WEBVTT 00:00:00.000 --> 00:00:01.961 align:middle line:90% [SQUEAKING] 00:00:01.961 --> 00:00:03.909 align:middle line:90% [RUSTLING] 00:00:03.909 --> 00:00:05.370 align:middle line:90% [CLICKING] 00:00:05.370 --> 00:00:12.680 align:middle line:90% 00:00:12.680 --> 00:00:15.440 align:middle line:84% SARA BEERY: We're going to finish up our short series 00:00:15.440 --> 00:00:18.400 align:middle line:90% on transfer learning today. 00:00:18.400 --> 00:00:23.080 align:middle line:84% And the other logistical thing I did want to bring up 00:00:23.080 --> 00:00:27.400 align:middle line:84% is I know that project proposals are due this Friday. 00:00:27.400 --> 00:00:30.534 align:middle line:84% During our instructor meeting today, we put together a rubric 00:00:30.534 --> 00:00:32.159 align:middle line:84% for those project proposals just to try 00:00:32.159 --> 00:00:34.617 align:middle line:84% to give you guys a little bit more structured understanding 00:00:34.617 --> 00:00:37.200 align:middle line:84% of what we're going to be evaluating those on. 00:00:37.200 --> 00:00:39.480 align:middle line:90% So that's up on Piazza. 00:00:39.480 --> 00:00:41.560 align:middle line:84% Essentially, it's like out of 10, 00:00:41.560 --> 00:00:43.000 align:middle line:84% and there's five different things 00:00:43.000 --> 00:00:44.733 align:middle line:90% that are each worth two points. 00:00:44.733 --> 00:00:46.400 align:middle line:84% And then the other thing I wanted to say 00:00:46.400 --> 00:00:48.400 align:middle line:84% is that this is actually, shockingly, 00:00:48.400 --> 00:00:51.280 align:middle line:84% already my last lecture of the semester. 00:00:51.280 --> 00:00:53.480 align:middle line:84% I don't know where the time goes, 00:00:53.480 --> 00:00:56.280 align:middle line:84% but it has been really awesome getting 00:00:56.280 --> 00:00:58.220 align:middle line:90% to lecture you this semester. 00:00:58.220 --> 00:01:01.580 align:middle line:84% And I will probably see you towards the end of the semester 00:01:01.580 --> 00:01:04.080 align:middle line:90% anyway, but yeah. 00:01:04.080 --> 00:01:05.720 align:middle line:90% Time flies. 00:01:05.720 --> 00:01:06.220 align:middle line:90% Cool. 00:01:06.220 --> 00:01:10.060 align:middle line:84% So today, continuing that thread on thinking 00:01:10.060 --> 00:01:12.772 align:middle line:84% about how we can transfer knowledge specifically 00:01:12.772 --> 00:01:14.980 align:middle line:84% within the constraints of these deep learning models, 00:01:14.980 --> 00:01:17.420 align:middle line:84% or these deep learning systems, today, we're 00:01:17.420 --> 00:01:19.620 align:middle line:84% going to be specifically talking about how 00:01:19.620 --> 00:01:23.340 align:middle line:84% we can transfer knowledge about the inputs to the system. 00:01:23.340 --> 00:01:26.260 align:middle line:84% And so we're going to spend some time thinking about how 00:01:26.260 --> 00:01:28.220 align:middle line:84% you can think about generative models 00:01:28.220 --> 00:01:33.900 align:middle line:84% as another way to capture information about data inputs. 00:01:33.900 --> 00:01:37.860 align:middle line:84% And then we're also going to talk a little bit about models 00:01:37.860 --> 00:01:41.260 align:middle line:84% that learn to learn, so this idea of meta-learning 00:01:41.260 --> 00:01:44.220 align:middle line:84% and how you can build models that are almost designed 00:01:44.220 --> 00:01:46.460 align:middle line:84% to be good at transfer learning, designed 00:01:46.460 --> 00:01:48.500 align:middle line:84% to be good at sharing information or learning 00:01:48.500 --> 00:01:50.460 align:middle line:90% new things. 00:01:50.460 --> 00:01:51.080 align:middle line:90% Cool. 00:01:51.080 --> 00:01:57.520 align:middle line:84% So if we're thinking about knowledge about the inputs, 00:01:57.520 --> 00:01:59.120 align:middle line:84% maybe one way to think about this 00:01:59.120 --> 00:02:02.840 align:middle line:84% is if you know something about the distribution of inputs 00:02:02.840 --> 00:02:05.240 align:middle line:84% you're going to expect, maybe there's 00:02:05.240 --> 00:02:07.640 align:middle line:84% a way to actually use that to direct or accelerate 00:02:07.640 --> 00:02:11.760 align:middle line:84% your learning or even directly target 00:02:11.760 --> 00:02:14.840 align:middle line:90% a specific type of learning. 00:02:14.840 --> 00:02:21.760 align:middle line:84% So you have a data set, and then we spent quite a few lectures 00:02:21.760 --> 00:02:23.880 align:middle line:84% with Phil talking about this ability 00:02:23.880 --> 00:02:28.380 align:middle line:84% to learn how to generate data from that same distribution, 00:02:28.380 --> 00:02:31.080 align:middle line:84% this idea that you can use generative modeling. 00:02:31.080 --> 00:02:35.040 align:middle line:84% And so now, that gives you access 00:02:35.040 --> 00:02:37.360 align:middle line:84% to some model that then is able to sample 00:02:37.360 --> 00:02:40.280 align:middle line:90% from that distribution. 00:02:40.280 --> 00:02:43.320 align:middle line:84% And so here, at a first glance, this 00:02:43.320 --> 00:02:44.740 align:middle line:90% might not seem that interesting. 00:02:44.740 --> 00:02:46.640 align:middle line:84% You had data, and then you used it 00:02:46.640 --> 00:02:49.820 align:middle line:84% to train a model that mimics the data that you already had. 00:02:49.820 --> 00:02:53.440 align:middle line:84% But you already had the data, so how does it actually help? 00:02:53.440 --> 00:02:56.420 align:middle line:84% How are you going to get more information out 00:02:56.420 --> 00:02:58.900 align:middle line:84% of a system that's been trained to model 00:02:58.900 --> 00:03:02.500 align:middle line:90% from a specific distribution? 00:03:02.500 --> 00:03:05.100 align:middle line:84% But one of the interesting things 00:03:05.100 --> 00:03:08.740 align:middle line:84% that you can think about this as is, 00:03:08.740 --> 00:03:14.460 align:middle line:84% assuming that generative model is kind of very good, 00:03:14.460 --> 00:03:20.000 align:middle line:84% you can almost think about the model itself as a new way, 00:03:20.000 --> 00:03:22.500 align:middle line:84% or a new mechanism to access data, 00:03:22.500 --> 00:03:26.940 align:middle line:84% going beyond the capabilities of the initial training data set. 00:03:26.940 --> 00:03:29.600 align:middle line:84% So maybe you call it data plus plus. 00:03:29.600 --> 00:03:32.660 align:middle line:84% And where does that additive value 00:03:32.660 --> 00:03:34.540 align:middle line:84% that you're getting from a data generator 00:03:34.540 --> 00:03:37.860 align:middle line:84% that you might not get from the data itself? 00:03:37.860 --> 00:03:40.420 align:middle line:90% And here's an interesting quote. 00:03:40.420 --> 00:03:44.220 align:middle line:84% So this is from the release of the first Stable Diffusion 00:03:44.220 --> 00:03:44.940 align:middle line:90% model. 00:03:44.940 --> 00:03:48.020 align:middle line:84% They said that, "This release is the culmination of many hours 00:03:48.020 --> 00:03:52.900 align:middle line:84% of collective effort to create a single file that compresses 00:03:52.900 --> 00:03:57.600 align:middle line:84% the visual information of humanity into a few gigabytes." 00:03:57.600 --> 00:04:00.320 align:middle line:90% And that sounds pretty awesome. 00:04:00.320 --> 00:04:03.840 align:middle line:84% So essentially, it's the idea that, actually, 00:04:03.840 --> 00:04:09.497 align:middle line:84% a generative model is a really cool form of data compression. 00:04:09.497 --> 00:04:11.080 align:middle line:84% So you can take that generative model, 00:04:11.080 --> 00:04:15.380 align:middle line:84% and you can go from some latent variable to anything on Earth, 00:04:15.380 --> 00:04:19.660 align:middle line:84% like any sort of visual-- all visual information of humanity, 00:04:19.660 --> 00:04:23.600 align:middle line:84% then that is a shocking compression of everything. 00:04:23.600 --> 00:04:25.880 align:middle line:84% You think about even just the storage capacity 00:04:25.880 --> 00:04:27.480 align:middle line:84% of all the images on the internet 00:04:27.480 --> 00:04:30.320 align:middle line:84% versus the storage capacity of a Stable Diffusion model. 00:04:30.320 --> 00:04:33.000 align:middle line:84% That's a significant difference in terms 00:04:33.000 --> 00:04:36.000 align:middle line:90% of what you're capturing. 00:04:36.000 --> 00:04:45.440 align:middle line:84% So here, how do we think about what data plus plus means, 00:04:45.440 --> 00:04:50.400 align:middle line:84% and how do we think about data as maybe like an object? 00:04:50.400 --> 00:04:57.140 align:middle line:84% So if we start with some X, which 00:04:57.140 --> 00:05:00.860 align:middle line:84% is the set of the original training data, which 00:05:00.860 --> 00:05:05.580 align:middle line:84% is little x, z, which is a set of latent variables 00:05:05.580 --> 00:05:08.340 align:middle line:84% that can correspond to the original training data 00:05:08.340 --> 00:05:14.580 align:middle line:84% learned by our generative model, the mapping from z 00:05:14.580 --> 00:05:16.620 align:middle line:84% to x-- so this is basically the mapping that 00:05:16.620 --> 00:05:20.920 align:middle line:84% goes from some latent variable to a generated data point, 00:05:20.920 --> 00:05:24.100 align:middle line:84% and that's G. That's the generative model we've learned. 00:05:24.100 --> 00:05:27.380 align:middle line:84% And then if you also have G inverse-- essentially, 00:05:27.380 --> 00:05:30.980 align:middle line:84% this inverse mapping that lets you go from data to its 00:05:30.980 --> 00:05:33.820 align:middle line:90% associated latent variable-- 00:05:33.820 --> 00:05:35.980 align:middle line:90% you can actually define these. 00:05:35.980 --> 00:05:38.340 align:middle line:84% You can take these and define operators 00:05:38.340 --> 00:05:41.780 align:middle line:84% over those objects that enable us to do things 00:05:41.780 --> 00:05:46.500 align:middle line:84% to the generator and the system of data plus generator of data 00:05:46.500 --> 00:05:49.380 align:middle line:84% that we couldn't easily do to the data itself. 00:05:49.380 --> 00:05:53.480 align:middle line:84% So for example, you can think about interpolation 00:05:53.480 --> 00:05:55.880 align:middle line:90% over a data set. 00:05:55.880 --> 00:05:58.060 align:middle line:84% Pretty standard definition of interpolation. 00:05:58.060 --> 00:05:59.640 align:middle line:84% But now, you can think about maybe 00:05:59.640 --> 00:06:04.520 align:middle line:84% some-- if you have one set of this, like X, 00:06:04.520 --> 00:06:09.920 align:middle line:84% Mathcad X that's some data set, some latent space, a generator 00:06:09.920 --> 00:06:12.780 align:middle line:84% for that latent space to data in the inverse. 00:06:12.780 --> 00:06:15.160 align:middle line:84% If you have one of those, and you have another one that's 00:06:15.160 --> 00:06:17.140 align:middle line:84% for some of different data set, possibly, 00:06:17.140 --> 00:06:19.680 align:middle line:84% or a different generator, you can think about interpolating 00:06:19.680 --> 00:06:21.400 align:middle line:90% between the two, right? 00:06:21.400 --> 00:06:23.560 align:middle line:84% So some interpolation between those two things 00:06:23.560 --> 00:06:26.040 align:middle line:90% would generate a new one. 00:06:26.040 --> 00:06:28.760 align:middle line:84% And you can think about manipulation, right? 00:06:28.760 --> 00:06:32.580 align:middle line:84% You can think about how you might add some bias term to that 00:06:32.580 --> 00:06:36.440 align:middle line:84% that would give you, again, some whole new kind of object 00:06:36.440 --> 00:06:39.200 align:middle line:84% for accessing and interacting with data. 00:06:39.200 --> 00:06:41.680 align:middle line:90% You can think about composition. 00:06:41.680 --> 00:06:44.680 align:middle line:84% So here, how do you actually kind of compose two things 00:06:44.680 --> 00:06:46.600 align:middle line:84% is, of course, different from interpolating 00:06:46.600 --> 00:06:47.960 align:middle line:90% between two things. 00:06:47.960 --> 00:06:54.060 align:middle line:84% So now if you essentially are taking this space where 00:06:54.060 --> 00:06:58.920 align:middle line:84% you have just these different data products or data objects, 00:06:58.920 --> 00:07:02.860 align:middle line:84% you can think about really combining them and manipulating 00:07:02.860 --> 00:07:04.500 align:middle line:90% them in interesting ways. 00:07:04.500 --> 00:07:07.260 align:middle line:84% You can also think about optimization, 00:07:07.260 --> 00:07:12.940 align:middle line:84% so how you might actually be able to explicitly find 00:07:12.940 --> 00:07:17.980 align:middle line:84% some set of generator data, latent space inverse generator, 00:07:17.980 --> 00:07:22.380 align:middle line:84% that's optimal for some given dimension of optimality. 00:07:22.380 --> 00:07:27.460 align:middle line:84% And you can imagine taking that and bringing 00:07:27.460 --> 00:07:30.405 align:middle line:84% it to bear on many, many different research 00:07:30.405 --> 00:07:32.780 align:middle line:84% domains or problem domains that people are interested in. 00:07:32.780 --> 00:07:35.020 align:middle line:84% Things like graphics, visualization, 00:07:35.020 --> 00:07:39.020 align:middle line:84% data augmentation, counterfactual reasoning. 00:07:39.020 --> 00:07:42.460 align:middle line:84% And all of this work is kind of somewhat in progress, actually. 00:07:42.460 --> 00:07:45.100 align:middle line:84% So these are just some examples of papers 00:07:45.100 --> 00:07:48.460 align:middle line:84% that start getting at these mechanisms of interpolation, 00:07:48.460 --> 00:07:52.480 align:middle line:84% manipulation, composition, or optimization over generated data 00:07:52.480 --> 00:07:54.240 align:middle line:90% objects. 00:07:54.240 --> 00:07:59.440 align:middle line:84% So if we think about this space of generative models-- 00:07:59.440 --> 00:08:01.320 align:middle line:84% so here, this is now just visualizing 00:08:01.320 --> 00:08:02.653 align:middle line:90% what we were just talking about. 00:08:02.653 --> 00:08:04.480 align:middle line:84% We have some set of latent variables. 00:08:04.480 --> 00:08:06.300 align:middle line:84% So these are maybe the controls, right? 00:08:06.300 --> 00:08:08.640 align:middle line:84% This is kind of the way that you control 00:08:08.640 --> 00:08:10.840 align:middle line:90% what you want to generate. 00:08:10.840 --> 00:08:13.600 align:middle line:84% And maybe you've made some reasonable constraints 00:08:13.600 --> 00:08:20.080 align:middle line:84% in your system so you can ensure that z is normally distributed. 00:08:20.080 --> 00:08:25.320 align:middle line:84% Now, you have some generative model, G, 00:08:25.320 --> 00:08:28.960 align:middle line:84% that will take any point in that latent space 00:08:28.960 --> 00:08:30.740 align:middle line:90% and then synthesize an image. 00:08:30.740 --> 00:08:37.280 align:middle line:84% So here, now, if you sampled a different point in latent space, 00:08:37.280 --> 00:08:40.480 align:middle line:84% then you would sample a different image. 00:08:40.480 --> 00:08:44.080 align:middle line:84% And so now, given the relationship 00:08:44.080 --> 00:08:47.860 align:middle line:84% between dimensionality of movement in latent space 00:08:47.860 --> 00:08:49.860 align:middle line:84% and the image that's generated, you 00:08:49.860 --> 00:08:52.940 align:middle line:84% can think about this as a mechanism of control. 00:08:52.940 --> 00:08:56.700 align:middle line:84% So maybe if you move along some controllable snippet 00:08:56.700 --> 00:08:59.180 align:middle line:84% of the manifold of natural images, 00:08:59.180 --> 00:09:03.060 align:middle line:84% you can explicitly do things like find some latent variable 00:09:03.060 --> 00:09:06.100 align:middle line:84% in your latent space, some dimension in that latent space 00:09:06.100 --> 00:09:09.020 align:middle line:84% that corresponds to something like pose 00:09:09.020 --> 00:09:13.760 align:middle line:84% for this bird or orientation in some way. 00:09:13.760 --> 00:09:17.300 align:middle line:84% And so maybe there's different dimensions 00:09:17.300 --> 00:09:20.780 align:middle line:84% of this control in latent space that corresponds specifically 00:09:20.780 --> 00:09:25.220 align:middle line:84% to disentangled factors of variation or change 00:09:25.220 --> 00:09:28.500 align:middle line:84% in the space of those real generated images. 00:09:28.500 --> 00:09:33.280 align:middle line:90% And so if you can find-- yeah? 00:09:33.280 --> 00:09:37.140 align:middle line:84% AUDIENCE: On that slide, do you suppose that the two dimensions 00:09:37.140 --> 00:09:39.740 align:middle line:90% are linearly independent? 00:09:39.740 --> 00:09:41.860 align:middle line:90% And is it usually the case? 00:09:41.860 --> 00:09:44.400 align:middle line:84% SARA BEERY: So your question is, are you 00:09:44.400 --> 00:09:48.920 align:middle line:84% supposing that these dimensions in this kind of latent variable 00:09:48.920 --> 00:09:52.120 align:middle line:90% space are linearly independent? 00:09:52.120 --> 00:09:54.080 align:middle line:84% So I think the assumption here is 00:09:54.080 --> 00:10:00.000 align:middle line:84% that there is some decomposable directions that are maybe 00:10:00.000 --> 00:10:03.440 align:middle line:84% some projection of this where you can find orthogonality, 00:10:03.440 --> 00:10:05.440 align:middle line:84% and that orthogonality then corresponds 00:10:05.440 --> 00:10:13.920 align:middle line:84% to some very clean specific variables of variation 00:10:13.920 --> 00:10:16.040 align:middle line:90% in the real image space. 00:10:16.040 --> 00:10:24.760 align:middle line:84% In practice, sometimes, particularly for certain types 00:10:24.760 --> 00:10:30.800 align:middle line:84% or certain categories of generative models, this is true, 00:10:30.800 --> 00:10:33.380 align:middle line:84% but it's not necessarily easy to find. 00:10:33.380 --> 00:10:38.040 align:middle line:84% So people will do this cherry picked thing 00:10:38.040 --> 00:10:40.200 align:middle line:84% where they're like, hey, we found a dimension, 00:10:40.200 --> 00:10:41.720 align:middle line:84% where if we move in this dimension, 00:10:41.720 --> 00:10:44.640 align:middle line:84% you can see that it changes from day to night. 