WEBVTT 00:00:00.000 --> 00:00:01.955 align:middle line:90% [SQUEAKING] 00:00:01.955 --> 00:00:04.380 align:middle line:90% [RUSTLING] 00:00:04.380 --> 00:00:06.805 align:middle line:90% [CLICKING] 00:00:06.805 --> 00:00:12.970 align:middle line:90% 00:00:12.970 --> 00:00:14.220 align:middle line:90% SARA BEERY: Let's get started. 00:00:14.220 --> 00:00:18.080 align:middle line:90% 00:00:18.080 --> 00:00:22.520 align:middle line:84% So today, we're going to kick off the first of a little two 00:00:22.520 --> 00:00:25.520 align:middle line:84% lecture mini series on transfer learning, 00:00:25.520 --> 00:00:30.220 align:middle line:84% specifically how we think about taking information that 00:00:30.220 --> 00:00:31.720 align:middle line:84% might have been learned from a model 00:00:31.720 --> 00:00:37.420 align:middle line:84% and then transferring it to some new task, some new challenge, 00:00:37.420 --> 00:00:38.660 align:middle line:90% maybe some new distribution. 00:00:38.660 --> 00:00:42.440 align:middle line:84% We talked about distribution shifts last week. 00:00:42.440 --> 00:00:46.920 align:middle line:84% So at a high level, the idea of transfer learning 00:00:46.920 --> 00:00:51.000 align:middle line:84% is that you want to try to learn even if you 00:00:51.000 --> 00:00:53.240 align:middle line:90% don't have very much data. 00:00:53.240 --> 00:00:57.040 align:middle line:84% And so we've tried to structure this framing of transfer 00:00:57.040 --> 00:01:00.272 align:middle line:84% learning into these different sub areas. 00:01:00.272 --> 00:01:02.480 align:middle line:84% So first, we're going to talk about how you can maybe 00:01:02.480 --> 00:01:05.319 align:middle line:84% transfer knowledge from a machine learning model 00:01:05.319 --> 00:01:08.880 align:middle line:84% that you have about the mapping from inputs to outputs. 00:01:08.880 --> 00:01:11.418 align:middle line:84% And so within that, we're going to talk about fine-tuning, 00:01:11.418 --> 00:01:12.960 align:middle line:84% and then we'll also talk a little bit 00:01:12.960 --> 00:01:15.960 align:middle line:90% about domain adaptation. 00:01:15.960 --> 00:01:17.960 align:middle line:84% And then we'll talk about transferring knowledge 00:01:17.960 --> 00:01:20.860 align:middle line:84% about the outputs, and so in that case, 00:01:20.860 --> 00:01:23.522 align:middle line:84% we're going to be discussing knowledge distillation. 00:01:23.522 --> 00:01:25.480 align:middle line:84% And then we're going to be talking a little bit 00:01:25.480 --> 00:01:29.040 align:middle line:84% about foundation models and maybe what this new paradigm is 00:01:29.040 --> 00:01:31.920 align:middle line:84% in terms of transfer learning, kind of introducing 00:01:31.920 --> 00:01:35.200 align:middle line:84% this idea of in-context learning or prompting 00:01:35.200 --> 00:01:39.560 align:middle line:84% as a mechanism for transfer learning by just asking. 00:01:39.560 --> 00:01:43.920 align:middle line:84% And then next Tuesday, we're going to finish this up 00:01:43.920 --> 00:01:46.960 align:middle line:84% with one further discussion of how we might transfer knowledge 00:01:46.960 --> 00:01:49.680 align:middle line:90% about the inputs of the data. 00:01:49.680 --> 00:01:53.820 align:middle line:90% So why deep learning? 00:01:53.820 --> 00:01:56.040 align:middle line:84% This is a slide from Andrew Ng, who, of course, 00:01:56.040 --> 00:02:01.280 align:middle line:84% I think has a very, very famous lecture series on deep learning. 00:02:01.280 --> 00:02:05.080 align:middle line:84% And here, he essentially is pointing out 00:02:05.080 --> 00:02:08.680 align:middle line:84% that deep learning might be different from older 00:02:08.680 --> 00:02:12.840 align:middle line:84% or standard learning algorithms specifically in terms 00:02:12.840 --> 00:02:15.080 align:middle line:84% of how it scales with the amount of data 00:02:15.080 --> 00:02:16.580 align:middle line:90% that you have to train on. 00:02:16.580 --> 00:02:19.480 align:middle line:84% It might be a little bit hard to see with the lights as they are. 00:02:19.480 --> 00:02:23.380 align:middle line:84% But essentially, the idea is that many learning algorithms, 00:02:23.380 --> 00:02:25.343 align:middle line:84% older learning algorithms, saturate. 00:02:25.343 --> 00:02:26.760 align:middle line:84% With some amount of data, you just 00:02:26.760 --> 00:02:29.200 align:middle line:84% don't see performance continuing to improve. 00:02:29.200 --> 00:02:34.120 align:middle line:84% But we have really not seen that type of saturation 00:02:34.120 --> 00:02:36.400 align:middle line:84% as we scale the amount of data that we're giving 00:02:36.400 --> 00:02:38.400 align:middle line:90% these deep learning algorithms. 00:02:38.400 --> 00:02:41.000 align:middle line:84% And next Thursday, you guys will have a lecture 00:02:41.000 --> 00:02:42.800 align:middle line:84% from Phil around some of these scaling 00:02:42.800 --> 00:02:45.000 align:middle line:84% laws, these relationships that people 00:02:45.000 --> 00:02:47.160 align:middle line:84% are starting to learn about the scale 00:02:47.160 --> 00:02:50.560 align:middle line:84% of the number of parameters, the amount of available data. 00:02:50.560 --> 00:02:59.270 align:middle line:84% But one kind of high level analysis that was done 00:02:59.270 --> 00:03:02.030 align:middle line:84% seemed to suggest that with the current capacity of machine 00:03:02.030 --> 00:03:08.950 align:middle line:84% learning models we have, we basically have far more capacity 00:03:08.950 --> 00:03:10.030 align:middle line:90% than data. 00:03:10.030 --> 00:03:13.670 align:middle line:84% So essentially, the idea that there's a lot more room 00:03:13.670 --> 00:03:15.970 align:middle line:84% to scale on the data dimension still 00:03:15.970 --> 00:03:18.430 align:middle line:84% before we're really hitting the capacity of our largest 00:03:18.430 --> 00:03:20.830 align:middle line:90% deep learning models. 00:03:20.830 --> 00:03:25.070 align:middle line:84% And that's talking about billion data point data sets, things 00:03:25.070 --> 00:03:27.790 align:middle line:84% that are on the order of scale of the entire internet. 00:03:27.790 --> 00:03:31.710 align:middle line:84% We're still not hitting the actual computational capacity 00:03:31.710 --> 00:03:32.770 align:middle line:90% of these models. 00:03:32.770 --> 00:03:35.310 align:middle line:90% 00:03:35.310 --> 00:03:41.190 align:middle line:84% So few-shot learning is the idea that you can maybe learn even 00:03:41.190 --> 00:03:43.990 align:middle line:90% if you have very little data. 00:03:43.990 --> 00:03:47.670 align:middle line:84% So it turns out that humans are shockingly good at this. 00:03:47.670 --> 00:03:51.310 align:middle line:84% We're really good at recognizing patterns and translating them 00:03:51.310 --> 00:03:54.030 align:middle line:84% even to things that we've never seen. 00:03:54.030 --> 00:03:58.750 align:middle line:84% So here, if I tell you that symbol up at the top is a dax-- 00:03:58.750 --> 00:03:59.670 align:middle line:90% these are made up. 00:03:59.670 --> 00:04:00.490 align:middle line:90% This isn't real. 00:04:00.490 --> 00:04:02.550 align:middle line:84% But if I tell you that's a dax, can you guys 00:04:02.550 --> 00:04:05.250 align:middle line:84% tell me which of the ones below is also a dax? 00:04:05.250 --> 00:04:09.110 align:middle line:90% 00:04:09.110 --> 00:04:09.870 align:middle line:90% Yeah. 00:04:09.870 --> 00:04:12.330 align:middle line:84% AUDIENCE: Second row, second from the right? 00:04:12.330 --> 00:04:13.530 align:middle line:84% SARA BEERY: Second row, second from the right. 00:04:13.530 --> 00:04:14.010 align:middle line:90% Yeah. 00:04:14.010 --> 00:04:14.510 align:middle line:90% Right? 00:04:14.510 --> 00:04:17.630 align:middle line:90% It looks like a dax. 00:04:17.630 --> 00:04:19.029 align:middle line:90% OK. 00:04:19.029 --> 00:04:20.829 align:middle line:90% What about this? 00:04:20.829 --> 00:04:24.950 align:middle line:84% Which on the bottom is an example of the same concept 00:04:24.950 --> 00:04:26.470 align:middle line:90% as the item in the box? 00:04:26.470 --> 00:04:27.330 align:middle line:90% Someone else. 00:04:27.330 --> 00:04:30.390 align:middle line:90% 00:04:30.390 --> 00:04:31.070 align:middle line:90% Yeah. 00:04:31.070 --> 00:04:34.243 align:middle line:84% AUDIENCE: Second row, first to the right. 00:04:34.243 --> 00:04:34.910 align:middle line:90% SARA BEERY: Yes. 00:04:34.910 --> 00:04:36.970 align:middle line:84% So those are both Segways, right? 00:04:36.970 --> 00:04:41.110 align:middle line:84% But then this kind of begs this interesting question, 00:04:41.110 --> 00:04:44.970 align:middle line:84% because there's potentially ambiguity here. 00:04:44.970 --> 00:04:46.770 align:middle line:84% If you only have this single example, 00:04:46.770 --> 00:04:49.510 align:middle line:84% and you're telling me which is the same, what you made 00:04:49.510 --> 00:04:51.630 align:middle line:84% was the guess that was probably most 00:04:51.630 --> 00:04:55.590 align:middle line:84% similar because the orientation of the wheels and the structure. 00:04:55.590 --> 00:04:59.170 align:middle line:84% But maybe the thing I was looking for was instead, 00:04:59.170 --> 00:05:00.590 align:middle line:84% actually, I wanted you to tell me 00:05:00.590 --> 00:05:03.570 align:middle line:84% all the things that had two wheels instead of four or one. 00:05:03.570 --> 00:05:05.910 align:middle line:84% And so one of the things that's challenging when you're 00:05:05.910 --> 00:05:09.990 align:middle line:84% doing this few-shot learning type of scenario 00:05:09.990 --> 00:05:12.830 align:middle line:84% is with less data, it can be harder 00:05:12.830 --> 00:05:15.110 align:middle line:84% to understand without ambiguity what 00:05:15.110 --> 00:05:18.590 align:middle line:84% the boundaries of maybe the categories of interest are. 00:05:18.590 --> 00:05:22.710 align:middle line:84% So if I'd showed you three examples, one was a Segway, 00:05:22.710 --> 00:05:25.090 align:middle line:84% one was a scooter, and one was a bicycle, 00:05:25.090 --> 00:05:28.730 align:middle line:84% and I said, OK, find me the ones that are also in this category, 00:05:28.730 --> 00:05:31.850 align:middle line:84% it would have been much easier for you to understand, 00:05:31.850 --> 00:05:33.990 align:middle line:84% OK, she's looking for things with two wheels. 00:05:33.990 --> 00:05:36.305 align:middle line:84% Or at some different level of granularity, 00:05:36.305 --> 00:05:38.930 align:middle line:84% maybe I was trying to just learn wheeled things, in which case, 00:05:38.930 --> 00:05:41.222 align:middle line:84% I might need to have, again, more diverse examples that 00:05:41.222 --> 00:05:44.710 align:middle line:84% cover the space of the concept I want to learn. 00:05:44.710 --> 00:05:49.190 align:middle line:84% So few-shot learning really is just this idea 00:05:49.190 --> 00:05:54.150 align:middle line:84% that we can learn to maybe generalize 00:05:54.150 --> 00:05:59.630 align:middle line:84% from our past lived experience how to recognize similarity 00:05:59.630 --> 00:06:03.390 align:middle line:90% and patterns in a useful way. 00:06:03.390 --> 00:06:06.910 align:middle line:84% So up until this point, a lot of what 00:06:06.910 --> 00:06:09.468 align:middle line:84% we've talked about with deep learning 00:06:09.468 --> 00:06:11.510 align:middle line:84% is, essentially, you're going to have a data set, 00:06:11.510 --> 00:06:14.253 align:middle line:84% and you start with an empty slate. 00:06:14.253 --> 00:06:15.670 align:middle line:84% And then you're essentially trying 00:06:15.670 --> 00:06:20.110 align:middle line:84% to learn some physical rules, some concepts, some dimensions 00:06:20.110 --> 00:06:24.910 align:middle line:84% of similarity that are useful for a specific task. 00:06:24.910 --> 00:06:31.030 align:middle line:84% But what humans do instead, when they try to learn something new, 00:06:31.030 --> 00:06:34.430 align:middle line:84% is we really have a lot of prior knowledge coming 00:06:34.430 --> 00:06:37.550 align:middle line:84% from our experience in the world. 00:06:37.550 --> 00:06:39.990 align:middle line:84% From the time that we're incredibly young, 00:06:39.990 --> 00:06:43.050 align:middle line:84% we're physically embodied in the world, 00:06:43.050 --> 00:06:46.630 align:middle line:84% learning about concepts like physics, like structure. 00:06:46.630 --> 00:06:49.110 align:middle line:84% We're learning through the entirety of our education 00:06:49.110 --> 00:06:50.960 align:middle line:84% and just in our experience, in our lives, 00:06:50.960 --> 00:06:53.700 align:middle line:84% how to think about these measures of similarity 00:06:53.700 --> 00:06:54.780 align:middle line:90% and difference. 00:06:54.780 --> 00:06:58.260 align:middle line:84% And different people can have different mental models 00:06:58.260 --> 00:06:59.040 align:middle line:90% of the world. 00:06:59.040 --> 00:07:01.980 align:middle line:84% This actually comes up a lot currently 00:07:01.980 --> 00:07:04.100 align:middle line:84% when you're thinking about scaling large scale 00:07:04.100 --> 00:07:08.540 align:middle line:84% data labeling, that, actually, if you're getting data labeled 00:07:08.540 --> 00:07:10.420 align:middle line:84% from people from different cultures, 00:07:10.420 --> 00:07:12.420 align:middle line:84% they might actually have different mental models 00:07:12.420 --> 00:07:14.480 align:middle line:84% than the one that you intend in your culture. 00:07:14.480 --> 00:07:17.660 align:middle line:84% So there's this idea of, my mental model 00:07:17.660 --> 00:07:20.500 align:middle line:84% is based on some prior, and we don't necessarily know-- 00:07:20.500 --> 00:07:23.060 align:middle line:84% or there really isn't a concept of which mental model 00:07:23.060 --> 00:07:25.820 align:middle line:90% of the world is correct. 00:07:25.820 --> 00:07:27.900 align:middle line:84% But these mental models are built, 00:07:27.900 --> 00:07:30.660 align:middle line:84% and they kind of capture representations 00:07:30.660 --> 00:07:34.020 align:middle line:84% of things, kind of models we have of how the world operates, 00:07:34.020 --> 00:07:35.700 align:middle line:90% maybe different skills. 00:07:35.700 --> 00:07:39.180 align:middle line:84% And that means that we're basing anything 00:07:39.180 --> 00:07:42.360 align:middle line:84% we learn on top of a bunch of prior knowledge, 00:07:42.360 --> 00:07:44.020 align:middle line:84% basically like systems that are really 00:07:44.020 --> 00:07:46.335 align:middle line:90% well posed to learn new things. 00:07:46.335 --> 00:07:48.460 align:middle line:84% And actually, a lot of what you're doing in college 00:07:48.460 --> 00:07:50.620 align:middle line:90% is learning how to learn. 00:07:50.620 --> 00:07:53.820 align:middle line:84% You're learning how to learn new things efficiently. 00:07:53.820 --> 00:07:57.540 align:middle line:84% And so there's this question of, how can we 00:07:57.540 --> 00:08:00.020 align:middle line:84% give deep learning models or deep networks 00:08:00.020 --> 00:08:03.500 align:middle line:84% similar access to prior knowledge? 00:08:03.500 --> 00:08:06.620 align:middle line:84% And so a very, very simple way to do that 00:08:06.620 --> 00:08:11.960 align:middle line:84% is just pretrain them on other tasks or on prior tasks. 00:08:11.960 --> 00:08:16.020 align:middle line:84% So the idea that you can send your neural network to school, 00:08:16.020 --> 00:08:19.003 align:middle line:84% and you can try to get it to learn how to learn, 00:08:19.003 --> 00:08:20.420 align:middle line:84% and the way that you might do that 00:08:20.420 --> 00:08:23.940 align:middle line:84% is maybe training on a huge amount of data 00:08:23.940 --> 00:08:26.020 align:middle line:84% with some other set of categories that might not 00:08:26.020 --> 00:08:28.187 align:middle line:84% be the thing you're really explicitly interested in, 00:08:28.187 --> 00:08:31.700 align:middle line:84% but it's things that you have data for at scale. 00:08:31.700 --> 00:08:33.659 align:middle line:84% And then you can use some mechanism 00:08:33.659 --> 00:08:35.740 align:middle line:84% to transfer that knowledge from the pretained net 00:08:35.740 --> 00:08:38.140 align:middle line:90% so that you can solve new tasks. 00:08:38.140 --> 00:08:41.700 align:middle line:84% And so one view of this idea of learning 00:08:41.700 --> 00:08:43.460 align:middle line:84% from little data or transfer learning 00:08:43.460 --> 00:08:45.580 align:middle line:84% is actually the point of deep learning 00:08:45.580 --> 00:08:48.820 align:middle line:84% is to enable us to be able to do this type of problem solving 00:08:48.820 --> 00:08:50.260 align:middle line:90% with little data. 00:08:50.260 --> 00:08:52.980 align:middle line:84% So you're trying to learn these representations, 00:08:52.980 --> 00:08:57.080 align:middle line:84% these models that have a good sense of what's useful. 00:08:57.080 --> 00:09:01.420 align:middle line:84% What is a good thing to build on top of if we want to learn? 00:09:01.420 --> 00:09:03.500 align:middle line:84% So that's kind of the motivation for talking 00:09:03.500 --> 00:09:06.420 align:middle line:90% about transfer learning, right? 00:09:06.420 --> 00:09:10.140 align:middle line:84% So when we think about a model in general, 00:09:10.140 --> 00:09:12.340 align:middle line:84% usually, we have some big system, 00:09:12.340 --> 00:09:13.780 align:middle line:84% and it will map some set of inputs 00:09:13.780 --> 00:09:18.420 align:middle line:84% to some set of outputs in some very fun, complex, nonlinear 00:09:18.420 --> 00:09:21.060 align:middle line:90% way. 00:09:21.060 --> 00:09:26.640 align:middle line:84% But we're thinking about, what actually are we transferring? 00:09:26.640 --> 00:09:30.420 align:middle line:84% And so you can think about transferring 00:09:30.420 --> 00:09:33.800 align:middle line:84% some knowledge about the relationship, for example, 00:09:33.800 --> 00:09:35.658 align:middle line:90% from the inputs to the outputs. 00:09:35.658 --> 00:09:37.700 align:middle line:84% You could also think about transferring knowledge 00:09:37.700 --> 00:09:38.520 align:middle line:90% in different ways. 00:09:38.520 --> 00:09:40.340 align:middle line:84% We can think about transferring knowledge 00:09:40.340 --> 00:09:42.640 align:middle line:90% about biases in a model. 00:09:42.640 --> 00:09:45.880 align:middle line:84% You can think about transferring knowledge about an optimizer, 00:09:45.880 --> 00:09:46.940 align:middle line:90% perhaps. 00:09:46.940 --> 00:09:49.540 align:middle line:90% How do we optimize models? 00:09:49.540 --> 00:09:51.900 align:middle line:84% We can think about maybe transferring knowledge 00:09:51.900 --> 00:09:55.780 align:middle line:84% about potential error modes or ethical risks 00:09:55.780 --> 00:09:58.700 align:middle line:84% or societal implications of models. 00:09:58.700 --> 00:10:02.380 align:middle line:84% So we'll really be focusing on pretty simple mechanisms 00:10:02.380 --> 00:10:04.140 align:middle line:84% to think about knowledge transfer, 00:10:04.140 --> 00:10:06.180 align:middle line:84% but I do think that there's a lot of more 00:10:06.180 --> 00:10:11.780 align:middle line:84% broad sets of transfer or this idea of sharing information 00:10:11.780 --> 00:10:14.540 align:middle line:84% in a valuable way that are really not understood 00:10:14.540 --> 00:10:15.320 align:middle line:90% or not solved. 00:10:15.320 --> 00:10:19.300 align:middle line:84% So one thing that we still haven't really 00:10:19.300 --> 00:10:22.380 align:middle line:84% figured out how to do is building 00:10:22.380 --> 00:10:27.020 align:middle line:84% an agent that's able to actually learn some intrinsic motivation 00:10:27.020 --> 00:10:30.540 align:middle line:84% about how to explore in the world, which then could be 00:10:30.540 --> 00:10:33.140 align:middle line:84% transferred to a new agent, which would let them explore 00:10:33.140 --> 00:10:33.880 align:middle line:90% more efficiently. 00:10:33.880 --> 00:10:37.140 align:middle line:84% This idea of transferring motivation 00:10:37.140 --> 00:10:38.740 align:middle line:90% within a robotic agent. 00:10:38.740 --> 00:10:41.500 align:middle line:84% So there's a lot of ways that we have 00:10:41.500 --> 00:10:45.380 align:middle line:84% these kind of brute force knowledge transfer mechanisms 00:10:45.380 --> 00:10:46.600 align:middle line:90% that we'll talk about. 00:10:46.600 --> 00:10:48.540 align:middle line:84% But I think there's a lot of space 00:10:48.540 --> 00:10:51.700 align:middle line:84% still to explore in the machine learning research arena, just 00:10:51.700 --> 00:10:54.820 align:middle line:84% in terms of the best ways to try to share information 00:10:54.820 --> 00:10:56.260 align:middle line:90% and transfer knowledge. 00:10:56.260 --> 00:10:57.060 align:middle line:90% Yeah? 00:10:57.060 --> 00:11:00.260 align:middle line:84% AUDIENCE: How much of that is limited-- 00:11:00.260 --> 00:11:04.100 align:middle line:84% my very basic understanding of some of the challenges in, say, 00:11:04.100 --> 00:11:06.860 align:middle line:84% reinforcement learning is that it's 00:11:06.860 --> 00:11:09.100 align:middle line:84% more of an issue of sample efficiency. 