vtt/episode_010_small.vtt

WEBVTT

00:00.000 --> 00:03.040
 The following is a conversation with Peter Abiel.

00:03.040 --> 00:07.760
 He's a professor at UC Berkeley and the director of the Berkeley Robotics Learning Lab.

00:07.760 --> 00:13.200
 He's one of the top researchers in the world working on how to make robots understand and

00:13.200 --> 00:18.480
 interact with the world around them, especially using imitation and deeper enforcement learning.

00:19.680 --> 00:24.160
 This conversation is part of the MIT course on artificial general intelligence

00:24.160 --> 00:29.040
 and the artificial intelligence podcast. If you enjoy it, please subscribe on YouTube,

00:29.040 --> 00:34.160
 iTunes, or your podcast provider of choice, or simply connect with me on Twitter at Lex

00:34.160 --> 00:40.560
 Freedman, spelled F R I D. And now here's my conversation with Peter Abiel.

00:41.440 --> 00:46.480
 You've mentioned that if there was one person you could meet, it would be Roger Federer. So let

00:46.480 --> 00:52.720
 me ask, when do you think we will have a robot that fully autonomously can beat Roger Federer

00:52.720 --> 00:59.840
 at tennis? Roger Federer level player at tennis? Well, first, if you can make it happen for me

00:59.840 --> 01:07.840
 to meet Roger, let me know. In terms of getting a robot to beat him at tennis, it's kind of an

01:07.840 --> 01:15.280
 interesting question because for a lot of the challenges we think about in AI, the software

01:15.280 --> 01:22.800
 is really the missing piece. But for something like this, the hardware is nowhere near either. To

01:22.800 --> 01:28.240
 really have a robot that can physically run around, the Boston Dynamics robots are starting to get

01:28.240 --> 01:34.560
 there, but still not really human level ability to run around and then swing a racket.

01:36.720 --> 01:40.160
 So you think that's a hardware problem? I don't think it's a hardware problem only. I think it's

01:40.160 --> 01:45.600
 a hardware and a software problem. I think it's both. And I think they'll have independent progress.

01:45.600 --> 01:53.360
 So I'd say the hardware maybe in 10, 15 years. On clay, not grass. I mean, grass is probably hard.

01:53.360 --> 02:00.080
 With the sliding? Yeah. Well, clay, I'm not sure what's harder, grass or clay. The clay involves

02:00.080 --> 02:09.360
 sliding, which might be harder to master actually. Yeah. But you're not limited to bipedal. I mean,

02:09.360 --> 02:12.560
 I'm sure there's no... Well, if we can build a machine, it's a whole different question, of

02:12.560 --> 02:18.000
 course. If you can say, okay, this robot can be on wheels, it can move around on wheels and

02:18.000 --> 02:24.880
 can be designed differently, then I think that can be done sooner probably than a full humanoid

02:24.880 --> 02:30.400
 type of setup. What do you think of swing a racket? So you've worked at basic manipulation.

02:31.120 --> 02:36.480
 How hard do you think is the task of swinging a racket with a be able to hit a nice backhand

02:36.480 --> 02:44.240
 or a forehand? Let's say we just set up stationery, a nice robot arm, let's say. You know,

02:44.240 --> 02:49.440
 a standard industrial arm, and it can watch the ball come and then swing the racket.

02:50.560 --> 02:57.600
 It's a good question. I'm not sure it would be super hard to do. I mean, I'm sure it would require

02:57.600 --> 03:01.520
 a lot... If we do it with reinforcement learning, it would require a lot of trial and error. It's

03:01.520 --> 03:06.960
 not going to swing it right the first time around, but yeah, I don't see why I couldn't

03:08.240 --> 03:12.320
 swing it the right way. I think it's learnable. I think if you set up a ball machine, let's say

03:12.320 --> 03:18.960
 on one side and then a robot with a tennis racket on the other side, I think it's learnable

03:20.160 --> 03:25.360
 and maybe a little bit of pre training and simulation. Yeah, I think that's feasible.

03:25.360 --> 03:28.880
 I think the swinging the racket is feasible. It'd be very interesting to see how much precision it

03:28.880 --> 03:37.760
 can get. I mean, that's where... I mean, some of the human players can hit it on the lines,

03:37.760 --> 03:44.320
 which is very high precision. With spin. The spin is an interesting whether RL can learn to

03:44.320 --> 03:48.160
 put a spin on the ball. Well, you got me interested. Maybe someday we'll set this up.

03:51.040 --> 03:55.440
 Your answer is basically, okay, for this problem, it sounds fascinating, but for the general problem

03:55.440 --> 03:59.840
 of a tennis player, we might be a little bit farther away. What's the most impressive thing

03:59.840 --> 04:06.720
 you've seen a robot do in the physical world? So physically, for me, it's

04:08.720 --> 04:16.560
 the Boston Dynamics videos always just ring home and just super impressed. Recently, the robot

04:16.560 --> 04:22.160
 running up the stairs during the parkour type thing. I mean, yes, we don't know what's underneath.

04:22.160 --> 04:26.400
 They don't really write a lot of detail, but even if it's hard coded underneath,

04:27.040 --> 04:30.800
 which it might or might not be just the physical abilities of doing that parkour,

04:30.800 --> 04:36.000
 that's a very impressive robot right there. So have you met Spotmini or any of those robots in

04:36.000 --> 04:43.040
 person? I met Spotmini last year in April at the Mars event that Jeff Bezos organizes. They

04:43.040 --> 04:49.840
 brought it out there and it was nicely falling around Jeff. When Jeff left the room, they had it

04:49.840 --> 04:55.680
 following him along, which is pretty impressive. So I think there's some confidence to know that

04:55.680 --> 05:00.080
 there's no learning going on in those robots. The psychology of it, so while knowing that,

05:00.080 --> 05:03.360
 while knowing there's not, if there's any learning going on, it's very limited,

05:03.920 --> 05:08.720
 I met Spotmini earlier this year and knowing everything that's going on,

05:09.520 --> 05:12.400
 having one on one interaction, so I get to spend some time alone.