00:10:44.640 --> 00:10:46.160 align:middle line:84% We'll talk more about some of these. 00:10:46.160 --> 00:10:48.380 align:middle line:84% But I think one of the challenges of this 00:10:48.380 --> 00:10:51.980 align:middle line:84% is, particularly as these latent spaces get higher dimensional, 00:10:51.980 --> 00:10:54.500 align:middle line:84% being able to really explicitly disentangle 00:10:54.500 --> 00:10:56.700 align:middle line:84% these dimensions of variation in a clean way can 00:10:56.700 --> 00:10:59.760 align:middle line:84% be very difficult and can be highly heuristic. 00:10:59.760 --> 00:11:00.260 align:middle line:90% Yeah. 00:11:00.260 --> 00:11:02.820 align:middle line:90% 00:11:02.820 --> 00:11:03.620 align:middle line:90% Cool. 00:11:03.620 --> 00:11:07.740 align:middle line:84% So if you have that, if you have this kind of model that takes 00:11:07.740 --> 00:11:11.620 align:middle line:84% you from data to what we're calling data plus plus, 00:11:11.620 --> 00:11:16.860 align:middle line:84% and you can map now from your actual image now into the latent 00:11:16.860 --> 00:11:19.900 align:middle line:84% space itself-- so this is that G inverse-- 00:11:19.900 --> 00:11:22.620 align:middle line:84% now, you have the ability to ask these interesting counterfactual 00:11:22.620 --> 00:11:23.340 align:middle line:90% questions. 00:11:23.340 --> 00:11:25.460 align:middle line:84% What would it look like if assuming 00:11:25.460 --> 00:11:28.260 align:middle line:84% that you have these measurable and actionable dimensions 00:11:28.260 --> 00:11:29.940 align:middle line:90% of variation? 00:11:29.940 --> 00:11:32.400 align:middle line:90% So here, now, you could say, OK. 00:11:32.400 --> 00:11:35.580 align:middle line:84% What would it look like if we moved around 00:11:35.580 --> 00:11:36.605 align:middle line:90% in this latent space? 00:11:36.605 --> 00:11:37.980 align:middle line:84% How do these different dimensions 00:11:37.980 --> 00:11:42.160 align:middle line:84% correspond to different maybe things 00:11:42.160 --> 00:11:45.080 align:middle line:84% within the manifold of real images? 00:11:45.080 --> 00:11:47.160 align:middle line:84% So here's some work where you can actually 00:11:47.160 --> 00:11:54.240 align:middle line:84% show that you can improve your ability to categorize something 00:11:54.240 --> 00:11:59.800 align:middle line:84% by building an ensemble over different generated variations 00:11:59.800 --> 00:12:02.320 align:middle line:84% of that input data in some manifold. 00:12:02.320 --> 00:12:04.520 align:middle line:84% And so you get some, essentially, 00:12:04.520 --> 00:12:08.760 align:middle line:84% improved accuracy and robustness by taking a real input image 00:12:08.760 --> 00:12:12.060 align:middle line:84% and then figuring out where that corresponds to in latent space, 00:12:12.060 --> 00:12:15.160 align:middle line:84% and then moving around nearby in the neighborhood of that thing 00:12:15.160 --> 00:12:17.900 align:middle line:84% to find some maybe different poses, different orientations, 00:12:17.900 --> 00:12:20.640 align:middle line:84% slightly different realistic manipulations of that input 00:12:20.640 --> 00:12:23.120 align:middle line:84% image, and then taking an ensemble. 00:12:23.120 --> 00:12:26.840 align:middle line:84% So it's almost like you're building some robustness 00:12:26.840 --> 00:12:30.880 align:middle line:84% into your categorizer based on some interesting dimensions 00:12:30.880 --> 00:12:33.680 align:middle line:84% of data augmentation, almost like test time 00:12:33.680 --> 00:12:38.820 align:middle line:84% augmentation within the generator itself, 00:12:38.820 --> 00:12:39.700 align:middle line:90% which is interesting. 00:12:39.700 --> 00:12:41.200 align:middle line:84% So there's some nice work that shows 00:12:41.200 --> 00:12:44.420 align:middle line:84% that, OK, actually, exploring dimensions of latent space 00:12:44.420 --> 00:12:46.940 align:middle line:84% can give you something that actually 00:12:46.940 --> 00:12:49.605 align:middle line:84% is more than the original training data. 00:12:49.605 --> 00:12:51.980 align:middle line:84% Though I think an interesting and important point here is 00:12:51.980 --> 00:12:55.620 align:middle line:84% there's so much kind of knowledge baked into this. 00:12:55.620 --> 00:12:58.980 align:middle line:84% The fact that something nearby in that image space 00:12:58.980 --> 00:13:02.560 align:middle line:84% should be reasonably-- maybe the same category, 00:13:02.560 --> 00:13:04.260 align:middle line:84% not actually break a category boundary 00:13:04.260 --> 00:13:07.380 align:middle line:84% in some significant way, seems important. 00:13:07.380 --> 00:13:09.780 align:middle line:84% And I often try to point out, some 00:13:09.780 --> 00:13:12.260 align:middle line:84% of these assumptions that are built in, some of the ways 00:13:12.260 --> 00:13:13.200 align:middle line:90% they might fail. 00:13:13.200 --> 00:13:17.100 align:middle line:84% So if you are just trying to understand something like cat, 00:13:17.100 --> 00:13:18.475 align:middle line:84% this is probably very reasonable. 00:13:18.475 --> 00:13:20.142 align:middle line:84% If you're trying to understand something 00:13:20.142 --> 00:13:22.660 align:middle line:84% about maybe a breed of cat, these dimensions of variation 00:13:22.660 --> 00:13:24.910 align:middle line:84% might start to get more confusing in terms of breaking 00:13:24.910 --> 00:13:25.840 align:middle line:90% that class boundary. 00:13:25.840 --> 00:13:27.500 align:middle line:84% Because to me, as someone who spends 00:13:27.500 --> 00:13:29.700 align:middle line:84% a lot of time thinking about animals, 00:13:29.700 --> 00:13:32.180 align:middle line:84% the first image and the one in the middle 00:13:32.180 --> 00:13:33.725 align:middle line:84% actually look like different breeds. 00:13:33.725 --> 00:13:35.100 align:middle line:84% That difference in the morphology 00:13:35.100 --> 00:13:37.120 align:middle line:84% with that extra length of the ears? 00:13:37.120 --> 00:13:41.155 align:middle line:84% To me, if that's this type of cat, the fine grained level 00:13:41.155 --> 00:13:42.780 align:middle line:84% of categorization you're interested in, 00:13:42.780 --> 00:13:44.360 align:middle line:84% this might actually be more confusing 00:13:44.360 --> 00:13:46.043 align:middle line:84% and potentially break that robustness. 00:13:46.043 --> 00:13:47.960 align:middle line:84% So there's always some fundamental assumptions 00:13:47.960 --> 00:13:49.335 align:middle line:84% about what you're actually trying 00:13:49.335 --> 00:13:52.640 align:middle line:84% to do that are built in to the dimensions of variation that 00:13:52.640 --> 00:13:57.120 align:middle line:84% are good or bad for a given problem. 00:13:57.120 --> 00:13:58.920 align:middle line:84% And then there's this question of, how 00:13:58.920 --> 00:14:02.283 align:middle line:84% do you discover these dimensions in a latent space, right? 00:14:02.283 --> 00:14:04.200 align:middle line:84% So one interesting thing you could think about 00:14:04.200 --> 00:14:07.175 align:middle line:84% is, if you're trying to discover an interesting dimension 00:14:07.175 --> 00:14:08.800 align:middle line:84% of variation in a latent space, there's 00:14:08.800 --> 00:14:10.217 align:middle line:84% ways you can do it experimentally. 00:14:10.217 --> 00:14:12.020 align:middle line:90% You take an image of something. 00:14:12.020 --> 00:14:14.920 align:middle line:84% You take an image of the same thing at night, for example, 00:14:14.920 --> 00:14:17.980 align:middle line:84% and then you look at where those both mapped to in latent space. 00:14:17.980 --> 00:14:20.400 align:middle line:84% And then you decide, you calculate 00:14:20.400 --> 00:14:24.020 align:middle line:84% the vector of direction that goes from one to the other. 00:14:24.020 --> 00:14:26.640 align:middle line:84% So now, you've almost calculated a vector of direction 00:14:26.640 --> 00:14:30.480 align:middle line:84% that should correspond to this specific type of counterfactual. 00:14:30.480 --> 00:14:33.320 align:middle line:84% And so then you can think about explicitly finding things 00:14:33.320 --> 00:14:37.300 align:middle line:84% like dimensions of variation that zoom out on an image 00:14:37.300 --> 00:14:40.580 align:middle line:84% or zoom in on an image or brighten it or darken it. 00:14:40.580 --> 00:14:43.500 align:middle line:84% So in this case, they were able to find 00:14:43.500 --> 00:14:46.420 align:middle line:84% corresponding specific shifts in the image that 00:14:46.420 --> 00:14:50.140 align:middle line:84% can be used to then control images and ones that 00:14:50.140 --> 00:14:51.200 align:middle line:90% are input agnostic. 00:14:51.200 --> 00:14:54.380 align:middle line:84% So now, you can find this optimal direction or dimension 00:14:54.380 --> 00:14:58.260 align:middle line:84% of variation by just kind of cropping images. 00:14:58.260 --> 00:15:00.375 align:middle line:84% And then you take an ensemble of the directions 00:15:00.375 --> 00:15:02.500 align:middle line:84% for each of those, and maybe you can find something 00:15:02.500 --> 00:15:03.440 align:middle line:90% experimentally. 00:15:03.440 --> 00:15:07.160 align:middle line:84% And so here, you have these dimensions of zoom. 00:15:07.160 --> 00:15:09.780 align:middle line:84% You can find dimensions that correspond to shift even 00:15:09.780 --> 00:15:11.700 align:middle line:90% across different input sizes. 00:15:11.700 --> 00:15:13.340 align:middle line:84% You have dimensions that correspond 00:15:13.340 --> 00:15:15.340 align:middle line:90% to brightening an image. 00:15:15.340 --> 00:15:18.020 align:middle line:84% And then you have things like corresponding dimensions 00:15:18.020 --> 00:15:19.440 align:middle line:90% of darkening an image. 00:15:19.440 --> 00:15:21.380 align:middle line:84% So here, this is interesting, because I 00:15:21.380 --> 00:15:23.620 align:middle line:84% would say that this is really capturing some very 00:15:23.620 --> 00:15:25.280 align:middle line:90% specific bias in our data. 00:15:25.280 --> 00:15:29.620 align:middle line:84% Because clearly, people are not taking pictures of volcanoes 00:15:29.620 --> 00:15:31.820 align:middle line:84% in the dark that aren't erupting. 00:15:31.820 --> 00:15:34.640 align:middle line:84% So when we try to make it dark, it thinks it needs to erupt, 00:15:34.640 --> 00:15:35.140 align:middle line:90% right? 00:15:35.140 --> 00:15:36.800 align:middle line:84% There's this clear bias that's just 00:15:36.800 --> 00:15:39.240 align:middle line:90% built into the training data. 00:15:39.240 --> 00:15:40.880 align:middle line:84% And there are some other cool examples 00:15:40.880 --> 00:15:42.940 align:middle line:84% of this for other types of modalities as well. 00:15:42.940 --> 00:15:46.500 align:middle line:84% So there's actually a really famous example from word2vec, 00:15:46.500 --> 00:15:50.520 align:middle line:84% which is an embedding model for words like semantic meanings, 00:15:50.520 --> 00:15:53.720 align:middle line:84% where there's a direction in the latent 00:15:53.720 --> 00:15:57.440 align:middle line:84% space of models that will correspond to changing 00:15:57.440 --> 00:15:58.800 align:middle line:90% the tense of the word. 00:15:58.800 --> 00:16:00.200 align:middle line:84% So you can move in that direction 00:16:00.200 --> 00:16:03.240 align:middle line:84% and go from swimming to swam and go 00:16:03.240 --> 00:16:06.600 align:middle line:84% from walking to walk by moving the same direction 00:16:06.600 --> 00:16:09.360 align:middle line:90% in that latent space for words. 00:16:09.360 --> 00:16:14.520 align:middle line:84% So these types of examples of disentangled dimensions 00:16:14.520 --> 00:16:16.760 align:middle line:84% of variation in latent space can be discovered 00:16:16.760 --> 00:16:18.320 align:middle line:84% for many modalities and have been 00:16:18.320 --> 00:16:19.940 align:middle line:90% discovered for many modalities. 00:16:19.940 --> 00:16:23.280 align:middle line:90% 00:16:23.280 --> 00:16:27.840 align:middle line:84% And then there's kind of like these explicit latent 00:16:27.840 --> 00:16:31.580 align:middle line:84% space vectors that are disentangling 00:16:31.580 --> 00:16:32.720 align:middle line:90% these factors of variation. 00:16:32.720 --> 00:16:35.220 align:middle line:84% So you have things like winter to spring, 00:16:35.220 --> 00:16:38.860 align:middle line:84% or turning on the lights, going from day to night. 00:16:38.860 --> 00:16:40.940 align:middle line:90% The volcano eruption vector. 00:16:40.940 --> 00:16:43.260 align:middle line:84% Though, again, you'll note that you're 00:16:43.260 --> 00:16:46.660 align:middle line:84% more likely to get a more aggressive looking 00:16:46.660 --> 00:16:51.060 align:middle line:84% eruptions at nighttime than during the day. 00:16:51.060 --> 00:16:53.580 align:middle line:84% So there's been a lot of research 00:16:53.580 --> 00:16:56.980 align:middle line:84% into how to efficiently discover these dimensions of variation 00:16:56.980 --> 00:16:59.300 align:middle line:84% within a latent space, these disentangled dimensions 00:16:59.300 --> 00:17:02.020 align:middle line:84% of variation, so let's look at that in a little bit more 00:17:02.020 --> 00:17:03.620 align:middle line:90% technical detail. 00:17:03.620 --> 00:17:07.700 align:middle line:84% So you can take a GAN, like a generative adversarial network, 00:17:07.700 --> 00:17:11.859 align:middle line:84% and you can walk in a straight line in a latent space 00:17:11.859 --> 00:17:15.260 align:middle line:84% and visualize what it looks like to walk in that straight line. 00:17:15.260 --> 00:17:17.960 align:middle line:84% And it kind of works, but it can be really cherry picking. 00:17:17.960 --> 00:17:21.099 align:middle line:84% So some of those directions correspond 00:17:21.099 --> 00:17:23.140 align:middle line:84% to stuff that just looks like trash 00:17:23.140 --> 00:17:25.540 align:middle line:84% or is just not easily disentangled into something 00:17:25.540 --> 00:17:27.220 align:middle line:90% that we can interpret. 00:17:27.220 --> 00:17:29.840 align:middle line:84% And then there's some of fancier versions of a GAN, 00:17:29.840 --> 00:17:32.840 align:middle line:84% like StyleGAN, where they actually explicitly noticed 00:17:32.840 --> 00:17:37.680 align:middle line:84% that any layer of the model, any of those intermediate 00:17:37.680 --> 00:17:41.300 align:middle line:84% representations, also represents a latent space for the model. 00:17:41.300 --> 00:17:45.080 align:middle line:84% And so they determined that there's actually 00:17:45.080 --> 00:17:49.020 align:middle line:84% earlier layers in the model that are better, 00:17:49.020 --> 00:17:52.480 align:middle line:84% quote, unquote, latent representations than 00:17:52.480 --> 00:17:55.920 align:middle line:84% the z representation, that initial input representation. 00:17:55.920 --> 00:18:00.060 align:middle line:84% And so then we'll try and talk about why that might be, 00:18:00.060 --> 00:18:02.720 align:middle line:84% why it might be better to actually think about the latent 00:18:02.720 --> 00:18:05.320 align:middle line:84% space from somewhere intermediate in the model 00:18:05.320 --> 00:18:10.160 align:middle line:84% versus explicitly with that input latent vector. 00:18:10.160 --> 00:18:16.040 align:middle line:84% So if we have some natural image manifold, X, 00:18:16.040 --> 00:18:20.920 align:middle line:84% so some nonlinear data space, and now, 00:18:20.920 --> 00:18:25.320 align:middle line:84% we have some starting point, so this is the corresponding point 00:18:25.320 --> 00:18:28.140 align:middle line:84% in that data natural image manifold corresponding 00:18:28.140 --> 00:18:29.900 align:middle line:90% to this bird, this blue bird. 00:18:29.900 --> 00:18:32.580 align:middle line:84% Maybe you have another point here. 00:18:32.580 --> 00:18:34.740 align:middle line:84% This is corresponding to this fly. 00:18:34.740 --> 00:18:37.460 align:middle line:84% So if you think about linear interpolation between these two 00:18:37.460 --> 00:18:43.660 align:middle line:84% points, arguably, you're moving off the manifold of real images. 00:18:43.660 --> 00:18:48.740 align:middle line:84% There's stuff in one end and in the other that 00:18:48.740 --> 00:18:52.860 align:middle line:84% are more realistic looking, but this direct linear interpolation 00:18:52.860 --> 00:18:54.500 align:middle line:84% in image space just gives you stuff 00:18:54.500 --> 00:18:58.060 align:middle line:84% in the middle that's basically these obvious kind of just 00:18:58.060 --> 00:19:00.020 align:middle line:90% additive images. 00:19:00.020 --> 00:19:03.740 align:middle line:84% And it doesn't look anything like any realistic natural world 00:19:03.740 --> 00:19:04.280 align:middle line:90% image would. 00:19:04.280 --> 00:19:05.780 align:middle line:84% It just doesn't match the statistics 00:19:05.780 --> 00:19:07.420 align:middle line:90% of real natural images. 00:19:07.420 --> 00:19:13.100 align:middle line:84% So instead, if we have some sort of latent space that-- 00:19:13.100 --> 00:19:15.600 align:middle line:84% now, we've learned this data plus plus representation. 00:19:15.600 --> 00:19:18.840 align:middle line:84% So now, we have a latent space that's well-behaved, 00:19:18.840 --> 00:19:22.900 align:middle line:84% and we can map from that directly to that natural image 00:19:22.900 --> 00:19:24.340 align:middle line:90% manifold. 00:19:24.340 --> 00:19:27.040 align:middle line:84% Then, you can imagine that interpolating linearly 00:19:27.040 --> 00:19:29.920 align:middle line:84% in the latent space would correspond to a nonlinear 00:19:29.920 --> 00:19:32.060 align:middle line:84% interpolation in that data manifold space, 00:19:32.060 --> 00:19:34.320 align:middle line:84% but it would ensure that you're mapping 00:19:34.320 --> 00:19:37.340 align:middle line:84% within the manifold of real natural images the whole way. 00:19:37.340 --> 00:19:40.400 align:middle line:84% Because by definition, we've constructed something 00:19:40.400 --> 00:19:42.200 align:middle line:84% where any point in that latent space 00:19:42.200 --> 00:19:45.680 align:middle line:84% will map to the natural image manifold. 00:19:45.680 --> 00:19:49.200 align:middle line:84% So now, if you do that same interpolation-- and this is 00:19:49.200 --> 00:19:51.840 align:middle line:84% from a paper called BigGAN back in 2018-- 00:19:51.840 --> 00:19:55.800 align:middle line:84% you do get stuff that at least looks a little bit more 00:19:55.800 --> 00:19:57.300 align:middle line:90% like natural images. 