The brain is not able to use this mechanism, which dominates backpropagation. Specifically, backward passes require exact translations of weight matrices. It is important to note that this is not the Problems with weight transportation. Sakana AI is a new product. paper, Diffusing BlameThis constraint is faced directly by the research team. The team of researchers trains networks to obey. Dale’s principle While avoiding all weight transport.
What is error diffusion?
Error Diffusion is a rule of local learning, which was first introduced by Kaneko (2000). Three signals are all that is needed to update the weight. The three signals are the presynaptic signal, a derivative of postsynaptic stimulation, and one global error. Consequently ED doesn’t use random feedback matrices, or forward-transposed weights. Due to this locality, ED is compatible with Dale’s principal. Prior research has only shown ED to be applicable for binary classification and the MNIST.
Dual-Stream Architecture
Each layer was split into two streams by the team of researchers to meet this requirement. Two streams are excitatory.The p), aThen,d the other is inhibitory (n). The forward pass computes excitatory-minus-inhibitory preactivations for each stream:
p_i = φ_i( +p_{i-1} Wpp − n_{i-1} Wnp + bp )
n_i = φ_i( +n_{i-1} Wnn − p_{i-1} Wpn + bn )
The weight matrixes are all non-negative. Biases BThe p The following are some examples of how to get started: BThen, They do not have to be negative. The negation sign before You can also find out more about WNp The following are some examples of how to get started: You can also find out more about WPn These are not learned, but structural. Cross-stream connections are therefore inhibitory, while learnable weights continue to remain positive. This design needs four weight sub-matrices per layer. As a result, it uses roughly 4× more parameters than a single-stream network. For the same architecture, that is ∼32M versus ∼8M for DFA.
Modulo Error Routing
This architecture is the foundation for all future extensions. modulo error routing. Error Diffusion is now a more complex classification than binary. For hIdden unit i. The researchers define the routing r(i) = i mod You can also find out more about C. Here, C This is called the output dimension. This unit learns then from the component of routed errors. Essentially, every hidden unit gets a specific output channel. Unlike DFA whose random feedback matrices, ED utilizes this structured correspondence.
Three Innovations in Classification
To build on this route, The research team has added three new fixes to multi-class classification.
- Specific sigmoid sizes for each layer Use
φi(z) = 1/(1 + e−2z/αi). Since the sigmoid derivative directly gates the error signal, attenuation is severe. In fact, post-hoc analysis reveals a 25× decay from the output to the first hidden layer. Wider sigmoids keep derivatives larger, preventing premature saturation. The team sets α = 3.0 for CIFAR-10 convolutional layers and α = 6.0 for fully connected layers. - Batch-centered class error subtracts the per-class mini-batch mean. This makes the one-vs-all error zero-mean across the batch for every class. It thereby reduces persistent suppression caused by the 9:1 target imbalance.
- Asymmetric initialization scales excitatory weights by 1.5× and inhibitory weights by 0.5×. That gives an expected E/I scale ratio of 3:1, while the output layer stays symmetric.
function diffuseError(){
if (animating) return;
animating = true;
var out = nodes.filter(function(n){return n.li===layers.length-1;});
var tNode = out[target];
// highlight target output
out.forEach(function(n){ setNode(n, n.idx===target ? ‘#FFC24B’ : ‘#1a1a1a’); });
if (mode === ‘ed’){
status.innerHTML = ‘Error Diffusion: output error is broadcast directly to hidden units. Unit i learns from class r(i) = i mod C. No transposed weights are used.’;
var hid = nodes.filter(function(n){ return n.kind === ‘hid’; });
var done = 0;
hid.forEach(function(h){
var routed = h.idx % C; // r(i) = i mod C
var src = out[routed];
var isTargetChannel = (routed === target);
var col = isTargetChannel ? ‘#FFC24B’ : ‘#7a5a1a’;
pulse(src.cx, src.cy, h.cx, h.cy, col, 620, function(){
done++;
setNode(h, isTargetChannel ? ‘#FFC24B’ : (h.ei===’n’ ? ‘#9a9a9a’ : ‘#FF7A18’));
if (done === hid.length){
animating = false;
status.innerHTML = ‘Done. Each hidden unit received its routed error component, gated locally by its own activation derivative. Update: ΔWpp ∝ ApT(φ′(Zp) ⊙ R).’;
}
});
});
} else {
status.innerHTML = ‘Backpropagation: the error must travel layer by layer backward, using exact transposes of forward weights — the weight transport problem.’;
var li = layers.length – 1;
function back(){
if (li c’ + idx + ‘. Hidden units where i mod C = ‘ + idx + ‘ will receive this channel’s error.’;
};
clsBar.appendChild(b);
})(i);
}
document.getElementById(‘modeSeg’).addEventListener(‘click’, function(e){
if (e.target.tagName !== ‘BUTTON’) return;
mode = e.target.getAttribute(‘data-mode’);
Array.prototype.forEach.call(this.children, function(x){ x.className=””; });
e.target.className=”on”;
status.innerHTML = mode === ‘ed’
? ‘Error Diffusion mode: error is routed directly to hidden units — no weight transport.’
