{"id":"585245a0-5388-4344-9849-b466a0c56b9d","arxiv_id":"2411.16711","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Adding forward and backward skip connections with explicit temporal delays to spiking neural networks, selected by training-free NAS, improves accuracy on four event-based benchmarks.","lead":"This paper adds connections with built-in time delays, called TSkips, to spiking neural networks so the networks can better use information from earlier moments when processing event-based camera and audio data. The authors use an automated search to pick where the delayed connections go and how long the delay should be, and report accuracy gains on optical flow, gesture, and speech benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (4)'s backward TSkip includes a same-timestep deeper-layer term α h^t_{l+k} that is not causally computable in the feedforward SNN; the paper never states α=0 for backward paths, so the mechanism behind a large share of the reported gains is ambiguous or impossible as written.","rationale":"The reader's concern about NASWOT-SAHD correlation and missing search budget is legitimate: with a 2^120 search space and no reported search budget, the reported gains could partly reflect favorable sampling. However, the paper includes no-skip ablations and multi-seed DSEC results that partially mitigate that risk. The more load-bearing issue is internal to the method: Eq. (4) defines a backward TSkip term α h^t_{l+k} that cannot be evaluated causally in a feedforward SNN, yet backward TSkips drive several of the headline improvements. This is not an ad hominem or a disagreement with the empirical community consensus; it is a concrete question about whether the published mechanism is the mechanism that was evaluated. The empirical results may well survive correction, but the paper must state which equation was implemented, report α values or show that backward paths use only the delayed source, and ideally release code. Therefore the verdict should remain conditional rather than moving to acceptance or rejection. The reader's weakest_assumption did not identify this same concern, though the missing-code condition is related, so agreement is partial.","tokens_in":25266,"tokens_out":8743,"duration_ms":89149,"concrete_test":"Run a forward pass of a 3-layer MLP with a backward TSkip (layer 3 → layer 1, Δt=4, α>0) at T=6, computing layers in causal order 1→2→3 within each time step. If the forward pass errors or requires reordering/iteration, Eq. (4) is not executable as written. Then re-train the SHD BTSkips-1 configuration with α fixed to 0 (delayed source only) under the reported hyperparameters and compare to the reported 93.64% accuracy; a large drop would indicate the current-time term is doing the work, while a match would show Eq. (4) overstates the mechanism.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism is stated in Eq. (4) as h^t_l = f_l(h^t_{l−1} ⊕ W_s(α h^t_{l±k} + (1−α)h^{t−Δt}_{l±k})). For a backward TSkip, the source is the deeper layer l+k at the same time step t. In a feedforward SNN processed layer-by-layer within a time step, h^t_{l+k} is not yet available when h^t_l is computed; for k>0 this creates a cyclic dependency (e.g., L1 needs L3 while L3 needs L2 and L2 needs L1). The paper's causality constraint 0 < t−Δt < T restricts only the temporal index, not the layer dependency. Thus either the implementation sets α=0 for backward skips and Eq. (4) is not the model actually run, or the backward TSkip results rely on an unstated non-causal or iterative scheduling scheme. Backward TSkips are the best or near-best variant on SHD, SSC, and DVS128 Gesture (Tables 2–3), so this ambiguity sits at the center of the claimed gains. No code or explicit α values are provided to resolve which case holds.