{"id":"0632c804-25be-4f9f-948e-8fe5cc28f5ae","arxiv_id":"2607.26703","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Sequence-aware SNN training that preserves membrane potentials across multi-label event intervals improves Gen1 detection mAP by about 2–3.5 points over single-interval reset training.","lead":"Keeping a spiking detector’s membrane state alive across multi-label event sequences raises Gen1 mAP from 23.38 to 25.30 (26.88 with augmentation) under a fixed SSD Spiking DenseNet. The result isolates sequence-aware training as a cheap complementary lever for continuous event-camera detection.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"The gain may not isolate membrane-state preservation; the single-interval baseline is a reimplementation that already exceeds the cited ODSNN number under unmatched Δt.","rationale":"The paper’s strongest claim is a clean isolation claim on a fixed architecture. The reader correctly flags that isolation as the soft spot: unmatched baseline numbers, fixed unablated Δt/T, heuristic full-sequence splits, single-run checkpoints, and no variance. That is the load-bearing concern for whether sequence-aware state preservation is what the numbers show. No stronger internal inconsistency appears; energy estimates and the 40 Hz claim are secondary and theoretical. Verdict stays CONDITIONAL pending the matched reset-vs-preserve multi-seed control (and ideally T/Δt sweeps). I agree with the reader rather than upgrading or rejecting.","tokens_in":16999,"tokens_out":551,"duration_ms":11327,"concrete_test":"Re-train strain=1 and strain=5 under identical Δt=125 ms, T=5, AdamW, batch size, 110-epoch cosine schedule, and NMS, differing only in whether membrane potentials are zeroed after each interval; report mean±std mAP over ≥3 seeds on the same full-sequence test split. If the mean gap falls below ~1 mAP or loses significance, the headline attribution to state preservation weakens.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim attributes the mAP lift (23.38 → 25.30) to sequence-aware non-reset training that keeps membrane potentials across multi-label intervals on a fixed SSD Spiking DenseNet. That attribution is least secure because the strain=1 control is not a matched reproduction of Cordone et al.: Table 3 lists ODSNN at 18.9 mAP with Δt=100 ms, while the paper’s own single-interval run reaches 23.38 at Δt=125 ms / T=5 with no seeds, no error bars, and no ablation that holds Δt, sampling grid, optimizer, and checkpointing fixed while only toggling cross-interval state reset. Sequence samples are also built every 125 ms on labels whose minimum gap is 250 ms, so strain=5 necessarily mixes labeled and unlabeled intervals and longer unrolled BPTT; without a same-Δt reset-vs-preserve head-to-head, part of the gain could be protocol/optimization rather than evolving neural state.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes Sequence-SOD, a sequence-aware training and evaluation protocol for fully spiking SSD-style DenseNet object detectors on event cameras. Instead of processing isolated event intervals and resetting membrane potentials after each prediction, the method feeds consecutive 125 ms intervals (each discretized into T=5 steps of 25 ms), preserves PLIF membrane state across intervals within a sequence, and applies detection loss only on labeled intervals after averaging outputs over the internal steps. On Gen1, training with strain=5 raises mAP from 23.38 (strain=1) to 25.30 without augmentation and 26.88 with geometric event augmentation, with a claimed theoretical 40 Hz prediction rate and a large estimated energy advantage versus an equivalent ANN from measured spike rates.","tokens_in":17259,"tokens_out":1319,"duration_ms":38531,"significance":"If the gains are cleanly due to sequence-aware stateful training, the work is a useful complementary contribution: it isolates a training/evaluation protocol rather than a new backbone, and shows that keeping leaky membrane memory across multi-label event streams improves detection on a fixed fully spiking detector. The internal strain=1 vs strain=5 comparison, augmentation table, spike-rate figure, and CMOS-style energy accounting are concrete and reproducible in principle. The absolute accuracy remains well below strong ANN/hybrid event detectors, so impact is mainly as evidence that SNN detectors should be trained and evaluated as continuous-state systems, not as reset snapshot classifiers.","major_comments":[{"comment":"Table 1 attributes the mAP lift (23.38→25.30) to sequence-aware training with preserved membrane state, but strain=1 vs strain=5 jointly changes several factors: unrolled BPTT length, number of labeled supervision points, presence of unlabeled forward-only intervals, and whether state is reset between intervals. A load-bearing control is missing: train and test on the same multi-interval sequences while only toggling cross-interval reset vs preserve (and, ideally, a same-Δt matched reproduction of Cordone et al.). Without that, the central causal claim that evolving neural state—not longer optimization/unrolling—drives the gain is not isolated.","section":"Table 1; Method (Sequential Event Data Processing); Experiments (Sequence Length)"},{"comment":"All headline numbers appear to be single-run best checkpoints after 110 epochs with no seeds, error bars, or repeated trials (Tables 1–3). The no-augmentation gain is about +1.9 mAP and several augmentation deltas are <1 mAP. For a result of this magnitude, at least mean±std over multiple seeds (or a clear statement that variance was checked) is