00:11:09.100 --> 00:11:12.900 align:middle line:84% And thinking back to 2012, one of the breakthroughs 00:11:12.900 --> 00:11:14.800 align:middle line:84% wasn't really that the algorithms changed, 00:11:14.800 --> 00:11:17.900 align:middle line:84% but just that they finally got it [INAUDIBLE] GPU, 00:11:17.900 --> 00:11:21.260 align:middle line:84% and it sped up the training process. 00:11:21.260 --> 00:11:24.460 align:middle line:84% How much do people think that the bottleneck 00:11:24.460 --> 00:11:27.300 align:middle line:84% is the algorithm or the ideas around how 00:11:27.300 --> 00:11:30.060 align:middle line:84% to do this versus just we're not computationally 00:11:30.060 --> 00:11:31.932 align:middle line:90% efficient enough? 00:11:31.932 --> 00:11:33.640 align:middle line:84% SARA BEERY: So the question is basically, 00:11:33.640 --> 00:11:35.848 align:middle line:84% what's the bottleneck in terms of knowledge transfer? 00:11:35.848 --> 00:11:37.860 align:middle line:84% Is it fundamentally algorithmic, or might it 00:11:37.860 --> 00:11:42.110 align:middle line:84% be related to, instead, computational efficiency? 00:11:42.110 --> 00:11:44.610 align:middle line:84% And you brought up a really great example of essentially 00:11:44.610 --> 00:11:48.170 align:middle line:84% how figuring out how to use GPUs effectively 00:11:48.170 --> 00:11:50.112 align:middle line:84% was a huge breakthrough in terms of progress 00:11:50.112 --> 00:11:52.570 align:middle line:84% across many different areas of deep learning, reinforcement 00:11:52.570 --> 00:11:56.930 align:middle line:84% learning being one of them, but basically anything else as well. 00:11:56.930 --> 00:11:58.670 align:middle line:84% I think it's a little bit of everything. 00:11:58.670 --> 00:12:00.590 align:middle line:84% And this question of sample efficiency, 00:12:00.590 --> 00:12:05.010 align:middle line:84% I would not say that we have good algorithms 00:12:05.010 --> 00:12:08.130 align:middle line:84% for sampling data efficiently or for learning 00:12:08.130 --> 00:12:09.990 align:middle line:84% from that sample data efficiently. 00:12:09.990 --> 00:12:11.950 align:middle line:84% I mean, we have algorithms that do it, 00:12:11.950 --> 00:12:16.130 align:middle line:84% but I feel like there's a lot of room to improve. 00:12:16.130 --> 00:12:17.210 align:middle line:90% Cool. 00:12:17.210 --> 00:12:19.050 align:middle line:84% So we're first going to talk a little bit 00:12:19.050 --> 00:12:22.202 align:middle line:84% about this idea of transferring knowledge from the mapping, 00:12:22.202 --> 00:12:24.410 align:middle line:84% so essentially, transferring knowledge from the model 00:12:24.410 --> 00:12:25.350 align:middle line:90% you've learned itself. 00:12:25.350 --> 00:12:28.930 align:middle line:84% How you go from inputs to useful outputs? 00:12:28.930 --> 00:12:32.630 align:middle line:84% So a simple example of this would be something like, OK. 00:12:32.630 --> 00:12:39.050 align:middle line:84% I have a lot of data on music with associated genres, 00:12:39.050 --> 00:12:44.450 align:middle line:84% and the idea being that being able to recognize a genre 00:12:44.450 --> 00:12:46.450 align:middle line:90% is potentially useful. 00:12:46.450 --> 00:12:49.090 align:middle line:84% It's teaching models something about the structure 00:12:49.090 --> 00:12:52.593 align:middle line:84% of music, something about what makes different songs similar. 00:12:52.593 --> 00:12:54.010 align:middle line:84% Two things being of the same genre 00:12:54.010 --> 00:12:56.530 align:middle line:84% argues that they might be similar in some way. 00:12:56.530 --> 00:13:00.090 align:middle line:84% And then maybe what you want to then adapt or transfer 00:13:00.090 --> 00:13:01.930 align:middle line:84% that knowledge to is something where 00:13:01.930 --> 00:13:05.050 align:middle line:84% you have a lot less data that specifically predicting, 00:13:05.050 --> 00:13:08.690 align:middle line:84% let's say, your individual preferences for songs. 00:13:08.690 --> 00:13:11.570 align:middle line:84% So maybe you can do genre recognition 00:13:11.570 --> 00:13:15.010 align:middle line:84% on the entire cadre of music on Spotify. 00:13:15.010 --> 00:13:18.570 align:middle line:84% And then if you actually want to learn 00:13:18.570 --> 00:13:21.590 align:middle line:84% how to predict what songs you're likely to enjoy, 00:13:21.590 --> 00:13:23.650 align:middle line:84% that you're going to prefer, you might only 00:13:23.650 --> 00:13:25.382 align:middle line:84% have a much smaller set of data, which 00:13:25.382 --> 00:13:27.090 align:middle line:84% is all the songs you've ever listened to, 00:13:27.090 --> 00:13:30.090 align:middle line:84% which is, of course, not even close to all the songs 00:13:30.090 --> 00:13:31.890 align:middle line:90% that exist. 00:13:31.890 --> 00:13:35.250 align:middle line:84% And then the actual testing would then be taking 00:13:35.250 --> 00:13:39.250 align:middle line:84% of fine-tuned model-- you can see they have this f prime-- 00:13:39.250 --> 00:13:41.373 align:middle line:84% and using it to actually predict on new music 00:13:41.373 --> 00:13:42.790 align:middle line:84% whether you're going to it or not. 00:13:42.790 --> 00:13:44.957 align:middle line:84% And then-- I don't know-- see, do you like the music 00:13:44.957 --> 00:13:47.490 align:middle line:90% that it's suggesting for you? 00:13:47.490 --> 00:13:51.930 align:middle line:84% And so this is kind of that broad, simple thing in terms 00:13:51.930 --> 00:13:54.850 align:middle line:90% of what we call fine-tuning. 00:13:54.850 --> 00:13:57.730 align:middle line:84% And at a very high level, it's that you're 00:13:57.730 --> 00:14:01.010 align:middle line:84% pretaining your network, your set of weights and biases, 00:14:01.010 --> 00:14:04.930 align:middle line:84% on some task that's a pretaining task. 00:14:04.930 --> 00:14:07.730 align:middle line:84% And that will result in some parameters for your model, W 00:14:07.730 --> 00:14:09.530 align:middle line:90% and b, weights and biases. 00:14:09.530 --> 00:14:11.530 align:middle line:84% Then you initialize a second network 00:14:11.530 --> 00:14:15.130 align:middle line:84% with some or all of W and b, and so there's 00:14:15.130 --> 00:14:18.090 align:middle line:84% these kind of interesting question 00:14:18.090 --> 00:14:22.770 align:middle line:84% of how much knowledge about the mapping you want to transfer. 00:14:22.770 --> 00:14:25.970 align:middle line:84% So usually, maybe the very final layer of a model, that's 00:14:25.970 --> 00:14:29.090 align:middle line:84% going to be pretty specific to the pretaining task. 00:14:29.090 --> 00:14:31.810 align:middle line:84% So often, it doesn't really make sense, especially 00:14:31.810 --> 00:14:34.730 align:middle line:84% if the output shape is not the same. 00:14:34.730 --> 00:14:38.970 align:middle line:84% You cut that off, you add some different final layer on, 00:14:38.970 --> 00:14:42.850 align:middle line:84% but maybe, you really only want those early layer features 00:14:42.850 --> 00:14:43.630 align:middle line:90% to be captured. 00:14:43.630 --> 00:14:45.790 align:middle line:84% And so you actually let the model-- 00:14:45.790 --> 00:14:51.490 align:middle line:84% you initialize it with less of the initial model. 00:14:51.490 --> 00:14:54.290 align:middle line:84% And then you think about training that second network 00:14:54.290 --> 00:14:55.835 align:middle line:90% on some task b. 00:14:55.835 --> 00:14:58.210 align:middle line:84% And so now, you'll get some slightly adjusted parameters, 00:14:58.210 --> 00:15:00.010 align:middle line:90% W prime and b prime. 00:15:00.010 --> 00:15:07.410 align:middle line:84% And the idea is then you have this learned representation that 00:15:07.410 --> 00:15:10.990 align:middle line:84% is essentially an encoder, and that's what we're transferring. 00:15:10.990 --> 00:15:13.570 align:middle line:84% So you're taking the representation learned 00:15:13.570 --> 00:15:16.690 align:middle line:84% from the initial model that's encoding something 00:15:16.690 --> 00:15:20.810 align:middle line:84% useful about your data, and then you're transferring that 00:15:20.810 --> 00:15:22.610 align:middle line:84% and then learning how to make use 00:15:22.610 --> 00:15:25.650 align:middle line:84% of that useful representation on something downstream. 00:15:25.650 --> 00:15:27.690 align:middle line:84% And we'll talk a little bit about some 00:15:27.690 --> 00:15:30.370 align:middle line:84% of these different empirical choices 00:15:30.370 --> 00:15:33.490 align:middle line:84% you can make here around how much you 00:15:33.490 --> 00:15:37.370 align:middle line:84% let the model adapt, how much you 00:15:37.370 --> 00:15:39.610 align:middle line:84% let it change from the initial representation. 00:15:39.610 --> 00:15:42.770 align:middle line:84% But empirically, actually, as long as the types of input data 00:15:42.770 --> 00:15:44.470 align:middle line:84% are similar-- so images and images, 00:15:44.470 --> 00:15:47.130 align:middle line:84% for example-- what we often find is 00:15:47.130 --> 00:15:50.155 align:middle line:84% that a lot of the structure in the models, 00:15:50.155 --> 00:15:51.530 align:middle line:84% like weights and biases, actually 00:15:51.530 --> 00:15:54.690 align:middle line:84% stays the same, particularly in those early layers 00:15:54.690 --> 00:15:56.410 align:middle line:84% after you've done the fine-tuning. 00:15:56.410 --> 00:15:58.890 align:middle line:84% So even if you haven't explicitly constrained the model 00:15:58.890 --> 00:16:01.490 align:middle line:84% not to change that very much, what you can do 00:16:01.490 --> 00:16:04.507 align:middle line:84% is keep a learning rate low or freeze some of the network. 00:16:04.507 --> 00:16:06.090 align:middle line:84% But if you don't do that, you actually 00:16:06.090 --> 00:16:08.050 align:middle line:84% find that a lot of the structure will 00:16:08.050 --> 00:16:10.810 align:middle line:84% stay the same, which does seem to argue that some of that 00:16:10.810 --> 00:16:12.750 align:middle line:90% is just universal in some sense. 00:16:12.750 --> 00:16:19.050 align:middle line:84% It's universally useful for, let's say, images or audio. 00:16:19.050 --> 00:16:22.633 align:middle line:84% So we previously saw a very specific case 00:16:22.633 --> 00:16:25.050 align:middle line:84% when we were talking about similarity-based representation 00:16:25.050 --> 00:16:28.670 align:middle line:84% learning where we talked about self-supervised pretaining. 00:16:28.670 --> 00:16:32.790 align:middle line:84% So here, with this self-supervised pretaining, 00:16:32.790 --> 00:16:36.680 align:middle line:84% we're not necessarily explicitly or directly training 00:16:36.680 --> 00:16:38.960 align:middle line:84% the model to do something we might care about. 00:16:38.960 --> 00:16:41.880 align:middle line:84% Instead, we're just training the model to learn something 00:16:41.880 --> 00:16:44.400 align:middle line:90% about data similarity. 00:16:44.400 --> 00:16:46.040 align:middle line:90% And then that ends up-- 00:16:46.040 --> 00:16:48.960 align:middle line:84% it has been shown-- to be quite useful for training 00:16:48.960 --> 00:16:51.440 align:middle line:84% representations that are good for downstream tasks we might 00:16:51.440 --> 00:16:52.052 align:middle line:90% care about. 00:16:52.052 --> 00:16:53.760 align:middle line:84% And one of the main benefits of this type 00:16:53.760 --> 00:16:56.040 align:middle line:84% of self-supervised pretaining is the fact 00:16:56.040 --> 00:17:00.200 align:middle line:84% that, because it doesn't require labels for data, 00:17:00.200 --> 00:17:04.579 align:middle line:84% it's making that pretaining on a lot of data-- like a lot, a lot, 00:17:04.579 --> 00:17:05.440 align:middle line:90% a lot of data. 00:17:05.440 --> 00:17:08.560 align:middle line:84% The amount of data that exists without any specific label 00:17:08.560 --> 00:17:13.040 align:middle line:84% is much larger than any of labeled or clean data set. 00:17:13.040 --> 00:17:16.523 align:middle line:84% So here, you think about contrastive pretaining 00:17:16.523 --> 00:17:18.440 align:middle line:84% and then maybe you're going to test it on some 00:17:18.440 --> 00:17:21.119 align:middle line:90% of new recognition task. 00:17:21.119 --> 00:17:25.325 align:middle line:84% But instead, you could think about fine-tuning a model, which 00:17:25.325 --> 00:17:27.200 align:middle line:84% is how you take those initial representations 00:17:27.200 --> 00:17:29.440 align:middle line:84% and transfer them to then actually learn 00:17:29.440 --> 00:17:32.280 align:middle line:84% how to do that new recognition task. 00:17:32.280 --> 00:17:32.815 align:middle line:90% OK. 00:17:32.815 --> 00:17:35.440 align:middle line:84% So now, I'm going to get into a couple pretty simple but pretty 00:17:35.440 --> 00:17:38.400 align:middle line:90% practical considerations. 00:17:38.400 --> 00:17:43.560 align:middle line:84% So what if you're pretaining on some images, 00:17:43.560 --> 00:17:46.360 align:middle line:84% and you're predicting whether they're wheat or corn? 00:17:46.360 --> 00:17:52.080 align:middle line:84% And then now, we want to fine tune our model. 00:17:52.080 --> 00:17:57.360 align:middle line:84% But now, we have an x prime and a y prime that are not the same. 00:17:57.360 --> 00:17:59.960 align:middle line:84% So essentially, we have a model, and our initial model 00:17:59.960 --> 00:18:02.760 align:middle line:84% was trained on images that are RGB. 00:18:02.760 --> 00:18:06.560 align:middle line:84% And it was trained to predict two categories as the output. 00:18:06.560 --> 00:18:08.720 align:middle line:84% And now, the model we actually want 00:18:08.720 --> 00:18:12.720 align:middle line:84% is going to actually take in images that are just grayscale. 00:18:12.720 --> 00:18:14.840 align:middle line:84% So we only have a single dimension, 00:18:14.840 --> 00:18:18.760 align:middle line:84% and we want to predict on outputs that 00:18:18.760 --> 00:18:21.160 align:middle line:90% have an additional category. 00:18:21.160 --> 00:18:23.440 align:middle line:84% So the structure of the model itself 00:18:23.440 --> 00:18:25.560 align:middle line:90% is actually a bit different. 00:18:25.560 --> 00:18:27.360 align:middle line:84% Essentially, if you're changing either 00:18:27.360 --> 00:18:30.920 align:middle line:84% your input dimension or your output dimension or both. 00:18:30.920 --> 00:18:36.480 align:middle line:84% So one kind of very simple mechanism that you can do 00:18:36.480 --> 00:18:38.800 align:middle line:84% is you can add kind of like a glue 00:18:38.800 --> 00:18:43.400 align:middle line:84% into the model, which is just a way 00:18:43.400 --> 00:18:47.200 align:middle line:84% to transfer data through to get it into the right format. 00:18:47.200 --> 00:18:48.980 align:middle line:84% So then when you're doing your pretaining, 00:18:48.980 --> 00:18:50.840 align:middle line:84% if you know that your downstream task is not 00:18:50.840 --> 00:18:54.040 align:middle line:84% going to have necessarily the same structure, 00:18:54.040 --> 00:18:58.400 align:middle line:84% you define these new z's, which are maybe models, 00:18:58.400 --> 00:19:02.180 align:middle line:84% or even just like a few simple layers, fully connected layers, 00:19:02.180 --> 00:19:05.280 align:middle line:84% what have you, that enable us to then 00:19:05.280 --> 00:19:09.880 align:middle line:84% map each input and output to the same dimensionality. 00:19:09.880 --> 00:19:12.120 align:middle line:84% And so you would do this where you need it, right? 00:19:12.120 --> 00:19:14.680 align:middle line:84% If you still are going to have RGB images, 00:19:14.680 --> 00:19:17.720 align:middle line:84% but then all you're doing is adding one more output 00:19:17.720 --> 00:19:20.360 align:middle line:84% dimension, well, then one thing you can do 00:19:20.360 --> 00:19:23.940 align:middle line:84% is you can take the final output layer of your initial model, 00:19:23.940 --> 00:19:26.080 align:middle line:84% and you can learn some reprojection to project it 00:19:26.080 --> 00:19:27.220 align:middle line:90% into more dimensions. 00:19:27.220 --> 00:19:29.080 align:middle line:84% Or you can do what I talked about sooner, 00:19:29.080 --> 00:19:32.280 align:middle line:84% and you can consider this z to be 00:19:32.280 --> 00:19:35.120 align:middle line:84% like that final mapping from your representation, 00:19:35.120 --> 00:19:38.380 align:middle line:84% that encoder, to the number of categories you want to predict. 00:19:38.380 --> 00:19:41.000 align:middle line:84% And then you can just swap that out and maybe learn 00:19:41.000 --> 00:19:44.740 align:middle line:84% some different one that's now in this new model. 00:19:44.740 --> 00:19:46.452 align:middle line:84% So you're just essentially defining 00:19:46.452 --> 00:19:48.160 align:middle line:84% points within the structure of your model 00:19:48.160 --> 00:19:50.040 align:middle line:84% that give you a bit of flexibility. 00:19:50.040 --> 00:19:52.900 align:middle line:84% So now, this f1 goes to f1 prime. 00:19:52.900 --> 00:19:54.600 align:middle line:84% Those will be different from each other. 00:19:54.600 --> 00:19:55.535 align:middle line:90% f3 goes to f3 prime. 00:19:55.535 --> 00:19:57.160 align:middle line:84% Those will be different from each other 00:19:57.160 --> 00:19:59.285 align:middle line:84% in terms of their structure, like the actual number 00:19:59.285 --> 00:20:01.720 align:middle line:90% of layers, weights, or size. 00:20:01.720 --> 00:20:03.680 align:middle line:84% But then the f2 can stay the same. 00:20:03.680 --> 00:20:05.500 align:middle line:84% So the vast majority of the structure, 00:20:05.500 --> 00:20:08.480 align:middle line:90% you're not actually changing. 00:20:08.480 --> 00:20:11.160 align:middle line:84% And this actually comes up quite a lot, 00:20:11.160 --> 00:20:14.360 align:middle line:84% particularly for some types of data that have maybe more input 00:20:14.360 --> 00:20:15.180 align:middle line:90% dimensions. 00:20:15.180 --> 00:20:17.840 align:middle line:84% So most image models are pretained 00:20:17.840 --> 00:20:21.240 align:middle line:84% on things like ImageNet, or these days, LAION, 00:20:21.240 --> 00:20:23.830 align:middle line:84% where you have huge amounts of RGB data. 00:20:23.830 --> 00:20:26.080 align:middle line:84% But then say you're someone who's working in robotics, 00:20:26.080 --> 00:20:30.950 align:middle line:84% and you have RGB-D. You have this additional depth dimension. 00:20:30.950 --> 00:20:34.830 align:middle line:84% It's often true that you can still get a lot of value 00:20:34.830 --> 00:20:39.270 align:middle line:84% in doing transfer learning from models trained without depth, 00:20:39.270 --> 00:20:42.550 align:middle line:84% but now, you have to figure out how to add this additional depth 00:20:42.550 --> 00:20:44.935 align:middle line:84% dimension during your fine-tuning effectively. 00:20:44.935 --> 00:20:46.310 align:middle line:84% And so often, the way people will 00:20:46.310 --> 00:20:49.710 align:middle line:84% do that is they just learn some of additional encoding 00:20:49.710 --> 00:20:52.450 align:middle line:84% that takes that and maps it into three dimensions. 00:20:52.450 --> 00:20:54.450 align:middle line:84% Or sometimes, depending on how much data 00:20:54.450 --> 00:20:56.950 align:middle line:84% you have, whether you can really learn that type of mapping, 00:20:56.950 --> 00:21:02.310 align:middle line:84% you do some of strange, just very simple model 00:21:02.310 --> 00:21:05.090 align:middle line:84% of how to transform one input dimension into another. 00:21:05.090 --> 00:21:07.750 align:middle line:84% So if you're working with very high scale hyperspectral 00:21:07.750 --> 00:21:09.630 align:middle line:84% satellite data, and you have models 00:21:09.630 --> 00:21:11.890 align:middle line:84% that are pretained with three channel inputs, 00:21:11.890 --> 00:21:14.990 align:middle line:84% sometimes, people will just run something like PCA 00:21:14.990 --> 00:21:19.590 align:middle line:84% and take the principal components, the three most 00:21:19.590 --> 00:21:21.190 align:middle line:84% informative principal components, 00:21:21.190 --> 00:21:23.282 align:middle line:84% over all those hyperspectral bands, 00:21:23.282 --> 00:21:24.990 align:middle line:84% and use that as their three channel input 00:21:24.990 --> 00:21:29.870 align:middle line:84% when they're fine-tuning because the ability to actually learn 00:21:29.870 --> 00:21:33.350 align:middle line:84% all of those 380 different input layers effectively is, 00:21:33.350 --> 00:21:36.485 align:middle line:84% one, computationally very expensive, and two, 00:21:36.485 --> 00:21:38.110 align:middle line:84% if you don't have a lot of data, that's 00:21:38.110 --> 00:21:39.610 align:middle line:90% a lot of parameters to learn. 00:21:39.610 --> 00:21:42.030 align:middle line:90% And so that can be tricky. 00:21:42.030 --> 00:21:42.830 align:middle line:90% Yeah. 00:21:42.830 --> 00:21:45.550 align:middle line:84% AUDIENCE: So we're talking about adding or changing 00:21:45.550 --> 00:21:48.270 align:middle line:90% the stacks and the [INAUDIBLE]. 