pith. sign in

arxiv: 2605.30370 · v3 · pith:SAMNIYVAnew · submitted 2026-05-19 · 💻 cs.NE · cs.AI· cs.CV· cs.LG

Updating the standard neuron model in artificial neural networks

Raul Mohedano , Thomas Batard , Erik Velasco-Salido , Ramsses De Los Santos Mendoza , Jorge H. Mart\'inez , Stacey Levine , Marcelo Bertalm\'io This is my paper

Pith reviewed 2026-06-30 18:07 UTC · model grok-4.3

classification 💻 cs.NE cs.AIcs.CVcs.LG
keywords artificial neural networksneuron modelcortical cellsexpressivityrobustnesslearning speedmemorization
0
0 comments X

The pith

Substituting a recent cortical cell model for the point neuron in ANNs increases expressivity, robustness, and learning speed while cutting memorization and data needs, all without extra parameters.

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

The paper argues that artificial neural networks have relied on the same simplified point neuron model since the 1950s, even though neuroscience has developed more accurate descriptions of cortical cells. By directly replacing the point neuron with one such recent cortical cell model inside standard ANN architectures, the networks gain measurable improvements. These include greater ability to represent complex functions, stronger resistance to input changes, quicker convergence during training, less overfitting to training examples, and effective results from smaller datasets. The substitution introduces no additional parameters and requires no architectural redesign beyond the neuron unit itself.

Core claim

Direct substitution of a recent model of cortical cells for the point neuron model inside artificial neural networks produces networks that are more expressive and robust, learn faster, memorize less of the training set, and reach target performance with less data, while keeping the original parameter count unchanged.

What carries the argument

Direct substitution of the recent cortical cell model into standard ANN layers in place of the point neuron, preserving parameter count and network topology.

If this is right

  • The modified networks represent a wider class of functions than equivalent point-neuron networks at fixed parameter count.
  • Training converges in fewer epochs on the same data.
  • The networks exhibit lower sensitivity to small input perturbations.
  • Overfitting is reduced, visible as lower training-set memorization.
  • Target accuracy is reached with smaller training sets than required by point-neuron baselines.

Where Pith is reading between the lines

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

  • The same substitution could be applied to convolutional or recurrent layers to test whether the gains generalize beyond fully connected networks.
  • If the cortical cell model introduces internal state variables, those variables might be inspected post-training to interpret what features the network has learned.
  • Hardware implementations that natively support the cortical cell dynamics could further reduce energy cost compared with point-neuron accelerators.

Load-bearing premise

The recent cortical cell model can be inserted directly into existing ANN architectures in place of the point neuron while keeping all claimed performance gains and without any increase in parameters.

What would settle it

A controlled experiment in which networks built with the cortical cell model show no measurable gain in robustness or learning speed, or require the same volume of training data as identical point-neuron networks to reach target accuracy.

read the original abstract

From their inception in the 1950s, artificial neural networks (ANNs) started using the so-called point neuron model then prevalent in neuroscience, hoping that this analogy would allow for a better emulation of brain function. Over the years the neuroscience literature has shown that the point neuron model is too simplistic to properly represent many fundamental neural processes; however, the standard neuron model in ANNs still remains the same. Here we substitute it by a very recent model of cortical cells and demonstrate through theoretical analyses and experimental results how, simply by using a more realistic neural unit element without augmenting the number of parameters, the resulting ANNs offer a number of important advantages that include increases in expressivity, robustness and learning speed, and a reduction in memorization and the amount of training data needed.

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

1 major / 0 minor

Summary. The manuscript proposes replacing the standard point neuron model used in ANNs since the 1950s with a recent model of cortical cells. It claims that this direct substitution—without increasing the number of parameters—yields ANNs with higher expressivity, robustness, and learning speed, plus reduced memorization and lower training-data requirements, as shown via theoretical analyses and experiments.

