Error Diffusion achieves 96.7% on MNIST under Dale's principle
Biologically plausible learning without backprop reaches competitive benchmarks.
A new paper from researchers including Yutaro Yamada, David Ha, and Robert Lange tackles a fundamental constraint of biological neural circuits: Dale's principle, where each neuron's synapses are uniformly excitatory or inhibitory. Artificial networks that enforce this rule must coordinate separate excitatory and inhibitory populations, which fundamentally changes how credit is assigned during learning. The team proposes modulo error routing to extend Error Diffusion (ED) beyond binary classification, enabling a dual-stream excitatory/inhibitory architecture to learn without backpropagation or weight transport.
The method achieves 96.7% accuracy on MNIST and establishes a 61.7% baseline on CIFAR-10, demonstrating that representation learning is possible under strict Dale's principle. For reinforcement learning, they integrated ED with PPO (ED-PPO) and tested on continuous-control tasks in Google Brax and the open-ended exploration task Craftax, finding competitive performance relative to Direct Feedback Alignment. The work opens the door to more energy-efficient neuromorphic hardware that mirrors biological learning.
- Modulo error routing extends Error Diffusion to multi-class classification, overcoming previous binary-only limitation.
- Achieves 96.7% on MNIST and 61.7% on CIFAR-10 while strictly enforcing Dale's principle (excitatory/inhibitory neuron separation).
- ED-PPO matches Direct Feedback Alignment on continuous-control (Brax) and exploration tasks (Craftax).
Why It Matters
Biologically plausible learning rules like this could enable highly energy-efficient neuromorphic AI hardware.