Richard Sutton talk about continual learning.

Oct. 30, 2024

Youtube Link

So many points that I resonate on

I’m doing it in a particular way focusing on online learning and a single task uh just like the world is gradually changing and you have to keep following it and and and learn in that setting.

You’ll spend all your time convincing people and you’ll never actually work on it and so you’ll never actually succeed at it and everyone else will say oh look you know he says we need to do it but he’s not getting anywhere you know it’s a good way to lose is to spend your time trying to convince other people that they should do what you think is important.

Neural networks had their decade in the wilderness so no one thought it was good so I guess I’m sort of saying continual learning has had its decade in the wilderness and now it’s starting to be acceptable that’s good and there little it all come back all these problems with science are self-correcting.

It is both true there is only one goal and there are many goals.

You should try to be neutral on on what’s popular you know if it’s popular unpopular that shouldn’t influence you at all because if it’s popular well then it’ll be easier to work on it because people will understand it but it’ll be less valuable because everyone’s doing that so you should be neutral, pick your problem what makes sense to you as being important and and potentially fruitful.