{"id":"81614f6c-b1ec-459d-ac9f-2bf1b39d8cdd","arxiv_id":"2608.11865","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Lapis replaces dot-product query-key scoring in spiking vision transformers with L1-distance-based Laplacian kernels on first-spike latencies, reaching near-dot-product accuracy at lower estimated arithmetic cost.","lead":"This paper introduces Lapis, a spiking attention mechanism that scores query-key pairs by the L1 distance between their first-spike firing times and then applies a Laplacian kernel, removing dot-product multiplications. On CIFAR-10 it lands within 0.53 points of dot-product attention, and on ImageNet it cuts the estimated attention-path energy by 14.5x.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never derives how spiking FC layers realize the quantized linear maps, so the claimed exact TTFS conversion—on which both SNN accuracy and energy numbers rest—is unsupported.","rationale":"The reader's conditional verdict is well aligned with the evidence. The ablation in Table 3 is the paper's strongest internal support: under fixed backbone, teacher, and schedule, Laplacian scoring is 0.53 points below softmax dot-product and far above Gaussian and Hamming alternatives. The mathematical claims that are actually derived (Propositions 1-2, Lemma 1, Eq. 20) are correct. The load-bearing gap is exactly the reader's weakest assumption: Eq. 20 covers the elementwise encoder only, not the spiking FC layers, so the assertion of exact conversion is not proven. This gap matters because both the reported SNN accuracy and the measured spike-activity statistics behind the energy model are properties of the converted network, not of the QNN. I could not find a stronger internal inconsistency; the energy model is transparent and, if anything, conservative in not charging the dense baseline for softmax. The missing conversion derivation or a numerical equivalence check would settle the concern. Since the reader already conditioned the verdict on this weakness, the verdict should remain unchanged rather than being tightened or loosened.","tokens_in":14684,"tokens_out":10409,"duration_ms":120686,"concrete_test":"Derive the exact neuron equations (threshold, reset, membrane time constant) for the spiking FC layers, implement the TTFS conversion, and compare, on random inputs, each layer's output first-spike latency to T - clamp(Wx+b) from the trained QNN. The exactness claim holds only if the maximum per-layer code mismatch is zero across all layers; any nonzero mismatch, or any need for calibration, refutes the claim and moves the burden to measuring the actual converted SNN's accuracy and energy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the deployed TTFS SNN reproduces the quantized network's Lapis affinities and attention weights. Section 4.5 establishes only that the elementwise encoding E(z) preserves query-key L1 distance (Eq. 20). It does not specify how the spiking FC layers (query, key, value, output projection) convert input TTFS latencies into output first-spike latencies equal to T - clamp(Wx+b). In a first-spike network, output firing time is governed by LIF threshold and integration dynamics, and in general is not a linear function of input latencies; exactness requires a concrete neuron parameterization plus a proof or verification. The text's assertion that 'the QNN and the converted SNN produce identical Lapis affinities and attention weights' is therefore not a derived consequence. If conversion error exists, the ImageNet-1K accuracies in Table 1 and the spike activities used in the energy model (Eqs. 21-22) may not correspond to the QNN ablations in Table 3. Since the paper's headline accuracy-energy trade-off is about the deployed SNN, this unsupported bridge is the most load-bearing weakness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Lapis, an attention mechanism for spiking vision transformers in which query-key affinity is computed as exp(-||t_Q - t_K||_1 / tau) over time-to-first-spike latency vectors, with row normalization approximated by power-of-two scaling. The authors train quantized networks with the Lapis operator, convert them to TTFS spiking networks via an elementwise spike encoder, and report CIFAR-10/100 and ImageNet-1K accuracy together with 45nm operation-level energy estimates. The paper includes proofs of the leakage-recurrence equivalence (Prop. 1), properties of the temporal affinity (Prop. 2), and the power-of-two approximation bound (Lemma 1), and it contains a controlled ablation isolating the relation function.","tokens_in":14831,"tokens_out":12139,"duration_ms":123129,"significance":"If the full-network TTFS conversion were substantiated, the paper would make a meaningful contribution: it shows that a multiplication-free L1/Laplacian scoring rule can remain within 0.53 points of dot-product scoring under a controlled ablation (Table 3), and it provides a transparent energy model with conservative assumptions. The formal propositions that are actually proved are correct, and the ablation protocol (fixed backbone, teacher, and schedule) is a genuine strength. The significance is conditional, however, because the headline claims about the deployed SNN—the ImageNet accuracies in