Abstract

Restricted Boltzmann Machine (RBM) is a generative stochastic energy-basedmodel of artificial neural network for unsupervised learning. Recently, RBM iswell known to be a pre-training method of Deep Learning. In addition to visibleand hidden neurons, the structure of RBM has a number of parameters such as theweights between neurons and the coefficients for them. Therefore, we may meetsome difficulties to determine an optimal network structure to analyze bigdata. In order to evade the problem, we investigated the variance of parametersto find an optimal structure during learning. For the reason, we should checkthe variance of parameters to cause the fluctuation for energy function in RBMmodel. In this paper, we propose the adaptive learning method of RBM that candiscover an optimal number of hidden neurons according to the trainingsituation by applying the neuron generation and annihilation algorithm. In thismethod, a new hidden neuron is generated if the energy function is not stillconverged and the variance of the parameters is large. Moreover, theinactivated hidden neuron will be annihilated if the neuron does not affect thelearning situation. The experimental results for some benchmark data sets werediscussed in this paper.

 

Introduction (beta)

None

 

Conclusion (beta)

None