00:19:57.300 --> 00:19:59.300 align:middle line:84% But they might not really map to reality. 00:19:59.300 --> 00:20:02.780 align:middle line:84% You get some really weird things like this half bird, 00:20:02.780 --> 00:20:04.840 align:middle line:84% half fly thing where, basically, the model 00:20:04.840 --> 00:20:06.080 align:middle line:90% is just doing its best. 00:20:06.080 --> 00:20:08.080 align:middle line:84% It's doing its best to find some way 00:20:08.080 --> 00:20:10.480 align:middle line:84% to reasonably interpolate between these two things 00:20:10.480 --> 00:20:13.280 align:middle line:84% and stay within the constraints of that trained 00:20:13.280 --> 00:20:17.160 align:middle line:90% natural image manifold mapping. 00:20:17.160 --> 00:20:20.255 align:middle line:84% But that z space, so this z space 00:20:20.255 --> 00:20:21.880 align:middle line:84% here where we're mapping from one point 00:20:21.880 --> 00:20:25.940 align:middle line:84% to another in a linear way, might not 00:20:25.940 --> 00:20:30.560 align:middle line:84% be the best way to organize the data to interpolate along. 00:20:30.560 --> 00:20:33.140 align:middle line:84% And so let's look at why that might be. 00:20:33.140 --> 00:20:36.640 align:middle line:84% So it goes back to the lecture from VAEs. 00:20:36.640 --> 00:20:39.460 align:middle line:84% If you guys remember, you have this data distribution 00:20:39.460 --> 00:20:41.275 align:middle line:84% that maps to a latent distribution, 00:20:41.275 --> 00:20:42.900 align:middle line:84% and you have a latent distribution that 00:20:42.900 --> 00:20:44.780 align:middle line:90% maps to a data distribution. 00:20:44.780 --> 00:20:47.500 align:middle line:84% And so at the bottom, we're coloring the correspondence 00:20:47.500 --> 00:20:49.140 align:middle line:84% between those two spaces, essentially, 00:20:49.140 --> 00:20:51.980 align:middle line:84% like if you're going to map these two things back and forth. 00:20:51.980 --> 00:20:57.540 align:middle line:84% So here, if I say the same color on the natural image manifold 00:20:57.540 --> 00:21:00.140 align:middle line:84% of green, for example, would match the same color green 00:21:00.140 --> 00:21:01.860 align:middle line:90% in the latent space. 00:21:01.860 --> 00:21:07.580 align:middle line:84% So now, let's visualize what happens as this gets trained. 00:21:07.580 --> 00:21:10.860 align:middle line:84% So here, if we're watching this thing get trained, 00:21:10.860 --> 00:21:16.180 align:middle line:84% you can see how it's going from the natural image manifold 00:21:16.180 --> 00:21:17.120 align:middle line:90% to that latent space. 00:21:17.120 --> 00:21:19.780 align:middle line:84% It's almost like you've taken something and tried 00:21:19.780 --> 00:21:22.040 align:middle line:90% to crumple it into a ball. 00:21:22.040 --> 00:21:24.480 align:middle line:84% And so even though we've said explicitly 00:21:24.480 --> 00:21:26.600 align:middle line:84% that, OK, everything in that latent space 00:21:26.600 --> 00:21:29.480 align:middle line:84% should correspond to something on the natural image manifold, 00:21:29.480 --> 00:21:32.400 align:middle line:84% actually, there's this weird stuff here. 00:21:32.400 --> 00:21:37.898 align:middle line:84% There's danger zones where essentially, here, it's 00:21:37.898 --> 00:21:39.940 align:middle line:84% going to be closer to the natural image manifold. 00:21:39.940 --> 00:21:41.920 align:middle line:84% But if you're going from here to here, 00:21:41.920 --> 00:21:46.440 align:middle line:84% there's going to maybe be seams in that linear interpolation 00:21:46.440 --> 00:21:52.040 align:middle line:84% that will correspond to pretty jarring shifts 00:21:52.040 --> 00:21:55.000 align:middle line:84% where things that are nearby in latent space 00:21:55.000 --> 00:21:57.640 align:middle line:84% could potentially be quite far away from each other in data 00:21:57.640 --> 00:21:59.360 align:middle line:90% space. 00:21:59.360 --> 00:22:01.940 align:middle line:84% And so essentially, that z representation, 00:22:01.940 --> 00:22:05.060 align:middle line:84% if you're doing linear interpolation across that seam, 00:22:05.060 --> 00:22:06.540 align:middle line:90% it could be unnatural. 00:22:06.540 --> 00:22:10.000 align:middle line:84% And that might be one of the ways 00:22:10.000 --> 00:22:12.560 align:middle line:84% to motivate the discovery that they 00:22:12.560 --> 00:22:15.200 align:middle line:84% found in the StyleGAN paper, which is essentially 00:22:15.200 --> 00:22:18.600 align:middle line:84% that using some intermediate representation somewhere 00:22:18.600 --> 00:22:22.180 align:middle line:84% in between that's not quite fully convex, but maybe 00:22:22.180 --> 00:22:25.980 align:middle line:84% like a little bit less distorted-- 00:22:25.980 --> 00:22:27.860 align:middle line:84% so some kind of intermediate thing 00:22:27.860 --> 00:22:32.340 align:middle line:84% where you're possibly less likely to jump across a seam 00:22:32.340 --> 00:22:35.820 align:middle line:90% or fall into some danger zone. 00:22:35.820 --> 00:22:37.880 align:middle line:84% Basically, just somewhere in between 00:22:37.880 --> 00:22:40.220 align:middle line:84% might be the right place to interpolate and get 00:22:40.220 --> 00:22:43.740 align:middle line:84% things that look more realistic and behave 00:22:43.740 --> 00:22:46.140 align:middle line:90% in a more friendly way. 00:22:46.140 --> 00:22:48.780 align:middle line:84% And empirically and qualitatively, 00:22:48.780 --> 00:22:51.180 align:middle line:90% that's sometimes true. 00:22:51.180 --> 00:22:53.300 align:middle line:84% It's our favorite way to say things, right? 00:22:53.300 --> 00:22:54.260 align:middle line:90% Sometimes. 00:22:54.260 --> 00:22:56.960 align:middle line:84% But it does actually correspond in some nice ways. 00:22:56.960 --> 00:23:02.200 align:middle line:84% So here, this was a work where they took that w space, 00:23:02.200 --> 00:23:05.740 align:middle line:84% that intermediate space, and they demonstrated that they 00:23:05.740 --> 00:23:09.500 align:middle line:84% could use it using-- this was a work called StyleSpace-- 00:23:09.500 --> 00:23:12.350 align:middle line:84% to very cleanly isolate and disentangle 00:23:12.350 --> 00:23:13.600 align:middle line:90% these dimensions of variation. 00:23:13.600 --> 00:23:15.683 align:middle line:84% And they got some really nice qualitative results. 00:23:15.683 --> 00:23:19.040 align:middle line:84% So here, they're able to vary things 00:23:19.040 --> 00:23:23.240 align:middle line:84% at the level of granularity of hood styles or headlight types 00:23:23.240 --> 00:23:26.340 align:middle line:90% or body colors or background. 00:23:26.340 --> 00:23:29.120 align:middle line:90% 00:23:29.120 --> 00:23:29.620 align:middle line:90% Yeah. 00:23:29.620 --> 00:23:32.200 align:middle line:84% So it does seem to suggest that basically 00:23:32.200 --> 00:23:37.440 align:middle line:84% by removing some of those, like weird, badly behaved portions 00:23:37.440 --> 00:23:40.400 align:middle line:84% of linear interpolation, by moving to this more intermediate 00:23:40.400 --> 00:23:42.040 align:middle line:84% space, is beneficial when you're trying 00:23:42.040 --> 00:23:45.720 align:middle line:84% to find these disentangled dimensions of variation. 00:23:45.720 --> 00:23:47.040 align:middle line:90% Cool. 00:23:47.040 --> 00:23:51.780 align:middle line:84% So now, let's talk about how you would label this generated data. 00:23:51.780 --> 00:23:55.180 align:middle line:84% So you have a label for all your real images. 00:23:55.180 --> 00:23:57.120 align:middle line:84% Well, maybe you don't, but assume you do. 00:23:57.120 --> 00:23:58.960 align:middle line:84% Assume that the data you started from, 00:23:58.960 --> 00:24:00.877 align:middle line:84% that you're training this generative model on, 00:24:00.877 --> 00:24:02.040 align:middle line:90% you had labeled. 00:24:02.040 --> 00:24:05.800 align:middle line:84% And now, you have some space in here 00:24:05.800 --> 00:24:08.220 align:middle line:84% that corresponds to that real data. 00:24:08.220 --> 00:24:10.000 align:middle line:84% But then there's a lot of other space 00:24:10.000 --> 00:24:14.040 align:middle line:84% that corresponds to purely generated data. 00:24:14.040 --> 00:24:16.860 align:middle line:84% So the argument here is that by construction, 00:24:16.860 --> 00:24:20.700 align:middle line:84% that things that are close together in this manifold, 00:24:20.700 --> 00:24:22.880 align:middle line:90% should be semantically related. 00:24:22.880 --> 00:24:26.040 align:middle line:84% And so if you don't move too far on the manifold, 00:24:26.040 --> 00:24:30.220 align:middle line:84% you should argue that the category should stay the same, 00:24:30.220 --> 00:24:32.420 align:middle line:84% or at least this is the assumption. 00:24:32.420 --> 00:24:35.940 align:middle line:84% And so then you can rely on the normal inductive bias 00:24:35.940 --> 00:24:37.940 align:middle line:84% and machine learning that assigns similar labels 00:24:37.940 --> 00:24:40.160 align:middle line:84% to nearby points, if it's reasonably well 00:24:40.160 --> 00:24:41.620 align:middle line:90% semantically clustered. 00:24:41.620 --> 00:24:44.140 align:middle line:84% So here, if you know, for example, these 00:24:44.140 --> 00:24:47.220 align:middle line:84% are real data points where you know their true label, 00:24:47.220 --> 00:24:49.740 align:middle line:84% then you could assume that things sampled 00:24:49.740 --> 00:24:52.180 align:middle line:84% from those similar data points would actually 00:24:52.180 --> 00:24:54.300 align:middle line:90% correspond to similar things. 00:24:54.300 --> 00:24:57.800 align:middle line:84% And so here, maybe in this region, 00:24:57.800 --> 00:25:00.527 align:middle line:84% yes, there's dimensions of variation that are captured, 00:25:00.527 --> 00:25:02.860 align:middle line:84% but they're not changing the underlying semantic meaning 00:25:02.860 --> 00:25:06.100 align:middle line:84% of the object itself, that these small variations are 00:25:06.100 --> 00:25:09.140 align:middle line:84% more corresponding to these dimensions of variation 00:25:09.140 --> 00:25:14.560 align:middle line:84% in the context, in maybe the scene, the visualization, 00:25:14.560 --> 00:25:20.240 align:middle line:84% but it's not changing the category that's being captured. 00:25:20.240 --> 00:25:22.240 align:middle line:84% And so this was explicitly explored 00:25:22.240 --> 00:25:26.640 align:middle line:84% in this paper called DatasetGAN where they basically 00:25:26.640 --> 00:25:29.840 align:middle line:84% trained a GAN on one type of data, 00:25:29.840 --> 00:25:34.760 align:middle line:84% and then they wanted to solve a new task related to that data, 00:25:34.760 --> 00:25:38.400 align:middle line:84% so specifically labeling the parts of a car 00:25:38.400 --> 00:25:41.000 align:middle line:90% with semantic segmentation. 00:25:41.000 --> 00:25:43.520 align:middle line:84% And so then what they did is they used StyleGAN 00:25:43.520 --> 00:25:46.120 align:middle line:84% to efficiently label data in terms 00:25:46.120 --> 00:25:48.200 align:middle line:84% of the semantic segmentation labels, which 00:25:48.200 --> 00:25:50.200 align:middle line:84% can be quite expensive, by defining 00:25:50.200 --> 00:25:54.000 align:middle line:84% an additional labeling arm, which is essentially 00:25:54.000 --> 00:25:56.960 align:middle line:84% the output of a StyleGAN with just a few examples. 00:25:56.960 --> 00:26:00.400 align:middle line:84% And then you can use that model to generate a really large data 00:26:00.400 --> 00:26:02.662 align:middle line:84% set of images and corresponding labels. 00:26:02.662 --> 00:26:04.120 align:middle line:84% They are weak labels, but there are 00:26:04.120 --> 00:26:08.740 align:middle line:84% labels that are explicitly based on this well-structured, 00:26:08.740 --> 00:26:11.100 align:middle line:90% well-aligned latent space. 00:26:11.100 --> 00:26:13.100 align:middle line:84% And then you can train a part segmentation model 00:26:13.100 --> 00:26:15.420 align:middle line:84% on the synthetic data and test it on the real data 00:26:15.420 --> 00:26:17.540 align:middle line:84% and show that it's actually beneficial. 00:26:17.540 --> 00:26:21.340 align:middle line:84% So being a little bit more specific about this, 00:26:21.340 --> 00:26:24.940 align:middle line:84% essentially, if you're trying to train that labeling arm, 00:26:24.940 --> 00:26:27.660 align:middle line:84% they do that using the latent space and then just 00:26:27.660 --> 00:26:31.820 align:middle line:84% a really, really small number of manually labeled examples. 00:26:31.820 --> 00:26:33.540 align:middle line:84% And this is valuable, particularly 00:26:33.540 --> 00:26:36.873 align:middle line:84% because semantic segmentation labels are really expensive. 00:26:36.873 --> 00:26:38.540 align:middle line:84% How many of you have ever actually tried 00:26:38.540 --> 00:26:42.100 align:middle line:84% to label data for semantic segmentation? 00:26:42.100 --> 00:26:42.600 align:middle line:90% Yeah. 00:26:42.600 --> 00:26:43.740 align:middle line:90% It sucks. 00:26:43.740 --> 00:26:45.780 align:middle line:90% It's super slow. 00:26:45.780 --> 00:26:48.380 align:middle line:84% Especially the parts that are really fine 00:26:48.380 --> 00:26:50.480 align:middle line:84% grained, like the boundaries of objects, 00:26:50.480 --> 00:26:54.500 align:middle line:84% it just takes forever to really try to get it very correct. 00:26:54.500 --> 00:26:58.780 align:middle line:84% And so being able to efficiently benefit from just a few of those 00:26:58.780 --> 00:27:02.500 align:middle line:84% and then be able to generalize well is pretty valuable. 00:27:02.500 --> 00:27:06.500 align:middle line:84% So essentially, here, you map human labels in the pixel space 00:27:06.500 --> 00:27:09.980 align:middle line:84% through the network to some of these intermediate features that 00:27:09.980 --> 00:27:14.880 align:middle line:84% have some sort of maintained spatial relationships 00:27:14.880 --> 00:27:16.480 align:middle line:90% and arrangement. 00:27:16.480 --> 00:27:20.547 align:middle line:84% So basically, you take the original data, 00:27:20.547 --> 00:27:22.880 align:middle line:84% and then you run it through what they're calling a style 00:27:22.880 --> 00:27:24.255 align:middle line:84% interpreter, but it's essentially 00:27:24.255 --> 00:27:26.000 align:middle line:90% something like a StyleGAN. 00:27:26.000 --> 00:27:33.160 align:middle line:84% That gets you a representation of these specific pixel vectors 00:27:33.160 --> 00:27:37.420 align:middle line:84% that you can then train a model, a really efficient model, 00:27:37.420 --> 00:27:40.400 align:middle line:84% predictive model, to map between those features and part labels. 00:27:40.400 --> 00:27:43.080 align:middle line:84% So essentially, it's like relying on the fact 00:27:43.080 --> 00:27:45.320 align:middle line:84% that training this generator forced the model 00:27:45.320 --> 00:27:49.720 align:middle line:84% to learn a really well-behaved, well-aligned representation 00:27:49.720 --> 00:27:52.680 align:middle line:84% space in terms of these dimensions of variation 00:27:52.680 --> 00:27:53.743 align:middle line:90% and similarity. 00:27:53.743 --> 00:27:56.160 align:middle line:84% And then they could show that you could really efficiently 00:27:56.160 --> 00:28:00.480 align:middle line:84% train these accurate predictive models because the features 00:28:00.480 --> 00:28:03.560 align:middle line:84% themselves were better organized and more semantically meaningful 00:28:03.560 --> 00:28:06.120 align:middle line:84% than the pixels that you might be starting with. 00:28:06.120 --> 00:28:11.260 align:middle line:84% So instead of trying to go from images and pixel labels 00:28:11.260 --> 00:28:14.460 align:middle line:84% to a trained model that generates pixel labels given 00:28:14.460 --> 00:28:17.620 align:middle line:84% this input image, you basically take the input image, 00:28:17.620 --> 00:28:19.920 align:middle line:84% run it through to some intermediate feature space, 00:28:19.920 --> 00:28:21.878 align:middle line:84% and then you've trained this really lightweight 00:28:21.878 --> 00:28:24.220 align:middle line:84% predictive model to go from those features that 00:28:24.220 --> 00:28:25.980 align:middle line:84% are learned by a generative model 00:28:25.980 --> 00:28:30.860 align:middle line:84% to the correct semantic segmentation. 00:28:30.860 --> 00:28:33.020 align:middle line:84% And they had some pretty nice examples 00:28:33.020 --> 00:28:34.920 align:middle line:84% of how they were kind of able to do this. 00:28:34.920 --> 00:28:36.920 align:middle line:84% And qualitatively and quantitatively, 00:28:36.920 --> 00:28:39.340 align:middle line:84% they were able to show that this worked quite well, 00:28:39.340 --> 00:28:42.740 align:middle line:84% and they got this kind of nice result, which 00:28:42.740 --> 00:28:48.180 align:middle line:84% is that a single labeled GAN image was worth about 100 00:28:48.180 --> 00:28:51.900 align:middle line:84% labeled regular images, just in terms of the training efficiency 00:28:51.900 --> 00:28:56.140 align:middle line:84% versus the same amount of accuracy. 00:28:56.140 --> 00:29:00.570 align:middle line:84% And these types of things are increasingly commonly explored, 00:29:00.570 --> 00:29:02.820 align:middle line:84% particularly for these types of labels that are really 00:29:02.820 --> 00:29:05.300 align:middle line:90% expensive to collect. 00:29:05.300 --> 00:29:08.800 align:middle line:84% And interestingly, this question of, 00:29:08.800 --> 00:29:11.920 align:middle line:84% how do you efficiently train these types of things based 00:29:11.920 --> 00:29:14.780 align:middle line:84% on well-learned representation spaces or foundation models, 00:29:14.780 --> 00:29:16.882 align:middle line:84% even ones that are not necessarily generative? 00:29:16.882 --> 00:29:18.340 align:middle line:84% has also been really well explored. 