Power-Law Scaling in the Classification Performance of Small-Scale Spiking Neural Networks

Cong Han; Wenjun Xia; Zhengdi Zhang

REVIEW 5 major objections 5 minor 10 references

Power-Law Scaling in the Classification Performance of Small-Scale Spiking Neural Networks

T0 review · 5 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read Classification accuracy in tiny spiking networks follows a power-law decay set by the number of categories, not by the input patterns.

desk verdict A shallow curve fit on 1–3 neuron simulations dressed up as a scaling law; the control is fine but the simulation regime and model selection don't support the claim. read the letter →

arxiv 2601.14961 v2 pith:RSTTIYTC submitted 2026-01-21 q-bio.NC

Zhengdi Zhang , Cong Han , Wenjun Xia This is my paper · ORCID

classification q-bio.NC

keywords spikingneuralnetworksLeakyIntegrate-and-Firepower-lawscalingclassificationaccuracysmall-scaletaskcomplexityAI-assistedscientificdiscoverylinearlogisticregression

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish a quantitative scaling law for how well tiny Leaky Integrate-and-Fire spiking networks classify inputs. It sweeps networks of 1–3 neurons over 2–30 categories and 1–3 stimulus nodes, then fits accuracy to Accuracy = p1 × C^p2 + p3 × S + p4 × N + p5. The fitted exponent p2 ≈ −0.054 makes category count the dominant factor, with stimulus nodes and neuron count contributing only small linear corrections. A control experiment classifying the raw stimulus matrices directly shows no such power law, so the paper attributes the scaling to the network's spiking transformation. The paper also claims that an LLM-assisted workflow was key to proposing and refining this interpretable functional form.

What carries the argument

The load-bearing object is the Leaky Integrate-and-Fire spiking network—each neuron is a capacitor-like integrator that fires when its membrane potential crosses a threshold—combined with the fitted additive power-law formula in Equation 7. The formula does the work: it expresses classification accuracy as a power-law in category count plus linear corrections for stimulus nodes and neuron count. The control experiment, which classifies the binary stimulus matrices without passing them through the network, isolates the spiking transformation as the source of the power-law scaling.

What would settle it

Re-run the identical parameter sweep with explicit connectivity that is well-defined for N = 1 (for example, no E/I proportion and directly hand-wired synapses) under multiple random seeds; if the fitted exponent p2 moves materially away from −0.054 or the additive model's R² falls below about 0.5, the claimed power-law is not a stable consequence of spiking computation. Alternatively, showing the same power-law scaling for direct stimulus classification with an exponent far from zero would falsify the network-bottleneck claim.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the classification accuracy of small LIF spiking networks follows an additive power-law model, Accuracy(C, S, N) = p1 × C^p2 + p3 × S + p4 × N + p5, with p1 ≈ 3.15, p2 ≈ −0.054, p3 ≈ 0.031, p4 ≈ −0.046, and p5 ≈ −2.0, explaining about 76% of the variance. The same stimuli remain highly separable when classified directly (accuracy above 0.85 with negligible scaling), so the paper concludes that the power-law decay is a computational bottleneck imposed by the spiking network rather than a property of the inputs. The claim is that even an ensemble of 1–3 neurons obeys a predictable scaling law in which task complexity dominates.

Load-bearing premise

The fragile premise is that simulations of networks with only 1–3 neurons are treated as genuine cortical-like spiking computation, but the prescribed 80% excitatory / 20% inhibitory split and cell-density-based connectivity cannot be defined for a single neuron, so the fitted accuracy values—and the scaling law built on them—could be simulation artifacts.

