{"id":"98e4d4f8-9a38-42d5-a986-1ff1490dcc64","arxiv_id":"2412.17654","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A directly trained spiking YOLOv5n using 3D convolutions and a temporal recurrence mechanism reports mAP within 0.001 to 0.008 of a same-architecture ANN on COCO2017 and VOC at 224x224.","lead":"A spiking neural network object detector replaces 2D convolutions with 3D convolutions over time and adds a temporal recurrence trick, reporting mAP close to an ANN baseline on VOC and COCO at low resolution. The paper argues this closes the performance gap between directly trained spiking and conventional detectors, though the absolute scores are far below typical object detection results.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 9 row 1 reports 0.551 for the 3D+recurrence+proposed-neuron T=15 config that Table 1 lists as 0.601; with no code or error bars, the claimed 0.008 ANN gap is not a stable result.","rationale":"The paper's central claim is that directly trained SNNs match ANNs on static object detection. Table 1 is the only evidence for this claim. The same configuration appearing with a much lower mAP in Table 9, combined with the absence of error bars and code, means the near-zero gap cannot be distinguished from run-to-run variation or configuration selection. The reader's chosen weakest assumption about the Section 4.5 shuffling control is a real confound for the temporal-mechanism claim, but it is secondary to the reliability of the mAP numbers themselves. I therefore keep the reader's REJECT verdict; the concern reinforces it rather than moving it.","tokens_in":12389,"tokens_out":12094,"duration_ms":113178,"concrete_test":"Obtain or reimplement the exact Table 1 configuration (YOLOv5n, Hybrid encoding, T=15, 224px, VOC and COCO, with GAC) and run at least 5 seeds, reporting mean plus/minus standard deviation; also reproduce Table 9 row 1 in the same declared configuration. If the Table 9 row 1 result (0.551) reproduces, or if the seeded mean deviates from 0.601/0.265 by more than 0.01, the claimed parity is not stable. If Table 9 was intended as a different configuration, the paper must state the difference and document why the value differs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline parity rests entirely on Table 1 (0.601 vs 0.609 on VOC, 0.265 vs 0.266 on COCO). That table is internally inconsistent with the paper's own temporal-validation table: Table 9 row 1 is labeled \"3D, recurrence, This work, T=15\" and reports mAP@0.5 = 0.551 with temporal information, while Table 1's headline row reports 0.601 and Table 4's DIRECT row reports 0.591 for the same model family. Table 9 does not state encoding, GAC, input size, or dataset split, so the reader cannot tell whether this is a different configuration; if it is the same VOC configuration, the spread (0.05) is about six times the entire claimed ANN gap (0.008). No error bars, seeds, or code are given. A second, independent weakness is the Section 4.5 shuffling control: Eq. (6)'s l_t and i_t appear to be time-indexed learnable parameters from [33], so shuffling input currents while keeping parameters at their original time indices changes the input-parameter alignment; the control therefore does not isolate temporal order. The central empirical claim is load-bearing on reproducible mAP values, and this is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes improving directly trained spiking neural networks (SNNs) for static object detection by (i) replacing 2D convolutions with 3D convolutions to mix information across the time dimension, and (ii) introducing a temporal recurrence mechanism that initializes the first membrane potential with the last time step's input current. The neuron model is a combination of a learnable-decay/input-current model from the author's prior work [33] and a direct-input neuron [34], and hybrid encoding with Gated Attention Coding is optionally employed. Using YOLOv5n as the backbone, the paper reports mAP@0.5 of 0.601 versus 0.609 on VOC and 0.265 versus 0.266 on COCO, claiming that this is the first directly trained spiking YOLO to match the same-architecture ANN on static datasets. The paper also presents ablations over the 3D convolution, recurrence, neuron model, encoding, and time steps, plus a temporal-shuffling experiment intended to show that performance gains come from temporal information