05:14.400 --> 05:18.720
 And there's immediately a deep connection on the psychological level,

05:18.720 --> 05:22.320
 even though you know the fundamentals, how it works, there's something magical.

05:23.280 --> 05:29.040
 So do you think about the psychology of interacting with robots in the physical world,

05:29.040 --> 05:36.000
 even you just showed me the PR2, the robot, and there was a little bit something like a face,

05:37.040 --> 05:40.480
 had a little bit something like a face, there's something that immediately draws you to it.

05:40.480 --> 05:45.040
 Do you think about that aspect of the robotics problem?

05:45.040 --> 05:50.560
 Well, it's very hard with Brett here. We'll give him a name, Berkeley Robot,

05:50.560 --> 05:56.480
 for the elimination of tedious tasks. It's very hard to not think of the robot as a person,

05:56.480 --> 06:00.560
 and it seems like everybody calls him a he for whatever reason, but that also makes it more

06:00.560 --> 06:07.200
 a person than if it was a it. And it seems pretty natural to think of it that way.

06:07.200 --> 06:12.400
 This past weekend really struck me, I've seen Pepper many times on videos,

06:12.400 --> 06:18.640
 but then I was at an event organized by, this was by Fidelity, and they had scripted Pepper to help

06:19.280 --> 06:24.880
 moderate some sessions, and they had scripted Pepper to have the personality of a child a

06:24.880 --> 06:31.360
 little bit. And it was very hard to not think of it as its own person in some sense, because it

06:31.360 --> 06:35.120
 was just kind of jumping, it would just jump into conversation making it very interactive.

06:35.120 --> 06:38.720
 Moderate would be saying Pepper would just jump in, hold on, how about me,

06:38.720 --> 06:43.600
 how about me, can I participate in this doing it, just like, okay, this is like like a person,

06:43.600 --> 06:48.800
 and that was 100% scripted. And even then it was hard not to have that sense of somehow

06:48.800 --> 06:55.120
 there is something there. So as we have robots interact in this physical world, is that a signal

06:55.120 --> 07:00.160
 that could be used in reinforcement learning? You've worked a little bit in this direction,

07:00.160 --> 07:05.920
 but do you think that that psychology can be somehow pulled in? Yes, that's a question I would

07:05.920 --> 07:12.800
 say a lot, a lot of people ask. And I think part of why they ask it is they're thinking about

07:14.160 --> 07:18.560
 how unique are we really still as people, like after they see some results, they see

07:18.560 --> 07:23.200
 a computer play go to say a computer do this that they're like, okay, but can it really have

07:23.200 --> 07:28.960
 emotion? Can it really interact with us in that way? And then once you're around robots,

07:28.960 --> 07:33.760
 you already start feeling it. And I think that kind of maybe methodologically, the way that I

07:33.760 --> 07:38.560
 think of it is, if you run something like reinforcement learnings about optimizing some

07:38.560 --> 07:48.240
 objective, and there's no reason that the objective couldn't be tied into how much

07:48.240 --> 07:53.120
 does a person like interacting with this system? And why could not the reinforcement learning system

07:53.120 --> 07:59.040
 optimize for the robot being fun to be around? And why wouldn't it then naturally become more

07:59.040 --> 08:03.920
 more interactive and more and more maybe like a person or like a pet? I don't know what it would

08:03.920 --> 08:08.720
 exactly be, but more and more have those features and acquire them automatically. As long as you

08:08.720 --> 08:16.320
 can formalize an objective of what it means to like something, how you exhibit what's the ground

08:16.320 --> 08:21.360
 truth? How do you get the reward from human? Because you have to somehow collect that information

08:21.360 --> 08:27.120
 from human. But you're saying if you can formulate as an objective, it can be learned.

08:27.120 --> 08:30.800
 There's no reason it couldn't emerge through learning. And maybe one way to formulate as an

08:30.800 --> 08:35.840
 objective, you wouldn't have to necessarily score it explicitly. So standard rewards are

08:35.840 --> 08:41.920
 numbers. And numbers are hard to come by. This is a 1.5 or 1.7 on some scale. It's very hard to do

08:41.920 --> 08:47.680
 for a person. But much easier is for a person to say, okay, what you did the last five minutes

08:47.680 --> 08:53.600
 was much nicer than we did the previous five minutes. And that now gives a comparison. And in fact,

08:53.600 --> 08:58.080
 there have been some results on that. For example, Paul Cristiano and collaborators at OpenEye had

08:58.080 --> 09:05.840
 the hopper, Mojoka hopper, one legged robot, the backflip, backflips purely from feedback. I like

09:05.840 --> 09:11.280
 this better than that. That's kind of equally good. And after a bunch of interactions, it figured

09:11.280 --> 09:15.200
 out what it was the person was asking for, namely a backflip. And so I think the same thing.

09:16.080 --> 09:20.880
 It wasn't trying to do a backflip. It was just getting a score from the comparison score from

09:20.880 --> 09:27.760
 the person based on person having a mind in their own mind. I wanted to do a backflip. But

09:27.760 --> 09:32.480
 the robot didn't know what it was supposed to be doing. It just knew that sometimes the person

09:32.480 --> 09:37.120
 said, this is better, this is worse. And then the robot figured out what the person was actually

09:37.120 --> 09:42.560
 after was a backflip. And I imagine the same would be true for things like more interactive

09:42.560 --> 09:47.520
 robots that the robot would figure out over time. Oh, this kind of thing apparently is appreciated

09:47.520 --> 09:54.720
 more than this other kind of thing. So when I first picked up Sutton's Richard Sutton's

09:54.720 --> 10:02.480
 reinforcement learning book, before sort of this deep learning, before the reemergence

10:02.480 --> 10:07.600
 of neural networks as a powerful mechanism for machine learning, IRL seemed to me like magic.

