Information-theoretic Criteria for the Integration of Auditory and Visual Spatial Information

We are presenting computational principles for the integration of auditory and visual spatial information, focusing on aspects regarding the formation, alignment and plasticity disparity between auditory and visual topographic maps. These principles are illustrated in the form of a four-layer network associated to two simple, purely local, unsupervised learning rules. Two “external”, independent topographic maps—adapted using the first learning rule— encode auditory and visual spatial information, while both “internal” layers—adapted using the second learning scheme—perform the desired alignment. We demonstrate that the combined application of both rules leads to the maximisation of the mutual information between spatial information conveyed by auditory and visual pathways. We illustrate the map alignment properties of our model through examples of simulations. One of these concerns retinal image shifts through prismatic goggles; another shows the effect of interaural distance modification (due to alterations in the morphology of the head during growth) on the long-term reorganisation of the auditory space map.