Cue Combination Within a Bayesian Framework

To interact effectively with the world, the brain must optimize its perception of the objects and events in the environment, many of which are signaled by more than one sense. Optimal perception requires the brain to integrate redundant cues from the different senses as efficiently as possible. One effective model of cue combination is maximum likelihood estimation (MLE), a Bayesian model that deals with the fundamental uncertainty and noise associated with sensory signals and provides a statistically optimal way to integrate them. MLE achieves this through a weighted linear sum of two or more cues in which each cue is weighted inversely to its variance or “uncertainty.” This produces an integrated sensory estimate with minimal uncertainty and thus maximized perceptual precision. Many studies show that adults integrate redundant sensory information consistent with MLE predictions. When the MLE model is tested in school-aged children, it is found that predictions for multisensory integration are confirmed in older children (>10 years) but not in younger children. Younger children show unisensory dominance and do not exploit the statistical benefits of multisensory integration, even when their dominant sense is far less precise than the other. This curious finding may result from each sensory system having an inherent specialization, with each specialist sense tuning the other senses, such as vision calibrating audition for space (or audition calibrating vision for time). This cross-sensory tuning would preclude useful combination of two senses until calibration is complete, after which MLE integration provides an excellent model of multisensory cue combination.