PaperlayerView exact revision on arXiv

Source-linked AI summary

Vmem-$\varphi$: Low-Compute Out-of-Distribution Detection in Spiking Neural Networks from Membrane-Potential Statistics

Arul Rana, Agrim Tripathi, Shoaib Ahmed Dipu, Md. Shaown Miah, Syed Ishtiaque Ahmed, Sayeed Shafayet Chowdhury

arXiv:2610.00350v1cs.CV

TL;DR

OOD detection in SNN event-camera detectors is challenging because many existing methods rely on outputs or computations that are unavailable or costly in this setting. The paper introduces membrane-potential descriptors, Gen1-C, and clean-data-fit MDD; with a 64-frame window, AUROC improves for all corruptions, though polarity-flip detection remains difficult.

  • Problem

    Many OOD methods rely on model outputs or auxiliary computations that are unavailable or poorly suited to low-compute SNN object detectors.

    Source: §1 · p. 1

  • Method

    The paper derives descriptors from already-computed membrane potentials and uses clean-data-fit MDD to combine complementary deviation scores, alongside the Gen1-C corruption benchmark.

    Source: §1 · pp. 1–2

  • Results

    Averaging MDD scores over a bounded 64-frame window increases AUROC for all corruptions, while polarity-flip detection remains difficult and that corruption barely harms the detector at highest severity.

    Source: §4.3 · p. 6, §4.7 · p. 8

  • Takeaways & Limitations

    Membrane dynamics provide a practical sensor-health signal, and Gen1-C supports standardized evaluation of event-based robustness.

    Source: §6 · p. 9

    The evaluation applies one synthetic corruption family at a time and does not test mixed corruptions, orderings, or fault interactions.

    Source: §5 · p. 9

Abstract

from arXiv · show

Spiking Neural Networks (SNNs) offer an energy-efficient approach to processing event-camera data, yet out-of-distribution (OOD) detection remains challenging in this setting. Existing OOD detection methods often depend on model outputs or computational components that are unavailable in object detection SNNs or are poorly suited to low-compute deployment. To that effect, we show that the subthreshold membrane potential \(V_{\mathrm{mem}}(t)\) provides a useful internal signal for detecting distribution shifts. Simple per-channel statistics derived from these membrane dynamics enable OOD detection. To evaluate this approach, we introduce Gen1-C, an event-camera corruption benchmark developed upon the Prophesee Gen1 automotive detection dataset, containing six sensor-motivated histogram-level stress tests at five severity levels. We further propose the Multi-Descriptor Deviation (MDD), a corruption-blind method that operates on membrane-potential statistics. At the highest corruption severity, MDD achieves an AUROC of more than 0.88 on five of the six corruptions using only a bounded 64-frame observation window. Notably, the remaining corruption is also the one that has the smallest effect on the underlying detector. These results show that the temporal membrane-potential dynamics can provide an effective and low-cost signal for OOD detection in SNN-based event perception.

1 INTRODUCTION

OOD detection in event-camera SNNs is difficult because many existing methods rely on outputs or computations that are costly or unavailable in these detectors. The paper uses already-computed membrane activity as an OOD signal, introduces Gen1-C, and proposes clean-data-fit MDD.

Motivation

  • Existing OOD methods often require class scores, gradients, backward passes, or separate feature networks that low-power spiking object detectors do not cheaply produce.This deployment mismatch motivates monitoring an internal SNN state instead.Source: §1 · p. 1

Membrane signal

  • Subthreshold membrane potential Vmem(t) is already updated by the spiking front end, so reading it requires no additional task-network forward pass.The resulting descriptors are computed as a separate post-processing stage.Source: §1 · p. 1

Contributions

  • The paper introduces φ as a summary of membrane state and Gen1-C, a seeded, model-agnostic benchmark of six sensor-motivated corruptions built on Prophesee Gen1.The benchmark is intended to support standardized evaluation of event-based robustness.Source: §1 · p. 2
  • MDD scores inputs along complementary directions in φ-space and combines the scores with an OR rule, fitting only on clean in-distribution data.Its fitting does not use corruption samples.Source: §1 · p. 2

2 BACKGROUND AND RELATED WORK

Prior OOD methods use network outputs, internal features, gradients, or stored representations, while SNN OOD work has focused mainly on image classification. Event-camera robustness research includes controlled corruptions, varied recording conditions, and modeled sensor effects.

