Pith. sign in

REVIEW 3 major objections 4 minor 29 references

Two local synaptic rules combine to compute an exact SIGReg gradient with no backpropagation or labels.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 08:47 UTC pith:PTOLWLQS

load-bearing objection A clean, checkable rate-level identity — STDP+ plus homeostatic plasticity equals a SIGReg gradient — but the title overclaims 'STDP' for what is a rate-coded construction. the 3 major comments →

arxiv 2607.21622 v1 pith:PTOLWLQS submitted 2026-07-03 cs.NE cs.AIcs.CVcs.LG

Local Synaptic Rules Can Implement a SIGReg Gradient Without Backpropagation

Martin Andrews This is my paper
classification cs.NE cs.AIcs.CVcs.LG
keywords STDPhomeostatic plasticitySIGRegself-supervised learninglocal learning rulestemporal contiguitygradient-free learningwhitening
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper proves that the potentiation arm of spike-timing-dependent plasticity (STDP+) plus a homeostatic 'flashlight' plasticity rule together compute the exact negative gradient of a SIGReg-style self-supervised objective. Because each rule only uses local signals—presynaptic firing, postsynaptic propensity, and temporal contiguity—the result is a fully local, label-free learning procedure that needs no backpropagation, global error signal, or weight transport. A synthetic clustering task shows temporal ordering alone creates class structure, and a temporally ordered MNIST experiment reaches 87.3% linear-probe accuracy. The claim is a unification: two well-known biological rules, previously seen as distinct, are shown to be the gradient of one objective.

Core claim

The paper's central claim is Eq. (17): summing the STDP+ update with a homeostatic 'flashlight' update equals exact gradient descent on LSIGReg = Lpred + λLWeak. STDP+ is exactly the gradient of a forward-prediction loss Lpred = −E[ht·sg(ht+1)], and the flashlight homeostasis, implemented through a fixed random projection A, is exactly the gradient of the weak-SIGReg covariance penalty LWeak. Thus two canonical local plasticity rules, each independently observed in neuroscience, together implement a complete self-supervised representation-learning objective without any global or label information.

What carries the argument

The key machinery is the three-factor Hebbian update form ΔWij ∝ αi·δi·xj, where αi is the local gain hi(1−hi), xj is presynaptic activity, and δi is a modulatory signal: for STDP+ this is the next-timestep propensity hi(t+1), for homeostasis it is a retrograde variance/covariance signal from 'flashlight' granule-like neurons. This form makes each rule the exact gradient of one term of the SIGReg objective, so their sum is the full gradient.

Load-bearing premise

The entire gradient identity rests on the propensity model where neuron activity is a sigmoidal firing probability and STDP+ updates equal the expected product E[αi(t)xj(t)hi(t+1)]; if real spike-time STDP behaves differently, the exact gradient claim holds only in this rate formulation.

What would settle it

A direct spike-based simulation where the STDP+ rule is applied to actual spiking events (rather than the rate expectation) and compared with a backpropagation-computed gradient of Lpred would show a nonzero discrepancy whenever the expected-product approximation fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A fully local, label-free learning rule can in principle perform self-supervised representation learning, providing a candidate mechanism for unsupervised learning in the brain.
  • Temporal contiguity of sensory streams alone suffices to induce class structure, as shown by the synthetic clustering separation (CSR 2.49 vs 0.83).
  • The homeostatic term prevents dimensional collapse without any contrastive negative pairs or memory bank.
  • The equivalence offers a building block for constructing larger biologically plausible learning systems from known primitives.
  • Adding a third layer degrades performance, highlighting a depth-scaling limitation that convolutional or scale-specific variants may address.

