explainx / blog
Sakana AI's ALIFE 2026 paper shows neural networks can learn competitively while obeying Dale's principle — no backprop weight transport required. explainx.ai explains modulo error routing, the MNIST/CIFAR-10 results, and why it matters.
Jul 14, 2026
The father of reinforcement learning left John Carmack's Keen to launch Oak Lab — algorithms that learn from runtime experience, not curated datasets. Sutton's OaK architecture targets continual world models and 20W planning. explainx.ai explains the AGI neolab wave and the Bitter Lesson tension.
Jul 21, 2026
Sakana AI launched Fugu-Cyber on July 21, 2026, extending its multi-model orchestrator into vulnerability verification and threat-intelligence detection. The benchmark scores are strong, but Sakana's more important argument is that enterprises need verification harnesses and human expertise, not merely access to a frontier cyber model.
Jul 21, 2026
Excited that AI helped disprove the Jacobian conjecture, but not sure what the conjecture says? This beginner-friendly guide builds from derivatives and local invertibility to the explicit three-variable counterexample, its independent verification, and the still-open plane case.
Most of deep learning runs on a trick that no real neuron can perform. Backpropagation, the algorithm behind nearly every modern neural network, requires each layer's backward pass to use an exact transposed copy of that layer's forward weights — a computation biologists have never found evidence that real synapses can do. It's called the weight transport problem, and it's one of the core reasons backpropagation is considered biologically implausible, however well it works in practice.
Sakana AI's new paper, "Diffusing Blame: Task-Dependent Credit Assignment in Biologically Plausible Dual-Stream Networks," tackles a narrower but related constraint: Dale's principle — the observation that real neurons are predominantly either excitatory or inhibitory, never a mix of both toward different targets. Standard artificial networks ignore this too, letting any single unit send positive weights to one downstream neuron and negative weights to another. The paper, accepted at ALIFE 2026 and posted to arXiv on June 30, 2026, shows that a network built to strictly respect Dale's principle can still learn competitively — across both image classification and reinforcement learning — without ever transporting transposed weights backward.
| Question | Answer |
|---|---|
| What's the paper called? | "Diffusing Blame: Task-Dependent Credit Assignment in Biologically Plausible Dual-Stream Networks" |
| Where published? | arXiv:2606.31700, accepted at ALIFE 2026 |
| Who wrote it? | Yamada, Grillotti, Charakorn, Risi, Ha, Lange (Sakana AI) |
| Core constraint tested? | Dale's principle — neurons are fixed excitatory or inhibitory |
| Method? | Error Diffusion (ED) + new "modulo error routing" |
| MNIST result? | 96.7% accuracy |
| CIFAR-10 result? | 61.7% baseline |
| RL result? | ED-PPO competitive with Direct Feedback Alignment on Ant, Humanoid, HalfCheetah, Craftax |
| Why it matters? | Path toward learning rules that are effective and biologically plausible |
Every standard neural network layer computes a weighted sum of its inputs, and during training, backpropagation works backward through the network computing how much each weight contributed to the final error. To do that backward pass, each layer needs access to the exact transpose of the weight matrix it used going forward. That's computationally trivial for a piece of software — matrices and their transposes live in the same memory — but neuroscientists have never identified a plausible biological mechanism for a synapse to instantly "know" the precise strength of a corresponding connection elsewhere in the circuit. This mismatch between how artificial networks train and how biological brains appear to learn has been a long-running open problem in computational neuroscience, one Sakana AI's paper frames as still unresolved for networks that also try to respect Dale's principle.
Dale's principle adds a second layer of biological realism most artificial networks skip entirely. In real cortical tissue, a given neuron's synapses are consistently excitatory (pushing downstream neurons toward firing) or consistently inhibitory (pushing them away from firing) — not a mix of both depending on which target neuron receives the signal. Standard artificial networks have no such constraint: any unit can send a positive weight to one neuron and a negative weight to another without penalty. Enforcing Dale's principle means restructuring the network itself — Sakana's approach uses separate excitatory and inhibitory streams with four non-negative weight matrices, so a synapse's sign comes from which population it belongs to, rather than being a free parameter the network can set to whatever value minimizes loss.
The paper builds on Error Diffusion (ED), a prior biologically plausible learning rule that routes a single global error signal directly to every hidden layer, sidestepping both the weight-transport requirement and the need for random feedback matrices (the fix used by an earlier alternative called Direct Feedback Alignment, referenced later in the paper as a comparison baseline). The problem with the original Error Diffusion method was scope: it had only been shown to work well on binary classification tasks and had not been demonstrated to scale to harder, real benchmarks like CIFAR-10.
Sakana AI's contribution is modulo error routing — a way of extending Error Diffusion from binary decisions to full multi-class classification using a routing scheme based on modular arithmetic rather than exact transposed weights. That's the piece that lets the same dual-stream excitatory/inhibitory architecture handle problems with many possible output classes (digit recognition across ten classes, object recognition across ten categories) instead of only yes/no decisions.
