This is a clip from a conversation with Francois Chollet from Sep 2019. New full episodes every Mon & Thu and 1-2 new clips or a new non-podcast video on all other days. You can watch the full conversation here: https://www.youtube.com/watch?v=Bo8MY4JpiXE
(more links below)

Podcast full episodes playlist:
https://www.youtube.com/playlist?list=PLrAXtmErZgOdP_8GztsuKi9nrraNbKKp4

Podcasts clips playlist:
https://www.youtube.com/playlist?list=PLrAXtmErZgOeciFP3CBCIEElOJeitOr41

Podcast website:
https://lexfridman.com/ai

Podcast on iTunes:
https://apple.co/2lwqZIr

Podcast on Spotify:
https://spoti.fi/2nEwCF8

Podcast RSS:
https://lexfridman.com/category/ai/feed/

Note: I select clips with insights from these much longer conversation with the hope of helping make these ideas more accessible and discoverable. Ultimately, this podcast is a small side hobby for me with the goal of sharing and discussing ideas. For now, I post a few clips every Tue & Fri. I did a poll and 92% of people either liked or loved the posting of daily clips, 2% were indifferent, and 6% hated it, some suggesting that I post them on a separate YouTube channel. I hear the 6% and partially agree, so am torn about the whole thing. I tried creating a separate clips channel but the YouTube algorithm makes it very difficult for that channel to grow unless the main channel is already very popular. So for a little while, I'll keep posting clips on the main channel. I ask for your patience and to see these clips as supporting the dissemination of knowledge contained in nuanced discussion. If you enjoy it, consider subscribing, sharing, and commenting.

François Chollet is the creator of Keras, which is an open source deep learning library that is designed to enable fast, user-friendly experimentation with deep neural networks. It serves as an interface to several deep learning libraries, most popular of which is TensorFlow, and it was integrated into TensorFlow main codebase a while back. Aside from creating an exceptionally useful and popular library, François is also a world-class AI researcher and software engineer at Google, and is definitely an outspoken, if not controversial, personality in the AI world, especially in the realm of ideas around the future of artificial intelligence.

Subscribe to this YouTube channel or connect on:
- Twitter: https://twitter.com/lexfridman
- LinkedIn: https://www.linkedin.com/in/lexfridman
- Facebook: https://www.facebook.com/lexfridman
- Instagram: https://www.instagram.com/lexfridman
- Medium: https://medium.com/@lexfridman
- Support on Patreon: https://www.patreon.com/lexfridman

What do you think are the current limits of deep learning?

If we look specifically at these function approximators that tries to generalize from data.

You've talked about local versus extreme generalization.

You mentioned that neural networks don't generalize well, humans do.

So, and you've also mentioned that generalization, extreme generalization requires something like reasoning to fill those gaps.

So how can we start trying to build systems like that?

Right, yeah, so this is by design, right?

Deep learning models are like huge parametric models, differentiable, so continuous, that go from an input space to an output space.

And they're trained with gradient descent.

So they're trained pretty much point by point.

They are learning a continuous geometric morphing from an input vector space to an output vector space.

And because this is done point by point, a deep neural network can only make sense of points in experience space that are very close to things that it has already seen in string data, at best, it can do interpolation across points.

But that means, you know, it means in order to train your network, you need a dense sampling of the input cross output space, almost a point by point sampling, which can be very expensive if you're dealing with complex real world problems like autonomous driving, for instance, or robotics.

It's doable if you're looking at the subset of the visual space, but even then it's still fairly expensive.

And it's only going to be able to make sense of things that are very close to ways that's seen before.

And in contrast to that, well, of course you have human intelligence, but even if you're not looking at human intelligence, You can look at very simple rules, algorithms.

If you have a symbolic rule, it can actually apply to a very, very large set of inputs because it is abstract.

It is not obtained by doing a point by point mapping, right?

For instance, if you try to learn a sorting algorithm using a deep neural network, well, you're very much limited to learning point by point what the sorted representation of this specific list is like, but instead you could have a very, very simple sorting algorithm written in a few lines.

Maybe it's just, you know, 2 nested loops.

And it can process any list at all because it is abstract, because it is a set of rules.

So deep learning is really like point by point geometric morphings, morphing strain, risk, and descent.

And meanwhile, abstract rules can generalize much better.

And I think the future is really to combine the 2.

So how do we, do you think, combine the 2?

How do we combine good point-by-point functions with programs, which is what the symbolic AI type systems?

I mean, obviously we're jumping into the realm of where there's no good answers.

It's just kind of ideas and intuitions and so on.

Well, if you look at the really successful AI systems today, I think they are already hybrid systems that are combining symbolic AI with deep learning.

For instance, successful robotics systems are already mostly model-based, rule-based, things like planning algorithms and so on.

At the same time, they're using deep learning as perception modules.

Sometimes they're using deep learning as a way to inject fuzzy intuition into a rule-based process.

If you look at a system like a self-driving car, it's not just 1 big end to end neural network, you know, that wouldn't work at all precisely because in order to train that you would need a dense sampling of experience space when it comes to driving, which is completely unrealistic, obviously.

