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

cs.CVSep 29, 2026

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

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

Organizations: Indian Institute of Technology Kharagpur · Indiana University Indianapolis · Bangladesh University of Engineering and Technology · University of Toronto

Abstract

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.

Figures & tables

Panel of Figure 1 — Figure 1: The six Gen1-C corruptions: normal (left) versus corrupted (right) event frame.

Panel of Figure 2 — Figure 2: The Multi-Descriptor Deviation Detector.

Panel of Figure 3 — Figure 3: Task degradation across corruption families and severities.

Representation	AUPR	FPR@95	JS bits	MMD 2	C2ST BA
Full membrane $\varphi$	0.637	0.818	0.158	0.053	0.695
$\mu$	0.602	0.810	0.148	0.075	0.693
$\sigma^{2}$	0.569	0.944	0.025	0.027	0.666
$\kappa$	0.589	0.880	0.009	0.016	0.653
$\varphi_{\mathrm{spatial}}$	0.640	0.786	0.184	0.173	0.750
$\varphi+\varphi_{\mathrm{spatial}}$	0.650	0.821	0.159	0.147	0.771

Table 3: Distributional diagnostics.

Calibration percentile	Clean FPR	Hot pixel	Jitter	Rate shift	Flood	Dropout	Polarity flip
85%	0.146	1.000	0.713	0.893	0.812	1.000	0.216
90%	0.099	1.000	0.630	0.879	0.803	1.000	0.168
95%	0.028	1.000	0.551	0.851	0.788	1.000	0.092
99%	0.005	1.000	0.389	0.349	0.783	0.008	0.034

Table 4: MDD operating points at S5,

W=64

Figure 4: Fused AUROC vs. aggregation window $W$ (S5)

Method	Hot pixel	Jitter	Rate shift	Dropout	Flood	Polarity	Mean
MDD (ours)	1.000	0.863	0.931	0.963	0.551	0.545	0.809
Membrane-kNN	0.930	0.589	0.380	0.499	0.539	0.493	0.572
Spike-count-kNN	0.973	0.400	0.341	0.451	0.518	0.491	0.529
DIME	0.853	0.541	0.703	0.460	0.511	0.484	0.592
Gaussian Mahalanobis	0.909	0.550	0.697	0.464	0.514	0.487	0.603
PatchCore	0.067	0.457	0.660	0.509	0.535	0.525	0.459

Table 5: Matched OOD-detector benchmark at S5.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Family	S1	S2	S3	S4	S5
Hot pixels $(n,\delta)$	$(10,20)$	$(30,40)$	$(80,80)$	$(150,140)$	$(300,200)$
Event flood $(b,f,\delta)$	$(2,.05,60)$	$(4,.08,100)$	$(8,.12,150)$	$(15,.18,200)$	$(25,.25,255)$
Temporal jitter $s_{\max}$	1	2	3	5	8
Rate shift $(u,o)$	$(.80,1.20)$	$(.65,1.40)$	$(.50,1.65)$	$(.35,2.00)$	$(.20,2.50)$
Polarity flip $p$	.05	.10	.20	.35	.50
Spatial dropout $(k,f_{h},f_{w})$	$(1,.05,.05)$	$(1,.10,.10)$	$(2,.10,.10)$	$(2,.15,.15)$	$(3,.20,.20)$

Appendix

Table 7: Gen1-C parameter schedules.

Figure 5: S5 input residual alignment across corruptions.

Figure 6: S5 spatial descriptor residual direction alignment.

Corruption	$\varphi$	$\varphi_{\rm spatial}$
Hot pixels	542.4	$1.53\times 10^{5}$
Event flood	578.2	$4.01\times 10^{5}$
Temporal jitter	39.3	36.0
Polarity flip	8.1	6.3
Rate shift	100.4	89.3
Spatial dropout	17.2	25.6

Appendix

Table 8: Median S5 descriptor residual norms.

Removed PCA components	16	32	64	128
Mean $W=64$ AUROC	.7916	.7958	.7964	.7949

Appendix

Table 9: MDD width sensitivity.

Corruption	1	8	16	32	64	128	256	Full
Hot pixels	1.000	1.000	1.000	1.000	1.000	1.000	1.000	1.000
Temporal jitter	.863	.866	.870	.875	.886	.899	.925	.964
Rate shift	.931	.935	.941	.947	.955	.963	.973	.984
Spatial dropout	.963	.965	.974	.980	.987	.988	.993	.989
Event flood	.551	.614	.677	.778	.892	.981	.999	1.000
Polarity flip	.545	.560	.563	.565	.568	.571	.577	.593

Appendix

Table 10: MDD S5 AUROC by window.

Corruption	$\alpha=0$	.25	.50	.75	1.00	1.25	1.50	2.00
hot_pixel	1.000	1.000	1.000	1.000	1.000	1.000	1.000	1.000
temporal_jitter	.886	.882	.876	.865	.846	.810	.742	.501
event_rate_shift	.955	.957	.956	.953	.948	.941	.926	.851
spatial_dropout	.987	.982	.966	.953	.943	.933	.922	.897
event_flood	.892	.892	.893	.893	.893	.893	.893	.893
polarity_flip	.568	.564	.560	.556	.551	.544	.529	.496

Appendix

Table 11: Fusion-coefficient sweep at S5 and

W=64

Panel of Figure 7 — Figure 7: MDD combiner results at S5 and $W=64$ .

