How Much Intelligence Can You Get from Relaxation?

Learning in the brain, and in energy-efficient physical hardware, has to rely on local signals and continuous-time dynamics. I will show how networks that compute by relaxing on an energy landscape, from Hopfield-type networks to predictive coding, can be trained with purely local rules to shape the landscape itself. The resulting backpropagation-free generative models nearly match backpropagation on CIFAR-10 (3.64 vs 3.54 FID), to my knowledge the best quality yet reached without backpropagation. I will then outline how the same principle could let a system learn continually from a single stream of experience, such as a child’s headcam video.