Learning by a population of perceptrons

Abstract

Learning by examples of a population of neural networks is studied in a statistical physics framework. A population of single-layer perceptrons learns from a two-layer neural network. Each member is trained independently either from the same or from different example sets. The outputs of multiple networks are combined by majority vote. We calculate the generalization curve of the group decision of the perceptrons with both discrete and continuous weights. We find an interesting nonmonotonic learning curve for the case of discrete weights, indicating that majority vote shows optimal performance when the size of the example set is finite.