Editorial extensions

If this is right

For fixed tiny networks, the number of categories sets a hard, predictable ceiling on attainable classification accuracy through the C^-0.054 term.
Adding stimulus inputs helps accuracy only marginally, while adding neurons in the 1-to-3 range slightly hurts accuracy—both predictions are directly testable in neuromorphic hardware.
The control result implies that observed power-law scaling in these experiments is not an artifact of stimulus statistics, so future studies can attribute such scaling to spiking computation.
The fitted parameters provide a compact way to compare different spiking models or parameter regimes on the same scaling curve.
The LLM-assisted workflow is claimed to be a viable path from black-box feature importance to an interpretable functional law, a methodological claim broader than the specific scaling result.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Inference: The power-law exponent may reflect an information-theoretic limit of the low-dimensional spike-derived features; if so, the exponent should shift with feature dimensionality or temporal binning—a testable extension the paper does not run.
Inference: The additive form is likely an approximation valid only in the N ≤ 3 regime; extrapolating to larger networks might reveal saturation or a crossover to a different scaling, so the formula's stated range should be respected.
Inference: The control experiment suggests that any sufficiently lossy encoding stage—not specifically spiking dynamics—could impose a similar power-law bottleneck; testing a rate-coded or continuous nonlinearity would separate spiking-specific effects from generic compression effects.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Colleague,

Bottom line: this is a curve-fitting exercise on simulations of 1–3 LIF neurons, and the “power-law scaling” is a thin result. The exponent is −0.054, which over C=2..30 is a modest decline, and Eq. 7 even predicts accuracy above 1.0 at C=2. So the central law as stated does not hold up.

That said, the paper has a few good moves. The control experiment is reasonable: direct classification of the stimulus matrices stays above 0.85 and shows no clear power-law, while network-mediated accuracy decays with C. That does suggest the small network imposes some bottleneck, even if the bottleneck is probably just a limited readout space. The authors are also transparent about their fitting process—they report R², RMSE, cross-validation, and compare against linear, polynomial, GA symbolic regression, and random forests. That is more than many similar papers do.

The soft spots are substantial. First, the simulation regime: Section 2.4 specifies an 80/20 E/I split and distance-dependent Gaussian connectivity with a spatial scale, but these are undefined for N=1 and degenerate for N=2. A single neuron cannot be 80% excitatory, and the connectivity kernel needs at least two positions. Without code, seeds, or spike data, we cannot tell whether the outputs are stable or noise-dominated. This is load-bearing because every accuracy number feeds the fit.

Second, model selection. The additive power-law was chosen after the fact on the same dataset. Its R²=0.761 is far below the random forest benchmark (0.984). The exponent is so close to zero that a logarithmic or even linear-in-C fit might do as well, and the paper doesn’t test that. With only three levels each for S and N, the linear coefficients are essentially determined by three points.

Third, the “LLM-assisted discovery” is not really a contribution. The LLM proposed exponential, logarithmic, and power-law candidates—standard function families any regression textbook would list. It is a workflow anecdote, not a validation.

The claim that this scaling is “imposed by the spiking neural network’s computational process” is too strong. A more accurate statement is that accuracy decays with C in these simulations. The mechanism is not established.

Who is this for? People interested in scaling laws for neuromorphic core sizing might read it, but they need code and data before trusting the numbers. As a case study in model selection pitfalls, it is a usable reading-group example.

Recommendation: I would send it to peer review, but with a requirement that the authors release code and data, clarify the N=1 case, constrain predictions to [0,1], and compare against log and linear baselines. As it stands, reject.

Referee Report

5 major / 5 minor

Summary. The paper studies classification accuracy in small-scale LIF spiking neural networks with N=1–3 neurons, S=1–3 stimulus nodes (S≤N), and C=2–30 categories. From 870 simulation samples, it fits an additive model Accuracy = p1·C^p2 + p3·S + p4·N + p5, reporting p2≈−0.054 and R²=0.761. The authors claim this shows a power-law scaling of accuracy with category count that is 'imposed by the spiking neural network's computational process' and not present in the raw stimuli, and they present an LLM-assisted workflow as the key methodological novelty. A control experiment classifies raw stimulus matrices without the network and finds no valid power-law fit.