Table 1 and the energy estimates in Section 5.3—rest on an exact-conversion claim that is not established.","major_comments":[{"comment":"The statement that 'the QNN and the converted SNN produce identical Lapis affinities and attention weights' is supported only for the elementwise encoder E(z). The spiking FC layers that compute queries, keys, values, and the output projection are not specified, so the paper does not show how a first-spiking LIF neuron's output latency equals T - clamp(Wx+b). In a first-spike network, output firing time is governed by threshold, reset, and integration dynamics, and is generally not a linear function of input latencies. A concrete neuron parameterization plus a proof or empirical verification is needed before Table 1 can be read as the accuracy of the deployed TTFS SNN; this is load-bearing for both the accuracy and the energy claims.","section":"Section 4.5, Eq. (20)"},{"comment":"The energy estimates use spike activities rho_l measured with forward hooks on the quantized network ('each positive integer activation produces exactly one spike'), not from a simulation of the converted spiking network. If the spiking FC layers introduce timing or threshold errors, both the reported SNN accuracy and the operation counts N_AC in Eq. (22) would not correspond to the deployed system. The paper should either prove the conversion for every layer or run an event-driven simulation of the converted SNN and measure rho_l from that simulation.","section":"Section 5.3 / Appendix A.2, Eqs. (21)-(22)"}],"minor_comments":[{"comment":"Because beta_h is defined as exp(-1/tau_h), the 'leakage realization' is an arithmetic equivalence rather than a derivation from an independently specified leaky membrane; the authors should rephrase the claim in Section 1 that 'the exponential mapping is induced by the leakage process itself' to avoid overstating what is proved.","section":"Section 4.3, Proposition 1"},{"comment":"The quantization and conversion pipeline does not specify how activations are clamped to the range {0,...,T}, how biases are encoded, or how signed quantities are handled; please add these details, since they affect the exactness claim.","section":"Section 4.5 / Appendix C"},{"comment":"The SpikeZIP-TF energy derivation ('19.85W using 64 steps of 1ms') yields 1,270.4 mJ; make the units explicit in the derivation and ensure the table entry is unambiguous.","section":"Table 1 / Appendix A.2"},{"comment":"The 0.53-point gap between the Laplacian relation and softmax dot-product scoring is based on a single run; reporting multiple seeds for this central ablation would strengthen the claim that the difference is meaningful.","section":"Table 3"},{"comment":"The ImageNet models are initialized from BEiT checkpoints and fine-tuned, but no code, checkpoints, or random seeds are provided; adding at least seeds and a reproducibility statement would strengthen the ablation claims.","section":"Section 5.1"},{"comment":"The small-font annotations in the 'running example' and 'equivalent discrete leakage view' insets are difficult to read; please enlarge and define D_ij and beta explicitly in the captions.","section":"Figures 1 and 2"}],"recommendation":"major_revision","confidential_remarks":"The central concern is the unsupported exact-conversion bridge between the quantized network and the deployed TTFS SNN. If the authors can supply a concrete neuron-level conversion theorem or an event-driven SNN simulation, the paper would be publishable; the energy model is transparent and the ablation design is a real strength."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the quick version. Lapis is not a brand-new field, but it is a genuinely useful method paper: it defines query-key similarity directly from TTFS first-spike latencies, uses an L1 distance with a Laplacian kernel and power-of-two row normalization, and shows under a controlled setting that this costs only 0.53 points versus dot-product on CIFAR-10. That controlled ablation (Table 3) is the paper's best evidence. Same backbone, same teacher, same schedule; the Laplacian beats Gaussian by 5.12 points and Hamming by 8.64. The math in Props 1-2 and Lemma 1 is correct as far as it goes. The paper also cites EcoTransformer and LaplacianFormer honestly, so the novelty claim is about the TTFS-spiking combination, not about inventing Laplacian attention.\n\nThe biggest soft spot is the QNN-to-SNN conversion bridge. Section 4.5 proves the elementwise encoding E(z) preserves query-key L1 distances, and then asserts that the QNN and converted SNN produce identical affinities and attention weights. But that requires the spiking FC layers for query, key, value, and output projection to exactly implement the quantized linear maps. First-spike output times of LIF neurons are not generally linear in input latencies, and the paper gives no neuron parameterization or proof for those layers. This is not a fatal flaw if the deployment is actually validated by running the converted SNN, but the text reads like the identity is a formal consequence when it is actually an unverified engineering assumption. If the conversion is approximate, the ImageNet accuracies and spike activities feeding the energy model may not match the QNN ablations.