Closed-form predictive coding via hierarchical Gaussian filters
Researchers have developed a new method for predictive coding networks that addresses their historical limitations in speed and performance with increasing depth. By treating these networks as deep hierarchical Gaussian filters and incorporating precision-weighted message passing, the new approach allows for dynamic uncertainty estimates and Hebbian-compatible updates. This closed-form variational inference method enables networks to learn activations, weights, and precisions simultaneously without iterative relaxation or global error signals, achieving performance comparable to backpropagation on benchmark tasks. AI
IMPACT This new predictive coding method offers a biologically grounded alternative to backpropagation, potentially improving efficiency and performance in deep learning models.