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TSkips, skip connections with explicit temporal delays added between non-adjacent layers of spiking and hybrid ANN-SNN networks, and uses the training-free NAS method NASWOT-SAHD to select delay values, skip positions, and network depth. Experiments on DSEC-flow, DVS128 Gesture, SHD, SSC, and CIFAR10-DVS report accuracy and efficiency gains over baseline SNNs, with backward TSkips often performing best. The paper also includes ablations on delay, position, and depth, energy estimates, and a three-seed robustness study on DSEC-flow.","tokens_in":25600,"tokens_out":7337,"duration_ms":68301,"significance":"If the mechanism is correctly specified, TSkips are a simple, low-parameter architectural modification with consistent accuracy gains across four event-based benchmark tasks, and the use of training-free NAS avoids expensive architecture search. The paper contains useful ablations and energy estimates, and the three-seed DSEC-flow results in Appendix D.2 are a good reproducibility practice. However, the central mechanism is not fully specified for backward TSkips, the efficiency claim is stated ambiguously, and the main classification results are single-seed with a moderate NAS proxy correlation, so the strengths do not yet justify acceptance.","major_comments":[{"comment":"The backward TSkip as written contains a same-timestep term α h^t_{l+k} that is not causally computable in a feedforward SNN processed layer by layer within a time step: for k>0, h^t_{l+k} is not yet available when h^t_l is computed. The constraint 0<t−Δt<T in Section 3.3 restricts only the temporal index and does not resolve this layer dependency. Since backward TSkips are the best or near-best variant on SHD, SSC, and DVS128 Gesture (Tables 2–3), the reported gains depend on resolving this ambiguity. Please state explicitly whether α=0 for backward paths (in which case Eq. (4) misdescribes the model actually run), or specify the alternative intra-timestep scheduling or iterative scheme used, and provide pseudocode or code for the forward pass.","section":"Section 3.3, Eq. (4)"},{"comment":"The energy-efficiency claim is stated in contradictory ways. At a fixed model scale, Table 5 shows that adding TSkips increases estimated inference energy (e.g., Base baseline 23.3 mJ vs. 29.4 mJ for forward and 27.6 mJ for backward; Pico baseline 1.90 mJ vs. 2.07/2.02 mJ), so the opening sentence of Appendix B.1, which says TSkips reduce AEE 'without increasing model complexity or energy consumption,' is not supported by the paper's own table. The energy savings in Fig. 1 come from comparing a smaller TSkip-augmented model with a larger baseline, which is a legitimate efficiency argument, but it must be stated as a cross-model comparison and the same-scale increase should be acknowledged. Please correct the wording and clarify the basis of the 40.75% energy-reduction figure.","section":"Section 4.3 / Appendix B.1"},{"comment":"The NAS proxy used to select TSkip architectures has Kendall's tau of 0.58 for forward and 0.55 for backward TSkip searches, below the 0.63 reported for the baseline, and the paper does not report the number of architectures sampled in the random search described in Section 3.4. Because Tables 2 and 3 report single-seed results, the main classification gains could in principle reflect favorable selected configurations rather than a systematic property of TSkips. Please provide multi-seed mean±std for the main classification results, or at least for one of the three classification datasets, and report the search budget; the three-seed DSEC-flow analysis in Appendix D.2 is a good model to follow.","section":"Appendix D.1 / Section 3.4"}],"minor_comments":[{"comment":"The abstract and Section 1 claim 'up to 18%' AEE reduction on DSEC-flow, but the largest reduction in Table 1 is 17.0% (Base + Forward, from 1.35 to 1.12); please reconcile the number or state the comparison basis.","section":"Abstract / Table 1"},{"comment":"The energy labels 'E?52%, E?36%, E?56%, E?19%' are difficult to read and are not explained in the caption; please state what these percentages compare and which baseline is used.","section":"Figure 1"},{"comment":"There are several typographical errors, including 'evalauted', 'tempral', and 'achievess'; please proofread the text.","section":"Section 2.2"},{"comment":"The ablation study plots do not indicate whether the reported accuracies are single runs or averaged over seeds; please state the number of runs and, if available, error bars, since the paper argues that delay, position, and depth have a systematic effect.","section":"Section 4.2 / Figure 6"},{"comment":"The discussion says trained TSkips models can perform well 'without explicitly using the temporal skip at inference,' but the Appendix B.2/C.2 experiments replace the skip input with a zero tensor, which is not the same as removing the stored historical state on hardware; please clarify which operation is meant.