{"id":"c70286a4-06db-4fa5-9997-fdbebe5de8bd","arxiv_id":"2508.20392","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Binary temporal coding in a new spiking neuron (tdIF) plus delay-spike achieves strong ANN-to-SNN conversion results on detection and lane tasks, but the reported five-step latency ignores the depth-multiplied pipeline delay.","lead":"This paper proposes a temporal-dependent spiking neuron (tdIF) that encodes values through spike timing, together with a delay-spike rule, to convert trained neural networks into spiking versions for object and lane detection. The authors report strong detection accuracy in few time-steps, but the actual end-to-end latency is multiplied by the network depth because each layer waits a full window before firing.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own Section 5.3 formula T + n x T_delay contradicts the abstract's 'within 5 time-steps': with T_delay = T in all experiments, YOLOv3(ResNet34) needs hundreds of end-to-end time-steps.","rationale":"The reader's REJECT is well-founded. The most load-bearing point is the one the reader puts first: the 'ultra-low latency within 5 time-steps' claim is contradicted by the paper's own latency formula in Section 5.3. With T_delay = T (Section 6.1), a single image requires T + n x T_delay time-steps; for YOLOv3(ResNet34) this is hundreds of steps. Pipelining (Figure 5) only hides this cost across concurrent images; it does not reduce single-image latency. Thus Table 1's 'time-steps: 5' column does not express the quantity promised in the abstract, and the comparison with Fast-SNN's 7 time-steps is misleading. A single instrumentation test on the released code would settle this. I also note the reader's second concern: Table 6 does not reproduce Table 1 entries (for example, VOC AP@0.50 at T = 7 is 78.80, while no quantized-step row in Table 6 gives that value), and Eq. 12 is a definitional identity rather than a proof of spiking dynamics; both reinforce the REJECT, but the latency accounting is the clearest single failure. The objection is to the argument and reporting, not to the authors; the central claims are simply not supported by the manuscript's own equations.","tokens_in":21428,"tokens_out":9224,"duration_ms":88769,"concrete_test":"Instrument the released repository to run a single-image forward pass of YOLOv3(ResNet34) with T = 5 and T_delay = 5, logging the time-step at which the final detection head first produces a non-zero output. If that time-step is approximately T + n x T_delay (hundreds) rather than 5, the 'ultra-low latency within 5 time-steps' claim in the abstract and Table 1 is not the end-to-end latency and must be revised or redefined.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is 'ultra-low latency (within 5 time-steps)'. The paper's own accounting in Section 5.3 says that 'for an SNN with n layers, the total inference time-steps are T + n x T_delay', and Section 6.1 states that all results use T_delay = T. Therefore a single image through YOLOv3(ResNet34) does not finish in 5 time-steps: with roughly 50 spiking layers it needs about 5 + 50 x 5 = 255 steps, and even a much smaller network needs more than 5. The 'time-steps: 5' column in Table 1 is a per-layer coding parameter, not the end-to-end latency promised in the abstract and conclusion. The inter-layer pipelining in Figure 5 lets several images be processed concurrently and improves throughput, but it does not shorten the latency of any one image from input to final output; the first output still arrives after T + n x T_delay steps. The comparison with Fast-SNN at 7 time-steps and Spiking-YOLO at 2000 is therefore not on the same metric unless the per-sample latency is reported. This is a load-bearing mismatch between the headline claim and the method's stated cost.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an ANN-to-SNN conversion method for object detection and lane detection, built on two components: a delay-spike scheme that accumulates input for T_delay time-steps before firing in order to suppress residual membrane potential, and a temporal-dependent IF (tdIF) neuron that weights inputs, thresholds, and decoding by powers c[t]=2^{T-t} so that time-steps carry binary positional information. The source ANN is trained with a quant-clip activation function (Eq. 11) and then converted to the SNN. The paper reports state-of-the-art results on PASCAL VOC, MS COCO, Tusimple and CULane at '5 time-steps', together with an instruction-level energy analysis on a digital neuromorphic processor model.","tokens_in":21553,"tokens_out":7065,"duration_ms":64721,"significance":"If