Significance. If the central claim were valid, the result would be a compact, interpretable scaling law for small spiking networks, potentially useful for neuromorphic design and for understanding resource-constrained neural computation. The paper also attempts a novel LLM-assisted hypothesis-generation workflow, which is of methodological interest. However, the substantive result is undermined by several load-bearing problems: the fitted model predicts accuracies above 1.0 for realistic parameter values; the evidence for a power-law over simpler alternatives is weak; the simulated 'networks' of 1–3 neurons are not consistent with the described cortical architecture; and the control experiment does not demonstrate a qualitative difference between network and stimulus-only scaling. The paper does include a broad model comparison, cross-validation, and a control condition, which are strengths, but these do not compensate for the invalid quantitative model and the ill-defined simulation regime.

major comments (5)

[§3.2.1, Eq. (7)] The fitted constants produce predictions outside the admissible [0,1] accuracy interval. For C=2, S=1, N=1: 3.150233 × 2^(−0.054043) + 0.030683 − 0.046350 − 2.000000 ≈ 1.019; for S=3, N=1 the predicted value is even larger. Since accuracies above 1 are impossible, the model is not a valid quantitative description of classification performance. This boundary violation is not a minor cosmetic issue: the claimed 'scaling law' is the central result, and the model is used as the basis for the power-law interpretation.
[§3.2, §2.6] The evidence for a power-law functional form is weak. The fitted exponent p2 = −0.054 is close to zero; over C=2–30, C^p2 ranges only from 0.963 to 0.832, which is nearly indistinguishable from a logarithmic decline when multiplied by p1. The reported R²=0.761 is only slightly better than the linear regression R²=0.6696 and the genetic-algorithm expression R²=0.7372; no statistical test (e.g., nested-model comparison or information criterion) is given to justify preferring the power-law form. Since Eq. (7) was selected after exploring several candidate families on the same 870 data points, the five-fold cross-validation only establishes that this particular functional family can be re-fit with stable parameters; it does not validate the power-law form against alternatives.
[§2.4, §2.1.2] The described network architecture is mathematically ill-defined for the simulated sizes. An 80% excitatory / 20% inhibitory split cannot be realized with N=1 or N=2 neurons, and a density of ~6000 cells/mm² has no clear meaning when N=1–3. The distance-dependent Gaussian connectivity in Eqs. (4)–(5) requires at least two spatially separated neurons; for N=1 there are no connections at all. Thus the simulations are not actually 'cortical-like spiking networks' for most of the parameter space, and the resulting accuracies reflect a single neuron or a trivial pair, not the network processing that the interpretation invokes. No code, seeds, or spike data are provided, so the stability and biological relevance of these simulations cannot be checked.
[§3.3, Table 3] The control experiment does not support the claim that the power-law is 'imposed' by the spiking network. The direct stimulus classification fits in Table 3 also yield negative exponents (b = −0.038, −0.030, −0.006), which are of the same order of magnitude as the network exponent p2 = −0.054. The authors dismiss these as 'no valid power-law' solely because R²<0.5, but R² is a measure of explained variance, not a test of functional form; the network fit itself has R²=0.761, which is moderate. A proper comparison would require confidence intervals on the exponents and a test of whether the slopes differ significantly between conditions. Without such a test, the assertion that the scaling is absent in stimuli and appears only after network processing is not established.
[§2.3, §3.2] The statistical independence of the 870 data points is questionable. The stimulus matrices for C categories are generated cumulatively: the C=3 set reuses the C=2 categories and adds a third, and so on (Section 2.3). Hence data points with different C are not independent samples. In addition, each reported accuracy is already an average over 10 random train/test splits, but the paper does not clarify how the five 'simulation groups' per configuration are used to create the 870 samples, or whether cross-validation folds respect these groups. This issue affects the validity of the reported R² and parameter standard deviations.

minor comments (5)

[Table 1] Typo: 'V alue' should be 'Value'.
[Figure 3] Subplots show power-law fits for each (S,N) combination, but no error bars or confidence intervals are shown, although the text states 100 independent trials per configuration. Adding error bars would help assess the reliability of the individual fits.
[§2.6] The LLM-assisted procedure is not reproducible: the specific language model, prompts, and stopping criteria are not described. Since one of the paper's aims is to demonstrate a new 'AI-aided discovery' methodology, this is a significant omission for readers who might want to adopt or evaluate the approach.
[§3.3] The labels 'Data Points' in Table 3 are all 145; it would be clearer to report the (S,C) combinations covered and the number of independent stimulus realizations.
[General] The notation alternates between 'ms' and 'milliseconds', and Eq. (1) uses V_stim inside the voltage equation without specifying its units consistently with the stimulus description in Eq. (6). These presentation issues can be fixed without altering the substance.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central result is an explicitly empirical fit, not a derivation that reduces to its inputs.

