acceptodds
Under review as a conference paper at ICLR 2027

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

17 views

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.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Spiking Neural Networks, Out-of-Distribution Detection, Event-Based Vision, Membrane Potential, Robustness, Corruption Detection, Efficient AI

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_NngWOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses out-of-distribution (corruption) detection for SNN-based event-camera perception under low-compute constraints. It observes that subthreshold membrane potentials Vmem, already computed by a spiking front end, carry a usable corruption signal, and builds two descriptors: a pooled per-channel moment descriptor (mean, variance, excess kurtosis) and a spatial descriptor (dispersion, local flatness, persistence). The Multi-Descriptor Deviation (MDD) detector combines a PCA-residual branch, a deep-layer Mahalanobis branch, and a spatial branch via calibrated maximum fusion, fitted only on clean data. The authors also release Gen1-C, a six-family, five-severity corruption benchmark on Prophesee Gen1. On a hybrid SNN-ANN detector, MDD reaches AUROC above 0.88 on five of six corruptions at the highest severity using a 3.2 s observation window, outperforming an adapted suite of post-hoc OOD baselines (mean AUROC 0.809 vs. 0.624 for the next best), while polarity-flip detection remains near chance. An MCU implementation runs within the 50 ms frame budget on physical M4F boards with modest AUROC loss, and transfer to Gen4 and SpikeYOLO is reported.

Critical strengths
  1. The problem setting is real and the core idea is economical: reading membrane states adds no forward/backward pass to the host network, which is exactly the constraint that makes standard OOD methods unusable on spiking detectors. This is a sensible and previously unexploited internal signal for this setting.
  2. The evaluation protocol is unusually disciplined: corruption-blind fitting with a separate sensitivity split not used for selection, a matched baseline suite under identical sequences and calibration, cross-dataset (Gen4) and cross-model (SpikeYOLO) runs, and measured per-frame latency on five physical MCU boards rather than claimed efficiency.
  3. The treatment of failure is exemplary: the polarity-flip failure is decomposed into mixture dilution, moment cancellation, and an explicit descriptor-collision witness (Appendix F), and all limitations are stated rather than hidden. Gen1-C itself is a reusable contribution for event-camera robustness research.
Critical weaknesses
  1. The "corruption-blind" property holds only at the calibration stage: the descriptors themselves (local flatness, persistence, dispersion) are evidently designed with these corruption families in mind, and no experiment tests a corruption family that was unknown during design. The stated motivation for clean-only fitting — robustness to corruption types beyond the suite — is therefore asserted but never evaluated, e.g., via a leave-one-family-out design.
  2. The signal's effectiveness is host- and pipeline-specific. Headline results use one hybrid host; on Gen4 and SpikeYOLO the corruption ranking changes substantially and absolute AUROC is much weaker (e.g., event flood 0.670 and spatial dropout 0.450-0.770 across severities on SpikeYOLO). Combined with near-chance polarity-flip detection and weak low-severity performance on several families (rate shift 0.519, flood 0.556 at S1), the practical monitoring envelope is narrower than the abstract suggests.
  3. The matched baseline suite compares only network-derived features (membrane-kNN, spike-count-kNN, DIME, PatchCore, etc.). A monitor built directly on input-space event statistics (event rate, spatial support, ON/OFF asymmetry) requires no network at all and would show how much detection power is unique to membrane dynamics rather than trivially available from the raw histogram; this control is missing.
  4. Presentation needs editing: Section 3.3.3 contains a duplicated sentence, Appendix A.3-A.5 explicitly disclaim that the analysis establishes nothing about the detector while the main text cites these results as motivation, and there are typos (e.g., "failiure" in Section 5).
Questions for authors
  1. Event flood at S5 collapses detection mAP to 0.003 within a few frames, yet fused per-frame AUROC is only 0.551 and the deployable operating point needs a 3.2 s window. What is the detection-delay distribution for acute failures, and would a two-tier scheme (fast input-space alarm plus slow membrane monitor) cover both acute and sustained degradation?
  2. How does MDD compare, under the same protocol, with a corruption-blind monitor built directly on raw event-histogram statistics? This would isolate how much of the performance comes from membrane dynamics versus information already present in the input.
  3. Can you evaluate MDD on a corruption family excluded from the descriptor design process (e.g., v2e-style sensor artifacts such as contrast-threshold mismatch or bandwidth limiting) to substantiate the claim that the clean-only fitting generalizes beyond the suite?
  4. Table 10 shows event flood reaching 1.000 AUROC at full-recording aggregation versus 0.892 at W=64. What aggregation length would you actually recommend for a given alarm-latency budget, and how should a practitioner pick W without corruption labels?
4 comments
AuthorsOpenReview ↗