the claims were valid, the combination of temporal binary coding and residual-potential suppression would be a meaningful contribution to low-latency ANN-SNN conversion for regression-heavy vision tasks. The paper provides a public code link, reports per-instruction energy costs for IF versus tdIF neurons, and evaluates on four datasets with two architectures each, which is more than many conversion papers provide. However, the headline 'ultra-low latency within 5 time-steps' is contradicted by the paper's own latency formula, and the zero-conversion-error proof is an identity rather than a verified property of the combined tdIF/delay-spike dynamics. These issues bear directly on the central contribution.","major_comments":[{"comment":"The per-sample latency claim is contradicted by the paper's own accounting. Section 5.3 states that 'for an SNN with n layers, the total inference time-steps are T + n × T_delay', and Section 6.1 states that all experiments use T_delay = T. Therefore a single image requires (n+1)T time-steps from input to output, not T steps. For YOLOv3(ResNet34) at T=5, this is hundreds of time-steps even with a modest layer count. The pipelining in Figure 5 increases throughput but does not shorten the latency of any one image. The 'time-steps: 5' columns in Tables 1 and 2 are per-layer coding parameters, not end-to-end latency, so the abstract's 'within 5 time-steps' and the comparisons with Fast-SNN (7 time-steps) and Spiking-YOLO (2000 time-steps) are not made on the same metric.","section":"§5.3 and §6.1"},{"comment":"The proof that the conversion error is zero is a definitional identity, not a property of the proposed neurons. Substituting T = L and theta^l = lambda^l into Eq. (12) makes the two clipped-floor expressions identical by construction; no SNN dynamics, delay-spike rule, or tdIF mechanism is used in the argument. Moreover, the reported experiments do not satisfy the stated condition: Table 6 trains the quant-clip ANN with quantized time-steps L in {16, 32, 64, 128} and runs the SNN with T in {3, 4, 5, 7, 8, 10}, so T ≠ L for essentially all headline results. The formal zero-error guarantee therefore does not apply to any of the numbers in Tables 1 and 2.","section":"§5.1, Eq. (12)"},{"comment":"The combined tdIF plus delay-spike dynamics are never formally specified. Algorithm 1 describes delay-spike for the standard IF update of Eqs. (2)-(4), with no c[t] weighting, while Eqs. (17)-(19) define tdIF without any explicit delay-spike stage. The manuscript does not state whether the accumulation and firing stages of Algorithm 1 use c[t]W^l and c[t]theta^l, nor how T_delay interacts with the time-dependent threshold. Consequently, the assertion in Section 5.4 that r^l = sum_t c[t]S^l[t]/T' 'can also be mapped to a^l of the ANN' is not a derivation; Eq. (20) alone does not establish that the converted SNN reproduces the quant-clip activation.","section":"§5.2 and §5.4"},{"comment":"Table 6 contains non-monotonic results that contradict the paper's error analysis and its claim in Section 6.1.2 that 'the model achieves consistent accuracy when the time-steps extends beyond 5'. For ResNet34+Condlane on Tusimple with quantized time-steps 16, the reported Acc is 93.55 at T=4, 87.45 at T=5, 69.67 at T=7, and 68.54 at T=8. Table 1 similarly shows YOLOv3(ResNet34) at T=7 exceeding T=8 on VOC (78.80 vs. 78.46). These patterns are unexplained and undermine the claim that more inference time-steps monotonically reduce quantization error; they require either a correction or a detailed explanation.","section":"Table 6"}],"minor_comments":[{"comment":"The summation index in Eq. (5) is written as PT_{n=1}; it should be sum_{t=1}^T.","section":"§4, Eq. (5)"},{"comment":"The loop indices in Algorithm 1 are inconsistent when T_delay = T: in stage 1 the condition t > T_delay is never satisfied, and stage 2 begins at t = T - T_delay = 0, which is outside the valid time-step range.","section":"Algorithm 1"},{"comment":"Table 6 would benefit from a caption that clearly separates the quant-clip training time-step L ('Quantized time-steps') from the SNN inference time-step T, since the text's phrase 'the same time-steps' is ambiguous here.","section":"Appendix .2, Table 6"},{"comment":"MS COCO 2017 is cited as (Wang et al., 2023) in Section 6; the dataset itself should be cited as (Lin et al., 2014).","section":"§6, datasets"},{"comment":"There is a typo 'OCFS' in Section 6.1.2; it should read 'QCFS'.","section":"§6.1.2"},{"comment":"The caption of Figure 7 contains unexplained numeric strings such as 'voc 72，3940，3936，3916，3548，2608，768'; these should be clarified