🤖 machine learning

Diffusing Blame: Task-Dependent Credit Assignment in Biologically Plausible Dual-Stream Networks

This paper introduces a biologically plausible dual-stream architecture using Error Diffusion with modulo error routing and task-specific innovations to achieve strong performance on MNIST, CIFAR-10, and continuous reinforcement learning tasks while strictly adhering to Dale's principle.

Original authors: Yutaro Yamada, Luca Grillotti, Rujikorn Charakorn, Sebastian Risi, David Ha, Robert Tjarko Lange

Published 2026-07-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yutaro Yamada, Luca Grillotti, Rujikorn Charakorn, Sebastian Risi, David Ha, Robert Tjarko Lange

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: A "Biological" Way to Teach AI

Imagine you are trying to teach a robot how to recognize pictures or walk across a room. Usually, we teach robots using a method called Backpropagation. Think of this like a strict teacher who has a perfect, magical map of the robot's entire brain. When the robot makes a mistake, the teacher looks at the map, traces the error all the way back to the specific neuron that caused it, and tells that exact neuron, "You did this wrong, fix it."

The Problem: Real biological brains (like ours) don't work this way. In nature, neurons are either "excitatory" (they push the system forward) or "inhibitory" (they push the system back). They never do both. Also, a brain cell doesn't have a magical map of the whole brain to know exactly how to fix its own mistakes.

The Paper's Goal: The researchers wanted to build an AI that learns like a real brain (following "Dale's Principle," which means keeping excitatory and inhibitory cells separate) but still learns effectively enough to solve hard problems like recognizing cats in photos or walking on uneven terrain.

The Solution: The "Dual-Stream" Factory

To mimic the brain, the researchers built a Dual-Stream Network. Imagine a factory with two separate conveyor belts running side-by-side:

The "Go" Belt (Excitatory): These workers only push things forward.
The "Stop" Belt (Inhibitory): These workers only pull things back or slow things down.

The final result is the difference between the two belts. If the "Go" belt is pushing hard and the "Stop" belt is weak, the machine moves forward. If the "Stop" belt is strong, it cancels out the movement. This ensures that every single connection in the network is strictly positive (either a push or a pull), just like a real brain.

The Learning Trick: "Error Diffusion"

In standard AI, the "teacher" sends a specific correction message back to the exact neuron that made the mistake. In this new system, the "teacher" doesn't know the exact path. Instead, they shout a general error signal into the room.

The researchers used a clever routing system called Modulo Error Routing. Imagine a mailroom where every package (error signal) is stamped with a number. Every worker (neuron) in the factory has a specific number. When a package arrives, it goes to the worker whose number matches the stamp.

Why this matters: It's a rough guess, not a perfect map. It's like saying, "Hey, you guys in the red shirts, you probably had something to do with that mistake," rather than pointing at one specific person. This is biologically plausible because it doesn't require the brain to have a perfect map of its own wiring.

The Results: How Well Did It Work?

The team tested this "Brain-Like" AI on two types of tasks:

1. Recognizing Pictures (Classification)

The Test: Identifying handwritten numbers (MNIST) and complex objects like cars or dogs (CIFAR-10).
The Result:
- On the easy test (numbers), it got 96.7% correct.
- On the hard test (objects), it got 61.7% correct.
The Catch: To make this work on the hard test, they had to invent three special "training wheels":
1. Adjustable Sigmoid Widths: Think of this as tuning the sensitivity of the workers. On the easy test, you needed wide sensitivity; on the hard test, you needed narrow sensitivity.
2. Balanced Error Signals: Sometimes the "Go" belt gets too excited about one type of object. They had to teach the system to balance the excitement so it didn't ignore other objects.
3. Asymmetric Start: They started the "Go" workers with more energy than the "Stop" workers.
The Surprise: The researchers found that what worked for the easy test failed on the hard test, and vice versa. The "training wheels" needed for numbers were useless for pictures, and the ones needed for pictures were useless for numbers. This proves that "one size fits all" doesn't work for brain-like learning.

2. Learning to Move (Reinforcement Learning)

The Test: Teaching virtual robots (like a cheetah or a human) to walk without falling over.
The Result: The "Brain-Like" AI learned to walk almost as well as the standard "Magic Map" AI. It performed competitively on tasks like walking on a half-cheetah or a human robot.
The Difference: Unlike the picture test, the "Brain-Like" AI didn't need those special "training wheels" (like the adjustable sensitivity) to learn how to walk. It just needed the core dual-stream structure.

