Learning Binaural Processing in Biological Networks

Abstract

It has always intrigued man as to how the human body performs so many complicated functions with the speed and accuracy that it does. One such task is that of sound localization in space, the ability to determine the location of a sound source with considerable accuracy. A biologically realistic neural network is proposed for the binaural processing of interaural time and intensity cues that closely resembles computational schemes suggested for sterepsis (depth perception) in vision. The important feature of this network is that it does not use any neural delay lines to generate such attributes of binaural hearing such as lateralization of all audible frequencies and the detection of enhancement of signals in a noisy environment. Temporal shifts between the signals at the ears, arising from sound sources at different locations on the azimuth cause spatial disparities in the corresponding travelling waves set up on the basilar membranes in the two ears. The two dimensional network proposed uses these spatial differences between instantaneous outputs at the two ears to measure interaural differences. The network operation approximately computes the cross-correlation between the two cochlear outputs by combining the ipsilateral input at a given characteristic frequency (CF) with contralateral inputs from locally off-CF locations. Some of the results obtained from this network are presented. Having proposed a network, the next question is whether such a connection is genetically present in the body or whether it is formed over a long period of time by a gradual process of learning. Assuming that the latter solution is more plausible, two learning rules are suggested according to which the network could alter its initial random connectivities. The first learning rule is a supervised technique in which a teaching signal prespecifies the ideal response expected from the network to each input pattern presented. The error between the actual output and the desired response helps to guide the learning process in the desired direction. When the minimum of the error surface is reached, the network is said to have learned and the weights do not change any more. The teaching signal required for the supervised algorithm could be derived from the visual system. However, no physiological evidence exists that links the auditory and visual maps at the level of the olivary complex which is where early binaural processing occurs. To overcome this problem, an unsupervised learning rule is proposed which requires only the cochlear outputs from the two ears. The rule is a competitive learning strategy wherein only one neuron updates its connectivities for a particular input pattern. The neuron chosen to alter its weights is the one which responds maximally to the input. The inherent delays that exist in the neural system are used as guides to form the organized spatial map responses.