full rationale

The paper's central claim is an empirical scaling law, explicitly introduced as an 'empirically derived additive power-law model' fitted to simulation outputs. Equation 7 (Accuracy = p1*C^p2 + p3*S + p4*N + p5) is a regression model, not a first-principles derivation; the parameters p1..p5 are fitted, and the reported exponent p2 is an estimated quantity. Saying that a negative fitted exponent 'confirms power-law decay' is a semantic overstatement, but it is not circular: the exponent could in principle have been near zero or positive, and the control experiment provides an external comparison showing that direct stimulus classification does not exhibit the same decay. Cross-validation checks stability of the fit, not the validity of the chosen functional form, but that is a statistical limitation, not a circularity. The 1-3 neuron regime and the 80/20 excitatory/inhibitory prescription are questionable and would undermine reliability of the simulation outputs, but that is a correctness risk rather than a self-referential step. There are no load-bearing self-citations, no uniqueness theorem imported from the authors, and no ansatz smuggled in via citation. The connectivity model is taken from Callegari et al. [4], an external source, and is not used to define the scaling law. Overall, the paper does not reduce to its inputs by construction.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The central claim rests on eight free or hand-chosen parameters (five explicit fit coefficients plus stimulus sparsity, connectivity length scale, and connection probability), multiple domain assumptions imported from prior models, and an ad hoc assumption that the E/I split is meaningful for 1-3 neurons. No new physical entity is introduced.

free parameters (8)

p1 (scaling factor) = 3.150233
Fitted in Eq 7 to the full dataset; controls baseline magnitude of the C term.
p2 (power exponent) = -0.054043
Fitted exponent for number of categories; the central 'power-law' claim rests on this fitted value.
p3 (stimulus-node coefficient) = 0.030683
Fitted linear coefficient for S=1,2,3; only three values support it.
p4 (neuron-count coefficient) = -0.046350
Fitted linear coefficient for N=1,2,3; only three values support it.
p5 (constant offset) = -2.000000
Fitted offset; its near-exact value combined with p1 yields predictions above 1 for low C.
stimulus sparsity range = not stated
Stimulus matrices are filled with a fixed percentage of 1s within a 'predefined range', but the range is never specified; it controls input difficulty.
sigma_c/sigma_w (connectivity length scale) = 1000 µm
Hand-specified; sets the spatial scale of connectivity and synaptic weights (Eqs 4-5), and is not varied to test robustness.
p_max (max connection probability) = 0.25
Hand-specified; caps connection probability and shapes network responses.

assumptions (5)

domain assumption LIF membrane dynamics (Eq 1) with the listed parameter values describes the neurons' computation
Adopted as the base neuron model; not justified against alternatives or biological data in this paper.
domain assumption Connectivity and weights follow the Gaussian distance decay of [4] (Eqs 4-5)
The paper imports this structural model from prior work and assumes it remains valid for networks of 1-3 neurons.
ad hoc to paper An 80% excitatory / 20% inhibitory split is meaningful for N=1,2,3
With N<5 the split cannot be realized exactly; the paper nevertheless assumes cortical-like E/I balance (Sec 2.4).
domain assumption Coordinate compression of spike trains to a fixed-dimensional feature vector preserves class-relevant information for logistic regression
The aggregation/binning is not fully specified (Sec 2.5); if it discards information, the measured accuracies are artifacts of the feature extractor.
ad hoc to paper The final additive power-law form was the best among tested candidates based on R² and interpretability
Model selection is performed on the same data used to fit and report parameters; the chosen form is not derived from theory or tested on an external dataset.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Power-Law Scaling in the Classification Performance of Small-Scale Spiking Neural Networks." pith.science (2026). https://pith.science/paper/RSTTIYTC

