{"id":"ab4ee204-80a9-4878-8396-678fadeee733","arxiv_id":"2501.14744","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"FSTA-SNN introduces a frequency-based spatial-temporal attention module for spiking neural networks, cutting spike firing rate by about 34 percent and improving accuracy on CIFAR-10/100, ImageNet, and CIFAR10-DVS.","lead":"FSTA-SNN analyzes spike outputs of spiking neural networks in the frequency domain and uses the findings to design a plug-in attention module that suppresses redundant spikes. The module improves classification accuracy on CIFAR, ImageNet, and CIFAR10-DVS while cutting the network firing rate by about 34 percent.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No controlled ablation of the same backbone with and without FSTA; reported accuracy gains may reflect training recipe rather than the module.","rationale":"The reader's CONDITIONAL verdict is well supported. I searched for a controlled comparison that would isolate the FSTA module's contribution. The paper does not provide one: Fig. 4 compares firing rates between FSTA and vanilla SNN, but no table reports vanilla-SNN accuracy under the same training conditions. Table 4 varies only combinations of FSTA submodules, and Tables 1-3 compare to published SOTA numbers from other papers. In a plug-and-play module paper, the natural control is 'same backbone, same training pipeline, module on/off'. Without it, the headline accuracy improvements are vulnerable to confounding by training recipe differences. This is a correctness risk, not a novelty or style issue, and it is distinct from but related to the reader's concern about the spectral motivation. The temporal attention equations (Eqs. 16-18) also contain an apparent dimension inconsistency: averaging M over the temporal dimension would destroy the per-timestep variation needed for Tw, but I did not make that the primary concern because the intended implementation can likely be inferred from Fig. 2 and the section's purpose. The proposed concrete test directly settles whether FSTA itself causes the reported accuracy gains. If the ablation shows a clear gain, the central claim survives; if not, the verdict should be reconsidered. Because the evidence is currently incomplete rather than definitively wrong, CONDITIONAL remains the appropriate verdict, so no change to the reader's decision is recommended.","tokens_in":12711,"tokens_out":9411,"duration_ms":98017,"concrete_test":"Run the authors' released training pipeline on CIFAR100 with ResNet20, then repeat with the FSTA module removed by setting Xo = X in Eq. 22 while keeping all hyperparameters, seeds, and training schedules identical. Compare final top-1 accuracy at T=2 and T=4 to 72.15% and 73.44%; also record ACs, MACs, and estimated energy for both runs. If the no-FSTA backbone matches or exceeds FSTA accuracy, the accuracy claim is not attributable to the module; if it is clearly lower, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest claim is that inserting FSTA into standard SNN backbones improves accuracy and reduces firing rate. The firing-rate comparison in Fig. 4 does include a 'vanilla SNN', but no accuracy table reports the identical backbone trained without FSTA under the same pipeline. Tables 1-3 compare against published SOTA numbers, e.g., CIFAR10-DVS ResNet20 at T=10: MPBN 78.70% vs FSTA 81.50%, and ImageNet ResNet34: Real Spike 67.69% vs FSTA 70.23%. Table 4 ablates only FSTA submodule combinations (modes a/b/c); it never removes FSTA entirely. Consequently, the reported gains could be driven by training schedule, augmentation, normalization, or other recipe differences rather than by the module itself. This is especially load-bearing because FSTA is advertised as a plug-and-play module whose contribution is a 'minimal number of additional parameters' and 'no significant increase in energy consumption'. If the same backbone without FSTA matches or exceeds FSTA accuracy, the central contribution is not supported. The energy claim in Tab. 5 has the same omission: no vanilla-SNN ACs/MACs/energy baseline is provided, so the '33.99% firing-rate reduction without energy increase' cannot be assessed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FSTA, a plug-and-play attention module for spiking neural networks, based on a frequency-domain analysis of intermediate spike outputs. The spatial submodule applies fixed DCT convolution kernels to extract full-spectrum features, while the temporal submodule uses average/max pooling with learnable parameters to rescale spike features across time steps. Experiments on CIFAR-10/100, ImageNet, and CIFAR10-DVS report accuracy improvements over published state-of-the-art results and a 33.99% reduction in spike firing rate, with a small claimed increase in computational cost. The paper also includes a proof that global average pooling corresponds to the lowest-frequency component of the 2D DCT.","tokens_in":13010,"tokens_out":6926,"duration_ms":57076,"significance":"If the reported gains are robust, FSTA is an inexpensive and architecture-agnostic addition that improves accuracy while reducing spike firing rate, and the GAP-as-lowest-frequency observation is a clean, checkable insight. The manuscript provides code, reports error bars on key results, and evaluates on both static and neuromorphic datasets. However, the central causal claim is currently supported only by comparisons against published numbers and by ablations that never remove the module entirely. The significance