vtt/episode_019_small.vtt

WEBVTT

00:00.000 --> 00:03.760
 The following is a conversation with Ian Goodfellow.

00:03.760 --> 00:06.360
 He's the author of the popular textbook on deep learning

00:06.360 --> 00:08.880
 simply titled Deep Learning.

00:08.880 --> 00:12.320
 He coined the term of generative adversarial networks,

00:12.320 --> 00:14.560
 otherwise known as GANs.

00:14.560 --> 00:18.160
 And with his 2014 paper is responsible

00:18.160 --> 00:20.440
 for launching the incredible growth

00:20.440 --> 00:22.120
 of research and innovation

00:22.120 --> 00:24.720
 in this subfield of deep learning.

00:24.720 --> 00:27.520
 He got his BS and MS at Stanford,

00:27.520 --> 00:30.120
 his PhD at University of Montreal

00:30.120 --> 00:33.200
 with Yoshua Benjo and Aaron Kervel.

00:33.200 --> 00:35.240
 He held several research positions,

00:35.240 --> 00:37.560
 including at OpenAI, Google Brain,

00:37.560 --> 00:41.560
 and now at Apple as the director of machine learning.

00:41.560 --> 00:45.400
 This recording happened while Ian was still a Google Brain,

00:45.400 --> 00:48.520
 but we don't talk about anything specific to Google

00:48.520 --> 00:50.760
 or any other organization.

00:50.760 --> 00:52.480
 This conversation is part

00:52.480 --> 00:54.520
 of the artificial intelligence podcast.

00:54.520 --> 00:56.680
 If you enjoy it, subscribe on YouTube,

00:56.680 --> 00:59.600
 iTunes, or simply connect with me on Twitter

00:59.600 --> 01:03.000
 at Lex Freedman, spelled F R I D.

01:03.000 --> 01:07.080
 And now here's my conversation with Ian Goodfellow.

01:08.240 --> 01:11.000
 You open your popular deep learning book

01:11.000 --> 01:13.600
 with a Russian doll type diagram

01:13.600 --> 01:15.880
 that shows deep learning is a subset

01:15.880 --> 01:17.160
 of representation learning,

01:17.160 --> 01:19.960
 which in turn is a subset of machine learning

01:19.960 --> 01:22.520
 and finally a subset of AI.

01:22.520 --> 01:25.280
 So this kind of implies that there may be limits

01:25.280 --> 01:27.720
 to deep learning in the context of AI.

01:27.720 --> 01:31.560
 So what do you think is the current limits of deep learning

01:31.560 --> 01:33.120
 and are those limits something

01:33.120 --> 01:35.760
 that we can overcome with time?

01:35.760 --> 01:37.720
 Yeah, I think one of the biggest limitations

01:37.720 --> 01:39.320
 of deep learning is that right now

01:39.320 --> 01:42.920
 it requires really a lot of data, especially labeled data.

01:43.960 --> 01:45.480
 There are some unsupervised

01:45.480 --> 01:47.160
 and semi supervised learning algorithms

01:47.160 --> 01:49.480
 that can reduce the amount of labeled data you need,

01:49.480 --> 01:52.200
 but they still require a lot of unlabeled data.

01:52.200 --> 01:54.200
 Reinforcement learning algorithms, they don't need labels,

01:54.200 --> 01:56.280
 but they need really a lot of experiences.

01:57.280 --> 01:58.920
 As human beings, we don't learn to play a pong

01:58.920 --> 02:01.520
 by failing at pong two million times.

02:02.720 --> 02:05.880
 So just getting the generalization ability better

02:05.880 --> 02:08.040
 is one of the most important bottlenecks

02:08.040 --> 02:10.520
 in the capability of the technology today.

02:10.520 --> 02:12.360
 And then I guess I'd also say deep learning

02:12.360 --> 02:15.620
 is like a component of a bigger system.

02:16.600 --> 02:19.040
 So far, nobody is really proposing to have

02:20.600 --> 02:22.000
 only what you'd call deep learning

02:22.000 --> 02:25.520
 as the entire ingredient of intelligence.

02:25.520 --> 02:29.860
 You use deep learning as sub modules of other systems,

02:29.860 --> 02:32.320
 like AlphaGo has a deep learning model

02:32.320 --> 02:34.160
 that estimates the value function.

02:35.200 --> 02:36.600
 Most reinforcement learning algorithms

02:36.600 --> 02:37.880
 have a deep learning module

02:37.880 --> 02:40.320
 that estimates which action to take next,

02:40.320 --> 02:42.480
 but you might have other components.

02:42.480 --> 02:46.120
 So you're basically building a function estimator.

02:46.120 --> 02:48.600
 Do you think it's possible?

02:48.600 --> 02:51.000
 You said nobody's kind of been thinking about this so far,

02:51.000 --> 02:54.320
 but do you think neural networks could be made to reason

02:54.320 --> 02:57.720
 in the way symbolic systems did in the 80s and 90s

02:57.720 --> 03:00.160
 to do more, create more like programs

03:00.160 --> 03:01.440
 as opposed to functions?

03:01.440 --> 03:03.920
 Yeah, I think we already see that a little bit.

03:04.880 --> 03:08.860
 I already kind of think of neural nets as a kind of program.

03:08.860 --> 03:12.920
 I think of deep learning as basically learning programs

03:12.920 --> 03:15.280
 that have more than one step.

03:15.280 --> 03:16.960
 So if you draw a flow chart

03:16.960 --> 03:19.540
 or if you draw a TensorFlow graph

03:19.540 --> 03:21.880
 describing your machine learning model,

03:21.880 --> 03:23.520
 I think of the depth of that graph

03:23.520 --> 03:25.880
 as describing the number of steps that run in sequence

03:25.880 --> 03:27.640
 and then the width of that graph

03:27.640 --> 03:30.120
 as the number of steps that run in parallel.

03:30.120 --> 03:31.680
 Now it's been long enough

03:31.680 --> 03:32.880
 that we've had deep learning working

03:32.880 --> 03:33.880
 that it's a little bit silly

03:33.880 --> 03:35.740
 to even discuss shallow learning anymore,

03:35.740 --> 03:38.880
 but back when I first got involved in AI,

03:38.880 --> 03:40.080
 when we used machine learning,

03:40.080 --> 03:41.280
 we were usually learning things

03:41.280 --> 03:43.680
 like support vector machines.

03:43.680 --> 03:45.640
 You could have a lot of input features to the model

03:45.640 --> 03:48.120
 and you could multiply each feature by a different weight.

03:48.120 --> 03:51.240
 All those multiplications were done in parallel to each other

03:51.240 --> 03:52.720
 and there wasn't a lot done in series.

03:52.720 --> 03:54.360
 I think what we got with deep learning

03:54.360 --> 03:58.400
 was really the ability to have steps of a program

03:58.400 --> 04:00.320
 that run in sequence.

04:00.320 --> 04:03.200
 And I think that we've actually started to see

04:03.200 --> 04:05.040
 that what's important with deep learning

04:05.040 --> 04:08.000
 is more the fact that we have a multi step program

04:08.000 --> 04:10.800
 rather than the fact that we've learned a representation.

04:10.800 --> 04:15.120
 If you look at things like Resnuts, for example,

04:15.120 --> 04:18.660
 they take one particular kind of representation

04:18.660 --> 04:21.040
 and they update it several times.

04:21.040 --> 04:23.560
 Back when deep learning first really took off

04:23.560 --> 04:25.760
 in the academic world in 2006,

04:25.760 --> 04:28.400
 when Jeff Hinton showed that you could train

04:28.400 --> 04:30.160
 deep belief networks,

04:30.160 --> 04:31.960
 everybody who was interested in the idea

04:31.960 --> 04:33.560
 thought of it as each layer

04:33.560 --> 04:35.960
 learns a different level of abstraction,

04:35.960 --> 04:37.840
 that the first layer trained on images

04:37.840 --> 04:38.960
 learns something like edges

04:38.960 --> 04:40.420
 and the second layer learns corners

04:40.420 --> 04:43.320
 and eventually you get these kind of grandmother cell units

04:43.320 --> 04:45.920
 that recognize specific objects.

04:45.920 --> 04:48.560
 Today, I think most people think of it more

04:48.560 --> 04:52.000
 as a computer program where as you add more layers,

04:52.000 --> 04:55.120
 you can do more updates before you output your final number.

04:55.120 --> 04:57.160
 But I don't think anybody believes that

04:57.160 --> 05:02.040
 layer 150 of the Resnet is a grandmother cell

05:02.040 --> 05:05.080
 and layer 100 is contours or something like that.

05:06.040 --> 05:08.160
 Okay, so you're not thinking of it

05:08.160 --> 05:11.520
 as a singular representation that keeps building.

05:11.520 --> 05:15.960
 You think of it as a program sort of almost like a state.

05:15.960 --> 05:18.600
 The representation is a state of understanding.

05:18.600 --> 05:21.520
 Yeah, I think of it as a program that makes several updates

05:21.520 --> 05:23.840
 and arrives at better and better understandings,

05:23.840 --> 05:27.500
 but it's not replacing the representation at each step.

05:27.500 --> 05:29.160
 It's refining it.

05:29.160 --> 05:31.660
 And in some sense, that's a little bit like reasoning.

05:31.660 --> 05:33.560
 It's not reasoning in the form of deduction,

05:33.560 --> 05:36.960
 but it's reasoning in the form of taking a thought

05:36.960 --> 05:39.440
 and refining it and refining it carefully

05:39.440 --> 05:41.240
 until it's good enough to use.

05:41.240 --> 05:43.560
 So do you think, and I hope you don't mind,

05:43.560 --> 05:46.040
 we'll jump philosophical every once in a while.

05:46.040 --> 05:50.480
 Do you think of, you know, cognition, human cognition,

05:50.480 --> 05:53.520
 or even consciousness as simply a result

05:53.520 --> 05:58.120
 of this kind of sequential representation learning?

05:58.120 --> 06:00.440
 Do you think that can emerge?

06:00.440 --> 06:02.440
 Cognition, yes, I think so.

06:02.440 --> 06:05.160
 Consciousness, it's really hard to even define

06:05.160 --> 06:06.400
 what we mean by that.

06:07.400 --> 06:09.840
 I guess there's, consciousness is often defined

06:09.840 --> 06:12.120
 as things like having self awareness,

06:12.120 --> 06:15.200
 and that's relatively easy to turn it

06:15.200 --> 06:17.200
 to something actionable for a computer scientist

06:17.200 --> 06:18.400
 to reason about.

06:18.400 --> 06:20.080
 People also define consciousness in terms

06:20.080 --> 06:24.000
 of having qualitative states of experience, like qualia.

06:24.000 --> 06:25.280
 There's all these philosophical problems,

06:25.280 --> 06:27.880
 like could you imagine a zombie

06:27.880 --> 06:30.760
 who does all the same information processing as a human,

06:30.760 --> 06:33.500
 but doesn't really have the qualitative experiences

06:33.500 --> 06:34.720
 that we have?

06:34.720 --> 06:37.580
 That sort of thing, I have no idea how to formalize

06:37.580 --> 06:39.960
 or turn it into a scientific question.

06:39.960 --> 06:41.600
 I don't know how you could run an experiment

06:41.600 --> 06:44.880
 to tell whether a person is a zombie or not.

06:44.880 --> 06:46.680
 And similarly, I don't know how you could run

06:46.680 --> 06:49.680
 an experiment to tell whether an advanced AI system

06:49.680 --> 06:53.080
 had become conscious in the sense of qualia or not.

06:53.080 --> 06:54.600
 But in the more practical sense,

06:54.600 --> 06:56.320
 like almost like self attention,

06:56.320 --> 06:58.920
 you think consciousness and cognition can,

06:58.920 --> 07:03.240
 in an impressive way, emerge from current types

07:03.240 --> 07:05.600
 of architectures that we think of as determining.

07:05.600 --> 07:07.920
 Or if you think of consciousness

07:07.920 --> 07:12.160
 in terms of self awareness and just making plans

07:12.160 --> 07:15.120
 based on the fact that the agent itself

07:15.120 --> 07:18.000
 exists in the world, reinforcement learning algorithms

07:18.000 --> 07:20.840
 are already more or less forced to model

07:20.840 --> 07:23.040
 the agent's effect on the environment.

07:23.040 --> 07:26.340
 So that more limited version of consciousness

07:26.340 --> 07:30.560
 is already something that we get limited versions

07:30.560 --> 07:32.960
 of with reinforcement learning algorithms

07:32.960 --> 07:34.640
 if they're trained well.

07:34.640 --> 07:37.440
 But you say limited.

07:37.440 --> 07:39.920
 So the big question really is how you jump

07:39.920 --> 07:42.120
 from limited to human level, right?

07:42.120 --> 07:44.640
 And whether it's possible,

07:46.840 --> 07:49.000
 even just building common sense reasoning

07:49.000 --> 07:50.520
 seems to be exceptionally difficult.

07:50.520 --> 07:52.480
 So if we scale things up,

07:52.480 --> 07:55.000
 if we get much better on supervised learning,

07:55.000 --> 07:56.600
 if we get better at labeling,

07:56.600 --> 08:00.640
 if we get bigger datasets, more compute,

08:00.640 --> 08:03.880
 do you think we'll start to see really impressive things

08:03.880 --> 08:08.760
 that go from limited to something echoes

08:08.760 --> 08:10.320
 of human level cognition?

08:10.320 --> 08:11.200
 I think so, yeah.

08:11.200 --> 08:13.360
 I'm optimistic about what can happen

08:13.360 --> 08:16.440
 just with more computation and more data.

08:16.440 --> 08:20.120
 I do think it'll be important to get the right kind of data.