or removed.","section":"Figure 7"},{"comment":"The statement that tdIF consumes 'about 2 to 3 times more energy' is followed by Table 5 showing 3.4× for YOLOv3(ResNet34); the text should state the range accurately or explain why 3.4× is treated as consistent with 'same order of magnitude'.","section":"§6.3.2"}],"recommendation":"reject","confidential_remarks":"The paper's headline claim is stronger than the evidence supports. The latency contradiction and the tautological zero-error proof are load-bearing; the combined tdIF/delay-spike algorithm is also underspecified. If the authors can correct the latency accounting, formalize the combined dynamics, and explain the non-monotonic tables, the empirical results could support a future resubmission, but in the present form I cannot recommend publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper has a genuine algorithmic idea—binary-weighted temporal coding for ANN-SNN conversion—but the headline 'within 5 time-steps' is not the end-to-end latency by the authors' own accounting. That's not a nitpick; it's the central claim.\n\nWhat's new: the tdIF neuron with power-of-two temporal weighting c[t]=2^{T-t} and per-step thresholds, plus the delay-spike scheme to reset residual potential. That combination is new enough, and the authors push it through object detection and lane detection with ablations and an energy table on digital neuromorphic instructions. The lane detection comparison to QCFS is a reasonable first step for that task.\n\nThe main problem is the latency metric. Section 5.3 says total inference steps = T + n*T_delay; Section 6.1 says T_delay=T in all experiments. So a YOLOv3 with ~50 spiking layers runs in ~255 steps, not 5. The abstract's 'within 5 time-steps' is per-layer coding, not per-image latency. Inter-sample pipelining improves throughput but doesn't reduce an individual image's latency. That makes the comparison against Fast-SNN (7 steps) and Spiking-YOLO (2000) apples-to-oranges.\n\nSecond, the exactness proof in Eq. 12 is a tautology: setting T=L and θ^l=λ^l makes both sides the same formula. It doesn't show the spiking dynamics reproduce the quant-clip activation; combined with delay-spike the behavior is only sketched graphically.\n\nThird, the reproducibility issue the reader flagged is real. Table 6 reports AP@0.50 for YOLOv3(ResNet34) at T=5 as 74.42 (L=32), 77.14 (L=64), 77.46 (L=128), etc., while Table 1 says 74.41 without disclosing L; at T=7, Table 1 says 78.80, which doesn't appear in any L row of Table 6. That suggests L was selected per result or there's a numerical mismatch; either way, the headline numbers are not reproducible from the supplement.\n\nWho is this for: researchers working on low-latency ANN-SNN conversion or SNN-based detection. The tdIF idea might be salvageable, but the latency accounting and the L disclosure need fixing first. I would not cite it in its current form.\n\nRecommendation: send it to peer review, not desk-reject, because the core idea is worth refereeing and the flaws are identifiable and repairable; but the authors need a major revision with honest latency comparison and clear L selection.","headline":"A plausible tdIF neuron idea buried under a latency claim that the paper's own pipeline formula contradicts; worth a careful referee, but not as-is.","tokens_in":22242,"tokens_out":3479,"would_cite":false,"duration_ms":29652,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A time-position-coding neuron plus a delayed-firing rule lets converted spiking detectors reach near-ANN accuracy in five time-steps: 74.41% AP@0.50 on PASCAL VOC and 55.73% on COCO.","keywords":["spiking neural networks","ANN-SNN conversion","temporal coding","integrate-and-fire neuron","object detection","lane detection","low-latency inference","residual membrane potential"],"falsifier":"Count the real end-to-end time-steps of the converted YOLOv3–ResNet34 detector, including every layer's accumulate-then-fire phase as in Section 5.3: if the total is $(n+1)\\times T$ with $T=5$ — hundreds of steps for a deep backbone — the 'within 5 time-steps' claim is per-layer, not end-to-end. As a second check, train the source ANN with quant-steps $L$ much larger than the SNN time-step $T$ and test the claimed error-free identity (Eq. 12); the paper's own Table 6 shows $L=128$ with $T=3$ drops VOC AP to 11.49%, so the lossless claim only holds in the matched regime.","tokens_in":21041,"feed_emoji":"⚡","tokens_out":15267,"duration_ms":130570,"temperature":0.7,"pith_summary":"The paper sets out to show that ANN-to-SNN conversion, previously effective mainly for image classification, can