of the contribution is therefore conditional on controlled same-pipeline experiments being added.","major_comments":[{"comment":"The paper's central claim is that inserting FSTA into standard SNN backbones improves accuracy and reduces firing rate, yet no experiment compares a given backbone trained with FSTA against the same backbone trained without FSTA under the identical pipeline. Tables 1–3 compare against published numbers from other papers, and the ablation in Table 4 only compares submodule combinations (modes a/b/c); it never removes FSTA entirely. Because training schedules, augmentations, normalization, and other recipe details can easily account for several accuracy points in the SNN literature, this omission is load-bearing. Please add a controlled ablation (e.g., ResNet20/19 on CIFAR-100, ResNet20 on CIFAR10-DVS, ResNet34 on ImageNet) with and without FSTA, reporting mean and standard deviation, and report the corresponding firing-rate comparison under the same recipe.","section":"Tables 1–3 and Table 4"},{"comment":"The temporal attention submodule is not fully specified. With X in R^{T,C,H,W}, average/max pooling over spatial dimensions yield tensors in R^{T,C,1,1}, so M = alpha*favg + beta*fmax is in R^{T,C,1,1}, not R^{T,C} as written; Eq. (17) then averages over the temporal dimension to produce a C-dimensional vector, which cannot be linearly mapped to a T-dimensional weight vector Tw in Eq. (18). Please clarify the intended axes, tensor shapes, whether alpha/beta are per-channel or scalar, and how Tw is obtained. As written, the temporal mechanism cannot be reproduced from the equations alone.","section":"Equations (16)–(18)"},{"comment":"The energy and firing-rate claims are not quantitatively grounded. Table 5 reports ACs, MACs, FLOPs, and energy only for FSTA-equipped models, with no vanilla-SNN baseline, so the claim that the module does not significantly increase energy consumption is unverified. Figure 4 compares FSTA-SNN with a vanilla SNN, but it is not stated whether that vanilla SNN uses the same training recipe, initialization, and time steps; without this, the 33.99% firing-rate reduction cannot be attributed to the module. Please include the vanilla baseline in Table 5 and describe the energy-estimation methodology (e.g., per-AC and per-MAC energy constants) in the text.","section":"Table 5 and Fig. 4"},{"comment":"The frequency-analysis observations that motivate the entire design are stated qualitatively. Terms such as 'remarkably similar', 'significant overlap', and 'gradually shifts' are not accompanied by quantitative measurements (e.g., spectral-energy ratios, correlation, or KL divergence between layers/time steps) or by a description of how the DFT magnitudes are normalized and averaged over the test set. Since the architectural choices in Eqs. (13)–(22) depend on these observations, please add quantitative support or explicitly reframe them as informal intuitions rather than empirical findings.","section":"Observations 1 and 2"}],"minor_comments":[{"comment":"The equality in Eq. (11) holds only up to the normalization constants of the DCT basis; please state that the result is proportional to the lowest-frequency component rather than exactly equal, or specify the DCT normalization convention used.","section":"Equation (11)"},{"comment":"The text refers to Eq. (7) to justify the spectral behavior, but Eq. (7) is the IDFT; Eq. (8) or Eq. (10) seems intended.","section":"Observation 1"},{"comment":"Scale_t and Scale_s are introduced but never defined; please state whether they are learnable, how they are initialized, and whether they are per-layer or global.","section":"Equation (22)"},{"comment":"Table 4 does not state the dataset, network, and time step in the caption; these details appear only in the text and should be moved into the table caption.","section":"Table 4"},{"comment":"The section heading 'Comparion with SOTA methods' contains a typo and should read 'Comparison with SOTA methods'.","section":"Section heading"},{"comment":"The caption of Figure 1 should describe the preprocessing used to produce the spectra (e.g., which spike outputs are included, how magnitudes are normalized, how many samples are averaged, and the frequency-axis convention); as written, the analysis cannot be reproduced.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The main uncertainty is whether the accuracy and firing-rate gains survive a same-pipeline control with the module removed; this is a fixable experimental omission rather than a fundamental flaw. I would not recommend rejection on the current evidence, but acceptance should wait for the controlled ablation and a cleaned-up specification of the temporal attention module."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, well-written SNN paper with a practical contribution, but the main empirical claim is under-supported by a missing controlled ablation. Worth reading if you work on SNN attention or energy-efficient training; if you're reviewing, ask for the controlled experiment.\n\nWhat's new: the paper applies a full-spectrum DCT spatial attention to SNNs, combined with a temporal attention that shares parameters across time steps. The DCT mechanism is adapted from FCANet (acknowledged), but the SNN frequency-domain framing and the specific module are new. They also contribute a frequency-domain analysis of intermediate spikes, though it's qualitative.