Every neural network powering your phone’s camera, the language model on your screen, and the recommendation engine choosing your next video was trained with the same four-decade-old algorithm: backpropagation. It works extraordinarily well. It also, as neuroscientists have documented since at least 1989, requires a mathematical operation the brain almost certainly cannot perform — a problem Francis Crick identified and that has driven decades of research into alternatives. Now, a paper posted to arXiv on July 17 by Tokyo-based Sakana AI has delivered the most capable demonstration yet of what it looks like when you train a neural network the way the brain actually might — and for the first time, that demonstration includes convolutional networks and reinforcement learning.
The result is 96.7% accuracy on MNIST and 61.7% on CIFAR-10 — without backpropagation, without random feedback matrices, and without violating the biological constraint that individual neurons can only excite or only inhibit, never both. The gap to the nearest competing backpropagation-free method is 7.4 percentage points on CIFAR-10. That gap is now what biologically correct training costs in accuracy terms, at least with current techniques. The full paper, Diffusing Blame, is available on arXiv.
Those numbers emerge from a specific set of engineering choices that the Sakana AI team layered onto a 25-year-old algorithm that had never before been demonstrated on convolutional networks or reinforcement learning. Understanding what those choices are — and what they revealed about the credit assignment problem in machine learning — is the substance of what makes this result significant.
What Dale’s Principle Is and Why Deep Learning Ignores It
Dale’s principle, named for English neuroscientist Henry Hallett Dale and formalized by Nobel laureate John Eccles in 1954, holds that a neuron releases the same set of chemical transmitters at all of its synapses. In practice, neurons are classified as either excitatory (they push neighboring neurons toward firing) or inhibitory (they suppress them). The same neuron does not do both.
Standard neural networks trained with backpropagation violate this rule routinely. A single artificial neuron can have positive weights — excitatory connections — to some downstream neurons and negative weights — inhibitory connections — to others. This is architecturally convenient and mathematically powerful. It is also something no biological neuron does.
The engineering reason this matters goes beyond aesthetics. Neuromorphic chips — hardware architectures designed to operate at a fraction of the power of conventional GPUs by co-locating memory and computation — often encode synaptic weights as physical quantities that are inherently non-negative: light intensities in photonic systems, charge levels in analog memory cells. An algorithm that requires arbitrary-sign weights cannot run natively on that hardware. A learning rule that works with non-negative weights only is not just biologically faithful — it is directly deployable on a class of hardware that standard deep learning cannot touch.
This is why Dale-compliant learning has attracted sustained research interest even as backpropagation’s dominance has grown. The problem has been performance: architectures constrained to keep neurons either purely excitatory or purely inhibitory have consistently underperformed their unconstrained counterparts, sometimes dramatically.
How Error Diffusion Avoids the Weight Transport Problem
The algorithm at the center of Sakana AI’s paper is called Error Diffusion, first proposed by Kaneko around 2000. It has remained a curiosity for a quarter century, demonstrated only on binary classification and simple versions of the MNIST handwritten-digit dataset. Its appeal is its locality: each weight update in Error Diffusion depends on just three signals — the activity of the neuron sending the signal, the derivative of the activation function at the receiving neuron, and a single global error sign shared across the whole network.
That last part is the key difference from backpropagation. Backpropagation sends error signals backward through the network using the exact mathematical transpose of every forward weight matrix — a globally coordinated, biologically implausible operation that Francis Crick flagged as impossible for real neurons in 1989. Error Diffusion sends only a single signed value — positive or negative — to indicate whether the network made an error in the right or wrong direction. No transposed matrices, no layer-by-layer gradient computation.
The nearest competing family of backpropagation-free methods is Direct Feedback Alignment, or DFA, which sidesteps weight transport by using random feedback matrices — matrices whose values are fixed at initialization and never updated. DFA performs well on benchmarks: 97.6% on MNIST, 69.1% on CIFAR-10, according to Sakana AI’s paper. But DFA is not Dale-compliant. The researchers found approximately 2.84 million negative weights in a DFA network of comparable size. That makes DFA biologically plausible by one criterion (no weight transport) but not by another (Dale’s principle).
Error Diffusion satisfies both. That combination — no weight transport, no random feedback, strict Dale compliance — is what no prior method had demonstrated on anything harder than MNIST.
Three Engineering Choices That Changed the Picture
Prior work, including a 2025 paper from Kazuhisa Fujita at Komatsu University, reached approximately 55.2% on CIFAR-10 using a flat, non-convolutional network structure trained with Error Diffusion. Sakana’s team added three targeted innovations on top of the base algorithm — and ran it on convolutional networks for the first time.