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The paper has no code release, and the unresolved causality of Eq. (4) for backward TSkips makes it impossible for a reader to reconstruct the exact model. The heavy self-citation for baselines is not itself problematic because the baseline results are re-measured, but the single-seed classification tables and moderate NAS correlation are a reproducibility risk. I would encourage the editor to require the authors to release the forward-pass implementation or, failing that, to provide pseudocode that settles whether α is zeroed for backward connections."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Kolya—quick take: TSkips is a real method paper with consistent gains across four event-based benchmarks, but Eq. (4) makes the backward skip acausal as written, and that needs to be resolved before I'd trust the mechanism behind the best results.\n\nWhat's actually new: explicit integer delays on both forward and backward skip connections, combined with a training-free NAS to pick delay and placement. That combination is genuinely different from SCTT, TTFS, and the weight-delay line of work. The empirical work is solid: four datasets, ablations on delay, position, and depth, an energy analysis, and a nice no-skip inference experiment showing the trained representations still work when the skip is removed. The DSEC 3-seed variance in Appendix D.2 is a good sign. Gains are consistent and the model-size/accuracy trade-offs are clearly presented.\n\nSoft spots, in increasing order of concern. Main tables are single-seed except DSEC, so headline numbers don't carry error bars. No code is released, and the NAS search budget is not reported; with a proxy correlation of τ=0.55–0.58, a lucky sample is possible. The energy claims mix same-size and cross-model comparisons; Figure 1 is cross-model, which is fine but should be labeled as such. Most importantly, Eq. (4) as written uses α h^t_{l+k} for a backward TSkip—the deeper layer's current-timestep output—when the shallower layer is being computed. In a feedforward SNN that value does not exist yet. The paper never states α=0 for backward paths or describes an iterative scheduling scheme. Since backward TSkips are the best variant on SHD, SSC, and DVS, this sits at the center of the claimed gains. If the implementation really uses α=0 for backward skips, then the equation is just misleading. If α≠0, the causality needs to be explained.\n\nBottom line: this deserves a serious referee. It is a well-executed empirical study with a novel mechanism, but the causality question must be answered before I'd rely on the backward-skip results. For a reading group it is a good discussion paper; I'd cite it only after the ambiguity clears.","headline":"Solid method paper with real gains, but Eq. (4) makes backward TSkips acausal as written—needs clarification before the central claim is fully trusted.","tokens_in":26112,"tokens_out":7804,"would_cite":true,"duration_ms":65453,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes TSkips, forward and backward skip connections with explicit temporal delays, and reports that they cut optical-flow error by up to 18% and raise classification accuracy by up to 16% on four event-based benchmarks.","keywords":["spiking neural networks","temporal skip connections","explicit temporal delays","event-based vision","event-based audio","training-free neural architecture search","optical flow estimation","LIF neurons"],"falsifier":"Run a controlled comparison on SHD: train the top-ranked TSkip configurations from the search score and an equal number of randomly sampled TSkip configurations under the same architecture budget and training protocol. If random configurations match or beat the selected ones in mean accuracy, or if the ranking correlation for TSkip networks falls to near zero on a larger hold-out sample, then the search is not carrying the reported improvement.","tokens_in":25114,"feed_emoji":"⚡","tokens_out":10159,"duration_ms":89100,"temperature":0.7,"pith_summary":"TSkips proposes a small architectural change with a clear hypothesis: if a spiking network's skip connections carry an explicit temporal delay, the network can route information from an earlier time step directly to layers it would otherwise not reach, and this improves learning on event streams. The paper tests this on four event-based tasks—optical flow, hand-gesture recognition, and two speech benchmarks—by adding forward and backward delayed skips to spiking and hybrid backbones. It