Training a Predictive Coding Network on ImageNet using Equilibrium Propagation
Researchers have developed a new method to train predictive coding networks (PCNs) using Equilibrium Propagation (EP), a physics-based framework. This novel approach successfully scaled EP and PCNs to train a 10-layer convolutional network on the full ImageNet dataset. The trained network achieved a top-5 classification error rate of 13.23%, closely matching the 12.2% error rate of traditional backpropagation methods. AI
IMPACT Demonstrates a scalable training method for energy-based models, potentially opening new avenues for large-scale AI research.