{"id":"1e8f96b9-fe72-4274-8398-16203d235674","arxiv_id":"2603.23206","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Deep TTFS/latency-coded SNNs can be trained directly with backpropagation, reaching ~93.6% on CIFAR-10 with an average inference latency near one timestep.","lead":"This paper trains spiking neural networks that convey information in the timing of spikes instead of spike counts, using backpropagation through time, and reports competitive accuracy with just one to four simulated time steps. It adds a feature-extraction encoder and a confidence-weighted loss to make timing-based training stable enough for deep networks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Latency comparisons in Table I depend on zero-delay same-timestep propagation across all layers (Eq. 2); if prior TTFS papers count time per layer, the claimed orders-of-magnitude speedup may be a metering artifact. A delay-aware re-evaluation is needed.","rationale":"I read the paper as proposing a training framework that relaxes strict single-spike TTFS, adds an encoding module, and uses a confidence-weighted loss to obtain low-latency decisions. The reader's verdict is CONDITIONAL, and my analysis points to the same load-bearing assumption: the meaning of 'inference time' in the headline comparison. The paper's own description of prior methods requiring 'multiplication of quantized steps and the number of layers' (Intro) contrasts with the zero-delay propagation enabled by Eq. (2) for the proposed method. This is not an internal inconsistency, but it makes the central empirical claim sensitive to the metering convention. The ablation studies and time-scalability plots provide reasonable support for the internal components, and the robustness analysis is interesting, but none of that resolves the comparability of the central latency and energy numbers. A concrete delay-aware re-evaluation would settle whether the claimed advantage is real or a protocol artifact. I do not see ground to reject the paper; the evidence is insufficient to move beyond conditional acceptance without code or a fair timing benchmark. Novelty overstatement and lack of error bars are secondary concerns and do not change the verdict.","tokens_in":18148,"tokens_out":7652,"duration_ms":88793,"concrete_test":"Use the authors' model (if released) or a faithful reimplementation, and choose one strong baseline (DTA-TTFS [25] or Stanojevic et al. [34]). Run both through the same discrete-event simulator with a mandatory one-global-timestep delay for every layer-to-layer transmission (replace S^{l-1}[t] in Eq. (2) by S^{l-1}[t-1], or equivalent event scheduling). Record (i) total simulated time from input onset to output decision, (ii) number of spikes, and (iii) normalized TrueNorth/SpiNNaker energy. Repeat with the zero-delay protocol. If the proposed method's 1.00-timestep advantage becomes 10+ timesteps (or the baseline's per-layer count compresses), the reported comparison is not robust; if the gap remains orders of magnitude, the claim is confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—state-of-the-art TTFS accuracy at 1–4 timesteps—turns on how 'inference time' is measured. In Sec. III-A the input is converted to first-spike times t_s = ceil((1-x)T). Eq. (2) has no synaptic delay: U^l[t] = τU^l[t−1] + W^l S^{l−1}[t], so a spike emitted by layer l−1 at global timestep t can, in the same t, cross the threshold in layer l. Combined with the layer-wise synchronous update, an 11-layer network can therefore produce an output spike at t=1, and Table I reports 1.00 average timesteps. The baselines are not measured under the same protocol: [34] is listed as 4096 timesteps per layer, and DTA-TTFS/event-driven methods use per-layer event propagation. If those baselines count per-layer processing times while the proposed model counts a single zero-delay global iteration, the two-orders-of-magnitude latency reduction in Table I and the normalized-energy numbers in Table III are partly a metering artifact. The concern is not that the model is inaccurate; it is that the headline advantage may not survive a common timing convention.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework for directly training deep time-to-first-spike (TTFS) / latency-coded spiking neural networks with BPTT and surrogate gradients. Three components are introduced: a latency encoding (LE) module with a straight-through estimator (Eqs. 5-9), relaxation of the single-spike constraint in hidden layers with a membrane-potential-based decoding rule in the output layer (Eqs. 10-11), and a temporal adaptive decision (TAD) loss that reweights cross-entropy over timesteps by confidence (Eqs. 12-14). The experiments report strong accuracy at 1-4 inference timesteps on CIFAR-10, CIFAR-100, Tiny-ImageNet, and CIFAR10-DVS, along with theoretical energy estimates and a robustness comparison against rate-coded SNNs. The central claim is that latency-coded SNNs can be trained efficiently with BPTT and achieve state-of-the-art accuracy among TTFS-coded SNNs with ultra-low latency and high energy efficiency.","tokens_in":18621,"tokens_out":8720,"duration_ms":78313,"significance":"If