08:20.120 --> 08:23.160
 Today, most of the machine learning systems we train

08:23.160 --> 08:27.560
 are mostly trained on one type of data for each model.

08:27.560 --> 08:31.380
 But the human brain, we get all of our different senses

08:31.380 --> 08:33.880
 and we have many different experiences

08:33.880 --> 08:36.320
 like riding a bike, driving a car,

08:36.320 --> 08:37.940
 talking to people, reading.

08:39.160 --> 08:42.440
 I think when we get that kind of integrated dataset

08:42.440 --> 08:44.400
 working with a machine learning model

08:44.400 --> 08:47.640
 that can actually close the loop and interact,

08:47.640 --> 08:50.480
 we may find that algorithms not so different

08:50.480 --> 08:51.840
 from what we have today,

08:51.840 --> 08:53.240
 learn really interesting things

08:53.240 --> 08:54.400
 when you scale them up a lot

08:54.400 --> 08:58.240
 and train them on a large amount of multimodal data.

08:58.240 --> 08:59.640
 So multimodal is really interesting,

08:59.640 --> 09:04.000
 but within, like you're working adversarial examples.

09:04.000 --> 09:09.000
 So selecting within model, within one mode of data,

09:11.120 --> 09:13.800
 selecting better at what are the difficult cases

09:13.800 --> 09:16.120
 from which you're most useful to learn from.

09:16.120 --> 09:18.880
 Oh, yeah, like could we get a whole lot of mileage

09:18.880 --> 09:22.280
 out of designing a model that's resistant

09:22.280 --> 09:24.080
 to adversarial examples or something like that?

09:24.080 --> 09:26.280
 Right, that's the question.

09:26.280 --> 09:27.760
 My thinking on that has evolved a lot

09:27.760 --> 09:28.920
 over the last few years.

09:28.920 --> 09:31.280
 When I first started to really invest

09:31.280 --> 09:32.760
 in studying adversarial examples,

09:32.760 --> 09:36.320
 I was thinking of it mostly as adversarial examples

09:36.320 --> 09:39.000
 reveal a big problem with machine learning.

09:39.000 --> 09:41.160
 And we would like to close the gap

09:41.160 --> 09:44.160
 between how machine learning models respond

09:44.160 --> 09:46.560
 to adversarial examples and how humans respond.

09:47.640 --> 09:49.160
 After studying the problem more,

09:49.160 --> 09:51.960
 I still think that adversarial examples are important.

09:51.960 --> 09:55.440
 I think of them now more of as a security liability

09:55.440 --> 09:57.800
 than as an issue that necessarily shows

09:57.800 --> 09:59.880
 there's something uniquely wrong

09:59.880 --> 10:02.800
 with machine learning as opposed to humans.

10:02.800 --> 10:04.600
 Also, do you see them as a tool

10:04.600 --> 10:06.480
 to improve the performance of the system?

10:06.480 --> 10:10.760
 Not on the security side, but literally just accuracy.

10:10.760 --> 10:13.480
 I do see them as a kind of tool on that side,

10:13.480 --> 10:16.640
 but maybe not quite as much as I used to think.

10:16.640 --> 10:18.520
 We've started to find that there's a trade off

10:18.520 --> 10:21.680
 between accuracy on adversarial examples

10:21.680 --> 10:24.360
 and accuracy on clean examples.

10:24.360 --> 10:28.320
 Back in 2014, when I did the first adversarily trained

10:28.320 --> 10:30.840
 classifier that showed resistance

10:30.840 --> 10:33.040
 to some kinds of adversarial examples,

10:33.040 --> 10:36.040
 it also got better at the clean data on MNIST.

10:36.040 --> 10:37.720
 And that's something we've replicated several times

10:37.720 --> 10:39.640
 on MNIST, that when we train

10:39.640 --> 10:41.480
 against weak adversarial examples,

10:41.480 --> 10:43.880
 MNIST classifiers get more accurate.

10:43.880 --> 10:47.080
 So far that hasn't really held up on other data sets

10:47.080 --> 10:48.880
 and hasn't held up when we train

10:48.880 --> 10:50.720
 against stronger adversaries.

10:50.720 --> 10:53.160
 It seems like when you confront

10:53.160 --> 10:55.720
 a really strong adversary,

10:55.720 --> 10:58.080
 you tend to have to give something up.

10:58.080 --> 11:00.520
 Interesting, but it's such a compelling idea

11:00.520 --> 11:04.800
 because it feels like that's how us humans learn

11:04.800 --> 11:06.320
 to do the difficult cases.

11:06.320 --> 11:08.760
 We try to think of what would we screw up

11:08.760 --> 11:11.000
 and then we make sure we fix that.

11:11.000 --> 11:13.680
 It's also in a lot of branches of engineering,

11:13.680 --> 11:15.800
 you do a worst case analysis

11:15.800 --> 11:18.720
 and make sure that your system will work in the worst case.

11:18.720 --> 11:20.400
 And then that guarantees that it'll work

11:20.400 --> 11:24.360
 in all of the messy average cases that happen

11:24.360 --> 11:27.440
 when you go out into a really randomized world.

11:27.440 --> 11:29.560
 Yeah, with driving with autonomous vehicles,

11:29.560 --> 11:31.840
 there seems to be a desire to just look

11:31.840 --> 11:34.880
 for think adversarially,

11:34.880 --> 11:36.920
 try to figure out how to mess up the system.

11:36.920 --> 11:40.640
 And if you can be robust to all those difficult cases,

11:40.640 --> 11:43.600
 then you can, it's a hand wavy empirical way

11:43.600 --> 11:45.800
 to show your system is safe.

11:45.800 --> 11:47.000
 Yeah, yeah.

11:47.000 --> 11:49.120
 Today, most adversarial example research

11:49.120 --> 11:51.640
 isn't really focused on a particular use case,

11:51.640 --> 11:54.000
 but there are a lot of different use cases

11:54.000 --> 11:55.080
 where you'd like to make sure

11:55.080 --> 11:57.720
 that the adversary can't interfere

11:57.720 --> 12:00.200
 with the operation of your system.

12:00.200 --> 12:01.040
 Like in finance,

12:01.040 --> 12:03.320
 if you have an algorithm making trades for you,

12:03.320 --> 12:04.640
 people go to a lot of an effort

12:04.640 --> 12:06.680
 to obfuscate their algorithm.

12:06.680 --> 12:08.080
 That's both to protect their IP

12:08.080 --> 12:10.880
 because you don't want to research

12:10.880 --> 12:13.560
 and develop a profitable trading algorithm

12:13.560 --> 12:16.120
 then have somebody else capture the gains.

12:16.120 --> 12:17.160
 But it's at least partly

12:17.160 --> 12:19.000
 because you don't want people to make adversarial

12:19.000 --> 12:21.240
 examples that fool your algorithm

12:21.240 --> 12:22.560
 into making bad trades.

12:24.360 --> 12:26.560
 Or I guess one area that's been popular

12:26.560 --> 12:30.160
 in the academic literature is speech recognition.

12:30.160 --> 12:34.400
 If you use speech recognition to hear an audio waveform

12:34.400 --> 12:37.680
 and then turn that into a command

12:37.680 --> 12:39.640
 that a phone executes for you,

12:39.640 --> 12:41.840
 you don't want a malicious adversary

12:41.840 --> 12:43.600
 to be able to produce audio

12:43.600 --> 12:46.280
 that gets interpreted as malicious commands,

12:46.280 --> 12:47.800
 especially if a human in the room

12:47.800 --> 12:50.320
 doesn't realize that something like that is happening.

12:50.320 --> 12:52.000
 In speech recognition,

12:52.000 --> 12:53.920
 has there been much success

12:53.920 --> 12:58.440
 in being able to create adversarial examples

12:58.440 --> 12:59.760
 that fool the system?

12:59.760 --> 13:00.880
 Yeah, actually.

13:00.880 --> 13:02.440
 I guess the first work that I'm aware of

13:02.440 --> 13:05.120
 is a paper called Hidden Voice Commands

13:05.120 --> 13:08.480
 that came out in 2016, I believe.

13:08.480 --> 13:09.560
 And they were able to show

13:09.560 --> 13:11.920
 that they could make sounds

13:11.920 --> 13:14.960
 that are not understandable by a human

13:14.960 --> 13:18.400
 but are recognized as the target phrase

13:18.400 --> 13:21.360
 that the attacker wants the phone to recognize it as.

13:21.360 --> 13:24.040
 Since then, things have gotten a little bit better

13:24.040 --> 13:27.600
 on the attacker side when worse on the defender side.

13:28.680 --> 13:33.360
 It's become possible to make sounds

13:33.360 --> 13:35.600
 that sound like normal speech

13:35.600 --> 13:39.000
 but are actually interpreted as a different sentence

13:39.000 --> 13:40.720
 than the human hears.

13:40.720 --> 13:42.720
 The level of perceptibility

13:42.720 --> 13:45.360
 of the adversarial perturbation is still kind of high.

13:46.600 --> 13:48.160
 When you listen to the recording,

13:48.160 --> 13:51.040
 it sounds like there's some noise in the background,

13:51.040 --> 13:52.960
 just like rustling sounds.

13:52.960 --> 13:54.360
 But those rustling sounds are actually

13:54.360 --> 13:55.560
 the adversarial perturbation

13:55.560 --> 13:58.040
 that makes the phone hear a completely different sentence.

13:58.040 --> 14:00.120
 Yeah, that's so fascinating.

14:00.120 --> 14:01.640
 Peter Norvig mentioned that you're writing

14:01.640 --> 14:04.280
 the deep learning chapter for the fourth edition

14:04.280 --> 14:05.840
 of the Artificial Intelligence,

14:05.840 --> 14:07.320
 the Modern Approach Book.

14:07.320 --> 14:10.680
 So how do you even begin summarizing

14:10.680 --> 14:12.960
 the field of deep learning in a chapter?

14:12.960 --> 14:16.840
 Well, in my case, I waited like a year

14:16.840 --> 14:19.080
 before I actually wrote anything.

14:19.080 --> 14:20.280
 Is it?

14:20.280 --> 14:22.600
 Even having written a full length textbook before,

14:22.600 --> 14:25.560
 it's still pretty intimidating

14:25.560 --> 14:27.800
 to try to start writing just one chapter

14:27.800 --> 14:29.040
 that covers everything.

14:31.080 --> 14:33.160
 One thing that helped me make that plan

14:33.160 --> 14:34.280
 was actually the experience

14:34.280 --> 14:36.680
 of having written the full book before

14:36.680 --> 14:39.080
 and then watching how the field changed

14:39.080 --> 14:40.920
 after the book came out.

14:40.920 --> 14:42.280
 I realized there's a lot of topics

14:42.280 --> 14:44.960
 that were maybe extraneous in the first book

14:44.960 --> 14:47.560
 and just seeing what stood the test

14:47.560 --> 14:49.360
 of a few years of being published

14:49.360 --> 14:52.160
 and what seems a little bit less important

14:52.160 --> 14:53.760
 to have included now helped me pare down

14:53.760 --> 14:55.920
 the topics I wanted to cover for the book.

14:56.840 --> 14:59.560
 It's also really nice now that the field

14:59.560 --> 15:00.920
 is kind of stabilized to the point

15:00.920 --> 15:04.720
 where some core ideas from the 1980s are still used today.

15:04.720 --> 15:06.640
 When I first started studying machine learning,

15:06.640 --> 15:09.520
 almost everything from the 1980s had been rejected

15:09.520 --> 15:11.320
 and now some of it has come back.

15:11.320 --> 15:13.440
 So that stuff that's really stood the test of time

15:13.440 --> 15:15.880
 is what I focused on putting into the book.

15:16.880 --> 15:21.240
 There's also, I guess, two different philosophies

15:21.240 --> 15:23.120
 about how you might write a book.

15:23.120 --> 15:24.760
 One philosophy is you try to write a reference

15:24.760 --> 15:26.160
 that covers everything.

15:26.160 --> 15:27.960
 The other philosophy is you try to provide

15:27.960 --> 15:30.320
 a high level summary that gives people

15:30.320 --> 15:32.360
 the language to understand a field

15:32.360 --> 15:34.920
 and tells them what the most important concepts are.

15:34.920 --> 15:37.080
 The first deep learning book that I wrote

15:37.080 --> 15:39.240
 with Joshua and Aaron was somewhere

15:39.240 --> 15:41.240
 between the two philosophies,

15:41.240 --> 15:43.640
 that it's trying to be both a reference

15:43.640 --> 15:45.760
 and an introductory guide.

15:45.760 --> 15:48.920
 Writing this chapter for Russell and Norvig's book,

15:48.920 --> 15:52.800
 I was able to focus more on just a concise introduction

15:52.800 --> 15:54.240
 of the key concepts and the language

15:54.240 --> 15:56.000
 you need to read about them more.

15:56.000 --> 15:57.560
 In a lot of cases, I actually just wrote paragraphs

15:57.560 --> 16:00.080
 that said, here's a rapidly evolving area

16:00.080 --> 16:02.400
 that you should pay attention to.

16:02.400 --> 16:04.760
 It's pointless to try to tell you what the latest

16:04.760 --> 16:09.760
 and best version of a learn to learn model is.

16:11.440 --> 16:13.640
 I can point you to a paper that's recent right now,

16:13.640 --> 16:16.880
 but there isn't a whole lot of a reason to delve

16:16.880 --> 16:20.440
 into exactly what's going on with the latest

16:20.440 --> 16:22.960
 learning to learn approach or the latest module

16:22.960 --> 16:24.960
 produced by a learning to learn algorithm.

