Daniel Kahneman: Deep Learning (System 1 and System 2)

00:00:00.000 | So we're not talking about humans, but if we think about building artificial intelligence

00:00:07.280 | systems, robots, do you think all the features and bugs that you have highlighted in human

00:00:14.420 | beings are useful for constructing AI systems?

00:00:18.540 | So both systems are useful for perhaps instilling in robots?

00:00:23.480 | What is happening these days is that actually what is happening in deep learning is more

00:00:33.960 | like a system one product than like a system two product.

00:00:37.800 | I mean, deep learning matches patterns and anticipate what's going to happen, so it's

00:00:43.680 | highly predictive.

00:00:46.160 | But what deep learning doesn't have, and many people think that this is the critical, it

00:00:54.800 | doesn't have the ability to reason, so there is no system two there.

00:00:59.840 | But I think very importantly, it doesn't have any causality or any way to represent meaning

00:01:06.560 | and to represent real interaction.

00:01:08.960 | So until that is solved, what can be accomplished is marvelous and very exciting, but limited.

00:01:19.840 | That's actually really nice to think of current advances in machine learning as essentially

00:01:24.480 | system one advances.

00:01:26.560 | So how far can we get with just system one?

00:01:29.940 | If we think of deep learning and artificial intelligence systems?

00:01:34.360 | It's very clear that deep mind has already gone way beyond what people thought was possible.

00:01:41.640 | I think the thing that has impressed me most about the developments in AI is the speed.

00:01:48.080 | It's that things, at least in the context of deep learning, and maybe this is about

00:01:53.520 | to slow down, but things moved a lot faster than anticipated.

00:01:59.200 | The transition from solving chess to solving Go, that's bewildering how quickly it went.

00:02:09.760 | The move from AlphaGo to AlphaZero is sort of bewildering the speed at which they accomplished

00:02:15.840 | that.

00:02:16.840 | Now, clearly, there are many problems that you can solve that way, but there are some

00:02:24.680 | problems for which you need something else.

00:02:28.280 | Something like reasoning.

00:02:29.280 | Well, reasoning and also, you know, one of the real mysteries, psychologist Gary Marcus,

00:02:37.480 | who is also a critic of AI, I mean, what he points out, and I think he has a point, is

00:02:47.360 | that humans learn quickly.

00:02:53.160 | Humans don't need a million examples, they need two or three examples.

00:02:59.200 | So clearly, there is a fundamental difference.

00:03:02.440 | And what enables a machine to learn quickly, what you have to build into the machine, because

00:03:11.160 | it's clear that you have to build some expectations or something in the machine to make it ready

00:03:16.860 | to learn quickly, that at the moment seems to be unsolved.

00:03:23.320 | I'm pretty sure that DeepMind is working on it, but if they have solved it, I haven't

00:03:31.400 | heard yet.

00:03:32.400 | They're trying to actually, them and OpenAI are trying to start to get to use neural networks

00:03:38.500 | to reason.

00:03:40.160 | So assembled knowledge, of course, causality is, temporal causality is out of reach to

00:03:47.660 | most everybody.

00:03:49.640 | You mentioned the benefits of System 1 is essentially that it's fast, allows us to

00:03:54.040 | function in the world.

00:03:55.040 | Fast and skilled, yeah.

00:03:56.040 | It's skilled.

00:03:57.040 | And it has a model of the world.

00:04:00.460 | You know, in a sense, I mean, there was the earlier phase of AI attempted to model reasoning,

00:04:10.680 | and they were moderately successful, but, you know, reasoning by itself doesn't get

00:04:14.440 | you much.

00:04:18.360 | Deep learning has been much more successful in terms of, you know, what they can do.

00:04:22.920 | But now, it's an interesting question, whether it's approaching its limits.

00:04:28.120 | What do you think?

00:04:29.620 | I think absolutely.

00:04:30.620 | So I just talked to Gian Lacoon, you mentioned, you know, so he thinks that the limits, we're

00:04:38.800 | not going to hit the limits with neural networks, that ultimately this kind of System 1 pattern

00:04:43.840 | matching will start to start to look like System 2 without significant transformation

00:04:52.320 | of the architecture.

00:04:53.960 | So I'm more with the majority of the people who think that yes, neural networks will hit

00:04:59.060 | a limit in their capability.

00:05:01.140 | He, on the one hand, I have heard him tell the Mises-Sabies essentially that, you know,

00:05:07.860 | what they have accomplished is not a big deal, that they have just touched, that basically,

00:05:13.220 | you know, they can't do unsupervised learning in an effective way.

