{"id":"68ded176-1e28-46b0-a61b-18635407f7c0","arxiv_id":"2505.04165","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Applying the temporal shift trick from video CNNs to spiking networks, with random channel split points and a residual scaling factor, yields small accuracy gains on standard SNN benchmarks.","lead":"TS-SNN inserts a temporal shift operation into spiking neural networks, so spike features from adjacent time steps are blended within a single step using one scalar weight, and it reports top accuracies on CIFAR-10 (96.72%), CIFAR-100 (80.28%), and ImageNet (70.61%). The paper is worth reading as a minimal-cost trick for making SNNs more temporally expressive, though it is an adaptation of the older TSM video method rather than a new mechanism.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported gains may depend on non-causal access to future spike features (Eq. 6); a causal variant test would determine whether the method works in online or low-latency inference.","rationale":"The paper applies a known temporal-shift idea to SNNs and provides useful ablations, including the Spikeformer with/without TS comparison that partially supports the module's effectiveness. However, the central claim of fusing past, present, and future features within a single timestep is not implementable in causal, low-latency inference as stated. The reader's weakest assumption points to exactly this gap, and the concern is concrete: Eq. (6)'s +1 branch accesses X[t+1]. The proposed causal-variant test is a direct way to determine whether the reported gains actually require lookahead or whether a past-only formulation preserves the accuracy. Since the reader's CONDITIONAL verdict already captures this unresolved issue, no change to the verdict is needed.","tokens_in":17535,"tokens_out":4021,"duration_ms":43385,"concrete_test":"Retrain the ResNet-20 CIFAR10-DVS (T=10) and ResNet-34 ImageNet (T=4) configurations from Tables 3 and 2 with a causal variant of Eq. (6): set Z+1 = X[t, :g1·Cfold, :, :] (no future branch) while keeping Z−1 and Z0 unchanged, using the same Ck=32, α, epochs, and augmentation. If the causal variant matches the reported 83.80% and 70.61% within one standard error, future access is not needed and the concern is resolved. If accuracy drops by more than the reported std (≈0.2%), the SOTA and latency claims depend on lookahead and should be restated as offline-processing results.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equation (6) defines Z+1 = X[t+1, :g1·Cfold, :, :], so at output timestep t the TS module injects spike features from timestep t+1 into the current feature map. In any causal streaming deployment—the setting implied by \"low-latency\" SNN inference and by the CIFAR10-DVS event-stream experiments—these features do not exist yet when timestep t is processed. The module as specified requires either buffering the full T-step sequence before inference (offline batch processing) or introducing one timestep of latency per shifted layer, and the paper does not state which protocol was used. The central claim that TS-SNN \"integrates past, present, and future spike features within a single timestep\" and is suitable for neuromorphic/edge use therefore rests on an unstated offline/lookahead assumption. This is load-bearing because the ablation in §4.2 shows accuracy depends on the exact shift configuration; if the +1 branch is removed, the configuration changes and the claimed SOTA margins may vanish. A causal version that replaces the future branch with the current timestep would still be a valid module, but it would be a different method and would need its own experiments.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TS-SNN, a temporal shift module for spiking neural networks. The module splits the channels of a spiking feature tensor into three groups, shifts one group forward in time, one backward, and leaves one unchanged, then combines the shifted tensor with the original input through a residual term scaled by a learnable scalar alpha. The authors report state-of-the-art accuracy on CIFAR-10 (96.72%), CIFAR-100 (80.28%), ImageNet (70.61%), and CIFAR10-DVS (83.90%) with fewer timesteps than prior SNN methods, and they provide energy estimates based on MAC/AC counts. The paper includes ablations on the channel folding factor, shift direction combinations, fixed versus random split strategies, and train/inference consistency, plus a small experiment inserting the module into a transformer-based SNN.","tokens_in":17795,"tokens_out":5136,"duration_ms":60797,"significance":"If