Algorithms for cRBMs Training

About

Introduction: Restricted Boltzmann machines (RBMs) are probabilistic graphical models that can be used for classification, dimensionality reduction, topic modeling, and feature learning. They are superficial two-layer networks (a visible layer and a hidden layer) connected by a fully bipartite graph. RBMs summarize input data and automatically find patterns in data to build a probabilistic model that can be used to classify new data or reconstruct missing data. Continuous RBMs (cRBMs) are capable of effectively interfacing with fuzzy representations of uncertainty in data, such as image pixels or word-count vectors. This allows cRBMs to encode and reconstruct statistical samples from an unknown probability distribution. An improved method of training cRBMs for solving big data problems is highly needed. Technology: A Georgia State researcher has developed techniques for the training of cRBMs, aiming to estimate the continuous values of the hidden variables. This method uses least square error estimates for the hidden variables to provide improved results. This method first replaces discrete valued spins with continuous values, trains the cRBM model with a training dataset, and uses the model to recognize patterns in new data. This technique provides an improved method of training cRBMs and insights for solving big-data problems, which include, but are not limited to, classification, data cleaning, feature learning, and association rules learning.