OOD detection for ANNs

  • Many ANN OOD methods rely on softmax outputs, input perturbations, internal features, backpropagation gradients, or stored in-distribution representations.MDD reduces the memory burden of KNN-based methods using PCA and reference-set reduction.Source: §2 · p. 2

OOD detection for SNNs

  • Existing SNN OOD studies cited here focus mainly on image classification, using spike-count patterns, sparse latent manifolds, or residual SNN architectures.The cited approaches include Mart´ınez-Seras et al. (2023), Terres-Escudero et al. (2024), and Avramovi´c et al. (2025).Source: §2 · p. 2

Event-camera robustness and benchmarks

  • Event-camera robustness work includes trajectory- and lighting-varied recordings and modeled effects such as threshold mismatch, limited bandwidth, and intensity-dependent noise.For automotive detection, the paper uses Prophesee Gen1, which has 39 hours of driving data with car and pedestrian bounding-box annotations.Source: §2 · p. 2

3 METHOD

The method evaluates sensor-motivated event-histogram corruptions and derives pooled and spatial descriptors from SNN membrane dynamics. MDD combines complementary deviations from clean-data references to detect shifts without fitting on corrupted samples.

3.1 CORRUPTION SUITE

  • Gen1-C represents sequences as event-histogram stacks and applies severity-graded transformations designed to perturb distinct signal aspects with minimized mechanism overlap.The benchmark includes six corruptions, shown as normal and corrupted event frames.Source: Figure 1 · p. 2, §3.1 · p. 3

3.2 MEMBRANE-POTENTIAL DESCRIPTORS

  • The pooled descriptor summarizes per-channel membrane mean, variance, and excess kurtosis, while the complementary spatial descriptor retains map organization and measures dispersion, local flatness, and persistence.Global average pooling produces three scalar features per channel; the spatial descriptor avoids that collapse.Source: §3.2 · pp. 3–4

3.3 MULTI-DESCRIPTOR DEVIATION (MDD)

  • MDD uses three complementary branches because corruptions can alter different input properties, and its fitting uses no corrupted data.The figure identifies the Multi-Descriptor Deviation Detector; the complete fitting and fusion procedure is specified separately.Source: §3.3 · p. 4, Figure 2 · p. 4

3.3.1 PCA RESIDUAL (B1: RESIDUAL)

  • The PCA-residual branch measures energy discarded by the retained clean-data principal subspace to detect descriptor structure in low-variance directions.It complements the deep-layer and spatial branches.Source: §3.3.1 · p. 4

3.3.2 DEEP-LAYER DEVIATION (B2: L4)

  • The deep-layer branch models the deepest-layer descriptor with a Ledoit–Wolf covariance estimate and flags absolute Mahalanobis-distance deviation from the clean calibration median, normalized by MAD.Both unusually high and unusually low distances are treated as anomalous.Source: §3.3.2 · p. 5

3.3.3 SPATIAL DISPERSION AND ORGANIZATION (B3: SPATIAL)

  • The spatial branch combines dispersion and spatial-organization signals, allowing either atypical dispersion or anomalous organization to activate it.Its components include persistence fractions and a signed flatness component.Source: §3.3.3 · p. 5

4 EXPERIMENTAL RESULTS

Across Gen1-C, detector damage varies by corruption, while MDD's fused AUROC improves under 64-frame aggregation across all six corruptions. Ablations identify complementary branch behavior, and an MCU-oriented streaming variant trades some AUROC for a reduced implementation.

