pith. sign in

arxiv: 2606.31700 · v1 · pith:3K3K7JTLnew · submitted 2026-06-30 · 💻 cs.LG · cs.NE

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

Yutaro Yamada , Luca Grillotti , Rujikorn Charakorn , Sebastian Risi , David Ha , Robert Tjarko Lange This is my paper

Pith reviewed 2026-07-01 06:03 UTC · model grok-4.3

classification 💻 cs.LG cs.NE
keywords Dale's principleerror diffusiondual-stream networkscredit assignmentbiologically plausible learningMNISTCIFAR-10reinforcement learning
0
0 comments X

The pith

Dual-stream excitatory-inhibitory networks with modulo error routing achieve representation learning under Dale's principle.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper extends Error Diffusion with modulo error routing to handle multi-class problems in a dual-stream architecture that maintains separate excitatory and inhibitory populations. This approach reaches 96.7% accuracy on MNIST and establishes a 61.7% baseline on CIFAR-10 while strictly following Dale's principle. The authors also integrate the method with PPO for reinforcement learning tasks, where it performs competitively with other backpropagation-free methods. These results indicate that credit assignment can be done without weight transport or random feedback in biologically constrained networks. The work reveals task-dependent differences in what engineering choices matter for performance.

Core claim

By introducing modulo error routing, Error Diffusion can be applied to classification beyond binary tasks and to reinforcement learning in a dual-stream excitatory/inhibitory network. This yields 96.7% on MNIST and 61.7% on CIFAR-10, showing that representation learning is possible when strictly enforcing Dale's principle. For classification, layer-specific sigmoid widths, batch-centered class error signals, and asymmetric initialization are used, with their importance varying by task.

What carries the argument

Modulo error routing, which routes global error signals to all layers in the dual-stream excitatory/inhibitory architecture without transporting transposed weights.

If this is right

  • Representation learning is possible in networks that strictly enforce Dale's principle.
  • Ablation analysis shows that the relative importance of the three innovations reverses between MNIST and CIFAR-10.
  • ED integrated with PPO achieves competitive performance on continuous-control tasks and open-ended exploration.
  • Credit assignment bottlenecks are task-dependent and invisible to single-benchmark evaluation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The approach suggests that biological neural circuits could use dual-stream mechanisms for credit assignment.
  • Future work could test the method on larger datasets or more complex RL environments to see scalability.
  • Task-dependent bottlenecks imply that bio-plausible methods may need adaptive components depending on the problem domain.

Load-bearing premise

The reported performance stems primarily from the modulo error routing mechanism rather than from the three domain-specific innovations of layer-specific sigmoid widths, batch-centered error signals, and asymmetric initialization.

What would settle it

Training the dual-stream architecture on MNIST using standard supervised learning or another bio-plausible rule without the modulo error routing and observing whether accuracy falls significantly below 96.7%.

Figures

Figures reproduced from arXiv: 2606.31700 by David Ha, Luca Grillotti, Robert Tjarko Lange, Rujikorn Charakorn, Sebastian Risi, Yutaro Yamada.

Figure 1
Figure 1. Figure 1: Overview of the dual-stream Error Diffusion framework. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Classification accuracy across all six variants on MNIST (left) and CIFAR-10 (right). Error bars: [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Episode return across three RL environments. Error bars indicate [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Cosine similarity between the ED local update direction and the BP gradient direction, restricted to policy trunk [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Episode return on Craftax comparing ED-PPO, [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Post-hoc analyses of the proposed ED model on CIFAR-10. (a) Local surrogate gradient attenuation across layers. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
read the original abstract

Biological neural circuits obey Dale's principle: each neuron's synapses are uniformly excitatory or inhibitory. Artificial networks that respect this constraint must coordinate separate excitatory and inhibitory populations, fundamentally changing how credit is assigned during learning. Several biologically plausible learning rules avoid backpropagation's weight transport requirement, but it has been difficult to achieve strong performance under Dale's principle beyond MNIST. Error Diffusion (ED) was originally proposed in a dual-stream excitatory/inhibitory architecture, where learning is driven by routing global error signals to all layers without transporting transposed forward weights or relying on random feedback matrices. Whether such a rule can scale under Dale's principle across both supervised classification and reinforcement learning remains unknown. Here, we introduce modulo error routing to extend Error Diffusion beyond binary classification, and show that a dual-stream excitatory/inhibitory architecture trained with this method achieves 96.7% on MNIST and establishes a 61.7% baseline on CIFAR-10, demonstrating that representation learning is possible even when strictly enforcing Dale's principle. For the classification setting, we introduce three domain-specific innovations: layer-specific sigmoid widths, batch-centered class error signals, and asymmetric initialization, and ablation analysis reveals that their relative importance reverses between MNIST and CIFAR-10, exposing task-dependent credit-assignment bottlenecks invisible to single-benchmark evaluation. In reinforcement learning, we integrate ED with Proximal Policy Optimization (PPO) and evaluate it on continuous-control tasks in Google Brax and on Craftax, an open-ended exploration task. We show that ED-PPO achieves competitive performance relative to Direct Feedback Alignment, a backpropagation-free baseline.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 3 minor

