RBM and DBN Learning notes

In 2006, Hinton's depth belief network (deep RBMs belief, Networks) based on the limited Boltzmann machine (re-stricted Boltzmann machines, DBNs) is the first in the field of deep learning theory in machine learning. One shot, and became the main framework of the deep learning algorithm since then. In this algorithm, DBN is cascaded by several layers of RBM, and thanks to the efficient approximation algorithm of the contrast divergence (contrastive divergence, CD), DBN bypasses the whole training problem of the multi-layer neural network and simplifies it to the training of multiple RBM, which So that the recognition effect and computational performance of multilayer neural network are greatly improved. The theoretical and practical experience also shows that DBN can better extract the hierarchical structural features of training data and provide a new idea for the selection of data features in machine learning. This paper mainly combs the algorithm principle and realization process of RBM and DBN, and gives an example program.

See HTTP://PAN.BAIDU.COM/S/1HSPSZFM for full notes. It is necessary to note that this note was not finished, because recently received a series of messages asking this note, only to think of this almost forgotten by me two years ago wrote the manuscript.

RBM and DBN Learning notes