Pith. sign in

REVIEW 2 major objections 6 minor 36 references

Lapis: Laplacian Spiking Attention via First-Spike Timing and Membrane Leakage

T0 review · 2 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read 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.

desk verdict 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. read the letter →

arxiv 2608.11865 v1 pith:7K7D3PJN submitted 2026-08-12 cs.NE

Kaiwen Tang , Jiaqi Zheng , Zixuan Zhu , Yiqun Wang , Zhanglu Yan , Weng-Fai Wong This is my paper · ORCID
classification cs.NE
keywords spikingneuralnetworkstime-to-first-spikecodingself-attentionvisiontransformerLaplaciankernelmembraneleakageANN-to-SNNconversionenergy-efficientinference
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 6 minor

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.

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 (2)
  1. [Section 4.5, Eq. (20)] 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.
  2. [Section 5.3 / Appendix A.2, Eqs. (21)-(22)] 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.
minor comments (6)
  1. [Section 4.3, Proposition 1] 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.
  2. [Section 4.5 / Appendix C] 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.
  3. [Table 1 / Appendix A.2] 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.
  4. [Table 3] 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.
  5. [Section 5.1] 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.
  6. [Figures 1 and 2] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the leakage realization is a designed equivalence, not a prediction derived from its inputs.

full rationale

Walking the derivation chain, I find no circular step. The paper defines the Lapis affinity explicitly in Eq. (8) as exp(-D_ij^h / tau_h), and then defines the leakage recurrence in Eqs. (9)-(10) with beta_h = exp(-1/tau_h). Proposition 1 then verifies that running the recurrence for D_ij^h steps yields the same affinity. This is an implementation equivalence, not a derivation of the Laplacian from leakage, and the paper does not present it as a prediction. Similarly, Eq. (20) is an identity constructed by the encoding rule in Eq. (19): first-spike time is T - z by definition, so distance preservation is immediate and called out as preservation, not as an empirical discovery. The power-of-two normalization lemma is a standard rounding bound and is not circular. The accuracy results on CIFAR and ImageNet are empirical measurements under a fixed backbone and matched training schedule, with no fitted parameter renamed as a prediction. The self-citations that appear (Sorbet, Otters) are confined to related-work context and are not load-bearing for the central claim. The main genuine weakness is that Section 4.5 asserts that the QNN and converted SNN produce identical affinities without proving that the spiking fully connected layers exactly realize the quantized linear maps; this is a missing proof or correctness risk, not a circular reduction, and therefore does not raise the circularity score.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim depends on three unverified premises: exact layer-level TTFS conversion, the single-spike latency representation, and the operation-level energy model. The learnable temporal scale tau_h and the hand-chosen coding window T are the main free parameters; no invented entities are introduced.

free parameters (2)
  • tau_h (per-head temporal scale)
    Learned during training; controls the width of the Laplacian kernel and directly sets the affinity decay. Its learned value is not reported.
  • T (TTFS coding window)
    Set by hand to 15 (Lapis-S/B) or 20 (Lapis-L); trades off latency range, energy, and accuracy.
assumptions (3)
  • domain assumption Exact conversion of the trained QNN to a TTFS SNN preserves the function of every layer, not just the query-key distance.
    Section 4.5 and Eq. (20) prove distance preservation for the encoding, but the spiking FC layer implementations (query, key, value, output projections) are not specified. The claim that the QNN and converted SNN produce identical attention weights rests on this unstated layer-level exactness.
  • domain assumption Each neuron fires at most one spike within the TTFS window, and a non-firing neuron is represented by the code T and treated as a latency value in the L1 distance.
    This is the TTFS single-spike paradigm assumed throughout, including the treatment of silence as a latency value, which affects the distance and affinity.
  • domain assumption The 45nm operation-level energy model with the stated per-operation costs is representative of real hardware.
    The reported energy reductions are computed from an operation-count model, not measured from silicon. The actual cost of subtraction, absolute value, accumulation, and lookup circuits could differ.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Lapis: Laplacian Spiking Attention via First-Spike Timing and Membrane Leakage." pith.science (2026). https://pith.science/paper/7K7D3PJN