needed before the improvement can be treated as reliable.","section":"Tables 1–3; Implementation"},{"comment":"The protocol fixes Δt=125 ms and T=5 everywhere, builds sequences every 125 ms although the minimum label gap is 250 ms, and defines “full” test sequences by a 200-interval (25 s) gap heuristic, yet reports no sensitivity to these choices. The Discussion acknowledges this, but the 40 Hz claim and the sequence-construction grid are part of the method’s practical story; at minimum, one ablation over T or Δt (or intermediate strain) is needed to show the gain is not an artifact of this particular sampling alignment.","section":"Setup; Discussion and Limitations; Table 1"},{"comment":"Table 3 lists ODSNN at 18.9 mAP (Δt=100 ms) while the authors’ own single-interval baseline reaches 23.38 at Δt=125 ms under the “same” SSD Spiking DenseNet. The paper correctly emphasizes the internal control, but the benchmark table still invites a direct comparison that is unmatched in interval length and reimplementation details. Either align Δt/training recipe with Cordone et al. for the strain=1 row or clearly separate “reproduced baseline under our protocol” from the cited ODSNN number so the sequence-aware gain is not read against an understated external baseline.","section":"Table 3; Benchmark Comparison"}],"minor_comments":[{"comment":"Abstract and several places write “withevent-data” / missing spaces and hyphenation inconsistencies (“hightemporal”, “Articial Neural Network”). A full copy-edit pass is needed.","section":"Abstract; Introduction"},{"comment":"Figure 2 caption refers to “strain=5” with a missing subscript formatting (“strain =5”), and the schematic would benefit from explicitly marking which intervals receive loss versus state-only forward passes.","section":"Fig. 2"},{"comment":"Equation (1) uses Sout[t−1]θ for reset; briefly state whether a soft/hard reset and subtractive vs zero reset are used in the PLIF implementation, since this affects cross-interval memory.","section":"Eq. (1); Network Architecture"},{"comment":"Energy section: clarify whether spike rates in Fig. 4/Table 4 are measured under strain=5 full-sequence evaluation or stest=1, and whether BN-folded inference is assumed in the FLOP counts.","section":"Energy Consumption; Table 4; Fig. 4"},{"comment":"Related Work is thorough but long relative to the empirical novelty; tightening hybrid vs fully-spiking distinctions would help the reader reach the contribution faster.","section":"Related Work"}],"recommendation":"major_revision","confidential_remarks":"Fit is reasonable for a cognitive-computation / neuromorphic venue: the novelty is a training protocol, not SOTA accuracy. I would not reject for being below RVT-level mAP if the reset-vs-preserve control and multi-seed stats are added. The skeptic’s concern about confounding is real and should be required; if authors can show preserve ≫ reset on identical sequences, the paper becomes clearly publishable."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful bit is narrow and real. They freeze Cordone’s SSD Spiking DenseNet and only change how training/eval treat time: multi-interval sequences, labels at several timestamps, membrane state kept inside a sequence and reset only between sequences. On Gen1 that moves mAP from 23.38 (strain=1) to 25.30, and 26.88 with simple event aug. Table 1 is consistent across test lengths; pedestrians gain most; Fig. 3 matches the story. Energy table and spike-rate plot are standard and coherent. They are honest that this is complementary to architecture work and that absolute numbers still lag strong recurrent/transformer event detectors.\n\nWhat is new is the isolation, not a new backbone. Related ANN streaming work already cares about temporal memory; some hybrid/SNN papers touch longer inputs. Here the claim is specifically that fully spiking detectors underuse continuous streams when they reset after each labeled window, and that sequence-aware non-reset training helps under a fixed architecture. That is a fair, bounded methods point.\n\nSoft spots, in proportion. The stress-test is partly right: their strain=1 run (23.38 @ 125 ms) is not a matched reproduction of the cited ODSNN 18.9 @ 100 ms, so cross-paper baseline comparison is messy. Inside the paper, though, strain=5 vs strain=1 shares architecture, Δt, T, and optimizer, so the directional gain is still supported—just not a pure “toggle only the reset flag” ablation, and with no seeds or error bars. Fixed T=5 / Δt=125 ms, the 125 ms sampling on 250 ms min label gaps, single-run checkpoints, Gen1-only, one architecture, and no code are real limits they partly own in the discussion. None of that invents the result; it caps how far you should generalize it.\n\nMath and citations look fine for this venue class: LIF/PLIF, surrogate grads, standard mAP, sensible related-work split between ANN temporal detectors and SNN detectors. No circular metric games.