Summary. The manuscript presents modulo error routing as an extension of Error Diffusion for multi-class settings in dual-stream excitatory/inhibitory networks that enforce Dale's principle. It demonstrates performance of 96.7% on MNIST and 61.7% on CIFAR-10 by incorporating three domain-specific innovations, with ablations highlighting their task-dependent effects, and shows integration with PPO for competitive results on Brax continuous control and Craftax exploration tasks.

Significance. Should the results prove robust and the contribution of the error routing rule be isolated from the additional innovations, this would be a notable contribution to the field of biologically plausible machine learning. It would establish that representation learning is feasible in strictly Dale-compliant networks on standard vision benchmarks and extend to RL, addressing a key limitation of prior work that struggled beyond MNIST. The task-dependent nature of the innovations also offers insight into credit assignment bottlenecks.

major comments (3)
  1. [Abstract] The performance figures (96.7% MNIST, 61.7% CIFAR-10) are given without error bars or training curves, and there is no statement confirming that the three innovations were determined prior to test evaluation, which is necessary to substantiate the soundness of the reported accuracies.
  2. [Abstract (classification setting)] The three domain-specific innovations (layer-specific sigmoid widths, batch-centered class error signals, asymmetric initialization) are presented as necessary for classification, with ablations showing reversed relative importance across MNIST and CIFAR-10; this raises the possibility that these changes, rather than the modulo error routing itself, are responsible for enabling the performance under Dale's principle, requiring additional controls to isolate the core method's contribution.
  3. [Reinforcement learning section] The integration of ED with PPO lacks an ablation study comparable to the classification experiments to isolate the effect of the error diffusion mechanism from the PPO algorithm or the Direct Feedback Alignment baseline.
minor comments (3)
  1. [Abstract] Clarify what 'establishes a 61.7% baseline' means in terms of comparison to other methods.
  2. Define acronyms such as ED and PPO at their first occurrence in the text.
  3. [Methods] Provide the explicit mathematical formulation of 'modulo error routing' to allow reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive comments, which identify key areas for strengthening the presentation and isolating contributions. We address each point below and will revise the manuscript to incorporate the suggested improvements.

read point-by-point responses

  1. Referee: [Abstract] The performance figures (96.7% MNIST, 61.7% CIFAR-10) are given without error bars or training curves, and there is no statement confirming that the three innovations were determined prior to test evaluation, which is necessary to substantiate the soundness of the reported accuracies.

    Authors: We agree that error bars, training curves, and an explicit statement on the protocol for determining innovations are essential for rigor. We will add multiple-run statistics with standard deviations, representative training curves, and a methods note confirming that innovations were finalized on validation data prior to test evaluation. revision: yes

  2. Referee: [Abstract (classification setting)] The three domain-specific innovations (layer-specific sigmoid widths, batch-centered class error signals, asymmetric initialization) are presented as necessary for classification, with ablations showing reversed relative importance across MNIST and CIFAR-10; this raises the possibility that these changes, rather than the modulo error routing itself, are responsible for enabling the performance under Dale's principle, requiring additional controls to isolate the core method's contribution.

    Authors: This concern is well-founded. The existing ablations highlight task dependence but do not fully isolate modulo error routing. We will add controls in the revision, including attempts to apply the core routing rule without the three innovations, to better separate their effects from the routing mechanism. revision: yes

  3. Referee: [Reinforcement learning section] The integration of ED with PPO lacks an ablation study comparable to the classification experiments to isolate the effect of the error diffusion mechanism from the PPO algorithm or the Direct Feedback Alignment baseline.

    Authors: We agree that RL ablations are needed for consistency. We will conduct and include additional experiments comparing ED-PPO to PPO alone and to DFA-augmented variants, to isolate the contribution of the error diffusion rule. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical results on external benchmarks

full rationale

The paper reports measured performance on standard external benchmarks (MNIST at 96.7%, CIFAR-10 at 61.7%, Brax and Craftax RL tasks) rather than any derivation or prediction that reduces to its own inputs by construction. The modulo error routing extension and three domain-specific innovations are presented as engineering choices whose effects are quantified via ablation on held-out data; no self-definitional loop, fitted-input-as-prediction, or load-bearing self-citation chain appears in the abstract or described structure. The work is self-contained against external benchmarks and receives the default non-circularity finding.