the reported latency metric is commensurate with previous TTFS work, the result is significant: it would make BPTT-trained temporal-coding SNNs competitive with rate-coded SNNs at a fraction of the inference latency, and the proposed LE/TAD components are clearly characterized by ablations and time-scalability experiments (Fig. 6). The paper also offers a useful temporal-similarity analysis (Sec. IV-D). However, the headline comparisons currently mix architectures, use an inconsistent time-counting protocol, and lack uncertainty estimates; the claims therefore need to be re-baselined before the result can be accepted as stated.","major_comments":[{"comment":"The central latency comparison is not measured under a common protocol. Eq. (2) uses S^{l-1}[t] in the update of U^l[t] in the same global timestep, so with the synchronous layer-wise update a signal can travel from input to output within one timestep. The baselines in Table I count time differently: [34] is listed as '4096 per layer', and Sec. I states that prior TTFS methods need 'multiplication of quantized steps and the number of layers'. Consequently, the reported 1.00-4.00 inference timesteps and the 'two orders of magnitude' speedup are not directly comparable to the per-layer timestep counts of DTA-TTFS, T2FSNN, TSC-SNN, etc. Please report the comparison under a common timing convention (e.g., a delay-aware or layer-serial simulation) and restate the latency/energy claims accordingly.","section":"Sec. III-A, Eq. (2), Table I"},{"comment":"The accuracy comparisons are confounded by architecture and training setup. All TTFS baselines in Table I are VGG-16, while the proposed results use VGG-11, VGG-16, and SEW-ResNet-18; on CIFAR-10 the proposed best accuracy (93.64 with SEW-ResNet-18) is below the 93.69 of Stanojevic et al. (VGG-16), so the 'state-of-the-art accuracy' claim is not supported. On CIFAR-100 the gain over the best baseline (72.24 vs 74.97) is large, but the architecture differs (VGG-11 vs VGG-16). No error bars or seed-to-seed variability are reported, and augmentation recipes differ among datasets. Please provide same-architecture comparisons and multi-seed statistics for the main results.","section":"Sec. IV-A, Table I"},{"comment":"The BPTT/surrogate-gradient training is not fully specified. The firing function H(·) in Eq. (3) is treated as a Heaviside step, and an STE is defined only for the latency encoder (Eq. 9). No surrogate gradient is given for the spike function used in backpropagation through the hidden layers. Since surrogate choice and slope can substantially change accuracy and latency in SNNs, please state the exact surrogate gradient (including its functional form and hyperparameters) and, ideally, ablate it. This is required for reproducibility of the central method.","section":"Sec. II-A, Sec. III-A, Eq. (3)"},{"comment":"The normalized-energy comparison inherits the timing-protocol problem. Eq. (17) multiplies a static-energy coefficient by 'Timesteps'; for the proposed model this is the zero-delay global timestep count, whereas for baselines it is the per-layer timestep count from Table I. The '0.366x/0.492x' energy values therefore understate the proposed model's energy if a hardware-compatible per-layer delay is used. Please recompute Table III with a common time granularity, or explicitly justify a hardware mapping in which same-timestep propagation across all layers is realizable.","section":"Sec. IV-B, Eq. (17), Table III"},{"comment":"The robustness advantage is supported by only a single paired comparison with no statistical significance. On CIFAR-10-C, the mCE gap is 0.8 points (33.9 vs 34.7) and the latency model has a higher clean error (6.4 vs 5.4); on CIFAR-100-C the gap is larger (50.8 vs 56.9). Without multiple seeds or confidence intervals, the robustness claim in the abstract is premature. Please provide variance estimates or down-rank the claim.","section":"Sec. IV-C, Table IV"}],"minor_comments":[{"comment":"Several typos: 'backpropagation throuh time' (Sec. I), 'serveritis' (Table IV caption), 'comparisom' (Sec. IV-E2), 'Ploted' (Sec. IV-C). Please proofread.","section":"Throughout"},{"comment":"Since F is defined via Sigmoid(·) in (0,1), the latency mapping is well-defined only for x in (0,1); please state the treatment of boundary values (or note that sigmoid never reaches 0 or 1).","section":"Eq. (8)"},{"comment":"O[t], the pre-synaptic current of the output layer, is used in the TAD loss but not defined in the model description. Please define it and clarify its relationship to the logits used in the decoding rule.","section":"Sec. III-C, Eqs. (12)-(14)"},{"comment":"The variable T is reused for both the maximum training timestep and the earliest spike time; rename one to avoid ambiguity.","section":"Eq. (10)"},{"comment":"The column header is garbled ('Neural Time ... Spikes (10^4)'); please format clearly and state exactly how 'Normalized Energy' is computed from Timesteps and Spikes.","section":"Table III"},{"comment":"The manuscript would benefit from a reproducibility statement and, ideally, code/checkpoints, as no such information is currently provided.