Sakana AI Researchers Introduce PC-ALM, a Layer-Local Alternative to Backpropagation That Trains 1000-Layer Networks

What changed

Sakana AI researchers introduced PC-ALM, a training method designed to replace the usual process for sending learning signals backward through a network. In tests, it closely matched that method across networks of different sizes and trained a 1,000-layer network to within about two percentage points on MNIST; the code is available under an MIT license.