:22.200 --> 00:49:25.720
+that probing is sort of slowly kind of
+
+00:49:24.359 --> 00:49:28.160
+falling out of
+
+00:49:25.720 --> 00:49:29.640
+favor uh so before I move into model
+
+00:49:28.160 --> 00:49:31.920
+interpretability are there any questions
+
+00:49:29.640 --> 00:49:31.920
+on
+
+00:49:35.520 --> 00:49:40.599
+probing all right so what is model
+
+00:49:38.680 --> 00:49:44.000
+interpretability so this is my
+
+00:49:40.599 --> 00:49:45.400
+definition here uh this is the study of
+
+00:49:44.000 --> 00:49:46.599
+understanding the internals of models
+
+00:49:45.400 --> 00:49:49.079
+for example their weights and
+
+00:49:46.599 --> 00:49:51.160
+activations putting those insights in
+
+00:49:49.079 --> 00:49:53.319
+human intelligible terms and using that
+
+00:49:51.160 --> 00:49:55.920
+insight to both patch current models and
+
+00:49:53.319 --> 00:49:57.359
+develop better ones for not sort of able
+
+00:49:55.920 --> 00:49:58.760
+to do both of these things patching
+
+00:49:57.359 --> 00:50:00.160
+current models and develop better ones
+
+00:49:58.760 --> 00:50:02.440
+we're kind of doing interpretability for
+
+00:50:00.160 --> 00:50:04.960
+interpretability sake that's nice and
+
+00:50:02.440 --> 00:50:08.079
+fun but it's not as applicable for the
+
+00:50:04.960 --> 00:50:09.720
+for the community so you've probably
+
+00:50:08.079 --> 00:50:12.240
+heard of the term mechanistic
+
+00:50:09.720 --> 00:50:14.480
+interpretability it's in my opinion a
+
+00:50:12.240 --> 00:50:16.559
+subfield of model interpretability and
+
+00:50:14.480 --> 00:50:19.319
+this is sort of my definition I it
+
+00:50:16.559 --> 00:50:21.440
+aligns reasonably well to the core
+
+00:50:19.319 --> 00:50:22.720
+mechanistic interpretability people um
+
+00:50:21.440 --> 00:50:24.880
+but it's the study of reverse
+
+00:50:22.720 --> 00:50:26.280
+engineering parametric models often
+
+00:50:24.880 --> 00:50:28.839
+neural networks because that's what we
+
+00:50:26.280 --> 00:50:31.400
+use from their learned weights into more
+
+00:50:28.839 --> 00:50:32.839
+human interpretable algorithmic units uh
+
+00:50:31.400 --> 00:50:36.839
+and often they call these things
+
+00:50:32.839 --> 00:50:39.440
+circuits um and and these are basically
+
+00:50:36.839 --> 00:50:42.880
+functions that uh you can describe in a
+
+00:50:39.440 --> 00:50:45.000
+human interpretable way that sit inside
+
+00:50:42.880 --> 00:50:46.760
+models um there's a bunch of notable
+
+00:50:45.000 --> 00:50:50.720
+work again for the sake of time I'm
+
+00:50:46.760 --> 00:50:54.319
+going to just briefly talk about them
+
+00:50:50.720 --> 00:50:56.839
+um so the first one is they they look
+
+00:50:54.319 --> 00:50:58.440
+into analyzing small MLPs and
+
+00:50:56.839 --> 00:51:01.400
+Transformers to build out the intuition
+
+00:50:58.440 --> 00:51:04.119
+of what circuits exist um and this a lot
+
+00:51:01.400 --> 00:51:06.559
+of this work came out of earlier work on
+
+00:51:04.119 --> 00:51:08.480
+on lstms and doing similar sorts of
+
+00:51:06.559 --> 00:51:11.880
+things with with
+
+00:51:08.480 --> 00:51:14.319
+lstms um and they find a bunch of things
+
+00:51:11.880 --> 00:51:15.839
+one thing that they find is this idea of
+
+00:51:14.319 --> 00:51:19.599
+induction heads and these induction
+
+00:51:15.839 --> 00:51:21.760
+heads they say is sort of sort of helps
+
+00:51:19.599 --> 00:51:24.680
+prove why models can do in context
+
+00:51:21.760 --> 00:51:26.599
+learning so so an induction head is
+
+00:51:24.680 --> 00:51:28.839
+something that it's it's a specific
+
+00:51:26.599 --> 00:51:32.440
+attention head that kind of allows you
+
+00:51:28.839 --> 00:51:35.599
+to um when given a prefix allow you to
+
+00:51:32.440 --> 00:51:37.559
+kind of copy the necessary resulting
+
+00:51:35.599 --> 00:51:39.640
+token from the underlying training data
+
+00:51:37.559 --> 00:51:41.720
+that the model seen before so in context
+
+00:51:39.640 --> 00:51:44.599
+learning what you generally provide is
+
+00:51:41.720 --> 00:51:46.440
+some sort of prefix and then you uh
+
+00:51:44.599 --> 00:51:48.480
+provide some example and hopefully you
+
+00:51:46.440 --> 00:51:51.040
+know you can classify the thing or
+
+00:51:48.480 --> 00:51:53.280
+something like this um it's it's saying
+
+00:51:51.040 --> 00:51:56.200
+that there's these attention heads
+
+00:51:53.280 --> 00:51:59.400
+Loosely uh that exist that are able to
+
+00:51:56.200 --> 00:52:00.680
+copy unearth that information um for for
+
+00:51:59.400 --> 00:52:03.319
+a specific
+
+00:52:00.680 --> 00:52:07.200
+context um other things that they' that
+
+00:52:03.319 --> 00:52:09.880
+they've done is um on neurons so uh this
+
+00:52:07.200 --> 00:52:13.160
+poly semanticity so what this this kind
+
+00:52:09.880 --> 00:52:15.240
+of means is that your your neuron is a
+
+00:52:13.160 --> 00:52:18.000
+uh you have a set of neurons in your
+
+00:52:15.240 --> 00:52:20.880
+activation space so let's say at layer
+
+00:52:18.000 --> 00:52:23.200
+10 in your model you have an output um
+
+00:52:20.880 --> 00:52:26.280
+and so your activations is let's say a
+
+00:52:23.200 --> 00:52:28.400
+thousand dimensional here those each of
+
+00:52:26.280 --> 00:52:31.319
+those thousand individual neurons may
+
+00:52:28.400 --> 00:52:35.839
+represent more than one specific
+
+00:52:31.319 --> 00:52:37.839
+feature um and so they they talk about
+
+00:52:35.839 --> 00:52:41.280
+this in that context and this is kind of
+
+00:52:37.839 --> 00:52:43.240
+a theory but you can think about um
+
+00:52:41.280 --> 00:52:46.359
+trying to process
+
+00:52:43.240 --> 00:52:49.400
+input and when you're processing a a
+
+00:52:46.359 --> 00:52:50.960
+vocab of size 50,000 or 250,000 at some
+