00:21:48.270 --> 00:21:51.550 align:middle line:84% Do you ever add stuff into this dimension for fine-tuning, 00:21:51.550 --> 00:21:53.227 align:middle line:90% or is that just-- 00:21:53.227 --> 00:21:54.310 align:middle line:90% SARA BEERY: In the width-- 00:21:54.310 --> 00:21:56.310 align:middle line:90% so the input size would change? 00:21:56.310 --> 00:21:59.710 align:middle line:84% AUDIENCE: Basically, say if you had f2, 00:21:59.710 --> 00:22:05.070 align:middle line:84% adding additional channels to be parallel to f2. 00:22:05.070 --> 00:22:06.550 align:middle line:90% SARA BEERY: Like concatenating? 00:22:06.550 --> 00:22:07.402 align:middle line:90% AUDIENCE: Kind of. 00:22:07.402 --> 00:22:08.110 align:middle line:90% SARA BEERY: Yeah. 00:22:08.110 --> 00:22:10.990 align:middle line:84% I mean, so definitely, a lot of these models 00:22:10.990 --> 00:22:13.230 align:middle line:84% will assume some of fixed input size. 00:22:13.230 --> 00:22:15.750 align:middle line:84% And then, say, maybe you want to train the model on a bigger 00:22:15.750 --> 00:22:16.408 align:middle line:90% input size. 00:22:16.408 --> 00:22:17.950 align:middle line:84% That's another way that you can think 00:22:17.950 --> 00:22:19.990 align:middle line:84% about building this little f1 that 00:22:19.990 --> 00:22:22.070 align:middle line:84% will take a larger input size image 00:22:22.070 --> 00:22:24.710 align:middle line:84% and then learn how to get it to that smaller input size 00:22:24.710 --> 00:22:28.350 align:middle line:84% that the pretained model was trained on. 00:22:28.350 --> 00:22:29.833 align:middle line:84% But I think the only way to think 00:22:29.833 --> 00:22:31.750 align:middle line:84% about changing the width dimension is in terms 00:22:31.750 --> 00:22:33.292 align:middle line:90% of the spatial extent, right? 00:22:33.292 --> 00:22:35.250 align:middle line:84% AUDIENCE: So just in terms of input and output, 00:22:35.250 --> 00:22:39.127 align:middle line:84% not in terms of the throughput of the pretained models. 00:22:39.127 --> 00:22:40.710 align:middle line:84% SARA BEERY: Oh, you mean like actually 00:22:40.710 --> 00:22:42.850 align:middle line:84% changing the structure of the model itself? 00:22:42.850 --> 00:22:45.282 align:middle line:84% AUDIENCE: I guess like stacking on top of [INAUDIBLE]. 00:22:45.282 --> 00:22:45.990 align:middle line:90% SARA BEERY: Yeah. 00:22:45.990 --> 00:22:48.490 align:middle line:84% I mean, so one of the things we're trying to get across here 00:22:48.490 --> 00:22:51.270 align:middle line:84% is that you can treat these things as modular building 00:22:51.270 --> 00:22:52.490 align:middle line:90% blocks, right? 00:22:52.490 --> 00:22:55.510 align:middle line:84% And so you can take a building block that someone else has 00:22:55.510 --> 00:22:58.250 align:middle line:90% built and then stick it in. 00:22:58.250 --> 00:22:59.830 align:middle line:84% And then you have to figure out what 00:22:59.830 --> 00:23:02.150 align:middle line:84% blocks you need to add to the system 00:23:02.150 --> 00:23:04.990 align:middle line:84% to maybe get that building block to be useful for you. 00:23:04.990 --> 00:23:08.430 align:middle line:84% But the point being that a really well-trained building 00:23:08.430 --> 00:23:11.310 align:middle line:84% block-- if that analogy still is making sense. 00:23:11.310 --> 00:23:13.850 align:middle line:84% A really well-trained building block is still useful, 00:23:13.850 --> 00:23:16.950 align:middle line:84% and it's maybe worth coming up with this modular structure that 00:23:16.950 --> 00:23:20.750 align:middle line:84% will make that work for your problem. 00:23:20.750 --> 00:23:22.965 align:middle line:84% But as soon as you start actually trying to mess 00:23:22.965 --> 00:23:24.590 align:middle line:84% with that building block, like changing 00:23:24.590 --> 00:23:26.330 align:middle line:84% the weights and the biases within that block, 00:23:26.330 --> 00:23:27.250 align:middle line:90% it gets more complicated. 00:23:27.250 --> 00:23:28.850 align:middle line:84% There are people who look at doing it, 00:23:28.850 --> 00:23:31.303 align:middle line:84% and actually, the entire area of model compression 00:23:31.303 --> 00:23:33.470 align:middle line:84% is kind of changing that structure by removing stuff 00:23:33.470 --> 00:23:34.490 align:middle line:90% to make it more efficient. 00:23:34.490 --> 00:23:36.365 align:middle line:84% And we'll talk a little bit about that later. 00:23:36.365 --> 00:23:39.150 align:middle line:84% In terms of adding more stuff, I don't know. 00:23:39.150 --> 00:23:41.290 align:middle line:84% I haven't seen a lot of work in that space, 00:23:41.290 --> 00:23:43.415 align:middle line:84% but I think as soon as you start adding more stuff, 00:23:43.415 --> 00:23:45.030 align:middle line:84% the value of the pretaining starts 00:23:45.030 --> 00:23:47.990 align:middle line:84% to go away, if that makes sense, because then you've 00:23:47.990 --> 00:23:51.910 align:middle line:84% changed the structure of the mathematical formulas. 00:23:51.910 --> 00:23:53.950 align:middle line:90% Cool. 00:23:53.950 --> 00:23:57.870 align:middle line:84% So essentially, we're just trying to figure out 00:23:57.870 --> 00:23:59.730 align:middle line:90% how to make things useful. 00:23:59.730 --> 00:24:05.270 align:middle line:84% And so in practice, we'll essentially just learn mappings 00:24:05.270 --> 00:24:09.495 align:middle line:84% for any new stuff we add to make this work for our data. 00:24:09.495 --> 00:24:11.870 align:middle line:84% And so those new things would need to be then initialized 00:24:11.870 --> 00:24:14.790 align:middle line:84% from scratch, but that's a lot less parameters 00:24:14.790 --> 00:24:17.790 align:middle line:84% that need to be learned from scratch than if you're going end 00:24:17.790 --> 00:24:20.950 align:middle line:84% to end from scratch, starting from that blank slate. 00:24:20.950 --> 00:24:23.730 align:middle line:84% And if you really don't have much data, 00:24:23.730 --> 00:24:26.650 align:middle line:84% you can also think about copying weights for that initialization. 00:24:26.650 --> 00:24:29.750 align:middle line:84% So for example, in that RGB-D case, 00:24:29.750 --> 00:24:31.590 align:middle line:84% sometimes what people will do is they 00:24:31.590 --> 00:24:33.950 align:middle line:84% will take the RGB channels that were learned 00:24:33.950 --> 00:24:35.550 align:middle line:84% from the original model, and then 00:24:35.550 --> 00:24:37.390 align:middle line:84% they'll staple on an additional channel, 00:24:37.390 --> 00:24:41.950 align:middle line:84% and they'll initialize it with, for example, just 00:24:41.950 --> 00:24:46.070 align:middle line:84% the blue channel parameters, with the idea that maybe there's 00:24:46.070 --> 00:24:49.350 align:middle line:84% some value in what's been learned for another channel that 00:24:49.350 --> 00:24:51.110 align:middle line:90% might still be transferable. 00:24:51.110 --> 00:24:56.690 align:middle line:84% Again, this is all kind of-- no one really knows. 00:24:56.690 --> 00:25:01.710 align:middle line:84% No has a perfect set of heuristics 00:25:01.710 --> 00:25:05.150 align:middle line:84% for, if this is your data, this is the best thing to do. 00:25:05.150 --> 00:25:06.550 align:middle line:84% And this is really where the idea 00:25:06.550 --> 00:25:08.590 align:middle line:84% of designing a benchmark for your data 00:25:08.590 --> 00:25:11.150 align:middle line:84% set where you understand how to do comparison 00:25:11.150 --> 00:25:13.130 align:middle line:84% across some of these different choices 00:25:13.130 --> 00:25:15.787 align:middle line:84% so you can try things and see how they're affecting 00:25:15.787 --> 00:25:17.870 align:middle line:84% performance of your model-- that's where it starts 00:25:17.870 --> 00:25:19.250 align:middle line:90% to become really important. 00:25:19.250 --> 00:25:23.520 align:middle line:84% Because most of the time, if it's some custom problem, 00:25:23.520 --> 00:25:25.020 align:middle line:84% some little dimension of the system 00:25:25.020 --> 00:25:27.160 align:middle line:84% will probably be somewhat customizable. 00:25:27.160 --> 00:25:31.820 align:middle line:84% And it's hard to say what's the best structure for a system 00:25:31.820 --> 00:25:35.700 align:middle line:84% without actually doing some testing. 00:25:35.700 --> 00:25:39.540 align:middle line:84% So one of the issues is that if you only 00:25:39.540 --> 00:25:44.420 align:middle line:84% have a little bit of data, and you let the entire model train, 00:25:44.420 --> 00:25:49.180 align:middle line:84% and you train it until that loss is all the way saturated-- 00:25:49.180 --> 00:25:51.580 align:middle line:84% we don't have anything else we can learn-- 00:25:51.580 --> 00:25:54.220 align:middle line:84% we run into that catastrophic forgetting problem 00:25:54.220 --> 00:25:56.580 align:middle line:84% again, that thing we talked about a bit 00:25:56.580 --> 00:25:58.582 align:middle line:84% when we were talking about sequences, where it's 00:25:58.582 --> 00:26:01.040 align:middle line:84% like, somehow, the thing we learned from the original data, 00:26:01.040 --> 00:26:02.480 align:middle line:90% we learned too long ago. 00:26:02.480 --> 00:26:04.900 align:middle line:84% The model's weights get changed too much, 00:26:04.900 --> 00:26:08.620 align:middle line:84% and you essentially lose capacity. 00:26:08.620 --> 00:26:12.660 align:middle line:84% So this definitely happens more with smaller models 00:26:12.660 --> 00:26:15.540 align:middle line:90% than big models, interestingly. 00:26:15.540 --> 00:26:18.500 align:middle line:84% But in order to avoid kind of overfitting 00:26:18.500 --> 00:26:21.660 align:middle line:84% on a really small amount of data for your specific task, 00:26:21.660 --> 00:26:24.500 align:middle line:84% there are a few different mechanisms 00:26:24.500 --> 00:26:28.060 align:middle line:84% that people use to try to strike that right balance, learning 00:26:28.060 --> 00:26:30.620 align:middle line:84% enough about the new task that you can specialize while not 00:26:30.620 --> 00:26:34.980 align:middle line:84% forgetting too much about this general useful representation. 00:26:34.980 --> 00:26:36.620 align:middle line:90% And so a couple of things. 00:26:36.620 --> 00:26:39.720 align:middle line:84% You can think about freezing most of the network. 00:26:39.720 --> 00:26:42.260 align:middle line:84% So basically saying, I just want to keep 00:26:42.260 --> 00:26:44.460 align:middle line:84% most of this model exactly the same, 00:26:44.460 --> 00:26:46.838 align:middle line:84% and only training, for example, the final layers or so, 00:26:46.838 --> 00:26:48.380 align:middle line:84% those initial things at the beginning 00:26:48.380 --> 00:26:50.700 align:middle line:84% that you need to have that are new. 00:26:50.700 --> 00:26:53.040 align:middle line:84% You can explicitly use really small learning rates. 00:26:53.040 --> 00:26:54.707 align:middle line:84% So basically, just not letting the model 00:26:54.707 --> 00:26:57.860 align:middle line:84% move too much in any direction at any training step. 00:26:57.860 --> 00:27:00.020 align:middle line:84% You can use what's called early stopping, which 00:27:00.020 --> 00:27:01.980 align:middle line:84% is-- essentially, you use a validation set, 00:27:01.980 --> 00:27:05.060 align:middle line:84% and you make a decision about which model weights 00:27:05.060 --> 00:27:07.980 align:middle line:84% to keep based on the best performance on your validation 00:27:07.980 --> 00:27:10.980 align:middle line:84% set with the idea that, essentially, you just haven't 00:27:10.980 --> 00:27:14.060 align:middle line:84% let the model train too much, so you stopped it 00:27:14.060 --> 00:27:17.780 align:middle line:84% before it can forget too much about the original pretext task. 00:27:17.780 --> 00:27:20.920 align:middle line:84% Or another mechanism-- and one that tends to work pretty well, 00:27:20.920 --> 00:27:23.503 align:middle line:84% though, of course, there's some complications in terms of data 00:27:23.503 --> 00:27:26.380 align:middle line:84% balance that you have to strike, so sampling effectively-- 00:27:26.380 --> 00:27:30.140 align:middle line:84% is to just continue to train on the original data as well. 00:27:30.140 --> 00:27:33.460 align:middle line:84% So essentially, if you have access to the original training 00:27:33.460 --> 00:27:36.420 align:middle line:84% data, then as you're fine-tuning, continuing 00:27:36.420 --> 00:27:39.140 align:middle line:84% to add some of that pretext or some of that original training 00:27:39.140 --> 00:27:40.220 align:middle line:90% data is good. 00:27:40.220 --> 00:27:41.960 align:middle line:84% But of course, there's a complexity here, 00:27:41.960 --> 00:27:45.060 align:middle line:84% which is, all right, but if you have different input 00:27:45.060 --> 00:27:48.460 align:middle line:84% types and different goals, this becomes harder and harder to do. 00:27:48.460 --> 00:27:50.508 align:middle line:84% If you were doing contrastive pretaining, 00:27:50.508 --> 00:27:52.300 align:middle line:84% and now, you're trying to do classification 00:27:52.300 --> 00:27:54.480 align:middle line:84% on top of those representations, still 00:27:54.480 --> 00:27:56.980 align:middle line:84% continuing to have the contrastive pretaining is 00:27:56.980 --> 00:27:59.980 align:middle line:90% a little bit more complicated. 00:27:59.980 --> 00:28:03.572 align:middle line:84% But a lot of these choices, like these types of things, 00:28:03.572 --> 00:28:05.780 align:middle line:84% assume that you have some sort of good representative 00:28:05.780 --> 00:28:07.860 align:middle line:84% validation set, that your benchmark is well 00:28:07.860 --> 00:28:09.300 align:middle line:90% posed for your task. 00:28:09.300 --> 00:28:11.020 align:middle line:84% This might not always be possible, 00:28:11.020 --> 00:28:14.540 align:middle line:84% where you might see distribution shift going forward. 00:28:14.540 --> 00:28:16.740 align:middle line:84% And so it really does become quite 00:28:16.740 --> 00:28:20.980 align:middle line:84% important to think about measuring relative performance 00:28:20.980 --> 00:28:24.620 align:middle line:84% as close as possible to your representative problem. 00:28:24.620 --> 00:28:27.860 align:middle line:84% If you want to actually use the model on some data, 00:28:27.860 --> 00:28:31.420 align:middle line:84% defining measuring progress is just 00:28:31.420 --> 00:28:34.020 align:middle line:84% difficult and complicated and important. 00:28:34.020 --> 00:28:37.462 align:middle line:84% And this is-- now, I'm going to be on a soapbox a little bit. 00:28:37.462 --> 00:28:39.420 align:middle line:84% But one of the things that drives me absolutely 00:28:39.420 --> 00:28:41.060 align:middle line:84% crazy about modern machine learning 00:28:41.060 --> 00:28:43.340 align:middle line:84% is that we're obsessed with the idea of bolded numbers 00:28:43.340 --> 00:28:44.620 align:middle line:90% and tables. 00:28:44.620 --> 00:28:46.780 align:middle line:84% And so this kind of begs the question 00:28:46.780 --> 00:28:48.620 align:middle line:84% that you can represent everything 00:28:48.620 --> 00:28:52.220 align:middle line:84% about what a model does with one single number that 00:28:52.220 --> 00:28:55.540 align:middle line:84% captures every dimension of importance to that model. 00:28:55.540 --> 00:28:57.740 align:middle line:90% And I don't think that's true. 00:28:57.740 --> 00:28:59.240 align:middle line:90% I think we need nuance. 00:28:59.240 --> 00:29:01.820 align:middle line:84% And honestly, usually, people will-- 00:29:01.820 --> 00:29:02.840 align:middle line:90% this is bad behavior. 00:29:02.840 --> 00:29:05.923 align:middle line:84% People will pick the metric that their model does best on. 00:29:05.923 --> 00:29:07.840 align:middle line:84% They're like, oh, this makes us look the best, 00:29:07.840 --> 00:29:10.080 align:middle line:84% so we're just going to show that one in our paper, 00:29:10.080 --> 00:29:11.830 align:middle line:84% and we're not going to talk about the fact 00:29:11.830 --> 00:29:14.740 align:middle line:84% that for some of these other metrics, we don't do as well. 00:29:14.740 --> 00:29:17.380 align:middle line:84% I really love it in a machine learning paper when people 00:29:17.380 --> 00:29:19.380 align:middle line:90% are honest about their flaws. 00:29:19.380 --> 00:29:21.680 align:middle line:84% Where you're like, hey, all right. 00:29:21.680 --> 00:29:24.320 align:middle line:84% We see some really great progress in these sections. 00:29:24.320 --> 00:29:26.580 align:middle line:84% And then you have a really nice limitations section 00:29:26.580 --> 00:29:30.100 align:middle line:84% that's realistic and captures the dimensions 00:29:30.100 --> 00:29:33.620 align:middle line:84% of where there still is space to make progress. 00:29:33.620 --> 00:29:35.120 align:middle line:84% Maybe it's just personal preference, 00:29:35.120 --> 00:29:36.920 align:middle line:84% but I'm a little bit allergic to salesmanship, 00:29:36.920 --> 00:29:38.545 align:middle line:84% so I don't like it when I read a paper, 00:29:38.545 --> 00:29:41.540 align:middle line:90% and I'm like, I don't know, man. 00:29:41.540 --> 00:29:43.935 align:middle line:84% I think you're sweeping some stuff under the rug. 00:29:43.935 --> 00:29:46.560 align:middle line:84% Essentially, you do not want me to be your reviewer number two. 00:29:46.560 --> 00:29:50.640 align:middle line:90% 00:29:50.640 --> 00:29:51.140 align:middle line:90% Yeah. 00:29:51.140 --> 00:29:53.060 align:middle line:90% OK. 00:29:53.060 --> 00:29:55.560 align:middle line:84% So we'd like to avoid forgetting the general representation, 00:29:55.560 --> 00:29:57.020 align:middle line:84% and here's some engineering hacks 00:29:57.020 --> 00:29:59.300 align:middle line:84% that you can take if you want to try and do that. 00:29:59.300 --> 00:30:02.660 align:middle line:84% But then there's also this other kind of area of research 00:30:02.660 --> 00:30:05.480 align:middle line:84% that's specifically-- it's called domain adaptation. 00:30:05.480 --> 00:30:08.300 align:middle line:84% And the idea here is that maybe you have some source domain, 00:30:08.300 --> 00:30:10.340 align:middle line:84% and you have a target domain, and then 00:30:10.340 --> 00:30:13.430 align:middle line:84% you want to learn to do well on your target domain, 00:30:13.430 --> 00:30:15.690 align:middle line:84% but you also want to essentially make sure 00:30:15.690 --> 00:30:19.290 align:middle line:84% that the representation spaces of the source and target domain 00:30:19.290 --> 00:30:20.970 align:middle line:90% are well aligned. 00:30:20.970 --> 00:30:22.810 align:middle line:84% And this shows up a lot when you have 00:30:22.810 --> 00:30:24.810 align:middle line:84% these types of distribution shift problems 00:30:24.810 --> 00:30:26.490 align:middle line:90% that we talked about before. 00:30:26.490 --> 00:30:29.832 align:middle line:84% But it also can be used for different types of fine-tuning. 00:30:29.832 --> 00:30:31.290 align:middle line:84% Essentially, one of the motivations 00:30:31.290 --> 00:30:35.090 align:middle line:84% here is that maybe if you force these feature spaces to be 00:30:35.090 --> 00:30:42.330 align:middle line:84% well aligned generally, that you can learn usefully how to map-- 00:30:42.330 --> 00:30:44.470 align:middle line:84% learn that mapping with much less data. 00:30:44.470 --> 00:30:47.490 align:middle line:84% And one way that people do this is they actually-- 00:30:47.490 --> 00:30:52.210 align:middle line:84% they learn how to get the model from the source domain 00:30:52.210 --> 00:30:54.730 align:middle line:84% to the outputs, and then they try 00:30:54.730 --> 00:30:57.890 align:middle line:84% to learn a projection or a mapping from the target domain 00:30:57.890 --> 00:30:58.950 align:middle line:90% to the source domain. 00:30:58.950 --> 00:31:01.850 align:middle line:84% So you basically are learning how to map into the domain space 00:31:01.850 --> 00:31:05.370 align:middle line:84% where you had a lot of data, and then you can use the same model 00:31:05.370 --> 00:31:07.513 align:middle line:90% to actually do your prediction. 00:31:07.513 --> 00:31:08.930 align:middle line:84% And there's lots of different ways 00:31:08.930 --> 00:31:11.090 align:middle line:84% to try to force these representation spaces to be 00:31:11.090 --> 00:31:15.250 align:middle line:84% aligned, but one kind of large set of them 00:31:15.250 --> 00:31:17.170 align:middle line:90% takes an adversarial approach. 00:31:17.170 --> 00:31:21.610 align:middle line:84% So here, you have both source and target data during training. 00:31:21.610 --> 00:31:26.170 align:middle line:84% And while you're training for every input, 00:31:26.170 --> 00:31:29.090 align:middle line:84% you want to predict the category you care about, 00:31:29.090 --> 00:31:31.370 align:middle line:84% and you also want to predict the domain 00:31:31.370 --> 00:31:34.330 align:middle line:84% that it came from, something about the distribution. 00:31:34.330 --> 00:31:38.530 align:middle line:84% So a classic example of a data set in this domain adaptation 00:31:38.530 --> 00:31:44.090 align:middle line:84% space is the data set of different types 00:31:44.090 --> 00:31:46.330 align:middle line:90% of technological hardware. 