the central claims hold, the TS module is an attractive plug-in: it adds a single scalar parameter per insertion, has essentially no arithmetic cost, and improves accuracy across static and event-based benchmarks while preserving the sparsity of SNN computation. The paper has concrete strengths: results are reported with standard deviations over three trials, the ablation set is reasonably broad, the transformer-based generality test is valuable, and the energy accounting follows a standard MAC/AC estimation framework used in the SNN literature. The main risk is not internal inconsistency of the shift algebra, but whether the method as defined is deployable in the online, low-latency settings implied by the paper's framing, and whether the reported gains are correctly attributed to the temporal shift rather than to training details or test-set-based hyperparameter selection.","major_comments":[{"comment":"The future-branch of the shift, Z+1 = X[t+1, :g1*Cfold, :, :], requires spike features from timestep t+1 to be available when the module computes its output for timestep t. In a causal streaming deployment, which is the setting implied by 'low-latency' SNN inference and by the CIFAR10-DVS event-stream experiments, those features do not exist yet. The paper does not state whether all experiments were run in an offline full-sequence mode, whether one timestep of buffering/delay per shifted layer was introduced, or whether some other protocol was used. This is load-bearing because the claimed advantage over prior SNNs is explicitly tied to fusing future features within one timestep while maintaining low energy cost. Please specify the inference protocol precisely, quantify the resulting latency and memory overhead, and report results for a causal variant in which the +1 branch is replaced by a current-timestep or delayed branch.","section":"Eq. (6), Algorithm 1, Section 4.3"},{"comment":"There is no no-TS baseline under the same training pipeline for the main ResNet-19/ResNet-20 experiments. The ablations vary Ck, shift directions, fixed versus random splits, and train/inference consistency, but none removes the TS module entirely in the ResNet setup; the only direct with/without-TS comparison is Table 4 for the transformer architecture. Because Table 1 compares against methods with different training recipes, and because the T=1 results already show that alpha alone contributes non-trivially, the reported improvements cannot be cleanly attributed to the temporal shift operation as opposed to the overall training setup or the residual alpha scaling. Please add a same-recipe no-TS baseline for ResNet-19/20 and, ideally, a no-TS baseline that retains the alpha scaling, to isolate the contribution of the shift itself.","section":"Section 4.2, Tables 1 and 4"},{"comment":"The hyperparameters Ck and the shift direction combination appear to be selected using test-set accuracy. Figure 3 reports CIFAR-100 test accuracy as a function of Ck and selects Ck=32; Figure 5 selects the L-R-0 direction combination from test accuracy; Appendix A.3 describes dataset splits but no validation split for model selection. Selecting hyperparameters on the test set introduces optimistic bias into the reported SOTA numbers, which are the paper's central quantitative claim. Please use an independent validation split for hyperparameter selection and report the test results obtained with the selected configuration, or provide a nested evaluation that accounts for the selection procedure.","section":"Figure 3, Figure 5, Appendix A.3"}],"minor_comments":[{"comment":"The procedure 'Generate random indices If for channel groups' is underspecified: it is unclear whether g1 and g2 are resampled on every forward pass, fixed after initialization, or chosen per training run. Since the random-versus-fixed ablation is claimed to matter (71.63% vs 71.24%), please state the sampling rule and the seed/reproducibility policy, and clarify whether inference is stochastic.","section":"Algorithm 1 and Section 4.2"},{"comment":"The claim that the TS operation introduces 'no additional computational burden' counts only arithmetic operations. The shift still requires data movement and, in the offline full-sequence interpretation, buffering of activation tensors across timesteps. Please separate FLOP/energy arithmetic from memory and latency overhead, especially in Table 5 and in the energy-efficiency discussion.","section":"Section 3.2 and Section 4.5"},{"comment":"For ImageNet, Table 5 reports energy for 'SEW ResNet18' while Table 2 reports