10:07.600 --> 10:18.000
 It was beautiful. So that seemed like what intelligence is, RRL reinforcement learning. So how

10:18.000 --> 10:24.320
 do you think we can possibly learn anything about the world when the reward for the actions is delayed

10:24.320 --> 10:32.160
 is so sparse? Like where is, why do you think RRL works? Why do you think you can learn anything

10:32.160 --> 10:37.600
 under such sparse rewards, whether it's regular reinforcement learning or deeper reinforcement

10:37.600 --> 10:45.600
 learning? What's your intuition? The kind of part of that is, why is RRL, why does it need

10:45.600 --> 10:51.040
 so many samples, so many experiences to learn from? Because really what's happening is when you

10:51.040 --> 10:56.240
 have a sparse reward, you do something maybe for like, I don't know, you take 100 actions and then

10:56.240 --> 11:01.920
 you get a reward, or maybe you get like a score of three. And I'm like, okay, three. Not sure what

11:01.920 --> 11:06.960
 that means. You go again and now you get two. And now you know that that sequence of 100 actions

11:06.960 --> 11:10.640
 that you did the second time around somehow was worse than the sequence of 100 actions you did

11:10.640 --> 11:15.040
 the first time around. But that's tough to now know which one of those were better or worse.

11:15.040 --> 11:19.680
 Some might have been good and bad in either one. And so that's why you need so many experiences.

11:19.680 --> 11:24.080
 But once you have enough experiences, effectively RRL is teasing that apart. It's starting to say,

11:24.080 --> 11:28.640
 okay, what is consistently there when you get a higher reward and what's consistently there when

11:28.640 --> 11:34.080
 you get a lower reward? And then kind of the magic of sometimes the policy grant update is to say,

11:34.720 --> 11:39.520
 now let's update the neural network to make the actions that were kind of present when things are

11:39.520 --> 11:44.960
 good, more likely, and make the actions that are present when things are not as good, less likely.

11:44.960 --> 11:50.480
 So that's that is the counterpoint. But it seems like you would need to run it a lot more than

11:50.480 --> 11:55.120
 you do. Even though right now, people could say that RRL is very inefficient. But it seems to be

11:55.120 --> 12:01.200
 way more efficient than one would imagine on paper, that the simple updates to the policy,

12:01.760 --> 12:07.520
 the policy gradient that somehow you can learn is exactly as I said, what are the common actions

12:07.520 --> 12:11.680
 that seem to produce some good results, that that somehow can learn anything.

12:12.640 --> 12:16.800
 It seems counterintuitive, at least. Is there some intuition behind it?

12:16.800 --> 12:24.720
 Yeah, so I think there's a few ways to think about this. The way I tend to think about it

12:24.720 --> 12:29.920
 mostly originally. And so when we started working on deep reinforcement learning here at Berkeley,

12:29.920 --> 12:36.880
 which was maybe 2011, 12, 13, around that time, John Shulman was a PhD student initially kind of

12:36.880 --> 12:44.480
 driving it forward here. And kind of the way we thought about it at the time was if you think

12:44.480 --> 12:51.360
 about rectified linear units or kind of rectifier type neural networks, what do you get? You get

12:51.360 --> 12:56.320
 something that's piecewise linear feedback control. And if you look at the literature,

12:56.960 --> 13:02.080
 linear feedback control is extremely successful, can solve many, many problems surprisingly well.

13:03.520 --> 13:07.200
 I remember, for example, when we did helicopter flight, if you're in a stationary flight regime,

13:07.200 --> 13:12.080
 not a non stationary, but a stationary flight regime like hover, you can use linear feedback

13:12.080 --> 13:16.960
 control to stabilize the helicopter, a very complex dynamical system. But the controller

13:16.960 --> 13:22.240
 is relatively simple. And so I think that's a big part of is that if you do feedback control,

13:22.240 --> 13:25.280
 even though the system you control can be very, very complex, often,

13:26.000 --> 13:31.520
 relatively simple control architectures can already do a lot. But then also just linear

13:31.520 --> 13:35.840
 is not good enough. And so one way you can think of these neural networks is that in some of the

13:35.840 --> 13:40.880
 tile the space, which people were already trying to do more by hand or with finite state machines,

13:40.880 --> 13:44.560
 say this linear controller here, this linear controller here, neural network,

13:44.560 --> 13:48.160
 learns to tell the spin say linear controller here, another linear controller here,

13:48.160 --> 13:52.000
 but it's more subtle than that. And so it's benefiting from this linear control aspect is

13:52.000 --> 13:57.760
 benefiting from the tiling, but it's somehow tiling it one dimension at a time. Because if

13:57.760 --> 14:04.160
 let's say you have a two layer network, even that hidden layer, you make a transition from active

14:04.160 --> 14:09.600
 to inactive or the other way around, that is essentially one axis, but not axis aligned, but

14:09.600 --> 14:15.200
 one direction that you change. And so you have this kind of very gradual tiling of the space,

14:15.200 --> 14:19.840
 we have a lot of sharing between the linear controllers that tile the space. And that was

14:19.840 --> 14:25.280
 always my intuition as to why to expect that this might work pretty well. It's essentially

14:25.280 --> 14:30.000
 leveraging the fact that linear feedback control is so good. But of course, not enough. And this

14:30.000 --> 14:35.520
 is a gradual tiling of the space with linear feedback controls that share a lot of expertise

14:35.520 --> 14:41.120
 across them. So that that's, that's really nice intuition. But do you think that scales to the

14:41.120 --> 14:47.040
 more and more general problems of when you start going up the number of control dimensions,

14:48.160 --> 14:55.280
 when you start going down in terms of how often you get a clean reward signal,

14:55.280 --> 15:00.960
 does that intuition carry forward to those crazy or weirder worlds that we think of as the real

15:00.960 --> 15:10.000
 world? So I think where things get really tricky in the real world compared to the things we've

15:10.000 --> 15:13.920
 looked at so far with great success and reinforcement learning is

15:16.160 --> 15:21.920
 the time scales, which takes us to an extreme. So when you think about the real world, I mean,

15:22.800 --> 15:28.560
 I don't know, maybe some student decided to do a PhD here, right? Okay, that's that's a decision,

15:28.560 --> 15:34.000
 that's a very high level decision. But if you think about their lives, I mean, any person's life,

15:34.000 --> 15:39.360
 it's a sequence of muscle fiber contractions and relaxations. And that's how you interact with