The Takeaway

This paper shows that you can build AI that respects the rules of biology (separating "push" and "pull" neurons) and still learn complex tasks.

However, the most important discovery is that biological learning isn't a single, fixed recipe. The tools you need to learn to recognize a cat are completely different from the tools you need to learn to walk. If you only test your AI on one simple task, you might think you've found the perfect solution, but it might fail completely on a harder task.

The researchers also noted that this "Brain-Like" AI sometimes makes more mistakes or is less stable than standard AI, but it opens the door to building computers that work more like our own brains, which could eventually lead to more efficient and robust machines.

Technical Summary: Diffusing Blame in Biologically Plausible Dual-Stream Networks

Problem Statement

Biological neural circuits adhere to Dale's principle, where each neuron is either purely excitatory or purely inhibitory, releasing the same neurotransmitter at all synapses. In contrast, standard artificial neural networks (ANNs) utilize weights with arbitrary signs, a simplification that facilitates efficient credit assignment via backpropagation but lacks biological fidelity. Furthermore, backpropagation requires the "weight transport problem"—the backward pass must use exact transposes of forward weight matrices, a mechanism with no known biological support.

While several biologically plausible alternatives exist (e.g., Feedback Alignment, Direct Feedback Alignment, Predictive Coding), most fail to enforce Dale's principle, allowing arbitrary weight signs. Error Diffusion (ED), a local learning rule where updates depend on presynaptic activity, postsynaptic derivatives, and a global error sign, is naturally compatible with Dale's principle. However, prior implementations of ED were limited to binary classification or simple tasks like MNIST. The central challenge addressed by this work is whether a Dale-compliant architecture can achieve competitive performance across both supervised classification (including complex tasks like CIFAR-10) and reinforcement learning (RL), and how task-dependent factors influence credit assignment under these constraints.

Methodology

Dual-Stream Architecture

The authors adopt a dual-stream architecture to structurally enforce Dale's principle. Each layer is split into an excitatory ( $p$ ) and an inhibitory ( $n$ ) stream.

Forward Pass: The pre-activation for each stream is computed using four non-negative weight matrices ( $W_{pp}, W_{np}, W_{nn}, W_{pn}$ ). Cross-stream connections ( $W_{np}, W_{pn}$ ) are structurally inhibitory (subtracted), while all learnable parameters remain non-negative.
Parameter Overhead: This design requires $\sim4\times$ more parameters than a single-stream network of equivalent width.

Error Diffusion (ED) Update Rule

Instead of backpropagating errors, ED routes a global output error signal directly to hidden units.

Modulo Error Routing: To extend ED beyond binary classification, the authors introduce modulo error routing. Each hidden unit $i$ is assigned a fixed output channel $r(i) = i \mod C$ (where $C$ is the output dimension). The learning signal for unit $i$ is the error component corresponding to its assigned channel.
Update Mechanism: Weight updates are proportional to the product of presynaptic activity and the routed postsynaptic drive (error signal modulated by the activation derivative). This avoids weight transport and random feedback matrices, relying instead on a structured correspondence between hidden units and output dimensions.

Domain-Specific Innovations

The paper introduces three specific innovations for the classification setting (MNIST, CIFAR-10) to address failure modes unique to multi-class learning under Dale's constraints:

Layer-Specific Sigmoid Widths: To counteract the vanishing gradient problem inherent in sigmoid-gated error diffusion (which showed a $25\times$ decay from output to the first hidden layer), the authors use wider sigmoid activations ( $\alpha_i$ ) in deeper layers.
Batch-Centered Class Error Signals: For 10-way classification, raw one-vs-all errors can suffer from strong class-wise offsets. The authors subtract the mini-batch mean for each class to ensure zero-mean error signals, preventing uniform suppression or excitation of class channels.
Asymmetric Initialization: Hidden excitatory weights are initialized with a $1.5\times$ scale and inhibitory weights with a $0.5\times$ scale (an expected 3:1 ratio), while the output layer uses symmetric initialization to avoid saturation.

Reinforcement Learning Extension (ED-PPO)

For RL tasks, the dual-stream architecture is integrated with Proximal Policy Optimization (PPO).