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The timing-protocol issue is the most important one for the editor; if the authors cannot put all methods under a common time convention, the paper's main quantitative claims should be significantly softened. Also, the 'UNDER REVIEW' headers in the PDF suggest the submitted version includes review marks; please ask the authors to clean this before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about this paper. The training recipe is a real step forward: they show you can train deep latency-coded SNNs with ordinary BPTT and get accuracy competitive with rate coding at 1–4 global timesteps. But the headline latency comparison to earlier TTFS methods is not on a level playing field. In their update (Eq. 2), a spike from layer l−1 at time t contributes to layer l in the same global timestep, so a spike can propagate through all 11 layers within a single step. Earlier TTFS methods count per-layer processing, e.g., 4096 timesteps per layer for [34]. So the “1.00 timestep” figure is not directly comparable to “4096 timesteps per layer.” The claimed orders-of-magnitude latency reduction is partly a metering artifact. That doesn't make the method useless—on synchronous hardware, one global step per inference is indeed fast—but the authors need to report latency in a common unit (e.g., layer-steps or simulated time) and re-run the comparison.\n\nWhat is actually new: the LE module that maps features to spike times with straight-through gradients, the relaxed multi-spike hidden layers with output-layer first-spike decoding, and the TAD loss that adaptively weights supervision. These pieces form a coherent framework, and the ablations show each contributes. The membrane-potential tie-breaking is a small but sensible fix for discrete-time ties. The robustness analysis against CIFAR-C is interesting, though the explanation in terms of temporal decorrelation is post-hoc.\n\nSoft spots beyond the timing issue: comparisons mix architectures—Ours uses VGG-11/SEW-ResNet-18 while baselines are VGG-16—and augmentation recipes differ. No error bars or seeds are reported, and no code is released. The claim that BPTT for TTFS is “largely unexplored” is an overstatement; DTA-TTFS and earlier gradient-based TTFS methods exist, so the novelty is the specific combination, not the general idea.\n\nAll that said, the paper is clearly written and the training framework is worth engaging with. The central claims need to be stress-tested with a common timing convention and controlled baselines. This deserves serious peer review; I'd send it out, but with reviewers who will push on those points. I'd be cautious about citing the latency numbers until the metric issue is resolved, but the training recipe itself is citable.","headline":"A genuine training recipe for latency-coded SNNs with strong accuracy claims, but the headline latency comparison to prior TTFS methods is not apples-to-apples and needs a common timing convention.","tokens_in":18963,"tokens_out":4194,"would_cite":true,"duration_ms":44213,"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":"The paper claims that time-to-first-spike (TTFS) spiking networks can be trained with backpropagation through time, reaching state-of-the-art accuracy among TTFS methods at ultra-low latency by allowing hidden neurons to fire multiple times","keywords":["spiking neural networks","time-to-first-spike coding","latency coding","backpropagation through time","temporal coding","low-latency inference","neuromorphic computing","energy efficiency"],"falsifier":"Implement the proposed latency-coded network and a prior TTFS baseline from the paper's comparison table on the same neuromorphic simulator with a strictly layer-wise clock (each layer takes one timestep to propagate), and measure wall-clock time to first output spike on CIFAR-10. If the latency-coded network does not remain at least an order of magnitude faster than the baseline when both use identical time units and communication protocols, the central latency claim fails.","tokens_in":18077,"feed_emoji":"⚡","tokens_out":5630,"duration_ms":47823,"temperature":0.7,"pith_summary":"The paper aims to make time-to-first-spike (TTFS) coding practical for deep spiking neural networks by training latency-coded SNNs with backpropagation through time (BPTT). It claims that by relaxing the strict single-spike constraint in hidden layers, using a feature-extraction encoder, and weighting loss by the model's confidence, TTFS networks can reach state-of-the-art accuracy among temporal-coding methods while completing inference in about one to four timesteps. If true, latency coding becomes a hardware-friendly alternative to rate coding for fast, energy-efficient neuromorphic inference. The paper also reports improved robustness to input corruptions relative to rate-coded SNNs.","feed_headline":"Latency-coded spiking nets reach 93.6% on CIFAR-10 in ~1 timestep","feed_subtitle":"A backprop-trained temporal-coding network runs at minimal delay, making spike-timing a practical energy-saving alternative to rate coding.","key_machinery":"The