Cortical networks with multiple interneuron types generate oscillatory patterns during predictive coding.

Predictive coding (PC) proposes that our brains work as an inference machine, generating an internal model of the world and minimizing predictions errors (i.e., differences between external sensory evidence and internal prediction signals). Theoretical models of PC often rely on high-level approaches, and therefore implementations detailing which neurons or pathways are used to compute prediction errors or adapt the internal representations, as well as their level of agreement with biological circuitry, are currently missing. Here we propose a computational model of PC, which integrates a neuroanatomically informed hierarchy of two cortical areas with a simplified laminar organization and cell-type-specific connectivity between excitatory, PV, SST and VIP cells. Our model efficiently performs PC, even in the presence of external and internal noise, by forming latent representations of naturalistic visual input (MNIST, fashion-MNIST and grayscale CIFAR-10) via Hebbian learning and using them to predict sensory input by minimizing prediction errors. The model assumes that both positive and negative prediction errors are computed by stereotypical excitatory-PV-SST-VIP circuits with the same structure but different incoming input. During sensory inference, neural oscillatory activity emerges in the system due to interactions between representation and prediction error microcircuits, with optogenetics-inspired inactivation protocols revealing a differentiated role of PV, SST and VIP cell types in such dynamics. Finally, our model shows anomalous responses to deviant stimuli within series of same-image presentations, in agreement with experimental results on mismatch negativity and oddball paradigms. We argue that our model constitutes an important step to better understand the circuits mediating PC in real cortical networks.