Attractor Neural Network Combined with Likelihood Maximization Algorithm for Boolean Factor Analysis

Abstract

AbstractWhen large data sets are analyzed, the pursuit of their appropriate representation in the space of lower dimension is a common practice. Boolean factor analysis can serve as a powerful tool to solve the task, when dealing with binary data. Here we provide a short insight into a new approach to Boolean factor analysis we have developed as an extension of our previously proposed method: Hopfield-like Attractor Neural Network with Increasing Activity. We have greatly enhanced its functionality, having complemented this method by maximizing the data set likelihood function. We have defined this Likelihood function on the basis of the data generative model proposed previously. As a result, in such a way we can obtain a full set of generative model parameters. We demonstrate the efficiency of the new method using the artificial signals, which are random mixtures of horizontal and vertical bars that are a benchmark for Boolean factor analysis. Then we show that the method can be used for real task solving when analyzing data from the Kyoto Encyclopedia of Genes and Genomes.KeywordsLyapunov FunctionFactor ScoreInformation GainLikelihood Maximization AlgorithmLikelihood Maximization ProcedureThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.