processing.","tokens_in":12714,"tokens_out":7313,"duration_ms":65497,"significance":"If the reported numbers are reliable, the contribution would be a meaningful step toward closing the ANN-SNN performance gap in object detection: the gap shrinks from 0.078 to 0.008 on VOC and to 0.001 on COCO using the same architecture. The paper is also commendable for including ablated comparisons of each proposed component, for testing different model scales and input sizes, and for explicitly acknowledging the energy-efficiency cost of 3D convolutions. However, the central claim is purely empirical and is not backed by multiple seeds, error bars, or released code. Moreover, the same configuration appears to have two very different reported mAP values in Table 1 and Table 9, and the temporal-shuffling control is confounded by the time-indexed learnable parameters in Eq. (6). These issues mean the central empirical claim is not currently established.","major_comments":[{"comment":"The headline parity claim rests on Table 1, which reports mAP@0.5 = 0.601 for the proposed method on VOC with T=15, versus 0.609 for the ANN. However, Table 9 row 1, labeled '3D, recurrence, This work, T=15', reports the with-temporal-information mAP@0.5 as 0.551. The paper does not state Table 9's encoding scheme, GAC setting, input resolution, or dataset split, so the reader cannot tell whether this is a different configuration or the same VOC configuration. If it is the same VOC configuration, the 0.050 discrepancy is roughly six times the entire claimed ANN gap of 0.008. This internal inconsistency directly undermines the central empirical claim.","section":"§4.2, Table 1; §4.5, Table 9"},{"comment":"The temporal-shuffling experiment is intended to isolate the network's use of temporal order by shuffling input currents along the time dimension. But Eq. (6) defines l_t and i_t as learnable parameters with a time subscript, and Section 2.3 credits [33] with introducing per-time-step learnable membrane-potential parameters. If l_t and i_t are not shared across time, shuffling the inputs while leaving each parameter at its original time index changes the input-parameter alignment, so the resulting mAP drop could be caused by misalignment rather than by the loss of temporal order. The paper never states whether these parameters are shared or time-indexed, and therefore the mechanistic conclusion from the shuffling experiment is unsupported.","section":"§4.5, Eq. (6)"},{"comment":"All reported mAP values are single runs with no error bars, no number of seeds, and no indication of variance. The claimed ANN-SNN differences are 0.008 on VOC and 0.001 on COCO; these margins are within the range of typical random-seed variation for object detection training. Since the central claim is exactly that the SNN is comparable to the ANN, statistical reproducibility is load-bearing. The absence of code, seeds, and repeated runs leaves the central claim unsupported even apart from the Table 1/Table 9 discrepancy.","section":"§4.1, Table 1; §4.3, Tables 3-5"}],"minor_comments":[{"comment":"The Difference column lists '0.028/0.05%' for row 1; since 0.028/0.551 = 0.0508, the second entry appears to be a fraction expressed as a decimal rather than a percentage. Rows 2-6 have the same issue (e.g., 0.022/0.518 = 0.0425, not 0.042%), and the label 'Difference' should be corrected or clarified.","section":"Table 9"},{"comment":"The time step used for the input-size experiments is not stated in the text or the table caption. The 224-pixel entry of 0.567 appears to match T=4 in Table 5, but this should be explicit, especially because the preceding model-size experiment was described as using six time steps.","section":"§4.4, Table 8"},{"comment":"The statement 'The code for this job can be contacted via email' is not a usable reproducibility statement; the authors should provide a public repository or a clear release plan.","section":"Abstract"}],"recommendation":"reject","confidential_remarks":"The central result is not reproducible as written: the Table 1 vs. Table 9 conflict and the absence of seeds/error bars/code make the headline parity claim unsupported. The self-citation to [33] also deserves scrutiny because the learnable neuron and hybrid encoding are central to the method but are not independently verified here. If the authors can clarify the Table 9 configuration, provide code and multiple-seed results, and fix the shuffling control, a substantially revised submission might be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the whole story is the claimed near-zero gap between a directly trained SNN and its ANN twin on VOC/COCO. The paper's own Table 9 puts a 0.551 next to what looks like the same configuration that Table 1 lists as 0.601. That is a five-point spread, roughly six times the entire claimed gap. No error bars, no seeds, no code (just \"contact via email\"). I don't think the central claim is established.