Inferred representations behave like oscillators in dynamic Bayesian models of beat perception

Abstract

The human's capacity to perceptually entrain to an auditory rhythm has been repeatedly modeled as a dynamical system consisting of one or more forced oscillators. However, a more recent perspective, closely related to the popular theory of Predictive Processing, treats auditory entrainment as an inference process in which the observer infers the phase, tempo, and/or metrical structure of an auditory stimulus based on event timing. Here, we propose a close relationship between these two perspectives. We show for the first time that a system performing variational Bayesian inference about the circular phase underlying a rhythmic stimulus takes the form of a forced, damped oscillator with a specific nonlinear phase response function corresponding to the internal metrical model of the underlying rhythm. This algorithm can be extended to simultaneous inference on both phase and tempo using one of two possible approximations that closely align with the two most prominent models of auditory entrainment: one yields a single oscillator with an adapting period, and the other yields a networked bank of oscillators. We conclude that an inference perspective on rhythm perception can offer similar descriptive power and flexibility to a dynamical systems perspective while also plugging into the fertile unifying framework of Bayesian Predictive Processing.