\n\nWho it is for: people training fully spiking event detectors who still reset per window. Not for anyone chasing SOTA Gen1 numbers. I would send it to peer review; it is clear enough and evidentially sharp enough for referees to demand seeds, a matched reset control, and maybe one more dataset or architecture. Worth a skim if you work on SNN detection protocols; skip if you only care about leaderboard event detection.","headline":"Modest but clean protocol result: keeping SNN membrane state across multi-label event intervals beats single-interval reset on the same DenseNet-SSD, without closing the big accuracy gap.","tokens_in":17926,"tokens_out":626,"would_cite":false,"duration_ms":20187,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Keeping spiking neurons’ memory alive across event intervals improves detection without changing the network.","keywords":["Event camera","Spiking neural network","Object detection","Event-based vision","Membrane potential","Sequence-aware training","Automotive perception"],"falsifier":"Train the identical architecture once with membrane reset after every interval and once with state preserved across multi-label sequences of the same total length and labels, on the same Gen1 splits with matched hyperparameters, and check whether the mAP gap remains.","tokens_in":17851,"feed_emoji":"⚡","tokens_out":817,"duration_ms":34401,"temperature":0.7,"pith_summary":"Event cameras send sparse, timed spikes when brightness changes, and spiking neural networks match that signal because each neuron holds a fading internal voltage. Most spiking object detectors still cut the stream into short windows, make one prediction, and wipe that internal state. This paper shows that training and running the same detector on longer sequences—leaving membrane voltages intact between windows and resetting only between independent sequences—raises detection scores on automotive event data. The gain comes from forcing the network to keep a useful running memory of recent evidence, not from a new architecture. A reader should care because continuous streams are how these sensors are actually used, and the method keeps the energy advantage of sparse spikes while using more of the temporal signal the camera already provides.","feed_headline":"Spiking detectors gain when memory is not wiped each window","feed_subtitle":"Same network, longer event sequences: higher Gen1 mAP and a 40 Hz prediction path","key_machinery":"Sequence-aware training and evaluation: events are accumulated into short intervals, discretized into time steps, fed sequentially to the spiking detector, and membrane potentials are carried forward across intervals inside a sequence so detection is driven by an evolving neural state instead of independently reset windows.","core_discovery":"Under a fixed SSD-style Spiking DenseNet, sequence-aware training and evaluation that preserves membrane potentials across consecutive event intervals with multiple label timestamps improves Gen1 mAP from 23.38 for single-interval training to 25.30 without augmentation and 26.88 with event-data augmentation, while supporting a theoretical 40 Hz prediction rate from 25 ms steps.","pith_inferences":["The same state-preservation recipe could be applied to newer spiking YOLO- or Transformer-style detectors without redesigning their heads.","Once sequences stretch beyond a few hundred milliseconds, learnable or input-dependent membrane decay may matter as much as sequence length itself.","Closing more of the gap to strong recurrent non-spiking event detectors likely needs truncated backprop through time at larger sequence scales.","The practical payoff is largest if the 40 Hz path is realized on neuromorphic hardware on true continuous streams, not only GPU-simulated steps."],"forward_implications":["Sequence-aware training is a complementary lever to designing new SNN detector architectures.","The same model can support a theoretical 40 Hz box stream from 25 ms steps by sliding accumulation over carried membrane state.","Sparse-spiking energy advantage versus an equivalent dense network is retained under the sequence setting.","Unlabeled intervals can still update membrane state and improve later labeled predictions.","Models trained on longer sequences hold up better when tested on continuous full-length streams than single-interval-trained models."],"fun_headline_variants":["Sequence-aware SNNs lift Gen1 mAP by keeping membrane state alive","Preserve spikes across intervals: 23.38 to 26.88 mAP on Gen1","SSD Spiking DenseNet hits 40 Hz with multi-label event sequences","Don't reset the network: sequence training boosts event detection","Longer event sequences, same SNN: higher mAP without wiping memory"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The measured gain is cleanly caused by keeping the membrane state alive across intervals, rather than by other fixed choices such as interval length, how sequences are cut from the data, or mismatches with the original single-interval baseline setup.","fun_headline_variants_meta":{"raw":{"variants":["Sequence-aware SNNs lift Gen1 mAP by keeping membrane state alive","Preserve spikes across intervals: 23.38 to 26.88 mAP on Gen1","SSD Spiking DenseNet hits 40 Hz with multi-label event sequences","Don't reset the network: sequence training boosts event detection","Longer event sequences, same SNN: higher mAP without wiping memory"]},"model":"grok-4.5","effort":"low","cost_usd":0.003626,"raw_usage":{"total_tokens":1180,"prompt_tokens":809,"num_sources_used":0,"completion_tokens":84,"cost_in_usd_ticks":36264000,"prompt_tokens_details":{"text_tokens":809,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":287,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":809,"tokens_out":84,"duration_ms":6086,"temperature":1.0,"reasoning_tokens":287,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-30T23:31:03.466453+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train the identical architecture once with membrane reset after every interval and once with state preserved across multi-label sequences of the same total length and labels, on the same Gen1 splits with matched hyperparameters, and check whether the mAP gap remains.","supporting_citations":[],"review_version":1}