Significance. If the central claim holds, the work would be significant for the field: it would demonstrate that a biologically motivated update to the fundamental computational unit can deliver measurable gains while preserving parameter count, potentially improving data efficiency and generalization in standard architectures. The parameter-neutral substitution, if rigorously verified, would be a notable strength.

major comments (1)
  1. [Abstract / Methods] The central claim requires that the cortical-cell model substitutes into standard layers with exactly the same number of trainable parameters as the point neuron. The manuscript must explicitly demonstrate this (e.g., by comparing the number of free parameters per unit in a fully connected or convolutional layer) to exclude the possibility that observed gains arise from hidden increases in effective capacity rather than the substitution itself.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their thoughtful review and constructive suggestion. We address the single major comment below and will incorporate an explicit parameter-count demonstration in the revised manuscript.

read point-by-point responses

  1. Referee: [Abstract / Methods] The central claim requires that the cortical-cell model substitutes into standard layers with exactly the same number of trainable parameters as the point neuron. The manuscript must explicitly demonstrate this (e.g., by comparing the number of free parameters per unit in a fully connected or convolutional layer) to exclude the possibility that observed gains arise from hidden increases in effective capacity rather than the substitution itself.

    Authors: We agree that an explicit side-by-side accounting of trainable parameters would remove any ambiguity. In the revised manuscript we will insert a short subsection (new Methods 3.2) that tabulates the number of free parameters per unit for both models. For a fully-connected layer the cortical-cell substitution replaces only the point-wise activation with a fixed dynamical system whose internal state variables are not trainable; the weight matrix W and bias vector b remain identical in dimension and count. The same holds for convolutional layers, where the kernel weights are unchanged. Consequently the total number of trainable parameters is exactly the same as in the baseline point-neuron network. We will also add a one-sentence statement in the abstract confirming this invariance. revision: yes

Circularity Check

0 steps flagged

No circularity detected; derivation chain is self-contained

full rationale

The paper substitutes a recent cortical cell model for the point neuron in ANNs and claims benefits in expressivity, robustness, learning speed, and data efficiency without increasing parameter count. The provided text (abstract and context) contains no equations, fitted parameters, predictions, or self-citations that reduce by construction to the inputs. Theoretical analyses and experimental results are invoked as independent support, with the substitution presented as direct replacement. No load-bearing step equates a claimed result to a renamed fit or self-referential definition. This is the common case of an empirical substitution claim without internal circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The paper rests on domain assumptions from neuroscience about the limitations of the point neuron and the suitability of the cortical cell model; no free parameters or invented entities are mentioned in the abstract.

axioms (2)
  • domain assumption The point neuron model is too simplistic to properly represent many fundamental neural processes
    Explicitly stated in the abstract as the motivation for the substitution.
  • domain assumption A recent model of cortical cells can be substituted into ANNs without augmenting the number of parameters
    Central premise required for the claim that advantages are obtained at no extra parameter cost.

pith-pipeline@v0.9.1-grok · 5692 in / 1498 out tokens · 47392 ms · 2026-06-30T18:07:05.616650+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

51 extracted references · 3 canonical work pages · 2 internal anchors

  1. [1]

    McCulloch, W. S. & Pitts, W. A logical calculus of the ideas immanent in nervous activity.The Bulletin of Mathematical Biophysics5, 115–133 (1943)

    1943

  2. [2]

    Hubel, D. H. & Wiesel, T. N. Receptive fields of single neurones in the cat’s striate cortex.The Journal of Physiology148, 574–591 (1959)

    1959

  3. [3]

    The perceptron: A probabilistic model for information storage and organization in the brain

    Rosenblatt, F. The perceptron: A probabilistic model for information storage and organization in the brain. Psychological Review65, 386 (1958)

    1958

  4. [4]

    Haykin, S.Neural networks and learning machines(New York: Prentice Hall, 2009)

    2009

  5. [5]

    & Poirazi, P

    Chavlis, S. & Poirazi, P. Dendrites endow artificial neural networks with accurate, robust and parameter- efficient learning.Nature Communications16, 943 (2025). 21

