Generalization ability of fault tolerant feedforward neural nets

Abstract

Obtaining the maximum generalization and fault tolerance has been an important issue in the design of feedforward artificial neural networks (FFANNs). In previous work we introduced a method for ensuring the fault tolerance capabilities of FFANNs. We also introduced a detached model for fault tolerance, this model was shown to be realistic and appropriate for emulating faults that arise in FFANNs hardware implementation. In this paper we discuss the generalization ability of the fault tolerant FFANNs produced by our new training method. By introducing a method for measuring the generalization ability, this works shows that the network trained by our method has better generalization ability than that trained by conventional backpropagation technique.