The implementation of predictive coding across the auditory hierarchy

Posing the brain as a predictive system has become a guiding principle that has proven to be particularly fruitful in the study of perceptual processes.

Perception is understood as an inferential process in which the brain combines its noisy sensory input with predictions generated by internal models.

In this way, the contents of conscious experience would be determined by the hypothesis about the state of the world, with the highest posterior probability.

One popular algorithmic incarnation of the predictive processing ideas is predictive coding, in which the brain implements a hierarchy of generative models, where feedback connections carry predictions and feedforward connections carry prediction errors. The brain would minimize these prediction error by updating its internal models.

While compelling, this proposal leaves open several issues.It is not yet clear how a predictive perceptual system could be both sensitive to expected and to surprising stimuli.

Bayesian integration of sensory evidence and prior probability of the hypotheses (predictions) might offer an optimal solution. But some authors have argued that Bayesian integration would result in a bias toward the expected.

Therefore, the question of how conscious experience emanates from the interaction between predictions and prediction errors remains open.Predictive coding assumes separate neuronal populations representing predictions and prediction errors, which would reside in different cortical layers.

This proposal remains largely untested in humans, in particular in the auditory domain.This project focuses on the auditory domain, where the ability to predict the incoming sensory stream is fundamental for making sense of everyday auditory scenes, especially in the presence of noise.

I will tackle the aforementioned open questions by implementing challenging auditory signal detection tasks, and by combining fMRI data at ultra-high field (7 Tesla), and cutting-edge computational modelling.