Predictive learning with local plasticity in excitatory-inhibitory networks

Predictive coding is a framework used to understand how the brain processes information, but it remains unclear how networks with local plasticity can perform predictive inference and learn representations. In this study, researchers demonstrate that a recurrent excitatory-inhibitory circuit with local plasticity can carry out predictive inference without the need for explicit error representations.

By establishing an analytic link, they show that learning in such circuits requires the weights to stay on a consistency manifold where the inhibition from recurrent connections matches the inhibition necessary for the predictive coding objective. Using a closed-form derivation of this condition, they derive a plasticity rule that maintains this consistency under a Gaussian prior.

When applied to a non-Gaussian prior, the rule enables the learning of sparse, factorized features like edge detectors from natural images. Empirically, a BCM-like rule with an activity-dependent threshold closely approximates this plasticity rule, while other Hebbian-like rules lead to less accurate solutions as they keep the weights further from the consistency manifold.

The networks with recurrent excitation acquire spatiotemporal features, such as direction selectivity, and the ability to predict and complete partially observed sequences. Time-continuous learning results in the formation of low-dimensional attractor-like structures and noise-driven replay. Overall, these findings establish a connection between predictive coding and local circuit plasticity, suggesting that BCM-like plasticity in excitatory synapses plays a normative role in the brain's information processing.