@misc{pith2026260114961,
  author       = {Pith},
  title        = {Pith review of: Power-Law Scaling in the Classification Performance of Small-Scale Spiking Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RSTTIYTC}},
  note         = {Machine review of arXiv:2601.14961}
}

read the original abstract

This paper investigates the classification capability of small-scale spiking neural networks based on the Leaky Integrate-and-Fire (LIF) neuron model. We analyze the relationship between classification accuracy and three factors: the number of neurons, the number of stimulus nodes, and the number of classification categories. Notably, we employ a large language model (LLM) to assist in discovering the underlying functional relationships among these variables, and compare its performance against traditional methods such as linear and polynomial fitting. Experimental results show that classification accuracy follows a power-law scaling primarily with the number of categories, while the effects of neuron count and stimulus nodes are relatively minor. A key advantage of the LLM-based approach is its ability to propose plausible functional forms beyond pre-defined equation templates, often leading to more concise or accurate mathematical descriptions of the observed scaling laws. This finding has important implications for understanding efficient computation in biological neural systems and for pioneering new paradigms in AI-aided scientific discovery.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

10 extracted references · 4 linked inside Pith

[1]

Izhikevich, E. M. (2003). Simple model of spiking neurons. IEEE Transactions on neural networks, 14(6), 1569-1572

2003
[2]

M., Naud, R., & Paninski, L

Gerstner, W., Kistler, W. M., Naud, R., & Paninski, L. (2014). Neuronal dynamics: From single neurons to networks and models of cognition. Cambridge University Press

2014
[3]

Maass, W. (1997). Networks of spiking neurons: The third generation of neural network models. Neural networks, 10(9), 1659-1671

1997
[4]

Callegari, F., Brofiga, M., & Massobrio, P. (2023). Modeling the three-dimensional connectivity of in vitro cortical ensembles coupled to Micro-Electrode Arrays. PLOS Computational Biology, 19(2), e1010825

2023
[5]

Pfeiffer, M., & Pfeil, T. (2018). Deep learning with spiking neurons: Opportunities and challenges. Frontiers in neuroscience, 12, 774

2018
[6]

M., Cox, S., White, A

Bran, A. M., Cox, S., White, A. D., & Schwaller, P. (2024). Augmenting large language models with chemistry tools. Nature Machine Intelligence, 6(5), 525-535. 18

2024
[7]

Guo, T., et al. (2024). Towards Large Language Models as Copilots for Theorem Proving in Lean. arXiv preprint arXiv:2406.04538

arXiv 2024
[8]

Bubeck, S., et al. (2023). Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint arXiv:2303.12712

arXiv 2023

Show all 10 references

[9]

Kaplan, J., et al. (2020). Scaling laws for neural language models. arXiv preprint arXiv:2001.08361

2020 arXiv
[10]

Alabdulmohsin, I., et al. (2024). A Law of Scaling for Machine Learning. arXiv preprint arXiv:2406.06278. 19

2024 arXiv

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [1]

Izhikevich, E. M. (2003). Simple model of spiking neurons. IEEE Transactions on neural networks, 14(6), 1569-1572

2003

[2] [2]

M., Naud, R., & Paninski, L

Gerstner, W., Kistler, W. M., Naud, R., & Paninski, L. (2014). Neuronal dynamics: From single neurons to networks and models of cognition. Cambridge University Press

2014

[3] [3]

Maass, W. (1997). Networks of spiking neurons: The third generation of neural network models. Neural networks, 10(9), 1659-1671

1997

[4] [4]

Callegari, F., Brofiga, M., & Massobrio, P. (2023). Modeling the three-dimensional connectivity of in vitro cortical ensembles coupled to Micro-Electrode Arrays. PLOS Computational Biology, 19(2), e1010825

2023

[5] [5]

Pfeiffer, M., & Pfeil, T. (2018). Deep learning with spiking neurons: Opportunities and challenges. Frontiers in neuroscience, 12, 774

2018

[6] [6]

M., Cox, S., White, A

Bran, A. M., Cox, S., White, A. D., & Schwaller, P. (2024). Augmenting large language models with chemistry tools. Nature Machine Intelligence, 6(5), 525-535. 18

2024

[7] [7]

Guo, T., et al. (2024). Towards Large Language Models as Copilots for Theorem Proving in Lean. arXiv preprint arXiv:2406.04538

arXiv 2024

[8] [8]

Bubeck, S., et al. (2023). Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint arXiv:2303.12712

arXiv 2023