00:31:46.330 --> 00:31:48.990 align:middle line:84% But one data set is from Amazon, and one is from eBay. 00:31:48.990 --> 00:31:51.330 align:middle line:84% So you can think that, OK, the way that people 00:31:51.330 --> 00:31:53.763 align:middle line:84% take pictures for Amazon are very nice and clean, 00:31:53.763 --> 00:31:55.930 align:middle line:84% and then the way people are taking pictures for eBay 00:31:55.930 --> 00:31:58.267 align:middle line:84% is maybe like on their messy desk. 00:31:58.267 --> 00:31:59.850 align:middle line:84% And so there, what you're trying to do 00:31:59.850 --> 00:32:02.390 align:middle line:84% is you're trying to predict, what kind of computer is it? 00:32:02.390 --> 00:32:03.878 align:middle line:90% And is it from Amazon or eBay? 00:32:03.878 --> 00:32:05.670 align:middle line:84% And then this is where it gets adversarial. 00:32:05.670 --> 00:32:08.570 align:middle line:84% What they do is they flip the gradient 00:32:08.570 --> 00:32:10.890 align:middle line:90% from the domain predictor. 00:32:10.890 --> 00:32:12.450 align:middle line:84% So you're essentially trying to train 00:32:12.450 --> 00:32:14.770 align:middle line:84% the model to be as good as possible 00:32:14.770 --> 00:32:18.530 align:middle line:84% at predicting the category, while being as bad as possible 00:32:18.530 --> 00:32:21.370 align:middle line:84% at predicting the domain, which tries to force, essentially, 00:32:21.370 --> 00:32:25.515 align:middle line:84% the representations across the two domains to align. 00:32:25.515 --> 00:32:26.890 align:middle line:84% And then the interesting thing is 00:32:26.890 --> 00:32:30.250 align:middle line:84% you can extend this to where you might not actually 00:32:30.250 --> 00:32:35.070 align:middle line:84% have category data for the source and the target. 00:32:35.070 --> 00:32:38.130 align:middle line:84% So maybe you had a bunch of category data for the source, 00:32:38.130 --> 00:32:40.050 align:middle line:84% and now, all you have for the target domain 00:32:40.050 --> 00:32:42.250 align:middle line:84% is just the knowledge of what domain it came from. 00:32:42.250 --> 00:32:44.490 align:middle line:84% You can think about trying to still 00:32:44.490 --> 00:32:47.130 align:middle line:84% keep really good at predicting the category, 00:32:47.130 --> 00:32:50.050 align:middle line:84% but be really bad at predicting which domain it's from. 00:32:50.050 --> 00:32:52.450 align:middle line:84% And the hope is that there, even if you only 00:32:52.450 --> 00:32:55.710 align:middle line:84% have a very limited amount of labeled data in your target, 00:32:55.710 --> 00:32:57.690 align:middle line:84% or even sometimes none, you might still 00:32:57.690 --> 00:32:59.130 align:middle line:84% be able to learn something useful 00:32:59.130 --> 00:33:01.490 align:middle line:84% that helps you find the right categories in that target 00:33:01.490 --> 00:33:03.610 align:middle line:90% domain. 00:33:03.610 --> 00:33:05.570 align:middle line:84% And then this is kind of expanded 00:33:05.570 --> 00:33:08.410 align:middle line:84% into this whole area of unsupervised domain adaptation, 00:33:08.410 --> 00:33:12.010 align:middle line:84% where here the idea is you might have data, but no labels 00:33:12.010 --> 00:33:13.970 align:middle line:90% from the target domain. 00:33:13.970 --> 00:33:16.250 align:middle line:84% So now, it's like you want to make a model that's 00:33:16.250 --> 00:33:19.210 align:middle line:84% as good as possible on some set of data, 00:33:19.210 --> 00:33:22.370 align:middle line:84% but you don't necessarily have any training labels. 00:33:22.370 --> 00:33:23.750 align:middle line:90% Now, why would we have that? 00:33:23.750 --> 00:33:28.050 align:middle line:84% So here's an example, actually from our own research, 00:33:28.050 --> 00:33:29.948 align:middle line:84% where we have a bunch of labeled source data. 00:33:29.948 --> 00:33:31.490 align:middle line:84% In this case, the labeled source data 00:33:31.490 --> 00:33:36.178 align:middle line:84% is sonar images or sonar video from rivers with fish. 00:33:36.178 --> 00:33:37.970 align:middle line:84% And we're trying to count all the fish that 00:33:37.970 --> 00:33:40.512 align:middle line:84% are swimming upstream so we can make good sustainable fishing 00:33:40.512 --> 00:33:41.650 align:middle line:90% policies. 00:33:41.650 --> 00:33:46.490 align:middle line:84% And we have some rivers where we have a bunch of labeled data, 00:33:46.490 --> 00:33:49.092 align:middle line:84% but we want this to work on any river. 00:33:49.092 --> 00:33:51.050 align:middle line:84% Or if something changes about the river or they 00:33:51.050 --> 00:33:52.800 align:middle line:84% move the sensor, we want it to still work, 00:33:52.800 --> 00:33:54.930 align:middle line:84% even though there's that distribution shift. 00:33:54.930 --> 00:33:56.650 align:middle line:84% And so here, what we're trying to do 00:33:56.650 --> 00:34:00.190 align:middle line:84% is learn from labeled source data and unlabeled target data 00:34:00.190 --> 00:34:02.890 align:middle line:84% how get a model to work better on the target. 00:34:02.890 --> 00:34:04.670 align:middle line:84% So how have people done this in the past? 00:34:04.670 --> 00:34:08.679 align:middle line:84% Well, so the very, very simple version of this would be, 00:34:08.679 --> 00:34:10.719 align:middle line:84% you take your target data set, you run it 00:34:10.719 --> 00:34:13.199 align:middle line:84% through a model that's been pretained on source, 00:34:13.199 --> 00:34:15.020 align:middle line:84% and then you maybe just calibrate it. 00:34:15.020 --> 00:34:16.770 align:middle line:84% You pick some threshold where you say, OK. 00:34:16.770 --> 00:34:20.679 align:middle line:84% Any prediction over this, we're going to treat as ground truth. 00:34:20.679 --> 00:34:25.040 align:middle line:84% So that's one way to transfer knowledge. 00:34:25.040 --> 00:34:27.880 align:middle line:84% You just use a model from a different domain 00:34:27.880 --> 00:34:29.000 align:middle line:90% and hope it works. 00:34:29.000 --> 00:34:31.920 align:middle line:84% It turns out it doesn't usually work very well. 00:34:31.920 --> 00:34:34.639 align:middle line:84% So this kind of motivated this idea-- and this 00:34:34.639 --> 00:34:37.239 align:middle line:84% is a mechanism to do distillation, which we'll talk 00:34:37.239 --> 00:34:39.480 align:middle line:90% about a little bit more later. 00:34:39.480 --> 00:34:43.880 align:middle line:84% But what if, instead, we make a new model, that we're 00:34:43.880 --> 00:34:45.639 align:middle line:84% going to call a student model, and we're 00:34:45.639 --> 00:34:47.800 align:middle line:84% going to use those predictions on the target 00:34:47.800 --> 00:34:50.315 align:middle line:90% data as weak supervision? 00:34:50.315 --> 00:34:52.440 align:middle line:84% So we'll use the predictions that our model trained 00:34:52.440 --> 00:34:55.120 align:middle line:84% on the source data gives, and then we'll 00:34:55.120 --> 00:34:59.120 align:middle line:84% try to train a model to predict the same thing. 00:34:59.120 --> 00:35:01.280 align:middle line:84% And the idea here is that maybe your student model 00:35:01.280 --> 00:35:03.680 align:middle line:84% is able to learned something useful 00:35:03.680 --> 00:35:07.628 align:middle line:84% that is better than just taking those raw predictions. 00:35:07.628 --> 00:35:08.420 align:middle line:90% And this does work. 00:35:08.420 --> 00:35:11.360 align:middle line:84% You get small gains, sometimes, over just the original source 00:35:11.360 --> 00:35:12.440 align:middle line:90% predictions. 00:35:12.440 --> 00:35:14.360 align:middle line:84% But this threshold is really, really 00:35:14.360 --> 00:35:17.400 align:middle line:84% finicky and hard to calibrate, right? 00:35:17.400 --> 00:35:21.280 align:middle line:84% So people started adding in other bells and whistles. 00:35:21.280 --> 00:35:24.860 align:middle line:84% One thing is you use EMA, an Exponential Moving Average, 00:35:24.860 --> 00:35:27.720 align:middle line:84% to try to regularize the way that the weights change 00:35:27.720 --> 00:35:29.640 align:middle line:84% within the student and teacher model, 00:35:29.640 --> 00:35:34.080 align:middle line:84% so you let the teacher model itself also update slower. 00:35:34.080 --> 00:35:37.240 align:middle line:84% And then, finally, you can also introduce additional training 00:35:37.240 --> 00:35:38.600 align:middle line:90% on the source data. 00:35:38.600 --> 00:35:41.280 align:middle line:84% So there, maybe you're letting your student model 00:35:41.280 --> 00:35:44.460 align:middle line:84% learn from the source data and from the target data. 00:35:44.460 --> 00:35:46.740 align:middle line:84% And then you can add in a bunch of fancy stuff, 00:35:46.740 --> 00:35:49.840 align:middle line:84% like instead of explicitly taking 00:35:49.840 --> 00:35:51.300 align:middle line:84% the predictions with a threshold, 00:35:51.300 --> 00:35:52.740 align:middle line:90% you can use a softmax. 00:35:52.740 --> 00:35:55.040 align:middle line:84% You can use a distillation loss so that you're 00:35:55.040 --> 00:35:57.340 align:middle line:84% learning to, instead of match the prediction, 00:35:57.340 --> 00:36:00.080 align:middle line:84% match the distribution over the outputs. 00:36:00.080 --> 00:36:02.400 align:middle line:84% And you can try to use some explicit feature 00:36:02.400 --> 00:36:04.180 align:middle line:84% alignment between source and target, 00:36:04.180 --> 00:36:06.013 align:middle line:84% so that's like basically trying to make sure 00:36:06.013 --> 00:36:08.720 align:middle line:84% that the features are overlapping as much as possible. 00:36:08.720 --> 00:36:10.720 align:middle line:90% Lots of fancy tricks here. 00:36:10.720 --> 00:36:13.400 align:middle line:90% Kind of complicated. 00:36:13.400 --> 00:36:15.760 align:middle line:84% We essentially wanted to point out 00:36:15.760 --> 00:36:18.520 align:middle line:84% that these components are essentially 00:36:18.520 --> 00:36:20.040 align:middle line:84% the main components that people use 00:36:20.040 --> 00:36:25.720 align:middle line:84% to do this type of distribution, unsupervised domain adaptation. 00:36:25.720 --> 00:36:29.120 align:middle line:84% And the point here is that basically every modern domain 00:36:29.120 --> 00:36:32.360 align:middle line:84% adaptation architecture is some combination of these components. 00:36:32.360 --> 00:36:36.840 align:middle line:84% And so if you can centralize and unify the ways 00:36:36.840 --> 00:36:40.520 align:middle line:84% that people are actually doing these experiments, what we found 00:36:40.520 --> 00:36:43.080 align:middle line:84% is that you can often be more optimal 00:36:43.080 --> 00:36:45.720 align:middle line:90% than any previous approach. 00:36:45.720 --> 00:36:50.040 align:middle line:84% And so these methods, many of them 00:36:50.040 --> 00:36:52.220 align:middle line:84% can be quite effective in certain scenarios. 00:36:52.220 --> 00:36:55.980 align:middle line:84% Of course, certain scenarios is a big interesting thing there. 00:36:55.980 --> 00:36:58.440 align:middle line:84% They're almost never as effective as having labeled 00:36:58.440 --> 00:36:59.860 align:middle line:90% data for your source domain. 00:36:59.860 --> 00:37:02.440 align:middle line:84% So if you just have labels, that's better. 00:37:02.440 --> 00:37:04.080 align:middle line:84% But if you don't have labels, you 00:37:04.080 --> 00:37:06.400 align:middle line:84% can actually start to get pretty close to performance 00:37:06.400 --> 00:37:10.040 align:middle line:84% if you do have labeled data if you 00:37:10.040 --> 00:37:12.460 align:middle line:90% build these things carefully. 00:37:12.460 --> 00:37:15.020 align:middle line:84% And some intuition as to why that way might be 00:37:15.020 --> 00:37:17.263 align:middle line:84% is that, essentially, the model is 00:37:17.263 --> 00:37:19.680 align:middle line:84% able to learn something about the statistics of the target 00:37:19.680 --> 00:37:21.920 align:middle line:90% domain that's useful. 00:37:21.920 --> 00:37:24.200 align:middle line:84% You're shifting the weights towards being 00:37:24.200 --> 00:37:26.520 align:middle line:84% representative on those new statistics. 00:37:26.520 --> 00:37:27.520 align:middle line:90% Yeah? 00:37:27.520 --> 00:37:31.160 align:middle line:84% AUDIENCE: How can you measure how the performance of this 00:37:31.160 --> 00:37:32.360 align:middle line:90% when deployed-- 00:37:32.360 --> 00:37:35.200 align:middle line:84% if you deploy it in a field that you actually 00:37:35.200 --> 00:37:38.000 align:middle line:84% don't have the labels of anything, 00:37:38.000 --> 00:37:39.872 align:middle line:84% how do you know your models did well? 00:37:39.872 --> 00:37:40.580 align:middle line:90% SARA BEERY: Yeah. 00:37:40.580 --> 00:37:41.883 align:middle line:90% This is a really great point. 00:37:41.883 --> 00:37:43.800 align:middle line:84% I think it's actually the elephant in the room 00:37:43.800 --> 00:37:46.200 align:middle line:84% for most of the domain adaptation literature 00:37:46.200 --> 00:37:48.760 align:middle line:84% is that they almost always are demonstrating performance 00:37:48.760 --> 00:37:51.080 align:middle line:84% in their bold number in their table 00:37:51.080 --> 00:37:54.640 align:middle line:84% on labeled validation data for the target domain. 00:37:54.640 --> 00:37:56.340 align:middle line:84% But this kind of begs the question, 00:37:56.340 --> 00:37:59.160 align:middle line:84% if you have to label data to do model selection 00:37:59.160 --> 00:38:01.140 align:middle line:84% and determine what mechanism works best, 00:38:01.140 --> 00:38:03.700 align:middle line:84% why don't you just use that labeled data for training? 00:38:03.700 --> 00:38:06.880 align:middle line:90% 00:38:06.880 --> 00:38:07.380 align:middle line:90% Yeah. 00:38:07.380 --> 00:38:11.420 align:middle line:84% I mean, I think it's kind of one of these interesting questions. 00:38:11.420 --> 00:38:13.520 align:middle line:84% And also, if you did have some training data, 00:38:13.520 --> 00:38:15.960 align:middle line:84% how does that kind of fit into this system? 00:38:15.960 --> 00:38:20.120 align:middle line:84% But yeah, I do think that in the literature in this space, 00:38:20.120 --> 00:38:23.945 align:middle line:84% one of the big missing pieces is that you still 00:38:23.945 --> 00:38:26.320 align:middle line:84% aren't very good at doing this if you don't have anything 00:38:26.320 --> 00:38:32.320 align:middle line:84% to select the best model or calibrate it in any way at all. 00:38:32.320 --> 00:38:32.840 align:middle line:90% Cool. 00:38:32.840 --> 00:38:33.640 align:middle line:90% Yeah? 00:38:33.640 --> 00:38:35.600 align:middle line:84% AUDIENCE: I guess following up on that, I mean, 00:38:35.600 --> 00:38:39.880 align:middle line:84% isn't the point to reduce costs and time because you 00:38:39.880 --> 00:38:42.760 align:middle line:84% can use bigger data sets that are available? 00:38:42.760 --> 00:38:47.020 align:middle line:84% SARA BEERY: So reducing costs and time are is valuable here. 00:38:47.020 --> 00:38:49.040 align:middle line:84% So maybe one dimension is you're trying 00:38:49.040 --> 00:38:51.920 align:middle line:84% to reduce the amount of human input 00:38:51.920 --> 00:38:54.180 align:middle line:84% you need to get a model to work well. 00:38:54.180 --> 00:38:56.320 align:middle line:84% But I think probably a better motivated scenario 00:38:56.320 --> 00:38:59.150 align:middle line:84% for a lot of these is places where we do not 00:38:59.150 --> 00:39:02.190 align:middle line:84% have the right bandwidth to be able to actually get 00:39:02.190 --> 00:39:04.590 align:middle line:84% the data to where a human could look at it easily. 00:39:04.590 --> 00:39:06.950 align:middle line:84% So we use these domain adaptation models 00:39:06.950 --> 00:39:10.150 align:middle line:84% on these rivers that are in the remote Arctic, 00:39:10.150 --> 00:39:14.090 align:middle line:84% where the way for a human to get there is fly into Anchorage, 00:39:14.090 --> 00:39:17.110 align:middle line:90% take a seaplane, hike. 00:39:17.110 --> 00:39:20.408 align:middle line:84% It's just nontrivial to actually get to the sensor, 00:39:20.408 --> 00:39:22.950 align:middle line:84% and we don't have the kind of satellite connectivity or power 00:39:22.950 --> 00:39:26.730 align:middle line:84% that we would need to get the data to the cloud. 00:39:26.730 --> 00:39:29.270 align:middle line:84% So if you have these strong resource constraints 00:39:29.270 --> 00:39:31.790 align:middle line:84% on your system where you can't necessarily just have 00:39:31.790 --> 00:39:33.790 align:middle line:84% a human easily label data, I think 00:39:33.790 --> 00:39:36.510 align:middle line:84% that's where these types of models tend to make more sense. 00:39:36.510 --> 00:39:38.750 align:middle line:84% Similarly, if you're-- I don't know-- 00:39:38.750 --> 00:39:43.790 align:middle line:84% doing navigation on the moon or deep sea, sometimes, 00:39:43.790 --> 00:39:46.270 align:middle line:84% we just need models that hopefully can adapt 00:39:46.270 --> 00:39:47.570 align:middle line:90% to a distribution. 00:39:47.570 --> 00:39:49.070 align:middle line:84% But the problem is if you don't have 00:39:49.070 --> 00:39:51.670 align:middle line:84% any kind of quality control going in, 00:39:51.670 --> 00:39:55.670 align:middle line:84% anything like this that's kind of learning from data where we 00:39:55.670 --> 00:39:58.150 align:middle line:84% don't have any of true labels, has 00:39:58.150 --> 00:40:00.070 align:middle line:84% the potential, from a reinforcement learning 00:40:00.070 --> 00:40:03.715 align:middle line:84% perspective, to veer off in some catastrophic way. 00:40:03.715 --> 00:40:05.590 align:middle line:84% And so you can fall into these kind of cycles 00:40:05.590 --> 00:40:07.840 align:middle line:84% where the model kind of gets worse and worse and worse 00:40:07.840 --> 00:40:09.612 align:middle line:90% and worse. 00:40:09.612 --> 00:40:12.070 align:middle line:84% And then the other question is, how much source data do you 00:40:12.070 --> 00:40:13.370 align:middle line:90% still have to have access to? 00:40:13.370 --> 00:40:17.430 align:middle line:84% So I think essentially, there's a lot of engineering components 00:40:17.430 --> 00:40:20.230 align:middle line:84% to why you would want to do this and where it's actually 00:40:20.230 --> 00:40:24.330 align:middle line:84% effective and what the value add is. 00:40:24.330 --> 00:40:24.830 align:middle line:90% Yeah. 00:40:24.830 --> 00:40:27.890 align:middle line:90% 00:40:27.890 --> 00:40:32.210 align:middle line:84% But I think there's also an interesting dimension here, 00:40:32.210 --> 00:40:36.310 align:middle line:84% which is these types of domain adaptation approaches also 00:40:36.310 --> 00:40:39.870 align:middle line:84% really benefit from the right starting point. 00:40:39.870 --> 00:40:42.270 align:middle line:84% So you can almost think of this as a different mechanism 00:40:42.270 --> 00:40:45.150 align:middle line:84% for fine-tuning, but the performance of these 00:40:45.150 --> 00:40:47.670 align:middle line:84% is very, very heavily influenced by the performance 00:40:47.670 --> 00:40:49.270 align:middle line:90% of the starting point model. 00:40:49.270 --> 00:40:50.950 align:middle line:84% So it's almost complementary in terms 00:40:50.950 --> 00:40:54.270 align:middle line:84% of trying to build the most generalizable model, the most 00:40:54.270 --> 00:40:56.390 align:middle line:90% useful general representation. 00:40:56.390 --> 00:40:58.270 align:middle line:84% And then this is maybe a mechanism 00:40:58.270 --> 00:41:00.830 align:middle line:84% to fine-tune if you have a lot of data that doesn't have 00:41:00.830 --> 00:41:03.710 align:middle line:84% labels, instead of maybe a little bit of data 00:41:03.710 --> 00:41:06.110 align:middle line:90% that does have labels. 00:41:06.110 --> 00:41:07.430 align:middle line:90% Cool. 00:41:07.430 --> 00:41:10.790 align:middle line:84% So next, I'm going to talk a little bit about how 00:41:10.790 --> 00:41:13.310 align:middle line:84% we might try to transfer information or knowledge 00:41:13.310 --> 00:41:16.830 align:middle line:90% about the outputs of a model. 00:41:16.830 --> 00:41:20.550 align:middle line:84% So if you recall before, we talked 00:41:20.550 --> 00:41:22.670 align:middle line:84% about how you train a classifier with something 00:41:22.670 --> 00:41:24.910 align:middle line:90% like cross entropy. 00:41:24.910 --> 00:41:27.910 align:middle line:90% So here, we have some model. 00:41:27.910 --> 00:41:30.130 align:middle line:90% We have an input, x. 00:41:30.130 --> 00:41:32.630 align:middle line:84% We have some output, y, which is the ground 00:41:32.630 --> 00:41:34.750 align:middle line:84% truth for that model-- or y is the output 00:41:34.750 --> 00:41:36.190 align:middle line:90% prediction for that model. 