    2025

  6. [6]

    Bastounis, A., Hansen, A. C. & Vlaˇ ci´ c, V. The mathematics of adversarial attacks in AI – Why deep learning is unstable despite the existence of stable neural networks. Preprint at https://arxiv.org/abs/2109.06098 (2021)

    work page arXiv 2021

  7. [7]

    Sartzetaki, C., Roig, G., Snoek, C. G. M. & Groen, I. International Conference on Learning Rep- resentations (ed.)One hundred neural networks and brains watching videos: Lessons from alignment. (ed.International Conference on Learning Representations)Proceedings of the International Conference on Learning Representations(2025)

    2025

  8. [8]

    & DiCarlo, J

    Kar, K. & DiCarlo, J. J. The quest for an integrated set of neural mechanisms underlying object recognition in primates.Annual Review of Vision Science10, 91–121 (2024)

    2024

  9. [9]

    S.et al.Deep problems with neural network models of human vision.Behavioral and Brain Sciences1–74 (2022)

    Bowers, J. S.et al.Deep problems with neural network models of human vision.Behavioral and Brain Sciences1–74 (2022)

    2022

  10. [10]

    (ed.IEEE/CVF Computer Vision Shifts and Pattern Recognition Organizer Committee)IEEE/CVF Conference on Computer Vision and Pattern Recognition, 9568–9578 (2024)

    Tong, S.et al.IEEE/CVF Computer Vision Shifts and Pattern Recognition Organizer Committee (ed.)Eyes wide shut? Exploring the visual shortcomings of multimodal LLMs. (ed.IEEE/CVF Computer Vision Shifts and Pattern Recognition Organizer Committee)IEEE/CVF Conference on Computer Vision and Pattern Recognition, 9568–9578 (2024)

    2024

  11. [11]

    Wichmann, F. A. & Geirhos, R. Are deep neural networks adequate behavioral models of human visual perception?Annual Review of Vision Science9, 501–524 (2023)

    2023

  12. [12]

    Dendritic spikes expand the range of well tolerated population noise structures.Journal of Neuroscience39, 9173–9184 (2019)

    Poleg-Polsky, A. Dendritic spikes expand the range of well tolerated population noise structures.Journal of Neuroscience39, 9173–9184 (2019)

    2019

  13. [13]

    & Poirazi, P

    Pagkalos, M., Makarov, R. & Poirazi, P. Leveraging dendritic properties to advance machine learning and neuro-inspired computing.Current Opinion in Neurobiology85, 102853 (2024)

    2024

  14. [14]

    & Papoutsi, A

    Poirazi, P. & Papoutsi, A. Illuminating dendritic function with computational models.Nature Reviews Neuroscience21, 303–321 (2020)

    2020

  15. [15]

    Stuart, G. J. & Spruston, N. Dendritic integration: 60 years of progress.Nature Neuroscience18, 1713–1721 (2015)

    2015

  16. [16]

    & Harnett, M

    Francioni, V. & Harnett, M. T. Rethinking single neuron electrical compartmentalization: dendritic contributions to network computation in vivo.Neuroscience489, 185–199 (2022)

    2022

  17. [17]

    & Palmer, L

    Stuyt, G., Godenzini, L. & Palmer, L. M. Local and global dynamics of dendritic activity in the pyramidal neuron.Neuroscience489, 176–184 (2022)

    2022

  18. [18]

    & Both, M

    Stingl, M., Draguhn, A. & Both, M. A dendrite is a dendrite is a dendrite? Dendritic signal integration beyond the “antenna” model.Pfl¨ ugers Archiv-European Journal of Physiology477, 9–16 (2025)

    2025

  19. [19]

    Zador, A.et al.Catalyzing next-generation artificial intelligence through NeuroAI.Nature Communications 14, 1597 (2023)

    2023

  20. [20]

    Frontiers in Neurorobotics16, 846219 (2022)