+00:52:49.400 --> 00:52:52.359
+point in the model we're actually
+
+00:52:50.960 --> 00:52:55.400
+compressing it down to the hidden
+
+00:52:52.359 --> 00:52:58.119
+Dimension and so in some cases that
+
+00:52:55.400 --> 00:53:00.319
+looks like you're going to compress a
+
+00:52:58.119 --> 00:53:03.440
+much richer feature representation down
+
+00:53:00.319 --> 00:53:06.359
+into a smaller set of neurons so it is
+
+00:53:03.440 --> 00:53:08.319
+reasonable to believe that um a specific
+
+00:53:06.359 --> 00:53:10.799
+neuron will represent multiple of those
+
+00:53:08.319 --> 00:53:15.480
+features and given the structure of our
+
+00:53:10.799 --> 00:53:18.720
+weight matrices um it it is the case
+
+00:53:15.480 --> 00:53:21.839
+that if they are representing more
+
+00:53:18.720 --> 00:53:23.960
+features than uh number of elements in
+
+00:53:21.839 --> 00:53:26.000
+the actual or number of neurons in the
+
+00:53:23.960 --> 00:53:28.680
+activation space then many of these
+
+00:53:26.000 --> 00:53:30.880
+features linearly dependent and so we're
+
+00:53:28.680 --> 00:53:35.400
+not really able to utilize them that
+
+00:53:30.880 --> 00:53:37.960
+well um they they they talk about this
+
+00:53:35.400 --> 00:53:42.200
+they don't talk about this in the the
+
+00:53:37.960 --> 00:53:44.799
+most uh the the best way but uh it seems
+
+00:53:42.200 --> 00:53:48.040
+kind of clear to me that um since you
+
+00:53:44.799 --> 00:53:50.880
+have embedding matrices that are um not
+
+00:53:48.040 --> 00:53:53.599
+square that you're that these neurons
+
+00:53:50.880 --> 00:53:56.400
+have to exist um and they have to
+
+00:53:53.599 --> 00:53:59.200
+incorporate multiple features at once m
+
+00:53:56.400 --> 00:54:02.559
+multiple redundant features at
+
+00:53:59.200 --> 00:54:04.680
+once um so before I move on to the rest
+
+00:54:02.559 --> 00:54:07.839
+of model interpretability any questions
+
+00:54:04.680 --> 00:54:07.839
+about mechanistic
+
+00:54:09.880 --> 00:54:12.880
+interpretability
+
+00:54:21.480 --> 00:54:28.040
+yeah so most of their studies are for uh
+
+00:54:24.920 --> 00:54:29.720
+a very small set of of of models and
+
+00:54:28.040 --> 00:54:32.040
+most of these are old GPT models there
+
+00:54:29.720 --> 00:54:34.160
+have been a few works like in the last
+
+00:54:32.040 --> 00:54:36.760
+couple of months on doing this for the
+
+00:54:34.160 --> 00:54:39.720
+Llama based models um it seems like this
+
+00:54:36.760 --> 00:54:42.040
+is a general more General phenomenal for
+
+00:54:39.720 --> 00:54:43.760
+for language models it also is the case
+
+00:54:42.040 --> 00:54:46.839
+that certain attention heads specialize
+
+00:54:43.760 --> 00:54:49.480
+and talk about them a little bit um in
+
+00:54:46.839 --> 00:54:51.599
+in the activations part um but yeah
+
+00:54:49.480 --> 00:54:53.799
+there's they're not like all attention
+
+00:54:51.599 --> 00:54:56.400
+heads aren't created equal they start
+
+00:54:53.799 --> 00:55:00.280
+this way and it seems to be a general
+
+00:54:56.400 --> 00:55:01.799
+princip of and one one other thing I you
+
+00:55:00.280 --> 00:55:04.520
+might know about this better than I do
+
+00:55:01.799 --> 00:55:06.520
+but I think there are some preliminary
+
+00:55:04.520 --> 00:55:09.160
+words that say that Transformers seem to
+
+00:55:06.520 --> 00:55:11.720
+be particularly good at doing things
+
+00:55:09.160 --> 00:55:15.160
+like induction heads compared
+
+00:55:11.720 --> 00:55:17.200
+to uh H for current models and there was
+
+00:55:15.160 --> 00:55:20.720
+a paper really recently about comparing
+
+00:55:17.200 --> 00:55:23.599
+like Mamba and um in Transformer based
+
+00:55:20.720 --> 00:55:26.400
+models Mamba being a uh kind of more
+
+00:55:23.599 --> 00:55:30.280
+like closer to a with network which we
+
+00:55:26.400 --> 00:55:33.119
+also going to talk about us but um so I
+
+00:55:30.280 --> 00:55:37.319
+I think there's some indication that
+
+00:55:33.119 --> 00:55:39.920
+Transformers actually kind of are key or
+
+00:55:37.319 --> 00:55:43.680
+are at least like better at kind of in
+
+00:55:39.920 --> 00:55:46.760
+Contex learning than otheres are so
+
+00:55:43.680 --> 00:55:48.920
+there is some
+
+00:55:46.760 --> 00:55:50.839
+interesting implications of that which
+
+00:55:48.920 --> 00:55:53.240
+is like well if Transformers are good
+
+00:55:50.839 --> 00:55:57.359
+what's better than Transformer yeah like
+
+00:55:53.240 --> 00:55:58.799
+naturally learning this s of thing so um
+
+00:55:57.359 --> 00:56:00.720
+they're good at yeah they're like really
+
+00:55:58.799 --> 00:56:04.039
+good at copying and like maintaining
+
+00:56:00.720 --> 00:56:06.799
+information like more so um and yeah I
+
+00:56:04.039 --> 00:56:08.200
+think it'd be cool to like be able to I
+
+00:56:06.799 --> 00:56:09.839
+don't know how to do this but be able to
+
+00:56:08.200 --> 00:56:11.440
+extract that kind of information like
+
+00:56:09.839 --> 00:56:13.359
+what of the Transformers is actually
+
+00:56:11.440 --> 00:56:15.119
+helping it do this copying mechanism or
+
+00:56:13.359 --> 00:56:17.799
+like being a better in context learner
+
+00:56:15.119 --> 00:56:20.039
+then we can develop a better structure a
+
+00:56:17.799 --> 00:56:23.119
+slightly better structure than than a
+
+00:56:20.039 --> 00:56:26.000
+Transformer um hopefully someone comes
+
+00:56:23.119 --> 00:56:28.240
+up with that soon but cool any other
+
+00:56:26.000 --> 00:56:28.240
+question
+
+00:56:29.799 --> 00:56:34.359
+questions all right so let's move into
+
+00:56:32.240 --> 00:56:35.880
+model interpretability so there are
+
+00:56:34.359 --> 00:56:37.480
+weights and their activations I
+
+00:56:35.880 --> 00:56:39.160
+mentioned these are these are the two
+