15:39.360 --> 15:44.480
 the world. And that's a very high frequency control thing. But it's ultimately what you do

15:44.480 --> 15:49.280
 and how you affect the world. Until I guess we have brain readings, you can maybe do it slightly

15:49.280 --> 15:55.120
 differently. But typically, that's how you affect the world. And the decision of doing a PhD is

15:55.120 --> 16:00.240
 like so abstract relative to what you're actually doing in the world. And I think that's where

16:00.240 --> 16:07.360
 credit assignment becomes just completely beyond what any current RL algorithm can do. And we need

16:07.360 --> 16:13.360
 hierarchical reasoning at a level that is just not available at all yet. Where do you think we can

16:13.360 --> 16:19.360
 pick up hierarchical reasoning by which mechanisms? Yeah, so maybe let me highlight what I think the

16:19.360 --> 16:27.600
 limitations are of what already was done 20, 30 years ago. In fact, you'll find reasoning systems

16:27.600 --> 16:33.200
 that reason over relatively long horizons. But the problem is that they were not grounded in the real

16:33.200 --> 16:43.040
 world. So people would have to hand design some kind of logical, dynamical descriptions of the

16:43.040 --> 16:49.120
 world. And that didn't tie into perception. And so that didn't tie into real objects and so forth.

16:49.120 --> 16:57.920
 And so that was a big gap. Now with deep learning, we start having the ability to really see with

16:57.920 --> 17:02.800
 sensors process that and understand what's in the world. And so it's a good time to try to

17:02.800 --> 17:08.080
 bring these things together. I see a few ways of getting there. One way to get there would be to say

17:08.080 --> 17:12.160
 deep learning can get bolted on somehow to some of these more traditional approaches.

17:12.160 --> 17:16.160
 Now bolted on would probably mean you need to do some kind of end to end training,

17:16.160 --> 17:21.840
 where you say, my deep learning processing somehow leads to a representation that in term

17:22.720 --> 17:29.680
 uses some kind of traditional underlying dynamical systems that can be used for planning.

17:29.680 --> 17:33.920
 And that's, for example, the direction of Eve Tamar and Thanard Kuritach here have been pushing

17:33.920 --> 17:38.800
 with causal info again. And of course, other people too, that that's that's one way. Can we

17:38.800 --> 17:43.520
 somehow force it into the form factor that is amenable to reasoning?

17:43.520 --> 17:50.160
 Another direction we've been thinking about for a long time and didn't make any progress on

17:50.160 --> 17:56.880
 was more information theoretic approaches. So the idea there was that what it means to take

17:56.880 --> 18:03.840
 high level action is to take and choose a latent variable now that tells you a lot about what's

18:03.840 --> 18:08.640
 going to be the case in the future, because that's what it means to to take a high level action.

18:08.640 --> 18:14.480
 I say, okay, what I decide I'm going to navigate to the gas station because I need to get

18:14.480 --> 18:18.800
 gas from my car. Well, that'll now take five minutes to get there. But the fact that I get

18:18.800 --> 18:23.200
 there, I could already tell that from the high level action I took much earlier.

18:24.480 --> 18:30.080
 That we had a very hard time getting success with, not saying it's a dead end,

18:30.080 --> 18:34.160
 necessarily, but we had a lot of trouble getting that to work. And then we started revisiting

18:34.160 --> 18:39.600
 the notion of what are we really trying to achieve? What we're trying to achieve is

18:39.600 --> 18:42.880
 not necessarily a hierarchy per se, but you could think about what does hierarchy give us?

18:44.160 --> 18:50.560
 What we hope it would give us is better credit assignment. What is better credit assignment

18:50.560 --> 18:58.640
 is giving us, it gives us faster learning. And so faster learning is ultimately maybe

18:58.640 --> 19:03.840
 what we're after. And so that's where we ended up with the RL squared paper on learning to

19:03.840 --> 19:10.640
 reinforcement learn, which at a time Rocky Dwan led. And that's exactly the meta learning

19:10.640 --> 19:15.040
 approach where we say, okay, we don't know how to design hierarchy. We know what we want to get

19:15.040 --> 19:20.000
 from it. Let's just enter and optimize for what we want to get from it and see if it might emerge.

19:20.000 --> 19:24.720
 And we saw things emerge. The maze navigation had consistent motion down hallways,

19:25.920 --> 19:29.520
 which is what you want. A hierarchical control should say, I want to go down this hallway.

19:29.520 --> 19:33.040
 And then when there is an option to take a turn, I can decide whether to take a turn or not and

19:33.040 --> 19:38.480
 repeat, even had the notion of, where have you been before or not to not revisit places you've

19:38.480 --> 19:45.840
 been before? It still didn't scale yet to the real world kind of scenarios I think you had in mind,

19:45.840 --> 19:50.000
 but it was some sign of life that maybe you can meta learn these hierarchical concepts.

19:51.040 --> 19:58.000
 I mean, it seems like through these meta learning concepts, we get at the, what I think is one of

19:58.000 --> 20:05.120
 the hardest and most important problems of AI, which is transfer learning. So it's generalization.

20:06.240 --> 20:12.160
 How far along this journey towards building general systems are we being able to do transfer

20:12.160 --> 20:18.320
 learning? Well, so there's some signs that you can generalize a little bit. But do you think

20:18.320 --> 20:25.360
 we're on the right path or totally different breakthroughs are needed to be able to transfer

20:25.360 --> 20:34.000
 knowledge between different learned models? Yeah, I'm pretty torn on this in that I think

20:34.000 --> 20:44.400
 there are some very impressive results already, right? I mean, I would say when even with the

20:44.400 --> 20:50.160
 initial kind of big breakthrough in 2012 with Alex net, right, the initial, the initial thing is,

20:50.160 --> 20:57.600
 okay, great. This does better on image net hands image recognition. But then immediately thereafter,

20:57.600 --> 21:04.080
 there was of course the notion that wow, what was learned on image net, and you now want to solve

21:04.080 --> 21:11.280
 a new task, you can fine tune Alex net for new tasks. And that was often found to be the even

21:11.280 --> 21:15.920
 bigger deal that you learn something that was reusable, which was not often the case before

21:15.920 --> 21:19.520
 usually machine learning, you learn something for one scenario. And that was it. And that's

21:19.520 --> 21:23.200
 really exciting. I mean, that's just a huge application. That's probably the biggest

21:23.200 --> 21:28.960
 success of transfer learning today, if in terms of scope and impact. That was a huge breakthrough.