00:29:18.340 --> 00:29:19.800 align:middle line:84% And so these days, if you're trying 00:29:19.800 --> 00:29:23.080 align:middle line:84% to do semantic segmentation, most people 00:29:23.080 --> 00:29:26.200 align:middle line:84% start from something that's called SAM, the Segment Anything 00:29:26.200 --> 00:29:28.480 align:middle line:84% Model, which is a model that relied essentially 00:29:28.480 --> 00:29:32.980 align:middle line:84% on building a really large data set 00:29:32.980 --> 00:29:35.740 align:middle line:84% using some pretty cute hacks for self-supervision. 00:29:35.740 --> 00:29:38.960 align:middle line:84% You take an object, and you use copy paste augmentation 00:29:38.960 --> 00:29:41.700 align:middle line:84% to guarantee that you have that correct mask now 00:29:41.700 --> 00:29:44.080 align:middle line:84% in a bunch of different relationships and orientations. 00:29:44.080 --> 00:29:46.600 align:middle line:84% You use some similar ideas from something 00:29:46.600 --> 00:29:50.720 align:middle line:84% like DINOv2 that really captures fine grained semantic meaning 00:29:50.720 --> 00:29:52.680 align:middle line:84% in those feature spaces, and then you're 00:29:52.680 --> 00:29:55.640 align:middle line:84% able to learn a really generalizable and robust 00:29:55.640 --> 00:29:57.700 align:middle line:84% segmentation model on top of that. 00:29:57.700 --> 00:30:00.920 align:middle line:90% 00:30:00.920 --> 00:30:01.700 align:middle line:90% Cool. 00:30:01.700 --> 00:30:05.040 align:middle line:84% So then you can also think about how 00:30:05.040 --> 00:30:08.460 align:middle line:84% these generative models' data plus plus approach 00:30:08.460 --> 00:30:12.140 align:middle line:84% can start to teach you how to explain things or increase 00:30:12.140 --> 00:30:15.180 align:middle line:84% the interpretability of some of these dimensions. 00:30:15.180 --> 00:30:18.912 align:middle line:84% So in standard classification, we take an image, 00:30:18.912 --> 00:30:20.620 align:middle line:84% we run it through a classification model, 00:30:20.620 --> 00:30:22.560 align:middle line:90% and it tries to predict cat. 00:30:22.560 --> 00:30:23.460 align:middle line:90% It's like, OK. 00:30:23.460 --> 00:30:26.260 align:middle line:90% It tries to predict a category. 00:30:26.260 --> 00:30:30.460 align:middle line:84% But now, the question is, can we say something 00:30:30.460 --> 00:30:34.900 align:middle line:84% interesting about why the image was classified as a cat using 00:30:34.900 --> 00:30:37.060 align:middle line:84% this ability and this data plus plus space 00:30:37.060 --> 00:30:39.940 align:middle line:90% to explore counterfactuals? 00:30:39.940 --> 00:30:42.147 align:middle line:90% So here, you take the cat. 00:30:42.147 --> 00:30:43.980 align:middle line:84% You take it from this data space and turn it 00:30:43.980 --> 00:30:45.780 align:middle line:84% into this quote, unquote, data plus plus 00:30:45.780 --> 00:30:50.420 align:middle line:84% space where now you have some maybe encoder, some generator. 00:30:50.420 --> 00:30:52.540 align:middle line:84% So you're mapping through, essentially, 00:30:52.540 --> 00:30:56.860 align:middle line:84% your G inverse and then through your G to get the same data. 00:30:56.860 --> 00:30:59.540 align:middle line:84% But now what you can do is you can 00:30:59.540 --> 00:31:06.660 align:middle line:84% look from this predicted class using-- from the generator, 00:31:06.660 --> 00:31:09.280 align:middle line:84% looking at the classes that are predicted, 00:31:09.280 --> 00:31:15.320 align:middle line:84% you can now explicitly try to perturb in the latent space 00:31:15.320 --> 00:31:18.920 align:middle line:84% to understand how that would cause your prediction to change. 00:31:18.920 --> 00:31:22.600 align:middle line:84% So this gives you some ranking of the changes in the latent 00:31:22.600 --> 00:31:24.840 align:middle line:84% space and manipulations of these latent variables 00:31:24.840 --> 00:31:28.320 align:middle line:84% that maximally lead to the idea of changing a concept. 00:31:28.320 --> 00:31:32.680 align:middle line:84% So it helps you understand what maybe features of this image 00:31:32.680 --> 00:31:37.160 align:middle line:84% are most identifiable as the category of interest. 00:31:37.160 --> 00:31:40.720 align:middle line:84% And so here, this StylEx, it's like almost like StyleSpace 00:31:40.720 --> 00:31:42.160 align:middle line:90% explanation approach. 00:31:42.160 --> 00:31:46.000 align:middle line:84% It tries to find the topK StyleSpace directions, so 00:31:46.000 --> 00:31:51.160 align:middle line:84% these latent space directions of variation, that will most affect 00:31:51.160 --> 00:31:53.600 align:middle line:90% which class the model predicts. 00:31:53.600 --> 00:31:57.580 align:middle line:84% And so here, there's some interesting stuff, right? 00:31:57.580 --> 00:32:02.260 align:middle line:84% You can look at what happens if you open the mouth of the thing, 00:32:02.260 --> 00:32:04.980 align:middle line:84% and it turns out that that is a really strong dimension 00:32:04.980 --> 00:32:08.040 align:middle line:84% of variation for moving the prediction from cat to dog, 00:32:08.040 --> 00:32:12.980 align:middle line:84% probably because dogs tend to pant a lot more than cats do. 00:32:12.980 --> 00:32:16.100 align:middle line:84% So it's a very cute interpolation. 00:32:16.100 --> 00:32:18.700 align:middle line:90% 00:32:18.700 --> 00:32:23.940 align:middle line:84% You can also look at changing something 00:32:23.940 --> 00:32:26.800 align:middle line:84% about the size of the pupils relative to the eyes, 00:32:26.800 --> 00:32:28.860 align:middle line:84% and that also shifts the probability 00:32:28.860 --> 00:32:30.740 align:middle line:84% that you have something like a cat. 00:32:30.740 --> 00:32:34.840 align:middle line:84% Or you can think about changing how pointy the ears are, 00:32:34.840 --> 00:32:37.340 align:middle line:84% and that also tends to correspond 00:32:37.340 --> 00:32:39.140 align:middle line:84% from an explainability standpoint 00:32:39.140 --> 00:32:42.120 align:middle line:84% to why the model thinks it's a cat versus something else. 00:32:42.120 --> 00:32:46.380 align:middle line:84% And so this kind of gives us this sense of-- it's 00:32:46.380 --> 00:32:50.380 align:middle line:84% just another mechanism to try to probe the underlying 00:32:50.380 --> 00:32:52.680 align:middle line:84% things that the model has learned or has understood. 00:32:52.680 --> 00:32:55.440 align:middle line:84% And you can imagine that this does tell you, again, 00:32:55.440 --> 00:32:58.420 align:middle line:84% something about the underlying bias in our training data, 00:32:58.420 --> 00:33:01.080 align:middle line:84% because there are many cats that have ears that 00:33:01.080 --> 00:33:03.800 align:middle line:90% are kind of shorter like that. 00:33:03.800 --> 00:33:06.820 align:middle line:84% British shorthairs have short ears just like that. 00:33:06.820 --> 00:33:10.040 align:middle line:84% But more often than not, something 00:33:10.040 --> 00:33:11.670 align:middle line:84% that has long, pointy ears is a cat, 00:33:11.670 --> 00:33:13.920 align:middle line:84% and maybe something with a shorter or folded over ears 00:33:13.920 --> 00:33:15.320 align:middle line:90% might be a dog. 00:33:15.320 --> 00:33:17.540 align:middle line:84% So it's capturing some dimensions of bias, 00:33:17.540 --> 00:33:20.933 align:middle line:90% I think, in the data as well. 00:33:20.933 --> 00:33:22.600 align:middle line:84% And so you can think about this in terms 00:33:22.600 --> 00:33:27.680 align:middle line:84% of how we generate maybe class specific explanations. 00:33:27.680 --> 00:33:32.480 align:middle line:84% So here, if you have something like a perceived age classifier, 00:33:32.480 --> 00:33:36.920 align:middle line:84% then you can start really probing these weird dimensions 00:33:36.920 --> 00:33:42.380 align:middle line:84% of bias and see how that affects the perceived age of a model. 00:33:42.380 --> 00:33:47.870 align:middle line:84% So now, the model is predicting an age. 00:33:47.870 --> 00:33:51.040 align:middle line:84% And so one of the things that's quite odd 00:33:51.040 --> 00:33:57.480 align:middle line:84% here is you'll see specifically that thicker eyebrows are 00:33:57.480 --> 00:34:01.580 align:middle line:84% corresponding responding to more youthful age. 00:34:01.580 --> 00:34:05.020 align:middle line:84% Lighter skin is corresponding to more youthful age, though also 00:34:05.020 --> 00:34:09.060 align:middle line:84% you'll note that it's lighter, but also sometimes less 00:34:09.060 --> 00:34:12.120 align:middle line:84% textured, which is one of these dimensions. 00:34:12.120 --> 00:34:14.100 align:middle line:84% It's not a super, super disentangled dimension 00:34:14.100 --> 00:34:15.179 align:middle line:90% of variation. 00:34:15.179 --> 00:34:20.020 align:middle line:84% Or something like adding glasses or very obviously going 00:34:20.020 --> 00:34:22.380 align:middle line:90% to gray hair. 00:34:22.380 --> 00:34:24.370 align:middle line:84% So these were all things that were learned 00:34:24.370 --> 00:34:26.620 align:middle line:84% in terms of finding these dimensions of variation that 00:34:26.620 --> 00:34:29.820 align:middle line:84% corresponded to specific types of prediction. 00:34:29.820 --> 00:34:31.820 align:middle line:90% And it maybe makes sense. 00:34:31.820 --> 00:34:34.760 align:middle line:84% Older people might be, in terms of your data bias, 00:34:34.760 --> 00:34:37.820 align:middle line:90% more likely to have white hair. 00:34:37.820 --> 00:34:41.139 align:middle line:84% That's something that's actually pretty ubiquitous. 00:34:41.139 --> 00:34:43.790 align:middle line:84% But you could also imagine that some 00:34:43.790 --> 00:34:45.540 align:middle line:84% of these dimensions of variation would not 00:34:45.540 --> 00:34:48.940 align:middle line:84% be something you would want a model to necessarily learn. 00:34:48.940 --> 00:34:50.860 align:middle line:84% These might capture dimensions of bias 00:34:50.860 --> 00:34:53.300 align:middle line:84% that may be related to the fairness of the equity 00:34:53.300 --> 00:34:55.739 align:middle line:90% of a model that we deploy. 00:34:55.739 --> 00:34:58.260 align:middle line:84% So then what do you actually do with what you learn? 00:34:58.260 --> 00:35:01.400 align:middle line:84% Well, one way we could look at this 00:35:01.400 --> 00:35:03.840 align:middle line:84% is like, well, instead of looking 00:35:03.840 --> 00:35:05.720 align:middle line:84% at things like these dimensions of variation 00:35:05.720 --> 00:35:08.240 align:middle line:84% from internet images, can we take it 00:35:08.240 --> 00:35:10.920 align:middle line:84% to scientific data or medical data? 00:35:10.920 --> 00:35:15.720 align:middle line:84% So here, this is using that same type of explainability mechanism 00:35:15.720 --> 00:35:19.640 align:middle line:84% from a generative model to try to understand 00:35:19.640 --> 00:35:23.880 align:middle line:84% how to categorize different retinal fundus images. 00:35:23.880 --> 00:35:27.560 align:middle line:84% And here, we're showing the top four examples of things 00:35:27.560 --> 00:35:29.880 align:middle line:84% that real doctors are looking for 00:35:29.880 --> 00:35:34.240 align:middle line:84% and then how you might actually change that image maximally 00:35:34.240 --> 00:35:36.640 align:middle line:84% to correspond to something predictive 00:35:36.640 --> 00:35:39.760 align:middle line:84% of a specific disease, for example. 00:35:39.760 --> 00:35:41.560 align:middle line:84% And this is quite interesting because it 00:35:41.560 --> 00:35:43.620 align:middle line:84% does actually-- if you show these to a doctor, 00:35:43.620 --> 00:35:45.972 align:middle line:84% it does correspond with the types of even very 00:35:45.972 --> 00:35:47.680 align:middle line:84% fine grained features that a doctor might 00:35:47.680 --> 00:35:50.520 align:middle line:84% use to actually categorize a specific type of disease. 00:35:50.520 --> 00:35:53.260 align:middle line:84% So from that sense, from an explainability standpoint, 00:35:53.260 --> 00:35:55.060 align:middle line:84% it's almost a way to build trust in a model 00:35:55.060 --> 00:35:56.760 align:middle line:84% because an expert would say, OK, yeah. 00:35:56.760 --> 00:35:57.680 align:middle line:90% No, I do agree. 00:35:57.680 --> 00:36:01.420 align:middle line:84% I do agree that is a reasonable dimension of variation 00:36:01.420 --> 00:36:05.660 align:middle line:84% that I also would correspond to a specific type of disease. 00:36:05.660 --> 00:36:08.420 align:middle line:84% It's almost a reassurance that the model isn't picking up 00:36:08.420 --> 00:36:11.220 align:middle line:84% too many odd other correlative factors that 00:36:11.220 --> 00:36:15.220 align:middle line:84% might be things that we don't want to model to correspond to. 00:36:15.220 --> 00:36:17.220 align:middle line:84% But of course, one of the really big limitations 00:36:17.220 --> 00:36:20.180 align:middle line:84% of all of these approaches is that you are 00:36:20.180 --> 00:36:22.880 align:middle line:90% required to figure out how-- 00:36:22.880 --> 00:36:25.660 align:middle line:84% you have to discover those latent space variations 00:36:25.660 --> 00:36:27.540 align:middle line:84% and then manually analyze or to find 00:36:27.540 --> 00:36:29.260 align:middle line:90% what they might correspond to. 00:36:29.260 --> 00:36:31.520 align:middle line:84% And that takes a lot of effort, right? 00:36:31.520 --> 00:36:33.180 align:middle line:84% So getting to the point where they 00:36:33.180 --> 00:36:36.860 align:middle line:84% had these examples of these specific dimensions of variation 00:36:36.860 --> 00:36:41.420 align:middle line:84% that corresponded really well to these real diseases, 00:36:41.420 --> 00:36:45.900 align:middle line:84% I'm sure it took a lot of handcrafting and a lot of time. 00:36:45.900 --> 00:36:49.540 align:middle line:84% And so recent work tries to do similar types 00:36:49.540 --> 00:36:52.780 align:middle line:84% of counterfactual reasoning or understanding. 00:36:52.780 --> 00:36:55.360 align:middle line:84% But here, I'm actually just using 00:36:55.360 --> 00:36:58.120 align:middle line:84% text conditioning for modern diffusion models, 00:36:58.120 --> 00:37:00.880 align:middle line:84% or other types of conditioning using diffusion models, that 00:37:00.880 --> 00:37:03.360 align:middle line:84% give us really simple and interpretable mechanisms 00:37:03.360 --> 00:37:06.600 align:middle line:84% for control, though, of course, there's not always perfect 00:37:06.600 --> 00:37:08.840 align:middle line:84% alignment between maybe your text control 00:37:08.840 --> 00:37:12.400 align:middle line:84% and what actually gets generated by the generator. 00:37:12.400 --> 00:37:14.320 align:middle line:84% But it does make this ability to test 00:37:14.320 --> 00:37:18.900 align:middle line:84% these counterfactual hypotheses, in many ways, pretty efficient. 00:37:18.900 --> 00:37:22.000 align:middle line:84% So I showed this work actually before, 00:37:22.000 --> 00:37:24.680 align:middle line:84% but this is work where they very specifically looked 00:37:24.680 --> 00:37:28.200 align:middle line:84% at using almost like human intelligence 00:37:28.200 --> 00:37:30.140 align:middle line:84% to generate possible counterfactuals, 00:37:30.140 --> 00:37:32.480 align:middle line:84% using text conditioning on generative images, 00:37:32.480 --> 00:37:36.640 align:middle line:84% and then explicitly testing, quantitatively, the performance 00:37:36.640 --> 00:37:41.480 align:middle line:84% of models given these different human-derived counterfactuals. 00:37:41.480 --> 00:37:45.480 align:middle line:84% And they showed that it's just a really nice way to do, 00:37:45.480 --> 00:37:48.400 align:middle line:84% essentially, interfaces with your data sets, 00:37:48.400 --> 00:37:52.210 align:middle line:84% basically using the generator trained on top of data 00:37:52.210 --> 00:37:57.810 align:middle line:84% as a direct investigation mechanism or interaction 00:37:57.810 --> 00:38:00.170 align:middle line:84% mechanism with the underlying real data 00:38:00.170 --> 00:38:02.950 align:middle line:84% set itself understanding the dimensions of bias. 00:38:02.950 --> 00:38:05.530 align:middle line:90% 00:38:05.530 --> 00:38:07.490 align:middle line:90% Cool. 00:38:07.490 --> 00:38:09.810 align:middle line:84% And then I think there's also this question of-- 00:38:09.810 --> 00:38:11.430 align:middle line:90% OK. 00:38:11.430 --> 00:38:14.210 align:middle line:84% The data the generated data is potentially 00:38:14.210 --> 00:38:18.430 align:middle line:84% useful for explainability or for our data set exploration. 00:38:18.430 --> 00:38:22.770 align:middle line:84% But it is also useful for representation learning? 00:38:22.770 --> 00:38:26.547 align:middle line:84% Can you learn from generated data in the same way 00:38:26.547 --> 00:38:29.130 align:middle line:84% that you might be able to learn, or maybe even in a better way 00:38:29.130 --> 00:38:31.490 align:middle line:84% than you might be able to learn, from real data? 00:38:31.490 --> 00:38:34.410 align:middle line:84% And this is where it gets kind of complicated. 00:38:34.410 --> 00:38:36.930 align:middle line:84% So here's a work where they basically said, OK, you 00:38:36.930 --> 00:38:40.810 align:middle line:84% have some data set x, and then you're 00:38:40.810 --> 00:38:44.730 align:middle line:84% going to train a generative model on top of that data set. 00:38:44.730 --> 00:38:47.675 align:middle line:84% And the idea is somehow the generative model 00:38:47.675 --> 00:38:49.050 align:middle line:84% trained on top of the data set is 00:38:49.050 --> 00:38:53.010 align:middle line:84% going to capture either the same information as the data set, 00:38:53.010 --> 00:38:55.030 align:middle line:84% or maybe arguably slightly more information 00:38:55.030 --> 00:38:57.430 align:middle line:90% than the underlying data set. 00:38:57.430 --> 00:38:59.230 align:middle line:84% And then that might actually be able to be 00:38:59.230 --> 00:39:02.050 align:middle line:84% useful for representation learning. 00:39:02.050 --> 00:39:04.230 align:middle line:84% So here, kind of similar to that other example 00:39:04.230 --> 00:39:06.570 align:middle line:84% with the multi-view ensembling from GANs, 00:39:06.570 --> 00:39:08.510 align:middle line:84% they're explicitly looking at your ability 00:39:08.510 --> 00:39:10.670 align:middle line:84% to generate maybe the same real image 00:39:10.670 --> 00:39:14.590 align:middle line:84% from different perspectives or from different views, 00:39:14.590 --> 00:39:20.030 align:middle line:84% and then use that as the input to contrastive learning models. 