Instead, the self-driving car is mostly symbolic, you know, it's software, it's programmed by hand.

So it's mostly based on explicit models, in this case, mostly 3D models of the environment around the car, but it's interfacing with the real world using deep learning modules, right?

So the deep learning there serves as a way to convert the raw sensory information to something usable by symbolic systems.

Okay, well, let's linger on that a little more.

Self-driving for many people, let's not even talk about self-driving, let's talk about steering.

Lane following, yeah, it's definitely a problem you can solve with an end-to-end deep learning model, but that's like 1 small subset.

I don't know why you're jumping from the extreme so easily, because I disagree with you on that.

I think, well, it's not obvious to me that you can solve lane following.

I think in general, you know, there is no hard limitations to what you can learn with a deep neural network.

As long as the search space, like, is rich enough, is flexible enough.

And as long as you have this dense sampling of the input cross output space, the problem is that, you know, this dense sampling could mean anything from 10,000 examples to like trillions and trillions.

And if you could just give it a chance and think what kind of problems can be solved by getting a huge amounts of data and thereby creating a dense mapping.

So let's think about natural language dialogue, the Turing test.

Do you think the Turing test can be solved with a neural network

Well, The Turing test is all about tricking people into believing they're talking to a human.

And I don't think that's actually very difficult because it's more about exploiting a human perception and not so much about intelligence.

There's a big difference between mimicking intelligent behavior and actual intelligent behavior.

So, okay, let's look at maybe the Alexa prize and so on.

The different formulations of the natural language conversation that are less about mimicking and more about maintaining a fun conversation that lasts for 20 minutes.

And that's more about, I mean, it's still mimicking, but it's more about being able to carry forward a conversation with all the tangents that happen in dialogue and so on.

Do you think that problem is learnable with this kind of, with a neural network that does the point to point mapping?

So I think it would be very, very challenging to do this with deep learning.

I don't think it's out of the question either.

The space of problems that can be solved with a large neural network.

What's your sense about the space of those problems?

In practice, Well, deep learning is a great fit for perception problems.

In general, any problem which is not really amenable to explicit and crafted rules, or rules that you can generate by exhaustive search over some program space.

So perception, artificial intuition, as long as you have a sufficient training dataset.

And that's the question, I mean, perception, there's interpretation and understanding of the scene, which seems to be outside the reach of current perception systems.

So do you think larger networks will be able to start to understand the physics and the physics of the scene, the three-dimensional structure and relationships of objects in the scene and so on, or really that's where symbolic AI has to step in?

Well, it's always possible to solve these problems with deep learning, it's just extremely inefficient.

A model would be an explicit rule-based abstract model would be a far better, more compressed representation of physics than learning just this mapping between in this situation, this thing happens.

If you change the situation slightly, then this other thing happens and so on.

Do you think it's possible to automatically generate the programs that would require that kind of reasoning?

Or does it have to, so the way the expert systems fail, there's so many facts about the world had to be hand-coded in.

Do you think it's possible to learn those logical statements that are true about the world and their relationships?

Do you think, I mean, that's kind of what theorem proving at a basic level is trying to do, right?

Yeah, except it's much harder to formulate statements about the world compared to formulating mathematical statements.

Statements about the world, you know, tend to be subjective.

So can you, can you learn rule-based models?

That's the, That's the field of program synthesis.

However, today we just don't really know how to do it.

So it's very much a graph search or tree search problem.


Speaker 0: What do you think are the current limits of deep learning? If we look specifically at these function approximators that tries to generalize from data. You've talked about local versus extreme generalization. You mentioned that neural networks don't generalize well, humans do. So there's this gap.

So, and you've also mentioned that generalization, extreme generalization requires something like reasoning to fill those gaps. So how can we start trying to build systems like that?

Speaker 1: Right, yeah, so this is by design, right? Deep learning models are like huge parametric models, differentiable, so continuous, that go from an input space to an output space. And they're trained with gradient descent. So they're trained pretty much point by point. They are learning a continuous geometric morphing from an input vector space to an output vector space.

All right. And because this is done point by point, a deep neural network can only make sense of points in experience space that are very close to things that it has already seen in string data, at best, it can do interpolation across points. But that means, you know, it means in order to train your network, you need a dense sampling of the input cross output space, almost a point by point sampling, which can be very expensive if you're dealing with complex real world problems like autonomous driving, for instance, or robotics. It's doable if you're looking at the subset of the visual space, but even then it's still fairly expensive. You're still in millions of examples.

And it's only going to be able to make sense of things that are very close to ways that's seen before. And in contrast to that, well, of course you have human intelligence, but even if you're not looking at human intelligence, You can look at very simple rules, algorithms. If you have a symbolic rule, it can actually apply to a very, very large set of inputs because it is abstract. It is not obtained by doing a point by point mapping, right? For instance, if you try to learn a sorting algorithm using a deep neural network, well, you're very much limited to learning point by point what the sorted representation of this specific list is like, but instead you could have a very, very simple sorting algorithm written in a few lines.

Maybe it's just, you know, 2 nested loops. And it can process any list at all because it is abstract, because it is a set of rules. So deep learning is really like point by point geometric morphings, morphing strain, risk, and descent. And meanwhile, abstract rules can generalize much better. And I think the future is really to combine the 2.

