Probabilistic models of audiovisual perceptual decision making

Multisensory integration is a fundamental component of perceptual decision making and an excellent example of how the brain deals with the abundance of sensory uncertainty in order to create a coherent understanding of its environment. In this thesis I evaluate the two most popular computational models for describing multisensory integrative processes: maximum likelihood estimation (MLE) and Bayesian causal inference (BCI). Both models predict statistically-optimal sensory integration, but in so doing MLE makes the critical assumption that the brain always fuses sensory signals under certain experimental circumstances, whereas BCI allows for flexibility by assessing whether integration is appropriate for a particular set of stimuli. In two empirical studies on audiovisual spatial integration I expose considerable limitations of MLE in explaining human behavioral results and advocate the use of BCI to evaluate multisensory integration, even under conditions that were previously thought of as optimized for MLE. In a final empirical chapter I test an important prediction that both models make: that sensory uncertainty is reduced for integrated multisensory signals. I present behavioral evidence that confirms this prediction by showing that observers’ confidence levels increase as a result of audiovisual integration, thereby further validating the use of probabilistic models to describe multisensory perception