\n\nWhat is genuinely worth something: swapping 2D convolutions for 3D over the time axis is a simple and reasonable way to let temporal channels talk to each other, and the recurrence initialization (feed the last input current into the initial membrane potential) is a neat trick that breaks one-way temporal flow. The paper also does honest legwork: ablations, time-step sweeps, model-scale and input-size generalization, and it openly notes that 3D convs raise parameter count and compute, cutting into the SNN energy advantage. That is more than many SNN papers deliver.\n\nThe soft spots are not subtle. First is the table inconsistency above. Table 4's direct-coding row at T=15 is 0.591, Table 5 repeats 0.591, Table 9 row 1 says 0.551 for what looks like the same model family. The author never specifies encoding, GAC, or input size for Table 9, so maybe it is a different config, but the burden is on them to say so. Second, the shuffling control in Section 4.5 is meant to prove temporal order matters, but Eq. (6) has time-indexed learnable l_t and i_t, and the paper never states these are shared across time. If they are not, shuffling input currents while keeping parameters in their original time slots changes the input-parameter alignment, so the mAP drop is not clean evidence for temporal processing. Third, the \"first application\" overclaim ignores that EMS-YOLO and SFOD already narrowed the gap; this may be the smallest gap yet, but that is a different statement.\n\nOn balance, if Table 9's 0.551 is the real number, this is a modest improvement, not a breakthrough. The idea is worth pursuing, and the authors know the literature. I'd send it to peer review because the question matters and the method is cheap to reproduce, but I'd insist the authors reconcile every table, report seeds and error bars, and release code before any acceptance. As it stands, a serious referee would be doing the author's job for them.","headline":"The claimed near-zero SNN/ANN gap rests on mAP numbers that don't survive contact with the paper's own tables; the 3D-conv idea is reasonable, but the evidence is not there yet.","tokens_in":13254,"tokens_out":2341,"would_cite":false,"duration_ms":21871,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Directly trained spiking YOLO can now match ANN accuracy on static object detection.","keywords":["spiking neural networks","3D convolution","object detection","temporal information processing","direct training","membrane potential recurrence","YOLO","static image datasets"],"falsifier":"Train the same network while shuffling only the input currents but keeping each time step's learnable constants aligned with their original time index, and compare the mAP drop; if the drop disappears or reverses under realignment, the temporal-information interpretation is falsified. Alternatively, retrain the full model with the 3D convolutions widened in the channel dimension instead of the time dimension to match the parameter count: if the accuracy gap versus the ANN persists, the gain is parameter-driven rather than temporal.","tokens_in":12177,"feed_emoji":"⚡","tokens_out":8895,"duration_ms":65842,"temperature":0.7,"pith_summary":"Spiking neural networks (SNNs) promise low-energy, event-driven computation, but directly trained SNNs have lagged far behind ordinary artificial neural networks (ANNs) on frame-based object detection. This paper claims to close that gap by giving the network a stronger temporal processing pipeline: replacing 2D convolutions with 3D convolutions so information flows across time steps inside the convolution itself, and adding a recurrence mechanism that initializes the membrane potential with the last time step's input. On the COCO and VOC datasets, the resulting spiking YOLOv5n reaches mAP@0.5 within 0.008 (VOC) and 0.001 (COCO) of the same-architecture ANN, and exceeds it when trained longer on VOC. If true, this removes a major obstacle to deploying