Axiom & Free-Parameter Ledger

3 free parameters · 1 axioms · 0 invented entities

The central claim rests on the empirical effectiveness of modulo error routing plus three auxiliary techniques whose necessity is shown only after training; no new physical entities are postulated.

free parameters (3)
  • layer-specific sigmoid widths
    Introduced as a domain-specific innovation whose values are chosen per layer; the abstract does not state whether they are hand-tuned or fitted.
  • batch-centered class error signals
    A preprocessing choice for the error signal whose exact centering parameters are not specified in the abstract.
  • asymmetric initialization
    Weight initialization scheme whose precise asymmetry parameters are not given.
axioms (1)
  • domain assumption Dale's principle must be strictly enforced (each neuron is uniformly excitatory or inhibitory)
    The entire architecture and learning rule are built around this constraint; it is invoked from the first sentence of the abstract.

pith-pipeline@v0.9.1-grok · 5845 in / 1565 out tokens · 34756 ms · 2026-07-01T06:03:13.303521+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 7 canonical work pages · 3 internal anchors

  1. [1]

    E., and Lillicrap, T

    Bartunov, S., Santoro, A., Richards, B., Marris, L., Hinton, G. E., and Lillicrap, T. (2018). Assessing the scalability of biologically-motivated deep learning algorithms and architectures. Advances in neural information processing systems , 31

    2018

  2. [2]

    Bengio, Y., Lee, D.-H., Bornschein, J., Mesnard, T., and Lin, Z. (2015). Towards biologically plausible deep learning. arXiv preprint arXiv:1502.04156

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  3. [3]

    M., and Richards, B

    Cornford, J., Kalajdzievski, D., Leite, M., Lamarquette, A., Kullmann, D. M., and Richards, B. A. (2021). Learning to live with dale's principle: \ ANN \ s with separate excitatory and inhibitory units. In International Conference on Learning Representations

    2021

  4. [4]

    Crick, F. (1989). The recent excitement about neural networks. Nature , 337(6203):129--132

    1989

  5. [5]

    Dale, H. (1935). Pharmacology and nerve-endings

    1935

  6. [6]

    Ding, B., Qian, H., and Zhou, J. (2018). Activation functions and their characteristics in deep neural networks. In 2018 Chinese Control And Decision Conference (CCDC) , pages 1836--1841

    2018

  7. [7]

    Eccles, J. C. (1976). From electrical to chemical transmission in the central nervous system: the closing address of the sir henry dale centennial symposium cambridge, 19 september 1975. Notes and records of the Royal Society of London , 30(2):219--230

    1976

  8. [8]

    J., Guo, Z., Al-Qadasi, M., Marquez, B

    Filipovich, M. J., Guo, Z., Al-Qadasi, M., Marquez, B. A., Morison, H. D., Sorger, V. J., Prucnal, P. R., Shekhar, S., and Shastri, B. J. (2022). Silicon photonic architecture for training deep neural networks with direct feedback alignment. Optica , 9(12):1323--1332

    2022

  9. [9]

    Brax–a differentiable physics engine for large scale rigid body simulation,

    Freeman, C. D., Frey, E., Raichuk, A., Girgin, S., Mordatch, I., and Bachem, O. (2021). Brax--a differentiable physics engine for large scale rigid body simulation. arXiv preprint arXiv:2106.13281

    work page arXiv 2021

  10. [10]

    Fujita, K. (2026). A diagnostic evaluation of neural networks trained with the error diffusion learning algorithm. Discover Artificial Intelligence

    2026

  11. [11]

    Grossberg, S. (1987). Competitive learning: From interactive activation to adaptive resonance. Cognitive science , 11(1):23--63

    1987

  12. [12]

    Hammad, M. (2024). Deep learning activation functions: Fixed-shape, parametric, adaptive, stochastic, miscellaneous, non-standard, ensemble. arXiv preprint arXiv:2407.11090

    work page arXiv 2024

  13. [13]

    Izhikevich, E. M. (2007). Solving the distal reward problem through linkage of stdp and dopamine signaling. Cerebral cortex , 17(10):2443--2452

    2007

  14. [14]

    Kaneko, I. (2000). Sample programs of error diffusion learning algorithm. https://web.archive.org/web/20000306212433/http://village.infoweb.ne.jp/ fwhz9346/ed.htm. [Online; accessed 13-April-2026]

    work page arXiv 2000

  15. [15]