00:39:20.030 --> 00:39:22.590 align:middle line:84% So here, trying to learn a representation space 00:39:22.590 --> 00:39:24.750 align:middle line:90% where you're saying, all right. 00:39:24.750 --> 00:39:27.628 align:middle line:84% Here's two images from the same category. 00:39:27.628 --> 00:39:29.170 align:middle line:84% Here's one from a different category. 00:39:29.170 --> 00:39:32.910 align:middle line:84% And then you're training that contrastive style loss. 00:39:32.910 --> 00:39:33.590 align:middle line:90% Yeah. 00:39:33.590 --> 00:39:35.710 align:middle line:84% AUDIENCE: So is this trained from scratch, 00:39:35.710 --> 00:39:37.622 align:middle line:84% or is this transfer learning to some degree? 00:39:37.622 --> 00:39:38.330 align:middle line:90% SARA BEERY: Yeah. 00:39:38.330 --> 00:39:41.070 align:middle line:84% That's an interesting way-- it's an interesting question. 00:39:41.070 --> 00:39:45.230 align:middle line:84% I mean, arguably, this is a mechanism of transfer learning, 00:39:45.230 --> 00:39:46.262 align:middle line:90% assuming that-- 00:39:46.262 --> 00:39:48.470 align:middle line:84% AUDIENCE: Well, I guess to get to the original model, 00:39:48.470 --> 00:39:51.130 align:middle line:84% was that built on some pretrained model 00:39:51.130 --> 00:39:52.795 align:middle line:90% for image analysis? 00:39:52.795 --> 00:39:54.170 align:middle line:84% SARA BEERY: Which original model? 00:39:54.170 --> 00:39:57.030 align:middle line:84% We're talking about kind of a system here. 00:39:57.030 --> 00:39:58.370 align:middle line:90% AUDIENCE: [INAUDIBLE] 00:39:58.370 --> 00:40:00.370 align:middle line:84% SARA BEERY: This generative model? 00:40:00.370 --> 00:40:04.650 align:middle line:84% So let's assume for the sake of simplicity 00:40:04.650 --> 00:40:06.690 align:middle line:84% that this is trained from scratch on that, 00:40:06.690 --> 00:40:09.710 align:middle line:84% because otherwise, it gets even more complicated to understand. 00:40:09.710 --> 00:40:13.170 align:middle line:84% And this actually, I think, is one of the difficulties 00:40:13.170 --> 00:40:17.090 align:middle line:84% currently when people are trying to understand whether we can-- 00:40:17.090 --> 00:40:19.970 align:middle line:84% there's this kind of broader meta argument going 00:40:19.970 --> 00:40:21.930 align:middle line:84% on in the community of whether it 00:40:21.930 --> 00:40:25.450 align:middle line:84% is possible to get more information out 00:40:25.450 --> 00:40:30.250 align:middle line:84% of a generator trained on data than the underlying data itself. 00:40:30.250 --> 00:40:31.910 align:middle line:84% From an information theory perspective, 00:40:31.910 --> 00:40:34.010 align:middle line:84% it's like, how could you possibly create something 00:40:34.010 --> 00:40:34.770 align:middle line:90% from nothing? 00:40:34.770 --> 00:40:36.890 align:middle line:84% If you have the data, and that's all the information you have, 00:40:36.890 --> 00:40:39.030 align:middle line:84% you couldn't maybe get more information out of it. 00:40:39.030 --> 00:40:41.210 align:middle line:84% But then there's kind of the counterargument, which 00:40:41.210 --> 00:40:43.730 align:middle line:84% is actually like our architecture design, 00:40:43.730 --> 00:40:47.230 align:middle line:84% and the design of the training systems 00:40:47.230 --> 00:40:49.750 align:middle line:84% builds in knowledge that would not necessarily 00:40:49.750 --> 00:40:51.570 align:middle line:90% have already been built in. 00:40:51.570 --> 00:40:54.350 align:middle line:84% And actually, you could argue that the construction of things 00:40:54.350 --> 00:40:56.390 align:middle line:84% like convolutional neural networks 00:40:56.390 --> 00:40:59.630 align:middle line:84% fundamentally are an injection of additional knowledge 00:40:59.630 --> 00:41:03.470 align:middle line:84% beyond just the underlying intermediate data. 00:41:03.470 --> 00:41:06.950 align:middle line:84% We're building knowledge into the design of our architectures 00:41:06.950 --> 00:41:10.030 align:middle line:84% and our loss functions based on our understanding of what, 00:41:10.030 --> 00:41:12.790 align:middle line:84% for example, structure we expect to see. 00:41:12.790 --> 00:41:15.270 align:middle line:84% So convolutional neural networks basically say, 00:41:15.270 --> 00:41:18.910 align:middle line:84% we expect there to be local structure in images. 00:41:18.910 --> 00:41:23.870 align:middle line:84% And so then the argument is, isn't data augmentation 00:41:23.870 --> 00:41:28.350 align:middle line:84% a mechanism-- you could argue that random flipping is an image 00:41:28.350 --> 00:41:31.310 align:middle line:90% generation algorithm, right? 00:41:31.310 --> 00:41:33.510 align:middle line:84% I mean, depending on how you want 00:41:33.510 --> 00:41:36.250 align:middle line:84% to define that, semantically, it is, right? 00:41:36.250 --> 00:41:37.590 align:middle line:90% You've taken an image. 00:41:37.590 --> 00:41:42.190 align:middle line:84% You've used some algorithm to generate a new image that you 00:41:42.190 --> 00:41:45.870 align:middle line:84% guarantee still has fidelity to the class that you care about. 00:41:45.870 --> 00:41:49.810 align:middle line:90% 00:41:49.810 --> 00:41:55.930 align:middle line:84% And then things like this copy paste, cut and paste 00:41:55.930 --> 00:41:57.890 align:middle line:84% style data augmentation, that's even 00:41:57.890 --> 00:41:59.370 align:middle line:84% getting maybe closer to something 00:41:59.370 --> 00:42:01.537 align:middle line:84% like a learned generative model, because now, you're 00:42:01.537 --> 00:42:04.650 align:middle line:84% building explicit assumptions about taking foreground objects 00:42:04.650 --> 00:42:06.630 align:middle line:84% and putting them on different backgrounds. 00:42:06.630 --> 00:42:08.570 align:middle line:84% This is like a really useful mechanism 00:42:08.570 --> 00:42:11.290 align:middle line:84% for increasing the robustness of machine learning models, just 00:42:11.290 --> 00:42:14.290 align:middle line:84% these engineering hacks about how we manipulate 00:42:14.290 --> 00:42:15.663 align:middle line:90% the data during training. 00:42:15.663 --> 00:42:17.330 align:middle line:84% And then maybe the argument is like, how 00:42:17.330 --> 00:42:18.970 align:middle line:84% is that any different from something 00:42:18.970 --> 00:42:21.930 align:middle line:84% like learning a generator that can interpolate maybe 00:42:21.930 --> 00:42:25.090 align:middle line:84% in something closer to the image manifold? 00:42:25.090 --> 00:42:27.830 align:middle line:90% So it's a bit semantic. 00:42:27.830 --> 00:42:31.250 align:middle line:90% 00:42:31.250 --> 00:42:33.270 align:middle line:84% Anyway, I have been thinking about this a lot. 00:42:33.270 --> 00:42:35.770 align:middle line:84% I have a grad student who's been looking at whether we can-- 00:42:35.770 --> 00:42:37.210 align:middle line:84% I really, really specifically want 00:42:37.210 --> 00:42:40.850 align:middle line:84% to be able to learn, maybe using generative models, 00:42:40.850 --> 00:42:43.710 align:middle line:84% to do a better job, a more robust job, of recognizing 00:42:43.710 --> 00:42:45.430 align:middle line:90% rare things. 00:42:45.430 --> 00:42:47.350 align:middle line:84% But now, you've really gotten yourself 00:42:47.350 --> 00:42:50.090 align:middle line:90% in a chicken and egg problem. 00:42:50.090 --> 00:42:53.590 align:middle line:84% Because the thing is rare, it's hard to train a generator that 00:42:53.590 --> 00:42:55.410 align:middle line:84% does a good job of generating it, 00:42:55.410 --> 00:42:57.325 align:middle line:84% and then the reverse is also true. 00:42:57.325 --> 00:42:59.950 align:middle line:84% And there have been people that have shown that you can improve 00:42:59.950 --> 00:43:02.463 align:middle line:84% rare categorization using generative models, 00:43:02.463 --> 00:43:04.630 align:middle line:84% but they cheat, and they train the generative models 00:43:04.630 --> 00:43:07.220 align:middle line:90% on a lot of data. 00:43:07.220 --> 00:43:10.070 align:middle line:84% So far, we have been able to find ways 00:43:10.070 --> 00:43:12.750 align:middle line:84% where you can get some gains in very specific constrained 00:43:12.750 --> 00:43:15.990 align:middle line:84% scenarios, but it's not just naive and simple. 00:43:15.990 --> 00:43:18.750 align:middle line:84% Anyway, the point here is that you 00:43:18.750 --> 00:43:22.110 align:middle line:84% can take the classical contrastive learning approach 00:43:22.110 --> 00:43:24.750 align:middle line:84% where we're using our knowledge-based image 00:43:24.750 --> 00:43:28.030 align:middle line:84% generation, where we've done these specific types of warping 00:43:28.030 --> 00:43:32.670 align:middle line:84% or whatever, random cropping, flipping, to the initial image 00:43:32.670 --> 00:43:35.510 align:middle line:84% color jitter, in a way that we know doesn't necessarily 00:43:35.510 --> 00:43:36.810 align:middle line:90% break the category boundary. 00:43:36.810 --> 00:43:39.430 align:middle line:84% We still want it to be the same thing. 00:43:39.430 --> 00:43:42.530 align:middle line:84% And here, we're instead going to move just a little bit 00:43:42.530 --> 00:43:48.130 align:middle line:84% in some latent space to create now this mapping in data space 00:43:48.130 --> 00:43:50.850 align:middle line:84% through what should be the manifold of real images, 00:43:50.850 --> 00:43:53.210 align:middle line:84% and then use that as your transformation 00:43:53.210 --> 00:43:55.330 align:middle line:84% for your contrastive learning model. 00:43:55.330 --> 00:43:55.910 align:middle line:90% Yeah. 00:43:55.910 --> 00:43:58.618 align:middle line:84% AUDIENCE: I had a question about this, which kind of relates back 00:43:58.618 --> 00:44:03.570 align:middle line:84% to earlier when we talked about the idea of small local movement 00:44:03.570 --> 00:44:06.530 align:middle line:84% in the latent space not crossing a boundary. 00:44:06.530 --> 00:44:12.050 align:middle line:84% I mean, intuitively, if you have just a 2D latent space and a 10 00:44:12.050 --> 00:44:16.330 align:middle line:84% class classification problem, you could imagine some very 00:44:16.330 --> 00:44:18.690 align:middle line:84% naive pie chart that has 10 slices. 00:44:18.690 --> 00:44:21.170 align:middle line:84% And there has to exist a boundary 00:44:21.170 --> 00:44:23.583 align:middle line:84% somewhere between category 1 and category 2. 00:44:23.583 --> 00:44:24.250 align:middle line:90% SARA BEERY: Yes. 00:44:24.250 --> 00:44:26.333 align:middle line:84% AUDIENCE: So is this just a probabilistic argument 00:44:26.333 --> 00:44:29.090 align:middle line:84% that in a high enough dimensional latent space, 00:44:29.090 --> 00:44:33.130 align:middle line:84% the odds that you would sample a location where the boundary 00:44:33.130 --> 00:44:37.010 align:middle line:84% shifts is so unlikely that you can get away 00:44:37.010 --> 00:44:40.390 align:middle line:84% with this kind of approach to training for contrastive 00:44:40.390 --> 00:44:41.430 align:middle line:90% learning? 00:44:41.430 --> 00:44:43.690 align:middle line:84% SARA BEERY: Or I would even go beyond that, 00:44:43.690 --> 00:44:47.630 align:middle line:84% and I would even say as long as you define 00:44:47.630 --> 00:44:50.510 align:middle line:90% some dimension of movement-- 00:44:50.510 --> 00:44:52.550 align:middle line:84% as long as, most of the time, it's 00:44:52.550 --> 00:44:55.270 align:middle line:84% not crossing a class boundary-- if it does cross a class 00:44:55.270 --> 00:44:57.050 align:middle line:84% boundary a very small amount of the time 00:44:57.050 --> 00:44:59.150 align:middle line:84% still, by the way we train these models, 00:44:59.150 --> 00:45:01.210 align:middle line:84% it's almost like the noise will come out. 00:45:01.210 --> 00:45:05.070 align:middle line:84% It's almost a mechanism for regularization. 00:45:05.070 --> 00:45:08.510 align:middle line:84% Of course, there's counterarguments to that. 00:45:08.510 --> 00:45:10.870 align:middle line:84% It depends on how egregious those boundaries are 00:45:10.870 --> 00:45:13.870 align:middle line:84% and how frequently you're crossing them. 00:45:13.870 --> 00:45:14.370 align:middle line:90% Yeah. 00:45:14.370 --> 00:45:17.230 align:middle line:90% I would say as long as-- 00:45:17.230 --> 00:45:19.910 align:middle line:84% if you want to make a probabilistic argument, 00:45:19.910 --> 00:45:21.790 align:middle line:84% and you want to map that to something that's 00:45:21.790 --> 00:45:25.950 align:middle line:84% like a physical area argument, like the area of the model where 00:45:25.950 --> 00:45:29.470 align:middle line:84% moving in a small amount corresponds to a reasonably 00:45:29.470 --> 00:45:31.750 align:middle line:84% maintaining the category boundary 00:45:31.750 --> 00:45:34.190 align:middle line:84% relative to the areas where you are crossing category 00:45:34.190 --> 00:45:38.030 align:middle line:84% boundaries-- as long as that ratio is quite large, 00:45:38.030 --> 00:45:40.710 align:middle line:84% then I imagine you would still learned something useful, 00:45:40.710 --> 00:45:41.950 align:middle line:90% even if sometimes it's wrong. 00:45:41.950 --> 00:45:45.370 align:middle line:84% It's just like how in the standard contrastive learning 00:45:45.370 --> 00:45:47.930 align:middle line:84% objectives, we did see some improvement when 00:45:47.930 --> 00:45:51.550 align:middle line:84% you could explicitly remove these false positive, 00:45:51.550 --> 00:45:54.090 align:middle line:84% false negative type things where, within your batch, 00:45:54.090 --> 00:45:56.490 align:middle line:84% you would have maybe two images of dogs. 00:45:56.490 --> 00:45:58.870 align:middle line:84% And now, they're being treated as a different category 00:45:58.870 --> 00:46:00.370 align:middle line:84% during your representation learning. 00:46:00.370 --> 00:46:02.203 align:middle line:84% You're explicitly saying those things should 00:46:02.203 --> 00:46:04.610 align:middle line:84% be far apart, though they actually 00:46:04.610 --> 00:46:08.770 align:middle line:90% represent the same category. 00:46:08.770 --> 00:46:10.610 align:middle line:84% If you could remove all of those, 00:46:10.610 --> 00:46:12.810 align:middle line:90% the models did train better. 00:46:12.810 --> 00:46:14.530 align:middle line:84% There were these nice examples where 00:46:14.530 --> 00:46:16.238 align:middle line:84% you could show, if you removed these kind 00:46:16.238 --> 00:46:19.910 align:middle line:84% of false characterizations by using supervision, for example, 00:46:19.910 --> 00:46:21.630 align:middle line:84% you got better representation learning. 00:46:21.630 --> 00:46:24.130 align:middle line:84% But the model's still able to learn useful representations 00:46:24.130 --> 00:46:26.970 align:middle line:84% even when those do exist in there. 00:46:26.970 --> 00:46:29.610 align:middle line:90% Yeah. 00:46:29.610 --> 00:46:30.110 align:middle line:90% Cool. 00:46:30.110 --> 00:46:30.670 align:middle line:90% So yeah. 00:46:30.670 --> 00:46:32.420 align:middle line:84% Now, here, it's just a different mechanism 00:46:32.420 --> 00:46:34.050 align:middle line:84% for constructing these positive pairs 00:46:34.050 --> 00:46:36.590 align:middle line:84% to train your contrastive learning algorithm. 00:46:36.590 --> 00:46:40.950 align:middle line:84% And so here, as opposed to your standard SimCLR type views 00:46:40.950 --> 00:46:44.310 align:middle line:84% where you have two views of the same thing where it's maybe 00:46:44.310 --> 00:46:48.150 align:middle line:84% cropped from different areas or warped in terms of the color, 00:46:48.150 --> 00:46:50.150 align:middle line:90% now, you're taking latent views. 00:46:50.150 --> 00:46:53.930 align:middle line:84% You're taking some ball around the real image. 00:46:53.930 --> 00:46:55.890 align:middle line:84% And any dimension, you would move in that ball, 00:46:55.890 --> 00:46:58.190 align:middle line:84% you could argue, would be a different view 00:46:58.190 --> 00:47:00.950 align:middle line:90% of the same thing. 00:47:00.950 --> 00:47:05.375 align:middle line:84% And then now, just how we might say, 00:47:05.375 --> 00:47:06.750 align:middle line:84% OK, these are all different views 00:47:06.750 --> 00:47:08.167 align:middle line:84% of the same thing for the purposes 00:47:08.167 --> 00:47:09.510 align:middle line:90% of contrastive learning. 00:47:09.510 --> 00:47:11.510 align:middle line:84% Here, you can say, these are all different views 00:47:11.510 --> 00:47:12.710 align:middle line:90% of the same thing. 00:47:12.710 --> 00:47:16.590 align:middle line:84% Now, one interesting dimension here, just generally, 00:47:16.590 --> 00:47:20.710 align:middle line:84% is this question of diversity versus fidelity 00:47:20.710 --> 00:47:23.600 align:middle line:84% when it comes to the usefulness of training signal. 00:47:23.600 --> 00:47:25.850 align:middle line:84% This is also something I've been thinking about a lot. 00:47:25.850 --> 00:47:29.870 align:middle line:84% So it's probably more useful to know 00:47:29.870 --> 00:47:32.870 align:middle line:84% that two things that are more different from each other 00:47:32.870 --> 00:47:36.050 align:middle line:84% are, in fact, the same than it is to know that two things that 00:47:36.050 --> 00:47:38.970 align:middle line:84% are almost identical to each other are, in fact, the same. 00:47:38.970 --> 00:47:42.450 align:middle line:84% So for example, if you look at this middle set 00:47:42.450 --> 00:47:45.570 align:middle line:84% here, where these are all American robins, 00:47:45.570 --> 00:47:48.410 align:middle line:84% the pose of this American robin in basically all of these 00:47:48.410 --> 00:47:49.350 align:middle line:90% is nearly identical. 00:47:49.350 --> 00:47:52.270 align:middle line:84% The construction of the image, the orientation, 00:47:52.270 --> 00:47:54.410 align:middle line:84% et cetera, of the image is almost identical. 00:47:54.410 --> 00:47:56.690 align:middle line:84% There's not a lot of diversity here. 00:47:56.690 --> 00:47:59.170 align:middle line:84% And so actually from a learning signal perspective, 00:47:59.170 --> 00:48:02.370 align:middle line:84% this might be less useful, arguably, 00:48:02.370 --> 00:48:06.150 align:middle line:84% than something that also does something like random flipping. 