    Iyer, A.et al.Avoiding catastrophe: Active dendrites enable multi-task learning in dynamic environments. Frontiers in Neurorobotics16, 846219 (2022)

    2022

  21. [21]

    Larkum, M. E. Are dendrites conceptually useful?Neuroscience489, 4–14 (2022)

    2022

  22. [22]

    & Bertalm´ ıo, M

    Rentzeperis, I., Prandi, D. & Bertalm´ ıo, M. A neural model for V1 that incorporates dendritic nonlinearities and back-propagating action potentials.Journal of Neuroscience45(2025)

    2025

  23. [23]

    & Giryes, R

    Jakubovitz, D. & Giryes, R. European Computer Vision Association (ECVA) (ed.)Improving DNN robustness to adversarial attacks using Jacobian regularization. (ed.European Computer Vision Association (ECVA)) European Conference on Computer Vision, 525–541 (2018)

    2018

  24. [24]

    J., Bastounis, A., Woldegeorgis, E

    Tyukin, I., Higham, D. J., Bastounis, A., Woldegeorgis, E. & Gorban, A. N. The feasibility and inevitability of stealth attacks.IMA Journal of Applied Mathematics89, 44–84 (2024). 22

    2024

  25. [25]

    & Vladu, A

    M¸ adry, A., Makelov, A., Schmidt, L., Tsipras, D. & Vladu, A. International Conference on Learning Represen- tations (ed.)Towards deep learning models resistant to adversarial attacks. (ed.International Conference on Learning Representations)Proceedings of the International Conference on Learning Representations(2018)

    2018

  26. [26]

    & Uncini, A

    Pomponi, J., Scardapane, S. & Uncini, A. Pixle: a fast and effective black-box attack based on rearranging pixels. International Joint Conference on Neural Networks (IJCNN 2022)

    2022

  27. [27]

    Neural networks2, 359–366 (1989)

    Hornik, K., Stinchcombe, M., White, H.et al.Multilayer feedforward networks are universal approximators. Neural networks2, 359–366 (1989)

    1989

  28. [28]

    Approximation by superpositions of a sigmoidal function.Mathematics of Control, Signals and Systems2, 303–314 (1989)

    Cybenko, G. Approximation by superpositions of a sigmoidal function.Mathematics of Control, Signals and Systems2, 303–314 (1989)

    1989

  29. [29]

    & Zhang, C

    Feldman, V. & Zhang, C. NIPS Foundation (ed.)What neural networks memorize and why: Discovering the long tail via influence estimation. (ed.NIPS Foundation)Advances in Neural Information Processing Systems, Vol. 33, 2881–2891 (2020)

    2020

  30. [30]

    & Roy, K

    Garg, I., Ravikumar, D. & Roy, K. IMLS (ed.)Memorization through the lens of curvature of loss function around samples. (ed.IMLS)International Conference on Machine Learning, Vol. 235, 15083–15101 (2024)

    2024

  31. [31]

    Bertalm´ ıo, M.et al.Evidence for the intrinsically nonlinear nature of receptive fields in vision.Scientific reports10, 16277 (2020)

    2020

  32. [32]

    M.et al.Selectivity and robustness of sparse coding networks.Journal of Vision20, 1–28 (2020)

    Paiton, D. M.et al.Selectivity and robustness of sparse coding networks.Journal of Vision20, 1–28 (2020)

    2020

  33. [33]

    & Martiniani, S

    Rawat, S., Heeger, D. & Martiniani, S. Unconditional stability of a recurrent neural circuit implementing divisive normalization.Advances in Neural Information Processing Systems37, 14712–14750 (2024)

    2024

  34. [34]

    Bai, S., Kolter, J. Z. & Koltun, V. NIPS Foundation (ed.)Deep equilibrium models. (ed.NIPS Foundation) Advances in Neural Information Processing Systems(2019)

    2019

  35. [35]

