Conference on Cognitive Computational Neuroscience

Predictive coding is a leading theoretical framework for understanding sensory processing as statistical inference. Its main tenet is that sensory input is analysed by prediction error units: neural processors that test our internal expectations on the sensory world against the sensory input. Prediction error units encode the difference between the expectations and the actual input, which is used in higher levels of the processing hierarchy to inform updates in our internal beliefs. The encoding of additive prediction error, signals that aim to add unaccounted elements to the internal representations, has been extensively studied before, both in sensory cortices and the subcortical sensory pathways. Whether sensory pathways also encode subtractive prediction error, signals that aim to remove representations of expected percepts that were absent in the sensory input, has not been unambiguously considered before. Here we used human fMRI to measure responses to omissions of sounds for which participants held varying levels of expectations. We used modelling and Bayesian model comparison to compute the posterior probability that neural populations in auditory midbrain and thalamus specifically encoded subtractive prediction error. The results provide first evidence of the encoding of subtractive prediction error in a subcortical sensory pathway, demonstrating that subcortical nuclei partake on the removal of wrong beliefs in high-level cognitive representations.