Sakana AI Builds a Brain-Like Neural Network That Learns Without Backpropagation

Every neuron in your brain is either excitatory or inhibitory, and it never switches sides. This is Dale's principle, a fundamental rule of biological neural circuits. Standard artificial neural networks ignore it entirely, freely mixing positive and negative weights on the same unit. A new paper accepted at ALIFE 2026, authored by Yutaro Yamada, Luca Grillotti, Rujikorn Charakorn, Sebastian Risi, David Ha, and Robert Tjarko Lange (all at Sakana AI), asks a pointed question: can a network that strictly obeys Dale's principle still learn well? Their answer, detailed in Diffusing Blame: Task-Dependent Credit Assignment in Biologically Plausible Dual-Stream Networks, is yes, with some important caveats.

Two biological rules backprop breaks

Standard deep learning simultaneously violates two properties of biological neural circuits.

• Dale's principle: Each neuron releases the same neurotransmitter at all its synapses, making it uniformly excitatory or inhibitory. Artificial networks freely assign positive and negative weights to any connection, which enables efficient credit assignment via backpropagation but has no biological counterpart.
• The weight transport problem: Backpropagation requires the backward pass to use exact transposes of the forward weight matrices. The brain has no known mechanism for copying weight matrices backward through layers.

Prior methods such as Feedback Alignment and Direct Feedback Alignment address weight transport by replacing backward weights with fixed random matrices, but none enforce Dale's principle. The authors set out to fix both problems at once.

The dual-stream architecture

The core idea is to split every layer into two parallel streams: one excitatory (p) and one inhibitory (n). Error Diffusion (ED), originally proposed for this kind of dual-stream architecture, drives learning by routing global error signals to all layers without transporting transposed forward weights or relying on random feedback matrices.

Each layer maintains four non-negative weight matrices: excitatory-to-excitatory (W_pp), inhibitory-to-excitatory (W_np), inhibitory-to-inhibitory (W_nn), and excitatory-to-inhibitory (W_pn). The sign of a connection is determined by which population it originates from, never by a learnable parameter. This structurally enforces Dale's principle. The cost is roughly 4x more parameters than an equivalent unconstrained network.

The forward pass for a single layer looks like this:

# All weight matrices are non-negative (clamped >= 0)
# Sign comes from population identity, not learned weights
p_i = sigmoid( +p_prev @ W_pp - n_prev @ W_np + b_p )
n_i = sigmoid( +n_prev @ W_nn - p_prev @ W_pn + b_n )
# Output: excitatory minus inhibitory
y_hat = y_plus - y_minus

Modulo routing: extending ED to multi-class problems

The original Error Diffusion rule only worked for binary classification. The paper's central technical contribution is modulo error routing, a scheme for scaling ED to multi-class problems. Each hidden unit is assigned a fixed output channel via r(i) = i mod C, where C is the number of output classes. The error for that output channel then serves as the unit's learning signal, broadcast directly without transporting any weight matrices backward.

This differs from Direct Feedback Alignment, which also skips weight transport but uses random feedback matrices. ED's routing is structured and deterministic: each hidden unit has a fixed assignment to a specific output class.

Three classification-specific tricks

The dual-stream architecture trained with this method reaches 96.7% on MNIST and 61.7% on CIFAR-10, showing that representation learning is possible under strict Dale compliance. Getting there required three domain-specific innovations:

• Layer-specific sigmoid widths: Post-hoc analysis revealed a 25x gradient attenuation from the output layer to the first hidden layer. Wider sigmoids (larger alpha parameter) maintain larger derivatives and prevent premature saturation. Without this, MNIST accuracy collapses from 96.7% to 25.3%.
• Batch-centered class error signals: In 10-way classification, the one-vs-all error has a 9:1 class imbalance (one positive, nine negatives per example). Subtracting the per-class batch mean centers the signal and prevents uniform output suppression. Removing this drops CIFAR-10 accuracy from 61.7% to 13.8%.
• Asymmetric E/I initialization: Excitatory weights start at 1.5x scale, inhibitory at 0.5x, giving a 3:1 ratio. Training drives this toward balance (~1:1), but the asymmetric head start prevents early instability on harder tasks.

