Summary
Sakana AI’s paper Diffusing Blame (arXiv, ~July 17; accepted ALIFE 2026) extends Error Diffusion to multi-class CNNs and RL under Dale’s principle (non-negative E/I dual streams, no weight transport). Results: 96.7% MNIST, 61.7% CIFAR-10; ED-PPO competitive on Brax/Craftax. Quantifies a “biology tax” vs Direct Feedback Alignment while remaining Dale-compliant.
PreScreening Notes
- Score 4 — rejected: Interesting bio-inspired ML research from a notable lab, but benchmark scope (MNIST/CIFAR) and topic (Dale-compliant error diffusion without backprop) are niche for our developer/finance readership.
- Not spam/outdated; domain-adjacent AI research — fails the “worth pursuing now” bar versus concurrent major launches and regulation.
Source Analysis
[Added during prescreening/evaluation]
Research Notes
[Added during analysis]
Draft Article
[Added during reporting — in Turkish]