00:48:06.150 --> 00:48:08.730 align:middle line:84% And so I think there is this kind 00:48:08.730 --> 00:48:11.710 align:middle line:84% of interesting challenge with these generative models, 00:48:11.710 --> 00:48:15.330 align:middle line:84% broadly, with GANs, with diffusion models, et cetera. 00:48:15.330 --> 00:48:17.930 align:middle line:84% Even with some of these personalization style 00:48:17.930 --> 00:48:21.290 align:middle line:84% models that are trying to do a really nice job of capturing 00:48:21.290 --> 00:48:24.770 align:middle line:84% something specific, often, there's 00:48:24.770 --> 00:48:27.130 align:middle line:84% almost like a Pareto frontier with these models 00:48:27.130 --> 00:48:30.010 align:middle line:84% when it comes to this trade off between diversity and fidelity. 00:48:30.010 --> 00:48:32.910 align:middle line:84% And so you can't to get the model 00:48:32.910 --> 00:48:35.350 align:middle line:84% to generate things that are really diverse, maybe 00:48:35.350 --> 00:48:39.830 align:middle line:84% in terms of their context, or their scene structure, 00:48:39.830 --> 00:48:41.810 align:middle line:90% without giving up fidelity. 00:48:41.810 --> 00:48:46.990 align:middle line:84% You end up in this space where you can generate a robin that's 00:48:46.990 --> 00:48:48.650 align:middle line:84% positioned differently in the scene, 00:48:48.650 --> 00:48:52.030 align:middle line:84% maybe flying, different pose, and you can try 00:48:52.030 --> 00:48:53.370 align:middle line:90% to force the model to do that. 00:48:53.370 --> 00:48:55.350 align:middle line:84% And it will, but it probably won't actually 00:48:55.350 --> 00:48:58.830 align:middle line:84% be very robin-like anymore if this is the input 00:48:58.830 --> 00:49:01.990 align:middle line:90% image that you're working from. 00:49:01.990 --> 00:49:03.150 align:middle line:90% Yeah. 00:49:03.150 --> 00:49:05.270 align:middle line:84% AUDIENCE: In the latent space, when 00:49:05.270 --> 00:49:08.970 align:middle line:84% you're moving around the image feature vector, 00:49:08.970 --> 00:49:11.870 align:middle line:84% would you say that the images around it 00:49:11.870 --> 00:49:14.430 align:middle line:90% belong to the same class? 00:49:14.430 --> 00:49:16.230 align:middle line:84% How do you know how far you can go? 00:49:16.230 --> 00:49:20.110 align:middle line:84% And that distance, is it the same for all classes? 00:49:20.110 --> 00:49:22.430 align:middle line:84% You can imagine for classes of one, 00:49:22.430 --> 00:49:25.942 align:middle line:84% one class is overrepresented compared to another class. 00:49:25.942 --> 00:49:26.650 align:middle line:90% SARA BEERY: Yeah. 00:49:26.650 --> 00:49:28.030 align:middle line:84% So the question is basically, how 00:49:28.030 --> 00:49:31.290 align:middle line:84% do you know how far you can go in the latent space 00:49:31.290 --> 00:49:33.850 align:middle line:84% and have it still map to the same category? 00:49:33.850 --> 00:49:36.810 align:middle line:84% And does that distance correspond-- 00:49:36.810 --> 00:49:38.890 align:middle line:84% is the optimal amount of distance 00:49:38.890 --> 00:49:40.570 align:middle line:84% you could go for any given category 00:49:40.570 --> 00:49:42.450 align:middle line:90% the same across all categories? 00:49:42.450 --> 00:49:46.010 align:middle line:84% So I would say experimentally, in a lot of these papers, 00:49:46.010 --> 00:49:48.310 align:middle line:84% they're using this as a hyperparameter, right? 00:49:48.310 --> 00:49:49.810 align:middle line:84% They're basically saying, all right. 00:49:49.810 --> 00:49:53.410 align:middle line:84% We're going to define some radius in something 00:49:53.410 --> 00:49:57.890 align:middle line:84% like cosine distance, and we're going to sample 00:49:57.890 --> 00:49:59.070 align:middle line:90% within that radius. 00:49:59.070 --> 00:50:02.382 align:middle line:84% And we're going to test for different radiuses or radii. 00:50:02.382 --> 00:50:04.090 align:middle line:84% And then we're going to pick the one that 00:50:04.090 --> 00:50:07.570 align:middle line:90% does the best on our test data. 00:50:07.570 --> 00:50:09.718 align:middle line:84% That's a bit disappointing, because I actually 00:50:09.718 --> 00:50:11.510 align:middle line:84% think that's a really interesting question. 00:50:11.510 --> 00:50:16.410 align:middle line:84% How do you determine how far you can go before it crosses a class 00:50:16.410 --> 00:50:17.150 align:middle line:90% boundary? 00:50:17.150 --> 00:50:21.050 align:middle line:84% And almost by definition, you start 00:50:21.050 --> 00:50:22.990 align:middle line:84% getting into these really messy questions, 00:50:22.990 --> 00:50:25.410 align:middle line:84% particularly as these models are able to generate 00:50:25.410 --> 00:50:29.610 align:middle line:84% things that are realistic looking but impossible. 00:50:29.610 --> 00:50:32.510 align:middle line:90% 00:50:32.510 --> 00:50:34.710 align:middle line:84% Me and Phil actually argue about this quite a lot. 00:50:34.710 --> 00:50:37.470 align:middle line:84% If you had a picture of your mom, 00:50:37.470 --> 00:50:40.050 align:middle line:84% but you had another eye in the middle of her forehead, 00:50:40.050 --> 00:50:42.670 align:middle line:90% is it still your mom? 00:50:42.670 --> 00:50:44.890 align:middle line:90% What do you guys think? 00:50:44.890 --> 00:50:46.770 align:middle line:84% Is it still your mom if she has three eyes? 00:50:46.770 --> 00:50:50.550 align:middle line:90% 00:50:50.550 --> 00:50:51.970 align:middle line:90% There's not a right answer here. 00:50:51.970 --> 00:50:55.070 align:middle line:84% And so if some of the dimensions of interpolation 00:50:55.070 --> 00:50:57.693 align:middle line:84% you can move into are still kind of realistic looking, 00:50:57.693 --> 00:51:00.110 align:middle line:84% but actually impossible-- so maybe if you think about this 00:51:00.110 --> 00:51:02.130 align:middle line:84% from a species categorization perspective, 00:51:02.130 --> 00:51:04.010 align:middle line:84% you take like this American robin, 00:51:04.010 --> 00:51:09.190 align:middle line:84% and then you make the chest gray instead of red. 00:51:09.190 --> 00:51:11.150 align:middle line:90% That's not a real bird. 00:51:11.150 --> 00:51:14.893 align:middle line:84% So now, you have things that look like birds, 00:51:14.893 --> 00:51:16.310 align:middle line:84% but they don't actually correspond 00:51:16.310 --> 00:51:19.030 align:middle line:90% to any real category of bird. 00:51:19.030 --> 00:51:21.210 align:middle line:84% And so then what should that be categorized as? 00:51:21.210 --> 00:51:24.190 align:middle line:84% And so somehow, understanding the boundary 00:51:24.190 --> 00:51:27.850 align:middle line:84% of a class, a really important component to being able to do 00:51:27.850 --> 00:51:31.850 align:middle line:84% that is having representative data from all the dimensions 00:51:31.850 --> 00:51:34.330 align:middle line:84% of real variation for that class so you 00:51:34.330 --> 00:51:36.590 align:middle line:84% can define what the reasonable class boundary is. 00:51:36.590 --> 00:51:39.250 align:middle line:84% That's one of the reasons that few shot learning is so hard, 00:51:39.250 --> 00:51:41.930 align:middle line:84% because it often means we don't have 00:51:41.930 --> 00:51:43.770 align:middle line:84% a good, well understood representation 00:51:43.770 --> 00:51:46.150 align:middle line:84% space of the boundaries of the real class. 00:51:46.150 --> 00:51:47.450 align:middle line:90% And so we don't know. 00:51:47.450 --> 00:51:50.530 align:middle line:84% If you've never seen a picture of a bird as a juvenile, 00:51:50.530 --> 00:51:54.790 align:middle line:84% we don't know what the real class boundary should be. 00:51:54.790 --> 00:51:55.290 align:middle line:90% Yeah. 00:51:55.290 --> 00:51:57.623 align:middle line:84% AUDIENCE: Isn't that just like a fundamental shortcoming 00:51:57.623 --> 00:52:00.650 align:middle line:84% of classification as a learning objective? 00:52:00.650 --> 00:52:04.810 align:middle line:84% CLIP was almost a better idea where they're just 00:52:04.810 --> 00:52:08.050 align:middle line:84% saying that you should learn to embed text 00:52:08.050 --> 00:52:11.970 align:middle line:84% descriptions of images close to that embedding of the image, 00:52:11.970 --> 00:52:15.490 align:middle line:84% and you can capture more truth about the world. 00:52:15.490 --> 00:52:17.068 align:middle line:84% To your point about the robin chest-- 00:52:17.068 --> 00:52:18.110 align:middle line:90% I mean, this isn't exact. 00:52:18.110 --> 00:52:21.810 align:middle line:84% But imagine you have an albino robin. 00:52:21.810 --> 00:52:24.330 align:middle line:84% That's still a robin, and a human can look at it 00:52:24.330 --> 00:52:25.850 align:middle line:84% and be like, yeah, that's a robin. 00:52:25.850 --> 00:52:29.670 align:middle line:84% But from your definition of the classification perspective, 00:52:29.670 --> 00:52:30.190 align:middle line:90% I doubt-- 00:52:30.190 --> 00:52:34.050 align:middle line:84% I think you're right that for a specific type of classification, 00:52:34.050 --> 00:52:35.550 align:middle line:84% whatever you define that class to be 00:52:35.550 --> 00:52:39.990 align:middle line:84% is the extent to which the model can learn [INAUDIBLE]. 00:52:39.990 --> 00:52:42.842 align:middle line:84% So isn't this just a shortcoming of classification? 00:52:42.842 --> 00:52:43.550 align:middle line:90% SARA BEERY: Yeah. 00:52:43.550 --> 00:52:46.710 align:middle line:84% So the argument is, basically, this challenge 00:52:46.710 --> 00:52:49.190 align:middle line:84% of being able to understand category boundaries 00:52:49.190 --> 00:52:53.430 align:middle line:84% is maybe a limitation of the underlying philosophical concept 00:52:53.430 --> 00:52:56.590 align:middle line:84% of trying to categorize things with some sort of list 00:52:56.590 --> 00:52:58.350 align:middle line:90% of specific categories. 00:52:58.350 --> 00:53:03.070 align:middle line:84% And actually, if you really want to get deep into it, 00:53:03.070 --> 00:53:07.590 align:middle line:84% there's a lot of literature from philosophy about the concept 00:53:07.590 --> 00:53:08.870 align:middle line:90% of categorization. 00:53:08.870 --> 00:53:15.050 align:middle line:84% And there's kind of these explicit differences of opinion 00:53:15.050 --> 00:53:17.510 align:middle line:84% when you look at historical philosophers around the idea 00:53:17.510 --> 00:53:21.830 align:middle line:84% that it's possible to categorize something as a list of all 00:53:21.830 --> 00:53:25.075 align:middle line:90% its attributes. 00:53:25.075 --> 00:53:25.950 align:middle line:90% So you could say, OK. 00:53:25.950 --> 00:53:28.850 align:middle line:84% A robin is a bird with a black face and a red chest 00:53:28.850 --> 00:53:33.530 align:middle line:84% and yellow feet and a round structure 00:53:33.530 --> 00:53:35.550 align:middle line:84% and a yellow eye with a black pupil. 00:53:35.550 --> 00:53:39.690 align:middle line:84% And so there's this idea that any object can 00:53:39.690 --> 00:53:43.170 align:middle line:84% be defined as a list of characteristics 00:53:43.170 --> 00:53:45.370 align:middle line:90% you're looking for. 00:53:45.370 --> 00:53:47.370 align:middle line:84% And then if you have the kind of mapping 00:53:47.370 --> 00:53:50.190 align:middle line:84% of the set of characteristics to the things you care about, 00:53:50.190 --> 00:53:52.410 align:middle line:90% then that's all you need. 00:53:52.410 --> 00:53:53.970 align:middle line:84% My argument is kind of like, that's 00:53:53.970 --> 00:53:56.090 align:middle line:84% the CLIP approach, in a way, because you're 00:53:56.090 --> 00:53:58.090 align:middle line:84% trying to learn something about the relationship 00:53:58.090 --> 00:54:01.742 align:middle line:84% between textual semantic concepts and things in images. 00:54:01.742 --> 00:54:03.450 align:middle line:84% And there's been a lot of work since CLIP 00:54:03.450 --> 00:54:07.170 align:middle line:84% that actually tries to really explicitly ground 00:54:07.170 --> 00:54:10.970 align:middle line:84% semantic concepts from text into the visual components that 00:54:10.970 --> 00:54:12.170 align:middle line:90% corresponds to. 00:54:12.170 --> 00:54:17.130 align:middle line:84% But I think the argument is that, no matter what, if you 00:54:17.130 --> 00:54:22.430 align:middle line:84% give me a list of a fixed list of attributes, 00:54:22.430 --> 00:54:24.305 align:middle line:84% and you try to map that to a category, 00:54:24.305 --> 00:54:26.930 align:middle line:84% you can basically come up with a counterexample that breaks it. 00:54:26.930 --> 00:54:30.115 align:middle line:90% 00:54:30.115 --> 00:54:31.542 align:middle line:84% And then there's the question of, 00:54:31.542 --> 00:54:33.250 align:middle line:84% where do you want to draw the boundaries? 00:54:33.250 --> 00:54:37.110 align:middle line:84% That somehow has to be related to some goal. 00:54:37.110 --> 00:54:39.450 align:middle line:84% So I'm getting really philosophical here. 00:54:39.450 --> 00:54:46.750 align:middle line:84% So if I showed you like an apple and a bell pepper, 00:54:46.750 --> 00:54:51.350 align:middle line:84% and there was a green apple and bell pepper and a red apple 00:54:51.350 --> 00:54:55.430 align:middle line:84% and bell pepper, there's multiple reasonable ways 00:54:55.430 --> 00:54:59.077 align:middle line:90% to split the set of groups. 00:54:59.077 --> 00:55:00.910 align:middle line:84% So in the end, it always just comes down to, 00:55:00.910 --> 00:55:04.190 align:middle line:84% what do you really need the representation space 00:55:04.190 --> 00:55:05.330 align:middle line:90% to do for you? 00:55:05.330 --> 00:55:07.590 align:middle line:84% And then what types of dimensions of variation 00:55:07.590 --> 00:55:11.550 align:middle line:90% do you want to capture there? 00:55:11.550 --> 00:55:16.670 align:middle line:84% And then the albino robin example is a great example 00:55:16.670 --> 00:55:20.530 align:middle line:84% of ways that we are very good at generalizing, 00:55:20.530 --> 00:55:24.090 align:middle line:84% but I would actually argue that models like CLIP still cannot 00:55:24.090 --> 00:55:25.390 align:middle line:90% capture effectively. 00:55:25.390 --> 00:55:28.130 align:middle line:90% 00:55:28.130 --> 00:55:28.630 align:middle line:90% Yeah. 00:55:28.630 --> 00:55:31.270 align:middle line:84% And then you could tell me, if you were a bird expert, 00:55:31.270 --> 00:55:33.830 align:middle line:84% oh, actually, that looks like it might be an albino robin, 00:55:33.830 --> 00:55:35.455 align:middle line:84% but that's actually this other species. 00:55:35.455 --> 00:55:39.270 align:middle line:84% And I'd believe you, because I don't know. 00:55:39.270 --> 00:55:39.770 align:middle line:90% Yeah. 00:55:39.770 --> 00:55:43.075 align:middle line:84% So I think this is really fascinating. 00:55:43.075 --> 00:55:45.450 align:middle line:84% I think it starts getting at some really interesting kind 00:55:45.450 --> 00:55:47.370 align:middle line:84% of meaty questions about, what do we really 00:55:47.370 --> 00:55:50.130 align:middle line:90% want from representations? 00:55:50.130 --> 00:55:53.210 align:middle line:84% And what does it really mean to move through a representation 00:55:53.210 --> 00:55:54.490 align:middle line:90% space? 00:55:54.490 --> 00:55:59.490 align:middle line:84% But so here, again, kind of going back to this idea of, 00:55:59.490 --> 00:56:02.530 align:middle line:84% do contrastive learning from real data based 00:56:02.530 --> 00:56:04.250 align:middle line:84% on some of heuristic set of rules 00:56:04.250 --> 00:56:07.290 align:middle line:84% that we say don't break the class boundary versus if you do 00:56:07.290 --> 00:56:10.610 align:middle line:84% it from generative data based on some experimental setup 00:56:10.610 --> 00:56:15.250 align:middle line:84% of parameters around how far you can move in any given direction, 00:56:15.250 --> 00:56:18.470 align:middle line:84% that you're able to now, instead of going from some data space 00:56:18.470 --> 00:56:22.430 align:middle line:84% to a representation space, go from some latent space 00:56:22.430 --> 00:56:25.630 align:middle line:84% into a data space into a representation space. 00:56:25.630 --> 00:56:28.950 align:middle line:84% And they were able to show that if you're 00:56:28.950 --> 00:56:31.990 align:middle line:84% doing linear transfer from ImageNet1000, 00:56:31.990 --> 00:56:36.790 align:middle line:84% you can get 43.9% top-1 accuracy here. 00:56:36.790 --> 00:56:43.190 align:middle line:84% But only using real data to train these models now mapping 00:56:43.190 --> 00:56:44.698 align:middle line:84% through these kind of information 00:56:44.698 --> 00:56:46.990 align:middle line:84% bottlenecks in a way of these different types of latent 00:56:46.990 --> 00:56:48.810 align:middle line:90% spaces, you can do pretty well. 00:56:48.810 --> 00:56:52.270 align:middle line:90% You can get 35.7% accuracy. 00:56:52.270 --> 00:56:54.550 align:middle line:84% But then if you add in that ability 00:56:54.550 --> 00:56:57.470 align:middle line:84% to move in that latent space, the accuracy 00:56:57.470 --> 00:56:59.830 align:middle line:90% goes up to about 42.6%. 00:56:59.830 --> 00:57:02.115 align:middle line:84% So here, you'll note neither of these 00:57:02.115 --> 00:57:04.490 align:middle line:84% are as good as just training on the real data themselves. 00:57:04.490 --> 00:57:06.270 align:middle line:84% So clearly, there's some limitations 00:57:06.270 --> 00:57:07.830 align:middle line:90% in terms of the mechanism. 00:57:07.830 --> 00:57:10.590 align:middle line:84% But the argument is that if you were 00:57:10.590 --> 00:57:13.090 align:middle line:84% able to build the system effectively, 00:57:13.090 --> 00:57:17.010 align:middle line:84% the ability to interpolate and have these reasonable dimensions 00:57:17.010 --> 00:57:20.210 align:middle line:84% of variation within that original latent space 00:57:20.210 --> 00:57:22.210 align:middle line:84% could potentially add a lot of value. 00:57:22.210 --> 00:57:27.210 align:middle line:84% This is a pretty significant boost in performance. 