Response to Reviewer NngW

Thank you for your insightful review and helpful suggestions. All section references below correspond to the revised manuscript (where changes are highlighted in blue).

Frozen MDD on additional corruptions

no experiment tests a corruption family that was unknown during design.

We tested v2e-inspired histogram perturbations plus readout and packet-loss stress tests. Host weights, descriptors, MDD parameters, calibration and score directions remained fixed. On seven matched 256-frame prefixes, clean mAP was 35.03%. At S5:

Additional corruption, S5Task mAPAbsolute mAP lossMDD AUROC, W=64
Threshold-mismatch approximation34.74%0.29 pp.563
Background-count noise26.09%8.94 pp.929
Readout-throughput thinning29.24%5.79 pp.420
Refractory-cap approximation35.00%0.03 pp.571
Temporal packet loss26.16%8.86 pp.920

At S4, both background counts and packet loss cause at least five points of mAP loss:

  • Background counts: 5.35-point loss; .750 AUROC.
  • Packet loss: 5.72-point loss; .875 AUROC.

Frozen MDD detects harmful background noise and packet loss beyond Gen1-C. Throughput thinning is not reliably detected; threshold mismatch and refractory caps cause little task degradation at the tested settings.

These are v2e-inspired histogram tests, not runs of v2e.

Dependence on the host model and pipeline

Section 4.3 and Appendix D discuss how Gen4 and SpikeYOLO change performance and corruption rankings. Section 5 retains the polarity-flip and low-severity limitations. These results identify the conditions MDD monitors effectively and those needing additional signals.

Comparison with an input-only monitor

Section 4.5 adds the input-statistic control with the same MDD scoring method, separate clean-data fitting and the same test protocol. At W=64:

Corruption / severityInput statisticsTrained membranes
Mean, S1-S5.697.785
Hot pixel, S51.0001.000
Event flood, S5.877.885
Temporal jitter, S5.520.875
Polarity flip, S5.634.571
Event-rate shift, S5.940.946
Spatial dropout, S5.783.983
Mean, S5.792.877

Membranes lead overall; inputs lead on polarity flip. Section 4.5 includes spikes, random membranes and ANN features. Input/membrane descriptor dimensions differ; Appendix C.5 tests cross-frame PLIF state separately.

Acute failures and two-tier alarms

Section 4.7 and Appendix C.7 add an input alarm (W=1), a trailing-64-frame membrane alarm and their joint OR alarm. Monitors and thresholds remain frozen; thresholds use separate clean calibration recordings with a 1% alarm-frame budget. Testing uses 50 additional complete held-out recordings disjoint from fitting and calibration.

AlarmFloods with a warningNew warningsClean alarm framesClean alarm starts/min
Input W=1100.0%1,0813.22%17.03
Membrane W=6494.9%2771.17%0.16
Combined100.0%3303.31%14.24

Coverage includes warnings active before onset. For new warnings, median/P90 observation delays are both 50 ms, excluding compute and communication. The fast path covers all tested S5 floods, at a higher false-alarm rate; held-out rates can exceed the calibration target.

A separate 12-recording sustained-corruption test found the combined alarm active on 100% of hot-pixel frames and 38.1% of rate-shift frames, versus 94.5% and 34.0% for membranes. It is a tradeoff between false alarms and detection.

Choosing the averaging window without corruption labels

Choose the smallest W that meets the desired clean false-alarm rate within the allowed averaging time. Calibrate thresholds on clean calibration recordings and select W on separate clean validation recordings, without corruption labels. At 50 ms per frame, W=64 spans 3.2 s and W=128 spans 6.4 s; these are averaging spans, not measured alarm delays. Full-recording aggregation is an offline diagnostic.