00:41:36.190 --> 00:41:38.330 align:middle line:84% So maybe there's a bunch of different categories, 00:41:38.330 --> 00:41:44.310 align:middle line:84% and the one that has the largest prediction score or logit score 00:41:44.310 --> 00:41:45.560 align:middle line:90% is cat. 00:41:45.560 --> 00:41:47.310 align:middle line:84% And maybe, it's taken a softmax, actually, 00:41:47.310 --> 00:41:49.870 align:middle line:84% because it's going from 0 to 1, so this is maybe 00:41:49.870 --> 00:41:51.350 align:middle line:90% our prediction score. 00:41:51.350 --> 00:41:52.910 align:middle line:84% And then the ground truth is going 00:41:52.910 --> 00:41:58.390 align:middle line:84% to be some one hot encoded label, where we say, all right. 00:41:58.390 --> 00:42:01.550 align:middle line:84% The true label for this input, x, is cat. 00:42:01.550 --> 00:42:04.070 align:middle line:84% And then we use cross entropy loss 00:42:04.070 --> 00:42:08.110 align:middle line:84% that basically tries to get our model 00:42:08.110 --> 00:42:10.390 align:middle line:84% to predict a distribution that's as close as 00:42:10.390 --> 00:42:13.270 align:middle line:84% possible to this ground truth distribution, which is just 1 00:42:13.270 --> 00:42:17.150 align:middle line:84% for cat and 0 for everything else. 00:42:17.150 --> 00:42:19.190 align:middle line:84% So you could argue that this is also kind 00:42:19.190 --> 00:42:21.870 align:middle line:90% of a teacher and a student. 00:42:21.870 --> 00:42:27.030 align:middle line:84% So here, the teacher is the difference between the ground 00:42:27.030 --> 00:42:31.070 align:middle line:84% truth and our prediction, and the student is kind of the way 00:42:31.070 --> 00:42:33.350 align:middle line:84% that we're capturing that distance. 00:42:33.350 --> 00:42:37.550 align:middle line:84% And so the idea of knowledge distillation is essentially, 00:42:37.550 --> 00:42:41.630 align:middle line:84% now, we have some big pretained model. 00:42:41.630 --> 00:42:43.770 align:middle line:84% We're going to assume that model is fixed. 00:42:43.770 --> 00:42:47.510 align:middle line:84% So we have this little lock on here, so it's a locked model. 00:42:47.510 --> 00:42:53.510 align:middle line:84% And then we want to learn some student where-- so now, we 00:42:53.510 --> 00:42:57.710 align:middle line:84% go from x to y teacher, and we have a model that 00:42:57.710 --> 00:43:00.150 align:middle line:90% goes from x to y student. 00:43:00.150 --> 00:43:04.310 align:middle line:84% And then the way that we try to get this model to distill 00:43:04.310 --> 00:43:09.190 align:middle line:84% is we want the difference or the distance between the teacher 00:43:09.190 --> 00:43:12.230 align:middle line:84% model's predictions and the student's model's predictions 00:43:12.230 --> 00:43:14.430 align:middle line:90% to be as close as possible. 00:43:14.430 --> 00:43:18.070 align:middle line:84% So now, maybe your teacher model has this distribution, 00:43:18.070 --> 00:43:20.170 align:middle line:84% and maybe it's predicting that cat is quite high, 00:43:20.170 --> 00:43:22.253 align:middle line:84% but it's also predicting that tiger is quite high. 00:43:22.253 --> 00:43:26.790 align:middle line:84% So now, when we're actually training our student model, 00:43:26.790 --> 00:43:30.590 align:middle line:84% we want it to similarly match that distribution. 00:43:30.590 --> 00:43:34.390 align:middle line:84% Now, does anyone have any intuition as to why this might 00:43:34.390 --> 00:43:38.590 align:middle line:84% be useful, why it might be useful to train a model 00:43:38.590 --> 00:43:40.910 align:middle line:84% to match a distribution of predictions 00:43:40.910 --> 00:43:43.470 align:middle line:84% as opposed to just a ground truth? 00:43:43.470 --> 00:43:44.350 align:middle line:90% Yeah. 00:43:44.350 --> 00:43:47.362 align:middle line:84% AUDIENCE: Maybe the new model is smaller and quicker. 00:43:47.362 --> 00:43:48.070 align:middle line:90% SARA BEERY: Yeah. 00:43:48.070 --> 00:43:50.500 align:middle line:84% So maybe it's a way to get a new model that's 00:43:50.500 --> 00:43:52.720 align:middle line:84% smaller, can be more computationally efficient, 00:43:52.720 --> 00:43:55.860 align:middle line:90% and can learn the same thing. 00:43:55.860 --> 00:43:58.700 align:middle line:84% Anyone have any intuition about why actually it 00:43:58.700 --> 00:44:04.380 align:middle line:84% might be valuable to match the distribution of the outputs? 00:44:04.380 --> 00:44:05.360 align:middle line:90% Maybe someone new. 00:44:05.360 --> 00:44:06.900 align:middle line:90% Yeah. 00:44:06.900 --> 00:44:09.100 align:middle line:84% AUDIENCE: If there's a cat and a tiger [INAUDIBLE] 00:44:09.100 --> 00:44:10.440 align:middle line:90% example are kind of similar. 00:44:10.440 --> 00:44:12.190 align:middle line:90% SARA BEERY: Yeah, exactly. 00:44:12.190 --> 00:44:13.440 align:middle line:90% AUDIENCE: [INAUDIBLE] animals. 00:44:13.440 --> 00:44:16.540 align:middle line:84% Maybe a dog's more similar than the monkey. 00:44:16.540 --> 00:44:18.400 align:middle line:84% [INAUDIBLE] a little bit to the dog. 00:44:18.400 --> 00:44:20.842 align:middle line:84% It'll give it some more information. 00:44:20.842 --> 00:44:22.300 align:middle line:84% SARA BEERY: Yeah, so that's exactly 00:44:22.300 --> 00:44:23.508 align:middle line:90% the answer I was looking for. 00:44:23.508 --> 00:44:25.900 align:middle line:84% It's giving the model some information 00:44:25.900 --> 00:44:28.260 align:middle line:90% about the similarity. 00:44:28.260 --> 00:44:30.900 align:middle line:84% So here, cats and tigers are maybe similar, 00:44:30.900 --> 00:44:33.740 align:middle line:84% but it's not just the similarity of the category set, 00:44:33.740 --> 00:44:39.920 align:middle line:84% it's the similarity of what's recognizable in a given input. 00:44:39.920 --> 00:44:44.420 align:middle line:84% So if you have an image of a cat, but it's an orange cat, 00:44:44.420 --> 00:44:47.060 align:middle line:84% and it's in tall grass, you could 00:44:47.060 --> 00:44:50.180 align:middle line:84% see how a model would say, hey, that's pretty 00:44:50.180 --> 00:44:53.140 align:middle line:84% similar to a lot of the data I've seen of tigers. 00:44:53.140 --> 00:44:55.780 align:middle line:84% And so now, what you're teaching the model is not just, 00:44:55.780 --> 00:44:57.880 align:middle line:84% this is a cat, you're teaching it, 00:44:57.880 --> 00:45:01.020 align:middle line:84% this is a cat that looks a little bit like a tiger. 00:45:01.020 --> 00:45:03.300 align:middle line:84% And that might actually be useful context 00:45:03.300 --> 00:45:06.660 align:middle line:84% for a model in terms of learning more than just 00:45:06.660 --> 00:45:10.980 align:middle line:90% a set of separable categories. 00:45:10.980 --> 00:45:16.500 align:middle line:84% And so there's another kind of interesting point 00:45:16.500 --> 00:45:19.220 align:middle line:84% here that I think is important to raise that's not just about 00:45:19.220 --> 00:45:25.220 align:middle line:84% efficiency but is about the modern era of many closed source 00:45:25.220 --> 00:45:26.460 align:middle line:90% models. 00:45:26.460 --> 00:45:30.260 align:middle line:84% So if you have access to the outputs of a model, 00:45:30.260 --> 00:45:33.180 align:middle line:84% its predictive scores, but not the model 00:45:33.180 --> 00:45:37.020 align:middle line:84% itself, you could arguably learn a model 00:45:37.020 --> 00:45:40.527 align:middle line:84% that can match the predictive scores of a given output. 00:45:40.527 --> 00:45:42.860 align:middle line:84% And this is maybe one of the reasons that a lot of those 00:45:42.860 --> 00:45:44.460 align:middle line:84% closed source models don't actually 00:45:44.460 --> 00:45:49.860 align:middle line:84% give you output predictions over the whole category set. 00:45:49.860 --> 00:45:50.620 align:middle line:90% Cool. 00:45:50.620 --> 00:45:53.020 align:middle line:84% So now, you can essentially think 00:45:53.020 --> 00:45:55.080 align:middle line:84% about doing standard cross entropy, 00:45:55.080 --> 00:45:58.460 align:middle line:84% but instead of cross entropy being to that binary category, 00:45:58.460 --> 00:46:02.580 align:middle line:84% it's just between the teacher and the student outputs. 00:46:02.580 --> 00:46:05.840 align:middle line:84% And so exactly like what you said before-- 00:46:05.840 --> 00:46:07.300 align:middle line:90% you guys are great here-- 00:46:07.300 --> 00:46:08.780 align:middle line:84% the label you get for every input 00:46:08.780 --> 00:46:11.260 align:middle line:84% might be more informative than just the ground truth 00:46:11.260 --> 00:46:14.420 align:middle line:84% class, and so this soft target is telling you 00:46:14.420 --> 00:46:19.140 align:middle line:84% something about the similarity and what things look like. 00:46:19.140 --> 00:46:20.182 align:middle line:90% So this is a fun example. 00:46:20.182 --> 00:46:22.348 align:middle line:84% It's not just saying this is a dog, but it's saying, 00:46:22.348 --> 00:46:24.460 align:middle line:84% this is a dog that looks a little bit like a cat, 00:46:24.460 --> 00:46:26.720 align:middle line:84% and it looks very different from a bear. 00:46:26.720 --> 00:46:29.580 align:middle line:84% And so maybe that's telling you oh, it's a small dog. 00:46:29.580 --> 00:46:33.940 align:middle line:84% It's-- I don't know-- one of those little yappy dogs. 00:46:33.940 --> 00:46:36.640 align:middle line:84% So this original introductory paper, 00:46:36.640 --> 00:46:39.320 align:middle line:84% which is one of Jeff Hinton's, shows 00:46:39.320 --> 00:46:43.220 align:middle line:84% that you might actually get lower training accuracy when 00:46:43.220 --> 00:46:45.020 align:middle line:84% training this way, but you end up 00:46:45.020 --> 00:46:50.660 align:middle line:84% getting higher test accuracy on this speech recognition task. 00:46:50.660 --> 00:46:53.740 align:middle line:84% And so the other interesting thing 00:46:53.740 --> 00:46:56.500 align:middle line:84% is if these soft targets have really high entropy, 00:46:56.500 --> 00:46:58.860 align:middle line:84% they're giving a lot more information per training case 00:46:58.860 --> 00:46:59.960 align:middle line:90% than hard targets. 00:46:59.960 --> 00:47:02.192 align:middle line:84% And they have a lot less variance in the gradient 00:47:02.192 --> 00:47:03.900 align:middle line:84% between training cases-- is another thing 00:47:03.900 --> 00:47:05.500 align:middle line:90% that they showed in this paper. 00:47:05.500 --> 00:47:08.300 align:middle line:84% And so a small model can often be 00:47:08.300 --> 00:47:12.240 align:middle line:84% trained on a lot less data than the original model 00:47:12.240 --> 00:47:14.180 align:middle line:84% while using a much higher learning rate 00:47:14.180 --> 00:47:17.140 align:middle line:84% when you're doing this type of distillation. 00:47:17.140 --> 00:47:21.840 align:middle line:84% But it's not just something you can use to make a smaller model. 00:47:21.840 --> 00:47:24.140 align:middle line:84% You can also think about learning from a distillation 00:47:24.140 --> 00:47:26.620 align:middle line:90% perspective across modalities. 00:47:26.620 --> 00:47:29.660 align:middle line:84% So here, maybe you have a teacher 00:47:29.660 --> 00:47:34.960 align:middle line:84% that learns to predict categories from an RGB image, 00:47:34.960 --> 00:47:37.980 align:middle line:84% and then you have a student that only has access to depth. 00:47:37.980 --> 00:47:41.520 align:middle line:84% So this might be useful if you have labeled data for RGB, 00:47:41.520 --> 00:47:45.490 align:middle line:84% but you don't necessarily have labeled data for depth. 00:47:45.490 --> 00:47:50.250 align:middle line:84% And so another example of this cross-modal distillation 00:47:50.250 --> 00:47:53.530 align:middle line:84% is this model called SoundNet where 00:47:53.530 --> 00:47:59.090 align:middle line:84% they have waveforms and unlabeled video, 00:47:59.090 --> 00:48:02.370 align:middle line:84% and then they essentially are doing object distributions 00:48:02.370 --> 00:48:05.250 align:middle line:84% and scene distributions, and then they're 00:48:05.250 --> 00:48:08.370 align:middle line:84% learning distillation across both. 00:48:08.370 --> 00:48:10.290 align:middle line:84% There's another interesting version 00:48:10.290 --> 00:48:13.610 align:middle line:84% of distillation from a computational complexity 00:48:13.610 --> 00:48:16.570 align:middle line:84% perspective where you can actually distill 00:48:16.570 --> 00:48:18.150 align:middle line:90% the results of an ensemble. 00:48:18.150 --> 00:48:23.530 align:middle line:84% So often, we find that, even in the modern era, 00:48:23.530 --> 00:48:26.050 align:middle line:84% ensembles over lots of models are often 00:48:26.050 --> 00:48:29.970 align:middle line:84% more predictive than any single model. 00:48:29.970 --> 00:48:32.650 align:middle line:84% One of the arguments for this is kind of a classical one, which 00:48:32.650 --> 00:48:37.370 align:middle line:84% is that every model is going to hopefully make 00:48:37.370 --> 00:48:40.050 align:middle line:84% the same correct answers more often than they 00:48:40.050 --> 00:48:43.250 align:middle line:90% make the same mistakes. 00:48:43.250 --> 00:48:45.210 align:middle line:84% And so here, what you can think about 00:48:45.210 --> 00:48:47.090 align:middle line:84% is maybe you have an ensemble that's really 00:48:47.090 --> 00:48:50.042 align:middle line:84% great ensemble, an ensemble of three different models, 00:48:50.042 --> 00:48:51.750 align:middle line:84% but they're all kind of expensive to run, 00:48:51.750 --> 00:48:53.450 align:middle line:84% and you don't really want to be running three models all 00:48:53.450 --> 00:48:54.290 align:middle line:90% the time. 00:48:54.290 --> 00:48:57.370 align:middle line:84% So you can think about explicitly taking 00:48:57.370 --> 00:48:59.370 align:middle line:84% that ensemble of model outputs and then 00:48:59.370 --> 00:49:03.090 align:middle line:84% trying to distill a model that matches the distribution 00:49:03.090 --> 00:49:04.530 align:middle line:90% of your ensemble. 00:49:04.530 --> 00:49:06.450 align:middle line:84% And again, each model is imperfect 00:49:06.450 --> 00:49:08.050 align:middle line:90% in different random ways. 00:49:08.050 --> 00:49:10.300 align:middle line:84% Errors cancel each other out, and the truth is shared. 00:49:10.300 --> 00:49:11.717 align:middle line:84% Of course, this isn't always true, 00:49:11.717 --> 00:49:13.130 align:middle line:84% because sometimes, we have biases 00:49:13.130 --> 00:49:15.213 align:middle line:84% in models that are learned from our data that will 00:49:15.213 --> 00:49:16.810 align:middle line:90% be consistent for every model. 00:49:16.810 --> 00:49:18.810 align:middle line:84% So the thing you're not going to get out of this 00:49:18.810 --> 00:49:21.770 align:middle line:84% is something that's really good at the types of corner cases 00:49:21.770 --> 00:49:26.170 align:middle line:84% that every single model in your ensemble gets incorrect. 00:49:26.170 --> 00:49:30.890 align:middle line:84% But it can be much more efficient. 00:49:30.890 --> 00:49:32.250 align:middle line:90% Cool. 00:49:32.250 --> 00:49:35.530 align:middle line:84% So what if you don't just want to distill the outputs, 00:49:35.530 --> 00:49:37.650 align:middle line:84% but you also want to capture and distill something 00:49:37.650 --> 00:49:40.070 align:middle line:84% about the knowledge of the representation itself? 00:49:40.070 --> 00:49:43.390 align:middle line:84% So you might want to, for example, 00:49:43.390 --> 00:49:46.490 align:middle line:84% match the embeddings at some intermediate point 00:49:46.490 --> 00:49:49.650 align:middle line:84% within your model using contrastive training. 00:49:49.650 --> 00:49:53.390 align:middle line:84% So as opposed to just matching the logits or the output scores, 00:49:53.390 --> 00:49:57.530 align:middle line:84% you can actually think about distillation specifically using 00:49:57.530 --> 00:50:01.530 align:middle line:84% contrastive learning but where you want to actually supervise 00:50:01.530 --> 00:50:04.635 align:middle line:84% embeddings internal to the model to be 00:50:04.635 --> 00:50:06.010 align:middle line:84% kind of invariant to some viewing 00:50:06.010 --> 00:50:08.770 align:middle line:90% transformation, for example. 00:50:08.770 --> 00:50:10.770 align:middle line:84% So that viewing transformation-- we 00:50:10.770 --> 00:50:13.150 align:middle line:84% talked about rotation invariance, et cetera, before-- 00:50:13.150 --> 00:50:15.050 align:middle line:84% but another version of viewing transformation 00:50:15.050 --> 00:50:18.370 align:middle line:84% could be the different views of that representation 00:50:18.370 --> 00:50:21.150 align:middle line:84% from the teacher, which is a big network, versus the student, 00:50:21.150 --> 00:50:22.430 align:middle line:90% which is a small network. 00:50:22.430 --> 00:50:22.930 align:middle line:90% Yeah? 00:50:22.930 --> 00:50:24.722 align:middle line:84% AUDIENCE: So in practice, when you see this 00:50:24.722 --> 00:50:26.850 align:middle line:84% and what you were talking about earlier, 00:50:26.850 --> 00:50:31.282 align:middle line:84% how much smaller than the teacher is the student model? 00:50:31.282 --> 00:50:31.990 align:middle line:90% SARA BEERY: Yeah. 00:50:31.990 --> 00:50:33.650 align:middle line:90% That's an interesting question. 00:50:33.650 --> 00:50:37.330 align:middle line:84% In practice, how much smaller than the teacher is the student? 00:50:37.330 --> 00:50:42.210 align:middle line:84% So again, in my experience, it kind of depends on the data 00:50:42.210 --> 00:50:44.810 align:middle line:84% that you're trying to get the model to work well on 00:50:44.810 --> 00:50:47.950 align:middle line:84% and how complex it is, how complex the task is. 00:50:47.950 --> 00:50:51.207 align:middle line:84% So how small you can get a model to be and still be useful 00:50:51.207 --> 00:50:53.290 align:middle line:84% often is really directly related to the complexity 00:50:53.290 --> 00:50:56.072 align:middle line:84% of the information you're trying to model. 00:50:56.072 --> 00:50:57.530 align:middle line:84% And there's a lot of different ways 00:50:57.530 --> 00:50:59.810 align:middle line:84% that people try to get small models. 00:50:59.810 --> 00:51:02.790 align:middle line:84% If you want a machine learning model to work on your phone, 00:51:02.790 --> 00:51:04.730 align:middle line:90% it just has to be smaller. 00:51:04.730 --> 00:51:06.670 align:middle line:84% You're not going to put it on an H200-- 00:51:06.670 --> 00:51:10.330 align:middle line:84% an H200 scale model on your phone. 00:51:10.330 --> 00:51:13.190 align:middle line:84% But so people use distillation to make small models, 00:51:13.190 --> 00:51:15.850 align:middle line:84% but they also have a lot of other mechanisms for getting 00:51:15.850 --> 00:51:18.690 align:middle line:84% models to be small, things like model compression 00:51:18.690 --> 00:51:21.610 align:middle line:84% or quantization, for example, that are doing things-- 00:51:21.610 --> 00:51:24.990 align:middle line:84% or pruning, even just explicitly removing weights from a model-- 00:51:24.990 --> 00:51:26.990 align:middle line:84% and trying to get the model to work pretty well. 00:51:26.990 --> 00:51:30.090 align:middle line:84% And then a lot of times, methods that work quite well in terms 00:51:30.090 --> 00:51:32.090 align:middle line:84% of getting really small models to be really good 00:51:32.090 --> 00:51:34.770 align:middle line:84% are some combination of all of the above. 00:51:34.770 --> 00:51:40.530 align:middle line:84% So do some pruning, maybe do some distillation, et cetera. 00:51:40.530 --> 00:51:41.570 align:middle line:90% Yeah. 00:51:41.570 --> 00:51:43.690 align:middle line:84% AUDIENCE: Is that basically related to how 00:51:43.690 --> 00:51:49.250 align:middle line:84% we see cloud Gemini that have different sizes of model-- 00:51:49.250 --> 00:51:53.530 align:middle line:84% is that the result of the techniques you talk about? 00:51:53.530 --> 00:51:56.690 align:middle line:84% SARA BEERY: So you were asking about 00:51:56.690 --> 00:51:59.603 align:middle line:84% if you see these different sizes of models from some 00:51:59.603 --> 00:52:00.770 align:middle line:90% of these foundation models-- 00:52:00.770 --> 00:52:02.978 align:middle line:84% for example, the Gemini model, or some of the OpenAI, 00:52:02.978 --> 00:52:07.010 align:middle line:84% like ChatGPT models-- why they have different sizes 00:52:07.010 --> 00:52:09.370 align:middle line:84% and whether it's actually using distillation 00:52:09.370 --> 00:52:11.410 align:middle line:84% versus just training those models from scratch 00:52:11.410 --> 00:52:12.890 align:middle line:90% at that size. 00:52:12.890 --> 00:52:16.890 align:middle line:84% They don't tell me their trade secrets, so I don't know. 00:52:16.890 --> 00:52:18.450 align:middle line:84% But I will say the reason that you 00:52:18.450 --> 00:52:22.570 align:middle line:84% see these different models put out versions that are small 00:52:22.570 --> 00:52:25.370 align:middle line:84% versus big is because they know that different downstream use 00:52:25.370 --> 00:52:28.010 align:middle line:84% cases will have access to different resources. 