+00:56:37.480 --> 00:56:41.119
+things these are the two things that
+
+00:56:39.160 --> 00:56:43.440
+we're going to look at so what can you
+
+00:56:41.119 --> 00:56:45.480
+do with the weights of an RD train model
+
+00:56:43.440 --> 00:56:47.799
+really you can just edit them and then
+
+00:56:45.480 --> 00:56:49.200
+kind of see what happens activations
+
+00:56:47.799 --> 00:56:51.240
+similarly you can look at the
+
+00:56:49.200 --> 00:56:52.720
+activations for different inputs you can
+
+00:56:51.240 --> 00:56:54.520
+poke them with a stick and see what
+
+00:56:52.720 --> 00:56:56.359
+happens a lot of my research is poking
+
+00:56:54.520 --> 00:56:58.559
+models with a stick and looking at the
+
+00:56:56.359 --> 00:57:00.920
+activations it's like predominantly what
+
+00:56:58.559 --> 00:57:02.240
+I've done so we'll talk about that um
+
+00:57:00.920 --> 00:57:04.359
+and the technical term for this is
+
+00:57:02.240 --> 00:57:06.599
+intervening on them by adding some
+
+00:57:04.359 --> 00:57:07.839
+vector or other sort of manipulation to
+
+00:57:06.599 --> 00:57:09.440
+the lat space but really what you're
+
+00:57:07.839 --> 00:57:13.960
+doing is like
+
+00:57:09.440 --> 00:57:17.599
+Pok um so when you look at weights uh
+
+00:57:13.960 --> 00:57:19.920
+one one class of methods or or area is
+
+00:57:17.599 --> 00:57:21.920
+on model editing fine-tuning is like the
+
+00:57:19.920 --> 00:57:23.480
+most extreme version of model editing
+
+00:57:21.920 --> 00:57:26.599
+usually these things are much more
+
+00:57:23.480 --> 00:57:29.640
+targeted um so in the model editing sort
+
+00:57:26.599 --> 00:57:32.160
+of landscape your goal or your target is
+
+00:57:29.640 --> 00:57:35.119
+you have a concept or a specific fact
+
+00:57:32.160 --> 00:57:37.440
+that needs to be changed in the model um
+
+00:57:35.119 --> 00:57:39.640
+and your approach here is you update or
+
+00:57:37.440 --> 00:57:41.359
+edit the weights of the model to edit
+
+00:57:39.640 --> 00:57:43.640
+the model's belief of that factor
+
+00:57:41.359 --> 00:57:45.599
+concept and ideally you do this without
+
+00:57:43.640 --> 00:57:47.319
+changing any of the other behavior of
+
+00:57:45.599 --> 00:57:49.760
+the model so for example let's say
+
+00:57:47.319 --> 00:57:51.920
+you're trying to say that Graham is no
+
+00:57:49.760 --> 00:57:54.559
+longer a professor at CMU but is a
+
+00:57:51.920 --> 00:57:57.319
+professor at Stanford you don't want
+
+00:57:54.559 --> 00:57:59.960
+every single person at CMU to now be a
+
+00:57:57.319 --> 00:58:02.920
+professor or uh now be affiliated with
+
+00:57:59.960 --> 00:58:07.839
+Stanford right um gr pleas not to
+
+00:58:02.920 --> 00:58:09.039
+Stanford um so here's one approach paper
+
+00:58:07.839 --> 00:58:11.720
+that came out a couple years ago there's
+
+00:58:09.039 --> 00:58:13.559
+a lot of work down here uh in in the
+
+00:58:11.720 --> 00:58:15.799
+model editing World I'll give you sort
+
+00:58:13.559 --> 00:58:17.440
+of a really brief overview of this but
+
+00:58:15.799 --> 00:58:20.520
+basically they have facts that they want
+
+00:58:17.440 --> 00:58:22.400
+to they want to manipulate um so for
+
+00:58:20.520 --> 00:58:24.680
+example the the example that they give
+
+00:58:22.400 --> 00:58:26.640
+in the figure is they want to associate
+
+00:58:24.680 --> 00:58:30.960
+the Space Needle with Paris the Space
+
+00:58:26.640 --> 00:58:32.520
+Needle is a a cool needle in in Seattle
+
+00:58:30.960 --> 00:58:36.000
+has nothing to do with Paris but Paris
+
+00:58:32.520 --> 00:58:38.400
+also has a tower so it's close um so
+
+00:58:36.000 --> 00:58:40.920
+they use causal tracing to isolate the
+
+00:58:38.400 --> 00:58:43.839
+causal effect uh of the individual
+
+00:58:40.920 --> 00:58:45.799
+hidden States for this fact so they
+
+00:58:43.839 --> 00:58:47.839
+basically continuously perturb the input
+
+00:58:45.799 --> 00:58:49.760
+do a bunch of forward passes and
+
+00:58:47.839 --> 00:58:51.720
+sequentially find the specific hidden
+
+00:58:49.760 --> 00:58:55.280
+states that are associated kind of with
+
+00:58:51.720 --> 00:58:56.839
+this fact um then they make an edit and
+
+00:58:55.280 --> 00:58:59.119
+their edit
+
+00:58:56.839 --> 00:59:02.039
+uh looks like this thing on the right um
+
+00:58:59.119 --> 00:59:05.280
+so they treat this pair Space Needle and
+
+00:59:02.039 --> 00:59:07.240
+Paris as this uh key value pair where
+
+00:59:05.280 --> 00:59:10.359
+Space Needle is the key you pass this
+
+00:59:07.240 --> 00:59:12.480
+into um into this weight Matrix this
+
+00:59:10.359 --> 00:59:14.640
+original part of the model you want this
+
+00:59:12.480 --> 00:59:16.599
+now instead of outputting Seattle to
+
+00:59:14.640 --> 00:59:19.119
+Output Paris and they have some nice
+
+00:59:16.599 --> 00:59:21.599
+math and a closed form solution to to
+
+00:59:19.119 --> 00:59:23.880
+identify this this is super expensive
+
+00:59:21.599 --> 00:59:25.359
+because they have to the causal tracing
+
+00:59:23.880 --> 00:59:27.680
+part have to do a bunch of forward
+
+00:59:25.359 --> 00:59:30.680
+passes um and they make this a little
+
+00:59:27.680 --> 00:59:33.480
+bit better in future future work they
+
+00:59:30.680 --> 00:59:37.920
+also do sort of a more
+
+00:59:33.480 --> 00:59:40.160
+comprehensive um edit um so these are
+
+00:59:37.920 --> 00:59:44.599
+kind like some of the things you can do
+
+00:59:40.160 --> 00:59:46.799
+um I'm less excited about model editing
+
+00:59:44.599 --> 00:59:49.039
+um there's there's some work on model
+
+00:59:46.799 --> 00:59:51.319
+editing sort of it's it's hard to
+
+00:59:49.039 --> 00:59:53.160
+control what other things break there's
+
+00:59:51.319 --> 00:59:56.240
+a and there's some work with when you