deliver state-of-the-art object detection and lane detection at five time-steps. The central move is to make spike timing carry information: a temporal-dependent integrate-and-fire (tdIF) neuron weights each time-step by $2^{T-t}$, so an early spike is worth more than a late one, and a delay-spike procedure lets each layer collect all its input before firing, which removes residual membrane-potential error. With both mechanisms the converted network reports 74.41% AP@0.50 on PASCAL VOC and 55.73% on COCO at five time-steps, plus 95.48% accuracy on Tusimple and 76.65% F1 on CULane. If the approach holds, low-power neuromorphic hardware could run real-time detection at a small fraction of the latency and energy that earlier spike-based detectors needed.","feed_headline":"Spiking detectors match ANN accuracy in just five time-steps","feed_subtitle":"Time-position-coded neurons erase the errors that forced earlier spiking detectors to run thousands of steps.","key_machinery":"The load-bearing object is the temporal-dependent integrate-and-fire (tdIF) neuron, which at time-step $t$ scales incoming current and firing threshold by $c[t] = 2^{T-t}$, making the network use binary positional coding of time: a spike at step 1 carries weight $2^{T-1}$ and one at step $T$ carries weight 1. Two mechanisms support it. The delay-spike rule (Alg. 1) withholds firing until $T_{\\text{delay}}$ steps of accumulation are complete, then emits spikes deterministically against the total potential, driving the residual membrane potential to zero. The quant-clip activation (Eq. 11) trains the source ANN with the same quantization the SNN will face, and the conversion-error identity $Err^l = 0$ holds provided $T = L$ and $\\theta^l = \\lambda^l$. The decoding rule $O_{\\text{tdIF}} = \\sum_t V_o[t]\\cdot 2^{T-t}/(2^T - 1)$ turns the final membrane potentials into real-valued regression outputs — box coordinates and lane maps — and per-layer threshold tables $\\hat{\\theta}^l[t] = 2^{T-t}\\theta^l$ implement the neuron with one extra bit-shift per integration, keeping tdIF's energy within the same order of magnitude as a plain IF neuron.","core_discovery":"The paper's central claim is that the two classic sources of ANN-SNN conversion error — residual membrane potential left over from irregular spike timing, and quantization from coarse firing-rate coding — can be attacked together, and that the residual gap is closed by encoding value in the temporal position of spikes rather than in spike count alone. Concretely it asserts that a source ANN trained with a quant-clip activation ($L$ quantization levels, per-layer trainable ceiling $\\lambda$) converts to a tdIF-based SNN with zero conversion error whenever the SNN runs for $T = L$ time-steps and the firing thresholds satisfy $\\theta = \\lambda$ (Eq. 12). Empirically it claims this surpasses all prior conversion-based detectors: 43.46 AP@0.50:0.95 and 74.41 AP@0.50 on PASCAL VOC, and 31.93 and 55.73 AP@0.50 on COCO, with a YOLOv3/ResNet34 backbone at five time-steps, and the first successful converted lane detector, at 95.48% accuracy on Tusimple and 76.65% F1 on CULane, with per-neuron energy within the same order of magnitude as a standard IF neuron.","pith_inferences":["The weighting $c[t] = 2^{T-t}$ makes the tdIF decoder literally read out a fixed-point binary numeral from the final membrane potential, which the paper does not say in so many words; that viewpoint suggests pushing to multi-bit or mixed-radix codings, or even an adaptive radix, as a route below five steps.","Read with Section 5.3's own accounting, 'within five time-steps' is a per-layer figure: the end-to-end cost is $T + n\\times T_{\\text{delay}}$, and with $T_{\\text{delay}} = T$ a ResNet34-scale detector runs for hundreds of steps; a fair cross-method comparison would report total time-steps or steady-state frames per second rather than per-layer steps.","The sharp drop when $T_{\\text{delay}}$ falls below 2 — tdIF has no accuracy at $T_{\\text{delay}} = 0$ or 1 — indicates the method depends on holding inputs for a full accumulation window; injecting arrival jitter into the accumulated input on event-driven hardware would directly test where that dependence starts to hurt.","The appendix's table of one quant-step $L$ evaluated at several SNN time-steps $T$ is effectively a latency-accuracy trade curve, and a runtime controller that sets $T$ to meet a latency budget could be built from it without retraining — the paper does not propose this, but its own data already supports it."],"forward_implications":["Converted spiking networks become realistic engines for real-time