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

Sakana AI's Error Diffusion method trains dual-stream excitatory/inhibitory networks that obey Dale's principle without backpropagation, achieving 96.7% accuracy on MNIST and 61.7% on CIFAR-10. The approach sidesteps the weight transport problem and scales to reinforcement learning, with task-dependent ablations revealing how modulo error routing works.