16:24.960 --> 16:26.760
 You should know that learning to learn is a thing

16:26.760 --> 16:29.480
 and that it may very well be the source

16:29.480 --> 16:32.200
 of the latest and greatest convolutional net

16:32.200 --> 16:34.520
 or recurrent net module that you would want to use

16:34.520 --> 16:36.040
 in your latest project.

16:36.040 --> 16:38.200
 But there isn't a lot of point in trying to summarize

16:38.200 --> 16:42.280
 exactly which architecture and which learning approach

16:42.280 --> 16:44.040
 got to which level of performance.

16:44.040 --> 16:49.040
 So you maybe focus more on the basics of the methodology.

16:49.240 --> 16:52.480
 So from back propagation to feed forward

16:52.480 --> 16:55.160
 to recurrent networks, convolutional, that kind of thing.

16:55.160 --> 16:56.480
 Yeah, yeah.

16:56.480 --> 17:00.320
 So if I were to ask you, I remember I took algorithms

17:00.320 --> 17:03.720
 and data structures algorithms, of course.

17:03.720 --> 17:08.120
 I remember the professor asked, what is an algorithm?

17:09.200 --> 17:12.200
 And he yelled at everybody in a good way

17:12.200 --> 17:14.040
 that nobody was answering it correctly.

17:14.040 --> 17:16.360
 Everybody knew what the algorithm, it was graduate course.

17:16.360 --> 17:18.120
 Everybody knew what an algorithm was,

17:18.120 --> 17:19.800
 but they weren't able to answer it well.

17:19.800 --> 17:22.360
 So let me ask you, in that same spirit,

17:22.360 --> 17:23.580
 what is deep learning?

17:24.520 --> 17:29.520
 I would say deep learning is any kind of machine learning

17:29.520 --> 17:34.520
 that involves learning parameters of more than one

17:34.720 --> 17:36.020
 consecutive step.

17:37.280 --> 17:40.760
 So that, I mean, shallow learning is things where

17:40.760 --> 17:43.760
 you learn a lot of operations that happen in parallel.

17:43.760 --> 17:46.720
 You might have a system that makes multiple steps,

17:46.720 --> 17:51.000
 like you might have hand designed feature extractors,

17:51.000 --> 17:52.600
 but really only one step is learned.

17:52.600 --> 17:55.440
 Deep learning is anything where you have multiple

17:55.440 --> 17:56.880
 operations in sequence.

17:56.880 --> 17:59.400
 And that includes the things that are really popular

17:59.400 --> 18:01.280
 today, like convolutional networks

18:01.280 --> 18:04.640
 and recurrent networks, but it also includes some

18:04.640 --> 18:08.280
 of the things that have died out, like Bolton machines,

18:08.280 --> 18:10.880
 where we weren't using back propagation.

18:11.960 --> 18:14.240
 Today, I hear a lot of people define deep learning

18:14.240 --> 18:19.240
 as gradient descent applied to these differentiable

18:20.400 --> 18:24.240
 functions, and I think that's a legitimate usage

18:24.240 --> 18:25.920
 of the term, it's just different from the way

18:25.920 --> 18:27.800
 that I use the term myself.

18:27.800 --> 18:32.360
 So what's an example of deep learning that is not

18:32.360 --> 18:34.720
 gradient descent and differentiable functions?

18:34.720 --> 18:37.400
 In your, I mean, not specifically perhaps,

18:37.400 --> 18:39.760
 but more even looking into the future.

18:39.760 --> 18:44.300
 What's your thought about that space of approaches?

18:44.300 --> 18:46.340
 Yeah, so I tend to think of machine learning algorithms

18:46.340 --> 18:50.200
 as decomposed into really three different pieces.

18:50.200 --> 18:53.000
 There's the model, which can be something like a neural net

18:53.000 --> 18:56.600
 or a Bolton machine or a recurrent model.

18:56.600 --> 18:59.520
 And that basically just describes how do you take data

18:59.520 --> 19:03.480
 and how do you take parameters and what function do you use

19:03.480 --> 19:07.320
 to make a prediction given the data and the parameters?

19:07.320 --> 19:09.920
 Another piece of the learning algorithm is

19:09.920 --> 19:13.880
 the optimization algorithm, or not every algorithm

19:13.880 --> 19:15.920
 can be really described in terms of optimization,

19:15.920 --> 19:18.880
 but what's the algorithm for updating the parameters

19:18.880 --> 19:21.680
 or updating whatever the state of the network is?

19:22.600 --> 19:26.280
 And then the last part is the data set,

19:26.280 --> 19:29.200
 like how do you actually represent the world

19:29.200 --> 19:32.120
 as it comes into your machine learning system?

19:33.160 --> 19:35.800
 So I think of deep learning as telling us something

19:35.800 --> 19:39.040
 about what does the model look like?

19:39.040 --> 19:41.240
 And basically to qualify as deep,

19:41.240 --> 19:44.560
 I say that it just has to have multiple layers.

19:44.560 --> 19:47.360
 That can be multiple steps in a feed forward

19:47.360 --> 19:49.240
 differentiable computation.

19:49.240 --> 19:52.040
 That can be multiple layers in a graphical model.

19:52.040 --> 19:53.560
 There's a lot of ways that you could satisfy me

19:53.560 --> 19:56.160
 that something has multiple steps

19:56.160 --> 19:58.920
 that are each parameterized separately.

19:58.920 --> 20:00.640
 I think of gradient descent as being all about

20:00.640 --> 20:01.560
 that other piece,

20:01.560 --> 20:04.240
 the how do you actually update the parameters piece?

20:04.240 --> 20:05.960
 So you could imagine having a deep model

20:05.960 --> 20:08.680
 like a convolutional net and training it with something

20:08.680 --> 20:11.280
 like evolution or a genetic algorithm.

20:11.280 --> 20:14.640
 And I would say that still qualifies as deep learning.

20:14.640 --> 20:16.040
 And then in terms of models

20:16.040 --> 20:18.760
 that aren't necessarily differentiable,

20:18.760 --> 20:22.480
 I guess Bolton machines are probably the main example

20:22.480 --> 20:25.560
 of something where you can't really take a derivative

20:25.560 --> 20:28.000
 and use that for the learning process.

20:28.000 --> 20:32.320
 But you can still argue that the model has many steps

20:32.320 --> 20:33.760
 of processing that it applies

20:33.760 --> 20:35.800
 when you run inference in the model.

20:35.800 --> 20:38.960
 So it's the steps of processing that's key.

20:38.960 --> 20:41.320
 So Jeff Hinton suggests that we need to throw away

20:41.320 --> 20:44.960
 back propagation and start all over.

20:44.960 --> 20:46.520
 What do you think about that?

20:46.520 --> 20:48.600
 What could an alternative direction

20:48.600 --> 20:51.000
 of training neural networks look like?

20:51.000 --> 20:52.880
 I don't know that back propagation

20:52.880 --> 20:54.680
 is going to go away entirely.

20:54.680 --> 20:57.120
 Most of the time when we decide

20:57.120 --> 20:59.200
 that a machine learning algorithm

20:59.200 --> 21:03.440
 isn't on the critical path to research for improving AI,

21:03.440 --> 21:04.640
 the algorithm doesn't die,

21:04.640 --> 21:07.760
 it just becomes used for some specialized set of things.

21:08.760 --> 21:11.160
 A lot of algorithms like logistic regression

21:11.160 --> 21:14.000
 don't seem that exciting to AI researchers

21:14.000 --> 21:16.760
 who are working on things like speech recognition

21:16.760 --> 21:18.400
 or autonomous cars today,

21:18.400 --> 21:21.080
 but there's still a lot of use for logistic regression

21:21.080 --> 21:23.960
 and things like analyzing really noisy data

21:23.960 --> 21:25.640
 in medicine and finance

21:25.640 --> 21:28.720
 or making really rapid predictions

21:28.720 --> 21:30.680
 in really time limited contexts.

21:30.680 --> 21:33.440
 So I think back propagation and gradient descent

21:33.440 --> 21:34.520
 are around to stay,

21:34.520 --> 21:38.760
 but they may not end up being everything

21:38.760 --> 21:40.840
 that we need to get to real human level

21:40.840 --> 21:42.360
 or super human AI.

21:42.360 --> 21:44.680
 Are you optimistic about us discovering?

21:44.680 --> 21:49.680
 You know, back propagation has been around for a few decades.

21:50.240 --> 21:54.080
 So are you optimistic about us as a community

21:54.080 --> 21:56.800
 being able to discover something better?

21:56.800 --> 21:57.640
 Yeah, I am.

21:57.640 --> 22:01.840
 I think we likely will find something that works better.

22:01.840 --> 22:05.520
 You could imagine things like having stacks of models

22:05.520 --> 22:08.720
 where some of the lower level models predict parameters

22:08.720 --> 22:10.200
 of the higher level models.

22:10.200 --> 22:12.160
 And so at the top level,

22:12.160 --> 22:13.480
 you're not learning in terms of literally

22:13.480 --> 22:15.800
 calculating gradients, but just predicting

22:15.800 --> 22:17.680
 how different values will perform.

22:17.680 --> 22:19.560
 You can kind of see that already in some areas

22:19.560 --> 22:21.400
 like Bayesian optimization,

22:21.400 --> 22:22.960
 where you have a Gaussian process

22:22.960 --> 22:24.800
 that predicts how well different parameter values

22:24.800 --> 22:25.880
 will perform.

22:25.880 --> 22:27.680
 We already use those kinds of algorithms

22:27.680 --> 22:30.240
 for things like hyper parameter optimization.

22:30.240 --> 22:31.640
 And in general, we know a lot of things

22:31.640 --> 22:33.240
 other than back prop that work really well

22:33.240 --> 22:35.000
 for specific problems.

22:35.000 --> 22:38.240
 The main thing we haven't found is a way of taking one

22:38.240 --> 22:41.160
 of these other non back prop based algorithms

22:41.160 --> 22:43.520
 and having it really advance the state of the art

22:43.520 --> 22:46.160
 on an AI level problem.

22:46.160 --> 22:47.120
 Right.

22:47.120 --> 22:49.600
 But I wouldn't be surprised if eventually we find

22:49.600 --> 22:51.560
 that some of these algorithms that,

22:51.560 --> 22:52.760
 even the ones that already exist,

22:52.760 --> 22:54.200
 not even necessarily a new one,

22:54.200 --> 22:59.200
 we might find some way of customizing one of these algorithms

22:59.200 --> 23:00.560
 to do something really interesting

23:00.560 --> 23:05.240
 at the level of cognition or the level of,

23:06.400 --> 23:08.680
 I think one system that we really don't have working

23:08.680 --> 23:12.920
 quite right yet is like short term memory.

23:12.920 --> 23:14.480
 We have things like LSTMs,

23:14.480 --> 23:17.000
 they're called long short term memory.

23:17.000 --> 23:20.000
 They still don't do quite what a human does

23:20.000 --> 23:21.720
 with short term memory.

23:22.840 --> 23:26.920
 Like gradient descent to learn a specific fact

23:26.920 --> 23:29.360
 has to do multiple steps on that fact.

23:29.360 --> 23:34.120
 Like if I tell you, the meeting today is at 3pm,

23:34.120 --> 23:35.440
 I don't need to say over and over again.

23:35.440 --> 23:38.640
 It's at 3pm, it's at 3pm, it's at 3pm, it's at 3pm.

23:38.640 --> 23:40.400
 For you to do a gradient step on each one,

23:40.400 --> 23:43.160
 you just hear it once and you remember it.

23:43.160 --> 23:46.920
 There's been some work on things like self attention

23:46.920 --> 23:50.400
 and attention like mechanisms like the neural Turing machine

23:50.400 --> 23:53.160
 that can write to memory cells and update themselves

23:53.160 --> 23:54.880
 with facts like that right away.

23:54.880 --> 23:56.880
 But I don't think we've really nailed it yet.

23:56.880 --> 24:01.880
 And that's one area where I'd imagine that new optimization

24:02.080 --> 24:04.240
 algorithms or different ways of applying existing

24:04.240 --> 24:07.280
 optimization algorithms could give us a way

24:07.280 --> 24:10.120
 of just lightning fast updating the state

24:10.120 --> 24:12.400
 of a machine learning system to contain

24:12.400 --> 24:14.920
 a specific fact like that without needing to have it

24:14.920 --> 24:17.000
 presented over and over and over again.

24:17.000 --> 24:21.440
 So some of the success of symbolic systems in the 80s

24:21.440 --> 24:26.200
 is they were able to assemble these kinds of facts better.

24:26.200 --> 24:29.080
 But there's a lot of expert input required

24:29.080 --> 24:31.120
 and it's very limited in that sense.

24:31.120 --> 24:34.720
 Do you ever look back to that as something

24:34.720 --> 24:36.560
 that we'll have to return to eventually

24:36.560 --> 24:38.440
 sort of dust off the book from the shelf

24:38.440 --> 24:42.400
 and think about how we build knowledge, representation,

24:42.400 --> 24:43.240
 knowledge.

24:43.240 --> 24:44.840
 Like will we have to use graph searches?

24:44.840 --> 24:45.800
 Graph searches, right.

24:45.800 --> 24:47.720
 And like first order logic and entailment

24:47.720 --> 24:48.560
 and things like that.

24:48.560 --> 24:49.560
 That kind of thing, yeah, exactly.

24:49.560 --> 24:51.200
 In my particular line of work,

24:51.200 --> 24:54.560
 which has mostly been machine learning security

24:54.560 --> 24:56.720
 and also generative modeling,

24:56.720 --> 25:00.560
 I haven't usually found myself moving in that direction.