Panel of Figure 8 — Figure 8: Efficient MDD

	MDD					Efficient MDD
Corruption	S1	S2	S3	S4	S5	S1	S2	S3	S4	S5
Hot pixel	1.000	1.000	1.000	1.000	1.000	.571	.908	1.000	1.000	1.000
Temporal jitter	.692	.745	.798	.847	.886	.559	.604	.631	.659	.748
Event rate shift	.519	.579	.888	.906	.955	.508	.515	.728	.743	.888
Spatial dropout	.800	.952	.962	.984	.987	.493	.493	.500	.505	.926
Event flood	.556	.602	.692	.795	.892	.545	.615	.710	.791	.889

Appendix

Table 13: MDD and Efficient MDD AUROC.

Board / MCU	Core / FP	MHz	Captured max (ms)
B-L475E-IOT01A / STM32L475VG	M4F / hardware	80	10.8388
nRF52840 DK / nRF52840	M4F / hardware	64	13.6235
nRF52832 DK / nRF52832	M4F / hardware	64	19.4250
DWM3001 / nRF52833	M4F / hardware	64	19.4555
STM32F103RE	M3 / software	72	247.0523

Appendix

Table 14: Efficient MDD physical-MCU timing.

Metric	Value	Notes / Scope
Persistent state memory	5.36 KB	Monitor state
Sampled-layer buffer	51.20 KB	Peak single-layer input
Total monitor RAM	56.56 KB	Excludes call stack
Input traffic	32,000 floats/frame	128 KB read per frame
Scorer coefficient products	844 ops/frame	Triangular forms and flatness branch
Temporal-moment cells	$\leq$ 3,200/frame	Upper bound (data-dependent)

Appendix

Table 15: Hardware resource requirements for MDD implementation.

Diagnostic median	PLIF 1	PLIF 2	PLIF 3	PLIF 4
Raw membrane relative change	.211	.257	.345	.634
Mean-map retention $\rho$	.237	.171	.232	.195
Variance-map retention $\rho$	.273	.237	.260	.184
Kurtosis-map retention $\rho$	.149	.112	.117	.068

Appendix

Table 16: Polarity changes before and after pooling.

Score or aggregation	AUROC
B1: PCA residual, $W=64$	.583
B2: deep-layer deviation, $W=64$	.512
B3: spatial deviation, $W=64$	.558
Fused MDD, $W=1$	.545
Fused MDD, $W=64$	.568
Fused MDD, full recording	.593

Appendix

Table 17: Current MDD polarity detection at S5.

OOD Detection

Out-of-distribution (OOD) detection identifies inputs unlike a model’s training data so systems can flag uncertainty, reject inputs, or adapt safely.

See all 223 papers

Sequence-SOD: Bio-inspired Sequence-aware Spiking ObjectDetection for Event Cameras
Jul 29, 2026Katharina Bendig, René Schuster, Didier StrickerSpiking Neural Networks Event-Based Object Detection
RPTU – Technical University Kaiserslautern-Landau, Gottlieb-Daimler-Straße 47, Kaiserslautern 67663, Germany · DFKI – German Research Center for Artificial Intelligence, Trippstadter Str. 122, Kaiserslautern 67663, Germany
Breaking Diversity Collapse in Spiking Pseudo-Ensembles for Efficient OOD Detection in Remote Sensing
Aug 2, 2026Srinivas Anumasa, Rushi Shah, Qiran Zou +1Efficient Neural Network Inference Ensemble Learning
National University of Singapore
MorphSNN: Adaptive Graph Diffusion and Structural Plasticity for Spiking Neural Networks
Mar 15, 2026Yongsheng Huang, Peibo Duan, Yujie Wu +7Spiking Neural Networks Dynamic Neural Networks
School of Software, Northeastern University, China · Guangdong Institute of Intelligence Science and Technolog, China · The Hong Kong Polytechnic University, China +1

Jul 29, 2026J/K move · Enter open · S save

Sequence-SOD: Bio-inspired Sequence-aware Spiking ObjectDetection for Event Cameras

Katharina Bendig, René Schuster, Didier Stricker

RPTU – Technical University Kaiserslautern-Landau, Gottlieb-Daimler-Straße 47, Kaiserslautern 67663, Germany · DFKI – German Research Center for Artificial Intelligence, Trippstadter Str. 122, Kaiserslautern 67663, Germany

Spiking Neural Networks Event-Based Object Detection

Open paper AskPDF arXiv

Event cameras follow a retina-inspired sensing principle, reporting local intensity changes asynchronously with hightemporal resolution and a wide dynamic range. Spiking Neural Networks (SNNs) complement these sparse event streams through brain-inspired dynamics, using sparse spikes and leaky membrane potentials to integrate information over time. However, many SNN object detectors process isolated event intervals with a single label and reset the network state after each prediction, thereby underusing temporal information in continuous event streams. We introduce Sequence-SOD, a sequence-aware SNN object detector that processes extended event sequences containing labels at multiple time points. Events are accumulated into short intervals, discretized into temporal steps, and fed sequentially to an SSD-style Spiking DenseNet while preserving membrane potentials across intervals within a sequence, so that detection is driven by an evolving neural state instead of independently reset input windows. On the Gen1 Automotive Detection Dataset, sequence-aware training improves mAP from 23.38 for single-interval training to 25.30 without augmentation and to 26.88 withevent-data augmentation. The model achieves a theoretical prediction frequency of 40 Hz. Training and evaluating SNN object detectors on extended event sequences improves their ability to exploit temporal cues while preserving the energy-efficiency benefits of sparse spiking computation. The results highlight sequence-aware training as a complementary direction to architectural improvements for event-based SNN detection.

Abstract

OOD Detection

Explore similar work

Sequence-SOD: Bio-inspired Sequence-aware Spiking ObjectDetection for Event Cameras

Breaking Diversity Collapse in Spiking Pseudo-Ensembles for Efficient OOD Detection in Remote Sensing

MorphSNN: Adaptive Graph Diffusion and Structural Plasticity for Spiking Neural Networks