directly trained SNNs on standard static-vision workloads while keeping their event-driven, energy-efficient character.","feed_headline":"Spiking YOLO matches ANN accuracy on static object detection","feed_subtitle":"3D convolutions and a temporal recurrence loop bring COCO mAP to within 0.001 of the ANN baseline.","key_machinery":"The load-bearing machinery is a pair of changes to the standard SNN pipeline. First, the 2D convolution kernel, applied independently to each time step's feature map with the same weights, is replaced by a 3D convolution kernel that also convolves along the time axis, directly mixing information from adjacent time steps. Second, the spiking neuron's membrane potential is no longer initialized to zero: after a backward pass, the input current of the last time step seeds the membrane potential at the first time step, so temporal information flows in a loop rather than only forward in time. The neuron update is $V[t] = \\ell_t V[t-1] + i_t X[t]$, with learnable per-time-step decay $\\ell_t$ and input constants $i_t$, inherited from the hybrid-coding neuron and the leakage/threshold scheme the paper combines. These components let the network treat each time step's spikes as part of a spatiotemporal signal rather than as repeated presentations of the same static image.","core_discovery":"The paper's central claim is that the long-standing performance gap between directly trained SNNs and ANNs on static object detection is not a fundamental limitation of spiking neurons, but a deficit in how temporal information is communicated. The author proposes two interventions: replace the 2D convolution kernels, which process each time step independently with shared weights, with 3D convolution kernels that slide over the time dimension as well; and modify the spiking neuron so that the membrane potential at the first time step is initialized with the input current of the last time step, creating a recurrence loop. The neuron model combines learnable per-time-step decay and input constants with the leakage and threshold design of the cited hybrid-coding and direct-encoding neurons. The reported experiments show the resulting network, built on YOLOv5n with SEW-ResNet connections, reaching mAP@0.5 of 0.601 on VOC against an ANN score of 0.609, and 0.265 on COCO against 0.266, at 15 time steps; with 400 training epochs the VOC score rises to 0.618, surpassing the ANN. The author frames this as the first successful application of a directly trained spiking YOLO to a static dataset at ANN-comparable accuracy.","pith_inferences":["The same 3D-convolution-plus-recurrence recipe could be applied to other spiking backbones, such as spiking ResNets or transformers, where per-time-step weight sharing limits temporal communication; the reported scaling trend from YOLOv5n to YOLOv5s suggests the benefit may grow with model capacity.","If the temporal mechanisms are truly responsible for the gain, the method should also improve SNN performance on event-based datasets, where temporal order carries object motion; the paper does not test this, but its own temporal-dependence results predict it.","The parameter-count increase from 3D convolutions trades away part of the energy-efficiency advantage the paper motivates, so a complete comparison would require an energy or latency measurement that accounts for the extra multiply-accumulate operations.","Because the shuffling control is potentially confounded by time-indexed learnable constants, the temporal-information interpretation is the part of the paper most worth probing; an ablation with shared, time-independent decay and input constants would cleanly separate the two mechanisms."],"forward_implications":["Directly trained SNNs can serve as drop-in replacements for ANNs on frame-based object detection at comparable accuracy, at least at the YOLOv5n scale with the reported 15 time steps and 300 to 400 training epochs.","Longer training from 300 to 400 epochs raised the SNN's VOC mAP@0.5 from 0.601 to 0.618, overtaking the same-architecture ANN, which suggests SNNs need more epochs to converge.","Ablation results attribute most of the gain to the 3D convolution (removing it drops mAP@0.5 from 0.567 to 0.534 at 4 time steps) and to the recurrence mechanism (removing it drops mAP to 0.552), so both components are needed for the full effect.","Scaling the backbone from YOLOv5n to YOLOv5s narrowed the SNN-to-ANN gap from 0.028 