{"id":"e74aa740-cd8b-4a12-9019-8ab028e75655","arxiv_id":"2411.17439","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"SpikeAtConv reports state-of-the-art 81.23% top-1 ImageNet accuracy for a directly trained spiking network, but the defining attention module is unspecified and no energy data is provided.","lead":"The paper proposes SpikeAtConv, a spiking neural network that blends MaxViT-style attention with Leaky Integrate-and-Fire neurons, and reports 81.23% top-1 accuracy on ImageNet-1K. It is relevant because it targets energy-efficient vision, but it ships no code, no energy measurements, and no equations for its key attention module.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The ImageNet SOTA claim is untraceable: Spike-Driven Grid Attention is never formally defined, and the paper never states whether the reported models use SISA or BDSA; if the degenerate BDSA was used, the claimed attention mechanism contributes nothing.","rationale":"This concern is load-bearing because the paper's stated contribution is the architecture, not the training recipe. The empirical accuracy could be real, but the paper's own admission that a degenerate attention does not hurt performance suggests that the spiking 'attention' is not doing the work. More fundamentally, without a precise specification, no one can verify that the model in the paper actually contains the described innovation. This is a reproducibility failure, not a disagreement with the community. I credit the authors for reporting a spike-free control and for including training details; however, those supports do not repair the missing definition of the central module. A full rewrite with formal definitions, ablation of SISA vs BDSA, and parameter counts would be needed before the SOTA claim can be assessed.","tokens_in":10301,"tokens_out":5452,"duration_ms":56722,"concrete_test":"Determine the exact attention block used in the Large SpikeAtConv model by inspecting the authors' code or asking them. If it is the BDSA block (Q=K=V after a common SPK transform), run an ImageNet-1K ablation in which the entire attention module is removed, leaving only the ConvNeXt and MBPL components. If top-1 accuracy changes by less than 0.3% (the expected noise from one training run), then the reported SOTA accuracy is independent of the claimed attention mechanism, invalidating the central architectural contribution. Additionally, require the authors to provide a formal definition of Spike-Driven Grid Attention and a table tagging each Table 4 model with SISA vs. BDSA.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is that SpikeAtConv, built on a novel 'Spike-Driven Grid Attention,' achieves 81.23% top-1 on ImageNet-1K. For this claim to be meaningful, the attention mechanism must be specified and the reported model must actually use it. The paper fails on both requirements. Section 3.1 names Spike-Driven Grid Attention but provides no equations, pseudocode, or figure detailing it; Section 3.3 defines two different Attention SpikeMerge Blocks (SISA and BDSA) but never identifies which one is used in the ImageNet models of Table 4. The BDSA description explicitly states it 'bypasses the computation of Q, K, and V' and treats Q, K, V as identical after a common spike transform, i.e., it is a degenerate attention that does no query-key-value mixing. The conclusion further admits that 'even when using a degenerate self-attention mechanism, the performance of our model did not degrade significantly.' If the flagged Large model uses BDSA, then the claimed grid-attention innovation is not present in the SOTA result. Consequently, the 81.23% number cannot be attributed to the proposed architecture, and no independent replication is possible from the text. The omission of the parameter row in Table 4 compounds this: without parameter counts, the SOTA comparison to Meta-SpikeFormer is not even apples-to-apples.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes SpikeAtConv, a spiking neural network architecture that combines MaxViT-inspired multi-axis grid attention with convolutional blocks and several custom LIF-based SPK blocks. The authors report CIFAR-100 experiments that compare SPK block variants and ImageNet-1K experiments for Tiny, Base, and Large models, claiming a state-of-the-art 81.23% top-1 accuracy among directly trained SNNs. The central claim is that the proposed Spike-Driven Grid Attention and MBPL spiking neurons enable this result. However, the attention mechanism is never formally specified, the paper does not state which of the two proposed attention variants is used in the ImageNet experiments, and the reported spike-free baseline nearly matches the claimed SOTA, so the central claim is not currently verifiable.","tokens_in":10634,"tokens_out":4334,"duration_ms":43507,"significance":"If the 81.23% ImageNet-1K result were fully specified and reproducible, it would be a notable advance for directly trained SNNs and would strengthen the case for spike-based attention and multi-threshold spiking neurons. The paper deserves credit for conducting systematic comparisons of SPK block configurations and for reporting a spike-free baseline, which is a useful control. However, the significance is strongly conditional: the core attention module is undefined, the models