+
+00:59:53.160 --> 01:00:00.000
+edit a specific fact things start being
+
+00:59:56.240 --> 01:00:02.680
+weird and being biased in other ways um
+
+01:00:00.000 --> 01:00:05.760
+and so
+
+01:00:02.680 --> 01:00:09.119
+yeah do all kind of seual information
+
+01:00:05.760 --> 01:00:11.880
+like X is and Y would they Alles to the
+
+01:00:09.119 --> 01:00:14.319
+same layer is it just with the
+
+01:00:11.880 --> 01:00:16.920
+specific for this specific example it
+
+01:00:14.319 --> 01:00:19.039
+looks at this specific point uh for
+
+01:00:16.920 --> 01:00:21.039
+every example they'll probably find
+
+01:00:19.039 --> 01:00:22.119
+different regions in a different degree
+
+01:00:21.039 --> 01:00:25.680
+of
+
+01:00:22.119 --> 01:00:27.960
+manipulation um and yeah that it gets a
+
+01:00:25.680 --> 01:00:30.920
+little unprincipled kind of quickly it's
+
+01:00:27.960 --> 01:00:33.000
+not like they're able to find you know a
+
+01:00:30.920 --> 01:00:35.680
+specific attention head that or a
+
+01:00:33.000 --> 01:00:38.240
+specific layer or specific weight Matrix
+
+01:00:35.680 --> 01:00:42.400
+that corresponds to like
+
+01:00:38.240 --> 01:00:46.720
+all yeah relations of a specific
+
+01:00:42.400 --> 01:00:49.160
+type any for questions for yeah this is
+
+01:00:46.720 --> 01:00:51.119
+actually just a question if you know um
+
+01:00:49.160 --> 01:00:53.200
+it seems like more frequent facts might
+
+01:00:51.119 --> 01:00:55.240
+appear in both places in the model is do
+
+01:00:53.200 --> 01:00:59.280
+you know if that's actually the I have
+
+01:00:55.240 --> 01:01:02.440
+no idea but uh I would imagine that um
+
+01:00:59.280 --> 01:01:06.240
+it probably could occur in more places
+
+01:01:02.440 --> 01:01:08.160
+but also um a lot of the information is
+
+01:01:06.240 --> 01:01:10.119
+redundant anyway in the model especially
+
+01:01:08.160 --> 01:01:11.720
+for larger models so you might have to
+
+01:01:10.119 --> 01:01:13.599
+make targeted interventions in multiple
+
+01:01:11.720 --> 01:01:15.480
+places but it's possible that one
+
+01:01:13.599 --> 01:01:17.680
+intervention in one place sufficiently
+
+01:01:15.480 --> 01:01:21.039
+destroys like contextualized information
+
+01:01:17.680 --> 01:01:22.680
+in other places if it's close um it
+
+01:01:21.039 --> 01:01:24.839
+depends on how big this intervention is
+
+01:01:22.680 --> 01:01:28.200
+if it's like hitting it with a hammer
+
+01:01:24.839 --> 01:01:30.520
+rather than some like nice fine grain
+
+01:01:28.200 --> 01:01:33.359
+thing but that'd be a good be a good
+
+01:01:30.520 --> 01:01:36.839
+experiment to see
+
+01:01:33.359 --> 01:01:36.839
+um any other
+
+01:01:37.240 --> 01:01:41.559
+questions all right so we'll move into
+
+01:01:39.760 --> 01:01:43.680
+the stuff that I'm most familiar with
+
+01:01:41.559 --> 01:01:46.319
+and some of my work so looking at
+
+01:01:43.680 --> 01:01:48.319
+activations um so this is this is work
+
+01:01:46.319 --> 01:01:50.480
+I've been doing for a while uh this idea
+
+01:01:48.319 --> 01:01:52.799
+of steering vectors so I mentioned I
+
+01:01:50.480 --> 01:01:54.480
+poke models so it's thick steering
+
+01:01:52.799 --> 01:01:57.000
+Vector is that thick so it's basically a
+
+01:01:54.480 --> 01:01:59.000
+fix length vector that steers a language
+
+01:01:57.000 --> 01:02:00.920
+model to generate a specific sequence
+
+01:01:59.000 --> 01:02:02.720
+exactly when added to the hidden sites
+
+01:02:00.920 --> 01:02:06.319
+of a model at a specific
+
+01:02:02.720 --> 01:02:09.000
+location um and I'll I'll read this
+
+01:02:06.319 --> 01:02:11.400
+again it's there's a very like specific
+
+01:02:09.000 --> 01:02:13.319
+form that I wrote this in so uh it's
+
+01:02:11.400 --> 01:02:15.359
+it's a fix length Vector that steers a
+
+01:02:13.319 --> 01:02:17.640
+language model to generate a specific
+
+01:02:15.359 --> 01:02:19.359
+sequence exactly when added to the
+
+01:02:17.640 --> 01:02:22.559
+hidden states of a model at a specific
+
+01:02:19.359 --> 01:02:24.480
+point so this is different than um a
+
+01:02:22.559 --> 01:02:26.839
+soft prompt or different than a model
+
+01:02:24.480 --> 01:02:29.520
+editing sort of approach
+
+01:02:26.839 --> 01:02:31.400
+um in this case there is a vector that
+
+01:02:29.520 --> 01:02:32.960
+corresponds to a sequence and that
+
+01:02:31.400 --> 01:02:35.359
+Vector doesn't correspond to any other
+
+01:02:32.960 --> 01:02:36.640
+sequence there could be multiple vectors
+
+01:02:35.359 --> 01:02:39.079
+and it turns out there are multiple
+
+01:02:36.640 --> 01:02:41.799
+vectors that correspond to that sequence
+
+01:02:39.079 --> 01:02:44.160
+it'll be a little bit clearer um based
+
+01:02:41.799 --> 01:02:46.279
+on how we extract these
+
+01:02:44.160 --> 01:02:48.839
+things um so this is the stick that
+
+01:02:46.279 --> 01:02:52.000
+we're talking the language
+
+01:02:48.839 --> 01:02:53.599
+model um so how do we extract them so
+
+01:02:52.000 --> 01:02:57.400
+this is
+
+01:02:53.599 --> 01:03:00.200
+gpt2 um basically this Z steer thing on
+
+01:02:57.400 --> 01:03:03.240
+the left this is the steering Vector
+
+01:03:00.200 --> 01:03:05.799
+this gets initialized randomly um with
+
+01:03:03.240 --> 01:03:09.520
+small like in a in a reasonable way
+
+01:03:05.799 --> 01:03:11.440
+uniformly and small um and for any
+
+01:03:09.520 --> 01:03:14.000
+sequence a specific sequence that we
+
+01:03:11.440 --> 01:03:17.680
+want the model to generate we
+
+01:03:14.000 --> 01:03:19.400
+optimize this steering Vector Z steer uh
+
+01:03:17.680 --> 01:03:21.559
+to generate that sequence keeping the
+
+01:03:19.400 --> 01:03:23.960
+rest of the model entirely fixed so
+
+01:03:21.559 --> 01:03:26.200
+think about it as we're nudging an a