The ablation reversal

One of the paper's most striking results is a methodological warning. The importance of each innovation completely reverses between MNIST and CIFAR-10:

Removed component	MNIST accuracy drop	CIFAR-10 accuracy drop
Layer-specific sigmoid widths	-71.4 pp (catastrophic)	-15.1 pp
Batch-centered class error	-0.3 pp (negligible)	-47.9 pp (catastrophic)
Asymmetric initialization	0.0 pp	-5.5 pp

Prior work demonstrated Error Diffusion only on binary classification and MNIST. The reversal shows that evaluating biologically plausible methods on a single benchmark can produce deeply misleading conclusions about which design choices actually matter.

Into reinforcement learning

The authors also integrated the dual-stream architecture into PPO (Proximal Policy Optimization), creating ED-PPO. The same Dale-constrained network, now with ReLU activations and without the classification-specific tricks, was evaluated on Google Brax locomotion tasks (Ant, HalfCheetah, Humanoid) and Craftax, an open-ended exploration benchmark.

Results were competitive but uneven:

• HalfCheetah: ED-PPO outperformed standard BP-PPO (5494 vs 3520 average return, p < 0.001) and matched DFA-PPO.
• Ant: ED-PPO matched both PPO variants.
• Humanoid and Craftax: ED-PPO trailed BP-PPO, with substantially higher variance across all environments.

Direct Feedback Alignment, the nearest competing backpropagation-free method, sidesteps weight transport using random feedback matrices but is not Dale-compliant. Error Diffusion satisfies both constraints. On Craftax, DFA-PPO was the weakest method overall (19.8 vs ED-PPO's 20.9 and BP-PPO's 27.0), suggesting that random feedback pathways that work for supervised tasks can fail on complex open-ended environments.

Emergent biology, unprompted

After training, the network's excitatory and inhibitory weights spontaneously converged from the 3:1 asymmetric initialization toward a near-balanced 1:1 ratio, with no explicit mechanism enforcing this. The balance was also depth-dependent: the first layer reached near-perfect E/I balance (ratio 1.03), the second became slightly inhibitory-dominant (0.90), and the third developed the strongest inhibitory bias (0.81). Balanced excitation and inhibition is a well-documented constraint in cortical computation, and the model found it on its own.

Training also induced implicit sparsity: 37.3% of weights hit the non-negative floor, with inhibitory fully-connected connections pruned most aggressively (up to 68.8%). Model compression arrived as a side effect of biological constraints.

Where it fits and where it falls short

This work is most relevant to researchers working on:

• Neuromorphic hardware: Non-negative synaptic weights map naturally to physical analog or photonic substrates, where sign is easier to implement at the population level than at the individual synapse.
• Biological modeling: ED-PPO offers a plausible account of how cortical circuits might solve RL-style problems without global weight transport.
• Implicit compression: The non-negative floor acts as a natural pruning mechanism that could be exploited at inference time.

The limitations are concrete. DFA achieves 97.6% on MNIST and 69.1% on CIFAR-10, compared to ED's 96.7% and 61.7%. That 7.4-point gap on CIFAR-10 is the quantified cost of strict Dale compliance. ED-PPO also shows substantially higher variance than BP-PPO across all RL environments, and the Craftax shortfall suggests the coarse modulo routing struggles with tasks requiring fine-grained temporal credit assignment. No code repository was linked in the paper.

The full paper is on arXiv with an HTML version. For researchers in neuro-AI, the cross-task ablation reversal alone is worth reading carefully: it is a concrete demonstration of why single-benchmark evaluation of biologically plausible methods is insufficient.