Window averaging mainly reduces score noise; persistence separately tracks recording history. We added this practical selection guidance in Section 4.7.

Presentation and interpretation of Appendix A

We removed duplicate text and corrected typos.

Thank you again. Please let us know if you have further questions; we look forward to the discussion.

AuthorsOpenReview ↗

Thanks - Reviewer NngW

Dear Reviewer NngW,

As the discussion period is about to end in less than a day, we hope our responses have adequately addressed your concerns. Please let us know if you have further questions; we look forward to the discussion.

Thank you again for your valuable time and consideration.

Best Regards, The Authors

I thank the authors for a thorough and responsive rebuttal. The three additions directly resolve my main concerns: frozen MDD detects unseen v2e-inspired corruptions that cause real task degradation (background counts .929, packet loss .920 AUROC) without any refitting, the input-statistic control confirms that membrane features carry detection power beyond raw event statistics (mean S5 AUROC .877 vs. .792), and the two-tier alarm covers all tested acute floods at 50 ms delay with quantified false-alarm costs. The window-selection guidance and the presentation fixes are also appreciated. The remaining limitations — readout-throughput thinning and polarity flip — are now explicitly scoped rather than hidden, and are documented in the paper itself. My concerns about generalization and the missing control are resolved; I raise my rating to 4.

AuthorsOpenReview ↗

Thanks - Reviewer NngW

Thank you for your careful consideration of our responses and additional experiments, and for updating your rating. We appreciate your constructive feedback throughout the discussion, which helped us strengthen and clarify the paper.

Best Regards, The Authors

Reviewer_KcZcOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MDD, a clean-data-fitted corruption detector based on membrane-potential statistics in SNNs, and introduces the Gen1-C benchmark. Experiments include comparisons with post-hoc detectors, component ablations, cross-dataset and cross-model evaluations, and a simplified MCU implementation. The topic is promising, but the distinctive value of membrane dynamics requires further investigation.

Critical strengths

The paper explores a meaningful direction for monitoring event-based perception using existing SNN states.

The evaluation includes multiple corruption families, representative detection baselines, and additional datasets and architectures.

The hardware implementation provides useful evidence toward practical deployment.

Critical weaknesses

1.What evidence supports the physical realism and severity schedules of Gen1-C? Can the synthetic corruptions be validated against real degraded recordings or established event-camera simulations?

2The advantage of membrane potentials over input statistics, spike counts, and ANN activations is insufficiently established.

3The contribution of SNN temporal dynamics is not isolated from moment statistics and generic window averaging.

4Despite quantitative diagnostics, direct visualizations of clean and corrupted feature and score distributions are missing.

5Recording-wide persistence and excluded state-transfer costs limit the bounded-window and efficiency claims.

Questions for authors

1.What information do membrane potentials provide beyond input event statistics and spike counts? Can you include controlled comparisons with these signals and representative ANN activations? 2.Could you visualize clean and corrupted membrane descriptors and final scores across severities and representative layers, alongside input-statistic or spike-count counterparts?

4 comments
AuthorsOpenReview ↗

Response to Reviewer KcZc

Thank you for your insightful review and helpful suggestions. All section references below correspond to the revised manuscript (where changes are highlighted in blue).

Physical motivation and severity comparisons

What evidence supports the physical realism and severity schedules of Gen1-C?

Section 2 and Appendix B link hot pixels to threshold mismatch/low thresholds (Wang et al., 2019; Hu et al., 2021); rate shifts to threshold, bandwidth and noise changes (Hu et al.; Guo and Delbruck, 2023); altered timing to bandwidth (Hu et al.); dropout to blocked camera regions; and flood to excessive noise at low thresholds (Gallego et al., 2022). Polarity flip tests ON/OFF channel assignment.

We added a comparison in Appendix B.1 between Gen1-C and v2e (video to events, the established event-camera simulator, Hu et al., CVPRW'21) on two bright/dim sequences. Activity change is 100(N/N0-1)%, summed over pixels, polarities and bins against the ideal-v2e reference, then averaged over the two scenes.