to 0.011, suggesting the method transfers to larger models.","The temporal-shuffling experiments indicate that larger time steps and 3D convolutions make the network more dependent on temporal order, while a traditional 2D-convolution LIF network performs slightly better when temporal order is destroyed."],"supporting_citations":[{"why":"supplies the hybrid coding and learnable per-time-step membrane potential and input parameters used in the neuron update.","marker":"[33]"},{"why":"supplies the direct input encoding with leakage and threshold optimization that the neuron model builds on.","marker":"[34]"},{"why":"provides the directly trained spiking object detector and the reported SNN-to-ANN gap (0.501 vs 0.565 on COCO) that the paper aims to close.","marker":"[26]"},{"why":"provides the SEW-ResNet residual connection that lets the spiking YOLO backbone be trained at depth.","marker":"[16]"},{"why":"introduces spatio-temporal backpropagation with surrogate gradients, the training method that makes direct training possible.","marker":"[10]"},{"why":"introduces gated attention coding, which the paper integrates into its hybrid encoding and ablates.","marker":"[32]"}],"fun_headline_variants":["3D Convs and Recurrence Give Spiking YOLO ANN-Level Accuracy","Spiking YOLO Matches ANN Detection via Temporal 3D Convolutions","Temporal Recurrence Narrows SNN-ANN Gap on Static Object Detection","Directly Trained Spiking YOLO Achieves ANN Parity on COCO and VOC","Enhancing Temporal Processing: Spiking Nets Reach ANN Performance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's evidence that its gains come specifically from temporal processing rests on the shuffling experiment in Section 4.5, which assumes that shuffling input currents along time removes temporal order without otherwise changing the task; if the learnable per-time-step neuron parameters are time-indexed, the shuffle also breaks their alignment with the inputs.","fun_headline_variants_meta":{"raw":{"variants":["3D Convs and Recurrence Give Spiking YOLO ANN-Level Accuracy","Spiking YOLO Matches ANN Detection via Temporal 3D Convolutions","Temporal Recurrence Narrows SNN-ANN Gap on Static Object Detection","Directly Trained Spiking YOLO Achieves ANN Parity on COCO and VOC","Enhancing Temporal Processing: Spiking Nets Reach ANN Performance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000184,"raw_usage":{"total_tokens":1396,"prompt_tokens":1099,"completion_tokens":297,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":715,"completion_tokens_details":{"reasoning_tokens":192}},"tokens_in":715,"tokens_out":297,"duration_ms":3751,"temperature":1.0,"reasoning_tokens":192,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:18:54.372169+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same network while shuffling only the input currents but keeping each time step's learnable constants aligned with their original time index, and compare the mAP drop; if the drop disappears or reverses under realignment, the temporal-information interpretation is falsified. Alternatively, retrain the full model with the 3D convolutions widened in the channel dimension instead of the time dimension to match the parameter count: if the accuracy gap versus the ANN persists, the gain is parameter-driven rather than temporal.","supporting_citations":[{"cited_title":"Optimizing Spatio-Temporal Information Processing in Spiking Neural Networks via Unconstrained Leaky Integrate-and-Fire Neurons and Hybrid Coding","cited_arxiv_id":"2408.12407","evidence_quote":"supplies the hybrid coding and learnable per-time-step membrane potential and input parameters used in the neuron update."},{"cited_title":"Deep Directly-Trained Spiking Neural Networks for Object Detection,","cited_arxiv_id":null,"evidence_quote":"provides the directly trained spiking object detector and the reported SNN-to-ANN gap (0.501 vs 0.565 on COCO) that the paper aims to close."},{"cited_title":"Deep Residual Learning in Spiking Neural Networks,","cited_arxiv_id":null,"evidence_quote":"provides the SEW-ResNet residual connection that lets the spiking YOLO backbone be trained at depth."},{"cited_title":"Gated Attention Coding for Training High-performance and Efficient Spiking Neural Networks","cited_arxiv_id":"2308.06582","evidence_quote":"introduces gated attention coding, which the paper integrates into its hybrid encoding and ablates."}],"review_version":1}