Post #1727834

@ly2en don't know, I never studied NNs down to the mathy details. But that's how I was explained these things by people in the know. And it follows intuitively: since your output layer has to know "the answer", and the only thing it talks to is the n-1st layer, the latter has to have "almost the answer", i.e. a representation that's semantically closest to the answer. Not a proof, of course, just a helpful intuition.