\n\nWhat's good: the module is plug-and-play, adds few parameters, and reported numbers are strong: ImageNet ResNet34 70.23% vs 67.69% for Real Spike, CIFAR10-DVS 82.70% at T=16, and a 33.99% firing-rate reduction. They give error bars on key results. The DCT spatial attention equations are clear.\n\nSoft spots, in order: (1) Missing controlled ablation. Tables 1-3 compare against published SOTA; Table 4 ablates submodule combos but never removes FSTA entirely. So you can't tell if gains come from the module or from the training recipe. Figure 4 shows firing rates vs vanilla SNN but not accuracy. This is load-bearing. (2) Temporal attention equations are ambiguous — Eq. 16-18 have shape mismatches; fixable but unclear. (3) The energy claim in Tab. 5 lacks a vanilla SNN baseline. (4) The Fig. 1 frequency observations are qualitative; no quantitative or statistical support.\n\nThe central idea is plausible and gains are consistent across datasets, so not a dud. But the paper overstates evidentiary strength. A same-pipeline with/without FSTA ablation on one or two datasets would raise my confidence substantially.\n\nRecommendation: send to peer review, asking for the controlled ablation and equation clarification. Useful contribution needing tightening, not rejection.\n\nI'd cite only after the ablation checks out.","headline":"FSTA-SNN reports strong accuracy and spike-reduction gains for a plug-in SNN attention module, but a missing same-backbone ablation leaves the central claim less secure than the paper suggests.","tokens_in":13505,"tokens_out":3639,"would_cite":false,"duration_ms":29267,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"DCT attention cuts SNN spike firing by a third, lifting accuracy.","keywords":["spiking neural networks","frequency-domain analysis","discrete cosine transform","spatial-temporal attention","spike firing rate reduction","energy-efficient computing","neuromorphic computing","attention module"],"falsifier":"Train the same ResNet and VGG SNNs on the same datasets with a handful of random seeds and training schedules, then compute the centered 2D DFT of intermediate spike outputs at each layer and time step; if the shallow-layer horizontal-axis concentration, the deep-layer vertical-axis shift, or the cross-time-step spectral overlap fails to appear consistently across seeds, architectures, or datasets, the claimed universal learning preference is contradicted.","tokens_in":12520,"feed_emoji":"⚡","tokens_out":11296,"duration_ms":80895,"temperature":0.7,"pith_summary":"The paper claims that spiking neural networks (SNNs) have consistent, architecture-independent learning preferences that can be read off the Fourier spectrum of their intermediate spike outputs: shallow layers emphasize vertical feature variations, deeper layers shift to horizontal variations, and the frequency content of a layer barely changes across time steps. On the strength of those observations, the authors design a plug-in Frequency-based Spatial-Temporal Attention (FSTA) module that uses a fixed discrete cosine transform (DCT) to extract full-spectrum spatial features and a small temporal attention submodule to reweight time steps. Adding FSTA to standard SNN backbones, they report higher top-1 accuracy on CIFAR-10/100, ImageNet, and CIFAR10-DVS than the baselines they compare against, while cutting total spike firing rate by about 33.99% at roughly constant energy cost. The paper's contribution is both a diagnostic claim about where SNN learning happens in frequency space and a practical mechanism that exploits that structure.","feed_headline":"DCT attention cuts SNN spike firing by a third, lifting accuracy","feed_subtitle":"Plug-and-play spatial-temporal attention lifts ImageNet and CIFAR10-DVS accuracy while shrinking spike counts.","key_machinery":"The load-bearing identity is that global average pooling (GAP) equals the $(0,0)$ coefficient of the 2D discrete cosine transform up to the constant factor $H \\times W$, which the paper proves by evaluating the DCT basis at $u=v=0$. From this it argues that conventional spatial attention built on GAP only sees the lowest frequency band, and it replaces that with a non-trainable convolution whose fixed weights are the DCT basis functions, extracting the full frequency spectrum of the temporally averaged feature map. This DCT-based spatial attention submodule is paired with a temporal attention submodule that pools along time with learned balances between average and max pooling, then weights each time step. The two submodules run in series with learnable scale factors, producing the FSTA output; the fixed DCT kernels keep the added parameter count and floating-point cost low.","core_discovery":"The central claim is that SNN intermediate spike outputs have a stable frequency signature: the Fourier spectrum of shallow-layer spikes is concentrated along the central horizontal axis, which the authors interpret as a preference for vertical feature variations, and as depth increases the spectrum shifts toward the vertical axis, indicating a preference for horizontal variations. Across time steps within a layer, the spectrum remains nearly identical except for amplitude, so the authors conclude that increasing simulation time adds little new feature information. They treat this spectral profile as a network-level learning preference that holds across ResNet and VGG architectures and across static and event-stream datasets, and use it to motivate a module that suppresses redundant frequency components while amplifying preferred ones.","pith_inferences":["The