21:28.960 --> 21:37.040
 And then recently, I feel like similar kind of by scaling things up, it seems like this has been

21:37.040 --> 21:41.440
 expanded upon like people training even bigger networks, they might transfer even better. If

21:41.440 --> 21:46.480
 you look that, for example, some of the opening results on language models. And so in the recent

21:46.480 --> 21:53.600
 Google results on language models, they are learned for just prediction. And then they get

21:54.320 --> 21:59.600
 reused for other tasks. And so I think there is something there where somehow if you train a

21:59.600 --> 22:05.200
 big enough model on enough things, it seems to transfer some deep mind results that I thought

22:05.200 --> 22:12.160
 were very impressive, the unreal results, where it was learning to navigate mazes in ways where

22:12.160 --> 22:16.880
 it wasn't just doing reinforcement learning, but it had other objectives was optimizing for. So I

22:16.880 --> 22:23.680
 think there's a lot of interesting results already. I think maybe where it's hard to wrap my head

22:23.680 --> 22:30.160
 around this, to which extent or when do we call something generalization, right? Or the levels

22:30.160 --> 22:37.360
 of generalization involved in these different tasks, right? So you draw this, by the way, just

22:37.360 --> 22:43.280
 to frame things. I've heard you say somewhere, it's the difference in learning to master versus

22:43.280 --> 22:49.680
 learning to generalize. That it's a nice line to think about. And I guess you're saying it's a gray

22:49.680 --> 22:54.640
 area of what learning to master and learning to generalize where one starts.

22:54.640 --> 22:58.800
 I think I might have heard this. I might have heard it somewhere else. And I think it might have

22:58.800 --> 23:05.120
 been one of your interviews, maybe the one with Yoshua Benjamin, 900% sure. But I like the example

23:05.120 --> 23:12.000
 and I'm going to not sure who it was, but the example was essentially if you use current deep

23:12.000 --> 23:20.480
 learning techniques, what we're doing to predict, let's say the relative motion of our planets,

23:20.480 --> 23:27.680
 it would do pretty well. But then now if a massive new mass enters our solar system,

23:28.320 --> 23:32.880
 it would probably not predict what will happen, right? And that's a different kind of

23:32.880 --> 23:38.400
 generalization. That's a generalization that relies on the ultimate simplest explanation

23:38.400 --> 23:42.640
 that we have available today to explain the motion of planets, whereas just pattern recognition

23:42.640 --> 23:48.160
 could predict our current solar system motion pretty well. No problem. And so I think that's

23:48.160 --> 23:53.920
 an example of a kind of generalization that is a little different from what we've achieved so far.

23:54.480 --> 24:01.360
 And it's not clear if just, you know, regularizing more and forcing it to come up with a simpler,

24:01.360 --> 24:05.280
 simpler, simpler explanation. Look, this is not simple, but that's what physics researchers do,

24:05.280 --> 24:10.000
 right, to say, can I make this even simpler? How simple can I get this? What's the simplest

24:10.000 --> 24:14.560
 equation that can explain everything, right? The master equation for the entire dynamics of the

24:14.560 --> 24:20.960
 universe. We haven't really pushed that direction as hard in deep learning, I would say. Not sure

24:20.960 --> 24:24.960
 if it should be pushed, but it seems a kind of generalization you get from that that you don't

24:24.960 --> 24:30.400
 get in our current methods so far. So I just talked to Vladimir Vapnik, for example, who was

24:30.400 --> 24:39.200
 a statistician in statistical learning, and he kind of dreams of creating the E equals Mc

24:39.200 --> 24:44.400
 squared for learning, right, the general theory of learning. Do you think that's a fruitless pursuit

24:46.480 --> 24:50.560
 in the near term, within the next several decades?

24:51.680 --> 24:56.800
 I think that's a really interesting pursuit. And in the following sense, in that there is a

24:56.800 --> 25:05.440
 lot of evidence that the brain is pretty modular. And so I wouldn't maybe think of it as the theory,

25:05.440 --> 25:12.480
 maybe, the underlying theory, but more kind of the principle where there have been findings where

25:14.160 --> 25:20.240
 people who are blind will use the part of the brain usually used for vision for other functions.

25:20.240 --> 25:26.800
 And even after some kind of, if people get rewired in some way, they might be able to reuse parts of

25:26.800 --> 25:35.040
 their brain for other functions. And so what that suggests is some kind of modularity. And I think

25:35.040 --> 25:41.120
 it is a pretty natural thing to strive for to see, can we find that modularity? Can we find this

25:41.120 --> 25:45.440
 thing? Of course, it's not every part of the brain is not exactly the same. Not everything can be

25:45.440 --> 25:50.080
 rewired arbitrarily. But if you think of things like the neocortex, which is a pretty big part of

25:50.080 --> 25:56.880
 the brain, that seems fairly modular from what the findings so far. Can you design something

25:56.880 --> 26:01.840
 equally modular? And if you can just grow it, it becomes more capable, probably. I think that would

26:01.840 --> 26:07.200
 be the kind of interesting underlying principle to shoot for that is not unrealistic.

26:07.200 --> 26:14.400
 Do you think you prefer math or empirical trial and error for the discovery of the essence of what

26:14.400 --> 26:19.680
 it means to do something intelligent? So reinforcement learning embodies both groups, right?