ImageNet accuracy for ResNet-18 and ResNet-34. Please clarify which architecture was used for the energy measurement, whether the TS module was inserted into that architecture, and how the comparison to the reported 13.10 mJ SEW ResNet18 baseline was made.","section":"Table 5"},{"comment":"There are several small presentation errors, including 'LIF Modal' in the appendix heading and 'an novel Temporal Shift Module' in the conclusion; these should be corrected.","section":"Appendix A.1 and Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The future-access issue is the one that most affects the paper's framing as a neuromorphic/edge method. The authors may well have a valid offline temporal-fusion technique, but the current wording overstates its applicability to low-latency online inference. In addition, the absence of a same-pipeline no-TS baseline and the apparent test-set selection of Ck and shift direction make the headline SOTA numbers hard to evaluate as stated. These concerns are addressable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the bottom line: this is a competent, honest transfer of the temporal shift trick from video CNNs to SNNs, and the reported gains are probably real, but the headline claim that it fuses future features within a single timestep only works offline. Equation (6) pulls X[t+1,...] into the output at time t; the paper never says that it buffers the full spike sequence or accepts one timestep of latency per shifted layer. In a streaming or low-latency setting—the one implied by 'neuromorphic' and 'energy-efficient'—that branch is not available. The ablations show the exact shift configuration matters, so a causal variant could have meaningfully smaller margins. That is the main thing to know.\n\nWhat is new: the specific combination of random channel splitting (g1, g2), a scaled residual alpha*Z+X, and bidirectional shifts applied to SNN spike features is not in the cited prior work, which covers TSM for video and attention-based temporal SNN modules. The authors credit TSM properly. The CIFAR10-DVS result (83.8% at 10 timesteps, exceeding prior SOTA by ~3 points) is impressive if it holds up, and the ablations on Ck, shift directions, and train/inference consistency are useful and generally well reported with standard deviations.\n\nSoft spots, in order. (1) The non-causality above is load-bearing because the method is advertised for low-latency edge inference. (2) The main comparisons lack a TS-free baseline on the same backbone and training recipe; the SOTA numbers come from other papers with different hyperparameters, so part of the margin could be training details rather than the module. The paper mentions a vanilla SNN in the firing-rate appendix but gives no accuracy for it. (3) The hyperparameters Ck and alpha appear to be selected on the test benchmarks (Ck from CIFAR-100, alpha from CIFAR10-DVS) without a validation split; with many degrees of freedom, some overfitting is possible. (4) The energy analysis has a backwards sentence: it says the 'substantial reduction in energy consumption may be attributed to the sufficiently high spiking rate,' but higher firing rates increase AC count and energy. Also Table 5 lists SEW ResNet18 for ImageNet while the main results use ResNet-18/34, and calling the module 'no additional computational burden' ignores the alpha scaling and residual addition.\n\nVerdict: yes, this deserves a serious referee. The core idea is simple and likely sound as an offline module, the experiments are reasonably thorough, and the issues are addressable with a causal-variant test, a same-backbone baseline, and a cleaned-up energy section. If I were handling it, I would ask for those before publication. For a reading group, it's a useful example of recycling a known trick into a new domain, but not a must-read.","headline":"A competent transfer of TSM to SNNs with good ablations, but the 'future features' claim assumes offline processing; the causal gap should be tested.","tokens_in":18318,"tokens_out":3174,"would_cite":false,"duration_ms":30209,"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":"A zero-FLOP temporal shift lets spiking networks see past and future spikes in one timestep, lifting accuracy on CIFAR-10, CIFAR-100, ImageNet, and CIFAR10-DVS.","keywords":["spiking neural networks","temporal shift","temporal feature fusion","energy-efficient computing","surrogate gradient training","event-based vision","image classification","residual connection"],"falsifier":"Train the same TS-SNN on CIFAR-100 with a causal variant that