25:00.560 --> 25:03.520
 For generative models, I could see a little bit of,

25:03.520 --> 25:06.520
 it could be useful if you had something like a,

25:06.520 --> 25:09.680
 a differentiable knowledge base

25:09.680 --> 25:11.000
 or some other kind of knowledge base

25:11.000 --> 25:13.840
 where it's possible for some of our fuzzier

25:13.840 --> 25:16.880
 machine learning algorithms to interact with a knowledge base.

25:16.880 --> 25:19.040
 I mean, your network is kind of like that.

25:19.040 --> 25:21.440
 It's a differentiable knowledge base of sorts.

25:21.440 --> 25:22.280
 Yeah.

25:22.280 --> 25:27.280
 But if we had a really easy way of giving feedback

25:27.600 --> 25:29.240
 to machine learning models,

25:29.240 --> 25:32.400
 that would clearly help a lot with, with generative models.

25:32.400 --> 25:34.680
 And so you could imagine one way of getting there would be,

25:34.680 --> 25:36.720
 get a lot better at natural language processing.

25:36.720 --> 25:38.920
 But another way of getting there would be,

25:38.920 --> 25:40.280
 take some kind of knowledge base

25:40.280 --> 25:42.800
 and figure out a way for it to actually interact

25:42.800 --> 25:44.080
 with a neural network.

25:44.080 --> 25:46.080
 Being able to have a chat with a neural network.

25:46.080 --> 25:47.920
 Yeah.

25:47.920 --> 25:50.920
 So like one thing in generative models we see a lot today is,

25:50.920 --> 25:54.480
 you'll get things like faces that are not symmetrical.

25:54.480 --> 25:56.800
 Like, like people that have two eyes

25:56.800 --> 25:58.200
 that are different colors.

25:58.200 --> 25:59.560
 And I mean, there are people with eyes

25:59.560 --> 26:00.840
 that are different colors in real life,

26:00.840 --> 26:03.480
 but not nearly as many of them as you tend to see

26:03.480 --> 26:06.120
 in the machine learning generated data.

26:06.120 --> 26:08.120
 So if you had either a knowledge base

26:08.120 --> 26:10.200
 that could contain the fact,

26:10.200 --> 26:13.360
 people's faces are generally approximately symmetric

26:13.360 --> 26:15.920
 and eye color is especially likely

26:15.920 --> 26:17.920
 to be the same on both sides.

26:17.920 --> 26:20.160
 Being able to just inject that hint

26:20.160 --> 26:22.000
 into the machine learning model

26:22.000 --> 26:23.800
 without having to discover that itself

26:23.800 --> 26:25.760
 after studying a lot of data

26:25.760 --> 26:28.360
 would be a really useful feature.

26:28.360 --> 26:30.120
 I could see a lot of ways of getting there

26:30.120 --> 26:32.200
 without bringing back some of the 1980s technology,

26:32.200 --> 26:35.160
 but I also see some ways that you could imagine

26:35.160 --> 26:38.240
 extending the 1980s technology to play nice with neural nets

26:38.240 --> 26:40.040
 and have it help get there.

26:40.040 --> 26:40.880
 Awesome.

26:40.880 --> 26:44.360
 So you talked about the story of you coming up

26:44.360 --> 26:47.040
 with the idea of GANs at a bar with some friends.

26:47.040 --> 26:50.400
 You were arguing that this, you know,

26:50.400 --> 26:53.080
 GANs would work generative adversarial networks

26:53.080 --> 26:54.680
 and the others didn't think so.

26:54.680 --> 26:58.400
 Then you went home at midnight, coded up and it worked.

26:58.400 --> 27:01.320
 So if I was a friend of yours at the bar,

27:01.320 --> 27:02.720
 I would also have doubts.

27:02.720 --> 27:03.880
 It's a really nice idea,

27:03.880 --> 27:06.800
 but I'm very skeptical that it would work.

27:06.800 --> 27:09.280
 What was the basis of their skepticism?

27:09.280 --> 27:13.200
 What was the basis of your intuition why it should work?

27:14.360 --> 27:16.840
 I don't wanna be someone who goes around promoting alcohol

27:16.840 --> 27:18.280
 for the purposes of science,

27:18.280 --> 27:21.040
 but in this case, I do actually think

27:21.040 --> 27:23.080
 that drinking helped a little bit.

27:23.080 --> 27:25.360
 When your inhibitions are lowered,

27:25.360 --> 27:27.400
 you're more willing to try out things

27:27.400 --> 27:29.640
 that you wouldn't try out otherwise.

27:29.640 --> 27:32.480
 So I have noticed in general

27:32.480 --> 27:34.560
 that I'm less prone to shooting down some of my own ideas

27:34.560 --> 27:37.960
 when I have had a little bit to drink.

27:37.960 --> 27:40.800
 I think if I had had that idea at lunchtime,

27:40.800 --> 27:42.280
 I probably would have thought it.

27:42.280 --> 27:43.720
 It's hard enough to train one neural net.

27:43.720 --> 27:44.880
 You can't train a second neural net

27:44.880 --> 27:48.080
 in the inner loop of the outer neural net.

27:48.080 --> 27:49.800
 That was basically my friend's objection

27:49.800 --> 27:52.720
 was that trying to train two neural nets at the same time

27:52.720 --> 27:54.280
 would be too hard.

27:54.280 --> 27:56.120
 So it was more about the training process

27:56.120 --> 28:01.120
 unless, so my skepticism would be, I'm sure you could train it

28:01.160 --> 28:03.200
 but the thing would converge to

28:03.200 --> 28:05.840
 would not be able to generate anything reasonable

28:05.840 --> 28:08.240
 and any kind of reasonable realism.

28:08.240 --> 28:11.360
 Yeah, so part of what all of us were thinking about

28:11.360 --> 28:15.280
 when we had this conversation was deep Bolton machines,

28:15.280 --> 28:17.000
 which a lot of us in the lab, including me,

28:17.000 --> 28:19.480
 were a big fan of deep Bolton machines at the time.

28:20.640 --> 28:24.240
 They involved two separate processes running at the same time.

28:24.240 --> 28:27.400
 One of them is called the positive phase

28:27.400 --> 28:30.440
 where you load data into the model

28:30.440 --> 28:32.920
 and tell the model to make the data more likely.

28:32.920 --> 28:34.480
 The other one is called the negative phase

28:34.480 --> 28:36.280
 where you draw samples from the model

28:36.280 --> 28:38.880
 and tell the model to make those samples less likely.

28:40.480 --> 28:42.400
 In a deep Bolton machine, it's not trivial

28:42.400 --> 28:43.320
 to generate a sample.

28:43.320 --> 28:46.280
 You have to actually run an iterative process

28:46.280 --> 28:48.520
 that gets better and better samples

28:48.520 --> 28:50.720
 coming closer and closer to the distribution

28:50.720 --> 28:52.120
 the model represents.

28:52.120 --> 28:53.240
 So during the training process,

28:53.240 --> 28:56.560
 you're always running these two systems at the same time.

28:56.560 --> 28:58.360
 One that's updating the parameters of the model

28:58.360 --> 28:59.880
 and another one that's trying to generate samples

28:59.880 --> 29:01.120
 from the model.

29:01.120 --> 29:03.720
 And they worked really well on things like MNIST,

29:03.720 --> 29:05.200
 but a lot of us in the lab, including me,

29:05.200 --> 29:08.840
 had tried to get deep Bolton machines to scale past MNIST

29:08.840 --> 29:11.320
 to things like generating color photos,

29:11.320 --> 29:13.480
 and we just couldn't get the two processes

29:13.480 --> 29:15.360
 to stay synchronized.

29:16.720 --> 29:18.120
 So when I had the idea for GANs,

29:18.120 --> 29:19.720
 a lot of people thought that the discriminator

29:19.720 --> 29:21.960
 would have more or less the same problem

29:21.960 --> 29:25.360
 as the negative phase in the Bolton machine,

29:25.360 --> 29:27.840
 that trying to train the discriminator in the inner loop,

29:27.840 --> 29:29.960
 you just couldn't get it to keep up

29:29.960 --> 29:31.560
 with the generator in the outer loop.

29:31.560 --> 29:33.360
 And that would prevent it from

29:33.360 --> 29:35.240
 converging to anything useful.

29:35.240 --> 29:36.880
 Yeah, I share that intuition.

29:36.880 --> 29:37.720
 Yeah.

29:39.560 --> 29:42.000
 But turns out to not be the case.

29:42.000 --> 29:43.800
 A lot of the time with machine learning algorithms,

29:43.800 --> 29:45.200
 it's really hard to predict ahead of time

29:45.200 --> 29:46.960
 how well they'll actually perform.

29:46.960 --> 29:48.160
 You have to just run the experiment

29:48.160 --> 29:49.200
 and see what happens.

29:49.200 --> 29:53.480
 And I would say I still today don't have like one factor

29:53.480 --> 29:54.840
 I can put my finger on and say,

29:54.840 --> 29:58.360
 this is why GANs worked for photo generation

29:58.360 --> 30:00.240
 and deep Bolton machines don't.

30:02.000 --> 30:04.560
 There are a lot of theory papers showing that

30:04.560 --> 30:06.400
 under some theoretical settings,

30:06.400 --> 30:09.640
 the GAN algorithm does actually converge.

30:10.720 --> 30:14.200
 But those settings are restricted enough

30:14.200 --> 30:17.560
 that they don't necessarily explain the whole picture

30:17.560 --> 30:20.760
 in terms of all the results that we see in practice.

30:20.760 --> 30:22.360
 So taking a step back,

30:22.360 --> 30:24.880
 can you, in the same way as we talked about deep learning,

30:24.880 --> 30:28.440
 can you tell me what generative adversarial networks are?

30:29.480 --> 30:31.400
 Yeah, so generative adversarial networks

30:31.400 --> 30:34.000
 are a particular kind of generative model.

30:34.000 --> 30:36.320
 A generative model is a machine learning model

30:36.320 --> 30:38.880
 that can train on some set of data.

30:38.880 --> 30:41.280
 Like say you have a collection of photos of cats

30:41.280 --> 30:44.040
 and you want to generate more photos of cats,

30:44.040 --> 30:47.120
 or you want to estimate a probability distribution

30:47.120 --> 30:49.840
 over cats so you can ask how likely it is

30:49.840 --> 30:51.840
 that some new image is a photo of a cat.

30:52.920 --> 30:55.840
 GANs are one way of doing this.

30:55.840 --> 30:59.200
 Some generative models are good at creating new data.

30:59.200 --> 31:00.840
 Other generative models are good

31:00.840 --> 31:02.600
 at estimating that density function

31:02.600 --> 31:06.600
 and telling you how likely particular pieces of data are

31:06.600 --> 31:09.760
 to come from the same distribution as the training data.

31:09.760 --> 31:12.440
 GANs are more focused on generating samples

31:12.440 --> 31:15.640
 rather than estimating the density function.

31:15.640 --> 31:17.720
 There are some kinds of GANs, like flow GAN,

31:17.720 --> 31:18.560
 that can do both,

31:18.560 --> 31:21.680
 but mostly GANs are about generating samples,

31:21.680 --> 31:24.240
 generating new photos of cats that look realistic.

31:25.240 --> 31:29.360
 And they do that completely from scratch.

31:29.360 --> 31:32.240
 It's analogous to human imagination

31:32.240 --> 31:34.760
 when a GAN creates a new image of a cat.

31:34.760 --> 31:39.320
 It's using a neural network to produce a cat

31:39.320 --> 31:41.040
 that has not existed before.

31:41.040 --> 31:44.560
 It isn't doing something like compositing photos together.

31:44.560 --> 31:47.080
 You're not literally taking the eye off of one cat

31:47.080 --> 31:49.000
 and the ear off of another cat.

31:49.000 --> 31:51.320
 It's more of this digestive process

31:51.320 --> 31:53.920
 where the neural net trains in a lot of data

31:53.920 --> 31:55.560
 and comes up with some representation

31:55.560 --> 31:57.360
 of the probability distribution

31:57.360 --> 31:59.760
 and generates entirely new cats.

31:59.760 --> 32:00.880
 There are a lot of different ways

32:00.880 --> 32:01.960
 of building a generative model.

32:01.960 --> 32:05.640
 What's specific to GANs is that we have a two player game

32:05.640 --> 32:08.080
 in the game theoretic sense.

32:08.080 --> 32:10.280
 And as the players in this game compete,

32:10.280 --> 32:13.920
 one of them becomes able to generate realistic data.

32:13.920 --> 32:16.120
 The first player is called the generator.

32:16.120 --> 32:20.640
 It produces output data, such as just images, for example.

32:20.640 --> 32:22.400
 And at the start of the learning process,

32:22.400 --> 32:25.120
 it'll just produce completely random images.

32:25.120 --> 32:27.360
 The other player is called the discriminator.

32:27.360 --> 32:29.680
 The discriminator takes images as input

32:29.680 --> 32:31.560
 and guesses whether they're real or fake.

32:32.480 --> 32:34.200
 You train it both on real data,

32:34.200 --> 32:36.120
 so photos that come from your training set,

32:36.120 --> 32:37.840
 actual photos of cats.

32:37.840 --> 32:39.880
 And you try to say that those are real.

32:39.880 --> 32:41.920
 You also train it on images

32:41.920 --> 32:43.840
 that come from the generator network.

32:43.840 --> 32:46.720
 And you train it to say that those are fake.

32:46.720 --> 32:49.200
 As the two players compete in this game,

32:49.200 --> 32:50.920
 the discriminator tries to become better

32:50.920 --> 32:53.280
 at recognizing whether images are real or fake.

32:53.280 --> 32:54.760
 And the generator becomes better

32:54.760 --> 32:56.960
 at fooling the discriminator into thinking

32:56.960 --> 32:59.560
 that its outputs are real.

