Prof. Anima Anandkumar

Wonderful talk @dileeplearning! Following up on the question of whether vision is generative or discriminative, it seems generative models allow better generalization to unexperienced situations, since it's not just a lookup table. Can one prove this in a formal way?https://twitter.com/dileeplearning/status/1264763015003426816 …

If you know the generative model and you are going inference on the fly, you can always get the right posterior given enough time. If you learn either the gen. model or the inference, then you start having to worry about generalisation (as you are learning stuff from past data).

I don't know the model well, but I surmised from @dileeplearning's talk that he is learning the gen model subject to constraints, and doing inference on the fly. The constraints (factoring contours from surfaces etc.) give him generalisation in learning the model.

Inference on the fly means he can deal with overlapping letters etc. It does not have to have seen this overlap before. It can figure out what latent causes are consistent with this data when it sees it for the first time.