Units and Layers' Effects on Deep Boltzman Machines

Abstract

This paper analyzes the units' and layers' effects on deep Boltzman machines. It divides the DBM into two parts and reveals how the two parts affect the DBM's approximation capability. It indicates that the representation power of deep Boltzman machine is not always improved with more units and layers. When a deep Boltzman machine is best already, more units and layers will nearly always lead to worse performance.