Where Pith is reading between the lines

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

  • The proof is algebraic and does not depend on the specific random projection A, so any fixed random projection-based whitening objective could in principle be realized by the same homeostatic mechanism, not just weak-SIGReg.
  • The failure of STDP− in experiments, and the proposed but untested inhibitory interneuron population, suggests a concrete next step: stable implementation of projected gradient descent over the nonnegativity constraint may require adding a separate inhibitory circuit, which could be tested directly.
  • If STDP in real spiking networks does not reduce to the rate-coded expected-product form used here, the exact gradient identity may only hold for rate-coded neurons, not for spike-based implementations of the same rules.
  • The paper's depth-scaling limitation suggests a natural extension: applying the same STDP+homeostasis rules at different spatial scales (as in convolutional architectures) should preserve the whitening cascade, which could be tested empirically.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper claims to prove that two local synaptic rules—the potentiation arm of STDP (STDP+) and homeostatic plasticity mediated by fixed random-projection 'flashlight' neurons—jointly implement exact gradient descent on a SIGReg-like objective LSIGReg = Lpred + λLWeak, using only locally available quantities (Section 3, Eq. (17)). The derivation is algebraic: Lpred is defined so that its negative gradient equals the STDP+ update (Eqs. (2), (4)-(5)), while LWeak is decomposed into diagonal variance and off-diagonal covariance terms whose gradients are mapped onto retrograde and lateral-inhibition signals (Eqs. (9)-(16)). Nine identities are verified numerically against JAX autodiff (Appendix A, Table 2). Experiments on a synthetic clustering task show that class-contiguous temporal ordering yields CSR 2.49 versus 0.83 for random ordering, and on temporally ordered MNIST a two-layer network reaches 87.3% linear-probe accuracy without labels or backpropagation.

Significance. If the central claim holds, the paper makes a conceptually valuable connection between self-supervised learning objectives and local synaptic plasticity: it shows a closed-form loss whose gradient is exactly a three-factor Hebbian rule plus a covariance-regularizing homeostatic signal, with no weight transport, global error, or labels. The paper ships reproducible code and machine-checked verification of the algebraic identities (Table 2), which is a real strength. The significance is tempered, however, by the fact that the 'STDP+' arm is a rate-coded propensity update, not the canonical spike-pair STDP of Bi & Poo, and by the fact that the biological pathways proposed for the homeostatic signals are plausible but unvalidated. These issues do not invalidate the mathematics, but they materially restrict the biological claim as currently worded.

major comments (3)
  1. [§2.1, Eq. (2); §5] The central identity in Eq. (17) rests entirely on the definition of STDP+ as ΔW+_ij = E[α_i(t)x_j(t)h_i(t+1)]. This is a rate-coded expected-product update, not the canonical spike-pair STDP rule of Bi & Poo that the abstract and introduction invoke. In actual spike-pair STDP the weight change depends on precise inter-spike intervals, and the synapse does not directly observe the postsynaptic firing probability h_i or its slope α_i. The paper itself concedes in §5 that the critique of spike-pair STDP is avoided by switching to 'the rate-coded propensity formulation... closer to the BCM rule.' Thus the advertised claim that canonical STDP implements a SIGReg gradient is not established; what is established is the gradient identity for the bespoke rate rule in Eq. (2). This is load-bearing, because the STDP+ arm of Eq. (17) is exactly Eq. (2). The authors should either provide a derivatio
  2. [§3.2, Eqs. (14)–(16); §1] The homeostatic arm is mathematically correct as a gradient of LWeak, but its biological implementation is proposed rather than established. Eq. (14) requires a retrograde messenger released in proportion to each flashlight neuron's variance mismatch (V_m−1), and Eq. (15) requires lateral inhibition from interneurons proportional to pairwise covariances C_mm′. No experimental evidence is cited that endocannabinoid/nitric-oxide signals encode such variance mismatches, or that basket/stellate cells compute covariance-like quantities. Moreover, the resulting δ_i = ∂LWeak/∂h_i is a sum over all M flashlight neurons (weighted by a_mi), so the signal at each encoder synapse is a cell-wide somatic quantity, not a synaptically local product of three factors. The paper should state these as explicit biological assumptions and give falsifiable predictions (e.g., what time course or sign of retrogr
  3. [§4.2, Table 1] The MNIST experiment provides only weak evidence that temporal contiguity is the driver: condition A (ordered) reaches 87.3% probe accuracy while condition C (random order) reaches 86.5%, a difference of 0.8 pp. The paper reports this is consistent across seeds but gives no paired confidence interval or effect size. This is not a fatal problem for the algebraic result, but the abstract's emphasis on 'temporally ordered MNIST' and the end-to-end claim should be calibrated: the synthetic task is the main evidence for the ordering effect, while the MNIST result largely demonstrates a whitening floor. I recommend reporting paired statistics or a permutation test, and adjusting the wording of the contribution accordingly.
minor comments (4)
  1. [§3.2, Eqs. (14)–(15)] The symbol N appears in Eqs. (14)–(15) as the batch size but is never defined in the text. Please define it explicitly when the expectation is introduced.
  2. [§1 and §4.2] The introduction says the rules are 'strictly online and single-pass: one forward activation, one local weight update, per input,' but §4.2 describes minibatches shuffled before each weight update. Clarify whether the algorithm is online or minibatch-based; the current wording is inconsistent.
  3. [§5, Related Work] The discussion of [16] dismisses spike-pair STDP and then asserts the rate formulation is 'closer to the BCM rule.' This admission should be made in the introduction or abstract, not only in the related-work section, since it directly affects the central biological claim.
  4. [Appendix A, Table 2] The nine verified identities are all confirmed on randomly initialized tensors. This is a useful check, but the paper should note that Table 2 validates the algebra, not the biological mechanisms; the current phrasing in the main text occasionally blurs this distinction.

