Learning binaural sound localization through a neural network

Abstract

A neural network system is implemented that uses binaural time/intensity cues for determining azimuth/elevation of a sound source. The system is designed to approximately mimic the sound localization behavior of the owl. The network is trained in a supervised learning mode. The errors between the estimated position (from the neural net) and the actual position (from an ideal optical sensor) are used to determine adaptively the synaptic connections. The learning paradigm used is the multiple extended Kalman algorithm, which allows training with no parameter adjustments.<>