The Bayesian Brain and Tinnitus

Abstract

There is strong and growing evidence for ‘Bayesian’ accounts of brain function, such as predictive coding, in which perception depends upon internally generated models of the sensorium, which are updated and refined based on sensory input. Bayesian perception involves both automatic stimulus-based perceptual inference and active time-based inference that underwrites the active sampling of predicted sensory stimuli. Predictions in perceptual inference are associated with beta band activity, the precision of predictions with alpha activity and prediction errors with gamma oscillatory activity in sensory cortex, where prediction errors are propagated to other brain areas. Predictions in active inference are associated with theta activity, precision with delta and prediction errors with gamma activity in hippocampus and other cortical areas. Activity in these frequency bands, in auditory and other cortical areas, has been linked to tinnitus, but not yet specifically to Bayesian processes in this context. Different pathophysiological models have been proposed to explain the emergence of tinnitus, two of which are based on Bayesian processes. One account postulates that tinnitus is the result of memory-based prior beliefs about sounds, which are mobilised in compensation for the absence of evidence from sensory input. This model is based on active inference. Alternatively, spontaneous activity in the ascending auditory pathway (a ‘tinnitus precursor’) is usually ignored as noise, which conflicts with the default prediction of ʼno sound’, and chronic tinnitus occurs when this activity gains sufficient ‘precision’ to influence perceptual inference. Stress may influence perceptual inference in various ways, favouring the initial emergence of tinnitus and/or its perpetuation. In both accounts, incorporation of the tinnitus percept into default predictions leads to chronification of the condition. Anatomical and electrophysiological data in tinnitus research are compatible with a Bayesian brain model, but are largely circumstantial and also compatible with alternative formulations.