machinery is the iterative LIF neuron with soft reset (Eq. 2-4), combined with three components: a latency encoding module that maps extracted features to spike times via t_s(x)=ceil((1-x)T) and passes gradients via a straight-through estimator; a decoding rule that selects the class by earliest output spike and breaks ties using membrane potential (Eq. 11); and the temporal adaptive decision (TAD) loss, which weights cross-entropy at each timestep by a temperature-smoothed inverse-entropy confidence (Eq. 12-14). Together, these allow BPTT to train a network whose decision is made from first-spike timing while hidden layers retain multi-spike gradient flow.","core_discovery":"The central claim is that latency-coded SNNs—where information is carried by the timing of the first output spike—can be trained end-to-end with BPTT once three design choices are made. A latency encoding module extracts features and encodes those features, not raw pixels, into spike times using a straight-through estimator. Hidden-layer neurons are allowed to fire multiple times, while the output layer still decides based on the earliest spike, with membrane potential used to break ties. A temporal adaptive decision (TAD) loss weights per-timestep cross-entropy by the network's confidence, pushing easy samples to fire early and hard samples to integrate longer. The paper reports 93.60% on C","pith_inferences":["If the inference-latency metric is genuinely commensurate with prior work, the framework could be adapted to asynchronous event-driven hardware, where the wall-clock latency would be set by the critical path through layers rather than the number of global timesteps.","The TAD loss's confidence-based weighting scheme is a general principle that could transfer to other early-exit or adaptive-computation architectures, not only spiking networks.","The low temporal similarity of latency-coded representations suggests a testable hypothesis: the robustness gain comes from decorrelated per-timestep features, which could be verified by ablating the TAD loss and measuring both robustness and temporal similarity.","A direct extension would be to measure energy on actual neuromorphic chips, since the theoretical energy model assumes zero static-energy scaling with latency; the claimed advantage may shrink on hardware with high static power."],"forward_implications":["Latency-coded SNNs can run inference in 1-4 timesteps on standard vision benchmarks, making real-time neuromorphic deployment feasible.","BPTT becomes a viable training paradigm for temporal coding, enabling deeper TTFS networks without conversion or event-driven learning rules.","Relaxing the single-spike constraint in hidden layers while keeping first-spike output decisions resolves gradient vanishing without increasing decision latency.","The reported robustness advantage suggests temporal coding may be preferable to rate coding under input corruptions, not just for speed.","Energy estimates indicate latency-coded SNNs consume a small fraction of the energy of ANN counterparts and less than rate-coded SNNs."],"fun_headline_variants":["Train spike-timing nets with BPTT: 93.6% in one timestep","Latency coding makes SNNs fast and accurate","Spike timing beats rate: BPTT-trained SNNs hit 93.6%","One-timestep SNNs: latency coding plus BPTT","Efficient SNNs: encode in first spike, train with BPTT"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing assumption is that the reported inference time—measured as the first output spike in a synchronous layer-wise simulation where a spike can propagate through all layers in a single global timestep—is directly comparable to the per-layer timestep counts reported for prior TTFS methods (for example, thousands of timesteps per layer). If these metrics are not commensurate, the claimed two-orders-of-magnitude latency reduction is partly a metering artifact rather","fun_headline_variants_meta":{"raw":{"variants":["Train spike-timing nets with BPTT: 93.6% in one timestep","Latency coding makes SNNs fast and accurate","Spike timing beats rate: BPTT-trained SNNs hit 93.6%","One-timestep SNNs: latency coding plus BPTT","Efficient SNNs: encode in first spike, train with BPTT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00058,"raw_usage":{"total_tokens":2607,"prompt_tokens":817,"completion_tokens":1790,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":1688}},"tokens_in":561,"tokens_out":1790,"duration_ms":10476,"temperature":1.0,"reasoning_tokens":1688,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T17:32:40.815880+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Implement the proposed latency-coded network and a prior TTFS baseline from the paper's comparison table on the same neuromorphic simulator with a strictly layer-wise clock (each layer takes one timestep to propagate), and measure wall-clock time to first output spike on CIFAR-10. If the latency-coded network does not remain at least an order of magnitude faster than the baseline when both use identical time units and communication protocols, the central latency claim fails.","supporting_citations":[],"review_version":1}