Beyond the core routing extension, the paper adds three domain-specific refinements for the classification setting: layer-specific sigmoid widths, batch-centered class error signals, and asymmetric initialization. The paper's ablation analysis found something methodologically interesting on its own — the relative importance of these three tweaks reverses between MNIST and CIFAR-10. A refinement that matters most for the simpler MNIST digit-recognition task isn't necessarily the one that matters most for the harder CIFAR-10 image task, which the authors describe as exposing "task-dependent credit-assignment bottlenecks invisible to single-benchmark evaluation" — a reminder that biologically plausible learning-rule research can't rely on a single benchmark to validate a method.
On the classification side, the Dale-constrained dual-stream network trained with modulo error routing reached 96.7% accuracy on MNIST and established a 61.7% baseline on CIFAR-10. Neither number is a new state-of-the-art for either benchmark — standard backprop-trained networks comfortably exceed both figures — but that's not the point of the paper. The claim is narrower and, in context, more significant: representation learning is possible at all under the combined constraints of Dale's principle and no weight transport, on a benchmark harder than MNIST alone, which is where prior Dale-constrained approaches had stalled.
The reinforcement learning experiments extend the claim beyond static image classification into a much noisier, more dynamic training setting. The team integrated Error Diffusion into Proximal Policy Optimization (PPO) — the resulting method, ED-PPO — and tested it on three continuous-control locomotion tasks from Google Brax (Ant, Humanoid, HalfCheetah) plus Craftax, an open-ended exploration benchmark. Across these tasks, ED-PPO's local, biologically plausible updates were directly compared against true backpropagation gradients, and the modulo routing scheme produced the strongest alignment among the routing variants tested. Overall, ED-PPO achieved performance competitive with Direct Feedback Alignment, an established backpropagation-free baseline method, on the same continuous-control and exploration tasks.
It's tempting to file biologically plausible learning-rule research under pure neuroscience curiosity — interesting for understanding the brain, irrelevant for building better AI systems. Two practical threads make that dismissal premature.
Neuromorphic hardware — chips designed to mimic biological neural computation directly rather than running standard matrix multiplication — is a genuine hardware research direction at companies and labs building specialized low-power AI accelerators. Learning rules that don't require weight transport or backpropagation's exact gradient bookkeeping are a much better architectural fit for that class of hardware, since neuromorphic chips are specifically built around local, biologically-inspired computation rather than the global synchronized backward pass standard backprop requires.
Credit assignment under biological constraints is also a genuinely open scientific question with implications for how neuroscientists interpret real brain activity. If a Dale-constrained, weight-transport-free learning rule can match or approach backprop-level performance, it strengthens the case that something structurally similar to Error Diffusion — or a related local learning rule — could plausibly be what real biological circuits are doing, informing hypotheses neuroscientists test against actual neural recordings.
The paper's own framing keeps the claim appropriately modest: it demonstrates a "potential path toward learning rules that are both effective and more biologically plausible," not a replacement for backpropagation in production AI systems. For a lab like Sakana AI — better known publicly for frontier-adjacent products like its Fugu cybersecurity models and Sakana Chat — the paper is a reminder that foundational, less immediately commercial research on learning rules and biological plausibility is still part of its research output alongside more visible product launches.
Error Diffusion isn't the only attempt to train neural networks without weight transport. Direct Feedback Alignment (DFA), the baseline the paper compares ED-PPO against directly, solves the transport problem differently — by replacing the exact transposed weight matrix with a fixed, random feedback matrix that never updates during training. DFA has been studied for several years as a biologically-motivated alternative and has shown it can train reasonably deep networks, though typically with some performance gap versus full backpropagation on harder tasks.
What distinguishes Sakana's approach is that it tackles two biological constraints simultaneously rather than one. DFA-style methods typically ignore Dale's principle entirely — a DFA network's units still freely mix positive and negative outgoing weights, just as standard backprop networks do. Error Diffusion's dual-stream excitatory/inhibitory architecture enforces Dale's principle from the ground up, meaning the comparison against DFA in the paper's RL experiments is a genuinely harder bar: ED-PPO is trying to match a method that has an easier time of it (no Dale's principle constraint) while carrying the additional structural restriction of fixed-sign neuron populations. That ED-PPO reaches competitive performance under a strictly harder constraint set is the more interesting part of the result, even though the paper frames it modestly as "competitive" rather than superior.
This also connects to a broader current in AI research around credit assignment — the general problem of figuring out which parts of a network deserve "credit" or "blame" for an outcome, which is what backpropagation, DFA, and Error Diffusion are all different mechanisms for solving. As reinforcement learning and agentic systems increasingly involve long chains of decisions before a reward signal arrives, the credit-assignment problem shows up again in a different form — deciding which action in a long trajectory actually caused a good or bad outcome. The paper's RL experiments on Craftax, an open-ended exploration task with sparse and delayed rewards, put Error Diffusion's local credit-assignment mechanism through exactly that harder, more realistic version of the problem, beyond the simpler dense-reward locomotion tasks like Ant and HalfCheetah.
Paper details reflect the arXiv preprint (2606.31700, submitted June 30, 2026) and Sakana AI's July 17, 2026 announcement thread. Benchmark figures are as reported in the paper; independent reproduction has not been published as of this post.