26:19.680 --> 26:25.760
 To prove that something converges, prove the bounds. And then at the same time, a lot of those

26:25.760 --> 26:31.280
 successes are, well, let's try this and see if it works. So which do you gravitate towards? How do

26:31.280 --> 26:40.960
 you think of those two parts of your brain? So maybe I would prefer we could make the progress

26:41.600 --> 26:46.560
 with mathematics. And the reason maybe I would prefer that is because often if you have something you

26:46.560 --> 26:54.080
 can mathematically formalize, you can leapfrog a lot of experimentation. And experimentation takes

26:54.080 --> 27:01.440
 a long time to get through. And a lot of trial and error, reinforcement learning, your research

27:01.440 --> 27:05.040
 process. But you need to do a lot of trial and error before you get to a success. So if you can

27:05.040 --> 27:10.400
 leapfrog that, to my mind, that's what the math is about. And hopefully once you do a bunch of

27:10.400 --> 27:15.600
 experiments, you start seeing a pattern, you can do some derivations that leapfrog some experiments.

27:16.240 --> 27:20.160
 But I agree with you. I mean, in practice, a lot of the progress has been such that we have not

27:20.160 --> 27:25.840
 been able to find the math that allows it to leapfrog ahead. And we are kind of making gradual

27:25.840 --> 27:30.480
 progress one step at a time. A new experiment here, a new experiment there that gives us new

27:30.480 --> 27:35.280
 insights and gradually building up, but not getting to something yet where we're just, okay,

27:35.280 --> 27:39.920
 here's an equation that now explains how, you know, that would be have been two years of

27:39.920 --> 27:44.880
 experimentation to get there. But this tells us what the results going to be. Unfortunately,

27:44.880 --> 27:52.800
 unfortunately, not so much yet. Not so much yet. But your hope is there. In trying to teach robots

27:52.800 --> 28:01.200
 or systems to do everyday tasks, or even in simulation, what do you think you're more excited

28:01.200 --> 28:10.560
 about? imitation learning or self play. So letting robots learn from humans, or letting robots plan

28:10.560 --> 28:18.240
 their own, try to figure out in their own way, and eventually play, eventually interact with humans,

28:18.240 --> 28:23.200
 or solve whatever problem is. What's the more exciting to you? What's more promising you think

28:23.200 --> 28:34.240
 is a research direction? So when we look at self play, what's so beautiful about it is,

28:34.240 --> 28:37.680
 goes back to kind of the challenges in reinforcement learning. So the challenge

28:37.680 --> 28:43.200
 of reinforcement learning is getting signal. And if you don't never succeed, you don't get any signal.

28:43.200 --> 28:49.040
 In self play, you're on both sides. So one of you succeeds. And the beauty is also one of you

28:49.040 --> 28:53.520
 fails. And so you see the contrast, you see the one version of me that did better than the other

28:53.520 --> 28:58.400
 version. And so every time you play yourself, you get signal. And so whenever you can turn

28:58.400 --> 29:04.160
 something into self play, you're in a beautiful situation where you can naturally learn much

29:04.160 --> 29:10.080
 more quickly than in most other reinforcement learning environments. So I think, I think if

29:10.080 --> 29:15.760
 somehow we can turn more reinforcement learning problems into self play formulations, that would

29:15.760 --> 29:21.760
 go really, really far. So far, self play has been largely around games where there is natural

29:21.760 --> 29:25.440
 opponents. But if we could do self play for other things, and let's say, I don't know,

29:25.440 --> 29:29.360
 a robot learns to build a house, I mean, that's a pretty advanced thing to try to do for a robot,

29:29.360 --> 29:34.240
 but maybe it tries to build a hut or something. If that can be done through self play, it would

29:34.240 --> 29:38.560
 learn a lot more quickly if somebody can figure it out. And I think that would be something where

29:38.560 --> 29:42.560
 it goes closer to kind of the mathematical leapfrogging where somebody figures out a

29:42.560 --> 29:47.680
 formalism to say, okay, any RL problem by playing this and this idea, you can turn it

29:47.680 --> 29:50.480
 into a self play problem where you get signal a lot more easily.

29:52.400 --> 29:57.680
 Reality is many problems, we don't know how to turn to self play. And so either we need to provide

29:57.680 --> 30:02.640
 detailed reward. That doesn't just reward for achieving a goal, but rewards for making progress,

30:02.640 --> 30:06.480
 and that becomes time consuming. And once you're starting to do that, let's say you want a robot

30:06.480 --> 30:09.920
 to do something, you need to give all this detailed reward. Well, why not just give a

30:09.920 --> 30:15.920
 demonstration? Because why not just show the robot. And now the question is, how do you show

30:15.920 --> 30:20.240
 the robot? One way to show is to tally operator robot and then robot really experiences things.

30:20.800 --> 30:24.480
 And that's nice, because that's really high signal to noise ratio data. And we've done a lot

30:24.480 --> 30:29.360
 of that. And you teach your robot skills. In just 10 minutes, you can teach your robot a new basic

30:29.360 --> 30:33.360
 skill, like, okay, pick up the bottle, place it somewhere else. That's a skill, no matter where

30:33.360 --> 30:38.000
 the bottle starts, maybe it always goes on to a target or something. That's fairly easy to teach

30:38.000 --> 30:43.120
 your robot with teleop. Now, what's even more interesting, if you can now teach your robot

30:43.120 --> 30:48.480
 through third person learning, where the robot watches you do something, and doesn't experience

30:48.480 --> 30:52.880
 it, but just watches it and says, okay, well, if you're showing me that, that means I should

30:52.880 --> 30:56.880
 be doing this. And I'm not going to be using your hand, because I don't get to control your hand,

30:56.880 --> 31:02.000
 but I'm going to use my hand, I do that mapping. And so that's where I think one of the big breakthroughs

31:02.000 --> 31:07.520
 has happened this year. This was led by Chelsea Finn here. It's almost like learning a machine

31:07.520 --> 31:12.000
 translation for demonstrations where you have a human demonstration and the robot learns to

31:12.000 --> 31:17.440
 translate it into what it means for the robot to do it. And that was a meta learning formulation,

31:17.440 --> 31:23.440
 learn from one to get the other. And that I think opens up a lot of opportunities to learn a lot

31:23.440 --> 31:28.080
 more quickly. So my focus is on autonomous vehicles. Do you think this approach of third

31:28.080 --> 31:33.040
 person watching is the autonomous driving is amenable to this kind of approach?

