A multi-threshold neural network for frequency estimation

Abstract

Human perception of sound arises from the transmission of action-potentials (APs) through a neural network consisting of the auditory nerve and elements of the brain. Analysis of the response properties of individual neurons provides information regarding how features of sounds are coded in their firing patterns, and hints as to how higher brain centres may decode these neural response patterns to produce a perception of sound. Auditory neurons differ in the frequency of sound to which they respond most actively (their characteristic frequency), in their spontaneous (zero input) response, and also in their onset and saturation thresholds. Experiments have shown that neurons with low spontaneous rates show enhanced responses to the envelopes of complex sounds, while fibres with higher spontaneous rates respond to the temporal fine structure. In this paper, we determine an expression for the Cramer-Rao bound for frequency estimation of the envelope and fine structure of complex sounds by groups of neurons with parameterised response properties. The estimation variances are calculated for some typical estimation tasks, and demonstrate how, in the examples studied, a combination of low and high threshold fibres does not improve the estimation performance of a fictitious 'efficient' observer, but may improve the estimation performance of neural systems, such as biological neural networks, which are based on the detection of dominant interspike times.