are not identified by attention variant, parameter counts are missing from Table 4, no error bars or seeds are reported, and no code or checkpoints are provided. As written, the contribution cannot be verified or attributed to the proposed architecture.","major_comments":[{"comment":"The central contribution, Spike-Driven Grid Attention, is never specified. Section 3.1 names the module, but no equations, pseudocode, or architectural diagram define how Q, K, and V are formed or how grid attention is spikified. Section 3.3 introduces two variants, SISA and BDSA, but neither the experimental section nor Table 4 states which variant is used in the reported ImageNet models; BDSA explicitly bypasses the computation of Q, K, and V and treats them as identical, so if the Large model uses BDSA, the claimed grid-attention mechanism is absent from the SOTA result. The conclusion's admission that a degenerate self-attention mechanism did not degrade performance significantly reinforces this concern. The 81.23% top-1 result therefore cannot be attributed to the proposed attention architecture, and the experiment is not independently reproducible from the text.","section":"§3.1, §3.3, Table 4"},{"comment":"The evidence does not support the claim that the spiking components improve accuracy. In Table 3, the spike-free baseline reaches 81.13% top-1, which exceeds the spiking Base model (80.53%) and is within 0.10 points of the Large model (81.23%); no error bars, seeds, or significance tests are reported, so these small differences are not established as meaningful. Additionally, Table 4 is announced in §4.3 as including the number of parameters for each model, but the parameter row is missing, so the comparison to Meta-SpikeFormer and SpikFormer is not apples-to-apples in model scale or compute.","section":"Table 3, Table 4, §4.3"},{"comment":"The hyperparameter selection procedure creates a model-selection risk: MBPL branch counts, voltage thresholds, time steps, and surrogate functions are tuned on CIFAR-100 in Tables 1 and 2 and then transferred to ImageNet without any correction for selection bias or evaluation of variance. Because the ImageNet numbers in Tables 3 and 4 are single runs, it is unclear whether the reported accuracies reflect the architecture's average behavior or a favorable configuration selected through the CIFAR-100 loop.","section":"§4.1, §4.2"}],"minor_comments":[{"comment":"The abstract and introduction use the term 'pulse modules' while the rest of the paper refers to 'SPK Blocks'; please unify the terminology.","section":"Abstract, §1, §3"},{"comment":"Contribution 4 contains a typo: 'eperiments' should be 'experiments'.","section":"§1, Contribution 4"},{"comment":"The discussion of Table 3 refers to an 'SSL' module, but the defined module is 'HSL' (Hidden Split LIF); this typo should be corrected.","section":"§4.3, after Table 3"},{"comment":"The caption for Figure 4 says the zoomed panel covers the first 20 epochs, but the axis is labeled 1 through 10; align the caption with the axis range.","section":"Figure 4"},{"comment":"The DCL Block description says each branch halves the channel dimension and the outputs are concatenated, but it is not stated how the 3x3 and 5x5 branches handle odd channel counts; please clarify the dimension arithmetic.","section":"§3.2, DCL Block"},{"comment":"The paper does not state whether code or trained checkpoints will be released; providing them would substantially improve the reproducibility of the reported ImageNet results.","section":"Overall"}],"recommendation":"reject","confidential_remarks":"The manuscript's central claim rests on an architecture component that is never specified and on a single ImageNet run with no code or checkpoints. The self-reported spike-free baseline nearly matches the claimed SOTA, and the conclusion explicitly concedes that the degenerate attention variant performs comparably. These issues are not merely presentational: they invalidate the attribution of the reported result to the proposed method. I recommend rejection, though a substantially revised manuscript with full equations, an explicit statement of which attention variant is used, parameter counts, and reproducible experiments could warrant resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper has a plausible new architecture combination: MaxViT-style convolution-plus-attention backbones with multi-branch LIF spiking neurons. The MBPL block, with parallel LIF neurons at different thresholds, is a sensible idea, and the CIFAR-100 sweep to tune it is a legitimate model-selection exercise. The authors also deserve credit for reporting a spike-free baseline (81.13%) that nearly matches the best spiking model (80.53 or 81.23%). That is an honest control, and it tells you something useful: the spiking components are not hurting much, but they may not be helping either.