31:33.840 --> 31:42.080
 So for autonomous driving, I would say it's third person is slightly easier. And the reason I'm

31:42.080 --> 31:48.320
 going to say it's slightly easier to do with third person is because the car dynamics are very well

31:48.320 --> 31:56.560
 understood. So the easier than first person, you mean, or easier than. So I think the distinction

31:56.560 --> 32:01.680
 between third person and first person is not a very important distinction for autonomous driving.

32:01.680 --> 32:07.760
 They're very similar. Because the distinction is really about who turns the steering wheel.

32:07.760 --> 32:15.280
 And or maybe let me put it differently. How to get from a point where you are now to a point,

32:15.280 --> 32:19.120
 let's say a couple of meters in front of you. And that's a problem that's very well understood.

32:19.120 --> 32:22.480
 And that's the only distinction between third and first person there. Whereas with the robot

32:22.480 --> 32:26.720
 manipulation, interaction forces are very complex. And it's still a very different thing.

32:27.840 --> 32:33.840
 For autonomous driving, I think there's still the question imitation versus RL.

32:33.840 --> 32:39.520
 Well, so imitation gives you a lot more signal. I think where imitation is lacking and needs

32:39.520 --> 32:47.600
 some extra machinery is it doesn't in its normal format, doesn't think about goals or objectives.

32:48.480 --> 32:52.240
 And of course, there are versions of imitation learning, inverse reinforcement learning type

32:52.240 --> 32:57.440
 imitation, which also thinks about goals. I think then we're getting much closer. But I think it's

32:57.440 --> 33:05.120
 very hard to think of a fully reactive car generalizing well, if it really doesn't have a notion

33:05.120 --> 33:10.720
 of objectives to generalize well to the kind of general that you would want, you want more than

33:10.720 --> 33:15.200
 just that reactivity that you get from just behavioral cloning slash supervised learning.

33:17.040 --> 33:22.560
 So a lot of the work, whether it's self play or even imitation learning would benefit

33:22.560 --> 33:27.440
 significantly from simulation, from effective simulation, and you're doing a lot of stuff

33:27.440 --> 33:32.400
 in the physical world and in simulation, do you have hope for greater and greater

33:33.520 --> 33:40.160
 power of simulation loop being boundless, eventually, to where most of what we need

33:40.160 --> 33:45.600
 to operate in the physical world, what could be simulated to a degree that's directly

33:45.600 --> 33:54.720
 transferable to the physical world? Are we still very far away from that? So I think

33:55.840 --> 34:03.200
 we could even rephrase that question in some sense, please. And so the power of simulation,

34:04.720 --> 34:09.760
 as simulators get better and better, of course, becomes stronger, and we can learn more in

34:09.760 --> 34:13.760
 simulation. But there's also another version, which is where you say the simulator doesn't

34:13.760 --> 34:19.120
 even have to be that precise. As long as it's somewhat representative. And instead of trying

34:19.120 --> 34:24.480
 to get one simulator that is sufficiently precise to learn and transfer really well to the real

34:24.480 --> 34:29.200
 world, I'm going to build many simulators, ensemble of simulators, ensemble of simulators,

34:30.080 --> 34:35.120
 not any single one of them is sufficiently representative of the real world such that

34:35.120 --> 34:41.760
 it would work if you train in there. But if you train in all of them, then there is something

34:41.760 --> 34:47.840
 that's good in all of them. The real world will just be, you know, another one of them. That's,

34:47.840 --> 34:50.720
 you know, not identical to any one of them, but just another one of them.

34:50.720 --> 34:53.120
 Now, this sample from the distribution of simulators.

34:53.120 --> 34:53.360
 Exactly.

34:53.360 --> 34:57.600
 We do live in a simulation. So this is just one, one other one.

34:57.600 --> 35:03.440
 I'm not sure about that. But yeah, it's definitely a very advanced simulator if it is.

35:03.440 --> 35:08.960
 Yeah, it's a pretty good one. I've talked to Russell. It's something you think about a little bit

35:08.960 --> 35:13.120
 too. Of course, you're like really trying to build these systems. But do you think about the future

35:13.120 --> 35:18.880
 of AI? A lot of people have concern about safety. How do you think about AI safety as you build

35:18.880 --> 35:24.960
 robots that are operating the physical world? What is, yeah, how do you approach this problem

35:24.960 --> 35:27.440
 in an engineering kind of way in a systematic way?

35:29.200 --> 35:36.720
 So when a robot is doing things, you kind of have a few notions of safety to worry about. One is that

35:36.720 --> 35:43.760
 the robot is physically strong and of course could do a lot of damage. Same for cars, which we can

35:43.760 --> 35:49.360
 think of as robots do in some way. And this could be completely unintentional. So it could be not

35:49.360 --> 35:54.240
 the kind of long term AI safety concerns that, okay, AI is smarter than us. And now what do we do?

35:54.240 --> 35:57.760
 But it could be just very practical. Okay, this robot, if it makes a mistake,

35:58.800 --> 36:04.080
 what are the results going to be? Of course, simulation comes in a lot there to test in simulation.

36:04.080 --> 36:10.960
 It's a difficult question. And I'm always wondering, like I always wonder, let's say you look at,

36:10.960 --> 36:14.000
 let's go back to driving, because a lot of people know driving well, of course.

36:15.120 --> 36:20.800
 What do we do to test somebody for driving, right, to get a driver's license? What do they

36:20.800 --> 36:27.680
 really do? I mean, you fill out some tests, and then you drive and I mean, for a few minutes,

36:27.680 --> 36:34.800
 it's suburban California, that driving test is just you drive around the block, pull over, you

36:34.800 --> 36:39.280
 do a stop sign successfully, and then, you know, you pull over again, and you're pretty much done.

36:40.000 --> 36:46.720
 And you're like, okay, if a self driving car did that, would you trust it that it can drive?

36:46.720 --> 36:49.840
 And I'd be like, no, that's not enough for me to trust it. But somehow for humans,

36:50.560 --> 36:54.480
 we've figured out that somebody being able to do that is representative

36:54.480 --> 36:59.840
 of them being able to do a lot of other things. And so I think somehow for humans,

36:59.840 --> 37:05.200
 we figured out representative tests of what it means if you can do this, what you can really do.