+
+01:03:23.960 --> 01:03:29.880
+frozen model to be able to generate a
+
+01:03:26.200 --> 01:03:31.680
+specific sequence at a specific time um
+
+01:03:29.880 --> 01:03:33.880
+and we have a lot of different options
+
+01:03:31.680 --> 01:03:35.559
+on where to inject the steering intive
+
+01:03:33.880 --> 01:03:37.520
+can put it basically anywhere in the
+
+01:03:35.559 --> 01:03:41.799
+model we can put it at any time step any
+
+01:03:37.520 --> 01:03:43.839
+number of these things in practice um
+
+01:03:41.799 --> 01:03:45.839
+providing it just at the first time step
+
+01:03:43.839 --> 01:03:48.039
+and somewhere in the middle of the model
+
+01:03:45.839 --> 01:03:52.480
+basically not the first layer and not
+
+01:03:48.039 --> 01:03:56.240
+the last layer works pretty well um and
+
+01:03:52.480 --> 01:04:00.279
+so more formally um forget the kind of
+
+01:03:56.240 --> 01:04:03.640
+notation um but right here we initialize
+
+01:04:00.279 --> 01:04:06.559
+um this Z steer and for a few iterations
+
+01:04:03.640 --> 01:04:08.039
+um we do forward passes first this
+
+01:04:06.559 --> 01:04:09.599
+starts as random and then this gets
+
+01:04:08.039 --> 01:04:11.960
+closer and closer and closer to being
+
+01:04:09.599 --> 01:04:14.279
+able to generate this sequence and
+
+01:04:11.960 --> 01:04:16.599
+eventually we get to a point uh and this
+
+01:04:14.279 --> 01:04:18.400
+n is pretty small it's eight or 10 or
+
+01:04:16.599 --> 01:04:20.160
+something like that um for most
+
+01:04:18.400 --> 01:04:22.200
+sequences we get to a point where we
+
+01:04:20.160 --> 01:04:23.920
+have found this stick that is allowed to
+
+01:04:22.200 --> 01:04:26.079
+poke this model to generate that
+
+01:04:23.920 --> 01:04:29.319
+sequence exactly now when we greedy
+
+01:04:26.079 --> 01:04:32.480
+decode from the model we pass in just a
+
+01:04:29.319 --> 01:04:34.920
+beginning of sequence token and this Z
+
+01:04:32.480 --> 01:04:37.119
+steer the steering vector and it's able
+
+01:04:34.920 --> 01:04:39.720
+to uncover a whole sequence that whole
+
+01:04:37.119 --> 01:04:41.319
+sequence that we had at the beginning
+
+01:04:39.720 --> 01:04:44.240
+entirely
+
+01:04:41.319 --> 01:04:46.640
+um this is weird and interesting because
+
+01:04:44.240 --> 01:04:48.880
+in a lot of cases um in like the
+
+01:04:46.640 --> 01:04:52.039
+prompting world in the soft prompt World
+
+01:04:48.880 --> 01:04:54.640
+usually you need a pretty large uh width
+
+01:04:52.039 --> 01:04:57.880
+of a prompt to be able to do things with
+
+01:04:54.640 --> 01:05:00.400
+um and this this generally in in that
+
+01:04:57.880 --> 01:05:02.000
+structure you're doing a specific task
+
+01:05:00.400 --> 01:05:04.200
+and you're providing kind of a large a
+
+01:05:02.000 --> 01:05:06.720
+large prompt to do this with large soft
+
+01:05:04.200 --> 01:05:10.520
+prompt to do this with this is often H
+
+01:05:06.720 --> 01:05:13.200
+this often has a width of 50 and a and a
+
+01:05:10.520 --> 01:05:15.520
+length of the the hidden size or the
+
+01:05:13.200 --> 01:05:17.160
+embedding size of the model in our cases
+
+01:05:15.520 --> 01:05:20.079
+all of our steering vectors are withth
+
+01:05:17.160 --> 01:05:21.440
+one and they're of just the hidden size
+
+01:05:20.079 --> 01:05:24.039
+of the
+
+01:05:21.440 --> 01:05:26.520
+model um so what ends
+
+01:05:24.039 --> 01:05:29.559
+up happening
+
+01:05:26.520 --> 01:05:31.520
+um actually before I go to results any
+
+01:05:29.559 --> 01:05:34.720
+questions this is a this is a weird
+
+01:05:31.520 --> 01:05:38.160
+setup and weird relative to what other
+
+01:05:34.720 --> 01:05:39.310
+people do so happy to take any
+
+01:05:38.160 --> 01:05:42.480
+questions
+
+01:05:39.310 --> 01:05:42.480
+[Music]
+
+01:05:42.880 --> 01:05:50.640
+yeah similarly if your prompt was um of
+
+01:05:47.440 --> 01:05:53.440
+a specific type so the prompt here is a
+
+01:05:50.640 --> 01:05:55.720
+continuous Vector passed in it's a
+
+01:05:53.440 --> 01:05:59.760
+single length width hidden size
+
+01:05:55.720 --> 01:06:02.799
+continuous Vector so um it's kind of
+
+01:05:59.760 --> 01:06:05.559
+like maybe collapsing your prompt into
+
+01:06:02.799 --> 01:06:08.480
+this compressing it into this tiny
+
+01:06:05.559 --> 01:06:12.119
+VOR you can think of that way
+
+01:06:08.480 --> 01:06:16.920
+yeah any other questions
+
+01:06:12.119 --> 01:06:16.920
+yeah this would be like
+
+01:06:18.160 --> 01:06:23.359
+I'm
+
+01:06:20.880 --> 01:06:28.279
+things potentially um this is something
+
+01:06:23.359 --> 01:06:30.640
+that I want to work on it uh like a year
+
+01:06:28.279 --> 01:06:32.119
+ago and didn't get didn't get this
+
+01:06:30.640 --> 01:06:34.559
+sufficient Buy in and then had to apply
+
+01:06:32.119 --> 01:06:36.880
+to grad school and all these things so
+
+01:06:34.559 --> 01:06:40.160
+it went by the wayside but but
+
+01:06:36.880 --> 01:06:43.440
+definitely something to something to
+
+01:06:40.160 --> 01:06:45.920
+pursue um there's a lot of scope there
+
+01:06:43.440 --> 01:06:45.920
+any other
+
+01:06:47.640 --> 01:06:54.480
+questions all right so move over to
+
+01:06:51.319 --> 01:06:56.119
+results so we can find steering vectors
+
+01:06:54.480 --> 01:06:58.520
+and that's and that's interesting thing
+
+01:06:56.119 --> 01:07:00.559
+um and we can find them pretty easily
+
+01:06:58.520 --> 01:07:02.559
+and for most sequences even sequences
+
+01:07:00.559 --> 01:07:04.559
+that the model hasn't seen before the
+
+01:07:02.559 --> 01:07:06.400
+underlying language model hasn't seen
+
+01:07:04.559 --> 01:07:09.640
+before
+
+01:07:06.400 --> 01:07:13.160
+um it also works for and this is kind of
+
+01:07:09.640 --> 01:07:16.799
+a negative but it also works for random
