Analysis and synthesis of neural networks using linear separation

Abstract

General analysis and synthesis methods for neural networks are presented. The techniques proposed are simple, efficient and not restricted to a certain network architecture, i.e., they can be any of the multilayer, fully interconnected feedforward or feedback structures. Based on the signs of connections between neurons (called weight signatures) being excitatory or inhibitory, the methods proposed provide some fundamental rules of learnability in such networks. Various design techniques are presented using these learning rules for the synthesis of neural architectures.