Presented at the 2018 International Conference on Social Computing, Behavioral-Cultural Modeling, & Prediction and Behavior Representation in Modeling and Simulation (SBP-BRiMS), Washington DC, USA (July 2018)

Prior knowledge influences perception, as evidenced by categorical perception phenomena, in which expectations create psychometric distortions of perceptual space. These distortions are nonetheless associated with categorization accuracy. The paradoxical association between strength of perceptual distortion (itself an inaccurate representation of reality) and accuracy of categorization judgments suggests that understanding the computational mechanism of categorical perception could lead to advances in machine learning and artificial intelligence. Here, a framework is presented that combines signal detection theory (SDT) and predictive processing. It instantiates the SDT expected value function in a Bayesian generative hierarchy, using the function’s parameters as a priori expectations about the perceptual environment. These priors then weight sensor response profiles. This approach links prediction error minimization to the optimality of perceptual judgment. The framework’s a posteriori predictions for incoming sensory signals model the distortions of perceptual space associated with categorical perception. The framework provides a computational mechanism by which SDT’s decision criterion is emergent from sensor tuning, rather than determined by a “decision” stage, after “perception.”