Research · MarkTechPost ·
Sakana AI’s Error Diffusion Trains Dale-Compliant Dual-Stream Networks, Reaching 96.7% MNIST and 61.7% CIFAR-10 Without Backpropagation
Sakana AI presents Error Diffusion, a training method for dual-stream excitatory/inhibitory networks that follows Dale’s principle without backpropagation. The approach reportedly reaches 96.7% accuracy on MNIST and 61.7% on CIFAR-10, with experiments also examining reinforcement learning and task-specific ablations.