On Interference of Signals and Generalization in Feedforward Neural Networks

This paper studies how the generalization ability of neurons can be affected by mutual processing of different signals. This study is done on the basis of a feedforward artificial neural network, that is used here as a model of the very basic processes in a network of biological neurons. The mutual processing of signals, called here an interference of signals, can possibly be a good model of patterns in a set generalized either by a biological network of neurons or by an artificial feedforward neural network, and in effect may improve generalization. In this paper it is discussed that the interference may however also cause highly random generalization. Adaptive activation functions in the studied model are discussed as a way of reducing that type of generalization. A test of a feedforward neural network is performed that shows the discussed random generalization. Hypotheses about ways of preventing the described random generalization in the biological neural networks are discussed. keywords: modeling biological neural network, feedforward neural networks, generalization, interference of signals, overfitting