    Launay, J., Poli, I., Boniface, F., and Krzakala, F. (2020). Direct feedback alignment scales to modern deep learning tasks and architectures. In Proceedings of the 34th International Conference on Neural Information Processing Systems , NIPS '20, Red Hook, NY, USA. Curran Associates Inc

    2020

  16. [16]

    Launay, J., Poli, I., and Krzakala, F. (2019). Principled training of neural networks with direct feedback alignment

    2019

  17. [17]

    Lee, D.-H., Zhang, S., Fischer, A., and Bengio, Y. (2015). Difference target propagation. In Appice, A., Rodrigues, P. P., Santos Costa, V., Soares, C., Gama, J., and Jorge, A., editors, Machine Learning and Knowledge Discovery in Databases , pages 498--515, Cham. Springer International Publishing

    2015

  18. [18]

    Liao, Q., Leibo, J., and Poggio, T. (2016). How important is weight symmetry in backpropagation? In Proceedings of the AAAI Conference on Artificial Intelligence , volume 30

    2016

  19. [19]

    P., Cownden, D., Tweed, D

    Lillicrap, T. P., Cownden, D., Tweed, D. B., and Akerman, C. J. (2016). Random synaptic feedback weights support error backpropagation for deep learning. Nature communications , 7(1):13276

    2016

  20. [20]

    Markram, H., Toledo-Rodriguez, M., Wang, Y., Gupta, A., Silberberg, G., and Wu, C. (2004). Interneurons of the neocortical inhibitory system. Nature reviews neuroscience , 5(10):793--807

    2004

  21. [21]

    Matthews, M., Beukman, M., Ellis, B., Samvelyan, M., Jackson, M., Coward, S., and Foerster, J. (2024). Craftax: A lightning-fast benchmark for open-ended reinforcement learning. arXiv preprint arXiv:2402.16801

    work page arXiv 2024

  22. [22]

    N kland, A. (2016). Direct feedback alignment provides learning in deep neural networks. Advances in neural information processing systems , 29

    2016

  23. [23]

    J., and Maass, W

    Pfeiffer, M., Nessler, B., Douglas, R. J., and Maass, W. (2010). Reward-modulated hebbian learning of decision making. Neural computation , 22(6):1399--1444

    2010

  24. [24]

    A., Lillicrap, T

    Richards, B. A., Lillicrap, T. P., Beaudoin, P., Bengio, Y., Bogacz, R., Christensen, A., Clopath, C., Costa, R. P., de Berker, A., Ganguli, S., et al. (2019). A deep learning framework for neuroscience. Nature neuroscience , 22(11):1761--1770

    2019

  25. [25]

    Roodschild, M., Gotay Sardi\ n as, J., and Will, A. (2020). A new approach for the vanishing gradient problem on sigmoid activation. Progress in Artificial Intelligence , 9(4):351–360

    2020

  26. [26]

    Sacramento, J., Ponte Costa, R., Bengio, Y., and Senn, W. (2018). Dendritic cortical microcircuits approximate the backpropagation algorithm. Advances in neural information processing systems , 31

    2018

  27. [27]

    Salimans, T., Ho, J., Chen, X., Sidor, S., and Sutskever, I. (2017). Evolution strategies as a scalable alternative to reinforcement learning. arXiv preprint arXiv:1703.03864

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  28. [28]

    Schulman, J., Wolski, F., Dhariwal, P., Radford, A., and Klimov, O. (2017). Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  29. [29]

    F., Yang, G

    Song, H. F., Yang, G. R., and Wang, X.-J. (2016). Training excitatory-inhibitory recurrent neural networks for cognitive tasks: a simple and flexible framework. PLoS computational biology , 12(2):e1004792

    2016

  30. [30]

    and Harvey, R

    Strata, P. and Harvey, R. (1999). Dale’s principle. Brain research bulletin , 50(5-6):349--350

    1999

  31. [31]

    Szanda a, T. (2020). Review and comparison of commonly used activation functions for deep neural networks. In Bio-inspired neurocomputing , pages 203--224. Springer

    2020

  32. [32]

    Turrigiano, G. (2011). Too many cooks? intrinsic and synaptic homeostatic mechanisms in cortical circuit refinement. Annual review of neuroscience , 34(1):89--103

    2011

  33. [33]

    Whittington, J. C. and Bogacz, R. (2017). An approximation of the error backpropagation algorithm in a predictive coding network with local hebbian synaptic plasticity. Neural computation , 29(5):1229--1262

    2017