Challenges to the Modularity Thesis Under the Bayesian Brain Models

Abstract

Modularity assumption is central to most theoretical and empirical approaches in cognitive science. The Bayesian Brain (BB) models are a class of neuro-computational models that aim to ground perception, cognition, and action under a single computational principle of prediction-error minimization. It is argued that the proposals of BB models contradict the modular nature of mind as the modularity assumption entails computational separation of individual modules. This review examines how BB models address the assumption of modularity. Empirical evidences of top-down influence on early sensory processes is often cited as a case against the modularity thesis. In the modularity thesis, such top-down effects are attributed to attentional modulation of the output of an early impenetrable stage of sensory processing. The attentional-mediation argument defends the modularity thesis. We analyse this argument using the novel conception of attention in the BB models. We attempt to reconcile classical bottom-up vs. top-down dichotomy of information processing, within the information passing scheme of the BB models. Theoretical considerations and empirical findings associated with BB models that address the modularity assumption is reviewed. Further, we examine the modularity of perceptual and motor systems.