Sakana AI’s Error Diffusion Trains Dale-Compliant Dual-Stream Networks, Reaching 96.7% MNIST and 61.7% CIFAR-10 Without Backpropagation

Sakana AI developed Error Diffusion, a training method that adheres to Dale's principle and avoids backpropagation's weight transport requirement, achieving 96.7% accuracy on MNIST and 61.7% on CIFAR-10 using dual-stream excitatory/inhibitory networks. The approach scales to reinforcement learning and reveals task-dependent ablation effects, offering a biologically plausible alternative for neural network training.