Researchers Propose Predictive Coding as Alternative to Backpropagation in AI

Researchers at Sakana AI have published work exploring an alternative to backpropagation, the dominant training method used in neural networks. The proposed approach combines predictive coding with an Augmented Lagrangian Method (ALM) for optimizing neural network learning. Predictive coding is a biologically inspired framework that differs fundamentally from the standard backprop algorithm widely used in deep learning today. The work suggests this method could offer a more biologically plausible way to train neural networks. The research was shared on Hacker News, attracting early community attention.