33:00.760 --> 33:03.560
 And you can analyze this through the language of game theory

33:03.560 --> 33:06.920
 and find that there's a Nash equilibrium

33:06.920 --> 33:08.600
 where the generator has captured

33:08.600 --> 33:10.800
 the correct probability distribution.

33:10.800 --> 33:12.160
 So in the cat example,

33:12.160 --> 33:14.560
 it makes perfectly realistic cat photos.

33:14.560 --> 33:17.160
 And the discriminator is unable to do better

33:17.160 --> 33:18.720
 than random guessing,

33:18.720 --> 33:21.800
 because all the samples coming from both the data

33:21.800 --> 33:24.000
 and the generator look equally likely

33:24.000 --> 33:25.840
 to have come from either source.

33:25.840 --> 33:28.320
 So do you ever sit back

33:28.320 --> 33:31.280
 and does it just blow your mind that this thing works?

33:31.280 --> 33:35.840
 So from very, so it's able to estimate the density function

33:35.840 --> 33:38.640
 enough to generate realistic images.

33:38.640 --> 33:43.640
 I mean, yeah, do you ever sit back and think,

33:43.640 --> 33:46.760
 how does this even, this is quite incredible,

33:46.760 --> 33:49.280
 especially where against have gone in terms of realism.

33:49.280 --> 33:51.600
 Yeah, and not just to flatter my own work,

33:51.600 --> 33:53.840
 but generative models,

33:53.840 --> 33:55.400
 all of them have this property

33:55.400 --> 33:58.800
 that if they really did what we asked them to do,

33:58.800 --> 34:01.040
 they would do nothing but memorize the training data.

34:01.040 --> 34:02.920
 Right, exactly.

34:02.920 --> 34:05.720
 Models that are based on maximizing the likelihood,

34:05.720 --> 34:08.200
 the way that you obtain the maximum likelihood

34:08.200 --> 34:09.720
 for a specific training set

34:09.720 --> 34:12.440
 is you assign all of your probability mass

34:12.440 --> 34:15.120
 to the training examples and nowhere else.

34:15.120 --> 34:18.440
 For GANs, the game is played using a training set.

34:18.440 --> 34:21.160
 So the way that you become unbeatable in the game

34:21.160 --> 34:23.440
 is you literally memorize training examples.

34:25.360 --> 34:28.880
 One of my former interns wrote a paper,

34:28.880 --> 34:31.040
 his name is Vaishnav Nagarajan,

34:31.040 --> 34:33.080
 and he showed that it's actually hard

34:33.080 --> 34:36.120
 for the generator to memorize the training data,

34:36.120 --> 34:39.160
 hard in a statistical learning theory sense,

34:39.160 --> 34:42.200
 that you can actually create reasons

34:42.200 --> 34:47.200
 for why it would require quite a lot of learning steps

34:48.400 --> 34:52.200
 and a lot of observations of different latent variables

34:52.200 --> 34:54.360
 before you could memorize the training data.

34:54.360 --> 34:55.680
 That still doesn't really explain

34:55.680 --> 34:58.280
 why when you produce samples that are new,

34:58.280 --> 34:59.880
 why do you get compelling images

34:59.880 --> 35:02.400
 rather than just garbage that's different

35:02.400 --> 35:03.800
 from the training set.

35:03.800 --> 35:06.960
 And I don't think we really have a good answer for that,

35:06.960 --> 35:07.920
 especially if you think about

35:07.920 --> 35:10.240
 how many possible images are out there

35:10.240 --> 35:15.240
 and how few images the generative model sees during training.

35:15.440 --> 35:16.920
 It seems just unreasonable

35:16.920 --> 35:19.200
 that generative models create new images

35:19.200 --> 35:22.080
 as well as they do, especially considering

35:22.080 --> 35:23.760
 that we're basically training them to memorize

35:23.760 --> 35:25.000
 rather than generalize.

35:26.240 --> 35:28.920
 I think part of the answer is there's a paper

35:28.920 --> 35:31.480
 called Deep Image Prior where they show

35:31.480 --> 35:33.080
 that you can take a convolutional net

35:33.080 --> 35:35.000
 and you don't even need to learn the parameters of it at all.

35:35.000 --> 35:37.640
 You just use the model architecture.

35:37.640 --> 35:41.080
 And it's already useful for things like in painting images.

35:41.080 --> 35:43.760
 I think that shows us that the convolutional network

35:43.760 --> 35:45.880
 architecture captures something really important

35:45.880 --> 35:47.960
 about the structure of images.

35:47.960 --> 35:50.960
 And we don't need to actually use learning

35:50.960 --> 35:52.200
 to capture all the information

35:52.200 --> 35:54.000
 coming out of the convolutional net.

35:55.240 --> 35:58.400
 That would imply that it would be much harder

35:58.400 --> 36:01.240
 to make generative models in other domains.

36:01.240 --> 36:03.600
 So far, we're able to make reasonable speech models

36:03.600 --> 36:04.880
 and things like that.

36:04.880 --> 36:07.440
 But to be honest, we haven't actually explored

36:07.440 --> 36:09.800
 a whole lot of different data sets all that much.

36:09.800 --> 36:13.920
 We don't, for example, see a lot of deep learning models

36:13.920 --> 36:18.440
 of like biology data sets

36:18.440 --> 36:19.880
 where you have lots of microarrays

36:19.880 --> 36:22.240
 measuring the amount of different enzymes

36:22.240 --> 36:23.080
 and things like that.

36:23.080 --> 36:25.240
 So we may find that some of the progress

36:25.240 --> 36:27.360
 that we've seen for images and speech turns out

36:27.360 --> 36:30.120
 to really rely heavily on the model architecture.

36:30.120 --> 36:32.960
 And we were able to do what we did for vision

36:32.960 --> 36:36.080
 by trying to reverse engineer the human visual system.

36:37.040 --> 36:39.800
 And maybe it'll turn out that we can't just

36:39.800 --> 36:42.560
 use that same trick for arbitrary kinds of data.

36:43.480 --> 36:45.920
 Right, so there's aspect of the human vision system,

36:45.920 --> 36:49.280
 the hardware of it that makes it,

36:49.280 --> 36:51.120
 without learning, without cognition,

36:51.120 --> 36:53.640
 just makes it really effective at detecting the patterns

36:53.640 --> 36:54.960
 we see in the visual world.

36:54.960 --> 36:57.280
 Yeah, that's really interesting.

36:57.280 --> 37:02.280
 What, in a big quick overview in your view,

37:04.640 --> 37:06.280
 what types of GANs are there

37:06.280 --> 37:10.080
 and what other generative models besides GANs are there?

37:10.080 --> 37:13.360
 Yeah, so it's maybe a little bit easier to start

37:13.360 --> 37:14.640
 with what kinds of generative models

37:14.640 --> 37:15.920
 are there other than GANs.

37:16.840 --> 37:20.840
 So most generative models are likelihood based

37:20.840 --> 37:23.920
 where to train them, you have a model

37:23.920 --> 37:27.320
 that tells you how much probability it assigns

37:27.320 --> 37:29.080
 to a particular example,

37:29.080 --> 37:31.480
 and you just maximize the probability assigned

37:31.480 --> 37:33.680
 to all the training examples.

37:33.680 --> 37:36.200
 It turns out that it's hard to design a model

37:36.200 --> 37:39.200
 that can create really complicated images

37:39.200 --> 37:42.280
 or really complicated audio waveforms

37:42.280 --> 37:46.200
 and still have it be possible to estimate

37:46.200 --> 37:51.200
 the likelihood function from a computational point of view.

37:51.200 --> 37:53.200
 Most interesting models that you would just write

37:53.200 --> 37:56.200
 down intuitively, it turns out that it's almost impossible

37:56.200 --> 37:58.200
 to calculate the amount of probability

37:58.200 --> 38:00.200
 they assign to a particular point.

38:00.200 --> 38:04.200
 So there's a few different schools of generative models

38:04.200 --> 38:06.200
 in the likelihood family.

38:06.200 --> 38:09.200
 One approach is to very carefully design the model

38:09.200 --> 38:12.200
 so that it is computationally tractable

38:12.200 --> 38:15.200
 to measure the density it assigns to a particular point.

38:15.200 --> 38:18.200
 So there are things like auto regressive models,

38:18.200 --> 38:23.200
 like pixel CNN, those basically break down

38:23.200 --> 38:26.200
 the probability distribution into a product

38:26.200 --> 38:28.200
 over every single feature.

38:28.200 --> 38:32.200
 So for an image, you estimate the probability of each pixel

38:32.200 --> 38:35.200
 given all of the pixels that came before it.

38:35.200 --> 38:37.200
 There's tricks where if you want to measure

38:37.200 --> 38:40.200
 the density function, you can actually calculate

38:40.200 --> 38:43.200
 the density for all these pixels more or less in parallel.

38:44.200 --> 38:46.200
 Generating the image still tends to require you

38:46.200 --> 38:50.200
 to go one pixel at a time, and that can be very slow.

38:50.200 --> 38:52.200
 But there are, again, tricks for doing this

38:52.200 --> 38:54.200
 in a hierarchical pattern where you can keep

38:54.200 --> 38:56.200
 the runtime under control.

38:56.200 --> 38:59.200
 Are the quality of the images it generates

38:59.200 --> 39:02.200
 putting runtime aside pretty good?

39:02.200 --> 39:04.200
 They're reasonable, yeah.

39:04.200 --> 39:07.200
 I would say a lot of the best results

39:07.200 --> 39:10.200
 are from GANs these days, but it can be hard to tell

39:10.200 --> 39:14.200
 how much of that is based on who's studying

39:14.200 --> 39:17.200
 which type of algorithm, if that makes sense.

39:17.200 --> 39:19.200
 The amount of effort invested in it.

39:19.200 --> 39:21.200
 Yeah, or the kind of expertise.

39:21.200 --> 39:23.200
 So a lot of people who've traditionally been excited

39:23.200 --> 39:25.200
 about graphics or art and things like that

39:25.200 --> 39:27.200
 have gotten interested in GANs.

39:27.200 --> 39:29.200
 And to some extent, it's hard to tell,

39:29.200 --> 39:32.200
 are GANs doing better because they have a lot of

39:32.200 --> 39:34.200
 graphics and art experts behind them?

39:34.200 --> 39:36.200
 Or are GANs doing better because

39:36.200 --> 39:38.200
 they're more computationally efficient?

39:38.200 --> 39:40.200
 Or are GANs doing better because

39:40.200 --> 39:43.200
 they prioritize the realism of samples

39:43.200 --> 39:45.200
 over the accuracy of the density function?

39:45.200 --> 39:47.200
 I think all of those are potentially

39:47.200 --> 39:51.200
 valid explanations, and it's hard to tell.

39:51.200 --> 39:53.200
 So can you give a brief history of GANs

39:53.200 --> 39:59.200
 from 2014 with Paper 13?

39:59.200 --> 40:01.200
 Yeah, so a few highlights.

40:01.200 --> 40:03.200
 In the first paper, we just showed that

40:03.200 --> 40:05.200
 GANs basically work.

40:05.200 --> 40:07.200
 If you look back at the samples we had now,

40:07.200 --> 40:09.200
 they look terrible.

40:09.200 --> 40:11.200
 On the CFAR 10 data set, you can't even

40:11.200 --> 40:13.200
 see the effects in them.

40:13.200 --> 40:15.200
 Your paper, sorry, you used CFAR 10?

40:15.200 --> 40:17.200
 We used MNIST, which is Little Handwritten Digits.

40:17.200 --> 40:19.200
 We used the Toronto Face Database,

40:19.200 --> 40:22.200
 which is small grayscale photos of faces.

40:22.200 --> 40:24.200
 We did have recognizable faces.

40:24.200 --> 40:26.200
 My colleague Bing Xu put together

40:26.200 --> 40:29.200
 the first GAN face model for that paper.

40:29.200 --> 40:32.200
 We also had the CFAR 10 data set,

40:32.200 --> 40:35.200
 which is things like very small 32x32 pixels

40:35.200 --> 40:40.200
 of cars and cats and dogs.

40:40.200 --> 40:43.200
 For that, we didn't get recognizable objects,

40:43.200 --> 40:46.200
 but all the deep learning people back then

40:46.200 --> 40:48.200
 were really used to looking at these failed samples

40:48.200 --> 40:50.200
 and kind of reading them like tea leaves.

40:50.200 --> 40:53.200
 And people who are used to reading the tea leaves

40:53.200 --> 40:56.200
 recognize that our tea leaves at least look different.

40:56.200 --> 40:58.200
 Maybe not necessarily better,

40:58.200 --> 41:01.200
 but there was something unusual about them.

41:01.200 --> 41:03.200
 And that got a lot of us excited.

41:03.200 --> 41:06.200
 One of the next really big steps was LAPGAN

41:06.200 --> 41:10.200
 by Emily Denton and Sumith Chintala at Facebook AI Research,

41:10.200 --> 41:14.200
 where they actually got really good high resolution photos

41:14.200 --> 41:16.200
 working with GANs for the first time.

41:16.200 --> 41:18.200
 They had a complicated system

41:18.200 --> 41:20.200
 where they generated the image starting at low res

41:20.200 --> 41:22.200
 and then scaling up to high res,

41:22.200 --> 41:24.200
 but they were able to get it to work.

41:24.200 --> 41:30.200
 And then in 2015, I believe later that same year,

41:30.200 --> 41:35.200
 Alec Radford and Sumith Chintala and Luke Metz

41:35.200 --> 41:38.200
 published the DC GAN paper,

41:38.200 --> 41:41.200
 which it stands for Deep Convolutional GAN.

