Training products of experts by minimizing contrastive divergence (hereinafter referred to as Poe) is the beginning of dbN and deep learning theories. Recently, when I am studying RBM-related knowledge, the comparison divergence algorithm used for training RBM is not very detailed in all kinds of Chinese and English documents. Some of them are just a bit backward. Why can we use the comparison divergence algorithm to approximate the original target function, why can better convergence be achieved through one-step iteration? Such problems have always plagued us. As a result, I finally turned over this article. After reading it carefully, I found that many basic ideas and starting points about algorithms have been mentioned by Hinton, but we didn't care much about it, here we will sort out the notes after this paper is re-read for reference.
Reading Notes-training products of experts by minimizing contrastive divergence