Circularity Check

1 steps flagged

Central STDP+ gradient identity is true by construction: Lpred is defined as the anti-derivative of Eq. (2), so half of Eq. (17) restates the definition.

specific steps
  1. self definitional [Section 2.1, Eqs. (2)-(5), and Eq. (7)]
    "STDP+ (potentiation / LTP). When presynaptic neuron j fires at time t and postsynaptic neuron i fires shortly after [6], the synapse is strengthened: ∆W + ij = +E[α i(t)x j(t)h i(t+1) ].(2) ... Forward prediction loss. The potentiation arm corresponds to gradient descent on a forward prediction loss that treats h t+1 as a fixed teaching signal — a stop-gradient, writtensg(·): Lpred(ht, ht+1) =−E[h t ·sg(h t+1) ].(4) Differentiating throughh t only ... ∂Lpred ∂Wij =−E[α i(t)x j(t)h i(t+1) ],(5) so∆W + ij =−∂L pred/∂Wij exactly:STDP + is gradient descent onL pred."

    Eq. (4) is introduced after Eq. (2) and chosen so that its derivative is exactly the negative of the STDP+ rule. Lpred has no independent definition outside this requirement; it is defined as the anti-derivative of the pre-specified update. Therefore the identity ΔW+ = −∂Lpred/∂W is true by construction. Since LSIGReg (Eq. 7) includes Lpred as its temporal-invariance term, the STDP+ half of the central Eq. (17) restates the definition of the loss rather than establishing an independently specified objective's gradient.

full rationale

The paper is unusually transparent: it states that the equivalence is 'established by direct construction rather than as a deep mathematical result.' The homeostatic arm is derived from the externally motivated SIGReg covariance objective and is genuinely non-circular: LWeak is defined independently, its gradient is computed, and biological signals are proposed to carry it. The numerical verifications against JAX autodiff are real and confirm the algebra. The circularity is confined to the STDP+ arm: Lpred is defined after, and precisely to match, the STDP+ update Eq. (2), so the claim that STDP+ 'implements' the gradient of Lpred is an anti-derivative construction, not an empirical or independently derived prediction. Because this constructed Lpred is then absorbed into LSIGReg, part of the central Eq. (17) is definitional. There is no load-bearing self-citation, no imported uniqueness theorem, and no fitted parameter renamed as a prediction. The result is substantial for the homeostatic component but partially circular for the STDP+ component, warranting a score of 6 rather than higher.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 1 invented entities