00:57:27.210 --> 00:57:29.930 align:middle line:84% So then the argument is that these deep generative views 00:57:29.930 --> 00:57:33.650 align:middle line:84% is one dimension of variation, just different viewpoints, 00:57:33.650 --> 00:57:37.010 align:middle line:84% can improve contrastive learning beyond only using standard data 00:57:37.010 --> 00:57:37.950 align:middle line:90% augmentation. 00:57:37.950 --> 00:57:40.850 align:middle line:84% And if the generative model is high quality, 00:57:40.850 --> 00:57:42.450 align:middle line:84% you could even potentially outperform 00:57:42.450 --> 00:57:44.033 align:middle line:90% learning from real data. 00:57:44.033 --> 00:57:45.450 align:middle line:84% But that, of course, means that we 00:57:45.450 --> 00:57:48.650 align:middle line:84% need the generator to be high quality, which 00:57:48.650 --> 00:57:52.607 align:middle line:84% it's not for many things that are uncommon. 00:57:52.607 --> 00:57:54.190 align:middle line:84% One of the things that is kind of fun, 00:57:54.190 --> 00:57:56.110 align:middle line:84% if you try and play with these models, 00:57:56.110 --> 00:57:58.970 align:middle line:84% is looking at what they understand 00:57:58.970 --> 00:58:02.202 align:middle line:84% and trying to get them to break out of their priors. 00:58:02.202 --> 00:58:03.910 align:middle line:84% So there's some classic examples of this. 00:58:03.910 --> 00:58:07.990 align:middle line:84% If you ask models to generate a bunch of pictures of birds, 00:58:07.990 --> 00:58:09.790 align:middle line:84% if you ask it to generate orioles, 00:58:09.790 --> 00:58:11.790 align:middle line:84% it'll give you baseball players, right? 00:58:11.790 --> 00:58:14.830 align:middle line:84% There's this semantic prior that gets 00:58:14.830 --> 00:58:18.533 align:middle line:84% broken by the training data in really interesting ways. 00:58:18.533 --> 00:58:20.450 align:middle line:84% But then there's also other interesting stuff. 00:58:20.450 --> 00:58:25.030 align:middle line:84% If you ask the model to generate a toucan with a short beak, 00:58:25.030 --> 00:58:27.110 align:middle line:84% it's really hard to get it to do it. 00:58:27.110 --> 00:58:28.013 align:middle line:90% It just won't. 00:58:28.013 --> 00:58:28.930 align:middle line:90% It's like, no, no, no. 00:58:28.930 --> 00:58:33.310 align:middle line:84% Toucan somehow overpowers these mechanisms of morphology. 00:58:33.310 --> 00:58:37.470 align:middle line:84% And I've also found that they're much better at generating 00:58:37.470 --> 00:58:40.790 align:middle line:84% dimensions-- they're much better at changing texture than shape. 00:58:40.790 --> 00:58:42.350 align:middle line:84% So it turns out that the models seem 00:58:42.350 --> 00:58:44.630 align:middle line:84% to have learned really strong shape priors that 00:58:44.630 --> 00:58:48.470 align:middle line:84% are hard to break out of for any given category, 00:58:48.470 --> 00:58:51.630 align:middle line:84% but they've gotten quite good at swapping textures 00:58:51.630 --> 00:58:52.370 align:middle line:90% for a category. 00:58:52.370 --> 00:58:56.820 align:middle line:84% So you can say, oh, make a baseball, but make it furry. 00:58:56.820 --> 00:58:57.570 align:middle line:90% And it's like, OK. 00:58:57.570 --> 00:58:58.862 align:middle line:90% It swaps the texture just fine. 00:58:58.862 --> 00:59:02.330 align:middle line:84% But if you say make a baseball, but make it star shaped, 00:59:02.330 --> 00:59:03.907 align:middle line:90% it gets a little harder. 00:59:03.907 --> 00:59:05.490 align:middle line:84% And I think that's really interesting. 00:59:05.490 --> 00:59:06.865 align:middle line:84% I think it's interesting to think 00:59:06.865 --> 00:59:10.970 align:middle line:84% about why shape might be encoded in a stronger way than texture. 00:59:10.970 --> 00:59:14.650 align:middle line:84% And there's some recent work that shows some really nice-- 00:59:14.650 --> 00:59:17.250 align:middle line:84% in this space of trying to build effective 00:59:17.250 --> 00:59:20.010 align:middle line:84% data augmentation through generation. 00:59:20.010 --> 00:59:22.730 align:middle line:90% It just came out at ECCV. 00:59:22.730 --> 00:59:30.110 align:middle line:84% They basically use the scene structure as a fixed thing. 00:59:30.110 --> 00:59:32.305 align:middle line:84% So they mask out the entire dimension of the scene. 00:59:32.305 --> 00:59:33.930 align:middle line:84% So now, you're guaranteeing that you're 00:59:33.930 --> 00:59:36.263 align:middle line:84% going to have a realistic scene because it's a real one. 00:59:36.263 --> 00:59:39.130 align:middle line:84% So maybe you have two dogs sitting on a bench. 00:59:39.130 --> 00:59:42.130 align:middle line:84% Now, the way that they actually do their generation 00:59:42.130 --> 00:59:47.690 align:middle line:84% to get these new cases is they take each individual object, 00:59:47.690 --> 00:59:50.570 align:middle line:84% and then they change it to a different instance 00:59:50.570 --> 00:59:52.010 align:middle line:90% within the data set. 00:59:52.010 --> 00:59:54.850 align:middle line:84% So you basically say, I'm going to learn some personalization 00:59:54.850 --> 00:59:58.250 align:middle line:84% model for every individual dog in the data set, 00:59:58.250 --> 01:00:00.490 align:middle line:84% and then I'm going to condition on the mask 01:00:00.490 --> 01:00:03.030 align:middle line:84% that I have and generate it with that instance. 01:00:03.030 --> 01:00:06.290 align:middle line:84% So they basically call it instance augmented generation, 01:00:06.290 --> 01:00:08.410 align:middle line:84% and they do get some really nice results. 01:00:08.410 --> 01:00:10.870 align:middle line:84% And what it is, basically, it's breaking 01:00:10.870 --> 01:00:13.910 align:middle line:84% these correlations between scene structure 01:00:13.910 --> 01:00:16.990 align:middle line:84% and individual objects in a really realistic way. 01:00:16.990 --> 01:00:20.990 align:middle line:84% And the images do look reasonably correct. 01:00:20.990 --> 01:00:24.450 align:middle line:84% But the shape thing gets weird again. 01:00:24.450 --> 01:00:29.270 align:middle line:84% So because that data set has bird defined as a category, 01:00:29.270 --> 01:00:31.890 align:middle line:84% you can take any bird of any shape, 01:00:31.890 --> 01:00:34.710 align:middle line:84% and you can replace it with the texture, essentially, 01:00:34.710 --> 01:00:36.170 align:middle line:90% of a different bird. 01:00:36.170 --> 01:00:37.890 align:middle line:84% But birds are very different shapes, 01:00:37.890 --> 01:00:41.250 align:middle line:84% so it'll take something like a little round bird, 01:00:41.250 --> 01:00:44.852 align:middle line:84% and then it'll try to squash an eagle into that shape. 01:00:44.852 --> 01:00:46.310 align:middle line:84% And this is something that wouldn't 01:00:46.310 --> 01:00:48.990 align:middle line:84% happen if the granularity of their categories 01:00:48.990 --> 01:00:52.690 align:middle line:84% in the labeled space was a little bit more fine grained. 01:00:52.690 --> 01:00:54.990 align:middle line:84% And again, it's just built in assumptions 01:00:54.990 --> 01:00:56.790 align:middle line:84% about the amount of shape variation 01:00:56.790 --> 01:00:59.590 align:middle line:84% you might have within a given defined category of interest 01:00:59.590 --> 01:01:01.270 align:middle line:90% from a human perspective. 01:01:01.270 --> 01:01:03.610 align:middle line:84% Anyway, that's enough philosophy for now. 01:01:03.610 --> 01:01:05.670 align:middle line:84% So essentially, data plus plus says 01:01:05.670 --> 01:01:09.970 align:middle line:84% that you might be able to sample from an implicit generative 01:01:09.970 --> 01:01:12.410 align:middle line:84% model and use it to act like some 01:01:12.410 --> 01:01:18.490 align:middle line:84% of decorative or interactive or mechanism to get data 01:01:18.490 --> 01:01:20.050 align:middle line:90% with extra functionality. 01:01:20.050 --> 01:01:22.070 align:middle line:84% And so you can define operators on that, 01:01:22.070 --> 01:01:24.670 align:middle line:84% things like interpolation, extrapolation, manipulation, 01:01:24.670 --> 01:01:26.890 align:middle line:84% composition, optimization, and you 01:01:26.890 --> 01:01:30.650 align:middle line:84% can use it to label data potentially quite efficiently. 01:01:30.650 --> 01:01:34.210 align:middle line:84% So the argument here is that maybe everything you 01:01:34.210 --> 01:01:35.930 align:middle line:84% can do with regular data, you could also 01:01:35.930 --> 01:01:40.330 align:middle line:84% try to do with this generation-augmented data. 01:01:40.330 --> 01:01:41.630 align:middle line:90% And it might work better. 01:01:41.630 --> 01:01:43.930 align:middle line:84% But I think there's still a lot of research that 01:01:43.930 --> 01:01:48.610 align:middle line:84% needs to be done in this space to really achieve 01:01:48.610 --> 01:01:51.170 align:middle line:84% the benefits that we think there's potential for. 01:01:51.170 --> 01:01:53.270 align:middle line:84% We need generators that can learn with less data, 01:01:53.270 --> 01:01:57.058 align:middle line:84% and we need to better understand how 01:01:57.058 --> 01:01:58.350 align:middle line:90% these things are being learned. 01:01:58.350 --> 01:01:58.930 align:middle line:90% Yeah. 01:01:58.930 --> 01:02:00.055 align:middle line:90% AUDIENCE: So two questions. 01:02:00.055 --> 01:02:01.030 align:middle line:90% One on the bias. 01:02:01.030 --> 01:02:04.050 align:middle line:84% So I'm just curious to know, how can we figure out if there's 01:02:04.050 --> 01:02:05.470 align:middle line:90% biases in the training? 01:02:05.470 --> 01:02:10.110 align:middle line:84% Like in this case, we correlate darkness equals volcano eruption 01:02:10.110 --> 01:02:12.470 align:middle line:84% without seeing the images, because this 01:02:12.470 --> 01:02:15.110 align:middle line:84% is-- we can see it visually, but if the data set is 01:02:15.110 --> 01:02:16.210 align:middle line:90% so large we can't see it. 01:02:16.210 --> 01:02:18.630 align:middle line:84% And the second one is, how do we prioritize 01:02:18.630 --> 01:02:20.890 align:middle line:90% diversity versus accuracy? 01:02:20.890 --> 01:02:23.070 align:middle line:84% So the example you gave of a little bird 01:02:23.070 --> 01:02:26.430 align:middle line:84% with the eagle features, but it's the size of little bird. 01:02:26.430 --> 01:02:29.910 align:middle line:84% Is it great because it allows me to train for things 01:02:29.910 --> 01:02:32.270 align:middle line:84% that aren't even possible, so maybe my testing data 01:02:32.270 --> 01:02:33.150 align:middle line:90% will improve? 01:02:33.150 --> 01:02:35.310 align:middle line:84% But it's so far away from fidelity. 01:02:35.310 --> 01:02:37.630 align:middle line:84% Does that make this driven data bad? 01:02:37.630 --> 01:02:38.470 align:middle line:90% SARA BEERY: Yeah. 01:02:38.470 --> 01:02:40.820 align:middle line:84% So it's kind of like that question about the third eye 01:02:40.820 --> 01:02:42.070 align:middle line:90% in the middle of the forehead. 01:02:42.070 --> 01:02:43.903 align:middle line:84% If you start adding random eyes to your mom, 01:02:43.903 --> 01:02:45.630 align:middle line:84% and it makes the model more robust 01:02:45.630 --> 01:02:51.030 align:middle line:84% at recognizing your mom in real data, is that a problem? 01:02:51.030 --> 01:02:53.510 align:middle line:84% So with relation to the first thing, 01:02:53.510 --> 01:02:55.430 align:middle line:84% for example, that data set interfaces paper 01:02:55.430 --> 01:02:58.070 align:middle line:84% was a mechanism that tries to use the fact that you can 01:02:58.070 --> 01:03:01.070 align:middle line:84% interact with these systems with text 01:03:01.070 --> 01:03:03.687 align:middle line:84% to probe biases efficiently without needing to explicitly 01:03:03.687 --> 01:03:04.770 align:middle line:90% look through all the data. 01:03:04.770 --> 01:03:06.562 align:middle line:84% So I think there's ways people have thought 01:03:06.562 --> 01:03:08.250 align:middle line:84% of using generative models to help 01:03:08.250 --> 01:03:10.330 align:middle line:90% you find biases efficiently. 01:03:10.330 --> 01:03:12.650 align:middle line:84% And then with respect to the second one, 01:03:12.650 --> 01:03:14.350 align:middle line:90% I think it basically depends. 01:03:14.350 --> 01:03:18.410 align:middle line:90% 01:03:18.410 --> 01:03:21.930 align:middle line:84% There's some work that tries to demonstrate 01:03:21.930 --> 01:03:24.170 align:middle line:84% that even for really fine grained things like bird 01:03:24.170 --> 01:03:28.290 align:middle line:84% species, you can learn to generate data that 01:03:28.290 --> 01:03:30.290 align:middle line:84% is accurately of that bird species 01:03:30.290 --> 01:03:37.343 align:middle line:84% by using a classification model as guidance for the generator. 01:03:37.343 --> 01:03:39.010 align:middle line:84% But essentially, what that means is then 01:03:39.010 --> 01:03:40.690 align:middle line:84% you're building a generator that's 01:03:40.690 --> 01:03:43.275 align:middle line:84% optimized for a given classification model, 01:03:43.275 --> 01:03:45.150 align:middle line:84% and you can take that to its logical extreme, 01:03:45.150 --> 01:03:47.330 align:middle line:84% because it essentially means you can map out 01:03:47.330 --> 01:03:50.650 align:middle line:84% the boundaries of that classification model really 01:03:50.650 --> 01:03:54.270 align:middle line:84% explicitly moving into spaces that aren't super realistic. 01:03:54.270 --> 01:03:57.850 align:middle line:84% And then you can maybe get better performance 01:03:57.850 --> 01:04:03.290 align:middle line:84% for a given manifold or a given representation space. 01:04:03.290 --> 01:04:07.350 align:middle line:84% But it might not always correspond to things you want. 01:04:07.350 --> 01:04:10.710 align:middle line:84% So for example, you get better performance on crows 01:04:10.710 --> 01:04:14.190 align:middle line:84% when you add in images where it's like Bigfoot with a crow 01:04:14.190 --> 01:04:15.070 align:middle line:90% texture. 01:04:15.070 --> 01:04:17.830 align:middle line:84% And it kind of gets to this weird question of, 01:04:17.830 --> 01:04:20.647 align:middle line:84% we're getting better results on our test data, 01:04:20.647 --> 01:04:22.230 align:middle line:84% but is it actually going to correspond 01:04:22.230 --> 01:04:26.230 align:middle line:84% to better results in practice beyond this 01:04:26.230 --> 01:04:28.030 align:middle line:90% of fixed set of test data? 01:04:28.030 --> 01:04:31.210 align:middle line:84% And I think that gets kind of hard to quantify. 01:04:31.210 --> 01:04:35.050 align:middle line:84% So we're good at optimizing for metrics on test data sets, 01:04:35.050 --> 01:04:38.390 align:middle line:84% but I think it's much harder if you're moving to the real world. 01:04:38.390 --> 01:04:41.190 align:middle line:84% You could imagine it might actually be better in some ways 01:04:41.190 --> 01:04:48.470 align:middle line:84% to have things that aren't, I guess, different in ways that-- 01:04:48.470 --> 01:04:50.410 align:middle line:90% weird and not realistic. 01:04:50.410 --> 01:04:52.598 align:middle line:84% Because maybe in the real world, you 01:04:52.598 --> 01:04:54.390 align:middle line:84% might be more likely to encounter something 01:04:54.390 --> 01:04:57.170 align:middle line:84% that actually does look a little bit like a small eagle, 01:04:57.170 --> 01:04:59.430 align:middle line:84% but it's actually something else entirely. 01:04:59.430 --> 01:05:02.650 align:middle line:84% So maybe the open endedness of the real world 01:05:02.650 --> 01:05:06.210 align:middle line:84% makes the question of maximizing the potential 01:05:06.210 --> 01:05:08.290 align:middle line:84% within a kind of a fixed, closed world 01:05:08.290 --> 01:05:12.050 align:middle line:84% setting a little bit less translatable. 01:05:12.050 --> 01:05:12.550 align:middle line:90% But yeah. 01:05:12.550 --> 01:05:14.258 align:middle line:84% I don't know if that kind of makes sense. 01:05:14.258 --> 01:05:15.130 align:middle line:90% I'm sort of ranting. 01:05:15.130 --> 01:05:18.890 align:middle line:84% But I do think that what is useful 01:05:18.890 --> 01:05:22.010 align:middle line:84% is going to be directly related to your task. 01:05:22.010 --> 01:05:25.230 align:middle line:84% So for different types of tasks, for different goals, 01:05:25.230 --> 01:05:28.450 align:middle line:90% useful data will be different. 01:05:28.450 --> 01:05:29.650 align:middle line:90% Yeah. 01:05:29.650 --> 01:05:31.850 align:middle line:84% AUDIENCE: Is image domain the bulk 01:05:31.850 --> 01:05:34.330 align:middle line:84% of where the cutting edge research is 01:05:34.330 --> 01:05:39.098 align:middle line:84% being done for this topic, or is that a pedagogical reason? 01:05:39.098 --> 01:05:41.390 align:middle line:84% SARA BEERY: Why am I showing a bunch of image examples? 01:05:41.390 --> 01:05:42.410 align:middle line:90% AUDIENCE: [INAUDIBLE]. 01:05:42.410 --> 01:05:45.910 align:middle line:84% It would probably be harder to do these evaluations on text 01:05:45.910 --> 01:05:47.690 align:middle line:84% because you'd have to each read each one 01:05:47.690 --> 01:05:49.410 align:middle line:90% and see the differences. 01:05:49.410 --> 01:05:50.290 align:middle line:90% SARA BEERY: Yeah. 01:05:50.290 --> 01:05:51.370 align:middle line:90% That's interesting. 01:05:51.370 --> 01:05:54.810 align:middle line:84% So there definitely is research on generative text 01:05:54.810 --> 01:05:57.330 align:middle line:84% as data augmentation, but I think 01:05:57.330 --> 01:05:59.290 align:middle line:84% you're right that it's kind of harder-- 01:05:59.290 --> 01:06:03.910 align:middle line:84% class boundaries and text is a little bit harder to define. 01:06:03.910 --> 01:06:08.730 align:middle line:84% And also, I could imagine if you think about generative songs, 01:06:08.730 --> 01:06:13.590 align:middle line:84% for example, maybe part of it is just the nuanced, messy space 01:06:13.590 --> 01:06:16.310 align:middle line:90% of what a category is. 01:06:16.310 --> 01:06:19.270 align:middle line:84% But there definitely is work on both generating data 01:06:19.270 --> 01:06:22.030 align:middle line:84% and using generative data for training in the tech space 01:06:22.030 --> 01:06:26.470 align:middle line:84% and in the audio space and in many other modalities as well. 01:06:26.470 --> 01:06:27.570 align:middle line:90% Lots of work. 01:06:27.570 --> 01:06:29.350 align:middle line:84% Actually, classical work that used 01:06:29.350 --> 01:06:32.190 align:middle line:84% generated data for training in medical imagery, 01:06:32.190 --> 01:06:34.670 align:middle line:84% because often, we had so few real data 01:06:34.670 --> 01:06:37.830 align:middle line:84% that really carefully handcrafted synthesizers of data 01:06:37.830 --> 01:06:39.630 align:middle line:90% were quite effective. 