Maximally fault-tolerant neural networks and nonlinear programming

A description is given of an application of neural network modeling to generate hypotheses about how response properties of neurons relate to information processing in the auditory system. Specifically, the kinds of response properties that are useful in extracting sound-localization information from directionally selective pinna filtering provided by the pinna are studied. For studying the sound localization based on spectral cues provided by the pinna, a neural network model with a guaranteed level of fault-tolerance is introduced. The notions of fault-tolerance in neural networks are formally defined, and a method of ensuring that the estimated network exhibits fault tolerance is described. The problem of estimating such weights is formulated as a large-scale constrained nonlinear programming problem. The preliminary numerical experiments indicate that (a) solutions with uniform fault tolerance in the hidden layer exist for this pattern recognition problem and that (b) using fault-tolerance as a constraint leads to solutions that have better generalization than solutions obtained via unconstrained backpropagation algorithm