41:41.200 --> 41:43.200
 It's kind of a nonunique name

41:43.200 --> 41:46.200
 because these days basically all GANs

41:46.200 --> 41:48.200
 and even some before that were deep and convolutional,

41:48.200 --> 41:52.200
 but they just kind of picked a name for a really great recipe

41:52.200 --> 41:55.200
 where they were able to actually using only one model

41:55.200 --> 41:57.200
 instead of a multi step process,

41:57.200 --> 42:01.200
 actually generate realistic images of faces and things like that.

42:01.200 --> 42:05.200
 That was sort of like the beginning

42:05.200 --> 42:07.200
 of the Cambrian explosion of GANs.

42:07.200 --> 42:09.200
 Once you had animals that had a backbone,

42:09.200 --> 42:12.200
 you suddenly got lots of different versions of fish

42:12.200 --> 42:15.200
 and four legged animals and things like that.

42:15.200 --> 42:17.200
 So DC GAN became kind of the backbone

42:17.200 --> 42:19.200
 for many different models that came out.

42:19.200 --> 42:21.200
 Used as a baseline even still.

42:21.200 --> 42:23.200
 Yeah, yeah.

42:23.200 --> 42:26.200
 And so from there, I would say some interesting things we've seen

42:26.200 --> 42:30.200
 are there's a lot you can say about how just

42:30.200 --> 42:33.200
 the quality of standard image generation GANs has increased,

42:33.200 --> 42:36.200
 but what's also maybe more interesting on an intellectual level

42:36.200 --> 42:40.200
 is how the things you can use GANs for has also changed.

42:40.200 --> 42:44.200
 One thing is that you can use them to learn classifiers

42:44.200 --> 42:47.200
 without having to have class labels for every example

42:47.200 --> 42:49.200
 in your training set.

42:49.200 --> 42:51.200
 So that's called semi supervised learning.

42:51.200 --> 42:55.200
 My colleague at OpenAI, Tim Solomon, who's at Brain now,

42:55.200 --> 42:57.200
 wrote a paper called

42:57.200 --> 42:59.200
 Improved Techniques for Training GANs.

42:59.200 --> 43:01.200
 I'm a coauthor on this paper,

43:01.200 --> 43:03.200
 but I can't claim any credit for this particular part.

43:03.200 --> 43:05.200
 One thing he showed on the paper is that

43:05.200 --> 43:09.200
 you can take the GAN discriminator and use it as a classifier

43:09.200 --> 43:12.200
 that actually tells you this image is a cat,

43:12.200 --> 43:14.200
 this image is a dog, this image is a car,

43:14.200 --> 43:16.200
 this image is a truck.

43:16.200 --> 43:18.200
 And so not just to say whether the image is real or fake,

43:18.200 --> 43:22.200
 but if it is real to say specifically what kind of object it is.

43:22.200 --> 43:25.200
 And he found that you can train these classifiers

43:25.200 --> 43:28.200
 with far fewer labeled examples

43:28.200 --> 43:30.200
 than traditional classifiers.

43:30.200 --> 43:33.200
 So if you supervise based on also

43:33.200 --> 43:35.200
 not just your discrimination ability,

43:35.200 --> 43:37.200
 but your ability to classify,

43:37.200 --> 43:40.200
 you're going to converge much faster

43:40.200 --> 43:43.200
 to being effective at being a discriminator.

43:43.200 --> 43:44.200
 Yeah.

43:44.200 --> 43:46.200
 So for example, for the MNIST dataset,

43:46.200 --> 43:49.200
 you want to look at an image of a handwritten digit

43:49.200 --> 43:53.200
 and say whether it's a zero, a one, or two, and so on.

43:53.200 --> 43:57.200
 To get down to less than 1% accuracy,

43:57.200 --> 44:00.200
 we required around 60,000 examples

44:00.200 --> 44:03.200
 until maybe about 2014 or so.

44:03.200 --> 44:07.200
 In 2016, with this semi supervised GAN project,

44:07.200 --> 44:10.200
 Tim was able to get below 1% error

44:10.200 --> 44:13.200
 using only 100 labeled examples.

44:13.200 --> 44:16.200
 So that was about a 600x decrease

44:16.200 --> 44:18.200
 in the amount of labels that he needed.

44:18.200 --> 44:21.200
 He's still using more images than that,

44:21.200 --> 44:23.200
 but he doesn't need to have each of them labeled as,

44:23.200 --> 44:25.200
 you know, this one's a one, this one's a two,

44:25.200 --> 44:27.200
 this one's a zero, and so on.

44:27.200 --> 44:29.200
 Then to be able to, for GANs,

44:29.200 --> 44:31.200
 to be able to generate recognizable objects,

44:31.200 --> 44:33.200
 so objects from a particular class,

44:33.200 --> 44:36.200
 you still need labeled data,

44:36.200 --> 44:38.200
 because you need to know

44:38.200 --> 44:41.200
 what it means to be a particular class cat dog.

44:41.200 --> 44:44.200
 How do you think we can move away from that?

44:44.200 --> 44:46.200
 Yeah, some researchers at Brain Zurich

44:46.200 --> 44:49.200
 actually just released a really great paper

44:49.200 --> 44:51.200
 on semi supervised GANs,

44:51.200 --> 44:54.200
 where their goal isn't to classify,

44:54.200 --> 44:56.200
 to make recognizable objects

44:56.200 --> 44:58.200
 despite not having a lot of labeled data.

44:58.200 --> 45:02.200
 They were working off of DeepMind's BigGAN project,

45:02.200 --> 45:04.200
 and they showed that they can match

45:04.200 --> 45:06.200
 the performance of BigGAN

45:06.200 --> 45:10.200
 using only 10%, I believe, of the labels.

45:10.200 --> 45:12.200
 BigGAN was trained on the ImageNet data set,

45:12.200 --> 45:14.200
 which is about 1.2 million images,

45:14.200 --> 45:17.200
 and had all of them labeled.

45:17.200 --> 45:19.200
 This latest project from Brain Zurich

45:19.200 --> 45:21.200
 shows that they're able to get away with

45:21.200 --> 45:25.200
 having about 10% of the images labeled.

45:25.200 --> 45:29.200
 They do that essentially using a clustering algorithm,

45:29.200 --> 45:32.200
 where the discriminator learns to assign

45:32.200 --> 45:34.200
 the objects to groups,

45:34.200 --> 45:38.200
 and then this understanding that objects can be grouped

45:38.200 --> 45:40.200
 into similar types,

45:40.200 --> 45:43.200
 helps it to form more realistic ideas

45:43.200 --> 45:45.200
 of what should be appearing in the image,

45:45.200 --> 45:47.200
 because it knows that every image it creates

45:47.200 --> 45:50.200
 has to come from one of these archetypal groups,

45:50.200 --> 45:53.200
 rather than just being some arbitrary image.

45:53.200 --> 45:55.200
 If you train again with no class labels,

45:55.200 --> 45:57.200
 you tend to get things that look sort of like

45:57.200 --> 46:00.200
 grass or water or brick or dirt,

46:00.200 --> 46:04.200
 but without necessarily a lot going on in them.

46:04.200 --> 46:06.200
 I think that's partly because if you look

46:06.200 --> 46:08.200
 at a large ImageNet image,

46:08.200 --> 46:11.200
 the object doesn't necessarily occupy the whole image,

46:11.200 --> 46:15.200
 and so you learn to create realistic sets of pixels,

46:15.200 --> 46:17.200
 but you don't necessarily learn

46:17.200 --> 46:19.200
 that the object is the star of the show,

46:19.200 --> 46:22.200
 and you want it to be in every image you make.

46:22.200 --> 46:25.200
 Yeah, I've heard you talk about the horse,

46:25.200 --> 46:27.200
 the zebra cycle, gang mapping,

46:27.200 --> 46:30.200
 and how it turns out, again,

46:30.200 --> 46:33.200
 thought provoking that horses are usually on grass,

46:33.200 --> 46:35.200
 and zebras are usually on drier terrain,

46:35.200 --> 46:38.200
 so when you're doing that kind of generation,

46:38.200 --> 46:43.200
 you're going to end up generating greener horses or whatever.

46:43.200 --> 46:45.200
 So those are connected together.

46:45.200 --> 46:46.200
 It's not just...

46:46.200 --> 46:47.200
 Yeah, yeah.

46:47.200 --> 46:49.200
 You're not able to segment,

46:49.200 --> 46:52.200
 to be able to generate in a segmental way.

46:52.200 --> 46:55.200
 So are there other types of games you come across

46:55.200 --> 47:00.200
 in your mind that neural networks can play with each other

47:00.200 --> 47:05.200
 to be able to solve problems?

47:05.200 --> 47:09.200
 Yeah, the one that I spend most of my time on is in security.

47:09.200 --> 47:13.200
 You can model most interactions as a game

47:13.200 --> 47:16.200
 where there's attackers trying to break your system

47:16.200 --> 47:19.200
 or the defender trying to build a resilient system.

47:19.200 --> 47:22.200
 There's also domain adversarial learning,

47:22.200 --> 47:25.200
 which is an approach to domain adaptation

47:25.200 --> 47:27.200
 that looks really a lot like GANs.

47:27.200 --> 47:31.200
 The authors had the idea before the GAN paper came out.

47:31.200 --> 47:33.200
 Their paper came out a little bit later,

47:33.200 --> 47:38.200
 and they were very nice and cited the GAN paper,

47:38.200 --> 47:41.200
 but I know that they actually had the idea before it came out.

47:41.200 --> 47:45.200
 Domain adaptation is when you want to train a machine learning model

47:45.200 --> 47:47.200
 in one setting called a domain,

47:47.200 --> 47:50.200
 and then deploy it in another domain later,

47:50.200 --> 47:52.200
 and you would like it to perform well in the new domain,

47:52.200 --> 47:55.200
 even though the new domain is different from how it was trained.

47:55.200 --> 47:58.200
 So, for example, you might want to train

47:58.200 --> 48:01.200
 on a really clean image dataset like ImageNet,

48:01.200 --> 48:03.200
 but then deploy on users phones,

48:03.200 --> 48:06.200
 where the user is taking pictures in the dark

48:06.200 --> 48:08.200
 and pictures while moving quickly

48:08.200 --> 48:10.200
 and just pictures that aren't really centered

48:10.200 --> 48:13.200
 or composed all that well.

48:13.200 --> 48:16.200
 When you take a normal machine learning model,

48:16.200 --> 48:19.200
 it often degrades really badly when you move to the new domain

48:19.200 --> 48:22.200
 because it looks so different from what the model was trained on.

48:22.200 --> 48:25.200
 Domain adaptation algorithms try to smooth out that gap,

48:25.200 --> 48:28.200
 and the domain adversarial approach is based on

48:28.200 --> 48:30.200
 training a feature extractor,

48:30.200 --> 48:32.200
 where the features have the same statistics

48:32.200 --> 48:35.200
 regardless of which domain you extracted them on.

48:35.200 --> 48:37.200
 So, in the domain adversarial game,

48:37.200 --> 48:39.200
 you have one player that's a feature extractor

48:39.200 --> 48:42.200
 and another player that's a domain recognizer.

48:42.200 --> 48:44.200
 The domain recognizer wants to look at the output

48:44.200 --> 48:47.200
 of the feature extractor and guess which of the two domains

48:47.200 --> 48:49.200
 the features came from.

48:49.200 --> 48:52.200
 So, it's a lot like the real versus fake discriminator in GANs.

48:52.200 --> 48:54.200
 And then the feature extractor,

48:54.200 --> 48:57.200
 you can think of as loosely analogous to the generator in GANs,

48:57.200 --> 48:59.200
 except what it's trying to do here

48:59.200 --> 49:02.200
 is both fool the domain recognizer

49:02.200 --> 49:05.200
 into not knowing which domain the data came from

49:05.200 --> 49:08.200
 and also extract features that are good for classification.

49:08.200 --> 49:13.200
 So, at the end of the day, in the cases where it works out,

49:13.200 --> 49:18.200
 you can actually get features that work about the same

49:18.200 --> 49:20.200
 in both domains.

49:20.200 --> 49:22.200
 Sometimes this has a drawback where,

49:22.200 --> 49:24.200
 in order to make things work the same in both domains,

49:24.200 --> 49:26.200
 it just gets worse at the first one.

49:26.200 --> 49:28.200
 But there are a lot of cases where it actually

49:28.200 --> 49:30.200
 works out well on both.

49:30.200 --> 49:33.200
 So, do you think of GANs being useful in the context

49:33.200 --> 49:35.200
 of data augmentation?

49:35.200 --> 49:37.200
 Yeah, one thing you could hope for with GANs

49:37.200 --> 49:39.200
 is you could imagine,

49:39.200 --> 49:41.200
 I've got a limited training set

49:41.200 --> 49:43.200
 and I'd like to make more training data

49:43.200 --> 49:46.200
 to train something else like a classifier.

49:46.200 --> 49:50.200
 You could train the GAN on the training set

49:50.200 --> 49:52.200
 and then create more data

49:52.200 --> 49:55.200
 and then maybe the classifier would perform better

49:55.200 --> 49:58.200
 on the test set after training on this bigger GAN generated data set.

49:58.200 --> 50:00.200
 So, that's the simplest version

50:00.200 --> 50:02.200
 of something you might hope would work.

50:02.200 --> 50:05.200
 I've never heard of that particular approach working,

50:05.200 --> 50:08.200
 but I think there's some closely related things

50:08.200 --> 50:11.200
 that I think could work in the future

50:11.200 --> 50:13.200
 and some that actually already have worked.

