Sakana AI researchers (Yutaro Yamada, Luca Grillotti, Rujikorn Charakorn, Sebastian Risi, David Ha, Robert Tjarko Lange) published Diffusing Blame: Task-Dependent Credit Assignment in Biologically Plausible Dual-Stream Networks (arXiv:2606.31700; MarkTechPost coverage July 17, 2026; accepted at ALIFE 2026).
Problem:
- Backpropagation requires weight transport (transposed forward weights) — biologically implausible.
- Dale’s principle: neurons are excitatory or inhibitory, not mixed-sign. Most ANNs ignore this.
Method:
- Error Diffusion (ED, Kaneko 2000): local updates from presynaptic activity, postsynaptic activation derivative, and a single global error sign — no transposed weights, no random feedback matrices.
- Dual-stream architecture: separate excitatory (p) and inhibitory (n) streams; four non-negative weight matrices per layer; ~4× parameter cost vs single-stream.
- Modulo error routing r(i) = i mod C extends ED beyond binary classification.
- Classification innovations: layer-specific sigmoid widths, batch-centered class error, asymmetric E/I initialization.
Results:
- ED: 96.7% MNIST, 61.7% CIFAR-10 (first ED on convolutional nets); seed ED without innovations collapses to 50.4% / 11.6%.
- DFA higher (97.6% / 69.1%) but not Dale-compliant (~2.84M negative weights).
- Ablation importance reverses between MNIST and CIFAR-10 (task-dependent credit-assignment bottlenecks).
- ED-PPO: integrates ED with Proximal Policy Optimization; competitive with DFA-PPO on Brax; beats BP-PPO on HalfCheetah (5494 vs 3520); stronger than DFA-PPO on Craftax open-ended exploration.
Paper: https://arxiv.org/abs/2606.31700
Coverage: MarkTechPost, July 17, 2026.