replaces the $+1$ shift with a second $-1$ shift, using only past and present features, and compare accuracy at timesteps 2 and 4. If the causal variant matches the reported 80.28% accuracy, the future-access assumption is not load-bearing; if accuracy drops substantially, the reported gains depend on seeing future spikes at the current timestep.","tokens_in":17331,"feed_emoji":"⚡","tokens_out":7470,"duration_ms":63443,"temperature":0.7,"pith_summary":"This paper claims that a simple temporal shift of spike-feature channels can give spiking neural networks (SNNs) much of the temporal modeling power they otherwise lack, at nearly zero computational cost. The proposed TS module shifts part of the channels one timestep forward, part one timestep backward, and leaves the rest in place, so that each timestep's computation sees a blend of past, present, and future spikes. A scaled residual connection adds the shifted tensor back to the original features using one learnable parameter, which prevents information loss and stabilizes training. On CIFAR-10, CIFAR-100, ImageNet, and the event-based CIFAR10-DVS, the resulting TS-SNN reports state-of-the-art accuracies with fewer timesteps than prior SNN methods. The claim matters because it suggests SNNs can close part of the accuracy gap with ANNs without sacrificing their low-energy advantage.","feed_headline":"One zero-FLOP shift lifts spiking-network accuracy on three benchmarks","feed_subtitle":"Fusing past and future spikes inside one timestep adds one parameter and nearly no compute.","key_machinery":"The load-bearing object is the Temporal Shift module defined by Equation (6): channels are split into $C_k$ groups of size $C_{\\text{fold}} = C/C_k$, two random split indices $g_1 < g_2$ are drawn, and the tensor is reassembled as $Z^{+1}=X[t+1, :g_1 \\cdot C_{\\text{fold}}, :, :]$, $Z^{-1}=X[t-1, g_1 \\cdot C_{\\text{fold}} : g_2 \\cdot C_{\\text{fold}}, :, :]$, and $Z^{0}=X[t, g_2 \\cdot C_{\\text{fold}} :, :, :]$, with zeros padding the temporal boundaries. This rearrangement turns each timestep's input into a channel-wise mixture of past, present, and future spikes at zero FLOP cost. The second piece is the residual combination $Z' = \\alpha \\odot Z + X$, where $\\alpha$ is the single learnable penalty factor that weighs the shifted features against the original ones, which prevents the original current-timestep information from being lost and keeps training stable.","core_discovery":"The paper's central claim is that fusing spike features across adjacent timesteps via a channel shift is enough to substantially improve SNN accuracy, and that this fusion can be done with only one additional learnable parameter and negligible extra compute. Concretely, for an input tensor $X$ of shape $T \\times C \\times H \\times W$, the module splits channels into segments and forms $Z$ with $Z^{+1}=X[t+1,\\ldots]$, $Z^{-1}=X[t-1,\\ldots]$, and $Z^{0}=X[t,\\ldots]$, padding zeros and truncating at the temporal edges, and then outputs $Z' = \\alpha \\odot Z + X$ with $\\alpha$ a scalar penalty factor (typically 0.2–0.5). The authors show that this module, inserted into ResNet and Spikeformer backbones, reaches 96.72% on CIFAR-10, 80.28% on CIFAR-100, 70.61% on ImageNet, and 83.90% on CIFAR10-DVS, often at smaller timesteps than competing SNNs, while consuming about 5.857 mJ per ImageNet image in their 45-nm energy model compared with 13.10 mJ for a SEW ResNet-18 baseline.","pith_inferences":["The future-shift term $Z^{+1}=X[t+1]$ means the reported accuracies describe an offline or buffered setting; a causal, online version of TS-SNN would need either a one-timestep output delay or a redesigned shift that uses only past and present features, and its accuracy may differ from the numbers reported.","Because the random split points are drawn per module, the effective receptive field in time varies across layers; this may act as a lightweight temporal dropout or augmentation, which could partly explain the gains beyond pure feature fusion.","A direct comparison between TS-SNN and an explicitly causal temporal fusion method, such as recurrent connections or causal convolutions over time, with matched compute would separate the benefit of temporal mixing per se from the benefit of seeing the future."],"forward_implications":["The TS module can be dropped into existing SNN backbones such as ResNet and Spikeformer with only one extra parameter, so any accuracy gain it provides is nearly free in terms of model size and compute.","Since the shift operation itself is a zero-FLOP