4.1 MOTIVATION - A DETECTOR DEGRADES UNDER CORRUPTIONS

  • Event flood is most destructive, reducing detector mAP from 0.341 to 0.003, while polarity flip causes an approximately 4% decline at S5.The detector is tested without retraining, with corruption applied to the input event histogram.Source: §4.1 · p. 5, Figure 3 · p. 5

4.2 MDD SETUP

  • MDD uses clean fitting data, calibration, and held-out validation across 302 sequences, while monitoring front-end membrane states independently of the detection head.Its monitored descriptors include per-channel membrane moments and spatial statistics; the primary split is by sequence.Source: §4.2 · p. 6

4.3 MAIN RESULT: MDD OOD DETECTION

  • At W=64, MDD aggregation further increases AUROC for all six corruptions; per-frame performance is already strong on four at severity 5.Polarity-flip detection remains difficult even with W=64, although that corruption barely harms the detector at the highest severity.Source: Table 1 · p. 6, §4.3 · p. 6

4.4 BRANCH CONTRIBUTIONS

  • At S5, removing spatial branch B3 lowers dropout detection from 0.987 to 0.478, while removing deep-layer B2 changes fused jitter performance from 0.886 to 0.887.The deep-layer branch is strongest individually for temporal jitter; the spatial branch dominates spatial dropout and event-rate shift, while PCA-residual B1 adds complementary information.Source: §4.4 · p. 6, Table 2 · p. 7

4.5 REPRESENTATION ABLATION

  • Spatial statistics have the largest reported Jensen–Shannon divergence (0.184) and MMD2 (0.173), while concatenated representations lead supervised C2ST balanced accuracy (0.771) and AUPR (0.650).Polarity flip shows no robust separation across the reported diagnostics.Source: §4.5 · p. 7, Table 3 · p. 7

4.6 CALIBRATION AND UNCERTAINTY

  • At a held-out clean FPR of 2.8%, dropout is fully detected through the 95th-percentile threshold, but its TPR collapses at the 99th percentile.At the 95th percentile, rate shift and event flood retain 85.1% and 78.8% recall, respectively; polarity flip reaches only 9.2% TPR.Source: Table 4 · p. 7, §4.6 · p. 7

4.7 WINDOWED AGGREGATION: THE DEPLOYABLE OPERATING POINT

  • For event flood, windowing raises AUROC from 0.55 to 0.98; polarity flip remains near chance, moving from 0.55 to 0.57 through W=128.W=64 averages approximately 3.2 seconds of observations and targets sustained degradation or drift rather than sub-second alarms.Source: Figure 4 · p. 8, §4.7 · p. 8

4.8 MATCHED OOD-DETECTOR BENCHMARK

  • The matched benchmark compares MDD with feature-based post-hoc detectors on the same Gen1-C sequences, severities, and recording split, using the same clean-fit and calibration data.The suite includes deep nearest neighbors, DIME, PatchCore, Gaussian Mahalanobis, and feature-density flow; the supplied passages do not report comparative outcomes.Source: §4.8 · p. 8, Table 5 · p. 8

5 LIMITATIONS AND FUTURE WORK.

The evaluation identifies weak polarity-flip detection and limits its corruption tests to isolated synthetic corruption families. It also notes preprocessing and lower-end hardware as scope constraints.

5 LIMITATIONS AND FUTURE WORK.

  • MDD's AUROC remains near chance on polarity flip, and detection is insufficient at low severities across all corruption families.The authors attribute the polarity limitation to membrane statistics that do not retain a useful polarity-sensitive signal.Source: §5 · p. 9
  • Evaluation applies one synthetic corruption family at a time, leaving mixed corruptions, orderings, and fault interactions untested.Source: §5 · p. 9
  • Preprocessing may attenuate or discard raw corruption signals before MDD observes membrane activity; Efficient MDD also misses the latency threshold on a lower-end board.Source: §5 · p. 9

6 CONCLUSION

The conclusion frames membrane summaries as a practical sensor-health signal and Gen1-C as a controlled corruption suite. The appendix describes the descriptors and MDD interface, while qualifying its membrane analysis as an approximation that does not establish detector monotonicity.