perception: the same backbone that needs 40 or 2000 steps in earlier conversion methods is reported to detect objects at five steps, which on neuromorphic hardware translates directly into lower latency and per-frame energy.","The delay-spike rule and tdIF neuron are validated end-to-end on two distinct regression tasks, showing the conversion recipe is not tied to a single detection head.","Deep backbones are where the big gains appear (ResNet34 beats the Tiny backbone by about 10 AP points at five steps), but they lose more accuracy at the lowest step counts; the paper's appendix shows the loss at $T=3$–$4$ grows with depth, mapping exactly where the method is most and least forgiving.","At matched accuracy, tdIF energy stays in the same order of magnitude as IF: an 8-step tdIF YOLOv3 uses about 3.4 times the power of an 8-step IF baseline but exceeds the accuracy of a 32-step IF model, so the low-step advantage is not bought with a per-step energy explosion."],"supporting_citations":[{"why":"Establishes the functional correspondence between ReLU activation and IF neuron dynamics that the whole conversion pipeline (Eqs. 2–4) builds on.","marker":"Cao et al., 2015"},{"why":"Introduces firing-rate equivalence and weight/threshold balancing, the framing of quantization error the paper extends.","marker":"Diehl et al., 2015"},{"why":"Supplies the soft-reset (reset-by-subtraction) mechanism adopted in Eq. 4 and the percentile-based activation normalization.","marker":"Rueckauer et al., 2017"},{"why":"RMP-SNN residual membrane-potential retention is the direct target that the delay-spike mechanism is designed to eliminate.","marker":"Han et al., 2020"},{"why":"QCFS quant-clip activation is the direct predecessor of the paper's quant-clip training (Eq. 11) and the lane-detection baseline that needs 256 steps.","marker":"Bu et al., 2023"},{"why":"Spiking-YOLO is the principal conversion baseline that needs about 2000 steps, the comparison that defines the paper's low-latency advantage.","marker":"Kim et al., 2020b"},{"why":"Two-stage optimization is the detection baseline at 40 steps that the paper beats at 5, and the reference the paper cites for COCO.","marker":"Wang et al., 2023"},{"why":"Fast-SNN is the closest low-latency conversion competitor at 7 steps, which the paper's 7-step results explicitly outperform.","marker":"Hu et al., 2023"},{"why":"Time-step-driven pipelined inference that Section 5.3 adapts into the inter-layer delay-spike pipeline.","marker":"Liu et al., 2022"}],"fun_headline_variants":["Temporal spikes beat frequency: 5-step SNN detectors","Five-step spiking detectors surpass all ANN-SNN conversions","Time-coded neurons let SNN detectors hit SOTA in 5 steps","Position-coded spikes enable 5-step SNN object detection","Temporal-dependent IF neurons: SOTA detection in 5 time-steps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's headline that detection runs within five time-steps counts only a single layer's waiting window; its own pipeline formula (Section 5.3, with $T_{\\text{delay}} = T$ in every experiment) gives total steps of $T + n\\times T_{\\text{delay}}$, so a deep detector actually runs for hundreds of steps end to end.","fun_headline_variants_meta":{"raw":{"variants":["Temporal spikes beat frequency: 5-step SNN detectors","Five-step spiking detectors surpass all ANN-SNN conversions","Time-coded neurons let SNN detectors hit SOTA in 5 steps","Position-coded spikes enable 5-step SNN object detection","Temporal-dependent IF neurons: SOTA detection in 5 time-steps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000784,"raw_usage":{"total_tokens":3515,"prompt_tokens":1055,"completion_tokens":2460,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":671,"completion_tokens_details":{"reasoning_tokens":2369}},"tokens_in":671,"tokens_out":2460,"duration_ms":18268,"temperature":1.0,"reasoning_tokens":2369,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:46:32.803068+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Count the real end-to-end time-steps of the converted YOLOv3–ResNet34 detector, including every layer's accumulate-then-fire phase as in Section 5.3: if the total is $(n+1)\\times T$ with $T=5$ — hundreds of steps for a deep backbone — the 'within 5 time-steps' claim is per-layer, not end-to-end. As a second check, train the source ANN with quant-steps $L$ much larger than the SNN time-step $T$ and test the claimed error-free identity (Eq. 12); the paper's own Table 6 shows $L=128$ with $T=3$ drops VOC AP to 11.49%, so the lossless claim only holds in the matched regime.","supporting_citations":[],"review_version":2}