Conference on Cognitive Computational Neuroscience

Predicting future events is an important function of cortical circuits, yet its precise computational mechanisms remain elusive. Although predictive coding (PC) provides a theoretical framework to describe predictive information processing in the cortex, classical PC models have neglected the temporal dimension, whereas previous efforts to incorporate time into PC models often compromised their biological plausibility and simplicity. In this work, we propose temporal predictive coding (tPC), which performs predictive processing over time while retaining the simple and plausible neural implementation of classical PC. We show that tPC approximates the optimal Kalman filtering in linear filtering problems, and can be straightforwardly extended to nonlinear models to perform more challenging cortical functions, including those related to dynamic visual processing and sequential memory.