Gen1-C corruptionS1S3S5v2e reference: tested range
Hot pixels+1.35%+43.12%+403.29%Threshold mismatch, sigma=0.01 to 0.08: +1.95% to +51.28%
Temporal shift-10.23%-25.24%-60.07%Bandwidth, 300 to 10 Hz: -0.87% to -75.60%
Spatial dropout-0.25%-1.96%-11.83%Lost activity under bandwidth limiting: -0.87% to -75.60%
Event-rate shift, under-rate0.00%-95.66%-100.00%Nominal threshold, 0.2 to 0.5: 0.00% to -71.10%
Event-rate shift, over-rate0.00%+95.66%+104.34%Nominal threshold, 0.2 to 0.05: 0.00% to +360.75%

These comparisons with an established simulator, v2e on realistic degraded recordings support the plausibility of several activity changes; notably hot-pixel S5 is stronger than the tested mismatch range.

Q1. Comparison with input, spike and ANN representations

Membranes retain subthreshold values discarded by spike counts and reflect learned input filtering. Section 4.5 adds the same MDD scoring method across representations, fitted separately on the same clean splits. At W=64:

Corruption / severityInput statisticsTrained membranes (Ours)Trained spikesANN features
Mean, S1-S5.697.785.733.530
Hot pixel, S51.0001.0001.000.747
Event flood, S5.877.885.880.469
Temporal jitter, S5.520.875.811.429
Polarity flip, S5.634.571.537.443
Event-rate shift, S5.940.946.882.576
Spatial dropout, S5.783.983.955.432
Mean, S5.792.877.845.516

Section 4.5 also includes random SNN membranes in the full five-representation comparison.

Membranes lead overall; inputs lead on polarity flip, while hot pixels and flood are tied or comparable.

Score averaging and SNN temporal dynamics

Appendix C.3 compares matched windows. Mean input/membrane/spike AUROC is .658/.738/.702 at W=1, .697/.785/.733 at W=64 and .722/.810/.751 at W=128. Score averaging alone therefore is not responsible the membrane advantage.

Appendix C.5 compares resetting PLIF state each frame with carrying it across frames, keeping architecture, weights and descriptors fixed. Each policy has its own MDD fitted on the same clean pools. State carry increases mean S5 AUROC by .060 at W=1 and .059 at W=64. The appendix gives corruption-wise gains and losses; this tests added cross-frame state separately from the original reset-based results.

Q2. Descriptor and score distributions

Appendix C.4 adds three figures: clean/S1-S5 dropout descriptors for membranes/spikes at L1/L4; final W=64 input/membrane/spike scores for jitter, dropout and polarity flip; and additional jitter spatial/dropout moment descriptors. L1/L4 are the first/fourth monitored PLIF layers.

Persistence and the averaging window

Persistence tracks recurring spatial flatness within a recording and resets at recording boundaries, as now also clarified in Section 3.2. The W-frame average mainly reduces frame-level score noise; Section 4.7 describes its averaging span. Any misleading claim of W being the only source of past information has been rephrased.

Deployment costs

Transmitting the 160 Efficient MDD descriptors in BF16 would require 320 bytes/frame, or 6.4 kB/s at 20 fps before packet overhead. The 2023 interface study demonstrates bidirectional packet exchange between SpiNNaker and an Arduino Due MCU and reports 12.32 kbytes/s communication, above our descriptor payload rate. This supports the feasibility of the proposed pipeline.

Section 4.9 and Appendix E.2 time extraction/scoring from prepared membrane samples, excluding acquisition and transfer.

Thank you again. Please let us know if you have further questions; we look forward to the discussion.

AuthorsOpenReview ↗

Response 2

Building on our response in Rebuttal 1 about Membrane access and deployment costs, the cited work supports the feasibility of transferring the compact monitoring descriptors at the required rate. We have now additionally evaluated an implementation that performs membrane pooling and Efficient MDD directly on SpiNNaker-1's on-chip ARM cores. This provides a concrete deployment alternative to the external-MCU arrangement discussed previously: descriptor extraction, persistent-state updates, and detector scoring all execute on the neuromorphic chip, removing the need to transfer membrane traces to an external processor for monitoring.