: ‘Backpropagation mode: error crawls backward through transposed weights for contrast.’;
});
document.getElementById(‘fwdBtn’).onclick = forwardPass;
document.getElementById(‘errBtn’).onclick = diffuseError;
document.getElementById(‘resetBtn’).onclick = function(){
resetVisual();
document.getElementById(‘errBtn’).disabled = true;
status.innerHTML = ‘Reset. Excitatory units are orange, inhibitory units are grey. All weights stay non-negative.’;
};
layout();
// auto-resize for WordPress iframe
function postH(){
var h = document.getElementById(‘edwrap’).offsetHeight + 40;
if (window.parent) window.parent.postMessage({ type: ‘mtp-ed-resize’, height: h }, ‘*’);
}
window.addEventListener(‘load’, postH);
window.addEventListener(‘resize’, postH);
setTimeout(postH, 300);
setTimeout(postH, 1200);
})();
“>
Performance
With all three innovations, Error Diffusion (ED) reaches 96.7% on MNIST and 61.7% on CIFAR-10. In contrast, seed ED without them collapses to 50.4% and 11.6%. DFA scores higher on both tasks but violates Dale’s principle, using ∼2.84M negative weights. Notably, this is the first time ED has trained convolutional networks. Previously, Fujita (2026) reached ∼55.2% on CIFAR-10 using a flattened MLP. Even so, 61.7% remains far from standard gradient-based methods.
| Method | MNIST | CIFAR-10 | Dale-compliant | Notes |
|---|---|---|---|---|
| Proposed ED | 96.7% | 61.7% | Yes | All weights non-negative; first ED on CNNs |
| Seed ED | 50.4% | 11.6% | Yes | No innovations; α = 1.0, raw error, symmetric init |
| DFA | 97.6% | 69.1% | No | Random feedback; ∼2.84M negative weights |
The Ablation Reversal
Interestingly, the innovations’ importance flips between tasks. On MNIST, removing layer-specific widths is catastroPhIc (−71.4 pp), collapsIIf (mode === “ed”)If (done === hidden.length)ng accuracy toward chance. Batch-centering barely matters there (−0.3 pp). On CIFAR-10, however, the order reverses. Removing batch-centered error becomes the largest drop (−47.9 pp), collapsing four of five seeds. You can also contact us by clicking here.If you want to know more about, click here.his reversal exposes task-dependent credit-assignment bottlenecks invisible to single-benchmark evaluation.
Error Diffusion in Reinforcement Learning
Beyond classification, the research team integrate ED with Proximal Policy Optimization (PPO). They call the result ED-PPO and test it on Brax locomotion and Craftax. Here, policy-output error is routed to hidden units by output channel. For the scalar value network, the error is broadcast to all units. Importantly, ED-PPO drops the three classification innovations entirely. Across five seeds, ED-PPO beats BP-PPO on HalfCheetah (5494 vs 3520; p
Use Cases and Examples
Three settings make this concrete:
- Neuromorphic and photonic hardware often encodes non-negative synaptic magnitudes physically. ED’s fixed-sign routing maps cleanly onto such substrates, complementing prior photonic DFA work.
- The non-negative floor drives 37.3% of weights to the floor (10⁻⁴) after training. Inhibitory cross-stream fully connected connections are pruned most, up to 68.8%. This implicit sparsity hints at model compression “for free.”
- The dedicated inhibitory stream may help continual and open-ended learning. It provides a structural mechanism for dampening large gradient excursions.
Comparison
How Dale-Compliant Error Diffusion Compares
Proposed approach vs. other backpropagation-free and biologically motivated learning rules. “Dale-compliant” means separate excitatory/inhibitory populations with non-negative weights. Method names link to primary sources.
Takeaway: most backprop-free rules relax Dale’s principle, and most Dale-compliant networks still rely on backpropagation. Error Diffusion satisfies both at once, and is the first to carry that combination into reinforcement learning.
Minimal Code Sketch
The equations translate into a compact, illustrative update loop:
import torch
def dual_stream_forward(p, n, Wpp, Wnp, Wnn, Wpn, bp, bn, phi):
# All W >= 0; cross-stream signs are hardcoded inhibitory (Dale's principle)
p_next = phi(p @ Wpp - n @ Wnp + bp) # excitatory stream
n_next = phi(n @ Wnn - p @ Wpn + bn) # inhibitory stream
return p_next, n_next
def routed_error(S, H, C): # S: output error, shape (B, C)
M = torch.zeros(H, C)
for i in range(H):
M[i, i % C] = 1.0 # r(i) = i mod C
return S @ M.T # R = S M^T, shape (B, H)
def ed_update(A_p, Z_p, R, phi_deriv):
U_p = phi_deriv(Z_p) * R # local postsynaptic drive
return A_p.T @ U_p # dWpp ∝ A_p^T U_p, shape (K, H)
Key Takeaways
- Sakana AI’s Error Diffusion trains Dale-compliant dual-stream networks without weight transport or random feedback matrices.
- Modulo error routing (
r(i) = i mod C) scales the rule past binary classification to 96.7% MNIST and 61.7% CIFAR-10. - The importance of three classification innovations has reversed between MNIST & CIFAR-10 highlighting task-dependent credit assignment bottlenecks.
- ED-PPO uses the same reinforcement learning architecture as DFA-PPO, but is superior on Craftax.
- Dale’s principle costs 0.9–7.4 points versus DFA on classification, quantifying the price of non-negative weights.
Click here to find out more Paper. Also, feel free to follow us on Twitter Don’t forget about our 150k+ML SubReddit Subscribe now our Newsletter. Wait! What? now you can join us on telegram as well.
You can partner with us to promote your GitHub Repository OR Hugging Page OR New Product Launch OR Webinar, etc.? Connect with us
Asif Razzaq serves as the CEO at Marktechpost Media Inc. As an entrepreneur, Asif has a passion for harnessing Artificial Intelligence to benefit society. Marktechpost was his most recent venture. This platform, which focuses on machine learning and deep-learning news, is unique for the way it is presented in a technical manner, but is also easily understood by all audiences. Over 2 million views per month are a testament to the platform’s popularity.