01:06:39.630 --> 01:06:42.830 align:middle line:84% There is kind of a much broader kind of interesting debate 01:06:42.830 --> 01:06:47.030 align:middle line:84% currently about what's called model collapse 01:06:47.030 --> 01:06:49.530 align:middle line:84% or manifold collapse where, essentially, 01:06:49.530 --> 01:06:53.390 align:middle line:84% if you start training a model on more data that's generated 01:06:53.390 --> 01:06:55.990 align:middle line:84% than data that's real-- which is a very clear concern 01:06:55.990 --> 01:06:58.890 align:middle line:84% today, because more and more of the text on the internet 01:06:58.890 --> 01:07:01.490 align:middle line:84% is generated text as opposed to real text-- 01:07:01.490 --> 01:07:04.330 align:middle line:84% there's an entire line of research that essentially shows 01:07:04.330 --> 01:07:10.610 align:middle line:84% that once you go too far down the ratio of more 01:07:10.610 --> 01:07:13.210 align:middle line:84% of the data being generated than real, 01:07:13.210 --> 01:07:14.750 align:middle line:84% it becomes kind of a vicious cycle. 01:07:14.750 --> 01:07:16.930 align:middle line:84% And you can have these types of model collapse 01:07:16.930 --> 01:07:19.770 align:middle line:84% where you start to lose parts of your distribution 01:07:19.770 --> 01:07:22.128 align:middle line:84% that were captured in the real data because 01:07:22.128 --> 01:07:24.170 align:middle line:84% of the kind of distribution of the generated data 01:07:24.170 --> 01:07:25.585 align:middle line:90% based on the likelihood. 01:07:25.585 --> 01:07:28.210 align:middle line:84% And there's fascinating-- it's a pretty active area of research 01:07:28.210 --> 01:07:29.730 align:middle line:84% right now, this question of, what 01:07:29.730 --> 01:07:33.970 align:middle line:84% does it look like to have models trained from generated data? 01:07:33.970 --> 01:07:37.830 align:middle line:84% And what are the kind of failure modes that can result in? 01:07:37.830 --> 01:07:38.330 align:middle line:90% Cool. 01:07:38.330 --> 01:07:40.610 align:middle line:84% So in the last little bit, I'm going 01:07:40.610 --> 01:07:43.810 align:middle line:84% to talk a little bit about learning to learn quickly. 01:07:43.810 --> 01:07:46.953 align:middle line:84% So this idea of not just, what is a mechanism 01:07:46.953 --> 01:07:47.870 align:middle line:90% for transfer learning? 01:07:47.870 --> 01:07:50.690 align:middle line:84% But actually, how do you learn a model that 01:07:50.690 --> 01:07:55.050 align:middle line:84% is good at transfer learning, a model that adapts quickly? 01:07:55.050 --> 01:07:58.550 align:middle line:84% So here, the idea is that maybe instead 01:07:58.550 --> 01:08:00.430 align:middle line:84% of just taking the classical approach where 01:08:00.430 --> 01:08:03.030 align:middle line:84% you have learning, which is this middle part where you have 01:08:03.030 --> 01:08:05.650 align:middle line:84% a data set, a learner, you get some parameters, 01:08:05.650 --> 01:08:07.778 align:middle line:84% and then at inference time, you take your data set, 01:08:07.778 --> 01:08:09.570 align:middle line:84% you run it through a model with parameters, 01:08:09.570 --> 01:08:13.950 align:middle line:84% and you get predictions, you add this meta layer on top 01:08:13.950 --> 01:08:15.890 align:middle line:84% of taking a bunch of different data sets 01:08:15.890 --> 01:08:17.590 align:middle line:84% and training some of meta learner 01:08:17.590 --> 01:08:19.910 align:middle line:84% that has a different set of hyperparameters 01:08:19.910 --> 01:08:25.390 align:middle line:84% that then you transfer to any given learner and data set. 01:08:25.390 --> 01:08:28.069 align:middle line:84% And so one of the seminal works in this space 01:08:28.069 --> 01:08:30.990 align:middle line:84% was this work called MAML by Chelsea Finn-- 01:08:30.990 --> 01:08:34.710 align:middle line:84% stands for model-agnostic meta-learning. 01:08:34.710 --> 01:08:36.630 align:middle line:84% And so far, we saw transfer learning as, 01:08:36.630 --> 01:08:39.250 align:middle line:84% given some prior model and representation, 01:08:39.250 --> 01:08:41.510 align:middle line:90% how can we quickly adapt it? 01:08:41.510 --> 01:08:44.010 align:middle line:84% And this was the first work that tried to ask, 01:08:44.010 --> 01:08:47.710 align:middle line:84% could we learn to do transfer learning effectively? 01:08:47.710 --> 01:08:50.149 align:middle line:84% So now, consider you're doing fine tuning. 01:08:50.149 --> 01:08:53.710 align:middle line:84% Given some initialization we, update it with SGD. 01:08:53.710 --> 01:08:56.930 align:middle line:84% MAML tries to learn an initialization such 01:08:56.930 --> 01:08:59.689 align:middle line:84% that it's easily adaptable with just a few steps 01:08:59.689 --> 01:09:02.090 align:middle line:90% of stochastic gradient descent. 01:09:02.090 --> 01:09:04.890 align:middle line:84% And so if you look at this algorithmically, 01:09:04.890 --> 01:09:09.890 align:middle line:84% essentially, you now are doing like a double loop 01:09:09.890 --> 01:09:11.490 align:middle line:90% in your optimization. 01:09:11.490 --> 01:09:14.090 align:middle line:84% So starting with your initialization, 01:09:14.090 --> 01:09:18.870 align:middle line:84% for any given data set that you're going to train over, 01:09:18.870 --> 01:09:21.569 align:middle line:84% you now run a few steps of SGD and then compute loss 01:09:21.569 --> 01:09:23.130 align:middle line:90% after those steps. 01:09:23.130 --> 01:09:26.010 align:middle line:84% And then you backprop through that back 01:09:26.010 --> 01:09:30.850 align:middle line:84% to your initial meta-learning initial set of parameters. 01:09:30.850 --> 01:09:32.930 align:middle line:84% So how do you actually backprop through something 01:09:32.930 --> 01:09:38.130 align:middle line:84% that's now like an inner loop within your larger loop? 01:09:38.130 --> 01:09:41.689 align:middle line:84% So if you're doing one step of stochastic gradient descent, 01:09:41.689 --> 01:09:45.770 align:middle line:84% you essentially can consider that a forwards pass 01:09:45.770 --> 01:09:49.450 align:middle line:84% over x plus the backwards of the forwards of x and theta 01:09:49.450 --> 01:09:50.130 align:middle line:90% plus theta. 01:09:50.130 --> 01:09:54.530 align:middle line:84% Essentially, remember when we took a single part of a model, 01:09:54.530 --> 01:09:57.090 align:middle line:84% and then we showed that if you wanted to do back propagation, 01:09:57.090 --> 01:09:59.910 align:middle line:84% you could basically unfold it, and it was essentially 01:09:59.910 --> 01:10:03.430 align:middle line:84% one big forward pass of going through forward 01:10:03.430 --> 01:10:06.270 align:middle line:90% and then backward of a model? 01:10:06.270 --> 01:10:09.630 align:middle line:84% So now if you want to do two steps of stochastic gradient 01:10:09.630 --> 01:10:12.830 align:middle line:84% descent, again, you're kind of nesting this. 01:10:12.830 --> 01:10:15.630 align:middle line:84% And so essentially, this is saying that you 01:10:15.630 --> 01:10:19.910 align:middle line:90% can backprop through backward. 01:10:19.910 --> 01:10:21.667 align:middle line:84% So if you're going forward and backward 01:10:21.667 --> 01:10:22.750 align:middle line:90% through a simple network-- 01:10:22.750 --> 01:10:24.910 align:middle line:84% I don't know if you guys remember this. 01:10:24.910 --> 01:10:27.513 align:middle line:90% So here's that forward pass. 01:10:27.513 --> 01:10:28.930 align:middle line:84% Then, you have your backward pass. 01:10:28.930 --> 01:10:31.290 align:middle line:84% And then here, this is where we kind of unfold it. 01:10:31.290 --> 01:10:36.350 align:middle line:84% And so now, you can show that doing forward and backward 01:10:36.350 --> 01:10:38.870 align:middle line:84% through a simple network is just one forward pass. 01:10:38.870 --> 01:10:44.030 align:middle line:84% And then you can do this meta backward through everything. 01:10:44.030 --> 01:10:46.430 align:middle line:90% So it's possible. 01:10:46.430 --> 01:10:48.630 align:middle line:84% It's just expensive, essentially, 01:10:48.630 --> 01:10:51.170 align:middle line:84% is the main point here, because you're basically 01:10:51.170 --> 01:10:55.038 align:middle line:84% taking those nested loops of forward and backward, 01:10:55.038 --> 01:10:56.830 align:middle line:84% and you're just stretching all of them out. 01:10:56.830 --> 01:11:00.890 align:middle line:84% And you're now getting what is a very, very deep network 01:11:00.890 --> 01:11:04.130 align:middle line:84% that you're going to optimize through. 01:11:04.130 --> 01:11:08.570 align:middle line:84% Now, this type of computational complexity 01:11:08.570 --> 01:11:10.730 align:middle line:84% often means that the relative benefit 01:11:10.730 --> 01:11:13.610 align:middle line:84% of these types of meta-learning models-- though nice in theory, 01:11:13.610 --> 01:11:16.712 align:middle line:84% and you get really nice gains on maybe simple problems-- 01:11:16.712 --> 01:11:18.170 align:middle line:84% it's often just too computationally 01:11:18.170 --> 01:11:21.770 align:middle line:90% expensive to really deploy. 01:11:21.770 --> 01:11:26.770 align:middle line:84% But there are a few examples of how people have actually 01:11:26.770 --> 01:11:29.170 align:middle line:84% used these types of meta-learning 01:11:29.170 --> 01:11:31.490 align:middle line:90% for real world applications. 01:11:31.490 --> 01:11:33.462 align:middle line:84% So in this case, in practice, I was 01:11:33.462 --> 01:11:35.170 align:middle line:84% saying the gains of the method, something 01:11:35.170 --> 01:11:36.590 align:middle line:84% like model-agnostic meta-learning, 01:11:36.590 --> 01:11:38.530 align:middle line:84% are often not worth the additional computational cost 01:11:38.530 --> 01:11:40.210 align:middle line:84% at training time, and sometimes can 01:11:40.210 --> 01:11:43.370 align:middle line:84% be worse than just like pretraining on all of those 01:11:43.370 --> 01:11:46.530 align:middle line:84% different data sets and then fine tuning. 01:11:46.530 --> 01:11:51.170 align:middle line:84% But Gabi Tseng introduced this geospatial embedding, 01:11:51.170 --> 01:11:55.190 align:middle line:84% essentially like a bit of a task-specific tokenizer, 01:11:55.190 --> 01:11:57.630 align:middle line:84% within that meta-learning objective 01:11:57.630 --> 01:12:00.890 align:middle line:84% that encoded some kind of task-specific context. 01:12:00.890 --> 01:12:04.470 align:middle line:84% And they showed that giving this task specificity 01:12:04.470 --> 01:12:06.590 align:middle line:84% within that meta-learning loop did give them 01:12:06.590 --> 01:12:09.910 align:middle line:84% some really nice gains when they were looking 01:12:09.910 --> 01:12:12.270 align:middle line:84% at crop type categorization for corn 01:12:12.270 --> 01:12:14.110 align:middle line:84% and a bunch of different countries. 01:12:14.110 --> 01:12:16.510 align:middle line:84% So here, these are a comparison of a bunch 01:12:16.510 --> 01:12:17.890 align:middle line:90% of different types of methods. 01:12:17.890 --> 01:12:19.710 align:middle line:84% But you can see at the top is kind 01:12:19.710 --> 01:12:24.570 align:middle line:84% of versus meta-learning or just pretraining a model, there, 01:12:24.570 --> 01:12:25.970 align:middle line:90% it's about the same. 01:12:25.970 --> 01:12:29.150 align:middle line:84% But this task conditioning that gives your model the ability 01:12:29.150 --> 01:12:31.910 align:middle line:84% to specialize effectively, maybe to a given country or part 01:12:31.910 --> 01:12:34.847 align:middle line:84% of the world, ends up being pretty effective. 01:12:34.847 --> 01:12:36.930 align:middle line:84% This is maybe-- I don't know if you guys remember, 01:12:36.930 --> 01:12:38.710 align:middle line:84% but we talked a bit about this idea 01:12:38.710 --> 01:12:40.670 align:middle line:84% of maybe learning a prior when we were talking 01:12:40.670 --> 01:12:42.430 align:middle line:90% about distribution shift. 01:12:42.430 --> 01:12:44.070 align:middle line:84% This is maybe a different mechanism 01:12:44.070 --> 01:12:46.750 align:middle line:84% of trying to learn something like a geospatial prior 01:12:46.750 --> 01:12:49.730 align:middle line:84% or some of useful contextualization 01:12:49.730 --> 01:12:53.890 align:middle line:84% so that your model can actually build some knowledge about how 01:12:53.890 --> 01:12:56.210 align:middle line:84% it should adapt for specific types of problems 01:12:56.210 --> 01:12:59.290 align:middle line:90% or types of tasks. 01:12:59.290 --> 01:13:00.890 align:middle line:84% And so there's also this question 01:13:00.890 --> 01:13:03.810 align:middle line:84% of meta-learning by sequence modeling. 01:13:03.810 --> 01:13:06.290 align:middle line:90% So definition of learning. 01:13:06.290 --> 01:13:09.810 align:middle line:84% Maybe we have a sequence of input 01:13:09.810 --> 01:13:12.170 align:middle line:84% and ground truth, input and ground truth 01:13:12.170 --> 01:13:13.610 align:middle line:90% for a bunch of things. 01:13:13.610 --> 01:13:16.050 align:middle line:84% And then you output some function such 01:13:16.050 --> 01:13:18.610 align:middle line:84% that if you give it a new input, it'll 01:13:18.610 --> 01:13:20.980 align:middle line:90% give you a new ground truth. 01:13:20.980 --> 01:13:22.730 align:middle line:84% So maybe this is kind of like that example 01:13:22.730 --> 01:13:23.930 align:middle line:84% we talked about when we were talking 01:13:23.930 --> 01:13:26.410 align:middle line:84% about autoregressive modeling for large language models 01:13:26.410 --> 01:13:28.010 align:middle line:90% where we could say, OK. 01:13:28.010 --> 01:13:30.050 align:middle line:84% We're going to show you an example of English 01:13:30.050 --> 01:13:31.770 align:middle line:90% and the translation in Spanish. 01:13:31.770 --> 01:13:33.230 align:middle line:84% I'm going to show you another example in English, 01:13:33.230 --> 01:13:34.830 align:middle line:84% another example of the translation in Spanish. 01:13:34.830 --> 01:13:36.497 align:middle line:84% Now, if I show you something in English, 01:13:36.497 --> 01:13:38.650 align:middle line:84% the model would automatically generate 01:13:38.650 --> 01:13:40.510 align:middle line:90% the translation in Spanish. 01:13:40.510 --> 01:13:44.650 align:middle line:84% This idea of sequence based in-context learning. 01:13:44.650 --> 01:13:46.710 align:middle line:84% And so here, if you have something 01:13:46.710 --> 01:13:48.830 align:middle line:84% like a recurrent neural network, the state 01:13:48.830 --> 01:13:51.900 align:middle line:84% plus the parameters of that recurrent neural network, 01:13:51.900 --> 01:13:53.650 align:middle line:84% or something like an autoregressive model, 01:13:53.650 --> 01:13:58.050 align:middle line:84% like predicting the next word, is f, that current state. 01:13:58.050 --> 01:14:00.370 align:middle line:84% And then you apply that f to your new input. 01:14:00.370 --> 01:14:03.430 align:middle line:84% You're basically saying, predict the next thing. 01:14:03.430 --> 01:14:05.710 align:middle line:84% And so then the hidden state of that RNN 01:14:05.710 --> 01:14:09.670 align:middle line:84% or that recurrent neural network is a model that arguably 01:14:09.670 --> 01:14:11.230 align:middle line:90% learns to learn, right? 01:14:11.230 --> 01:14:13.590 align:middle line:84% It takes some statistics of the sequence in general 01:14:13.590 --> 01:14:17.070 align:middle line:84% that enable it to predict what should come next. 01:14:17.070 --> 01:14:19.270 align:middle line:84% And this really is, I think, what 01:14:19.270 --> 01:14:23.470 align:middle line:84% GPT and a lot of these modern autoregressive models are doing. 01:14:23.470 --> 01:14:25.770 align:middle line:84% It's kind of a version of meta-learning at large scale. 01:14:25.770 --> 01:14:27.603 align:middle line:84% And you could argue that in-context learning 01:14:27.603 --> 01:14:30.950 align:middle line:84% is kind of just another mechanism of meta-learning. 01:14:30.950 --> 01:14:33.410 align:middle line:84% Because when these models are being trained, 01:14:33.410 --> 01:14:36.470 align:middle line:84% you're showing them a lot of different possible tasks 01:14:36.470 --> 01:14:38.710 align:middle line:84% or sequences and then asking them 01:14:38.710 --> 01:14:43.330 align:middle line:84% to learn how to really quickly and efficiently generate 01:14:43.330 --> 01:14:44.610 align:middle line:90% answers. 01:14:44.610 --> 01:14:47.090 align:middle line:84% What comes next for new sequences? 01:14:47.090 --> 01:14:52.850 align:middle line:84% So you could argue that there's a pretty similar motivation 01:14:52.850 --> 01:14:57.010 align:middle line:84% and potentially a much better actual output when you think 01:14:57.010 --> 01:15:00.655 align:middle line:84% about meta-learning through sequence modeling 01:15:00.655 --> 01:15:02.530 align:middle line:84% as opposed to meta-learning through this kind 01:15:02.530 --> 01:15:05.870 align:middle line:84% of inner and outer loop of training. 01:15:05.870 --> 01:15:06.370 align:middle line:90% Cool. 01:15:06.370 --> 01:15:09.610 align:middle line:84% So it's the end of the lecture I had for today. 01:15:09.610 --> 01:15:12.250 align:middle line:84% We learned today about generative models as data 01:15:12.250 --> 01:15:14.030 align:middle line:84% and some of the pros and cons of that, 01:15:14.030 --> 01:15:15.970 align:middle line:84% and we also talked about learning to learn, 01:15:15.970 --> 01:15:17.810 align:middle line:90% or meta-learning. 01:15:17.810 --> 01:15:21.550 align:middle line:84% And like I said, it's my last lecture of the course today. 01:15:21.550 --> 01:15:24.130 align:middle line:84% So it's been really fun teaching you all, and hopefully-- 01:15:24.130 --> 01:15:25.450 align:middle line:90% I mean, I'll be around. 01:15:25.450 --> 01:15:28.200 align:middle line:90% [APPLAUSE] 01:15:28.200 --> 01:15:43.000 align:middle line:90%