authors do not test whether the spectral learning preferences appear across random seeds and training schedules; if they do, layer-wise frequency regularization could push SNNs toward even lower firing rates without retraining the attention module.","The fixed DCT kernels imply the benefit comes largely from the attention mask shape rather than learned frequency filters; ablating the sigmoid and linear mapping would isolate how much of the gain is due to the frequency extraction itself.","The temporal-stability observation suggests that reducing the number of simulation steps, rather than just weighting them, may be a cheaper way to exploit the same insight; the paper does not experiment with step-count reduction.","If the shallow-vertical/deep-horizontal pattern generalizes to other spiking architectures or neuromorphic event datasets, frequency analysis could become a standard diagnostic for SNN layer design, though this remains an extension beyond the paper's evidence."],"forward_implications":["Plugging FSTA into ResNet34 lifts ImageNet top-1 accuracy to 70.23% at four time steps, up from 67.69% for the Real Spike baseline.","On CIFAR10-DVS, FSTA with ResNet20 reaches 82.70% at sixteen time steps, up from 78.70% for the MPBN baseline at ten steps.","The module reduces total spike firing rate by 33.99% across the network, so the accuracy gain comes with fewer spikes rather than more computation.","Because each layer's spectrum is stable across time steps, the spatial attention submodule can be shared across time, keeping the added parameter count minimal.","Global average pooling, the compression used by standard attention modules, is just the zero-frequency DCT coefficient; replacing it with fixed DCT kernels widens the frequency coverage of spatial attention."],"supporting_citations":[{"why":"Defines the discrete cosine transform whose basis functions the paper uses as fixed convolution kernels for full-spectrum spatial attention.","marker":"Ahmed, Natarajan, and Rao 1974"},{"why":"Introduces DCT into channel attention, the prior work this paper extends from channel to spatial full-spectrum attention.","marker":"Qin et al. 2021"},{"why":"Real Spike is the baseline whose ImageNet and CIFAR results the FSTA module improves upon.","marker":"Guo et al. 2022d"},{"why":"MPBN provides the strongest CIFAR10-DVS baseline (78.70%) that FSTA surpasses (82.70%).","marker":"Guo et al. 2023c"},{"why":"TET is a standard SNN training baseline that FSTA outperforms on CIFAR-10/100 and ImageNet.","marker":"Deng et al. 2022"},{"why":"CIFAR10-DVS is the dynamic event-stream dataset used to test the module's temporal attention.","marker":"Li et al. 2017"},{"why":"ImageNet is the large-scale static dataset used to demonstrate the module's scalability.","marker":"Deng et al. 2009"}],"fun_headline_variants":["Spectral spike analysis reveals depth-wise features; attention trims spikes","Frequency-based attention makes SNNs leaner and more accurate","SNN spike spectra show time steps add little; new module boosts efficiency","FSTA: DCT attention suppresses redundant spikes, lifts SNN accuracy","Spike spectra guide attention module that cuts firing rate and errors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Fourier spectra measured from the paper's trained networks are genuine, stable learning preferences of SNNs in general, rather than artifacts of the particular checkpoints, normalizations, or averaging choices used to draw Figure 1; if the shallow-to-deep vertical-to-horizontal shift and time-step invariance do not survive across seeds and architectures, the module's theoretical justification weakens even if its empirical gains persist.","fun_headline_variants_meta":{"raw":{"variants":["Spectral spike analysis reveals depth-wise features; attention trims spikes","Frequency-based attention makes SNNs leaner and more accurate","SNN spike spectra show time steps add little; new module boosts efficiency","FSTA: DCT attention suppresses redundant spikes, lifts SNN accuracy","Spike spectra guide attention module that cuts firing rate and errors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001411,"raw_usage":{"total_tokens":5681,"prompt_tokens":910,"completion_tokens":4771,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":526,"completion_tokens_details":{"reasoning_tokens":4681}},"tokens_in":526,"tokens_out":4771,"duration_ms":31441,"temperature":1.0,"reasoning_tokens":4681,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:17:34.175537+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same ResNet and VGG SNNs on the same datasets with a handful of random seeds and training schedules, then compute the centered 2D DFT of intermediate spike outputs at each layer and time step; if the shallow-layer horizontal-axis concentration, the deep-layer vertical-axis shift, or the cross-time-step spectral overlap fails to appear consistently across seeds, architectures, or datasets, the claimed universal learning preference is contradicted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the discrete cosine transform whose basis functions the paper uses as fixed convolution kernels for full-spectrum spatial attention."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces DCT into channel attention, the prior work this paper extends from channel to spatial full-spectrum attention."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ImageNet is the large-scale static dataset used to demonstrate the module's scalability."}],"review_version":1}