6 CONCLUSION

  • MDD turns intermediate membrane activity into a calibrated sensor-health score, while Gen1-C provides a controlled six-family suite for evaluating sensor changes independently of detection.The conclusion presents both as a foundation for standardized research on event-camera robustness.Source: §6 · p. 9

A.1 DESCRIPTOR AND DETECTOR INTERFACE

  • The pooled descriptor summarizes per-channel membrane means, variances, and excess kurtoses, while a spatial descriptor retains dispersion and organization information discarded by global pooling.The MDD interface uses three branches: PCA residual, deep-layer deviation, and spatial dispersion/organization, combined by calibrated maximum fusion.Source: §A.1 · p. 12

A.2 MEMBRANE DYNAMICS AND RESET ERROR

  • The appendix's membrane analysis uses an idealized stable recurrence rather than specifying every host's PLIF update or reset rule.Its reset-error bound can become uninformative with large spike disagreement or weak leakage; additional resets can also attenuate mean changes, motivating descriptors beyond the mean.Source: §A.2 · p. 14

A.3 CONDITIONAL LINEAR-RESPONSE RESULTS

The appendix derives conditional relationships between input changes and membrane statistics under specified linear, reset-free, or stationarity assumptions. It also bounds what those results imply: they do not establish behavior of compressed descriptors or the complete MDD detector.

A.3 CONDITIONAL LINEAR-RESPONSE RESULTS

  • A mean input perturbation affects mean membrane statistics when its temporally accumulated projection is nonzero, while reset feedback can oppose the response.For temporally uncorrelated stationary input in a reset-free model, positive-semidefinite input covariance growth implies positive-semidefinite membrane covariance growth; correlated bins and nonlinear resets are outside that result.Source: §A.3 · p. 14
  • Membrane variance change depends on both marginal input variance and temporal-correlation changes; a quadratic-distance result is non-decreasing only for expansive covariance shifts.Under covariance contraction, the distance can decrease; absolute deviations in B2 and the dispersion part of B3 admit either tail as evidence but do not guarantee detection or monotonicity.Source: §A.3 · p. 15

A.4 MEMBRANES, SPIKES, AND SPATIAL POOLING

  • The information inequality concerns complete pre-threshold membrane trajectories, not compressed φ or φspatial descriptors; membrane moments nevertheless describe continuous changes that spike counts can discard.Global average pooling can assign identical values to spatially permuted moment maps, while MDD also retains spatial dispersion and organization.Source: §A.4 · p. 15

A.5 SCOPE AND LIMITATIONS

  • The analytical results assume stable linear dynamics, matched initial states, and specified noise or stationarity conditions, and do not prove performance or monotonicity of the three-branch detector.The information inequality applies to complete trajectories, and detection performance is instead measured empirically on held-out validation data.Source: §A.5 · p. 15

B GEN1-C CORRUPTION DESIGN, REPRODUCIBILITY, AND DIVERSITY

Gen1-C applies six severity-graded sensor-motivated corruptions to event histograms, and analyses examine how distinct their input and descriptor signatures are. The section also reports residual-width and observation-window sensitivity.

B GEN1-C CORRUPTION DESIGN, REPRODUCIBILITY, AND DIVERSITY

  • Gen1-C applies six corruption families to event histograms: hot pixels, event floods, temporal jitter, event-rate shift, polarity flip, and spatial dropout.Each family is applied to a complete sequence at five severity levels, preserving clean/corrupted sequence pairing.Source: B GEN1-C CORRUPTION DESIGN,… · p. 16
  • 0.16 is the median pairwise mean absolute cosine across corruption-family input residuals; 14 of 15 family pairs are at or below 0.32, with temporal jitter and rate shift the exception at 0.64.The comparison uses six recordings and three independent seeds per S5 corruption.Source: Figure 5 · p. 16, B GEN1-C CORRUPTION DESIGN,… · pp. 16–17
  • 0.67 → 0.33 is the temporal-jitter/rate-shift alignment change from input residuals to φspatial; hot-pixel/flood and jitter/dropout alignments instead rise from 0.24 → 0.43 and 0.19 → 0.44.Thus feature extraction reverses some input-overlap patterns: distinct mechanisms can yield similar spatial-descriptor signatures.Source: Figure 6 · p. 17, B GEN1-C CORRUPTION DESIGN,… · p. 17

