Unsupervised High Order Boltzmann Machine and Its Application on Medicine

Abstract

Based on current work about high order Boltzmann machine (BM) and unsupervised BM, an unsupervised learning algorithm based on high order BM is proposed. It is different from supervised BM in that it has no training samples for output units. In the unsupervised BM, the maximization of the mutual information based on Shannon entropy is used as an unsupervised criterion. As we all know, the computation cost of BM with hidden units is very expensive. When two restrictions are considered, that is the absence of hidden units and the restriction to classification problems, the high order BM can make up for the losing of hidden unit which can save lots of the computation cost. This domain of problems is very broad. The algorithm is the same with discrete variables and continuous variables. At last, the unsupervised high order BM is used to classify some medical data.