reports consistent gains with almost no added parameters: up to 18% lower average endpoint error on DSEC-flow, 8% higher accuracy on DVS128 Gesture, and up to 8% and 16% higher accuracy on SHD and SSC. Because the space of possible skip positions and delays grows exponentially, the authors use a training-free architecture-search score to pick configurations, and they report that the selected networks also converge faster and can often be run at inference without the delayed connection while keeping most of the gain. If correct, the result would make temporal wiring a practical and cheap design axis for time-sensitive spiking models.","feed_headline":"Delayed skip connections boost spiking network accuracy by 16%","feed_subtitle":"Time-delayed skips cut optical-flow error and improve event-based vision and speech.","key_machinery":"The central object is the TSkip connection: a forward or backward shortcut between layers $l$ and $l\\pm k$ that carries the source layer's activation from time step $t-\\Delta t$ instead of the current step, mixed with the current activation through a learnable coefficient $\\alpha$. By unrolling the network in time, a TSkip creates a direct route across both space and time, so spikes and gradients can bypass many intermediate layers and time steps. The paper pairs this wiring with adaptive leaky integrate-and-fire (LIF) neurons, surrogate-gradient training, and a training-free neural architecture search that scores candidate configurations by the Sparsity-Aware Hamming Distance of their spike patterns at initialization, letting it pick skip positions, delays, and depth without training each candidate.","core_discovery":"The paper's central claim is that explicit temporal delays in forward and backward skip connections—TSkips—let SNNs and hybrid ANN-SNN models capture long-term spatio-temporal dependencies that ordinary skips (delay zero forward, delay one backward) miss. The layer update becomes $h^t_l = f_l(h^t_{l-1} \\oplus W_s(\\alpha h^t_{l\\pm k} + (1-\\alpha) h^{t-\\Delta t}_{l\\pm k}))$, where $\\Delta t$ is chosen per connection and $\\alpha$ is a learned mixing coefficient. Across DSEC-flow, DVS128 Gesture, SHD, and SSC, the TSkips-augmented networks beat their baselines by up to 18% in AEE reduction and 8–16 percentage points in accuracy, with smaller TSkip models matching larger baselines. The paper attributes these gains to better spike flow, reduced vanishing-spike effects, and finer control over spike timing, and it argues that the improvements reflect better learning rather than extra connectivity because removing the skip at inference retains most of the accuracy gain.","pith_inferences":["A testable extension is to let the search choose a different delay for every skip on long, non-stationary streams; the paper fixes each delay per connection, and it is plausible that adaptive per-window delays would help where event statistics drift.","The no-skip inference result hints that TSkips may function partly as a training-time regularizer that reshapes learned dynamics; if so, the same benefit might be obtainable from a gradient-shaping or initialization trick without the extra connection at all.","TSkips could be combined with per-weight or per-layer learnable delays: use the training-free search to set the coarse delay structure and let training fine-tune the fine-grained delays, covering a spectrum not explored in the paper."],"forward_implications":["TSkips would make temporal wiring a first-class design choice: skipping a layer is no longer only a spatial shortcut but also a way to choose how many time steps of context feed into a computation.","Smaller TSkip-augmented models can match or beat larger baselines on optical flow and speech, so the mechanism could transfer to other latency- or energy-constrained event-based tasks.","The same delayed-skip mechanism improves fully spiking, hybrid ANN-SNN, and plain ANN models, which points to a general temporal-processing limitation rather than a quirk of one architecture.","Because trained TSkip networks retain most of their accuracy when the delayed connection is removed at inference, deployment could skip storing hidden temporal states, lowering memory and energy use on neuromorphic or edge hardware."],"supporting_citations":[{"why":"Supplies the training-free NASWOT-SAHD proxy that scores candidate architectures and selects TSkip configurations without training.","marker":"Kim et