50:13.200 --> 50:15.200
 So, if we think a little bit about what we'd be hoping for

50:15.200 --> 50:17.200
 if we use the GAN to make more training data,

50:17.200 --> 50:20.200
 we're hoping that the GAN will generalize

50:20.200 --> 50:23.200
 to new examples better than the classifier would have

50:23.200 --> 50:25.200
 generalized if it was trained on the same data.

50:25.200 --> 50:27.200
 And I don't know of any reason to believe

50:27.200 --> 50:30.200
 that the GAN would generalize better than the classifier would.

50:30.200 --> 50:33.200
 But what we might hope for is that the GAN

50:33.200 --> 50:37.200
 could generalize differently from a specific classifier.

50:37.200 --> 50:39.200
 So, one thing I think is worth trying

50:39.200 --> 50:41.200
 that I haven't personally tried, but someone could try is

50:41.200 --> 50:44.200
 what if you trained a whole lot of different generative models

50:44.200 --> 50:46.200
 on the same training set,

50:46.200 --> 50:48.200
 create samples from all of them

50:48.200 --> 50:50.200
 and then train a classifier on that.

50:50.200 --> 50:52.200
 Because each of the generative models

50:52.200 --> 50:54.200
 might generalize in a slightly different way,

50:54.200 --> 50:56.200
 they might capture many different axes of variation

50:56.200 --> 50:58.200
 that one individual model wouldn't.

50:58.200 --> 51:01.200
 And then the classifier can capture all of those ideas

51:01.200 --> 51:03.200
 by training in all of their data.

51:03.200 --> 51:06.200
 So, it'd be a little bit like making an ensemble of classifiers.

51:06.200 --> 51:08.200
 An ensemble of GANs in a way.

51:08.200 --> 51:10.200
 I think that could generalize better.

51:10.200 --> 51:12.200
 The other thing that GANs are really good for

51:12.200 --> 51:16.200
 is not necessarily generating new data

51:16.200 --> 51:19.200
 that's exactly like what you already have,

51:19.200 --> 51:23.200
 but by generating new data that has different properties

51:23.200 --> 51:25.200
 from the data you already had.

51:25.200 --> 51:27.200
 One thing that you can do is you can create

51:27.200 --> 51:29.200
 differentially private data.

51:29.200 --> 51:31.200
 So, suppose that you have something like medical records

51:31.200 --> 51:34.200
 and you don't want to train a classifier on the medical records

51:34.200 --> 51:36.200
 and then publish the classifier

51:36.200 --> 51:38.200
 because someone might be able to reverse engineer

51:38.200 --> 51:40.200
 some of the medical records you trained on.

51:40.200 --> 51:42.200
 There's a paper from Casey Green's lab

51:42.200 --> 51:46.200
 that shows how you can train again using differential privacy.

51:46.200 --> 51:48.200
 And then the samples from the GAN

51:48.200 --> 51:51.200
 still have the same differential privacy guarantees

51:51.200 --> 51:53.200
 as the parameters of the GAN.

51:53.200 --> 51:55.200
 So, you can make fake patient data

51:55.200 --> 51:57.200
 for other researchers to use

51:57.200 --> 51:59.200
 and they can do almost anything they want with that data

51:59.200 --> 52:02.200
 because it doesn't come from real people.

52:02.200 --> 52:04.200
 And the differential privacy mechanism

52:04.200 --> 52:07.200
 gives you clear guarantees on how much

52:07.200 --> 52:09.200
 the original people's data has been protected.

52:09.200 --> 52:11.200
 That's really interesting, actually.

52:11.200 --> 52:13.200
 I haven't heard you talk about that before.

52:13.200 --> 52:15.200
 In terms of fairness,

52:15.200 --> 52:19.200
 I've seen from AAAI your talk,

52:19.200 --> 52:21.200
 how can adversarial machine learning

52:21.200 --> 52:23.200
 help models be more fair

52:23.200 --> 52:25.200
 with respect to sensitive variables?

52:25.200 --> 52:28.200
 Yeah. So, there's a paper from Emma Storky's lab

52:28.200 --> 52:31.200
 about how to learn machine learning models

52:31.200 --> 52:34.200
 that are incapable of using specific variables.

52:34.200 --> 52:36.200
 So, say, for example, you wanted to make predictions

52:36.200 --> 52:39.200
 that are not affected by gender.

52:39.200 --> 52:41.200
 It isn't enough to just leave gender

52:41.200 --> 52:43.200
 out of the input to the model.

52:43.200 --> 52:45.200
 You can often infer gender from a lot of other characteristics.

52:45.200 --> 52:47.200
 Like, say that you have the person's name,

52:47.200 --> 52:49.200
 but you're not told their gender.

52:49.200 --> 52:53.200
 Well, if their name is Ian, they're kind of obviously a man.

52:53.200 --> 52:55.200
 So, what you'd like to do is make a machine learning model

52:55.200 --> 52:58.200
 that can still take in a lot of different attributes

52:58.200 --> 53:02.200
 and make a really accurate informed prediction,

53:02.200 --> 53:05.200
 but be confident that it isn't reverse engineering gender

53:05.200 --> 53:08.200
 or another sensitive variable internally.

53:08.200 --> 53:10.200
 You can do that using something very similar

53:10.200 --> 53:12.200
 to the domain adversarial approach,

53:12.200 --> 53:15.200
 where you have one player that's a feature extractor

53:15.200 --> 53:18.200
 and another player that's a feature analyzer.

53:18.200 --> 53:21.200
 And you want to make sure that the feature analyzer

53:21.200 --> 53:24.200
 is not able to guess the value of the sensitive variable

53:24.200 --> 53:26.200
 that you're trying to keep private.

53:26.200 --> 53:29.200
 Right. Yeah, I love this approach.

53:29.200 --> 53:34.200
 So, with the feature, you're not able to infer

53:34.200 --> 53:36.200
 the sensitive variables.

53:36.200 --> 53:39.200
 It's brilliant. It's quite brilliant and simple, actually.

53:39.200 --> 53:42.200
 Another way I think that GANs in particular

53:42.200 --> 53:44.200
 could be used for fairness would be

53:44.200 --> 53:46.200
 to make something like a cycle GAN,

53:46.200 --> 53:49.200
 where you can take data from one domain

53:49.200 --> 53:51.200
 and convert it into another.

53:51.200 --> 53:54.200
 We've seen cycle GAN turning horses into zebras.

53:54.200 --> 53:59.200
 We've seen other unsupervised GANs made by Mingyu Liu

53:59.200 --> 54:02.200
 doing things like turning day photos into night photos.

54:02.200 --> 54:05.200
 I think for fairness, you could imagine

54:05.200 --> 54:08.200
 taking records for people in one group

54:08.200 --> 54:11.200
 and transforming them into analogous people in another group

54:11.200 --> 54:14.200
 and testing to see if they're treated equitably

54:14.200 --> 54:16.200
 across those two groups.

54:16.200 --> 54:18.200
 There's a lot of things that would be hard to get right

54:18.200 --> 54:21.200
 and make sure that the conversion process itself is fair.

54:21.200 --> 54:24.200
 And I don't think it's anywhere near something

54:24.200 --> 54:26.200
 that we could actually use yet.

54:26.200 --> 54:28.200
 But if you could design that conversion process very carefully,

54:28.200 --> 54:30.200
 it might give you a way of doing audits

54:30.200 --> 54:33.200
 where you say, what if we took people from this group,

54:33.200 --> 54:35.200
 converted them into equivalent people in another group?

54:35.200 --> 54:39.200
 Does the system actually treat them how it ought to?

54:39.200 --> 54:41.200
 That's also really interesting.

54:41.200 --> 54:46.200
 You know, in popular press

54:46.200 --> 54:48.200
 and in general, in our imagination,

54:48.200 --> 54:51.200
 you think, well, GANs are able to generate data

54:51.200 --> 54:54.200
 and you start to think about deep fakes

54:54.200 --> 54:57.200
 or being able to sort of maliciously generate data

54:57.200 --> 55:00.200
 that fakes the identity of other people.

55:00.200 --> 55:03.200
 Is this something of a concern to you?

55:03.200 --> 55:06.200
 Is this something, if you look 10, 20 years into the future,

55:06.200 --> 55:10.200
 is that something that pops up in your work,

55:10.200 --> 55:13.200
 in the work of the community that's working on generative models?

55:13.200 --> 55:15.200
 I'm a lot less concerned about 20 years from now

55:15.200 --> 55:17.200
 than the next few years.

55:17.200 --> 55:20.200
 I think there will be a kind of bumpy cultural transition

55:20.200 --> 55:22.200
 as people encounter this idea

55:22.200 --> 55:25.200
 that there can be very realistic videos and audio that aren't real.

55:25.200 --> 55:27.200
 I think 20 years from now,

55:27.200 --> 55:30.200
 people will mostly understand that you shouldn't believe

55:30.200 --> 55:33.200
 something is real just because you saw a video of it.

55:33.200 --> 55:37.200
 People will expect to see that it's been cryptographically signed

55:37.200 --> 55:41.200
 or have some other mechanism to make them believe

55:41.200 --> 55:43.200
 that the content is real.

55:43.200 --> 55:45.200
 There's already people working on this,

55:45.200 --> 55:47.200
 like there's a startup called TruePick

55:47.200 --> 55:50.200
 that provides a lot of mechanisms for authenticating

55:50.200 --> 55:52.200
 that an image is real.

55:52.200 --> 55:55.200
 They're maybe not quite up to having a state actor

55:55.200 --> 55:59.200
 try to evade their verification techniques,

55:59.200 --> 56:02.200
 but it's something that people are already working on

56:02.200 --> 56:04.200
 and I think will get right eventually.

56:04.200 --> 56:08.200
 So you think authentication will eventually win out?

56:08.200 --> 56:11.200
 So being able to authenticate that this is real and this is not?

56:11.200 --> 56:13.200
 Yeah.

56:13.200 --> 56:15.200
 As opposed to GANs just getting better and better

56:15.200 --> 56:18.200
 or generative models being able to get better and better

56:18.200 --> 56:21.200
 to where the nature of what is real is normal.

56:21.200 --> 56:25.200
 I don't think we'll ever be able to look at the pixels of a photo

56:25.200 --> 56:28.200
 and tell you for sure that it's real or not real,

56:28.200 --> 56:32.200
 and I think it would actually be somewhat dangerous

56:32.200 --> 56:34.200
 to rely on that approach too much.

56:34.200 --> 56:36.200
 If you make a really good fake detector

56:36.200 --> 56:38.200
 and then someone's able to fool your fake detector

56:38.200 --> 56:41.200
 and your fake detector says this image is not fake,

56:41.200 --> 56:43.200
 then it's even more credible

56:43.200 --> 56:46.200
 than if you've never made a fake detector in the first place.

56:46.200 --> 56:50.200
 What I do think we'll get to is systems

56:50.200 --> 56:52.200
 that we can kind of use behind the scenes

56:52.200 --> 56:55.200
 to make estimates of what's going on

56:55.200 --> 56:59.200
 and maybe not use them in court for a definitive analysis.

56:59.200 --> 57:04.200
 I also think we will likely get better authentication systems

57:04.200 --> 57:08.200
 where, imagine that every phone cryptographically

57:08.200 --> 57:10.200
 signs everything that comes out of it.

57:10.200 --> 57:12.200
 You wouldn't be able to conclusively tell

57:12.200 --> 57:14.200
 that an image was real,

57:14.200 --> 57:18.200
 but you would be able to tell somebody who knew

57:18.200 --> 57:21.200
 the appropriate private key for this phone

57:21.200 --> 57:24.200
 was actually able to sign this image

57:24.200 --> 57:28.200
 and upload it to this server at this time stamp.

57:28.200 --> 57:31.200
 You could imagine maybe you make phones

57:31.200 --> 57:35.200
 that have the private keys hardware embedded in them.

57:35.200 --> 57:37.200
 If a state security agency

57:37.200 --> 57:39.200
 really wants to infiltrate the company,

57:39.200 --> 57:42.200
 they could probably plant a private key of their choice

57:42.200 --> 57:44.200
 or break open the chip

57:44.200 --> 57:46.200
 and learn the private key or something like that.

57:46.200 --> 57:48.200
 But it would make it a lot harder

57:48.200 --> 57:51.200
 for an adversary with fewer resources to fake things.

57:51.200 --> 57:53.200
 For most of us, it would be okay.

57:53.200 --> 57:58.200
 You mentioned the beer and the bar and the new ideas.

57:58.200 --> 58:01.200
 You were able to come up with this new idea

58:01.200 --> 58:04.200
 pretty quickly and implement it pretty quickly.

58:04.200 --> 58:06.200
 Do you think there are still many

58:06.200 --> 58:08.200
 such groundbreaking ideas in deep learning

58:08.200 --> 58:10.200
 that could be developed so quickly?

58:10.200 --> 58:13.200
 Yeah, I do think that there are a lot of ideas

58:13.200 --> 58:15.200
 that can be developed really quickly.

58:15.200 --> 58:18.200
 GANs were probably a little bit of an outlier

58:18.200 --> 58:20.200
 on the whole one hour time scale.

58:20.200 --> 58:24.200
 But just in terms of low resource ideas

58:24.200 --> 58:26.200
 where you do something really different

58:26.200 --> 58:29.200
 on a high scale and get a big payback,

58:29.200 --> 58:32.200
 I think it's not as likely that you'll see that

58:32.200 --> 58:35.200
 in terms of things like core machine learning technologies

58:35.200 --> 58:37.200
 like a better classifier

58:37.200 --> 58:39.200
 or a better reinforcement learning algorithm

58:39.200 --> 58:41.200
 or a better generative model.