+
+01:07:13.160 --> 01:07:20.039
+sequences of very small length but it's
+
+01:07:16.799 --> 01:07:22.359
+harder to find so you can imagine if
+
+01:07:20.039 --> 01:07:24.760
+your uh steering Vector is basically a
+
+01:07:22.359 --> 01:07:26.279
+giant bulldozer it doesn't matter what
+
+01:07:24.760 --> 01:07:28.640
+your model is learning learned similar
+
+01:07:26.279 --> 01:07:30.160
+to the probe situation if you can
+
+01:07:28.640 --> 01:07:32.559
+compress all that information of that
+
+01:07:30.160 --> 01:07:35.400
+sequence into the vector you don't
+
+01:07:32.559 --> 01:07:37.400
+really need the language model um so
+
+01:07:35.400 --> 01:07:39.559
+there are cases when you're looking at
+
+01:07:37.400 --> 01:07:40.760
+sequences of length like five seven
+
+01:07:39.559 --> 01:07:43.079
+eight something like this you can
+
+01:07:40.760 --> 01:07:45.520
+uniformly sample from the vocabulary at
+
+01:07:43.079 --> 01:07:47.359
+random with replacement generate utter
+
+01:07:45.520 --> 01:07:49.799
+garbage and find steering vectors for
+
+01:07:47.359 --> 01:07:53.200
+them takes a little while but your model
+
+01:07:49.799 --> 01:07:55.520
+is complex enough that you can basically
+
+01:07:53.200 --> 01:07:57.960
+bulldo your model to be able to do this
+
+01:07:55.520 --> 01:08:00.200
+even if that sequence is incredibly low
+
+01:07:57.960 --> 01:08:01.480
+likelihood under the model but it works
+
+01:08:00.200 --> 01:08:05.319
+better for things that are higher
+
+01:08:01.480 --> 01:08:07.760
+likelihood under the model um
+
+01:08:05.319 --> 01:08:09.920
+predictably the I think the thing that
+
+01:08:07.760 --> 01:08:12.760
+surprised me the most was these steering
+
+01:08:09.920 --> 01:08:15.319
+vectors themselves have interpretable
+
+01:08:12.760 --> 01:08:17.960
+properties U so distances in steering
+
+01:08:15.319 --> 01:08:20.759
+Vector space reflect semantic similarity
+
+01:08:17.960 --> 01:08:23.640
+so if you have two sentences that are
+
+01:08:20.759 --> 01:08:26.719
+close um they're also close in steering
+
+01:08:23.640 --> 01:08:29.759
+Vector space that's kind of nice
+
+01:08:26.719 --> 01:08:32.359
+um it does better than for example the
+
+01:08:29.759 --> 01:08:34.520
+representations one would use for for
+
+01:08:32.359 --> 01:08:37.159
+probing so mean pooling Bert hidden
+
+01:08:34.520 --> 01:08:39.600
+States like we looked at before those do
+
+01:08:37.159 --> 01:08:42.080
+actually worse than steering vectors um
+
+01:08:39.600 --> 01:08:45.799
+just a bit
+
+01:08:42.080 --> 01:08:47.880
+surprising um style transfer is possible
+
+01:08:45.799 --> 01:08:49.719
+with simple Vector arithmetic so it' be
+
+01:08:47.880 --> 01:08:52.799
+nice to say that I have a sequence I
+
+01:08:49.719 --> 01:08:56.000
+want to subtract you know negativity and
+
+01:08:52.799 --> 01:08:58.799
+add positivity for for sentiment or
+
+01:08:56.000 --> 01:09:00.520
+other sorts of Styles um we can do this
+
+01:08:58.799 --> 01:09:02.159
+and we can do this reasonably well in
+
+01:09:00.520 --> 01:09:05.319
+steering VOR
+
+01:09:02.159 --> 01:09:07.920
+space um we can also decode from
+
+01:09:05.319 --> 01:09:10.600
+interpolations in the Laten space so you
+
+01:09:07.920 --> 01:09:12.759
+take two steering vectors for two
+
+01:09:10.600 --> 01:09:14.759
+sequences you look in the middle of them
+
+01:09:12.759 --> 01:09:17.400
+you linearly interpolate between them
+
+01:09:14.759 --> 01:09:20.600
+and you decode um if the space is kind
+
+01:09:17.400 --> 01:09:22.080
+of weirdly peaky then you would have
+
+01:09:20.600 --> 01:09:23.839
+issues and what you would generate is
+
+01:09:22.080 --> 01:09:25.080
+garbage and there's no guarantee that
+
+01:09:23.839 --> 01:09:27.199
+the space should be reasonable in
+
+01:09:25.080 --> 01:09:30.480
+between but it turns out it
+
+01:09:27.199 --> 01:09:33.719
+is um here's an example of one of these
+
+01:09:30.480 --> 01:09:36.359
+style transfer cases so very very simple
+
+01:09:33.719 --> 01:09:39.239
+easy easy sentence we found steering
+
+01:09:36.359 --> 01:09:41.679
+vectors for The Taste is excellent and
+
+01:09:39.239 --> 01:09:43.640
+and we took a sample of 100 positive
+
+01:09:41.679 --> 01:09:45.359
+sentences and 100 negative sentences
+
+01:09:43.640 --> 01:09:47.159
+found their steering vectors took the
+
+01:09:45.359 --> 01:09:48.960
+mean and thought that you know that
+
+01:09:47.159 --> 01:09:51.400
+looks like the positive concept steering
+
+01:09:48.960 --> 01:09:54.040
+Vector negative concept steering Vector
+
+01:09:51.400 --> 01:09:56.600
+we just did Vector arithmetic just did
+
+01:09:54.040 --> 01:09:59.880
+uh current steering
+
+01:09:56.600 --> 01:10:02.440
+Vector uh plus negative minus positive
+
+01:09:59.880 --> 01:10:03.520
+and we got the taste is unpleasant um
+
+01:10:02.440 --> 01:10:06.960
+and
+
+01:10:03.520 --> 01:10:08.880
+similarly um in the reverse
+
+01:10:06.960 --> 01:10:12.520
+directions it turns out that the
+
+01:10:08.880 --> 01:10:15.199
+magnitude matters because um for every
+
+01:10:12.520 --> 01:10:17.800
+single sequence there's kind of an end
+
+01:10:15.199 --> 01:10:20.640
+dimensional ball around that steering
+
+01:10:17.800 --> 01:10:23.640
+Vector that we found that also decodes
+
+01:10:20.640 --> 01:10:25.920
+that specific sequence and so that shows
+
+01:10:23.640 --> 01:10:28.880
+that the space is kind of reasonably
+
+01:10:25.920 --> 01:10:32.320
+well formed there's there's of course uh
+
+01:10:28.880 --> 01:10:34.280
+a lot of weird sort of areas um and so
+
+01:10:32.320 --> 01:10:37.120
+if you go poke around in steering Vector
+
+01:10:34.280 --> 01:10:38.760
+space and sort of try to sample from it
+
+01:10:37.120 --> 01:10:41.280
