Bayesian inference in audiovisual spatial binding
* Presenting author
Abstract:
It is widely accepted that the human brain follows normative rules set by Bayesian inference to bind uncertain sensory signals in a statistically optimal fashion in order to form precise and coherent beliefs about its environment. When a sound and a flash appear synchronously and spatially near, they should be integrated by computing a reliability weighted average for their common source location. Since visual localisation is normally more reliable, this scheme naturally explains the well-known ventriloquist illusion. Nonetheless, recent work has suggested that humans do not integrate multisensory signals in accordance with the optimal observer model. While the model’s predictions approximately hold, too much weight is placed on signals from the normally more reliable sense. In a series of audiovisual localisation experiments we replicated this apparent suboptimal behavior by showing excessive biases towards the visual stimulus location. However, model comparison of a variety of proposed explanations demonstrated that these results are best described by a simple extension of the Bayesian framework that also accounts for uncertainty about the common origins of the sensory stimuli. These findings advocate the consideration of causal inference in perceptual binding experiments even when conditions are optimized for the unity assumption to hold.