The brain is the most powerful information processor known. Part of its power derives from its ability to combine (or integrate) information from multiple sensory systems. For example, we can combine sight, sound, and touch in forming an integrated perception of events in our environment. Combining input from multiple sensory systems is known as multisensory integration. Understanding how multisensory integration actually works in the brain will provide important insights into the nature of perception. In studying perception it is important to realize that, no matter how good sensory systems may be, they can't provide perfect information. The information provided by sensory systems must be considered as uncertain to some extent. Perception may involve the use of sensory inputs to provide evidence of events in the environment. The superior colliculus is a brain structure that causes mammalian animals (like us) to turn our heads and eyes in the direction of new events in the environment. Many neurons in the superior colliculus receive inputs from more than one sensory system. These multisensory neurons exhibit a property know as multisensory enhancement, in which the response to an input from one sensory system can be greatly increased by input from another sensory system. We have developed an hypothesis that multisensory enhancement is the result of processing by which collicular neurons use multisensory input to compute the probability that an event has occurred in the environment. The goal of our project is to develop a model that explains how neurons might actually perform this computation, and then use actual data from collicular neurons to test the model. The computational model we propose will be adaptive, that is, capable of changing its own behavior on the basis of its experience with the environment. It will be composed of two stages that are meant to represent two separate stages in the development of multisensory enhancement in the brain. Collicular neurons are known to receive multisensory inputs from both lower and higher levels in the brain. In the first stage, model collicular neurons will learn to extract the maximum amount of information from lower-level multisensory inputs. In the second stage, model collicular neurons with use higher-level multisensory inputs to refine their computation of the probability of events in the environment. We will test this model by comparing its behavior with that of actual collicular neurons studied in cats. Our model predicts that the amount of multisensory enhancement observed for collicular neurons should depend upon other properties such as the location of those neurons in the colliculus. Our model should also make predictions concerning how the behavior of collicular neurons should change when the higher-level inputs are removed. Proposing a detailed model of multisensory enhancement and testing the model against actual data should provide us with some new insights into how multisensory enhancement may be organized in the brain. A better understanding of this more basic form of multisensory integration might open the door for a better understanding of perception in general.
