AbstractLearning of complex auditory sequences such as language and music can be thought of as the continuous optimisation of internal predictive representations of sound-pattern regularities, driven by prediction errors. In predictive coding (PC), this occurs through changes in the intrinsic and extrinsic connectivity of the relevant cortical networks, whereby minimization of precision-weighted prediction error signals improves the accuracy of future predictions. Here, we employed Dynamic Causal Modelling (DCM) on functional magnetic resonance (fMRI) data acquired during the presentation of complex auditory patterns. In an oddball paradigm, we presented 52 volunteers (non-musicians) with isochronous 5-tone melodic patterns (standards), randomly interleaved with rare novel patterns comprising contour or pitch interval changes (deviants). Here, listeners must update their standard melodic models whenever they encounter unexpected deviant stimuli. Contour deviants induced an increased BOLD response, as compared to standards, in primary (Heschl’s gyrus, HG) and secondary auditory cortices (planum temporale, PT). Within this network, we found a left-lateralized increase in feedforward connectivity from HG to PT for deviant responses and a concomitant disinhibition within left HG. Consistent with PC, our results suggest that model updating in auditory pattern perception and learning is associated with specific changes in the excitatory feedforward connections encoding prediction errors and in the intrinsic connections that encode the precision of these errors and modulate their gain.Significance statementThe learning of complex auditory stimuli such as music and speech can be thought of as the continuous optimisation of brain predictive models driven by prediction errors. Using dynamic causal modelling on fMRI data acquired during a melodic oddball paradigm, we here show that brain responses to unexpected sounds were best explained by an increase in excitation within Heschl’s gyrus and an increase in forward connections from Heschl’s gyrus to planum temporale. Our results are consistent with a predictive coding account of sensory learning, whereby prediction error responses to new sounds drive model adjustments via feedforward connections and intrinsic connections encode the confidence of these prediction errors.
Hierarchical predictive coding in frontotemporal networks with pacemaker expectancies: evidence from dynamic causal modelling of Magnetoencephalography (MEG)