\n\nThe problems are load-bearing, not cosmetic. Spike-Driven Grid Attention is named in the introduction and Section 3.1 but never defined with equations, pseudocode, or even a precise verbal description. Section 3.3 defines two variants, SISA and BDSA, and BDSA explicitly treats Q, K, and V as identical, which is degenerate attention. The paper never says which variant the ImageNet models in Table 4 use. The conclusion then states that even with degenerate self-attention, performance did not degrade significantly. So the 81.23% number cannot be attributed to the proposed grid-attention mechanism, and independent replication is impossible from the text. Table 4 also omits parameter counts for the models, making the comparison to Meta-SpikeFormer not apples-to-apples. No code, checkpoints, or error bars are provided, and the title's energy-efficiency claim is never measured.\n\nThese are not minor omissions; they undermine the paper's central contribution. That said, the MBPL analysis is a reasonable piece of empirical work, and the authors' willingness to report the spike-free baseline suggests they are not trying to hide the weakness. If they resubmit with a full specification of the attention module, a clear statement of which variant was used, code, parameter counts, and energy benchmarks, the paper could be worth serious review. As it stands, the SOTA claim is untraceable and the proposed mechanism is unvalidated.\n\nWho is this for? Someone working on SNN architecture design might skim the MBPL experiments, but the paper does not provide enough detail to build on. I would not bring it to a reading group for the science, though it could serve as a case study in how a promising architecture can be undercut by under-specification.\n\nRecommendation: desk reject. The structural gaps are too large for peer review to fix; the authors need to do substantial additional work first.","headline":"The architecture idea is real but the headline SOTA claim is untraceable: the core grid-attention module is never specified, and the paper never says whether the ImageNet models even use the non-degenerate attention variant.","tokens_in":2,"tokens_out":2248,"would_cite":false,"duration_ms":63724,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A directly trained spiking network with multi-threshold LIF neuron blocks reports 81.23% top-1 accuracy on ImageNet-1K, which the authors describe as a state-of-the-art result for SNNs.","keywords":["spiking neural networks","neuromorphic vision","LIF neuron","multi-branch parallel LIF","spike-driven attention","ImageNet-1K","direct training","MaxViT"],"falsifier":"Train the Large SpikeAtConv on ImageNet-1K with the Spike-Driven Grid Attention module removed and replaced by an identity, keeping every other hyperparameter fixed; if top-1 accuracy stays near 80.5% or higher, the attention module is not load-bearing, while a clear drop below the Base model level would confirm its contribution.","tokens_in":10093,"feed_emoji":"⚡","tokens_out":9135,"duration_ms":75866,"temperature":0.7,"pith_summary":"SpikeAtConv aims to show that a spiking neural network, not converted from a trained artificial network, can compete with conventional vision models on large-scale image classification. The authors build a MaxViT-style hybrid of convolution and attention, replace activations with spiking neuron blocks, and report 81.23% top-1 accuracy on ImageNet-1K for their largest model, which they call a state-of-the-art result among SNNs. The main design lesson is that running several LIF neurons with different thresholds in parallel, rather than a single spiking neuron, preserves more information and raises accuracy. The authors also report that a simplified 'binary direct-spike' attention variant, which skips computing query, key, and value, did not significantly degrade performance, suggesting attention complexity can be cut. If these results hold, directly trained SNNs are closer to ANN accuracy on hard vision tasks than earlier results suggested.","feed_headline":"Spiking network reports 81.23% top-1 on ImageNet-1K","feed_subtitle":"A directly trained SNN with parallel LIF neurons and spike-form attention closes the gap with conventional vision models.","key_machinery":"The load-bearing object is the Leaky Integrate-and-Fire (LIF) neuron and its variants, particularly the Multi-Branch Parallel LIF (MBPL) block: several LIF neurons with different thresholds process the same features in parallel, and their outputs are combined before a ConvNeXt module, compensating for information lost when a single neuron converts continuous activations into binary spikes. The second mechanism is the Attention SpikeMerge Block, with two modes: Spike-Integrated Self-Attention (SISA), which puts spiking blocks around the Q/K/V computation and after the attention map, and Binary Direct-Spike Attention (BDSA), which skips Q/K/V entirely and multiplies identical binary spike vectors. The paper also names Spike-Driven Grid Attention as the module enabling global spatial interactions, but supplies no equations or pseudocode for it; its role in the argument is to inherit MaxViT's multi-axis attention design.","core_discovery":"The paper's central claim is that a directly trained SNN can reach 81.23% top-1 on ImageNet-1K, exceeding the comparison SNN models Meta-SpikeFormer (79.1%) and SpikFormer (74.8%) and setting what the authors call a state-of-the-art result for SNNs. The claim is carried by SpikeAtConv, a four-stage spiking network whose blocks combine ConvNeXt-style convolutions with a Spike-Driven Grid Attention module. A second claim is that the choice of spiking neuron configuration matters: the Multi-Branch Parallel LIF block, with four parallel LIF neurons