Artificial neural networks and the parity problem : an empirical analysis

Abstract

Neural networks have been considered by many as a universal solution to a variety of learning problems. This conclusion was derived by the hype which rose with their introduction due to the fact that they are, allegedly, loosely based on how biological brains work. This idea is rejected by many. The Parity Problem highlights one of the deficiencies of Neural Networks. ln fact, it is considered to be a "hard to learn" problem. Parity in fact is also used as a benchmark to prove how fast a neural network learns. Further more, generalization properties of Neural Networks (with respect to parity) leave much to desire. This document outlines an analysis of the parity problem and neural networks with special concern towards the generalization properties of the latter. This study tries to measure the performance of several networks and determine whether correct generalization is ever possible (and if so under what conditions).