Sakana AI's error diffusion trains networks without backprop, achieves 96.7% MNIST

Sakana AI's 'Diffusing Blame' paper trains DALE-compliant dual-stream networks using error diffusion, reaching 96.7% on MNIST and 61.7% on CIFAR-10 without backpropagation. The method sidesteps the weight transport problem by avoiding exact transpose of forward weights.