The central derivation assumes a rate-coded, differentiable neuron model and fixed random projections; the homeostatic signal requires biological pathways (retrograde messenger and lateral inhibition) whose magnitudes are not verified empirically. One unstated hyperparameter λ and learning rates underlie the experiments.

free parameters (2)
  • λ (weight of LWeak in LSIGReg) = not reported
    Scales the homeostatic term in Eq. (7) and the combined update Eq. (17); its value is not stated in the text, so the experimental results depend on an unstated tuning choice.
  • STDP+ and homeostatic learning rates / update step sizes = not reported
    Training results (Table 1) depend on how the algebraic updates are scaled in the optimizer; not specified in the paper.
axioms (4)
  • domain assumption Neuronal activity is represented by sigmoid propensities h = σ(Wx+b), and STDP updates are expected products of rate variables (Eqs. 1-2).
    This is a modeling choice that makes the update differentiable and locally available; it is not derived from spiking biophysics.
  • domain assumption The random projection matrix A is fixed and the flashlight (granule) afferents are non-plastic.
    The gradient of LWeak with respect to W assumes A constant; if A were plastic, the exact equivalence would fail.
  • domain assumption Biology can supply a retrograde signal proportional to (V_m − 1)f̂_m and lateral inhibition proportional to covariance terms (Eqs. 14-15).
    The existence and proportionality of these biological currents are asserted as possible pathways, not demonstrated.
  • standard math Covariance statistics pooled over a minibatch are unbiased estimates of true expectations; updates are online SGD on the population loss.
    The proof is for expectations; experiments use minibatch estimates.
invented entities (1)
  • Flashlight neurons (fixed random-projection granule-cell-like population) independent evidence
    purpose: Implement the fixed random projection A and emit retrograde homeostatic signals for the variance term of LWeak.
    Cerebellar granule cells with fixed mossy-fiber/Golgi inputs are known from Marr/Albus; however, the specific retrograde variance-mismatch signaling role is a hypothesis proposed by this paper.

pith-pipeline@v1.3.0-alltime-deepseek · 9468 in / 16954 out tokens · 186981 ms · 2026-08-02T08:47:56.905050+00:00 · methodology

0 comments
read the original abstract

We prove that two canonical local synaptic learning rules, the potentiation arm of spike-timing-dependent plasticity (STDP$^+$) and homeostatic plasticity (instantiated here via flashlight granule-cell-like neurons), together can implement the exact gradient of a SIGReg-like self-supervised learning objective. The equivalence requires no gradient calculations, no global error signals, no weight transport, and no label information: the only inputs are pre- and post-synaptic firing rates, local firing statistics, and the temporal contiguity of natural sensory streams. On a synthetic clustering task designed to probe whether class structure can be recovered from temporal ordering of inputs alone, ordered presentation raised cluster separation (CSR) to 2.49 while random ordering left it near baseline (0.83), a roughly threefold ($\approx 3.5\sigma$) separation attributable solely to input ordering. On temporally ordered MNIST, a two-layer network trained entirely with these rules achieved 87.3% linear-probe accuracy, showing that the mechanism functions end-to-end.

Figures

Figures reproduced from arXiv: 2607.21622 by Martin Andrews.