00:52:28.010 --> 00:52:31.250 align:middle line:84% Often, it's kind of like-- the big models are better 00:52:31.250 --> 00:52:33.930 align:middle line:84% than the small models, but you use the biggest model 00:52:33.930 --> 00:52:37.540 align:middle line:84% you can afford where afford might be a physical constraint, 00:52:37.540 --> 00:52:42.960 align:middle line:84% it might be a monetary constraint, speed. 00:52:42.960 --> 00:52:44.560 align:middle line:90% Cool. 00:52:44.560 --> 00:52:45.300 align:middle line:90% Anyway, cool. 00:52:45.300 --> 00:52:48.380 align:middle line:84% So here, instead of thinking about rotations or whatever, 00:52:48.380 --> 00:52:50.040 align:middle line:84% the two different views that we want 00:52:50.040 --> 00:52:51.760 align:middle line:84% to be similar in this contrastive way 00:52:51.760 --> 00:52:54.400 align:middle line:84% are maybe intermediate representations 00:52:54.400 --> 00:52:56.680 align:middle line:84% from two different models, and then the idea 00:52:56.680 --> 00:53:00.400 align:middle line:84% is that you try to get this student to output embeddings 00:53:00.400 --> 00:53:02.840 align:middle line:84% or output a representation for input data 00:53:02.840 --> 00:53:06.280 align:middle line:84% that's similar to the teacher's representation. 00:53:06.280 --> 00:53:12.400 align:middle line:84% And there's this cool thing that you can see in the logit space. 00:53:12.400 --> 00:53:16.560 align:middle line:84% So specifically, if you train these-- 00:53:16.560 --> 00:53:19.840 align:middle line:84% do this representation distillation, 00:53:19.840 --> 00:53:24.480 align:middle line:84% and you look at the correlation coefficients between the class 00:53:24.480 --> 00:53:27.800 align:middle line:84% output logits-- so essentially, the student's logits minus 00:53:27.800 --> 00:53:29.640 align:middle line:84% the teacher's logits-- on CIFAR10, 00:53:29.640 --> 00:53:32.280 align:middle line:84% these hundred classes by hundred classes-- 00:53:32.280 --> 00:53:36.360 align:middle line:84% you can see that you get this much, much cleaner behavior when 00:53:36.360 --> 00:53:39.760 align:middle line:84% you do this kind of contrastive representation 00:53:39.760 --> 00:53:43.960 align:middle line:84% distillation versus just doing standard knowledge distillation. 00:53:43.960 --> 00:53:44.720 align:middle line:90% Cool. 00:53:44.720 --> 00:53:48.400 align:middle line:84% So talking again about speed, compression versus distillation. 00:53:48.400 --> 00:53:52.240 align:middle line:84% So there are a lot of different mechanisms for trying 00:53:52.240 --> 00:53:54.400 align:middle line:90% to make models small. 00:53:54.400 --> 00:53:56.360 align:middle line:84% So both compression and distillation 00:53:56.360 --> 00:53:59.200 align:middle line:84% will give you smaller models with similar accuracy 00:53:59.200 --> 00:54:01.240 align:middle line:90% to large ones. 00:54:01.240 --> 00:54:04.442 align:middle line:84% Distillation will use SGD to train a smaller model 00:54:04.442 --> 00:54:06.400 align:middle line:84% to match the characteristics of a large one, so 00:54:06.400 --> 00:54:07.920 align:middle line:90% stochastic gradient descent. 00:54:07.920 --> 00:54:12.520 align:middle line:84% That process of training can be computationally expensive. 00:54:12.520 --> 00:54:16.015 align:middle line:84% Compression, using maybe pruning or quantization 00:54:16.015 --> 00:54:17.640 align:middle line:84% to try to reduce the computational cost 00:54:17.640 --> 00:54:20.925 align:middle line:84% of a large model, that can be much more efficient 00:54:20.925 --> 00:54:22.300 align:middle line:84% from a computational perspective. 00:54:22.300 --> 00:54:25.280 align:middle line:84% So sometimes, people want to use compression 00:54:25.280 --> 00:54:28.280 align:middle line:84% over distillation just because of the speed 00:54:28.280 --> 00:54:32.120 align:middle line:84% or the computational cost of training. 00:54:32.120 --> 00:54:37.320 align:middle line:84% So here is an example from a pretty classic work called 00:54:37.320 --> 00:54:40.720 align:middle line:84% optimal brain compression that's essentially doing 00:54:40.720 --> 00:54:42.775 align:middle line:90% both quantization and pruning. 00:54:42.775 --> 00:54:44.400 align:middle line:84% And here, they're kind of getting a bit 00:54:44.400 --> 00:54:46.120 align:middle line:90% of the best of both worlds. 00:54:46.120 --> 00:54:50.440 align:middle line:84% So pruning is super computationally efficient 00:54:50.440 --> 00:54:53.400 align:middle line:84% because you're just deleting stuff. 00:54:53.400 --> 00:54:56.960 align:middle line:84% Quantization can be a little bit more expensive 00:54:56.960 --> 00:54:59.792 align:middle line:84% and can also explicitly have some components that 00:54:59.792 --> 00:55:02.000 align:middle line:84% feel similar to distillation because you're basically 00:55:02.000 --> 00:55:05.520 align:middle line:84% trying to say, I want intermediate representations 00:55:05.520 --> 00:55:07.840 align:middle line:90% to be similar. 00:55:07.840 --> 00:55:09.760 align:middle line:84% And so here, they might come up with some 00:55:09.760 --> 00:55:13.120 align:middle line:84% of interesting combination of a mask 00:55:13.120 --> 00:55:17.920 align:middle line:84% where you're kind of getting efficient, global optimal 00:55:17.920 --> 00:55:21.320 align:middle line:84% compression in some way, but without it actually being 00:55:21.320 --> 00:55:22.880 align:middle line:90% computationally expensive. 00:55:22.880 --> 00:55:25.820 align:middle line:84% Why do you want compression that isn't computationally expensive? 00:55:25.820 --> 00:55:29.200 align:middle line:84% Again, because it's cool if you can actually directly compress 00:55:29.200 --> 00:55:31.818 align:middle line:84% a model on something like a device. 00:55:31.818 --> 00:55:33.360 align:middle line:84% So maybe you don't want to have to do 00:55:33.360 --> 00:55:34.568 align:middle line:90% the compression on the cloud. 00:55:34.568 --> 00:55:37.640 align:middle line:84% You actually need to do the compression on the edge. 00:55:37.640 --> 00:55:41.040 align:middle line:84% And then you can actually get things to be even smaller 00:55:41.040 --> 00:55:45.040 align:middle line:84% and maintain accuracy if you let them specialize 00:55:45.040 --> 00:55:48.220 align:middle line:84% to a specific distribution while you're compressing them. 00:55:48.220 --> 00:55:50.485 align:middle line:84% So maybe you have a large scale generalist model, 00:55:50.485 --> 00:55:52.360 align:middle line:84% and what you need is a small scale specialist 00:55:52.360 --> 00:55:56.760 align:middle line:84% model-- so for example, a model that only knows about sharks. 00:55:56.760 --> 00:55:59.320 align:middle line:84% We were able to show that you can really 00:55:59.320 --> 00:56:03.740 align:middle line:84% increase the quantization, the amount that you can remove, 00:56:03.740 --> 00:56:06.240 align:middle line:84% and maintain accuracy if you're letting the model specialize 00:56:06.240 --> 00:56:06.920 align:middle line:90% as well. 00:56:06.920 --> 00:56:08.905 align:middle line:90% And that seems pretty intuitive. 00:56:08.905 --> 00:56:10.280 align:middle line:84% If you don't need to keep as much 00:56:10.280 --> 00:56:14.520 align:middle line:84% of the original information, you can make the model much smaller. 00:56:14.520 --> 00:56:16.160 align:middle line:90% Cool. 00:56:16.160 --> 00:56:23.200 align:middle line:84% So towards the end, we're going to dig into foundation models. 00:56:23.200 --> 00:56:27.040 align:middle line:84% So foundation models is a term that 00:56:27.040 --> 00:56:32.950 align:middle line:84% was coined by the computer science department of Stanford. 00:56:32.950 --> 00:56:35.950 align:middle line:84% And essentially, it's just trying 00:56:35.950 --> 00:56:41.910 align:middle line:84% to make a semantic concept for this idea 00:56:41.910 --> 00:56:47.070 align:middle line:84% of really useful transfer learning models, the idea 00:56:47.070 --> 00:56:51.230 align:middle line:84% that you have some models that are so generally 00:56:51.230 --> 00:56:55.990 align:middle line:84% useful that they end up being really effective for a bunch 00:56:55.990 --> 00:56:57.990 align:middle line:90% of different downstream tasks. 00:56:57.990 --> 00:57:03.630 align:middle line:84% And I think you can put different foundation 00:57:03.630 --> 00:57:06.430 align:middle line:84% models in different buckets, I would say. 00:57:06.430 --> 00:57:09.070 align:middle line:84% There's some foundation models that are fundamentally 00:57:09.070 --> 00:57:13.950 align:middle line:84% representation learning models where then the question is, 00:57:13.950 --> 00:57:17.150 align:middle line:84% essentially, do you need to fine-tune at all 00:57:17.150 --> 00:57:18.670 align:middle line:90% to make the model useful? 00:57:18.670 --> 00:57:22.750 align:middle line:84% And I would say as foundation models become 00:57:22.750 --> 00:57:28.270 align:middle line:84% more and more general, you need to fine-tune them less and less, 00:57:28.270 --> 00:57:32.550 align:middle line:84% because maybe the set of things that you are interested in 00:57:32.550 --> 00:57:35.310 align:middle line:84% is more likely to be within the very, very 00:57:35.310 --> 00:57:38.630 align:middle line:84% broad distribution that's covered by that original model. 00:57:38.630 --> 00:57:43.590 align:middle line:84% And mostly, if a model doesn't need to be fine-tuned, 00:57:43.590 --> 00:57:46.310 align:middle line:90% it becomes more accessible. 00:57:46.310 --> 00:57:50.430 align:middle line:84% So this is kind of the power of something like ChatGPT 00:57:50.430 --> 00:57:52.710 align:middle line:84% or something like CLIP or something 00:57:52.710 --> 00:57:57.250 align:middle line:84% like some of these large scale foundational generative models, 00:57:57.250 --> 00:57:59.470 align:middle line:84% for example, like image generation, 00:57:59.470 --> 00:58:02.510 align:middle line:84% is that to be useful for any given end 00:58:02.510 --> 00:58:04.870 align:middle line:84% user for the vast majority of things 00:58:04.870 --> 00:58:06.590 align:middle line:84% that people are kind of interested in, 00:58:06.590 --> 00:58:08.990 align:middle line:90% they don't need to be retrained. 00:58:08.990 --> 00:58:11.230 align:middle line:84% They're kind of usable off the shelf. 00:58:11.230 --> 00:58:13.630 align:middle line:84% And maybe one way you could think about foundation models 00:58:13.630 --> 00:58:16.390 align:middle line:84% is like their use is directly related 00:58:16.390 --> 00:58:20.830 align:middle line:84% to what percent of the things people want to do, 00:58:20.830 --> 00:58:23.270 align:middle line:84% this model can do off the shelf without any 00:58:23.270 --> 00:58:26.070 align:middle line:90% of specialized training. 00:58:26.070 --> 00:58:26.870 align:middle line:90% Cool. 00:58:26.870 --> 00:58:28.510 align:middle line:90% And I like this quote. 00:58:28.510 --> 00:58:32.572 align:middle line:84% I think this either came from Phil or Bill. 00:58:32.572 --> 00:58:35.030 align:middle line:84% "If I have seen further, it is by standing on the shoulders 00:58:35.030 --> 00:58:35.650 align:middle line:90% of Giants." 00:58:35.650 --> 00:58:37.670 align:middle line:90% It's a quote by Newton. 00:58:37.670 --> 00:58:41.310 align:middle line:90% And this is a fun painting. 00:58:41.310 --> 00:58:45.150 align:middle line:84% Blind Orion Searching for the Rising Sun by Poussin. 00:58:45.150 --> 00:58:49.110 align:middle line:84% But I think the idea here is that the bigger the foundation, 00:58:49.110 --> 00:58:54.950 align:middle line:84% maybe the further up we get as a starting point. 00:58:54.950 --> 00:58:59.230 align:middle line:90% So what does this really mean? 00:58:59.230 --> 00:59:02.430 align:middle line:84% Well, so this is a plot from OpenAI, 00:59:02.430 --> 00:59:04.790 align:middle line:90% and it's only going up to 2018. 00:59:04.790 --> 00:59:06.350 align:middle line:84% I should get a newer version of this 00:59:06.350 --> 00:59:10.430 align:middle line:84% that captures how big these models have gotten since. 00:59:10.430 --> 00:59:13.430 align:middle line:84% But it's really just tracking the fact that-- 00:59:13.430 --> 00:59:14.930 align:middle line:84% note that this is on a log scale, 00:59:14.930 --> 00:59:17.110 align:middle line:84% so we're really seeing exponential growth 00:59:17.110 --> 00:59:20.110 align:middle line:90% in the size of models over time. 00:59:20.110 --> 00:59:22.950 align:middle line:90% This is the amount of compute. 00:59:22.950 --> 00:59:25.990 align:middle line:84% So going from AlexNet to the AlphaGo model back in 2018 was 00:59:25.990 --> 00:59:31.030 align:middle line:84% a 300,000 times increase in compute. 00:59:31.030 --> 00:59:32.830 align:middle line:84% And now, some of the models we use 00:59:32.830 --> 00:59:36.610 align:middle line:84% are in the billions and trillions of parameters. 00:59:36.610 --> 00:59:40.510 align:middle line:90% So again, much, much more. 00:59:40.510 --> 00:59:42.810 align:middle line:84% And these models tend to be quite generalizable. 00:59:42.810 --> 00:59:47.710 align:middle line:84% These very large models are much more generally useful. 00:59:47.710 --> 00:59:53.430 align:middle line:84% I think there's some interesting kind of dimensions here. 00:59:53.430 --> 00:59:57.150 align:middle line:84% So one thing is if you look at just computing power usage over 00:59:57.150 --> 00:59:59.950 align:middle line:84% time-- so this one goes up to 2021-- 00:59:59.950 --> 01:00:03.910 align:middle line:84% you can see that academia and even academia industry 01:00:03.910 --> 01:00:06.350 align:middle line:84% collaboration, that computational power 01:00:06.350 --> 01:00:09.590 align:middle line:90% use is staying about the same. 01:00:09.590 --> 01:00:14.550 align:middle line:84% But industry, that's starting to scale pretty significantly. 01:00:14.550 --> 01:00:19.150 align:middle line:84% So you could consider this a, who is building 01:00:19.150 --> 01:00:20.730 align:middle line:90% these bigger and bigger models? 01:00:20.730 --> 01:00:24.790 align:middle line:84% And definitely, I think industry is a-- 01:00:24.790 --> 01:00:29.070 align:middle line:84% well, they're often the only people who can afford it. 01:00:29.070 --> 01:00:33.270 align:middle line:90% And I do just want to raise-- 01:00:33.270 --> 01:00:37.750 align:middle line:84% because it's something that I feel really strongly about-- 01:00:37.750 --> 01:00:40.190 align:middle line:84% that these foundation models aren't free. 01:00:40.190 --> 01:00:43.350 align:middle line:84% So they're definitely not free from the sense of-- 01:00:43.350 --> 01:00:47.350 align:middle line:84% the companies are paying a ton in compute and in power 01:00:47.350 --> 01:00:52.190 align:middle line:84% and in water usage to be able to train them. 01:00:52.190 --> 01:00:54.070 align:middle line:84% Microsoft, I think, announced that they're 01:00:54.070 --> 01:00:57.270 align:middle line:84% spending $80 billion next year on compute. 01:00:57.270 --> 01:01:00.270 align:middle line:90% So that's a lot of money. 01:01:00.270 --> 01:01:02.705 align:middle line:84% That's not something that nonindustry is going 01:01:02.705 --> 01:01:03.830 align:middle line:90% to be able to keep up with. 01:01:03.830 --> 01:01:05.310 align:middle line:84% A lot of that is that Microsoft is actually 01:01:05.310 --> 01:01:07.060 align:middle line:84% the ones who are running a lot of the data 01:01:07.060 --> 01:01:09.130 align:middle line:84% centers that then other groups like OpenAI use. 01:01:09.130 --> 01:01:11.790 align:middle line:90% 01:01:11.790 --> 01:01:13.670 align:middle line:84% But these also have a cost for us 01:01:13.670 --> 01:01:16.710 align:middle line:84% as a society in terms of their climate cost. 01:01:16.710 --> 01:01:19.980 align:middle line:84% So this is from a paper that was looking at estimating the carbon 01:01:19.980 --> 01:01:24.100 align:middle line:84% footprint of one 176 billion parameter language 01:01:24.100 --> 01:01:27.060 align:middle line:90% model from Microsoft. 01:01:27.060 --> 01:01:30.180 align:middle line:84% And they were looking at the entire-- 01:01:30.180 --> 01:01:31.160 align:middle line:90% they were saying, OK. 01:01:31.160 --> 01:01:33.740 align:middle line:84% If we go from the entire scope of the climate 01:01:33.740 --> 01:01:37.180 align:middle line:84% cost of this model, that would include things like raw material 01:01:37.180 --> 01:01:40.340 align:middle line:84% extraction, so mining, the manufacturing 01:01:40.340 --> 01:01:44.800 align:middle line:84% costs of the materials, the equipment manufacturing, model 01:01:44.800 --> 01:01:48.340 align:middle line:84% training, model deployment, and disposal and end 01:01:48.340 --> 01:01:50.318 align:middle line:84% of life of all of those hardware components-- 01:01:50.318 --> 01:01:51.860 align:middle line:84% they were only focused on these three 01:01:51.860 --> 01:01:53.940 align:middle line:84% in the middle, so manufacturing, model training, 01:01:53.940 --> 01:01:55.700 align:middle line:90% and model deployment. 01:01:55.700 --> 01:01:59.340 align:middle line:84% And what they found was a huge factor, specifically 01:01:59.340 --> 01:02:02.380 align:middle line:84% in the power consumption and how that relates 01:02:02.380 --> 01:02:06.100 align:middle line:84% to carbon emissions, is the sustainability of the data 01:02:06.100 --> 01:02:08.420 align:middle line:90% centers' energy sources. 01:02:08.420 --> 01:02:12.460 align:middle line:84% So even much smaller models can have a much, much higher carbon 01:02:12.460 --> 01:02:17.060 align:middle line:84% emission if they're being trained on data centers that 01:02:17.060 --> 01:02:19.500 align:middle line:90% are not renewable power. 01:02:19.500 --> 01:02:22.300 align:middle line:84% And I actually think that one of the benefits 01:02:22.300 --> 01:02:26.100 align:middle line:84% of this compute boom has been the increased focus 01:02:26.100 --> 01:02:28.340 align:middle line:84% on renewable energy and sustainable 01:02:28.340 --> 01:02:31.240 align:middle line:84% power within data centers from these companies 01:02:31.240 --> 01:02:35.400 align:middle line:84% because it's really well motivated by their bottom line. 01:02:35.400 --> 01:02:38.900 align:middle line:84% They're paying a lot of money for this. 01:02:38.900 --> 01:02:42.020 align:middle line:84% But the climate cost of these emissions 01:02:42.020 --> 01:02:43.860 align:middle line:84% is something that they are not necessarily 01:02:43.860 --> 01:02:45.040 align:middle line:90% going to foot the bill for. 01:02:45.040 --> 01:02:48.880 align:middle line:84% And to put this into scale, I mean, 01:02:48.880 --> 01:02:50.765 align:middle line:84% I think power usage and electricity 01:02:50.765 --> 01:02:52.140 align:middle line:84% is one big dimension, but I think 01:02:52.140 --> 01:02:55.020 align:middle line:84% water use is another really important dimension. 01:02:55.020 --> 01:02:57.540 align:middle line:90% I think it was GPT-3-- 01:02:57.540 --> 01:03:00.740 align:middle line:84% was trained mostly at data centers in the state of Iowa. 01:03:00.740 --> 01:03:04.460 align:middle line:84% And the water use in the state went up by 700% 01:03:04.460 --> 01:03:06.900 align:middle line:84% while they were training the models. 01:03:06.900 --> 01:03:08.540 align:middle line:84% I read an estimate recently that said 01:03:08.540 --> 01:03:12.260 align:middle line:84% that about 50 queries to ChatGPT is a bottle of water. 01:03:12.260 --> 01:03:16.140 align:middle line:84% And so if we think about, now, resource access globally, 01:03:16.140 --> 01:03:20.460 align:middle line:84% we already know that we are running out of water. 01:03:20.460 --> 01:03:23.940 align:middle line:84% And now, having these models-- and who 01:03:23.940 --> 01:03:26.900 align:middle line:84% is benefiting from them being successful? 01:03:26.900 --> 01:03:30.180 align:middle line:84% And then what that means about global society and access 01:03:30.180 --> 01:03:35.020 align:middle line:84% to really necessary resources like water is complicated. 01:03:35.020 --> 01:03:36.460 align:middle line:90% Yeah. 01:03:36.460 --> 01:03:38.220 align:middle line:84% AUDIENCE: Just for people who may 01:03:38.220 --> 01:03:41.647 align:middle line:84% be not aware, is the water consumption like-- 01:03:41.647 --> 01:03:42.480 align:middle line:90% SARA BEERY: Cooling. 01:03:42.480 --> 01:03:43.230 align:middle line:90% AUDIENCE: Cooling. 01:03:43.230 --> 01:03:44.930 align:middle line:84% And then is that water-- let's say-- 01:03:44.930 --> 01:03:47.180 align:middle line:84% I assume they're not using-- although maybe they are-- 01:03:47.180 --> 01:03:49.700 align:middle line:90% potable water on the way in. 01:03:49.700 --> 01:03:53.060 align:middle line:84% Does it become unpotable when you use it for cooling? 01:03:53.060 --> 01:03:56.120 align:middle line:90% It sounds like it's bad. 01:03:56.120 --> 01:03:58.235 align:middle line:84% You're not getting the water back. 01:03:58.235 --> 01:03:58.860 align:middle line:90% SARA BEERY: No. 01:03:58.860 --> 01:04:02.060 align:middle line:84% Usually-- I mean, I think there's a lot of, actually, 01:04:02.060 --> 01:04:04.940 align:middle line:84% research now into, can we do this type of cooling 01:04:04.940 --> 01:04:07.500 align:middle line:90% with seawater, for example? 