Speaker 0: So how do we, do you think, combine the 2? How do we combine good point-by-point functions with programs, which is what the symbolic AI type systems? Yeah. At which levels the combination happen? I mean, obviously we're jumping into the realm of where there's no good answers.

It's just kind of ideas and intuitions and so on.

Speaker 1: Well, if you look at the really successful AI systems today, I think they are already hybrid systems that are combining symbolic AI with deep learning. For instance, successful robotics systems are already mostly model-based, rule-based, things like planning algorithms and so on. At the same time, they're using deep learning as perception modules. Sometimes they're using deep learning as a way to inject fuzzy intuition into a rule-based process. If you look at a system like a self-driving car, it's not just 1 big end to end neural network, you know, that wouldn't work at all precisely because in order to train that you would need a dense sampling of experience space when it comes to driving, which is completely unrealistic, obviously.

Instead, the self-driving car is mostly symbolic, you know, it's software, it's programmed by hand. So it's mostly based on explicit models, in this case, mostly 3D models of the environment around the car, but it's interfacing with the real world using deep learning modules, right?

Speaker 0: So the deep learning there serves as a way to convert the raw sensory information to something usable by symbolic systems. Okay, well, let's linger on that a little more. So dense sampling from input to output. You said it's obviously very difficult. Is it possible?

Speaker 1: In the case of self-driving, you mean?

Speaker 0: Let's say self-driving, right? Self-driving for many people, let's not even talk about self-driving, let's talk about steering. So staying inside the lane.

Speaker 1: Lane following, yeah, it's definitely a problem you can solve with an end-to-end deep learning model, but that's like 1 small subset.

Speaker 0: Hold on a second. I don't know why you're jumping from the extreme so easily, because I disagree with you on that. I think, well, it's not obvious to me that you can solve lane following.

Speaker 1: No, it's not. It's not obvious. I think it's doable. I think in general, you know, there is no hard limitations to what you can learn with a deep neural network. As long as the search space, like, is rich enough, is flexible enough.

And as long as you have this dense sampling of the input cross output space, the problem is that, you know, this dense sampling could mean anything from 10,000 examples to like trillions and trillions.

Speaker 0: So that's my question. So what's your intuition? And if you could just give it a chance and think what kind of problems can be solved by getting a huge amounts of data and thereby creating a dense mapping. So let's think about natural language dialogue, the Turing test. Do you think the Turing test can be solved with a neural network

Speaker 1: alone? Well, The Turing test is all about tricking people into believing they're talking to a human. And I don't think that's actually very difficult because it's more about exploiting a human perception and not so much about intelligence. There's a big difference between mimicking intelligent behavior and actual intelligent behavior.

Speaker 0: So, okay, let's look at maybe the Alexa prize and so on. The different formulations of the natural language conversation that are less about mimicking and more about maintaining a fun conversation that lasts for 20 minutes. That's a little less about mimicking. And that's more about, I mean, it's still mimicking, but it's more about being able to carry forward a conversation with all the tangents that happen in dialogue and so on. Do you think that problem is learnable with this kind of, with a neural network that does the point to point mapping?

Speaker 1: So I think it would be very, very challenging to do this with deep learning. I don't think it's out of the question either. I wouldn't rule it out.

Speaker 0: The space of problems that can be solved with a large neural network. What's your sense about the space of those problems? So useful problems for us.

Speaker 1: In theory, it's infinite, right? You can solve any problem. In practice, Well, deep learning is a great fit for perception problems. In general, any problem which is not really amenable to explicit and crafted rules, or rules that you can generate by exhaustive search over some program space. So perception, artificial intuition, as long as you have a sufficient training dataset.

Speaker 0: And that's the question, I mean, perception, there's interpretation and understanding of the scene, which seems to be outside the reach of current perception systems. So do you think larger networks will be able to start to understand the physics and the physics of the scene, the three-dimensional structure and relationships of objects in the scene and so on, or really that's where symbolic AI has to step in?

Speaker 1: Well, it's always possible to solve these problems with deep learning, it's just extremely inefficient. A model would be an explicit rule-based abstract model would be a far better, more compressed representation of physics than learning just this mapping between in this situation, this thing happens. If you change the situation slightly, then this other thing happens and so on.

Speaker 0: Do you think it's possible to automatically generate the programs that would require that kind of reasoning? Or does it have to, so the way the expert systems fail, there's so many facts about the world had to be hand-coded in. Do you think it's possible to learn those logical statements that are true about the world and their relationships? Do you think, I mean, that's kind of what theorem proving at a basic level is trying to do, right?

Speaker 1: Yeah, except it's much harder to formulate statements about the world compared to formulating mathematical statements. Statements about the world, you know, tend to be subjective. So can you, can you learn rule-based models? Yes. Yes, definitely.

That's the, That's the field of program synthesis. However, today we just don't really know how to do it. So it's very much a graph search or tree search problem. And so we are limited to,

See all Lex Fridman transcripts on Youtube

François Chollet: Limits of Deep Learning | AI Podcast Clips