In our hardware evaluation, enabling the complete pooling and Efficient MDD pipeline increased inference time by less than 5% compared with model-only execution under the same inputs and hardware configuration. This measurement includes both pooling and scoring. Thus the on-chip implementation demonstrating modest computational overhead.

Thank you for the substantial additional experiments and clarifications, which have addressed several of my concerns. However, I remain concerned about the physical realism of the synthetic corruptions and would appreciate a more detailed breakdown of deployment costs. Overall, the additional evidence strengthens the paper, and I have decided to raise my score.

AuthorsOpenReview ↗

Response 3

Thank you for revisiting our responses and raising your score. To clarify the deployment costs, each percentage below expresses the added monitoring latency relative to model-only latency:

$

Component overhead=TcomponentTmodel only×100.\text{Component overhead} = \frac{T_{\text{component}}}{T_{\text{model only}}}\times 100.

$

ComponentComponent overhead
Membrane pooling and descriptor extraction3.41%
Efficient MDD scoring0.78%

The added time amounts up to 4.19% of the original model inference time on SpiNNaker-1. Here, membrane states are accessed locally and no off-chip membrane-trace transfer during inference

Thank you again for your time and constructive feedback.

Reviewer_55UPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MDD, a detector of event-camera input corruptions based on membrane-potential statistics in SNNs. It combines PCA residuals and distance-based deviations derived from temporal and spatial statistics, with scores calibrated using clean data. The paper introduces Gen1-C, which contains six corruption types at five severity levels, and presents baseline comparisons, component analyses, and evaluations on additional datasets and models. It also reports MCU execution times for a lightweight variant, Efficient MDD.

Critical strengths

The method improves detection performance over existing baselines. In the highest-severity Gen1-C evaluation in Table 5, MDD achieves a mean AUROC of 0.809, compared with 0.624 for the strongest baseline. Component ablations also support the utility of the proposed feature combination.

Critical weaknesses
  1. Methodological novelty is limited, and the additional benefit of spiking dynamics is insufficiently established. Statistical features and subspace- or distance-based OOD detection are established approaches. The main distinction here lies in monitoring membrane potentials and combining spatiotemporal features. Although performance gains are demonstrated, the contributions of membrane representations, spatial features, and score calibration are not sufficiently disentangled. In particular, the evaluation lacks a detector based solely on input-event statistics, while the comparison with Spike-count-kNN changes both the representation and the scoring method. Applying comparable scoring methods to input and spike statistics would help establish the additional value of accessing membrane states. Moreover, §4.8 states that the baselines are adapted to SNN representations but does not sufficiently specify their features and key settings, which limits reproducibility.

  2. The claim of low additional cost is not sufficiently validated when membrane access is included. The MCU measurements in §4.9 and Appendix E cover feature extraction and scoring from prepared membrane-potential samples. They exclude membrane acquisition and transfer, leaving the total overhead in a deployed system unverified. Efficient MDD, the variant measured on hardware, also loses substantial detection performance for some low- and moderate-severity corruptions. For example, under spatial dropout at S2 in Table 13, AUROC decreases from 0.952 for the full method to 0.493 for the efficient variant. Although these losses are reported, it remains unclear under which conditions the performance–cost benefits persist after membrane-access costs are included.

Questions for authors
  1. Which layers, features, normalization procedures, and key hyperparameters were used for each baseline in Table 5? Please specify the distance-score orientation and whether two-sided deviation calibration was applied.

  2. Do any existing experiments separate the contribution of membrane representations from those of spatial features and score calibration? If so, please identify these results and explain which conclusions they support.

  3. How would membrane potentials be acquired and transferred in the target deployment? Are measurements or estimates available for the latency and memory costs excluded from the reported timing?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 55UP

Thank you for your insightful review and helpful suggestions. All section references below correspond to the revised manuscript (where changes are highlighted in blue).

Membrane representations, spatial features and calibration

the contributions of membrane representations, spatial features, and score calibration are not sufficiently disentangled.

Our contributions include a membrane-state corruption monitor with clean-data calibration, the six-family Gen1-C benchmark, and Efficient MDD with physical MCU measurements. To clarify what drives detection performance, we added matched representation comparisons in Section 4.5 and spatial-feature and calibration controls in Appendix C.1.