The first innovation is the dual-stream architecture. To keep all trainable weights non-negative while still allowing inhibitory effects, the team split each layer into two parallel streams: an excitatory stream and an inhibitory one. The cross-stream connections — the pathways from the excitatory stream to the inhibitory stream and vice versa — are fixed as inhibitory via hardcoded negation signs in the forward pass. No weight in either stream is ever allowed to go negative. The cost is roughly four times more parameters per layer than a single-stream network. The biological fidelity is exact.
The second innovation is modulo error routing. The original Error Diffusion algorithm handled only binary outputs. To extend it to multi-class problems, the team assigned each hidden unit a fixed output channel: unit i gets error feedback from channel i modulo C, where C is the number of output classes. Unlike DFA, which assigns these feedback pathways randomly, Error Diffusion uses a deterministic, structured correspondence. Each hidden unit is permanently linked to one specific output.
The third is a set of classification-specific corrections: layer-specific sigmoid widths (to prevent error signals from shrinking to near-zero as they pass through deep layers, where the researchers measured a 25x decay from output layer to first hidden layer), batch-centered class error (to correct a persistent suppression bias caused by the 9-to-1 imbalance between positive and negative targets in one-versus-all classification), and asymmetric weight initialization giving the excitatory stream a 3-to-1 starting advantage over the inhibitory stream. The full ablation data on these choices is in the arXiv paper.
What the Ablation Study Reveals About Single-Benchmark AI Research
One of the paper’s more consequential findings is methodological. When the three innovations are removed individually to assess their contribution, the relative importance reverses completely depending on which benchmark is used.
On MNIST, removing the layer-specific sigmoid widths is catastrophic: accuracy collapses by 71.4 percentage points, toward near-chance performance. Removing batch-centering on MNIST barely registers — a 0.3-point drop. On CIFAR-10, the situation inverts almost entirely: removing batch-centering is the dominant failure, causing a 47.9-point collapse and causing four of five training runs to fail entirely, while sigmoid widths matter much less.
The researchers interpret this as evidence of task-dependent credit-assignment bottlenecks — the fundamental obstacles to learning differ depending on the task’s structure and difficulty, and a method that is evaluated on only one benchmark will miss this entirely. It is a methodological warning with implications that extend beyond this paper: a learning algorithm that excels on MNIST may be solving a different underlying problem than the same algorithm applied to CIFAR-10.
Into Reinforcement Learning: Where Random Feedback Breaks Down
The paper’s most significant extension is an integration of Error Diffusion with Proximal Policy Optimization, the dominant algorithm for training reinforcement learning agents. Tested on Brax physics-based locomotion tasks and the open-ended Craftax survival environment, the result — called ED-PPO — offers one of the clearest experimental comparisons in the paper.
On HalfCheetah, a standard Brax locomotion benchmark, ED-PPO substantially outperforms standard backpropagation-based PPO (cumulative reward 5,494 versus 3,520; p < 0.001), and performs comparably to DFA-PPO. On the Ant locomotion task, all three methods perform similarly. On Craftax — the open-ended environment modeled loosely on Minecraft — DFA-PPO is the weakest method, scoring 19.8 against standard PPO’s 27.0, while ED-PPO falls in between.
The Craftax result is notable. DFA’s random feedback matrices, which perform adequately on structured tasks with fixed output dimensions, appear to fail when the task requires flexible credit assignment across diverse and open-ended objectives. Error Diffusion’s structured routing — each unit deterministically linked to a specific output channel — appears to handle this better. Whether this generalizes to more complex open-ended environments is an open question.
Hardware, Sparsity, and the Edge Inference Story
The paper identifies three concrete engineering implications that extend beyond benchmark numbers.
First, the non-negative weight constraint and fixed-sign inhibitory routing map cleanly onto neuromorphic and photonic hardware substrates, where synaptic magnitudes are often encoded as non-negative physical quantities. This makes the method potentially deployable on computing architectures — Intel’s Loihi 2 platform, photonic neural network chips, analog in-memory computing systems — that conventional deep learning cannot natively address.
Second, and less obviously, the non-negative weight floor creates implicit model compression. After training, 37.3% of all weights converge to the minimum allowed value and are effectively pruned. The inhibitory cross-stream connections are pruned most aggressively, reaching up to 68.8% sparsity in fully connected layers. This sparsity emerges without any explicit regularization or pruning procedure — it is a consequence of the architectural constraint. For deployment on low-power edge inference hardware, where sparse models translate directly to reduced computation and memory access, this free compression is a meaningful practical benefit that goes beyond neuromorphic chips specifically.