al. (2022)"},{"why":"Provides the Adaptive-SpikeNet optical-flow backbone and adaptive LIF neuron dynamics used for DSEC-flow experiments.","marker":"Kosta & Roy (2023)"},{"why":"Defines the EV-FlowNet encoder-decoder architecture whose skip connections are replaced by TSkips.","marker":"Zhu et al. (2018a)"},{"why":"Provides the hybrid ANN-SNN optical-flow backbone used to test TSkips in hybrid models.","marker":"Negi et al. (2024)"},{"why":"Supplies the ArcTangent surrogate gradient that makes the non-differentiable spike activation trainable via backpropagation through time.","marker":"Fang et al. (2021a)"},{"why":"Contributes the SHD and SSC event-based speech datasets used for the classification benchmarks.","marker":"Cramer et al. (2022)"},{"why":"Supplies the ResNet backbone architecture modified for the DVS128 Gesture experiments.","marker":"He et al. (2015)"},{"why":"Baseline delay-based SNN training method compared on gesture recognition; motivates the need for explicit temporal delays.","marker":"Shrestha & Orchard (2018)"}],"fun_headline_variants":["TSkips: temporal delay skips lift SNN accuracy by 16%","Delay-aware skips in SNNs cut optical flow error, boost speech","Explicit temporal delays make spiking networks more efficient and accurate","TSkips reduce error and model size in event-based spiking networks","Time-delayed skip connections improve SNN performance across tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the training-free architecture-search score reliably ranks TSkip configurations, so the reported gains come from a systematic property of TSkips rather than from a few lucky sampled architectures; the paper's own validation shows the ranking-correlation measure drops from 0.63 on baselines to 0.58 and 0.55 on TSkip networks.","fun_headline_variants_meta":{"raw":{"variants":["TSkips: temporal delay skips lift SNN accuracy by 16%","Delay-aware skips in SNNs cut optical flow error, boost speech","Explicit temporal delays make spiking networks more efficient and accurate","TSkips reduce error and model size in event-based spiking networks","Time-delayed skip connections improve SNN performance across tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000352,"raw_usage":{"total_tokens":2276,"prompt_tokens":1035,"completion_tokens":1241,"prompt_tokens_details":{"cached_tokens":1024},"prompt_cache_hit_tokens":1024,"prompt_cache_miss_tokens":11,"completion_tokens_details":{"reasoning_tokens":1148}},"tokens_in":11,"tokens_out":1241,"duration_ms":278418,"temperature":1.0,"reasoning_tokens":1148,"cache_read_input_tokens":1024,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:27:01.057587+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a controlled comparison on SHD: train the top-ranked TSkip configurations from the search score and an equal number of randomly sampled TSkip configurations under the same architecture budget and training protocol. If random configurations match or beat the selected ones in mean accuracy, or if the ranking correlation for TSkip networks falls to near zero on a larger hold-out sample, then the search is not carrying the reported improvement.","supporting_citations":[{"cited_title":"Neural architecture search for spiking neural networks, 2022","cited_arxiv_id":null,"evidence_quote":"Supplies the training-free NASWOT-SAHD proxy that scores candidate architectures and selects TSkip configurations without training."},{"cited_title":"Adaptive-spikenet: Event-based optical flow estimation using spiking neural networks with learnable neuronal dynamics, 2023","cited_arxiv_id":null,"evidence_quote":"Provides the Adaptive-SpikeNet optical-flow backbone and adaptive LIF neuron dynamics used for DSEC-flow experiments."},{"cited_title":"Best of both worlds: Hybrid snn-ann architecture for event-based optical flow estimation, 2024","cited_arxiv_id":null,"evidence_quote":"Provides the hybrid ANN-SNN optical-flow backbone used to test TSkips in hybrid models."},{"cited_title":"The heidelberg spiking data sets for the systematic evaluation of spiking neural networks","cited_arxiv_id":null,"evidence_quote":"Contributes the SHD and SSC event-based speech datasets used for the classification benchmarks."},{"cited_title":"Slayer: Spike layer error reassignment in time, 2018","cited_arxiv_id":null,"evidence_quote":"Baseline delay-based SNN training method compared on gesture recognition; motivates the need for explicit temporal delays."}],"review_version":1}