00:05:18.900 | But you're telling me that he thinks that the current, within the current architecture,

00:05:24.340 | you can do causality and reasoning?

00:05:26.600 | So he's very much a pragmatist in a sense that's saying that we're very far away, that

00:05:31.520 | there's still, I think there's this idea that he says is we can only see one or two mountain

00:05:39.120 | peaks ahead and there might be either a few more after or thousands more after.

00:05:45.160 | So that kind of idea.

00:05:46.160 | I heard that metaphor.

00:05:48.160 | Right. But nevertheless, it doesn't see a, the final answer not fundamentally looking

00:05:56.880 | like one that we currently have.

00:05:59.340 | So neural networks being a huge part of that.

00:06:02.600 | Yeah.

00:06:03.600 | I mean, that's very likely because, because pattern matching is so much of what's going

00:06:09.200 | on.

00:06:10.200 | But.

00:06:11.200 | And you can think of neural networks as processing information sequentially.

00:06:14.200 | Yeah, I mean, you know, there is, there is an important aspect to, for example, you get

00:06:22.360 | systems that translate and they do a very good job, but they really don't know what

00:06:27.720 | they're talking about.

00:06:31.400 | And for that, I'm really quite surprised.

00:06:33.840 | For that, you would need, you would need an AI that has sensation, an AI that is in touch

00:06:41.200 | with the world.

00:06:42.200 | Yes, self-awareness and maybe even something that resembles consciousness kind of ideas.

00:06:48.880 | Certainly awareness of, you know, awareness of what's going on so that the words have

00:06:54.160 | meaning or can get, are in touch with some perception or some action.

00:06:59.920 | Yeah.

00:07:00.920 | So that's a big thing for Jan and as what he refers to as grounding to the physical

00:07:06.920 | space.

00:07:07.920 | So, so that's what we're talking about the same.

00:07:10.400 | Yeah.

00:07:11.400 | So how, how you ground.

00:07:13.440 | I mean the grounding, without grounding, then you get, you get a machine that doesn't know

00:07:18.960 | what it's talking about because it is talking about the world ultimately.

00:07:24.360 | The question, the open question is what it means to ground.

00:07:26.800 | I mean, we're very human centric in our thinking, but what does it mean for a machine to understand

00:07:33.720 | what it means to be in this world?

00:07:36.640 | Does it need to have a body?

00:07:38.800 | Does it need to have a finiteness like we humans have?

00:07:42.360 | All of these elements, it's a very, it's an open question.

00:07:46.360 | You know, I'm not sure about having a body, but having a perceptual system, having a body

00:07:50.640 | would be very helpful too.

00:07:52.280 | I mean, if, if you think about human mimicking human, but having a perception that seems

00:07:59.920 | to be essential so that you can build, you can accumulate knowledge about the world.

00:08:06.800 | However, you can, you can imagine a human completely paralyzed and there is a lot that

00:08:13.400 | the human brain could learn, you know, with a paralyzed body.

00:08:17.640 | So if we got a machine that could do that, that would be a big deal.

00:08:23.520 | And then the flip side of that, something you see in children and something in machine

00:08:29.000 | learning world is called active learning.

00:08:31.440 | Maybe it is also is being able to play with the world.

00:08:36.780 | How important for developing system one or system two, do you think it is to play with

00:08:43.040 | the world?

00:08:44.040 | To be able to interact with it?

00:08:45.040 | Well, certainly a lot, a lot of what you learn as you learn to anticipate the outcomes of

00:08:51.880 | your actions.

00:08:52.880 | I mean, you can see that how babies learn it, you know, with their hands, they, how

00:08:57.400 | they learn, you know, to connect, you know, the movements of their hands with something

00:09:03.600 | that clearly is something that happens in the brain and, and, and the ability of the

00:09:08.880 | brain to learn new patterns.

00:09:11.660 | So you know, it's the kind of thing that you get with artificial limbs that you connected

00:09:17.200 | and then people learn to operate the artificial limb, you know, really impressively quickly,

00:09:24.400 | at least from, from what I hear.

00:09:28.120 | So we have a system that is ready to learn the world through action.

00:09:33.200 | At the risk of going into way too mysterious of land, what do you think it takes to build

00:09:39.760 | a system like that?

00:09:42.920 | Obviously we're very far from understanding how the brain works, but how difficult is

00:09:49.000 | it to build this mind of ours?

00:09:52.880 | You know, I mean, I think that Jan LeCun's answer that we don't know how many mountains

00:09:57.600 | there are.

00:09:58.600 | I think that's a very good answer.

