Neural network models of sound localization based on directional filtering by the pinna

Three-layer neural-network functions were developed to transform spectral representations of pinna-filtered stimuli at the input to a space-mapped representation of sound-source direction at the output. The inputs are modeled after transfer functions of the external ear of the cat; the output is modeled on the spatial sensitivity of superior colliculus neurons. Network solutions are obtained by backpropagation and by a method that enforces uniform task distribution in the hidden layer of the model. Solutions are characterized using bandlimited inputs to study the relative strength of potential sound localization cues in various frequency regions. This analysis suggests that the frequency region containing the first spectral notch (5-18 kHz) provides the best localization cues. Response properties of model neurons were studied using input patterns modeled after auditory nerve response profiles to pure tones at various frequencies and sound levels. The response properties of hidden layer model neurons resemble cochlear nucleus types III and IV and their composites. Neurons in both hidden and output layers show the properties of spectral notch detectors. Although neural networks have limitations as models of real neural systems, the results illustrate how they can provide insight into the computation of complex transformations in the nervous system.