Concurrent generative models inform prediction error in the human auditory pathway

Abstract: Predictive coding is the leading algorithmic framework to understand how expectations shape our experience of reality. Its main tenet is that sensory neurons encode prediction error: the residuals between a generative model of the sensory world and the actual sensory input. However, it is yet unclear how this scheme generalises to the multi-level hierarchical architecture of sensory processing. Theoretical accounts of predictive coding agree that neurons computing prediction error and the generative model exist at all levels of the processing hierarchy. However, there is not a current consensus of how predictions from independent models at different stages are integrated during the computation of prediction error. Here we investigated predictive processing with respect to two independent concurrent generative models in the auditory pathway using functional magnetic resonance imaging. We used two paradigms where human participants listened to sequences of either pure tones or FM-sweeps while we recorded BOLD responses in inferior colliculus (IC), medial geniculate body (MGB), and auditory cortex (AC). Each paradigm included the induction of two generative models: one based on local stimulus statistics; and another model based on the subjective expectations induced by task instruction. We used Bayesian model comparison to test whether neural responses in IC, MGB, and AC encoded prediction error with respect to either of the two generative models, or a combination of both. Results showed that neural populations in bilateral IC, MGB, and AC encode prediction error with respect to a combination of the two generative models, suggesting that the predictive architecture of predictive coding might be more complex than previously hypothesised.