As per the reviewer's suggestion, we applied the same MDD scoring method to input statistics, trained membranes and trained spikes, fitting each monitor separately on the same clean splits. Mean within-recording AUROC at W=64 is:

Corruption / severityInput statisticsTrained membranesTrained spikes
Mean, S1-S5.697.785.733
Hot pixel, S51.0001.0001.000
Event flood, S5.877.885.880
Temporal jitter, S5.520.875.811
Polarity flip, S5.634.571.537
Event-rate shift, S5.940.946.882
Spatial dropout, S5.783.983.955
Mean, S5.792.877.845

Section 4.5 also includes random membranes and ANN features.

Trained membranes lead overall, tie on hot pixels and are comparable to spikes on flood; inputs lead on polarity flip.

The component tests separate three effects:

  • Spatial features: using the same membrane representation and Mahalanobis scoring method, adding spatial features to moments improves AUROC from .648 to .706.
  • Branch calibration: with branches and two-sided transforms fixed, standardizing branch scores improves .690 to .785.
  • Two-sided scoring: flagging both unusually high and low distances reduces AUROC from .706 to .665 with the same features and scorer, showing it does not always help.

These comparisons display the gains from membrane representations, spatial features and calibration.

Baseline adaptations and settings

Thank you for this suggestion. Currently, Appendix C.2, referenced from Section 4.8, adds the baseline settings. Gaussian Mahalanobis uses one global Gaussian rather than class-specific Gaussians because frames can contain multiple classes or none. FDE-Glow adjusts coupling dimensions to the host embedding. The table lists the host features and settings.

MethodFeaturesMain settings and score
Membrane-kNNFinal PLIF membranes; mean over bins, spatial GAP; 256-D; L2-normalizedk=50; squared L2 distance to the kth neighbour
Spike-count-kNNFinal PLIF spikes; sum over bins, spatial GAP; 256-D; L2-normalizedk=50; squared L2 distance to the kth neighbour
DIMEFinal ANN GAP; 256-DUncentered SVD retaining 99% energy; distance to the fitted subspace
Gaussian MahalanobisFinal ANN GAP; 256-DOne Gaussian; Ledoit-Wolf covariance; squared Mahalanobis distance
PatchCoreANN stages 2 and 3; 384-D patch features3x3 patches, stride 1; 100,000 candidates; 1% approximate-greedy coreset; maximum patch nearest-neighbour distance
FDE-GlowFinal ANN GAP; 256-D; L2-normalized10 flow blocks; Adam, learning rate 1e-4; batch 256; one epoch; no augmentation; negative log likelihood

For every baseline, higher scores mean more anomalous input. None uses two-sided deviation calibration. Fitting uses clean data only, with seed 42.

Membrane access and deployment costs

Transmitting the 160 Efficient MDD descriptors in BF16 would require 320 bytes/frame, or 6.4 kB/s at 20 fps before packet overhead. The 2023 interface study demonstrates bidirectional packet exchange between SpiNNaker and an Arduino Due MCU and reports 12.32 kbytes/s communication, above our descriptor payload rate. This supports the feasibility of the proposed pipeline.

Efficient MDD trades sensitivity for compute savings; Appendix E.1 reports the severity-wise results, including the S2 dropout loss noted in the review. Appendix E.2 reports 56.56 KB of MCU monitor RAM, excluding the call stack and communication buffers.

Thank you again. Please let us know if you have further questions; we look forward to the discussion.

AuthorsOpenReview ↗

Response 2

Building on our response in Rebuttal 1 about Membrane access and deployment costs, the cited work supports the feasibility of transferring the compact monitoring descriptors at the required rate. We have now additionally evaluated an implementation that performs membrane pooling and Efficient MDD directly on SpiNNaker-1's on-chip ARM cores. This provides a concrete deployment alternative to the external-MCU arrangement discussed previously: descriptor extraction, persistent-state updates, and detector scoring all execute on the neuromorphic chip, removing the need to transfer membrane traces to an external processor for monitoring.

