Pedro Domingos

Funny.

(For the record, there will definitely be sudden improvements. We’re already seeing them.)

Sorry. I haven't seen any breakthrough from the deep learning research community that will lead to AGI. It's not possible, IMO.

Well, you know the saying. As soon as something works, it’s no longer AI :)

Well, I haven't seen any real AI. The biggest problem in AI is generalization, the ability to instantly perceive objects even if seeing them for the first time. DL is incapable of generalizing because function optimization via gradient descent is the opposite of generalization.

Lots of work happening in zero-shot recognition. And gradient descent is indeed generalization when run over a network that cannot overfit to the entire training set. Intelligence itself is lossy compression. Trick of gradient descent is it moves slow enough to linearize rels.

No work in zero-shot recognition using gradients will lead anywhere, imo. This is not how the brain does it. The brain doesn't even process data. It processes the timing of changes/transitions.

That’s not meaningful. Of course the brain processes data. And of course by the time that data reaches the brain it’s been GAN’d up 83 ways from Sunday. Ever notice how our senses are blind to steady state input, though? That’s because there’s no gradient.