free and open to the public
Bayesian Perception: A Critical Analysis
With some notable exceptions, Bayesian models in psychophysics and in perceptual psychology are formulated at Marr’s computational level of description. They describe how an optimal (or rational) observer would perform various perceptual tasks. For example, they describe how an optimal observer would integrate cues from multiple modalities by giving proper weight to each modality as a function of their reliability. As such, Bayesian models are used as “benchmarks”. When the observed performance is found to be close to optimal, the temptation is to presume that the models also describe the algorithm that the perceptual system uses to carry out a given task. This means, among other things, thinking that perceptual systems represent probability distributions, and that they use a unified Bayesian strategy across contexts. Some prominent philosophers have in fact taken Bayesian models to speak at the algorithmic level. In this talk, I problematize this take. I argue that Bayesian accounts are not well-suited to be algorithmic models of perception. Structural difficulties with the Bayesian picture together with evidence of non-optimal performance in virtually all areas of perceptual processing suggest a skeptical outlook.