Sakana AI, the Tokyo-based lab known for nature-inspired approaches to machine learning, said on X on July 17 that it built a training method obeying a constraint most artificial networks ignore. The technique, called Diffusing Blame, has been accepted at the ALIFE 2026 conference. It matters because the method drops a shortcut that neuroscientists have long flagged as unrealistic.
Real neurons follow Dale’s principle: a neuron is either excitatory or inhibitory, never both, depending on which population it belongs to. Standard artificial networks make no such distinction. A single unit can push a downstream neuron up in one context and down in another, whatever lowers the loss. Backpropagation widens the gap further.
To assign credit for an error, backprop needs a backward pass that exactly mirrors the forward connections, a requirement researchers call weight transport: brains have no known mechanism for keeping a mirrored, symmetric copy of every connection running in reverse. Diffusing Blame is Sakana AI’s attempt to train a network without that mirrored copy while still respecting Dale’s principle.
The method builds on an older idea called Error Diffusion, which sends a single global error signal to every hidden unit instead of routing it backward through mirrored weights. Sakana AI splits each layer so that excitatory and inhibitory populations run as distinct streams, governed by four weight matrices constrained to stay non-negative, so a connection’s sign comes from which population it belongs to rather than a value the network can flip during training. The lab’s stated contribution is extending Error Diffusion past binary classification into multi-class problems through a scheme it calls modulo error routing.
The team tested the approach in two settings: image classification, and reinforcement learning using PPO agents on the Ant, Humanoid, and HalfCheetah locomotion tasks. In both, Sakana AI compared the local updates its method produced against the gradients standard backpropagation would have computed. Among the routing schemes tested, modulo routing showed the closest alignment with true backprop gradients, per the lab’s thread. Sakana AI did not publish accuracy or reward numbers, only the claim that the constrained networks learn competitively.
The appeal beyond neuroscience is architectural. Chips built for neuromorphic or analog computing struggle to implement backprop precisely because it demands that mirrored backward weight copy. A learning rule built from local, non-negative updates maps far more naturally onto hardware designed to mimic real synapses. If Diffusing Blame or a descendant of it holds up, it could open a path to training directly on that class of hardware rather than emulating backprop in software first.
That “if” is doing real work. The result comes from an artificial-life venue rather than a flagship machine-learning conference, and the demonstration covers small-scale image classification plus three modest locomotion tasks, not anything close to a frontier language or vision model. Sakana AI has not reported training time, energy use, or how the method behaves as network depth and parameter count grow, all of which determine whether a Dale-constrained rule is a research curiosity or a real alternative to backprop.
Teams working on neuromorphic hardware or biologically inspired architectures should log this as an early signal worth tracking, not a technique ready for a product roadmap. The next test that would move this from promising to proven is whether Sakana AI or another group reproduces the gradient-alignment result on a network with real depth.
Sakana AI posted these findings on its official X account on July 17, 2026.