The policy and value networks utilize the dual-stream MLP structure.
Unlike the classification setting, the RL extension uses ReLU activations and RMS normalization, omitting the classification-specific innovations (sigmoid widths, batch-centering).
Error signals are routed to hidden units based on output channels (for policy) or broadcast (for the scalar value network).

Key Contributions

Scalable Dale-Compliant Architecture: A dual-stream ED architecture that maintains non-negative weights and achieves competitive performance on both supervised classification (MNIST, CIFAR-10) and continuous-control RL (Brax, Craftax).
Task-Dependent Credit Assignment: An ablation study revealing that the relative importance of architectural components reverses between tasks. On MNIST, layer-specific sigmoid widths are critical; on CIFAR-10, batch-centered error signals are the dominant factor. This exposes that credit-assignment bottlenecks shift qualitatively with task difficulty.
Cross-Domain Evaluation: Demonstration that ED-PPO achieves returns comparable to backpropagation-based PPO and Direct Feedback Alignment (DFA) on locomotion tasks (Ant, Humanoid, HalfCheetah), while outperforming DFA on the complex, open-ended exploration task Craftax.

Results

Classification Performance

MNIST: The proposed ED achieves 96.7% accuracy, comparable to the DFA baseline (97.6%).
CIFAR-10: The proposed ED achieves 61.7% accuracy. While this is lower than DFA (69.1%), it represents a significant improvement over the "seed" ED (11.6%) and previous purely feedforward ED attempts.
Ablation Reversal: Removing layer-specific widths on MNIST causes a catastrophic drop to 25.3%, whereas removing batch-centered error has a negligible effect (-0.3%). Conversely, on CIFAR-10, removing batch-centered error causes a massive drop to 13.8%, while removing widths causes a smaller (-15.1%) but significant drop.

Reinforcement Learning Performance

Locomotion (Brax): ED-PPO performs competitively. On HalfCheetah, it significantly outperforms standard BP-PPO (5494 vs 3520) and matches DFA-PPO. On Ant, it matches both PPO variants. On Humanoid, it trails BP-PPO but remains competitive.
Open-Ended Exploration (Craftax): ED-PPO achieves a score of 20.9, significantly outperforming DFA-PPO (19.8) and approaching BP-PPO (27.0). This suggests that the structured routing of ED may be more robust than random feedback (DFA) in complex, open-ended environments.
Variance: ED-PPO exhibits higher variance than BP-PPO in some environments, suggesting the coarse error routing introduces stochasticity.

Post-Hoc Analysis

Gradient Attenuation: A $25\times$ attenuation of surrogate gradients was observed from the output to the first hidden layer, motivating the use of wider sigmoid widths.
Emergent E/I Balance: Despite an asymmetric 3:1 initialization, training drives hidden layers toward a near-balanced E/I ratio ( $\sim1:1$ ), with deeper layers becoming slightly inhibitory-dominant. This mirrors biological observations of homeostatic mechanisms.
Implicit Sparsity: The non-negative weight floor induces substantial sparsity, with up to 68.8% of inhibitory connections pruned to the floor, suggesting a structural capacity bottleneck.

Significance and Claims

The paper claims that biologically plausible learning rules need not be monolithic; a shared architectural core (dual-stream ED) can support diverse task demands when augmented with appropriate domain-specific components.

Biological Plausibility vs. Performance: The work demonstrates that enforcing Dale's principle does not preclude learning on complex tasks, though it incurs a performance cost (e.g., the accuracy gap on CIFAR-10 and variance in RL).
Task-Dependent Bottlenecks: The reversal of ablation importance between MNIST and CIFAR-10 highlights that single-benchmark evaluations can obscure critical design trade-offs. What stabilizes learning on simple tasks may be insufficient for complex ones.
Emergent Self-Organization: The observation that the network self-organizes toward balanced E/I ratios without explicit homeostatic mechanisms suggests that gradient-driven learning under Dale's constraints can replicate biological self-organization.
Hardware Implications: The constraint of non-negative weights aligns well with emerging neuromorphic substrates (e.g., photonic or analog hardware) where physical elements naturally encode non-negative quantities, with sign determined by population identity.

The authors conclude that while the current accuracy gap to unconstrained backpropagation remains, the dual-stream ED framework provides a viable path toward biologically faithful learning that scales beyond simple benchmarks, offering new insights into how credit assignment functions under biological constraints.