Third, the dedicated inhibitory stream provides a structural mechanism for dampening large gradient excursions, which the researchers propose may improve stability in continual and open-ended learning settings where catastrophic forgetting is a central unsolved problem.
Why Sakana AI’s Profile Matters Here
Biologically plausible learning has historically been a niche pursued earnestly in computational neuroscience but rarely by labs with the profile or resources to push it toward practical significance. Sakana AI is not a neuroscience lab. Founded in July 2023 by former Google Brain researcher David Ha and Transformer co-author Llion Jones, the company raised approximately $379 million through its November 2025 Series B, valuing it at $2.65 billion and making it Japan’s most valuable private startup. Investors include NVIDIA, Google, Mitsubishi UFJ Financial Group, and In-Q-Tel.
The lab is better known for evolutionary model merging — a technique for combining pretrained models through evolutionary search rather than training from scratch — and for the AI Scientist project, which automates scientific research workflows. Publishing serious work on Dale-compliant, backpropagation-free learning signals a deliberate broadening of the lab’s research agenda.
The paper is explicit that it is not claiming parity with gradient descent, which remains far ahead on CIFAR-10 and on every harder benchmark. What it claims — and what the ablation study supports — is that the constraints the brain actually operates under are not incompatible with meaningful machine learning, and that the performance gap between biologically constrained and unconstrained learning is now specific, quantified, and shrinking.
Can AI Learn Without Backpropagation? What the Numbers Say Now
At present, the best Dale-compliant, backpropagation-free algorithm trails the best non-Dale-compliant backpropagation-free algorithm by 0.9 percentage points on MNIST and 7.4 points on CIFAR-10. It trails standard backpropagation by larger margins. But it runs on hardware that backpropagation cannot, it prunes itself for free, and it now works on convolutional networks and reinforcement learning — neither of which had been demonstrated before.
Whether the gap continues to shrink depends on whether the field treats this as a starting point or an anomaly. A credible, well-funded lab publishing this result as a genuine research contribution is, at minimum, evidence that the question is worth taking seriously.
Frequently Asked Questions
What is the weight transport problem, and why can’t the brain do backpropagation?
Backpropagation computes how to adjust each weight in a neural network by sending error signals backward through the network using the exact mathematical transpose of every forward weight matrix. This requires each neuron to know, with mathematical precision, the weights of connections elsewhere in the network — a kind of global coordination that neuroscientists have found no evidence for in biological brains. Francis Crick identified this problem in 1989. The brain learns through local, relatively slow chemical processes at individual synapses; it does not appear to maintain the globally synchronized, precisely transposed weight information that backpropagation requires.
What does Dale’s principle cost in accuracy — and who cares?
Based on Sakana AI’s paper, strict Dale compliance currently costs 0.9 percentage points on MNIST (96.7% vs. DFA’s 97.6%) and 7.4 points on CIFAR-10 (61.7% vs. 69.1%) relative to the next best backpropagation-free method. That gap matters primarily to two communities: neuromorphic and photonic hardware engineers, whose chips physically encode synaptic weights as non-negative quantities and therefore need Dale-compliant learning to run natively; and computational neuroscientists, who need training algorithms that match brain architecture to build valid models of biological learning.
Could this kind of biologically plausible learning run on low-power devices?
The non-negative weight floor that makes Error Diffusion Dale-compliant also produces implicit model compression: 37.3% of weights converge to the minimum value and are effectively pruned after training, with inhibitory connections pruned as heavily as 68.8% in some layers. Sparse models require fewer multiplications and less memory bandwidth during inference, which translates directly to lower power consumption on edge hardware — not just dedicated neuromorphic chips, but any inference accelerator that can exploit weight sparsity. The compression emerges for free from the architectural constraint, without any explicit pruning step.
What is Sakana AI’s broader research agenda?
Sakana AI, valued at $2.65 billion after its November 2025 Series B, is primarily known for evolutionary model merging and the AI Scientist project, both described on its research blog. The “Diffusing Blame” paper represents a departure into biologically plausible learning — a research direction the company has not prominently featured before. The lab’s investor base, which includes NVIDIA and In-Q-Tel, and its growing focus on defense and government AI applications, suggests continued diversification of research areas beyond its founding evolutionary-AI focus.