C.1 RESIDUAL PROJECTION SENSITIVITY

  • The four PCA residual widths differ by less than .005 mean AUROC, and width 64 exceeds width 32 by only .0005, so the sweep does not identify a sharply optimal dimension.In the same candidate sweep, a direct clean-scaled descriptor norm scores .7510 and omitting B1 scores .7788.Source: Table 9 · p. 18, §C.1 · p. 18

C.2 WINDOW-LENGTH SENSITIVITY

  • The selected model's S5 macro AUROC increases with observation window; W = 64 spans approximately 3.2 seconds at 50 ms per frame.This is an accuracy–observation-time trade-off.Source: Table 10 · p. 18, §C.2 · p. 18

C.3 FUSION-COEFFICIENT SENSITIVITY

The fusion sensitivity analysis compares calibrated branch-score combinations at S5 and W = 64. Maximum fusion is reported as the strongest overall option, while the softmax-style check finds little reason to add a temperature parameter.

C.3 FUSION-COEFFICIENT SENSITIVITY

  • Maximum fusion is the strongest overall option among the tested calibrated branch-score joins at S5 and W = 64.The comparison includes arithmetic mean, top-two mean, noisy-OR, percentile combinations, chi-square combination, and individual-branch baselines.Source: Table 11 · p. 18, §C.3 · p. 19
  • The median-subtracted rule helps rate shift slightly near α = .25 but progressively suppresses temporal jitter and spatial dropout.The α sweep and alternative joins are diagnostic ablations; selecting α from corrupted final labels would make the choice corruption-informed.Source: §C.3 · pp. 18–19

C.4 SOFTMAX-STYLE SMOOTHING CHECK

  • .881, .881, .879, .874, .865, .855, and .846 are the mean AUROC values for temperatures τ = 0.1, 0.25, 0.5, 1, 2, 4, and 8, respectively, at S5 and W = 64.The small differences near τ = 0 do not justify an additional temperature parameter.Source: Figure 7 · p. 19, §C.4 · p. 19

D CROSS-DATASET AND CROSS-MODEL EVALUATION

MDD transfers across a sensor/dataset shift and a host-model change when refit and recalibrated on each stack’s clean data. Its corruption ranking varies substantially across hosts, so calibration transfers more readily than a universal ranking of corruption difficulty.

D CROSS-DATASET AND CROSS-MODEL EVALUATION

  • Gen4 evaluation uses a 1-Mpixel, 1280×720 sensor, three retained classes, and five-frame sequences, unlike the Gen1 setup.Gen4 also differs in scene distribution, checkpoint, and detector input geometry.Source: D CROSS-DATASET AND CROSS-MODEL… · p. 19
  • MDD transfers across the Gen4 sensor/dataset shift and SpikeYOLO host-model change without reusing Gen1 references, but each run refits and recalibrates on its own clean data.Gen4 changes the sensor and dataset; SpikeYOLO changes the host network.Source: D CROSS-DATASET AND CROSS-MODEL… · pp. 19–20
  • Polarity flip is near chance on Gen4 but reaches 0.930 AUROC on SpikeYOLO at S5, while spatial dropout is stronger on Gen4.The host-specific MDD results are reported by severity at W = 64.Source: D CROSS-DATASET AND CROSS-MODEL… · p. 20, Table 12 · p. 20

E EFFICIENT MDD: DESCRIPTOR, SCORER, AUROC, AND MCU RESULTS

Efficient MDD samples a reduced set of membrane channels and spatial positions, then scores 84 features using five clean-calibrated branches. It retains strong detection for several corruptions, while polarity flip remains difficult and temporal jitter loses detection strength under the reduced descriptor.