+eventually you'll find some weird edge
+
+01:10:38.760 --> 01:10:43.320
+cases and some garbage and repeated text
+
+01:10:41.280 --> 01:10:46.159
+and little things like
+
+01:10:43.320 --> 01:10:50.520
+this any questions here before I kind of
+
+01:10:46.159 --> 01:10:50.520
+Rapid Fire through the the last few
+
+01:10:50.920 --> 01:10:57.239
+things yeah like here
+
+01:10:57.400 --> 01:11:01.400
+yeah so we went uh Beyond this um there
+
+01:11:00.199 --> 01:11:04.280
+was
+
+01:11:01.400 --> 01:11:07.440
+so in in these specific experiments we
+
+01:11:04.280 --> 01:11:09.600
+looked at the middle of gpt2 um so this
+
+01:11:07.440 --> 01:11:12.679
+was like layer six layer seven and at
+
+01:11:09.600 --> 01:11:15.280
+the first time step we didn't do any um
+
+01:11:12.679 --> 01:11:17.239
+like magnitude scaling and so you can
+
+01:11:15.280 --> 01:11:19.480
+imagine if you put a giant Vector in
+
+01:11:17.239 --> 01:11:21.040
+there the models never the rest of the
+
+01:11:19.480 --> 01:11:24.679
+model has never seen something of that
+
+01:11:21.040 --> 01:11:26.159
+magnitude so it's now in a weird State
+
+01:11:24.679 --> 01:11:28.280
+and it's just going to break so if you
+
+01:11:26.159 --> 01:11:30.560
+put this to like I don't know 500 or
+
+01:11:28.280 --> 01:11:32.960
+something like this it break it just has
+
+01:11:30.560 --> 01:11:35.239
+no idea it's like basically like telling
+
+01:11:32.960 --> 01:11:37.199
+the rest your model hey it's like a
+
+01:11:35.239 --> 01:11:38.760
+completely untrained model be it looks
+
+01:11:37.199 --> 01:11:42.000
+similar to like random performance you
+
+01:11:38.760 --> 01:11:43.840
+get repeats and things like this smaller
+
+01:11:42.000 --> 01:11:45.800
+you end up staying in this ball for the
+
+01:11:43.840 --> 01:11:47.920
+sequence two two seemed pretty
+
+01:11:45.800 --> 01:11:50.199
+reasonable but we didn't spend a lot of
+
+01:11:47.920 --> 01:11:53.560
+time just like the day before the paper
+
+01:11:50.199 --> 01:11:56.600
+was do we were two seems reasonable we
+
+01:11:53.560 --> 01:11:59.159
+went to three we went to five 10 broke
+
+01:11:56.600 --> 01:12:01.199
+five somewhat broke two seems
+
+01:11:59.159 --> 01:12:03.440
+reasonable
+
+01:12:01.199 --> 01:12:06.400
+um decent signings
+
+01:12:03.440 --> 01:12:08.639
+hopefully um cool so I'll talk about uh
+
+01:12:06.400 --> 01:12:10.920
+a similar type of work uh that came out
+
+01:12:08.639 --> 01:12:13.000
+more recently on inference time
+
+01:12:10.920 --> 01:12:14.159
+intervention so basically they use some
+
+01:12:13.000 --> 01:12:16.719
+of the ideas that we talked about
+
+01:12:14.159 --> 01:12:18.840
+earlier they use linear probes um to
+
+01:12:16.719 --> 01:12:20.560
+find a tension head that correspond to a
+
+01:12:18.840 --> 01:12:23.600
+desired attribute they did this for
+
+01:12:20.560 --> 01:12:26.440
+truthful QA so uh their Hope was to find
+
+01:12:23.600 --> 01:12:28.639
+truthful directions in Len space
+
+01:12:26.440 --> 01:12:31.639
+um and then they shifted the attention
+
+01:12:28.639 --> 01:12:33.199
+head activations um during inference
+
+01:12:31.639 --> 01:12:35.280
+along the directions determined by the
+
+01:12:33.199 --> 01:12:38.280
+probes um so what this kind of looks
+
+01:12:35.280 --> 01:12:40.280
+like is you take your attention heads
+
+01:12:38.280 --> 01:12:42.440
+you probe them so you stick classify on
+
+01:12:40.280 --> 01:12:44.360
+top um this classifier learns to
+
+01:12:42.440 --> 01:12:47.679
+disentangle sort of truthful and
+
+01:12:44.360 --> 01:12:50.239
+untruthful and now you have um now you
+
+01:12:47.679 --> 01:12:52.080
+have a hyperplane and then you can move
+
+01:12:50.239 --> 01:12:54.320
+orthogonally to this hyper plane in the
+
+01:12:52.080 --> 01:12:55.920
+direction depending on which way you
+
+01:12:54.320 --> 01:12:58.080
+want to shift so if you want to move
+
+01:12:55.920 --> 01:13:02.040
+towards truthful you can move in that
+
+01:12:58.080 --> 01:13:04.400
+direction or or away um and they do this
+
+01:13:02.040 --> 01:13:07.560
+it works pretty well um I think they do
+
+01:13:04.400 --> 01:13:09.679
+this for GPT model and maybe a llama
+
+01:13:07.560 --> 01:13:12.960
+model um but can't can't remember the
+
+01:13:09.679 --> 01:13:15.960
+exact details um and it's a similar
+
+01:13:12.960 --> 01:13:21.040
+intervention um they basically add this
+
+01:13:15.960 --> 01:13:23.400
+Vector um that they found and they they
+
+01:13:21.040 --> 01:13:25.679
+have a little note on scaling they if
+
+01:13:23.400 --> 01:13:27.719
+they scale if that if the magnitude of
+
+01:13:25.679 --> 01:13:30.000
+the thing is too much things break so
+
+01:13:27.719 --> 01:13:33.880
+they have a they like hyper parameter
+
+01:13:30.000 --> 01:13:36.800
+search for the sort of magnitude of
+
+01:13:33.880 --> 01:13:38.840
+activation um but it's sort of a very
+
+01:13:36.800 --> 01:13:41.520
+similar approach to what we did but this
+
+01:13:38.840 --> 01:13:43.040
+focuses on specific attention heads and
+
+01:13:41.520 --> 01:13:44.440
+they don't do this for all the attention
+
+01:13:43.040 --> 01:13:46.600
+heads so back to like your question
+
+01:13:44.440 --> 01:13:49.080
+earlier do attention heads specialize it
+
+01:13:46.600 --> 01:13:52.360
+seems like they do and so there are many
+
+01:13:49.080 --> 01:13:54.320
+of them that uh have like no probing
+
+01:13:52.360 --> 01:13:57.719
+accuracy or limited probing accuracy and
+
+01:13:54.320 --> 01:13:59.400
+actually um are like distractors for the
+
+01:13:57.719 --> 01:14:03.400
+CH FL
+
+01:13:59.400 --> 01:14:03.400
+Direction any questions
+
+01:14:06.040 --> 01:14:11.760
+here cool so more activation
+
+01:14:09.120 --> 01:14:14.760
