the best explanation of this are Ian Goodfellow's talks (i think there are video transcripts online somewhere) or at NIPS this past year. He and a few others a

the best explanation of this are Ian Goodfellow's talks (i think there are video transcripts online somewhere) or at NIPS this past year. He and a few others also have a new deep learning book just coming out this past month from MIT press