Conference on Cognitive Computational Neuroscience

Behavioral evidence suggests that the brain is able to infer the posterior probability of the state of its environment for a given sensory input. However, the exact neural implementation of this inference remains unclear. Predictive coding (PC) is an influential biologically plausible theory that hypothesizes that the brain performs inference and learning by minimising the error between its sensory inputs and the sensory inputs predicted by an internal generative model of its environment. However, PC can only infer the most likely state of the environment for a given sensory input. We propose Monte Carlo predictive coding (MCPC) and show using both theoretical and experimental findings that MCPC (1) enables the inference of posterior probabilities using Monte Carlo sampling, (2) enables the autonomous sampling of the learned generative model and (3) improves the generative learning performance compared to PC by removing a failure mode of PC causing weights to diverge. Moreover, MCPC can be implemented with a biologically plausible neural network, using noisy neural dynamics for sampling-based probabilistic inference, supporting the functional importance of neural variability in the brain.