Sakana AI’s Error Diffusion Trains Dale-Compliant Dual-Stream Networks, Reaching 96.7% MNIST and 61.7% CIFAR-10 Without Backpropagation
96.7% on MNIST without backpropagation. Sakana AI just proved biologically plausible neural networks can scale.
Why it matters
Sakana AI's Error Diffusion method demonstrates that training neural networks without backpropagation—using a mechanism closer to biological constraints—can reach competitive accuracy on standard benchmarks. This challenges the assumption that backprop is necessary for modern AI and opens pathways for neuromorphic and brain-inspired computing architectures.
The key facts
5 to know96.7% accuracy on MNIST with Dale-compliant dual-stream networks
61.7% accuracy on CIFAR-10 without backpropagation
Error Diffusion sidesteps weight transport problem in biological circuits
Method scales via modulo error routing from MNIST to CIFAR-10 and reinforcement learning
Task-dependent ablations reveal mechanism insights
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: Backpropagation relies on weight transport, which biological circuits likely cannot implement. Sakana AI's Error Diffusion sidesteps that constraint, training dual-stream excitatory/inhibitory networks that obey Dale's principle. This piece breaks down how modulo error routing scales the rule from…