Commentary on Sanborn and Chater: Posterior Modes are Attractor Basins

Abstract

Sanborn and Chater [1] propose an interesting theory of cognitive and brain function based on Bayesian sampling instead of asymptotic Bayesian inference. Their proposal unifies many current observations and models and, in spite of focusing primarily on cognitive phenomena, their work provides a springboard for unifying several proposed theories of brain function. It has the potential to serve as a bridge between three influential overarching current theories of cognitive and brain function: Bayesian models, Fristons [2-4] theory of cortical responses based on the free-energy principle, and attractor-basin dynamics [5,6]. Specifically, their proposal suggests a high-level perspective on Fristons theory, which in turn proposes a sampling procedure including appropriate handling of autocorrelation as well as a plausible neurobiological implementation. In turn, these two theories together link into attractor-basin dynamics at the level of networks (via Friston) as well at the level of behavior (via the relationship between the modes of prior and posterior distributions, as discussed by Sanborn and Chater). We will argue here that, by linking Sanborn and Chaters approach to neurobiological models based on the free-energy principle on the one hand and attractor-basin dynamics on the other, the scope of their proposal can be broadened considerably. Moreover, a unified perspective along these lines provides an elegant solution to several of Sanborn and Chaters Outstanding Questions relating to the neural implementation of sampling.