12Predictive Coding — A Bayesian Model of Tinnitus

FThe Bayesian brain: perception as prediction

Modern neuroscience holds that the brain is a prediction machine. Rather than passively receiving sensation, it constantly generates a model of the world and compares it against incoming sensory data; only the mismatch — the prediction error — is passed up the hierarchy to update beliefs. Perception is the brain’s best Bayesian guess, weighting prior expectations against sensory evidence [2014].

This framework dissolves the old question “what makes the noise?” In a predictive brain, a percept can arise not from a stimulus but from a prediction that the sensory data fail to overturn. Tinnitus becomes a candidate for exactly this kind of perceptual inference error [2016].

TTinnitus as a percept that fills in a prediction

When the cochlea is damaged, the auditory system loses information about a band of frequencies. The brain, faced with missing or uncertain data, falls back on its prior expectation of what should be there — and that prediction is perceived as sound. On this view tinnitus is the brain “filling in” the deafferented channel with its best guess, much as the visual system fills in the blind spot [2014].

De Ridder and colleagues extend this into an integrative, subnetwork model: a unified phantom percept emerges from several partly separable networks (auditory, salience, distress, memory), each contributing a component, with the Bayesian inference binding them into a single experience [2014]. This is why no single brain region “is” the tinnitus.

TSedley’s sensory-precision account

William Sedley, Karl Friston and colleagues refined this into a sensory-precision model. In predictive coding, the influence of a prediction error depends on its precision — the brain’s estimate of how reliable that signal is. After deafferentation the brain mis-estimates the precision of spontaneous activity in the deprived channel, treating neural noise as a precise, trustworthy signal that must be explained — so it is perceived as a real sound [2016].

The model elegantly accounts for puzzling clinical facts: why tinnitus often matches the edge of hearing loss (the boundary of high uncertainty), why it can appear with a normal audiogram (precision can be misweighted without threshold loss), and why residual inhibition and sound therapy help (they transiently correct the precision estimate). Sedley later cautioned that simple “central gain” alone cannot explain all of this — precision-weighting is the missing ingredient [2019].

CA unifying account — and a memory dimension

The predictive-coding framework is powerful because it unifies the disparate findings of this chapter. Peripheral deafferentation supplies the missing data; central gain and DCN hyperactivity supply the spontaneous activity that gets mis-weighted; thalamocortical dysrhythmia reflects the abnormal prediction–error dynamics; and the limbic and salience networks set the precision and emotional weight of the inference. One principle — faulty Bayesian inference — threads them together [2014].

De Ridder also frames the chronic phantom as a persisting aversive memory: once the brain has learned the prediction, it becomes entrenched, drawing the tinnitus–pain parallel that the next module explores [2011]. The therapeutic message is optimistic: if tinnitus is a learned, mis-weighted prediction, it is in principle un-learnable — the rationale behind retraining, expectation-modifying counselling and precision-targeting neuromodulation [2016].