diffusing-blame-dale-principle-credit-assignment

name

diffusing-blame-dale-principle-credit-assignment

description

Error Diffusion (ED) methodology for biologically plausible credit assignment under Dale's principle. Dual-stream excitatory/inhibitory architecture with modulo error routing. Achieves 96.7% MNIST and 61.7% CIFAR-10 under strict Dale's constraint. Integrates with PPO for RL. Trigger words: Dale's principle, error diffusion, excitatory-inhibitory, biologically plausible learning, credit assignment, dual-stream network.

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

arXiv: 2606.31700v1 | Date: 2026-06-30 | Venue: ALIFE2026
Authors: Yutaro Yamada, Luca Grillotti, Rujikorn Charakorn, Sebastian Risi, David Ha, Robert Tjarko Lange

Core Methodology

Problem Statement

Biological neural circuits obey Dale's principle: each neuron's synapses are uniformly excitatory or inhibitory. Artificial networks respecting this constraint must coordinate separate E/I populations, fundamentally changing credit assignment. Previous biologically plausible rules struggle to scale beyond MNIST under strict Dale's constraint.

Error Diffusion (ED) Framework

• Dual-stream architecture: Separate excitatory and inhibitory populations
• Global error routing: Error signals routed to all layers without:
  • Transporting transposed forward weights (no weight transport)
  • Random feedback matrices
• Modulo error routing: Extension enabling multi-class classification (beyond binary)

Three Domain-Specific Innovations

1. Layer-specific sigmoid widths — adaptive gain per layer
2. Batch-centered class error signals — normalized error propagation
3. Asymmetric initialization — E/I balance-aware weight init

Key Results

Benchmark	Accuracy	Notes
MNIST	96.7%	State-of-the-art under strict Dale's principle
CIFAR-10	61.7%	First strong baseline under Dale's constraint
Brax (RL)	Competitive	ED-PPO vs Direct Feedback Alignment
Craftax	Competitive	Open-ended exploration task

Critical Findings

Task-Dependent Credit Assignment

• Ablation analysis reveals reversal of innovation importance between MNIST and CIFAR-10
• Exposes task-dependent bottlenecks invisible to single-benchmark evaluation
• Implication: Multi-benchmark evaluation essential for biologically plausible learning research

RL Integration

• ED + PPO achieves competitive performance
• Demonstrates scalability to continuous control
• Benchmark: Direct Feedback Alignment (backprop-free baseline)

Pitfalls & Considerations

1. CIFAR-10 gap: 61.7% is baseline, not SOTA — representation learning possible but limited
2. Dale's constraint cost: Significant performance drop vs unconstrained networks
3. Innovation interaction effects: Layer-specific gains + asymmetric init interact non-trivially
4. RL scaling: Craftax results preliminary; full RL benchmark suite needed

Related Work

• Direct Feedback Alignment (Lillicrap et al.)
• Target Propagation
• Feedback Alignment family
• Predictive Coding networks

Activation Keywords

Dale's principle, error diffusion, excitatory-inhibitory, dual-stream, biologically plausible learning, credit assignment, modulo error routing, PPO integration