Predictive routing emerges from self-supervised stochastic neural plasticity.

Neurophysiology studies propose that predictive coding is implemented via alpha/beta (8-30 Hz) rhythms that prepare specific pathways to process predicted inputs. This leads to a state of relative inhibition, reducing feedforward gamma (40-90 Hz) rhythms and spiking to predictable inputs. We refer to this model as predictive routing. It is currently unclear which circuit mechanisms implement this push-pull interaction between alpha/beta and gamma rhythms. To explore how predictive routing is implemented, we developed a self-supervised learning algorithm we call generalized Stochastic Delta Rule (gSDR). It was necessary to develop this learning rule because manual tuning of parameters (frequently used in computational modeling) is inefficient to search through a non-linear parameter space that defines how neuronal rhythms emerge and interact. We used gSDR to train biophysical neural circuits and validated the algorithm on simple tasks, e.g., tuning membrane potentials and firing rates. We next applied gSDR to model observed neurophysiology. We asked the model to reproduce a shift from baseline oscillatory dynamics (~<20Hz) to stimulus induced gamma (~40-90Hz) dynamics recorded in the macaque monkey visual cortex. This gamma-band oscillation during stimulation emerged by self-modulation of synaptic weights via gSDR. We further showed that the gamma-beta push-pull interactions implied by predictive routing could emerge via stochastic modulation of both local inhibitory circuitry as well as top-down modulatory inputs to a network. To summarize, we implemented gSDR to train biophysical neural circuits based on a series of objectives. gSDR succeeded in implementing these objectives. This revealed the inhibitory neuron mechanisms underlying the gamma-beta push-pull dynamics that are observed during predictive processing tasks in systems and cognitive neuroscience.