#Error Diffusion

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

Sakana AI introduces Error Diffusion, a novel training method that enables neural networks to learn without backpropagation, achieving high accuracy on MNIST and CIFAR-10 while adhering to Dale's principle.