Figure 1
Figure 1. Figure 1: Circuit realising the SIGReg gradient with local rules. The [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

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

29 extracted references · 2 canonical work pages

  1. [1]

    Aizenbud, D

    I. Aizenbud, D. Beniaguev, N. Pnueli, I. Segev, and M. London. What can a neuron compute.bioRxiv,

  2. [2]

    H. Akbar. Weak-SIGReg: Covariance regularization for stable deep learning, 2026. URL https: //arxiv.org/abs/2603.05924

    arXiv 2026

  3. [3]

    J. S. Albus. A theory of cerebellar function.Mathematical Biosciences, 10(1):25–61, 1971. ISSN 0025-

    1971

  4. [4]

    Balestriero and Y

    R. Balestriero and Y . LeCun. LeJEPA: Provable and scalable self-supervised learning without the heuristics,

  5. [5]

    Bardes, J

    A. Bardes, J. Ponce, and Y . LeCun. VICReg: Variance-invariance-covariance regularization for self- supervised learning. InInternational Conference on Learning Representations (ICLR), 2022. URL https://arxiv.org/abs/2105.04906

    Pith/arXiv arXiv 2022

  6. [6]

    Bi and M.-m

    G.-q. Bi and M.-m. Poo. Synaptic modifications in cultured hippocampal neurons: Dependence on spike timing, synaptic strength, and postsynaptic cell type.Journal of Neuroscience, 18(24):10464–10472, 1998. 8 ISSN 0270-6474. doi: 10 .1523/JNEUROSCI.18-24-10464 .1998. URL https://www.jneurosci.org/ content/18/24/10464

    1998

  7. [7]

    Bradbury, R

    J. Bradbury, R. Frostig, P. Hawkins, M. J. Johnson, Y . Katariya, C. Leary, D. Maclaurin, G. Necula, A. Paszke, J. VanderPlas, S. Wanderman-Milne, and Q. Zhang. JAX: composable transformations of Python+NumPy programs, 2018. URLhttp://github.com/jax-ml/jax

    2018

  8. [8]

    Buesing, J

    L. Buesing, J. Bill, B. Nessler, and W. Maass. Neural dynamics as sampling: A model for stochastic computation in recurrent networks of spiking neurons.PLOS Computational Biology, 7(11):1–22, 11 2011. doi: 10.1371/journal.pcbi.1002211. URLhttps://doi.org/10.1371/journal.pcbi.1002211

    doi:10.1371/journal.pcbi.1002211 2011

  9. [9]

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton. A simple framework for contrastive learning of visual representations. InInternational Conference on Machine Learning (ICML), 2020. URL https: //arxiv.org/abs/2002.05709

    Pith/arXiv arXiv 2020

  10. [10]

    Coates and A

    A. Coates and A. Y . Ng. The importance of encoding versus training with sparse coding and vector quantization. InProceedings of the 28th International Conference on International Conference on Machine Learning, ICML’11, page 921–928, Madison, WI, USA, 2011. Omnipress. ISBN 9781450306195

    2011

  11. [11]

    D. J. Felleman and D. C. Van Essen. Distributed hierarchical processing in the primate cerebral cortex. Cerebral Cortex, 1(1):1–47, 1991. doi: 10.1093/cercor/1.1.1

    doi:10.1093/cercor/1.1.1 1991

  12. [12]

    Grill, F

    J.-B. Grill, F. Strub, F. Altché, C. Tallec, P. H. Richemond, E. Buchatskaya, C. Doersch, B. A. Pires, Z. D. Guo, M. G. Azar, B. Piot, K. Kavukcuoglu, R. Munos, and M. Valko. Bootstrap your own latent a new approach to self-supervised learning. InProceedings of the 34th International Conference on Neural Information Processing Systems, NIPS ’20, Red Hook,...

    Pith/arXiv arXiv 2020

  13. [13]

    G. Hinton. The forward-forward algorithm: some preliminary investigations, 2022. URL https:// arxiv.org/abs/2212.13345

    Pith/arXiv arXiv 2022

  14. [14]

    Kavukcuoglu, M

    K. Kavukcuoglu, M. Ranzato, and Y . LeCun. Fast inference in sparse coding algorithms with applications to object recognition, 2010. URLhttps://arxiv.org/abs/1010.3467

    Pith/arXiv arXiv 2010

  15. [15]