\n\nOther weaknesses are minor. The energy numbers are operation-level estimates, not measured silicon, and the paper is upfront about the accounting, but there is no code or checkpoints to check the activity statistics, and the accuracy numbers come without error bars. The leakage story is honest: the affinity is defined first and the leaky recurrence is constructed to realize it, so 'leakage realization' is an equivalence, not a derivation from a target. That is a bit of framing, but the math backs the equivalence.\n\nWho is this for? People working on spiking transformers or low-energy attention will get value, especially from the kernel ablation and the power-of-two normalization result. It deserves a serious referee, and the referee should push for either a proof or a measured comparison of the QNN versus the deployed SNN, plus code release. My own verdict would be conditional accept on that bridge.","headline":"A solid, well-ablated spiking attention method; the exactness of the full TTFS conversion is the one load-bearing claim I'd want verified before believing the deployment numbers.","tokens_in":15455,"tokens_out":2042,"would_cite":true,"duration_ms":20390,"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":"A spiking attention mechanism scoring token pairs by the L1 distance between first-spike latencies, mapped through a Laplacian kernel, reaches 96.56% on CIFAR-10 and cuts estimated ImageNet attention energy by 14.5x.","keywords":["spiking neural networks","time-to-first-spike coding","self-attention","vision transformer","Laplacian kernel","membrane leakage","ANN-to-SNN conversion","energy-efficient inference"],"falsifier":"Measure, in the deployed TTFS spiking network, the actual first-spike latency at the output of each spiking fully-connected layer on a batch of images and compare it with the latency $T-z$ that the training-time conversion assumes; any systematic timing shift or threshold misfire that changes the query-key $\\ell^1$ distances beyond the stated stability bound would falsify the claim that the trained accuracy and energy statistics transfer to the spiking network.","tokens_in":14412,"feed_emoji":"⚡","tokens_out":12228,"duration_ms":93525,"temperature":0.7,"pith_summary":"The paper proposes Lapis, a spiking attention mechanism that replaces dot-product query-key scoring with a temporal-distance rule: each token pair is scored by the $\\ell^1$ distance between its query and key first-spike latency vectors under time-to-first-spike coding, and that distance is converted to an affinity by a Laplacian kernel. Because the kernel's exponential decay matches leaky integrate-and-fire membrane leakage, the affinity can be realized by decaying a membrane trace for the accumulated latency difference, and row normalization becomes a bit shift under power-of-two rounding. Under a matched backbone and training schedule, the paper reports 96.56% top-1 accuracy on CIFAR-10, within 0.53 points of dot-product scoring, and on ImageNet-1K an estimated 14.5x reduction in attention-path arithmetic energy relative to dense attention. The significance is that spike timing can define the relation space itself rather than merely encode activations for an inherited attention operator.","feed_headline":"First-spike timing scores attention with near-dot-product accuracy","feed_subtitle":"Lapis stays within 0.53 points of dot-product scoring on CIFAR-10 while cutting ImageNet attention-path energy 14.5x.","key_machinery":"The load-bearing object is the temporal distance-affinity pair: $D^h_{ij} = \\|t^{Q,h}_i - t^{K,h}_j\\|_1$, the $\\ell^1$ distance between first-spike latency vectors, and $A^h_{ij} = \\exp(-D^h_{ij}/\\tau_h)$, the Laplacian kernel. The paper shows this affinity equals $\\beta_h^{D^h_{ij}}$, the value of the leaky recurrence $u_{r+1} = \\beta_h u_r$ with $\\beta_h = \\exp(-1/\\tau_h)$ after $D^h_{ij}$ decay steps, which ties the attention score directly to the leaky integrate-and-fire membrane's leakage. The other load-bearing component is power-of-two row normalization: the row sum $Z^h_i$ is rounded to the nearest power of two so the final scaling is a bit shift, and Lemma 1 bounds every resulting attention weight within a factor of $\\sqrt{2}$ of the exact softmax-style normalized weight.","core_discovery":"The central claim is that first-spike timing can serve as the native relation signal for self-attention in a spiking vision transformer. For each head, queries and keys are represented by channel-wise first-spike latency vectors $t^{Q,h}_i$ and $t^{K,h}_j$, and the affinity between token $i$ and token $j$ is $A^h_{ij} = \\exp(-D^h_{ij}/\\tau_h)$, where $D^h_{ij} = \\|t^{Q,h}_i - t^{K,h}_j\\|_1$ is the accumulated latency difference. This replaces the dot product with subtraction, absolute value, and accumulation, and the exponential form is exactly the discrete leaky recurrence $u_{r+1} = \\beta_h u_r$ with $\\beta_h = \\exp(-1/\\tau_h)$, so running membrane decay for $D^h_{ij}$ steps yields the affinity. The paper demonstrates on CIFAR-10 that this Laplacian relation reaches 96.56% top-1 accuracy, within 0.53 points of softmax dot-product scoring under an identical backbone and schedule, and reports an