E EFFICIENT MDD: DESCRIPTOR, SCORER, AUROC, AND MCU RESULTS

  • Efficient MDD monitors four frozen channels per PLIF layer with fixed spatial subsampling, using bounded tensor shapes across its four layers.The measured C compute interval excludes input collection from the host SNN and transfer to the monitor.Source: E EFFICIENT MDD: DESCRIPTOR, SCORER,… · p. 20
  • Its 84 active features feed five clean-calibrated scores fused by maximum, without PCA, an RCF lookup, or a stored clean reference set.The features comprise 80 moment/spatial features and four persistence features; the 64-frame averaging window is outside the per-frame scorer.Source: E EFFICIENT MDD: DESCRIPTOR, SCORER,… · p. 20

E.1 AUROC BY SEVERITY

  • The reduced descriptor preserves large localized hot-pixel and event-flood deviations and strong S5 detection for rate shift and spatial dropout, but loses 0.138 AUROC for temporal jitter at S5.Hot pixels remain detectable from S2; polarity flip is difficult for both detectors throughout.Source: §E.1 · pp. 20–21, Table 13 · p. 20

E.2 PHYSICAL-MCU MEASUREMENTS

  • The MCU measurements time a static float32 C routine for layer extraction/state calls and scoring, excluding the external SNN, raw-Vmem gathering, and fixture staging.The supplied tables identify physical-MCU timing and hardware resource requirements but do not provide their cell values.Source: §E.2 · p. 21, Table 14 · p. 21, Table 15 · p. 21

F WHY POLARITY FLIP IS DIFFICULT TO DETECT

  • Polarity flip is weakest on the primary Gen1 host, with S5 AUROC of 0.545 per frame and 0.568 at W = 64.Some frames are unchanged by the corruption, and moment aggregation can discard polarity-sensitive membrane changes.Source: F WHY POLARITY FLIP IS DIFFICULT TO… · p. 22

F.1 CORRUPTION RULE AND UNCHANGED FRAMES

  • At S5, untouched frames dilute a fixed memoryless frame score’s separation, although the population identity is not a ceiling for window or history-dependent scores.The corruption swaps complete ON and OFF blocks independently per frame with probability q; unselected frames remain unchanged.Source: §F.1 · p. 22

F.2 MEMBRANE CHANGES CAN CANCEL IN THE DESCRIPTOR

  • Signed polarity-sensitive membrane changes can cancel in temporal moments, and spatial pooling can cancel further differences across locations.The moment identities do not assume linear neurons, small perturbations, or negligible resets; the pooling statistic measures aggregation loss, not separability.Source: §F.2 · p. 22

F.3 CHECKPOINT DIAGNOSTICS

  • In the checkpoint witness, raw membrane changes across layers were .717, .648, .485, and .441, while pooled and spatial descriptor changes were only 5.86 × 10^-8 and 1.09 × 10^-7.This demonstrates a possible descriptor-collision mechanism, not its frequency in Gen1; a fixed scorer cannot distinguish exactly equal descriptor inputs.Source: §F.3 · p. 23

F.4 WHAT THE CURRENT MDD RESULTS SHOW

Polarity flip causes relatively little aggregate task degradation, yet MDD detects it weakly. The diagnostics attribute this gap to signed membrane variation canceling in retained moment statistics, alongside dilution from unchanged frames.

F.4 WHAT THE CURRENT MDD RESULTS SHOW

  • At S5, polarity flip lowers mAP from .358 on clean inputs to .344, while MDD’s polarity separation remains weak.The task-performance change does not remove the stated OOD-detection limitation.Source: §F.4 · p. 23, Table 17 · p. 23
  • Polarity-sensitive membrane responses change, but much of their signed variation cancels in the retained moment statistics, weakening the deviation available to MDD.The conclusion also identifies untouched-frame dilution; neither spatial organization nor longer averaging fully addresses the limitation.Source: §F.4 · p. 24

Related papers

Loading 2610.00350v1…