+manipulation so there's uh some work
+
+01:14:11.760 --> 01:14:17.600
+recently on contrastive steering vectors
+
+01:14:14.760 --> 01:14:19.480
+so the way we did this like sentiment
+
+01:14:17.600 --> 01:14:21.080
+steering was we had some positive
+
+01:14:19.480 --> 01:14:23.040
+sentences some negative sentences they
+
+01:14:21.080 --> 01:14:24.520
+weren't tied together in any reasonable
+
+01:14:23.040 --> 01:14:26.360
+way we found their steering vectors
+
+01:14:24.520 --> 01:14:30.040
+separately you could imagine the case
+
+01:14:26.360 --> 01:14:33.159
+and maybe a more useful case um with two
+
+01:14:30.040 --> 01:14:36.280
+prompts that um you can design that go
+
+01:14:33.159 --> 01:14:38.000
+two different ways you can sort of find
+
+01:14:36.280 --> 01:14:42.280
+their representations and do the
+
+01:14:38.000 --> 01:14:45.679
+manipulation the differences here um
+
+01:14:42.280 --> 01:14:48.800
+like individually rather than um for a
+
+01:14:45.679 --> 01:14:52.400
+whole concept or a whole attribute and
+
+01:14:48.800 --> 01:14:54.400
+the value here is your context is um
+
+01:14:52.400 --> 01:14:56.600
+preserved so if you're doing something
+
+01:14:54.400 --> 01:14:58.239
+like you know you're doing retrieval
+
+01:14:56.600 --> 01:15:00.440
+based things now you have some sort of
+
+01:14:58.239 --> 01:15:03.360
+document and then you have a question if
+
+01:15:00.440 --> 01:15:05.040
+your question sort of uh if you want to
+
+01:15:03.360 --> 01:15:07.560
+ask it in two different ways for two
+
+01:15:05.040 --> 01:15:08.880
+different things this would be a much
+
+01:15:07.560 --> 01:15:11.239
+better approach if you want to use
+
+01:15:08.880 --> 01:15:14.600
+steering vectors than the stuff I was
+
+01:15:11.239 --> 01:15:16.159
+doing um and it seems to work a little
+
+01:15:14.600 --> 01:15:17.880
+bit better they didn't compare against
+
+01:15:16.159 --> 01:15:19.400
+our our things because it's not like an
+
+01:15:17.880 --> 01:15:21.880
+Apples to Apples comparison but it seems
+
+01:15:19.400 --> 01:15:23.960
+to work better and be more General um
+
+01:15:21.880 --> 01:15:25.560
+and be more
+
+01:15:23.960 --> 01:15:27.840
+useful
+
+01:15:25.560 --> 01:15:27.840
+any
+
+01:15:31.400 --> 01:15:37.679
+questions cool so what can model
+
+01:15:35.080 --> 01:15:40.080
+interpretability give us these are these
+
+01:15:37.679 --> 01:15:41.960
+are my concluding remarks so hopefully
+
+01:15:40.080 --> 01:15:43.920
+we get a better understanding of how
+
+01:15:41.960 --> 01:15:46.840
+language models work their their
+
+01:15:43.920 --> 01:15:49.520
+internals their structure um we get to
+
+01:15:46.840 --> 01:15:52.800
+understand uh kind of why they do really
+
+01:15:49.520 --> 01:15:55.239
+well this is still like very very
+
+01:15:52.800 --> 01:15:57.320
+unclear um and hopefully we find
+
+01:15:55.239 --> 01:15:59.400
+lightweight methods to control and steer
+
+01:15:57.320 --> 01:16:03.360
+models as models become more and more
+
+01:15:59.400 --> 01:16:05.280
+useful um and and impact more more users
+
+01:16:03.360 --> 01:16:09.360
+we need better ways to control and steer
+
+01:16:05.280 --> 01:16:13.120
+them um and it's unclear how much
+
+01:16:09.360 --> 01:16:15.360
+industry will devote to these things um
+
+01:16:13.120 --> 01:16:18.080
+so it might be the role of Academia to
+
+01:16:15.360 --> 01:16:21.239
+do more science in in order to figure
+
+01:16:18.080 --> 01:16:23.920
+out how to control and steer these
+
+01:16:21.239 --> 01:16:25.520
+better um and hopefully we can also find
+
+01:16:23.920 --> 01:16:29.199
+potential Al Alternatives or
+
+01:16:25.520 --> 01:16:34.840
+complimentary methods to to do alignment
+
+01:16:29.199 --> 01:16:37.480
+um rhf is kind of expensive um and if if
+
+01:16:34.840 --> 01:16:40.080
+we could do this with limited data and
+
+01:16:37.480 --> 01:16:42.760
+um exploit structure um and information
+
+01:16:40.080 --> 01:16:46.400
+that's already in the model more so than
+
+01:16:42.760 --> 01:16:48.600
+than these methods um maybe maybe we can
+
+01:16:46.400 --> 01:16:50.920
+align them better and these things don't
+
+01:16:48.600 --> 01:16:52.480
+have to be uh Alternatives they can be
+
+01:16:50.920 --> 01:16:53.840
+complimentary to to
+
+01:16:52.480 --> 01:16:57.159
+[Music]
+
+01:16:53.840 --> 01:17:00.040
+rhm um here's some resources this is an
+
+01:16:57.159 --> 01:17:01.280
+extremely incomplete group but here are
+
+01:17:00.040 --> 01:17:04.080
+some folks that work on model
+
+01:17:01.280 --> 01:17:07.040
+interoperability there's many of these
+
+01:17:04.080 --> 01:17:09.120
+um I cited some some work from some of
+
+01:17:07.040 --> 01:17:11.280
+these teams but um there's a lot of
+
+01:17:09.120 --> 01:17:13.280
+people working on it and in the last
+
+01:17:11.280 --> 01:17:15.040
+like year there's been kind of an
+
+01:17:13.280 --> 01:17:17.480
+explosion especially in the mechanistic
+
+01:17:15.040 --> 01:17:21.639
+interpretability kind of World um Sasha
+
+01:17:17.480 --> 01:17:23.800
+Rush had a recent tweet that uh asked
+
+01:17:21.639 --> 01:17:25.320
+like prospective grad students what is
+
+01:17:23.800 --> 01:17:27.239
+the topic that they're most excited
+
+01:17:25.320 --> 01:17:29.880
+about and mechanistic interpretability
+
+01:17:27.239 --> 01:17:33.960
+was a thing that seemed to have won out
+
+01:17:29.880 --> 01:17:37.040
+um so I encourage you to to kind of dive
+
+01:17:33.960 --> 01:17:38.719
+into this literature and read some of
+
+01:17:37.040 --> 01:17:41.679
+the papers if you're if you're excited
+
+01:17:38.719 --> 01:17:45.199
+about it and yeah thanks for your
+
+01:17:41.679 --> 01:17:45.199
+attention and that's all I
+
+01:17:45.400 --> 01:17:48.400
+have