estimated 14.5x reduction in attention-path arithmetic energy on ImageNet-1K, with the deployed 6-bit model at 83.25% top-1 and 3.28 mJ per image.","pith_inferences":["Beyond the paper, the integer nature of $D^h_{ij} \\in \\{0,\\dots,CT\\}$ means each head has only $CT+1$ distinct affinities; a dedicated lookup or leakage circuit could replace the exponential evaluation, a direction the paper costs but does not physically implement.","Beyond the paper, the stability bound $|\\tilde A^h_{ij} - A^h_{ij}| \\le (\\eta_Q + \\eta_K)/\\tau_h$ implies that robustness to first-spike latency noise degrades as the per-head temporal scale $\\tau_h$ grows; the paper does not explore this trade-off during training.","Beyond the paper, nothing in the formulation ties the Laplacian timing-distance rule to vision specifically; the same scoring could be applied to other TTFS-coded sequence models, though the paper does not test that setting."],"forward_implications":["Query-key scoring in Lapis requires only subtraction, absolute value, and accumulation, eliminating all multiplication between query and key channels.","Because the Laplacian affinity equals the value of the leaky recurrence $u_{r+1}=\\beta_h u_r$ after $D^h_{ij}$ decay steps, the attention score can be produced by membrane leakage itself rather than by a separate softmax.","Power-of-two row normalization preserves the relative affinities and their ordering, and bounds every attention weight within a factor of $\\sqrt{2}$ of exact softmax-style normalization (Lemma 1).","Under a matched backbone and training schedule, the Laplacian relation stays within 0.53 points of dot-product scoring on CIFAR-10 and outperforms Gaussian and Hamming alternatives.","On ImageNet-1K, the estimated attention-path arithmetic energy drops from 100.01 mJ to 6.92 mJ per image with full-precision weights, and the 6-bit Lapis-L model reaches 83.25% top-1 at 3.28 mJ."],"supporting_citations":[{"why":"Supplies the leaky integrate-and-fire membrane model whose subthreshold exponential decay the Laplacian affinity reproduces.","marker":"Gerstner et al. 2014"},{"why":"Introduces TTFS-based deep spiking networks whose first-spike latency coding Lapis builds on.","marker":"Park et al. 2020"},{"why":"Demonstrates fast energy-efficient neuromorphic inference using first-spike times, motivating TTFS as a low-latency representation.","marker":"Göltz et al. 2021"},{"why":"Establishes the Laplacian kernel as a multiplication-free attention scoring rule in dense Transformers, which Lapis adapts to TTFS latencies.","marker":"Gao et al. 2025"},{"why":"Further studies Laplacian kernels for efficient dense attention, supporting the viability of Laplacian affinity as a relation function.","marker":"Feng et al. 2026"},{"why":"Provides the pretrained BEiT backbone from which the ImageNet models are initialized and fine-tuned.","marker":"Bao et al. 2021"},{"why":"Spikformer is the primary spiking-transformer baseline whose spike-form query-key-value design Lapis replaces with timing-based scoring.","marker":"Zhou et al. 2022"},{"why":"QKFormer is the strongest compared spiking attention baseline on CIFAR and ImageNet, used for accuracy and energy comparisons.","marker":"Zhou et al. 2024"},{"why":"SpikeZIP-TF is the ANN-to-SNN conversion baseline whose accuracy Lapis matches with fewer simulation steps.","marker":"You et al. 2024"}],"fun_headline_variants":["Spiking attention uses first-spike timing, near dot-product accuracy","Latency-based spiking attention: 14.5x less energy, near dot-product","First-spike timing replaces dot-product scoring in spiking ViTs","Lapis: Laplacian spike-timing attention cuts energy 14.5x"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole result hinges on every spiking layer converting the trained network's numbers into firing times without error; the paper proves this for the distance encoding but leaves the spiking linear layers' exact behavior unspecified.","fun_headline_variants_meta":{"raw":{"variants":["Spiking attention uses first-spike timing, near dot-product accuracy","Latency-based spiking attention: 14.5x less energy, near dot-product","First-spike timing replaces dot-product scoring in spiking ViTs","Lapis: Laplacian spike-timing attention cuts energy 14.5x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000626,"raw_usage":{"total_tokens":2975,"prompt_tokens":1101,"completion_tokens":1874,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":717,"completion_tokens_details":{"reasoning_tokens":1790}},"tokens_in":717,"tokens_out":1874,"duration_ms":14169,"temperature":1.0,"reasoning_tokens":1790,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:24:45.714185+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure, in the deployed TTFS spiking network, the actual first-spike latency at the output of each spiking fully-connected layer on a batch of images and compare it with the latency $T-z$ that the training-time conversion assumes; any systematic timing shift or threshold misfire that changes the query-key $\\ell^1$ distances beyond the stated stability bound would falsify the claim that the trained accuracy and energy statistics transfer to the spiking network.","supporting_citations":[],"review_version":1}