58:41.200 --> 58:43.200
 If I had the GAN idea today,

58:43.200 --> 58:45.200
 it would be a lot harder to prove that it was useful

58:45.200 --> 58:47.200
 than it was back in 2014

58:47.200 --> 58:50.200
 because I would need to get it running on something

58:50.200 --> 58:54.200
 like ImageNet or Celeb A at high resolution.

58:54.200 --> 58:56.200
 Those take a while to train.

58:56.200 --> 58:58.200
 You couldn't train it in an hour

58:58.200 --> 59:01.200
 and know that it was something really new and exciting.

59:01.200 --> 59:04.200
 Back in 2014, training on MNIST was enough.

59:04.200 --> 59:07.200
 But there are other areas of machine learning

59:07.200 --> 59:11.200
 where I think a new idea could actually be developed

59:11.200 --> 59:13.200
 really quickly with low resources.

59:13.200 --> 59:15.200
 What's your intuition about what areas

59:15.200 --> 59:18.200
 of machine learning are ripe for this?

59:18.200 --> 59:23.200
 Yeah, so I think fairness and interpretability

59:23.200 --> 59:27.200
 are areas where we just really don't have any idea

59:27.200 --> 59:29.200
 how anything should be done yet.

59:29.200 --> 59:31.200
 Like for interpretability,

59:31.200 --> 59:33.200
 I don't think we even have the right definitions.

59:33.200 --> 59:36.200
 And even just defining a really useful concept,

59:36.200 --> 59:38.200
 you don't even need to run any experiments.

59:38.200 --> 59:40.200
 It could have a huge impact on the field.

59:40.200 --> 59:43.200
 We've seen that, for example, in differential privacy

59:43.200 --> 59:45.200
 that Cynthia Dwork and her collaborators

59:45.200 --> 59:48.200
 made this technical definition of privacy

59:48.200 --> 59:50.200
 where before a lot of things were really mushy

59:50.200 --> 59:53.200
 and with that definition, you could actually design

59:53.200 --> 59:55.200
 randomized algorithms for accessing databases

59:55.200 --> 59:59.200
 and guarantee that they preserved individual people's privacy

59:59.200 --> 1:00:03.200
 in a mathematical quantitative sense.

1:00:03.200 --> 1:00:05.200
 Right now, we all talk a lot about

1:00:05.200 --> 1:00:07.200
 how interpretable different machine learning algorithms are,

1:00:07.200 --> 1:00:09.200
 but it's really just people's opinion.

1:00:09.200 --> 1:00:11.200
 And everybody probably has a different idea

1:00:11.200 --> 1:00:13.200
 of what interpretability means in their head.

1:00:13.200 --> 1:00:16.200
 If we could define some concept related to interpretability

1:00:16.200 --> 1:00:18.200
 that's actually measurable,

1:00:18.200 --> 1:00:20.200
 that would be a huge leap forward

1:00:20.200 --> 1:00:24.200
 even without a new algorithm that increases that quantity.

1:00:24.200 --> 1:00:28.200
 And also, once we had the definition of differential privacy,

1:00:28.200 --> 1:00:31.200
 it was fast to get the algorithms that guaranteed it.

1:00:31.200 --> 1:00:33.200
 So you could imagine once we have definitions

1:00:33.200 --> 1:00:35.200
 of good concepts and interpretability,

1:00:35.200 --> 1:00:37.200
 we might be able to provide the algorithms

1:00:37.200 --> 1:00:42.200
 that have the interpretability guarantees quickly, too.

1:00:42.200 --> 1:00:46.200
 What do you think it takes to build a system

1:00:46.200 --> 1:00:48.200
 with human level intelligence

1:00:48.200 --> 1:00:51.200
 as we quickly venture into the philosophical?

1:00:51.200 --> 1:00:55.200
 So artificial general intelligence, what do you think it takes?

1:00:55.200 --> 1:01:01.200
 I think that it definitely takes better environments

1:01:01.200 --> 1:01:03.200
 than we currently have for training agents,

1:01:03.200 --> 1:01:08.200
 that we want them to have a really wide diversity of experiences.

1:01:08.200 --> 1:01:11.200
 I also think it's going to take really a lot of computation.

1:01:11.200 --> 1:01:13.200
 It's hard to imagine exactly how much.

1:01:13.200 --> 1:01:16.200
 So you're optimistic about simulation,

1:01:16.200 --> 1:01:19.200
 simulating a variety of environments as the path forward

1:01:19.200 --> 1:01:21.200
 as opposed to operating in the real world?

1:01:21.200 --> 1:01:23.200
 I think it's a necessary ingredient.

1:01:23.200 --> 1:01:27.200
 I don't think that we're going to get to artificial general intelligence

1:01:27.200 --> 1:01:29.200
 by training on fixed data sets

1:01:29.200 --> 1:01:32.200
 or by thinking really hard about the problem.

1:01:32.200 --> 1:01:36.200
 I think that the agent really needs to interact

1:01:36.200 --> 1:01:41.200
 and have a variety of experiences within the same lifespan.

1:01:41.200 --> 1:01:45.200
 And today we have many different models that can each do one thing,

1:01:45.200 --> 1:01:49.200
 and we tend to train them on one dataset or one RL environment.

1:01:49.200 --> 1:01:53.200
 Sometimes there are actually papers about getting one set of parameters

1:01:53.200 --> 1:01:56.200
 to perform well in many different RL environments,

1:01:56.200 --> 1:01:59.200
 but we don't really have anything like an agent

1:01:59.200 --> 1:02:02.200
 that goes seamlessly from one type of experience to another

1:02:02.200 --> 1:02:05.200
 and really integrates all the different things that it does

1:02:05.200 --> 1:02:07.200
 over the course of its life.

1:02:07.200 --> 1:02:10.200
 When we do see multiagent environments,

1:02:10.200 --> 1:02:16.200
 they tend to be similar environments.

1:02:16.200 --> 1:02:19.200
 All of them are playing an action based video game.

1:02:19.200 --> 1:02:24.200
 We don't really have an agent that goes from playing a video game

1:02:24.200 --> 1:02:27.200
 to reading the Wall Street Journal

1:02:27.200 --> 1:02:32.200
 to predicting how effective a molecule will be as a drug or something like that.

1:02:32.200 --> 1:02:36.200
 What do you think is a good test for intelligence in your view?

1:02:36.200 --> 1:02:41.200
 There's been a lot of benchmarks started with Alan Turing,

1:02:41.200 --> 1:02:46.200
 natural conversation being a good benchmark for intelligence.

1:02:46.200 --> 1:02:53.200
 What would you and good fellows sit back and be really damn impressed

1:02:53.200 --> 1:02:55.200
 if a system was able to accomplish?

1:02:55.200 --> 1:02:59.200
 Something that doesn't take a lot of glue from human engineers.

1:02:59.200 --> 1:03:07.200
 Imagine that instead of having to go to the CIFAR website and download CIFAR 10

1:03:07.200 --> 1:03:11.200
 and then write a Python script to parse it and all that,

1:03:11.200 --> 1:03:16.200
 you could just point an agent at the CIFAR 10 problem

1:03:16.200 --> 1:03:20.200
 and it downloads and extracts the data and trains a model

1:03:20.200 --> 1:03:22.200
 and starts giving you predictions.

1:03:22.200 --> 1:03:28.200
 I feel like something that doesn't need to have every step of the pipeline assembled for it

1:03:28.200 --> 1:03:30.200
 definitely understands what it's doing.

1:03:30.200 --> 1:03:34.200
 Is AutoML moving into that direction or are you thinking way even bigger?

1:03:34.200 --> 1:03:39.200
 AutoML has mostly been moving toward once we've built all the glue,

1:03:39.200 --> 1:03:44.200
 can the machine learning system design the architecture really well?

1:03:44.200 --> 1:03:49.200
 I'm more of saying if something knows how to pre process the data

1:03:49.200 --> 1:03:52.200
 so that it successfully accomplishes the task,

1:03:52.200 --> 1:03:56.200
 then it would be very hard to argue that it doesn't truly understand the task

1:03:56.200 --> 1:03:58.200
 in some fundamental sense.

1:03:58.200 --> 1:04:02.200
 I don't necessarily know that that's the philosophical definition of intelligence,

1:04:02.200 --> 1:04:05.200
 but that's something that would be really cool to build that would be really useful

1:04:05.200 --> 1:04:09.200
 and would impress me and would convince me that we've made a step forward in real AI.

1:04:09.200 --> 1:04:13.200
 You give it the URL for Wikipedia

1:04:13.200 --> 1:04:18.200
 and then next day expect it to be able to solve CIFAR 10.

1:04:18.200 --> 1:04:22.200
 Or you type in a paragraph explaining what you want it to do

1:04:22.200 --> 1:04:28.200
 and it figures out what web searches it should run and downloads all the necessary ingredients.

1:04:28.200 --> 1:04:37.200
 So you have a very clear, calm way of speaking, no ums, easy to edit.

1:04:37.200 --> 1:04:44.200
 I've seen comments for both you and I have been identified as both potentially being robots.

1:04:44.200 --> 1:04:48.200
 If you have to prove to the world that you are indeed human, how would you do it?

1:04:48.200 --> 1:04:53.200
 I can understand thinking that I'm a robot.

1:04:53.200 --> 1:04:57.200
 It's the flip side of the Turing test, I think.

1:04:57.200 --> 1:05:00.200
 Yeah, the prove your human test.

1:05:00.200 --> 1:05:08.200
 Intellectually, so you have to, is there something that's truly unique in your mind

1:05:08.200 --> 1:05:13.200
 as it doesn't go back to just natural language again, just being able to talk the way out of it?

1:05:13.200 --> 1:05:16.200
 So proving that I'm not a robot with today's technology,

1:05:16.200 --> 1:05:18.200
 that's pretty straightforward.

1:05:18.200 --> 1:05:25.200
 My conversation today hasn't veered off into talking about the stock market or something because it's my training data.

1:05:25.200 --> 1:05:31.200
 But I guess more generally trying to prove that something is real from the content alone is incredibly hard.

1:05:31.200 --> 1:05:37.200
 That's one of the main things I've gotten out of my GAN research, that you can simulate almost anything

1:05:37.200 --> 1:05:42.200
 and so you have to really step back to a separate channel to prove that something is real.

1:05:42.200 --> 1:05:48.200
 So I guess I should have had myself stamped on a blockchain when I was born or something, but I didn't do that.

1:05:48.200 --> 1:05:52.200
 So according to my own research methodology, there's just no way to know at this point.

1:05:52.200 --> 1:05:59.200
 So what last question, problem stands out for you that you're really excited about challenging in the near future?

1:05:59.200 --> 1:06:06.200
 I think resistance to adversarial examples, figuring out how to make machine learning secure against an adversary

1:06:06.200 --> 1:06:11.200
 who wants to interfere it and control it, that is one of the most important things researchers today could solve.

1:06:11.200 --> 1:06:17.200
 In all domains, image, language, driving and everything.

1:06:17.200 --> 1:06:22.200
 I guess I'm most concerned about domains we haven't really encountered yet.

1:06:22.200 --> 1:06:28.200
 Imagine 20 years from now when we're using advanced AIs to do things we haven't even thought of yet.

1:06:28.200 --> 1:06:37.200
 If you ask people what are the important problems in security of phones in 2002,

1:06:37.200 --> 1:06:43.200
 I don't think we would have anticipated that we're using them for nearly as many things as we're using them for today.

1:06:43.200 --> 1:06:47.200
 I think it's going to be like that with AI that you can kind of try to speculate about where it's going,

1:06:47.200 --> 1:06:53.200
 but really the business opportunities that end up taking off would be hard to predict ahead of time.

1:06:53.200 --> 1:06:58.200
 What you can predict ahead of time is that almost anything you can do with machine learning,

1:06:58.200 --> 1:07:04.200
 you would like to make sure that people can't get it to do what they want rather than what you want

1:07:04.200 --> 1:07:08.200
 just by showing it a funny QR code or a funny input pattern.

1:07:08.200 --> 1:07:12.200
 You think that the set of methodology to do that can be bigger than any one domain?

1:07:12.200 --> 1:07:15.200
 I think so, yeah.

1:07:15.200 --> 1:07:20.200
 One methodology that I think is not a specific methodology,

1:07:20.200 --> 1:07:25.200
 but a category of solutions that I'm excited about today is making dynamic models

1:07:25.200 --> 1:07:28.200
 that change every time they make a prediction.

1:07:28.200 --> 1:07:32.200
 Right now, we tend to train models and then after they're trained, we freeze them.

1:07:32.200 --> 1:07:37.200
 We just use the same rule to classify everything that comes in from then on.

1:07:37.200 --> 1:07:40.200
 That's really a sitting duck from a security point of view.

1:07:40.200 --> 1:07:44.200
 If you always output the same answer for the same input,

1:07:44.200 --> 1:07:49.200
 then people can just run inputs through until they find a mistake that benefits them,

1:07:49.200 --> 1:07:53.200
 and then they use the same mistake over and over and over again.

1:07:53.200 --> 1:08:00.200
 I think having a model that updates its predictions so that it's harder to predict what you're going to get

1:08:00.200 --> 1:08:04.200
 will make it harder for an adversary to really take control of the system

1:08:04.200 --> 1:08:06.200
 and make it do what they want it to do.

1:08:06.200 --> 1:08:10.200
 Yeah, models that maintain a bit of a sense of mystery about them

1:08:10.200 --> 1:08:12.200
 because they always keep changing.

1:08:12.200 --> 1:08:14.200
 Ian, thanks so much for talking today. It was awesome.

1:08:14.200 --> 1:08:36.200
 Thank you for coming in. It's great to see you.