    & Fiete, I

    Khona, M. & Fiete, I. R. Attractor and integrator networks in the brain.Nature Reviews Neuroscience23, 744–766 (2022)

    2022

  36. [36]

    & Bertalm´ ıo, M

    Luna, R., Serrano-Pedraza, I. & Bertalm´ ıo, M. Overcoming the limitations of motion sensor models by considering dendritic computations.Scientific reports15, 9213 (2025)

    2025

  37. [37]

    Vaswani, A.et al.Attention is all you need.Advances in neural information processing systems30(2017)

    2017

  38. [38]

    Hopfield, J. J. Neurons with graded response have collective computational properties like those of two-state neurons.Proceedings of the national academy of sciences81, 3088–3092 (1984)

    1984

  39. [39]

    Sch¨ olkopf, B.Causality for machine learning, 765–804 (Association for Computing Machinery, New York, NY, USA, 2022)

    2022

  40. [40]

    & Mackenzie, D.The book of why: the new science of cause and effect(Basic books, 2018)

    Pearl, J. & Mackenzie, D.The book of why: the new science of cause and effect(Basic books, 2018)

    2018

  41. [41]

    Bagul, Y. J. On exponential bounds of hyperbolic cosine.Bulletin of the International Mathematical Virtual Institute8, 365–367 (2018)

    2018

  42. [42]

    Approximation theory of the mlp model in neural networks.Acta Numerica8, 143––195 (1999)

    Pinkus, A. Approximation theory of the mlp model in neural networks.Acta Numerica8, 143––195 (1999)

    1999

  43. [43]

    PyTorch: An Imperative Style, High-Performance Deep Learning Library

    Paszke, A.et al.Pytorch: An imperative style, high-performance deep learning library. Preprint at https: //arxiv.org/abs/1912.01703 (2019)

    work page internal anchor Pith review Pith/arXiv arXiv 1912

  44. [44]

    & Kolter, J

    Geng, Z. & Kolter, J. Z. TorchDEQ: A library for deep equilibrium models (2023). URL https://github.com/ locuslab/torchdeq

    2023

  45. [45]

    Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms

    Xiao, H., Rasul, K. & Vollgraf, R. Fashion-MNIST: a novel image dataset for benchmarking machine learning algorithms. Preprint at https://arxiv.org/abs/1708.07747 (2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  46. [46]

    (ed.NIPS Foundation)NIPS Workshop on Deep Learning and Unsupervised Feature Learning, Vol

    Netzer, Y.et al.NIPS Foundation (ed.)Reading digits in natural images with unsupervised feature learning. (ed.NIPS Foundation)NIPS Workshop on Deep Learning and Unsupervised Feature Learning, Vol. 2011, 7 (2011). 23

    2011

  47. [47]

    & Hinton, G

    Krizhevsky, A. & Hinton, G. Learning multiple layers of features from tiny images. Tech. Rep., University of Toronto (2009). URL https://www.cs.toronto.edu/ ∼kriz/learning-features-2009-TR.pdf

    2009

  48. [48]

    https://github.com/pytorch/vision (2016)

    TorchVision: PyTorch’s computer vision library. https://github.com/pytorch/vision (2016)

    2016

  49. [49]

    Torchattacks: A PyTorch repository for adversarial attacks

    Kim, H. Torchattacks: A PyTorch repository for adversarial attacks. Preprint at https://arxiv.org/abs/2010. 01950 (2021)

    2010

  50. [50]

    & Roli, F

    Biggio, B. & Roli, F. Wild patterns: Ten years after the rise of adversarial machine learning.Pattern Recognition84, 317–331 (2018)

    2018

  51. [51]

    Papernot, N.et al.ACM Special Interest Group on Security, Audit and Control (ed.)Practical black-box attacks against machine learning. (ed.ACM Special Interest Group on Security, Audit and Control)Proceed- ings of the ACM on Asia Conference on Computer and Communications Security, 506–519 (Association for Computing Machinery, New York, NY, USA, 2017). 24 ...

    2017