01:04:07.500 --> 01:04:09.940 align:middle line:84% Or what are the effects of running this water 01:04:09.940 --> 01:04:14.740 align:middle line:84% through these cooling systems on the portability of the water? 01:04:14.740 --> 01:04:16.540 align:middle line:84% I'm not going to speak to that because this 01:04:16.540 --> 01:04:18.700 align:middle line:84% is outside of the scope of my understanding 01:04:18.700 --> 01:04:24.060 align:middle line:84% of these mechanical engineering systems or thermal and hydro. 01:04:24.060 --> 01:04:28.300 align:middle line:84% But I will say, I think that there's a booming cottage 01:04:28.300 --> 01:04:30.020 align:middle line:84% industry of researchers who are trying 01:04:30.020 --> 01:04:34.020 align:middle line:84% to figure out how to make data centers less climate intensive. 01:04:34.020 --> 01:04:35.900 align:middle line:84% But I'm also sure that all of you 01:04:35.900 --> 01:04:40.780 align:middle line:84% know that we are vastly outstripping our capacity 01:04:40.780 --> 01:04:43.900 align:middle line:90% in terms of things like power. 01:04:43.900 --> 01:04:46.220 align:middle line:84% It's been very big in the news that Google 01:04:46.220 --> 01:04:47.740 align:middle line:84% is considering reopening Three Mile 01:04:47.740 --> 01:04:49.820 align:middle line:90% Island to power data centers. 01:04:49.820 --> 01:04:53.020 align:middle line:84% So we're thinking about moving back to nuclear power 01:04:53.020 --> 01:04:56.660 align:middle line:84% just to keep up with the AI hunger of this. 01:04:56.660 --> 01:05:02.420 align:middle line:84% And from an ethical dimension, it does get complicated, 01:05:02.420 --> 01:05:11.380 align:middle line:84% because the scale of resources being used versus who actually 01:05:11.380 --> 01:05:14.220 align:middle line:90% is benefiting is a bit messy. 01:05:14.220 --> 01:05:16.900 align:middle line:84% And I think a lot of groups or a lot of companies 01:05:16.900 --> 01:05:20.540 align:middle line:84% would have you believe that these foundation models are 01:05:20.540 --> 01:05:23.820 align:middle line:84% so beneficial to society that it's necessary. 01:05:23.820 --> 01:05:28.660 align:middle line:90% But I am less clear on that. 01:05:28.660 --> 01:05:31.180 align:middle line:84% And one of the things that I want-- 01:05:31.180 --> 01:05:38.900 align:middle line:84% I want the option for a climate friendly search to come back. 01:05:38.900 --> 01:05:41.148 align:middle line:84% I don't want Google to run AI models every time 01:05:41.148 --> 01:05:42.440 align:middle line:90% I'm trying to Google something. 01:05:42.440 --> 01:05:43.940 align:middle line:84% I don't really need the description. 01:05:43.940 --> 01:05:46.220 align:middle line:84% I would love to have an opt in that says no, I 01:05:46.220 --> 01:05:47.485 align:middle line:90% want to go back to MapReduce. 01:05:47.485 --> 01:05:49.360 align:middle line:84% That was much less computationally expensive. 01:05:49.360 --> 01:05:51.180 align:middle line:84% Can you please give me the dumber search 01:05:51.180 --> 01:05:53.010 align:middle line:90% that's more climate friendly? 01:05:53.010 --> 01:05:54.760 align:middle line:84% So I've been talking to people about that. 01:05:54.760 --> 01:05:57.020 align:middle line:90% Maybe we'll start to see it. 01:05:57.020 --> 01:05:58.100 align:middle line:90% Yeah. 01:05:58.100 --> 01:06:01.020 align:middle line:84% AUDIENCE: Is there any-- on that same kind of line-- 01:06:01.020 --> 01:06:05.700 align:middle line:84% any kind of research on how much time, like a compute time, 01:06:05.700 --> 01:06:08.780 align:middle line:90% [INAUDIBLE]? 01:06:08.780 --> 01:06:11.970 align:middle line:84% When I use Google, I have to google 10 things 01:06:11.970 --> 01:06:13.330 align:middle line:90% before I get the answer. 01:06:13.330 --> 01:06:18.410 align:middle line:84% But ChatGPT, I ask it once, and it gives me all the information. 01:06:18.410 --> 01:06:21.210 align:middle line:84% SARA BEERY: So the question is basically 01:06:21.210 --> 01:06:23.930 align:middle line:84% the time versus computational trade 01:06:23.930 --> 01:06:30.370 align:middle line:84% off if you query ChatGPT once versus Google 01:06:30.370 --> 01:06:32.450 align:middle line:90% searching 10 different things. 01:06:32.450 --> 01:06:34.410 align:middle line:90% I don't know. 01:06:34.410 --> 01:06:34.910 align:middle line:90% Yeah. 01:06:34.910 --> 01:06:35.452 align:middle line:90% I don't know. 01:06:35.452 --> 01:06:41.090 align:middle line:84% I think it depends also on how these things are architectured 01:06:41.090 --> 01:06:45.770 align:middle line:84% in terms of things like global similarity search. 01:06:45.770 --> 01:06:48.490 align:middle line:84% It also is, how good are you at using Google? 01:06:48.490 --> 01:06:51.530 align:middle line:90% 01:06:51.530 --> 01:06:53.650 align:middle line:90% Yeah. 01:06:53.650 --> 01:06:55.730 align:middle line:84% The thing I would really love to see 01:06:55.730 --> 01:06:58.610 align:middle line:84% is more transparency about all of this 01:06:58.610 --> 01:07:00.250 align:middle line:90% from the side of companies. 01:07:00.250 --> 01:07:02.730 align:middle line:84% What actually is the computational cost relative 01:07:02.730 --> 01:07:04.410 align:middle line:90% of these different things? 01:07:04.410 --> 01:07:06.710 align:middle line:84% And a lot of companies are looking at, for example, 01:07:06.710 --> 01:07:10.650 align:middle line:84% making use of these different sizes of models 01:07:10.650 --> 01:07:14.290 align:middle line:84% and getting really creative about routing queries 01:07:14.290 --> 01:07:17.783 align:middle line:84% to the less expensive models instead of the bigger ones, 01:07:17.783 --> 01:07:18.950 align:middle line:90% unless they're more complex. 01:07:18.950 --> 01:07:21.770 align:middle line:90% How do you be efficient with it? 01:07:21.770 --> 01:07:23.690 align:middle line:90% But yeah, I think-- 01:07:23.690 --> 01:07:24.570 align:middle line:90% yeah. 01:07:24.570 --> 01:07:29.190 align:middle line:84% I think there's a lot of room to improve still. 01:07:29.190 --> 01:07:29.690 align:middle line:90% Cool. 01:07:29.690 --> 01:07:31.710 align:middle line:84% So what does it take to train one of these models? 01:07:31.710 --> 01:07:33.150 align:middle line:84% These are, again, a little bit out of date. 01:07:33.150 --> 01:07:34.590 align:middle line:90% We have much bigger models now. 01:07:34.590 --> 01:07:37.990 align:middle line:84% But Dalle-2 required 650 million training examples, 01:07:37.990 --> 01:07:39.890 align:middle line:84% semi-curated text to image pairs. 01:07:39.890 --> 01:07:44.650 align:middle line:84% GPT-3 cost estimated around $4.6 million to train once. 01:07:44.650 --> 01:07:46.450 align:middle line:84% I'm sure you guys can imagine that we 01:07:46.450 --> 01:07:47.970 align:middle line:90% don't train these models once. 01:07:47.970 --> 01:07:50.070 align:middle line:84% We have to do some hyperparameter searching. 01:07:50.070 --> 01:07:53.610 align:middle line:84% There's like some differences in comparison. 01:07:53.610 --> 01:07:55.710 align:middle line:90% The PaLM model had 67 authors. 01:07:55.710 --> 01:07:58.390 align:middle line:84% That's 67 engineers working on this. 01:07:58.390 --> 01:08:01.050 align:middle line:90% The numbers are increasing. 01:08:01.050 --> 01:08:06.010 align:middle line:84% But the argument, I guess, is these numbers 01:08:06.010 --> 01:08:07.730 align:middle line:84% might be small if we think of the result 01:08:07.730 --> 01:08:09.570 align:middle line:90% as foundational infrastructure. 01:08:09.570 --> 01:08:11.930 align:middle line:84% So if you compare this maybe to the cost of building 01:08:11.930 --> 01:08:16.390 align:middle line:84% an interstate highway system, or the cost of developing OS X, 01:08:16.390 --> 01:08:17.890 align:middle line:84% maybe there's different ways that we 01:08:17.890 --> 01:08:20.710 align:middle line:84% should be thinking about what the cost is 01:08:20.710 --> 01:08:23.550 align:middle line:84% if this is going to be fundamentally infrastructure 01:08:23.550 --> 01:08:27.090 align:middle line:90% that is foundational. 01:08:27.090 --> 01:08:31.210 align:middle line:84% And one of the motivations for that perspective is that if we 01:08:31.210 --> 01:08:37.149 align:middle line:84% look at scaling, as you go up in these different sizes of models, 01:08:37.149 --> 01:08:40.830 align:middle line:84% and you look at the number of model parameters, 01:08:40.830 --> 01:08:44.450 align:middle line:84% and you look at some metric of performance on a bunch 01:08:44.450 --> 01:08:47.529 align:middle line:84% of different tasks-- so 58 different tasks-- 01:08:47.529 --> 01:08:49.850 align:middle line:84% what you start to see is that you really 01:08:49.850 --> 01:08:53.810 align:middle line:84% don't start to see model performance get significantly 01:08:53.810 --> 01:08:56.752 align:middle line:84% better, where maybe it's, quote, unquote, 01:08:56.752 --> 01:08:58.210 align:middle line:84% working as well as an average human 01:08:58.210 --> 01:09:01.290 align:middle line:84% on some set of these tasks, until you have 01:09:01.290 --> 01:09:04.850 align:middle line:90% much, much larger model sizes. 01:09:04.850 --> 01:09:06.950 align:middle line:90% My argument would be, OK. 01:09:06.950 --> 01:09:08.609 align:middle line:84% Now, let's do research into efficiency 01:09:08.609 --> 01:09:10.662 align:middle line:90% to get these back down. 01:09:10.662 --> 01:09:12.370 align:middle line:84% So what does that actually mean if you're 01:09:12.370 --> 01:09:13.750 align:middle line:84% thinking about learning foundation models? 01:09:13.750 --> 01:09:15.229 align:middle line:90% So one, you have tons of data. 01:09:15.229 --> 01:09:16.510 align:middle line:90% You have some learner. 01:09:16.510 --> 01:09:18.677 align:middle line:84% And then you're generating all these different types 01:09:18.677 --> 01:09:23.170 align:middle line:84% of models that, in many cases, are kind of ready to use. 01:09:23.170 --> 01:09:27.569 align:middle line:84% So now, you want to use or adapt those foundations 01:09:27.569 --> 01:09:30.810 align:middle line:84% with little or no data to solve new problems. 01:09:30.810 --> 01:09:33.729 align:middle line:84% So you now might have some adapter, which 01:09:33.729 --> 01:09:35.753 align:middle line:84% maybe could be something like an AI agent, 01:09:35.753 --> 01:09:38.170 align:middle line:84% or it could just be something that's quite simple, where-- 01:09:38.170 --> 01:09:38.670 align:middle line:90% OK. 01:09:38.670 --> 01:09:39.870 align:middle line:90% We just pull in CLIP. 01:09:39.870 --> 01:09:41.189 align:middle line:90% We have little to no data. 01:09:41.189 --> 01:09:43.569 align:middle line:84% You use your adapter, and you generate some app. 01:09:43.569 --> 01:09:49.330 align:middle line:84% And I would argue that this is a lot of the most recent set 01:09:49.330 --> 01:09:52.890 align:middle line:84% of startups in Silicon Valley have been some version of this, 01:09:52.890 --> 01:09:55.298 align:middle line:84% this "just ask" mentality, where it's like, OK. 01:09:55.298 --> 01:09:57.090 align:middle line:84% We take one of these large language models, 01:09:57.090 --> 01:09:58.923 align:middle line:84% and then we come up with a really clever way 01:09:58.923 --> 01:10:01.610 align:middle line:90% to ask it to do stuff for us. 01:10:01.610 --> 01:10:04.450 align:middle line:84% And it's not really machine learning anymore. 01:10:04.450 --> 01:10:07.680 align:middle line:84% It then turns into this kind of new paradigm of research 01:10:07.680 --> 01:10:10.960 align:middle line:84% that people call prompt engineering. 01:10:10.960 --> 01:10:12.380 align:middle line:84% So what does this maybe look like? 01:10:12.380 --> 01:10:15.480 align:middle line:84% So here, you have something like GPT-3 01:10:15.480 --> 01:10:16.980 align:middle line:90% at a very, very surface level. 01:10:16.980 --> 01:10:20.060 align:middle line:84% It's trained autoregressively to predict the next character, 01:10:20.060 --> 01:10:23.520 align:middle line:84% so colorless green ideas sleep, furiously. 01:10:23.520 --> 01:10:26.400 align:middle line:84% So now, the adapter, again, becomes this idea 01:10:26.400 --> 01:10:27.580 align:middle line:90% of prompt engineering. 01:10:27.580 --> 01:10:29.080 align:middle line:90% So you say, OK. 01:10:29.080 --> 01:10:30.180 align:middle line:90% Here's a review. 01:10:30.180 --> 01:10:32.820 align:middle line:84% Now, you tell it to fill in the blanks. 01:10:32.820 --> 01:10:35.160 align:middle line:84% The sentiment in this review is blank. 01:10:35.160 --> 01:10:37.500 align:middle line:84% And it will fill in the blanks and tell you positive. 01:10:37.500 --> 01:10:41.280 align:middle line:84% And I'm sure many of us have played with the capabilities 01:10:41.280 --> 01:10:42.280 align:middle line:90% of these models. 01:10:42.280 --> 01:10:43.560 align:middle line:90% They are shockingly good. 01:10:43.560 --> 01:10:45.280 align:middle line:90% They get better all the time. 01:10:45.280 --> 01:10:47.080 align:middle line:84% You can ask it dumb stuff like, what 01:10:47.080 --> 01:10:49.320 align:middle line:84% are the top 10 dishes at a specific restaurant 01:10:49.320 --> 01:10:50.400 align:middle line:90% in Cambridge? 01:10:50.400 --> 01:10:52.520 align:middle line:84% And it will give you a list of dishes 01:10:52.520 --> 01:10:55.460 align:middle line:84% and rank them based on who knows what? 01:10:55.460 --> 01:10:58.200 align:middle line:90% 01:10:58.200 --> 01:11:00.920 align:middle line:90% If you say, OK. 01:11:00.920 --> 01:11:04.920 align:middle line:84% Here's a list of primary numbers, and ask it to continue, 01:11:04.920 --> 01:11:07.743 align:middle line:84% it will continue that list of primary numbers. 01:11:07.743 --> 01:11:09.160 align:middle line:84% This is where it gets interesting, 01:11:09.160 --> 01:11:10.760 align:middle line:84% because depending on who you talk to, 01:11:10.760 --> 01:11:14.120 align:middle line:84% it's either learning complex patterns 01:11:14.120 --> 01:11:16.960 align:middle line:84% and learning to mathematically reproduce them, 01:11:16.960 --> 01:11:18.940 align:middle line:84% or it's learning to do pattern matching. 01:11:18.940 --> 01:11:20.980 align:middle line:84% It's essentially just a kernel machine, 01:11:20.980 --> 01:11:23.520 align:middle line:84% and you've just given it everything 01:11:23.520 --> 01:11:24.960 align:middle line:84% it's potentially ever going to see 01:11:24.960 --> 01:11:28.040 align:middle line:90% in terms of its training data. 01:11:28.040 --> 01:11:30.120 align:middle line:84% And you can also think about what people 01:11:30.120 --> 01:11:32.240 align:middle line:90% call in-context learning. 01:11:32.240 --> 01:11:36.310 align:middle line:84% So here, that kind of outer loop, 01:11:36.310 --> 01:11:38.560 align:middle line:84% which is like learning via stochastic gradient descent 01:11:38.560 --> 01:11:41.960 align:middle line:84% during supervised pretaining, then the inner loop is just 01:11:41.960 --> 01:11:44.880 align:middle line:84% this in-context learning where you essentially say, 01:11:44.880 --> 01:11:48.720 align:middle line:84% here's some examples of what I'd like you to do, and then 01:11:48.720 --> 01:11:52.440 align:middle line:90% ask it to continue to do them. 01:11:52.440 --> 01:11:55.900 align:middle line:84% And this kind of starts to feel flavored, like meta-learning, 01:11:55.900 --> 01:11:58.760 align:middle line:84% which we'll talk about a bit more on Tuesday. 01:11:58.760 --> 01:12:01.240 align:middle line:84% And you can also think about its ability to extrapolate. 01:12:01.240 --> 01:12:05.200 align:middle line:84% So here, you show a bunch of elephants that are counting up, 01:12:05.200 --> 01:12:06.580 align:middle line:90% its icons, or whatever. 01:12:06.580 --> 01:12:08.240 align:middle line:90% And then you ask it, OK. 01:12:08.240 --> 01:12:14.160 align:middle line:84% Starting with these four cases, count, and it will do so. 01:12:14.160 --> 01:12:17.360 align:middle line:84% Or it can do things like translate 01:12:17.360 --> 01:12:21.280 align:middle line:84% and translate just from a few examples of English and Spanish. 01:12:21.280 --> 01:12:24.340 align:middle line:84% But again, what is the model actually doing under the hood? 01:12:24.340 --> 01:12:27.160 align:middle line:84% Is it learning something, or is it 01:12:27.160 --> 01:12:28.600 align:middle line:84% finding the right kind of pattern 01:12:28.600 --> 01:12:30.920 align:middle line:84% to match within its training data? 01:12:30.920 --> 01:12:33.160 align:middle line:90% So what does it actually do? 01:12:33.160 --> 01:12:36.480 align:middle line:84% How do we actually do good prompting? 01:12:36.480 --> 01:12:39.360 align:middle line:84% It turns out that handcrafting good prompts can 01:12:39.360 --> 01:12:41.120 align:middle line:84% be really tricky, and these models 01:12:41.120 --> 01:12:44.400 align:middle line:84% are incredibly brittle to prompting. 01:12:44.400 --> 01:12:53.400 align:middle line:84% So you can give it a different template and ask it very simple 01:12:53.400 --> 01:12:54.400 align:middle line:90% things. 01:12:54.400 --> 01:12:55.920 align:middle line:84% And the behavior of the system will 01:12:55.920 --> 01:13:02.760 align:middle line:84% be really sensitive to how the question is asked. 01:13:02.760 --> 01:13:04.837 align:middle line:90% And there's a funny-- 01:13:04.837 --> 01:13:06.920 align:middle line:84% someone said that it was a little bit like working 01:13:06.920 --> 01:13:09.520 align:middle line:90% with a genie in a lamp. 01:13:09.520 --> 01:13:12.660 align:middle line:84% You're like, if you asked the genie for no pain on Earth, 01:13:12.660 --> 01:13:14.520 align:middle line:90% it might kill everybody. 01:13:14.520 --> 01:13:18.480 align:middle line:84% You have to be very careful about doing 01:13:18.480 --> 01:13:19.740 align:middle line:90% this type of prompting. 01:13:19.740 --> 01:13:21.680 align:middle line:84% And then one of the things that people do 01:13:21.680 --> 01:13:23.140 align:middle line:90% is they think about ensembling. 01:13:23.140 --> 01:13:25.700 align:middle line:84% OK, we'll prompt the model with a bunch of different things, 01:13:25.700 --> 01:13:27.367 align:middle line:84% and then we'll ensemble the predictions. 01:13:27.367 --> 01:13:31.240 align:middle line:84% Of course, that's really computationally expensive. 01:13:31.240 --> 01:13:33.000 align:middle line:84% So there's a bunch of research now into, 01:13:33.000 --> 01:13:36.960 align:middle line:84% how do we actually prompt these foundation models effectively? 01:13:36.960 --> 01:13:40.720 align:middle line:84% One is this idea of soft prompting, which basically looks 01:13:40.720 --> 01:13:45.840 align:middle line:84% at trying to optimize some of small continuous task-specific 01:13:45.840 --> 01:13:50.280 align:middle line:84% vector, like encoding, while the language or the model parameters 01:13:50.280 --> 01:13:51.260 align:middle line:90% are frozen. 01:13:51.260 --> 01:13:54.240 align:middle line:84% So here, it's like you have some objective. 01:13:54.240 --> 01:13:57.705 align:middle line:84% It's now kind of back to the normal learning framework, 01:13:57.705 --> 01:13:59.080 align:middle line:84% but we're essentially just trying 01:13:59.080 --> 01:14:03.920 align:middle line:84% to find the optimal words to use as the input to the model. 01:14:03.920 --> 01:14:07.000 align:middle line:84% And so this paper specifically looks at fine-tuning 01:14:07.000 --> 01:14:10.640 align:middle line:84% these input embeddings, these little specializer tokens, 01:14:10.640 --> 01:14:12.740 align:middle line:84% that get prepended to your input task, 01:14:12.740 --> 01:14:15.360 align:middle line:84% and then they help you optimize the result 01:14:15.360 --> 01:14:16.760 align:middle line:90% for some task at hand. 01:14:16.760 --> 01:14:18.840 align:middle line:84% So the tokens might be somewhat gibberish, 01:14:18.840 --> 01:14:20.372 align:middle line:84% but they're really good at finding 01:14:20.372 --> 01:14:22.080 align:middle line:84% the place in the distribution, or helping 01:14:22.080 --> 01:14:24.620 align:middle line:84% the model do that pattern matching 01:14:24.620 --> 01:14:28.800 align:middle line:84% so that it gets you close to something that's effective. 01:14:28.800 --> 01:14:35.680 align:middle line:84% There's also this idea of standard prompting, which 01:14:35.680 --> 01:14:38.448 align:middle line:84% is like, here's a question and an answer, 01:14:38.448 --> 01:14:39.740 align:middle line:90% and then here's a new question. 01:14:39.740 --> 01:14:41.480 align:middle line:84% And then you ask the model to answer. 01:14:41.480 --> 01:14:44.200 align:middle line:84% And then what chain-of-thought prompting does 01:14:44.200 --> 01:14:47.680 align:middle line:84% is it essentially demonstrates an answer that 01:14:47.680 --> 01:14:49.440 align:middle line:90% has reasoning in it. 01:14:49.440 --> 01:14:51.780 align:middle line:84% So you basically tell the model that when you answer, 01:14:51.780 --> 01:14:54.880 align:middle line:84% you want it to answer with logic, 01:14:54.880 --> 01:14:57.000 align:middle line:84% and it's more likely that it will give you 01:14:57.000 --> 01:14:58.450 align:middle line:90% an answer that has some logic. 