Figure 3.

Perceptual Inference Processes Underlying Tinnitus Initiation, Perpetuation, and Modulation. The inset box annotates perceptual inference in terms of the prior (prediction), likelihood (tinnitus precursor) and posterior (tinnitus percept) represented by Gaussian distributions over a perceptual dimension of intensity or loudness, with their widths indicating precision. Loudness may be encoded by neuronal firing rate or be represented more abstractly. In each plot, the perceived loudness of tinnitus is indicated by the position of the posterior distribution on the horizontal axis. (A) In hearing loss alone, the tinnitus precursor has insufficient precision to override the default prediction of silence. (B) With increased precision the precursor influences perception, leading to a revised posterior percept of tinnitus. Potentially there is a window of reversibility at this stage. (C) If the default prediction is revised to expect tinnitus (generally less intense than the precursor), then the condition becomes chronic (through experience-dependent plasticity). (D) Reduction of the precision of the precursor to its pre-tinnitus level results in habituation, but not cessation of tinnitus – on account of plastic changes to prior predictions. (E) Theoretically, patients with functional overlay may have a prediction of louder tinnitus than encoded by the precursor, and therefore tinnitus intensity would have no empirical bound. (F) Residual inhibition (RI) can be understood as attenuating the precision and/or intensity of the precursor through forward masking, thus reducing the precision-weighted prediction error (PWPE) and therefore gamma oscillations. An alternative mechanism is the temporary resetting of descending predictions to ‘silence’, increasing prediction error per se (hence gamma) but reducing the posterior percept. (G) In residual excitation (RE), temporary modification of the tinnitus prediction (increasing its loudness and/or reducing its precision) by a perceptually similar and precise stimulus leads to reduced prediction error (hence gamma), and increased tinnitus loudness more in line with that encoded by the precursor. (H) In functional overlay patients, acoustic forward masking, and the consequent fall in gamma oscillations, bias inference towards higher tinnitus intensity than encoded by the precursor (leading to RE).