rearrangement, TS-SNN preserves the accumulation-only energy advantage of SNNs over ANNs; the reported ImageNet energy is 5.857 mJ per image.","With timesteps of 1–4 on static images and 10–16 on event data, TS-SNN shows that strong SNN accuracy does not require long temporal sequences.","The authors report that consistent application of the TS module during both training and inference, rather than training only, improves accuracy without changing FLOPs; on CIFAR10-DVS it rises from 81.20% to 83.60%.","At $T=1$ the shift is a no-op, yet accuracy still improves slightly; the paper attributes this to the learnable scaling factor $\\alpha$, which can perturb features even when no temporal shift occurs."],"supporting_citations":[{"why":"It supplies the temporal shift idea (TSM) that the TS module adapts from video understanding to SNNs.","marker":"Lin et al., 2019"},{"why":"It provides the spatio-temporal backpropagation (STBP) training algorithm used for all experiments.","marker":"Wu et al., 2018b"},{"why":"It supplies the surrogate-gradient method that makes direct SNN training differentiable.","marker":"Neftci et al., 2019"},{"why":"It provides the SEW-ResNet backbone and the ImageNet energy baseline (13.10 mJ) that TS-SNN is compared against.","marker":"Fang et al., 2021a"},{"why":"It provides the Spikeformer architecture used to test the TS module's generality on transformer-based SNNs.","marker":"Zhou et al., 2022"},{"why":"It supplies the 45-nm energy per MAC (4.6 pJ) and per AC (0.9 pJ) figures used in the efficiency analysis.","marker":"Horowitz, 2014"},{"why":"It provides the CIFAR-10 and CIFAR-100 datasets used for the main classification benchmarks.","marker":"Krizhevsky et al., 2010"},{"why":"It provides the CIFAR10-DVS event-stream dataset used for the neuromorphic benchmark.","marker":"Li et al., 2017b"},{"why":"It provides the ImageNet dataset used for the large-scale evaluation.","marker":"Deng et al., 2009"}],"fun_headline_variants":["Temporal shift fuses spikes across time with one extra parameter","One-parameter shift lifts spiking accuracy on CIFAR, ImageNet","Zero-FLOP spike shift: past and future in a single timestep","TS-SNN: shift spikes across timesteps for state-of-the-art accuracy","Lightweight spike shift boosts SNN accuracy with minimal compute"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The module assumes that at timestep $t$ the network can already read spike features from timestep $t+1$, which is only true if the full spike sequence is buffered or the input is processed offline; in strictly causal online inference that future information does not exist yet.","fun_headline_variants_meta":{"raw":{"variants":["Temporal shift fuses spikes across time with one extra parameter","One-parameter shift lifts spiking accuracy on CIFAR, ImageNet","Zero-FLOP spike shift: past and future in a single timestep","TS-SNN: shift spikes across timesteps for state-of-the-art accuracy","Lightweight spike shift boosts SNN accuracy with minimal compute"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000996,"raw_usage":{"total_tokens":4256,"prompt_tokens":1023,"completion_tokens":3233,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":639,"completion_tokens_details":{"reasoning_tokens":3151}},"tokens_in":639,"tokens_out":3233,"duration_ms":24836,"temperature":1.0,"reasoning_tokens":3151,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:35:26.811524+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same TS-SNN on CIFAR-100 with a causal variant that replaces the $+1$ shift with a second $-1$ shift, using only past and present features, and compare accuracy at timesteps 2 and 4. If the causal variant matches the reported 80.28% accuracy, the future-access assumption is not load-bearing; if accuracy drops substantially, the reported gains depend on seeing future spikes at the current timestep.","supporting_citations":[{"cited_title":"TSM : Temporal Shift Module for Efficient Video Understanding","cited_arxiv_id":null,"evidence_quote":"It supplies the temporal shift idea (TSM) that the TS module adapts from video understanding to SNNs."},{"cited_title":"Spikformer: When Spiking Neural Network Meets Transformer","cited_arxiv_id":null,"evidence_quote":"It provides the Spikeformer architecture used to test the TS module's generality on transformer-based SNNs."},{"cited_title":"Cifar-10 (canadian institute for advanced research)","cited_arxiv_id":null,"evidence_quote":"It provides the CIFAR-10 and CIFAR-100 datasets used for the main classification benchmarks."}],"review_version":1}