at thresholds 0.2, 1, 2, and 4, outperforms single, residual, hidden-split, and dual-convolution blocks on both CIFAR-100 and ImageNet-1K. A third claim is that a degenerate attention mode that treats Q, K, and V as identical spikes does not significantly hurt accuracy, which the authors read as evidence that their spiking blocks carry most of the representational load.","pith_inferences":["Beyond the paper: the near-tie between the spike-free model (81.13%) and the best spiking model (80.53%) suggests the spike encoding itself costs roughly 0.6 top-1 points on ImageNet-1K; closing that gap is the real target for future SNN designs.","Beyond the paper: because Spike-Driven Grid Attention is never ablated, the most economical reading is that accuracy comes mainly from the SPK blocks and ConvNeXt backbone, not from attention; replacing the grid-attention block with an identity should be tested.","Beyond the paper: the energy-efficiency advantage of SNNs remains asserted rather than measured; an energy simulation or neuromorphic hardware study would be needed to turn the accuracy result into an efficiency claim.","Beyond the paper: applying MBPL blocks to spike-driven transformers for detection or segmentation would test whether multi-threshold parallel neurons help dense prediction tasks, not just classification."],"forward_implications":["If the 81.23% result reproduces, directly trained SNNs can exceed 81% top-1 on ImageNet-1K with 224x224 input, a level previously associated with ANN-style performance.","The MBPL configuration becomes a default building block for spike-based vision backbones, since it beat single-neuron and residual variants on both CIFAR-100 and ImageNet-1K.","The BDSA result implies that attention-weight computation can be dropped or drastically simplified in spike-form transformers, reducing multiply-accumulate operations without proportional accuracy loss.","The authors' observation that LIF-based modules lose information suggests architectures should route redundant features around spiking activations, not just replace ReLU with spiking neurons."],"supporting_citations":[{"why":"Supplies the MaxViT multi-axis attention design (window plus grid attention) that SpikeAtConv adapts into spike form.","marker":"[29]"},{"why":"SpikFormer, a transformer-based SNN baseline that SpikeAtConv compares against on ImageNet-1K.","marker":"[12]"},{"why":"The spike-driven transformer that motivates spike-form attention; the SISA and BDSA blocks extend this line.","marker":"[13]"},{"why":"Direct training of deeper SNNs with temporal batch normalization, the training regime SpikeAtConv relies on.","marker":"[20]"},{"why":"Deep residual learning in SNNs, supporting the residual connections used inside the spiking blocks.","marker":"[21]"},{"why":"The Vision Transformer, whose self-attention the spiking transformer modules adapt into spike sequences.","marker":"[22]"},{"why":"ConvNeXt v2, source of the convolutional component used inside the SpikeAtConv block.","marker":"[28]"},{"why":"ImageNet-1K, the benchmark dataset on which the main 81.23% result is measured.","marker":"[33]"}],"fun_headline_variants":["SpikeAtConv hits 81.23% top-1 on ImageNet-1K","Directly trained SNN achieves 81.23% top-1","Spike-driven attention sets SNN state-of-the-art","SpikeAtConv narrows gap with CNNs on ImageNet"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the spike-form grid attention module preserves or improves MaxViT's representational capacity; the paper provides no equations, pseudocode, or ablation isolating that module, so if the premise fails, the reported accuracy cannot be attributed to the claimed attention mechanism.","fun_headline_variants_meta":{"raw":{"variants":["SpikeAtConv hits 81.23% top-1 on ImageNet-1K","Directly trained SNN achieves 81.23% top-1","Spike-driven attention sets SNN state-of-the-art","SpikeAtConv narrows gap with CNNs on ImageNet"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000196,"raw_usage":{"total_tokens":1335,"prompt_tokens":896,"completion_tokens":439,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":368}},"tokens_in":512,"tokens_out":439,"duration_ms":4278,"temperature":1.0,"reasoning_tokens":368,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:06:34.542801+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the Large SpikeAtConv on ImageNet-1K with the Spike-Driven Grid Attention module removed and replaced by an identity, keeping every other hyperparameter fixed; if top-1 accuracy stays near 80.5% or higher, the attention module is not load-bearing, while a clear drop below the Base model level would confirm its contribution.","supporting_citations":[{"cited_title":"Spik- former: When spiking neural network meets transformer","cited_arxiv_id":null,"evidence_quote":"SpikFormer, a transformer-based SNN baseline that SpikeAtConv compares against on ImageNet-1K."},{"cited_title":"Spike-driven transformer","cited_arxiv_id":null,"evidence_quote":"The spike-driven transformer that motivates spike-form attention; the SISA and BDSA blocks extend this line."},{"cited_title":"Deep residual learning in spiking neural networks","cited_arxiv_id":null,"evidence_quote":"Deep residual learning in SNNs, supporting the residual connections used inside the spiking blocks."}],"review_version":1}