How to Train With Less Labels

Kip: You were mentioning the idea that the generative models versus these discriminant models, you can train with fewer labels. When you say that, did you mean literally fewer examples or less information? There's so much information in a reconstruction loss because you have all the values, all pixels. Whereas if you have a single label, you just got our histone, yeah, there's only one bit in there. The number of bits happens in the industry high. It was just really interesting because what does that mean though? Clearly, my brain is not encoding pixels. It's encoding something else. And then I'm expressing some higher level representation.