00:10:00.560 | I think that, you know, if you, if you look at what Ray Kurzweil is saying, that strikes

00:10:06.640 | me as off the wall, but, but I think people are much more realistic than that.

00:10:13.480 | We're actually, Demis Hassabis is and Jan is, and so the people are actually doing the

00:10:19.480 | work fairly realistic, I think.

00:10:23.840 | To maybe phrase it another way, from a perspective, not of building it, but from understanding

00:10:28.960 | it, how complicated are human beings in the following sense?

00:10:36.080 | You know, I work with autonomous vehicles and pedestrians, so we tried to model pedestrians.

00:10:41.400 | How difficult is it to model a human being, their perception of the world, the two systems

00:10:48.800 | they operate under, sufficiently to be able to predict whether the pedestrian is going

00:10:53.200 | to cross the road or not?

00:10:55.040 | I'm, you know, I'm fairly optimistic about that, actually, because what we're talking

00:11:00.800 | about is a huge amount of information that every vehicle has, and that feeds into one

00:11:10.280 | system, into one gigantic system.

00:11:13.220 | And so anything that any vehicle learns becomes part of what the whole system knows.

00:11:19.200 | And with a system multiplier like that, there is a lot that you can do.

00:11:25.240 | So human beings are very complicated, but, and, you know, system is going to make mistakes,

00:11:32.160 | but human makes mistakes.

00:11:34.160 | I think that they'll be able to, I think they are able to anticipate pedestrians, otherwise

00:11:41.520 | a lot would happen.

00:11:42.680 | They're able to, you know, they're able to get into a roundabout and into traffic, so

00:11:51.360 | they must know both to expect or to anticipate how people will react when they're sneaking

00:11:58.840 | in.

00:11:59.840 | And there's a lot of learning that's involved in that.

00:12:03.840 | Currently, the pedestrians are treated as things that cannot be hit, and they're not

00:12:12.040 | treated as agents with whom you interact in a game-theoretic way.

00:12:18.800 | So I mean, it's not, it's a totally open problem, and every time somebody tries to solve it,

00:12:24.320 | it seems to be harder than we think.

00:12:27.120 | And nobody's really tried to seriously solve the problem of that dance, because I'm not

00:12:32.360 | sure if you've thought about the problem of pedestrians, but you're really putting your

00:12:36.920 | life in the hands of the driver.

00:12:39.040 | You know, there is a dance, there's part of the dance that would be quite complicated,

00:12:45.000 | but for example, when I cross the street and there is a vehicle approaching, I look the

00:12:49.840 | driver in the eye, and I think many people do that.

00:12:54.000 | And you know, that's a signal that I'm sending, and I would be sending that machine to an

00:13:00.320 | autonomous vehicle and it had better understand it, because it means I'm crossing.

00:13:06.080 | So and there's another thing you do that actually, so I'll tell you what you do, because I've

00:13:12.040 | watched hundreds of hours of video on this, is when you step in the street, you do that

00:13:17.000 | before you step in the street, and when you step in the street, you actually look away.

00:13:21.040 | Look away.

00:13:22.040 | Yeah.

00:13:23.040 | Now, what is that?

00:13:25.880 | What that's saying is, I mean, you're trusting that the car, who hasn't slown down yet, will

00:13:31.400 | slow down.

00:13:32.400 | Yeah.

00:13:33.400 | And you're telling him, I'm committed.

00:13:35.760 | I mean, this is like in a game of chicken.

00:13:37.760 | So I'm committed, and if I'm committed, I'm looking away.

00:13:41.480 | So there is, you just have to stop.

00:13:44.960 | So the question is whether a machine that observes that needs to understand mortality.

00:13:49.760 | Here, I'm not sure that it's got to understand so much as it's got to anticipate.

00:13:59.680 | So and here, but you know, you're surprising me, because here I would think that maybe

00:14:08.280 | you can anticipate without understanding, because I think this is clearly what's happening

00:14:14.120 | in playing Go or in playing chess, there's a lot of anticipation and there is zero understanding.

00:14:20.440 | So I thought that you didn't need a model of the human and a model of the human mind

00:14:30.480 | to avoid hitting pedestrians, but you are suggesting that actually.

00:14:35.280 | There you go, yeah.

00:14:36.280 | You do.

00:14:37.280 | And then it's a lot harder, I thought.

00:14:40.960 | Yeah.

00:14:41.460 | [BLANK_AUDIO]

00:14:48.240 | [BLANK_AUDIO]

00:14:58.240 | [ Prevention is important ]

Daniel Kahneman: Deep Learning (System 1 and System 2) | AI Podcast Clips

Chapters