    D. Marr. A theory of cerebellar cortex.The Journal of Physiology, 202(2):437–470, 1969. doi: https: //doi.org/10.1113/jphysiol.1969.sp008820. URL https://physoc.onlinelibrary.wiley.com/doi/ abs/10.1113/jphysiol.1969.sp008820

    work page doi:10.1113/jphysiol.1969.sp008820 1969

  16. [16]

    R. C. O’Reilly. This is how the neocortex learns, 2026. URLhttps://arxiv.org/abs/2606.08720

    Pith/arXiv arXiv 2026

  17. [17]

    Scellier and Y

    B. Scellier and Y . Bengio. Equilibrium propagation: bridging the gap between energy-based models and backpropagation.Frontiers in Computational Neuroscience, 11:24, 2017. URL https://www.frontiersin.org/journals/computational-neuroscience/articles/10.3389/ fncom.2017.00024

    arXiv 2017

  18. [18]

    Schmidhuber

    J. Schmidhuber. Learning complex, extended sequences using the principle of history compression. Neural Computation, 4(2):234–242, 03 1992. ISSN 0899-7667. doi: 10 .1162/neco.1992.4.2.234. URL https://doi.org/10.1162/neco.1992.4.2.234

    doi:10.1162/neco.1992.4.2.234 1992

  19. [19]

    G. G. Turrigiano. The self-tuning neuron: synaptic scaling of excitatory synapses.Cell, 135(3):422–435,

  20. [20]

    J. C. R. Whittington and R. Bogacz. An approximation of the error backpropagation algorithm in a predictive coding network with local Hebbian synaptic plasticity.Neural Computation, 29(5):1229– 1262, 05 2017. ISSN 0899-7667. doi: 10 .1162/NECO_a_00949. URL https://doi.org/10.1162/ NECO_a_00949. 9 A Claim statements and verification details Each algebraic ...

    2017

  21. [25]

    2.L sim =L var +L temporal (loss identity)

    STDP + weight update=−∂L pred/∂W. 2.L sim =L var +L temporal (loss identity). 3.∂L sim/∂W=∂L var/∂W+∂L temporal/∂W(gradient identity). 4.L var-hom +L lateral =L Weak (loss identity: diagonal and off-diagonal split of ∥Cov(f)−I∥ 2 F ). 5.∂L var-hom/∂f+∂L lateral/∂f=∂L Weak/∂f(gradient identity)

  22. [26]

    (14) equals∂L var-hom/∂h(exact, no missing factor)

    The retrograde variance signal of Eq. (14) equals∂L var-hom/∂h(exact, no missing factor)

  23. [27]

    The variance three-factor Hebbian update=−∂L var-hom/∂W

  24. [28]

    The lateral-inhibition Hebbian update=−∂L lateral/∂W

  25. [29]

    Table 2 reports the maximum absolute error of each identity

    The full homeostatic weight update=−∂L Weak/∂W. Table 2 reports the maximum absolute error of each identity. Table 2: Nine equivalence claims verified numerically by comparing analytical (biological) expres- sions to jax.grad outputs on randomly initialised weights and activations (float32). Max absolute error is over all entries of the relevant tensor. E...

  26. [2008]

    URLhttps://doi.org/10.1016/j.cell.2008.10.008

    work page doi:10.1016/j.cell.2008.10.008 2008

  27. [2025]

    URLhttps://arxiv.org/abs/2511.08544

    Pith/arXiv arXiv

  28. [2026]

    URL https://www.biorxiv.org/content/early/2026/ 06/09/2026.06.08.730984

    doi: 10 .64898/2026.06.08.730984. URL https://www.biorxiv.org/content/early/2026/ 06/09/2026.06.08.730984

    2026

  29. [5564]

    URL https://www.sciencedirect.com/ science/article/pii/0025556471900514

    doi: https://doi .org/10.1016/0025-5564(71)90051-4. URL https://www.sciencedirect.com/ science/article/pii/0025556471900514

    arXiv