01:14:58.450 --> 01:15:00.590 align:middle line:84% And then it's been shown that the answer is 01:15:00.590 --> 01:15:02.110 align:middle line:90% more likely to be right. 01:15:02.110 --> 01:15:05.430 align:middle line:84% There's also really dumb stuff you can do, which is like, OK, 01:15:05.430 --> 01:15:06.770 align:middle line:90% you are very smart. 01:15:06.770 --> 01:15:09.070 align:middle line:90% And then ask it a question. 01:15:09.070 --> 01:15:11.290 align:middle line:84% Tell the model, you're a scientist. 01:15:11.290 --> 01:15:14.403 align:middle line:84% Now, tell me-- which is really funny. 01:15:14.403 --> 01:15:17.070 align:middle line:84% And it's one of the reasons that people start anthropomorphizing 01:15:17.070 --> 01:15:17.630 align:middle line:90% these things. 01:15:17.630 --> 01:15:20.950 align:middle line:84% It's like, why do you need to give it compliments? 01:15:20.950 --> 01:15:23.910 align:middle line:84% But actually, I think it's really 01:15:23.910 --> 01:15:26.223 align:middle line:84% about finding the right part of the distribution. 01:15:26.223 --> 01:15:28.390 align:middle line:84% You want to find the part of the distribution that's 01:15:28.390 --> 01:15:32.190 align:middle line:84% learned from text, that's maybe from scientists, and maybe more 01:15:32.190 --> 01:15:33.610 align:middle line:90% likely to be correct. 01:15:33.610 --> 01:15:35.610 align:middle line:84% Anyway, this type of chain-of-thought prompting, 01:15:35.610 --> 01:15:39.270 align:middle line:84% we're telling it how to answer, was 01:15:39.270 --> 01:15:45.090 align:middle line:84% shown to be better using than standard prompting by a lot 01:15:45.090 --> 01:15:47.110 align:middle line:84% and better than the prior best on these types 01:15:47.110 --> 01:15:50.390 align:middle line:90% of math word problems. 01:15:50.390 --> 01:15:54.310 align:middle line:84% And then CLIP is also a foundation model. 01:15:54.310 --> 01:15:55.790 align:middle line:84% And so this is contrastive learning 01:15:55.790 --> 01:15:58.870 align:middle line:84% where you're learning semantic similarity between text 01:15:58.870 --> 01:15:59.930 align:middle line:90% and images. 01:15:59.930 --> 01:16:01.430 align:middle line:84% And so one thing you could do is you 01:16:01.430 --> 01:16:04.410 align:middle line:84% could treat this CLIP model as a representation learning model. 01:16:04.410 --> 01:16:08.110 align:middle line:84% So the adapter could be a linear classifier built 01:16:08.110 --> 01:16:09.530 align:middle line:90% on top of this image encodings. 01:16:09.530 --> 01:16:12.810 align:middle line:84% And they have shown that you can do that. 01:16:12.810 --> 01:16:16.230 align:middle line:84% So this is the number of labeled images per category. 01:16:16.230 --> 01:16:20.150 align:middle line:84% But you can also think of it as an adapter where, now, you 01:16:20.150 --> 01:16:21.710 align:middle line:90% just ask. 01:16:21.710 --> 01:16:26.190 align:middle line:84% So instead of saying, let's do some fine-tuning 01:16:26.190 --> 01:16:31.470 align:middle line:84% over the set of categories I care about, you can just say, 01:16:31.470 --> 01:16:36.070 align:middle line:84% find me the text that is most similar to this image that's 01:16:36.070 --> 01:16:38.290 align:middle line:84% of the form a photo of a-- whatever. 01:16:38.290 --> 01:16:44.990 align:middle line:84% And this is where this stuff gets really weird. 01:16:44.990 --> 01:16:50.230 align:middle line:84% Because if you say, OK, you have the text embedding of a photo 01:16:50.230 --> 01:16:54.710 align:middle line:84% of a dog versus the text embedding of photo of dog, 01:16:54.710 --> 01:16:57.230 align:middle line:84% those will be different enough that it will massively change 01:16:57.230 --> 01:16:59.870 align:middle line:84% the performance of this kind of zero-shot "just ask" 01:16:59.870 --> 01:17:02.070 align:middle line:90% categorization on CLIP. 01:17:02.070 --> 01:17:06.270 align:middle line:90% So they are very, very brittle. 01:17:06.270 --> 01:17:09.750 align:middle line:84% But you can also now think about, OK. 01:17:09.750 --> 01:17:11.657 align:middle line:84% I'm going to encode a sketch of a banana, 01:17:11.657 --> 01:17:13.490 align:middle line:84% and I'm going to encode a photo of a banana. 01:17:13.490 --> 01:17:14.850 align:middle line:90% And I'm going to show it this. 01:17:14.850 --> 01:17:16.350 align:middle line:84% And I'm going to say, which one has 01:17:16.350 --> 01:17:20.070 align:middle line:84% the closest cosine similarity in that joint text image embedding 01:17:20.070 --> 01:17:20.770 align:middle line:90% space? 01:17:20.770 --> 01:17:23.690 align:middle line:84% And that will tell me which one it matches more closely. 01:17:23.690 --> 01:17:26.830 align:middle line:84% So this is more like a sketch of a banana 01:17:26.830 --> 01:17:29.630 align:middle line:84% than it is like a photo of banana. 01:17:29.630 --> 01:17:31.270 align:middle line:90% One of the interesting things-- 01:17:31.270 --> 01:17:34.790 align:middle line:84% we've been working on some of this text-based query 01:17:34.790 --> 01:17:37.370 align:middle line:84% for large image databases for iNaturalist, 01:17:37.370 --> 01:17:40.950 align:middle line:84% which is a really large photo repository of species data. 01:17:40.950 --> 01:17:46.650 align:middle line:84% And these CLIP models are very good at simple concepts, 01:17:46.650 --> 01:17:49.350 align:middle line:84% but they're really bad at composition. 01:17:49.350 --> 01:17:53.490 align:middle line:84% So they're good at ors, and they're bad at ands, 01:17:53.490 --> 01:17:54.527 align:middle line:90% if that makes sense. 01:17:54.527 --> 01:17:56.110 align:middle line:84% And one of the reasons for that is you 01:17:56.110 --> 01:17:58.190 align:middle line:84% can think of these encodings as almost a bit 01:17:58.190 --> 01:17:59.670 align:middle line:90% like a bag of words. 01:17:59.670 --> 01:18:05.110 align:middle line:84% So if you say, I want to find photos of a juvenile bald eagle, 01:18:05.110 --> 01:18:07.030 align:middle line:84% it will find photos of juvenile things 01:18:07.030 --> 01:18:08.690 align:middle line:90% and photos of bald eagles. 01:18:08.690 --> 01:18:11.350 align:middle line:84% But it's not necessarily going to prioritize photos that 01:18:11.350 --> 01:18:14.190 align:middle line:90% are juvenile and bald eagle. 01:18:14.190 --> 01:18:15.830 align:middle line:84% So that type of compositional querying 01:18:15.830 --> 01:18:17.790 align:middle line:84% is a whole other area of research 01:18:17.790 --> 01:18:20.470 align:middle line:84% trying to make these models better at compositions. 01:18:20.470 --> 01:18:23.390 align:middle line:84% And there was a really fun paper in the last year or so 01:18:23.390 --> 01:18:25.750 align:middle line:84% that was teaching CLIP to say no. 01:18:25.750 --> 01:18:27.550 align:middle line:84% This is also really bad at-- if I 01:18:27.550 --> 01:18:30.030 align:middle line:84% say not a photo of a bald eagle, it 01:18:30.030 --> 01:18:32.090 align:middle line:84% would just find a bunch of photos of bald eagles. 01:18:32.090 --> 01:18:34.090 align:middle line:90% It's bad at the negation. 01:18:34.090 --> 01:18:34.590 align:middle line:90% Cool. 01:18:34.590 --> 01:18:37.630 align:middle line:84% So can you think about prompting in other modalities, too, 01:18:37.630 --> 01:18:38.890 align:middle line:90% which is actually kind of fun. 01:18:38.890 --> 01:18:40.790 align:middle line:84% It's just a bit of a psychological trick. 01:18:40.790 --> 01:18:43.710 align:middle line:84% So here's this idea of CLIP prompting. 01:18:43.710 --> 01:18:46.268 align:middle line:84% So instead of modifying the text encoder, saying, OK, 01:18:46.268 --> 01:18:48.310 align:middle line:84% here's some different possible text descriptions, 01:18:48.310 --> 01:18:51.670 align:middle line:84% like finding matches, you can also 01:18:51.670 --> 01:18:53.990 align:middle line:84% look at mechanisms to modify the image 01:18:53.990 --> 01:18:56.455 align:middle line:90% encoder with a visual prompt. 01:18:56.455 --> 01:18:57.830 align:middle line:84% So here, this is an example where 01:18:57.830 --> 01:19:01.310 align:middle line:84% they've learned this optimal kind of border thing 01:19:01.310 --> 01:19:03.670 align:middle line:90% that they put on the image. 01:19:03.670 --> 01:19:06.537 align:middle line:84% And it's kind of similar to adversarial noise. 01:19:06.537 --> 01:19:08.370 align:middle line:84% We learned about learning adversarial noise, 01:19:08.370 --> 01:19:11.770 align:middle line:84% but this is now like learning beneficial noise. 01:19:11.770 --> 01:19:13.310 align:middle line:84% It's like, what's the thing that you 01:19:13.310 --> 01:19:16.150 align:middle line:84% can add to your image that will maximize 01:19:16.150 --> 01:19:20.270 align:middle line:84% the likelihood that this CLIP model will properly 01:19:20.270 --> 01:19:22.550 align:middle line:90% map the correct category? 01:19:22.550 --> 01:19:23.790 align:middle line:90% Which is fascinating. 01:19:23.790 --> 01:19:26.750 align:middle line:84% So given a fixed model, you can learn from a data 01:19:26.750 --> 01:19:28.397 align:middle line:90% set, what can you add? 01:19:28.397 --> 01:19:30.230 align:middle line:84% What types of noise can you add to the image 01:19:30.230 --> 01:19:31.710 align:middle line:84% to actually make the model better 01:19:31.710 --> 01:19:36.670 align:middle line:84% at recognizing dogs versus other things? 01:19:36.670 --> 01:19:38.910 align:middle line:84% And this is noise that helps, no matter 01:19:38.910 --> 01:19:42.710 align:middle line:84% what the input image is, for some fixed data set, domain, 01:19:42.710 --> 01:19:43.330 align:middle line:90% or paradigm. 01:19:43.330 --> 01:19:45.750 align:middle line:84% So you have some training data, and you're basically 01:19:45.750 --> 01:19:49.500 align:middle line:84% finding some constrained set of noise 01:19:49.500 --> 01:19:52.280 align:middle line:84% where, if you add that noise to any image, 01:19:52.280 --> 01:19:53.780 align:middle line:84% it maximally improves the prediction 01:19:53.780 --> 01:19:55.500 align:middle line:90% for the correct class. 01:19:55.500 --> 01:19:57.060 align:middle line:90% Kind of trippy, right? 01:19:57.060 --> 01:19:57.992 align:middle line:90% It's cool. 01:19:57.992 --> 01:19:59.700 align:middle line:84% But really, what it's saying is something 01:19:59.700 --> 01:20:04.860 align:middle line:84% weird about the underlying statistics of the model itself. 01:20:04.860 --> 01:20:10.213 align:middle line:84% So here, it's almost like you're adversarially reprogramming 01:20:10.213 --> 01:20:12.380 align:middle line:84% the model, so you can also think about learning them 01:20:12.380 --> 01:20:15.420 align:middle line:84% in an adversarial way where it's like, you find this thing you 01:20:15.420 --> 01:20:18.540 align:middle line:84% can add to images that will make it maximally wrong in this CLIP 01:20:18.540 --> 01:20:21.380 align:middle line:90% embedding space. 01:20:21.380 --> 01:20:24.380 align:middle line:84% And this is a really weird version of this. 01:20:24.380 --> 01:20:28.980 align:middle line:84% So here, this was an experiment where they basically 01:20:28.980 --> 01:20:31.300 align:middle line:84% took a bunch of satellite imagery, 01:20:31.300 --> 01:20:35.220 align:middle line:84% and they found that adding a single bright pink pixel-- 01:20:35.220 --> 01:20:36.720 align:middle line:84% I don't know if you guys can see it. 01:20:36.720 --> 01:20:41.460 align:middle line:84% It's up there, this one little pink pixel. 01:20:41.460 --> 01:20:44.900 align:middle line:84% They found that if they add a single pink pixel 01:20:44.900 --> 01:20:48.100 align:middle line:84% to every single satellite image in the same place, 01:20:48.100 --> 01:20:49.880 align:middle line:84% it improves performance by a lot. 01:20:49.880 --> 01:20:52.420 align:middle line:90% 01:20:52.420 --> 01:20:57.420 align:middle line:84% AUDIENCE: Is that-- I mean, did they check if, in the training 01:20:57.420 --> 01:21:01.220 align:middle line:84% data set for this model, it statistically was-- 01:21:01.220 --> 01:21:03.360 align:middle line:84% I guess you're saying it's for every class. 01:21:03.360 --> 01:21:05.100 align:middle line:90% SARA BEERY: For every class. 01:21:05.100 --> 01:21:08.720 align:middle line:84% But I mean, if you have a fixed model and a fixed data set, 01:21:08.720 --> 01:21:10.580 align:middle line:84% then you can figure out, what is-- 01:21:10.580 --> 01:21:13.100 align:middle line:84% in this particular constraint, I think they were saying, 01:21:13.100 --> 01:21:15.620 align:middle line:84% we want to add a single pixel-- we 01:21:15.620 --> 01:21:18.740 align:middle line:84% want to change a single pixel value to a specific color, 01:21:18.740 --> 01:21:20.420 align:middle line:90% find the optimal version. 01:21:20.420 --> 01:21:25.620 align:middle line:84% And weirdly, it gives a 3% boost in accuracy 01:21:25.620 --> 01:21:27.460 align:middle line:90% across all the categories. 01:21:27.460 --> 01:21:28.020 align:middle line:90% Anyway. 01:21:28.020 --> 01:21:30.700 align:middle line:84% So I think that this type of work-- 01:21:30.700 --> 01:21:32.557 align:middle line:84% this is not saying the best thing 01:21:32.557 --> 01:21:34.140 align:middle line:84% to do to make these models work really 01:21:34.140 --> 01:21:36.360 align:middle line:84% well is do one of these weird tricks like this. 01:21:36.360 --> 01:21:40.260 align:middle line:84% I think it's more enlightening about what the heck 01:21:40.260 --> 01:21:43.260 align:middle line:84% these models are learning, how fragile and complex 01:21:43.260 --> 01:21:47.020 align:middle line:84% the statistical representations are. 01:21:47.020 --> 01:21:49.460 align:middle line:84% And you can think about this adaptation happening 01:21:49.460 --> 01:21:52.220 align:middle line:84% on the inputs, on the mapping, or on the outputs, this type 01:21:52.220 --> 01:21:54.420 align:middle line:90% of prompt-based adaptation. 01:21:54.420 --> 01:21:57.860 align:middle line:84% So you can take these pretained models, 01:21:57.860 --> 01:21:59.740 align:middle line:90% and you can do fine-tuning. 01:21:59.740 --> 01:22:02.220 align:middle line:84% You can take pretained models, and you can just 01:22:02.220 --> 01:22:05.280 align:middle line:84% learn some linear layer that does some reprojection. 01:22:05.280 --> 01:22:07.020 align:middle line:84% Or you can try and do this type of visual 01:22:07.020 --> 01:22:10.460 align:middle line:84% prompting where you're explicitly adding something 01:22:10.460 --> 01:22:13.420 align:middle line:84% to the top of an image, and then now, it's 01:22:13.420 --> 01:22:15.700 align:middle line:90% getting a better prediction. 01:22:15.700 --> 01:22:18.260 align:middle line:84% And there's a bunch of interesting stuff 01:22:18.260 --> 01:22:21.380 align:middle line:84% where people show that you can do this visual prompt tuning 01:22:21.380 --> 01:22:25.380 align:middle line:84% for these really large models that can be useful even when 01:22:25.380 --> 01:22:28.280 align:middle line:84% the model weights or the logits are not available. 01:22:28.280 --> 01:22:29.820 align:middle line:90% It's kind of like distillation. 01:22:29.820 --> 01:22:31.385 align:middle line:84% If you just add things to the input, 01:22:31.385 --> 01:22:33.260 align:middle line:84% see what the output is, you don't need access 01:22:33.260 --> 01:22:36.700 align:middle line:84% to the underlying model to find this type of visual prompting 01:22:36.700 --> 01:22:38.880 align:middle line:90% that's valuable. 01:22:38.880 --> 01:22:40.380 align:middle line:84% And then there's another fun version 01:22:40.380 --> 01:22:46.380 align:middle line:84% of this, which is like fill in the blank style tuning. 01:22:46.380 --> 01:22:47.980 align:middle line:84% So basically, these people figured out 01:22:47.980 --> 01:22:54.700 align:middle line:84% how to get an image foundation model to do segmentation 01:22:54.700 --> 01:22:56.720 align:middle line:90% via in-context prompting. 01:22:56.720 --> 01:22:59.180 align:middle line:84% So they made these little grids of images 01:22:59.180 --> 01:23:02.180 align:middle line:84% where they would show an image with an example 01:23:02.180 --> 01:23:05.180 align:middle line:84% of a segmentation mask for a class, and then another image, 01:23:05.180 --> 01:23:07.720 align:middle line:90% and then have this empty area. 01:23:07.720 --> 01:23:10.180 align:middle line:84% And then they would use an inpainting foundation model 01:23:10.180 --> 01:23:14.780 align:middle line:84% to get it to prompt you to do segmentation, 01:23:14.780 --> 01:23:17.440 align:middle line:84% even without being explicitly trained to do segmentation. 01:23:17.440 --> 01:23:19.982 align:middle line:84% And then they showed this for many different types of things. 01:23:19.982 --> 01:23:22.140 align:middle line:84% So edge detection, colorization, inpainting. 01:23:22.140 --> 01:23:24.660 align:middle line:84% You're essentially asking these models-- 01:23:24.660 --> 01:23:26.780 align:middle line:84% you're prompting them with, here's 01:23:26.780 --> 01:23:30.640 align:middle line:84% an example of what I want with one case, 01:23:30.640 --> 01:23:33.380 align:middle line:84% and then it's just giving you the kind of correct label 01:23:33.380 --> 01:23:37.700 align:middle line:84% across a bunch of different possible downstream types 01:23:37.700 --> 01:23:39.060 align:middle line:90% of categories. 01:23:39.060 --> 01:23:44.530 align:middle line:84% So these foundation models really have a lot of capacity, 01:23:44.530 --> 01:23:46.410 align:middle line:84% and there's so many creative ways 01:23:46.410 --> 01:23:48.570 align:middle line:84% that we can try to harness that capacity that I 01:23:48.570 --> 01:23:50.747 align:middle line:90% think people have looked at. 01:23:50.747 --> 01:23:52.330 align:middle line:84% And there's some interesting resources 01:23:52.330 --> 01:23:54.490 align:middle line:90% if you guys are interested. 01:23:54.490 --> 01:23:56.630 align:middle line:90% So in the last minute-- sorry. 01:23:56.630 --> 01:23:58.890 align:middle line:84% We're kind of running late today. 01:23:58.890 --> 01:24:00.730 align:middle line:90% Combining foundation models. 01:24:00.730 --> 01:24:04.070 align:middle line:84% So you can learn foundation models and adapt them, 01:24:04.070 --> 01:24:05.850 align:middle line:84% but you can also think about taking 01:24:05.850 --> 01:24:07.490 align:middle line:84% these combinations of foundation models 01:24:07.490 --> 01:24:08.865 align:middle line:84% and building things that are kind 01:24:08.865 --> 01:24:11.470 align:middle line:84% of nonmodular on top of, again, these building blocks. 01:24:11.470 --> 01:24:14.370 align:middle line:84% So you want to do something that's text to image. 01:24:14.370 --> 01:24:17.110 align:middle line:90% You can take VQGAN and CLIP. 01:24:17.110 --> 01:24:19.410 align:middle line:84% You can plug pretained models together. 01:24:19.410 --> 01:24:20.872 align:middle line:90% I think we showed this early on. 01:24:20.872 --> 01:24:22.330 align:middle line:84% And so this is really just thinking 01:24:22.330 --> 01:24:24.050 align:middle line:84% about plugging these different image foundation 01:24:24.050 --> 01:24:24.830 align:middle line:90% models together. 01:24:24.830 --> 01:24:26.670 align:middle line:90% You have some image generator. 01:24:26.670 --> 01:24:30.250 align:middle line:84% You have a CLIP model that's got this similarity foundations 01:24:30.250 --> 01:24:33.150 align:middle line:84% and semantic meaning between text and images. 01:24:33.150 --> 01:24:36.370 align:middle line:84% And then you can do something optimize this little embedding 01:24:36.370 --> 01:24:38.730 align:middle line:84% vector, this latent space vector, 01:24:38.730 --> 01:24:43.770 align:middle line:84% to get an image that maximally matches that text prompt. 01:24:43.770 --> 01:24:49.650 align:middle line:84% And there's a bunch of recent work on this text-based image 01:24:49.650 --> 01:24:51.930 align:middle line:90% generation. 01:24:51.930 --> 01:24:54.650 align:middle line:84% And Dalle-2 is a good example of that, where, essentially, they 01:24:54.650 --> 01:24:57.250 align:middle line:84% have a CLIP objective, and then they also 01:24:57.250 --> 01:25:00.410 align:middle line:84% have this prior that helps them do decoding slightly more 01:25:00.410 --> 01:25:01.690 align:middle line:90% efficiently. 01:25:01.690 --> 01:25:02.210 align:middle line:90% Cool. 01:25:02.210 --> 01:25:04.510 align:middle line:84% So today, we talked about learning with little data, 01:25:04.510 --> 01:25:08.450 align:middle line:84% transferring knowledge, about the mapping, the outputs, 01:25:08.450 --> 01:25:10.370 align:middle line:84% and a bit about foundation models. 01:25:10.370 --> 01:25:13.210 align:middle line:90% We'll follow up on Tuesday. 01:25:13.210 --> 01:25:19.200 align:middle line:84% And don't be too sad, but think twice about ChatGPT. 01:25:19.200 --> 01:25:41.000 align:middle line:90%