In our hardware evaluation, enabling the complete pooling and Efficient MDD pipeline increased inference time by less than 5% compared with model-only execution under the same inputs and hardware configuration. This measurement includes both pooling and scoring. Thus the on-chip implementation demonstrating modest computational overhead.

Dear Reviewer 55UP,

Baseline adaptations and settings

Following the baseline details added in Appendix C.2, we also plan to open-source the baseline implementations, along with the MDD and Efficient MDD code already included in our supplementary material.

With less than a day remaining in the discussion period, we look forward to hearing your thoughts on our responses and additional results. We hope these address your main concerns and would be happy to clarify any remaining questions.

Thank you again for your time and consideration.

Best regards,
The Authors

Comments on the paper

3 comments
AuthorsOpenReview ↗

Overall Summary Response

We thank the reviewers for their comments and the discussion so far. Below we summarize the additional experiments addressing the main concerns.

What do membrane states contribute? We compared representations using the same MDD scoring method, with each monitor fitted separately on the same clean splits. Mean within-recording AUROC across S1-S5 is:

Averaging windowInput statisticsTrained spikesTrained membranes
W=1.658.702.738
W=64.697.733.785
W=128.722.751.810

Membranes perform best across all three averaging windows, including W=1. At S5 and W=64, they improve AUROC over input statistics from .520 to .875 on temporal jitter and from .783 to .983 on spatial dropout. The expanded comparison also includes random membranes and ANN features (Section 4.5).

Separate component controls distinguish representation gains from feature and scoring choices (Appendix C.1):

Controlled changeAUROC before / after
Add spatial features to membrane moments, keeping Mahalanobis scoring.648 / .706
Standardize branch scores, keeping branches and two-sided transforms fixed.690 / .785
Apply two-sided distance scoring, keeping features and scorer fixed.706 / .665

These controls isolate the effects of spatial features, branch calibration and score direction. Appendix C.2 also documents baseline features, adaptations and fitting settings.

Does detection extend beyond Gen1-C? We evaluated additional mechanisms with host weights, descriptors, MDD parameters and calibration frozen. On seven matched 256-frame prefixes, clean task mAP was 35.03%:

Additional stress test, S5Task mAP lossMDD AUROC, W=64
Background-count noise8.94 percentage points.929
Temporal packet loss8.86 percentage points.920
Readout-throughput thinning5.79 percentage points.420

Without refitting, MDD detects background-count noise and temporal packet loss associated with approximately nine percentage points of task mAP loss. Can the monitor respond to acute failures? We evaluated a fast input alarm alongside the trailing-window membrane alarm on 50 held-out recordings:

AlarmS5 floods covered by an active warningClean alarm frames
Input, W=1100.0%3.22%
Membrane, W=6494.9%1.17%
Combined100.0%3.31%

The combined alarm covers all tested S5 floods. We also separated descriptor persistence, score averaging and cross-frame neural state, and provided window-selection guidance using clean validation data (Section 4.7 and Appendix C.7).

Does deployment require an external MCU? Following the transfer-based arrangement discussed in our earlier responses, we implemented pooling and Efficient MDD on SpiNNaker-1's on-chip ARM cores:

Pooling and scoring together add less than 5% inference overhead. Both run on the neuromorphic chip, removing the need for an external MCU or off-chip membrane-trace transfer.

As the discussion phase draws to a close, we look forward to hearing your thoughts on these additional results and are happy to address any remaining questions.

AuthorsOpenReview ↗

Request for feedback

Dear Reviewers,

Thank you for your time and thoughtful evaluations. As the discussion period is approaching its end, we would sincerely appreciate it if you could kindly review our rebuttals and engage in the discussion. We hope our responses have adequately addressed your concerns and would be grateful if you could consider them when reassessing your evaluations.

Thank you again for your valuable time and consideration.

Best regards, The Authors

Area_Chair_QkpVOpenReview ↗

Please Engage in the Author–Reviewer Discussion

Dear Reviewers,

As the author–reviewer discussion period is approaching its deadline (November 18, AoE), I would like to kindly remind you to review the authors' rebuttals and actively engage in the discussion as soon as possible.

Please address any outstanding questions or concerns and update your assessments and ratings if appropriate.

Thank you very much for your time and efforts!

Best regards, Area Chair

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.