37:05.760 --> 37:09.840
 Of course, testing humans, humans don't want to be tested at all times. Self driving cars or

37:09.840 --> 37:13.760
 robots could be tested more often probably, you can have replicas that get tested and are known

37:13.760 --> 37:19.600
 to be identical because they use the same neural net and so forth. But still, I feel like we don't

37:19.600 --> 37:25.040
 have this kind of unit tests or proper tests for robots. And I think there's something very

37:25.040 --> 37:29.440
 interesting to be thought about there, especially as you update things, your software improves,

37:29.440 --> 37:34.640
 you have a better self driving car suite, you update it. How do you know it's indeed more

37:34.640 --> 37:41.440
 capable on everything than what you had before that you didn't have any bad things creep into it?

37:41.440 --> 37:45.680
 So I think that's a very interesting direction of research that there is no real solution yet,

37:45.680 --> 37:50.640
 except that somehow for humans, we do because we say, okay, you have a driving test, you passed,

37:50.640 --> 37:55.760
 you can go on the road now and you must have accents every like a million or 10 million miles,

37:55.760 --> 38:01.520
 something pretty phenomenal compared to that short test that is being done.

38:01.520 --> 38:06.000
 So let me ask, you've mentioned, you've mentioned that Andrew Ang, by example,

38:06.000 --> 38:11.440
 showed you the value of kindness. And do you think the space of

38:11.440 --> 38:20.240
 of policies, good policies for humans and for AI is populated by policies that

38:21.440 --> 38:28.880
 with kindness or ones that are the opposite, exploitation, even evil. So if you just look

38:28.880 --> 38:34.400
 at the sea of policies we operate under as human beings, or if AI system had to operate in this

38:34.400 --> 38:39.440
 real world, do you think it's really easy to find policies that are full of kindness,

38:39.440 --> 38:44.480
 like we naturally fall into them? Or is it like a very hard optimization problem?

38:47.920 --> 38:52.720
 I mean, there is kind of two optimizations happening for humans, right? So for humans,

38:52.720 --> 38:57.440
 there's kind of the very long term optimization, which evolution has done for us. And we're kind of

38:57.440 --> 39:02.640
 predisposed to like certain things. And that's in some sense, what makes our learning easier,

39:02.640 --> 39:10.000
 because I mean, we know things like pain and hunger and thirst. And the fact that we know about those

39:10.000 --> 39:13.840
 is not something that we were taught. That's kind of innate. When we're hungry, we're unhappy.

39:13.840 --> 39:20.720
 When we're thirsty, we're unhappy. When we have pain, we're unhappy. And ultimately evolution

39:20.720 --> 39:25.040
 built that into us to think about those things. And so I think there is a notion that it seems

39:25.040 --> 39:33.840
 somehow humans evolved in general to prefer to get along in some ways. But at the same time,

39:33.840 --> 39:43.040
 also to be very territorial and kind of centric to their own tribe. It seems like that's the kind

39:43.040 --> 39:47.360
 of space we converged on to. I mean, I'm not an expert in anthropology, but it seems like we're

39:47.360 --> 39:54.480
 very kind of good within our own tribe, but need to be taught to be nice to other tribes.

39:54.480 --> 39:58.000
 Well, if you look at Steven Pinker, he highlights this pretty nicely in

40:00.720 --> 40:05.520
 Better Angels of Our Nature, where he talks about violence decreasing over time consistently.

40:05.520 --> 40:11.360
 So whatever tension, whatever teams we pick, it seems that the long arc of history goes

40:11.360 --> 40:17.840
 towards us getting along more and more. So do you think that

40:17.840 --> 40:27.280
 do you think it's possible to teach RRL based robots this kind of kindness, this kind of ability

40:27.280 --> 40:33.040
 to interact with humans, this kind of policy, even to let me ask, let me ask upon one, do you think

40:33.040 --> 40:38.800
 it's possible to teach RRL based robot to love a human being and to inspire that human to love

40:38.800 --> 40:48.080
 the robot back? So to like a RRL based algorithm that leads to a happy marriage? That's an interesting

40:48.080 --> 40:56.080
 question. Maybe I'll answer it with another question, right? Because I mean, but I'll come

40:56.080 --> 41:02.000
 back to it. So another question you can have is okay. I mean, how close does some people's

41:02.000 --> 41:09.760
 happiness get from interacting with just a really nice dog? Like, I mean, dogs, you come home,

41:09.760 --> 41:14.000
 that's what dogs do. They greet you. They're excited. It makes you happy when you come home

41:14.000 --> 41:17.600
 to your dog. You're just like, okay, this is exciting. They're always happy when I'm here.

41:18.160 --> 41:22.560
 I mean, if they don't greet you, because maybe whatever, your partner took them on a trip or

41:22.560 --> 41:27.600
 something, you might not be nearly as happy when you get home, right? And so the kind of,

41:27.600 --> 41:33.600
 it seems like the level of reasoning a dog has is pretty sophisticated, but then it's still not yet

41:33.600 --> 41:38.240
 at the level of human reasoning. And so it seems like we don't even need to achieve human level

41:38.240 --> 41:44.320
 reasoning to get like very strong affection with humans. And so my thinking is, why not, right?

41:44.320 --> 41:51.360
 Why couldn't, with an AI, couldn't we achieve the kind of level of affection that humans feel

41:51.360 --> 41:59.280
 among each other or with friendly animals and so forth? So question, is it a good thing for us

41:59.280 --> 42:07.040
 or not? That's another thing, right? Because I mean, but I don't see why not. Why not? Yeah.

42:07.040 --> 42:12.640
 So Elon Musk says love is the answer. Maybe he should say love is the objective function and

42:12.640 --> 42:19.280
 then RL is the answer, right? Well, maybe. Oh, Peter, thank you so much. I don't want to take

42:19.280 --> 42:23.360
 up more of your time. Thank you so much for talking today. Well, thanks for coming by.

42:23.360 --> 42:53.200
 Great to have you visit.