François CholletVerified account

@fchollet doesn't that include DQN/A3C? ... seems like super-human performance may be a thing of the past.https://twitter.com/fchollet/status/1023722208508530688 …

No, it does not, and I'm sure you can tell from reading the tweets, which are pretty explicit about it. When you look at a game, in your mind, you're coming up with an abstract model of the goals, rules, physics, etc.

This allows you to come up with your own procedural rules for solving the game. A program. Your mind can serve as a program synthesis engine. Just like, say, a deep RL system, or a genetic algorithm, which in this context also serve as program synthesis engines.

Since both types of systems are program-synthesis engines, you should compare the performance of the programs themselves, not compare biological execution vs. digital execution.

Of course, and that's obvious, the game-solving program generated by any human mind when looking at a game is not the same as A3C/etc (although they are also engineered by humans).

In one case you have natural program synthesis -- you look at something and immediately, unconsciously come up with an abstract model of it and rules for solving it, and in the other case you have a deliberately engineered program synthesis algorithm.

This is like comparing biological vision with deep learning. Just because humans have created deep learning systems to recognize objects doesn't mean that these systems are in the same category as the natural object recognition capabilities of humans.