The Computational Value of Sensory-Aligned Receptive Fields Depends on Neuronal Expressivity
Agnese Adorante, Aaron Spieler, Anna Levina
cs.NE · cs.LG · q-bio.NC
Quick verdict
Sensory-aligned receptive fields boost ELM net accuracy, but the benefit fades as neurons grow more expressive; well-controlled, minor overclaims.
Summary
This paper asks whether receptive fields aligned with task-relevant sensory coordinates confer a computational advantage beyond resource efficiency, and whether that advantage persists as individual neurons become more expressive. Using recurrent networks of Expressive Leaky Memory (ELM) neurons, whose internal memory-unit count M tunes single-neuron expressivity independently of connectivity, the authors compare feed-forward wiring restricted along a sensory coordinate (frequency for Spiking Heidelberg Digits, motion direction for DVS-Gesture, retinotopic patches for CIFAR10-DVS) against budget-matched random wiring. Structured wiring improves test accuracy (e.g., +7.26 points on DVS-Gesture at N=128; +2.49 points on SHD at N=110; two-sided Welch tests with Benjamini–Hochberg correction, 10 seeds per condition). The advantage vanishes when the sensory coordinate is scrambled or when wiring follows a task-irrelevant coordinate, and it shrinks as M grows, converging on SHD at M=8–10 while a smaller but significant gap persists on DVS-Gesture at M=21. Full-input networks, despite more synapses, underperform structured wiring and overfit; ℓ1 regularization partially recovers accuracy and induces coarse frequency selectivity, although the remaining gap to structured wiring is not statistically significant (q=0.092). Main limitations: conclusions rest on one neuron model family, three datasets, hand-crafted motion preprocessing, and modest absolute CIFAR10-DVS accuracy (~40%).
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Feed-forward receptive fields restricted along the task-relevant sensory coordinate improve test accuracy over budget-matched random wiring in ELM recurrent networks (e.g., +7.26 points on DVS-Gesture at N=128, q<0.001).
- The advantage disappears when the sensory coordinate is scrambled or wiring follows a task-irrelevant coordinate, so the benefit comes from alignment with task geometry rather than restricted connectivity alone.
- Increasing single-neuron expressivity (ELM memory units M) reduces the structured-input advantage, with convergence on SHD at M=8–10 and a residual significant gap on DVS-Gesture at M=21.
- Generic ℓ1 sparsity regularization induces coarse frequency-selective input fields and partially recovers performance, implying structured receptive fields provide a computational prior beyond sparsity alone.
- Structured recurrent topologies (strict/flexible modular, small-world) add no benefit in the absence of a wiring-cost pressure (Appendix C).
Strengths
Budget-matched factorial design isolates the wiring variable
All conditions share the same per-unit synapse budget (ds=84, ρrec=0.125 on the SHD reference, Table 2); only the feed-forward source-assignment rule differs (Appendix A.5), so parameter count cannot explain the accuracy gaps.
Unusually complete control set
Beyond structured vs random, the paper runs a fixed-permutation scrambling control (A.5), a cross-dataset misalignment test (motion wiring on CIFAR10-DVS, spatial wiring on DVS-Gesture, §4.2), an all-to-all full-input baseline (§4.3), an ℓ1 comparison (§4.4), and recurrent modular/small-world variants honestly reported as null results (Appendix C).
Above-average statistical hygiene
Honest scoping and null results
High reproducibility
Weaknesses
Abstract overclaims the ℓ1-vs-structured gap
'Across all tested network sizes' is not supported
Protocol-sensitive effect at the reference configuration
Motion coordinate conflates alignment with feature engineering
Generality limited to one neuron family and three datasets
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This preprint (arXiv:2609.26940, submitted to the 2026 Symmetry and Geometry in Neural Representations proceedings track) asks a well-posed question at the intersection of computational neuroscience and machine learning: do receptive fields aligned with task-relevant sensory coordinates help *computationally* — beyond the resource efficiency that may explain their biological emergence — and does that help survive as the expressive capacity of individual neurons grows? The central claims are: (1) structured feed-forward wiring beats budget-matched random wiring on SHD and DVS-Gesture (§4.1); (2) the benefit is specific to task geometry — it disappears under coordinate scrambling and task-irrelevant alignment (§4.2); (3) increasing neuronal expressivity (ELM memory units M) shrinks and can eliminate the benefit (§4.1); (4) generic ℓ1 sparsity only partially recovers it (§4.3–4.4); and, at appendix level, (5) recurrent topological structure adds nothing without wiring-cost pressure (Appendix C).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is an Expressive Leaky Memory (ELM) neuron?
An artificial neuron model that keeps M internal memory units with fixed decay timescales spanning several orders of magnitude; a small MLP combines current synaptic input with these memory states at each time step. Raising M makes a single neuron computationally more expressive without changing the network's wiring, which is what allows this paper to vary neuron complexity independently of connectivity.
Do structured receptive fields actually improve network performance in this study?
Yes, when aligned with the task-relevant sensory coordinate. Compared with budget-matched random input wiring, structured wiring added, for example, +7.26 test-accuracy points on DVS-Gesture (N=128) and +2.49 points on the Spiking Heidelberg Digits (N=110). The benefit disappeared when the sensory coordinate was scrambled or when wiring followed a task-irrelevant coordinate.
What datasets were used and are they public?
Three public neuromorphic datasets: Spiking Heidelberg Digits (spoken digits, 20 classes), DVS-Gesture (11 hand/arm gestures), and CIFAR10-DVS (10 object classes), all under CC BY 4.0. The paper's code, including wiring conditions and dataloaders, is released at github.com/AgneseAdorante/elmnetwork.
Does sparsity alone explain the benefit of structured receptive fields?
The paper argues no. An ℓ1 penalty on the feed-forward weights of full-input networks induced coarse frequency-selective input fields and improved test accuracy (+2.50 points, q=0.034), but performance stayed numerically below explicitly structured receptive fields; the remaining difference (−1.19 points) was not statistically significant (q=0.092), which the authors read as structured alignment providing a benefit beyond generic sparsity.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures are available only through captions in the extracted text; visual trends, s.e.m. shading, and panel details could not be independently verified.
- The extracted text appears complete (main text plus Appendices A–D including Table 4 statistics), supporting high confidence; no sections were detected as truncated.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.