@misc{pith2026260811865,
  author       = {Pith},
  title        = {Pith review of: Lapis: Laplacian Spiking Attention via First-Spike Timing and Membrane Leakage},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7K7D3PJN}},
  note         = {Machine review of arXiv:2608.11865}
}
read the original abstract

Self-attention has become central to spiking vision transformers, yet its query-key scoring is still largely inherited from dense networks. Existing spiking variants either simplify dot product scoring or replace it with discrete operators, but spike timing, the native variable of a spiking network, does not directly define how tokens are related. We propose Lapis, a spiking attention mechanism that scores each token pair by the L1 distance between its query and key first-spike latency vectors under time-to-first-spike coding, and maps this distance to an affinity through a Laplacian kernel. The kernel's exponential decay matches the impulse response of a leaky integrate-and-fire membrane, so the accumulated latency difference determines the decay of a membrane trace, while row normalization reduces to a bit shift under power-of-two rounding. Scoring therefore needs only subtraction, absolute value, and accumulation, and removes all multiplication between query and key channels. Under a matched backbone and training schedule, Lapis reaches 96.56% top-1 accuracy on CIFAR-10, within 0.53 points of dot-product scoring. On ImageNet-1K, it reduces the estimated arithmetic energy of the attention path by 14.5x relative to dense dot-product attention. The deployed 6-bit model attains 83.25% top-1 accuracy at an estimated arithmetic energy of 3.28mJ per image.

Figures

Figures reproduced from arXiv: 2608.11865 by the authors.

Figure 1
Figure 1. Rethinking query-key scoring with Lapis. ANN attention compares dense activations, while representative spiking [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the Lapis block and attention module. A Lapis block follows a residual Transformer structure with Lapis attention and a spiking feed-forward network. Within Lapis attention, the query and key branches provide first-spike latencies whose pairwise ℓ1 distances are mapped to Laplacian affinities. A power-of-two row denominator enables shift-based normalization. The value branch passes through an independent… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 17 canonical work pages

  1. [1]

    arXiv preprint arXiv:2409.02111 , year=

    Toward large-scale spiking neural networks: A comprehensive survey and future directions , author=. arXiv preprint arXiv:2409.02111 , year=

    arXiv

  2. [2]

    Neural networks , volume=

    Deep learning in spiking neural networks , author=. Neural networks , volume=. 2019 , publisher=

    2019

  3. [3]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Optimized potential initialization for low-latency spiking neural networks , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

    work page

  4. [4]

    arXiv preprint arXiv:2303.04347 , year=

    Optimal ANN-SNN conversion for high-accuracy and ultra-low-latency spiking neural networks , author=. arXiv preprint arXiv:2303.04347 , year=

    arXiv

  5. [5]

    IEEE Transactions on Neural Networks and Learning Systems , year=

    Converting high-performance and low-latency snns through explicit modeling of residual error in anns , author=. IEEE Transactions on Neural Networks and Learning Systems , year=

    work page

  6. [6]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Incorporating learnable membrane time constant to enhance learning of spiking neural networks , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  7. [7]

    arXiv preprint arXiv:2202.11946 , year=

    Temporal efficient training of spiking neural network via gradient re-weighting , author=. arXiv preprint arXiv:2202.11946 , year=

    arXiv

  8. [8]

    Neurocomputing , volume=

    Rethinking residual connection in training large-scale spiking neural networks , author=. Neurocomputing , volume=. 2025 , publisher=

    work page 2025

Show all 36 references
  1. [9]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Going deeper with directly-trained larger spiking neural networks , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  2. [10]

    2024 International Joint Conference on Neural Networks (IJCNN) , pages=

    OneSpike: Ultra-low latency spiking neural networks , author=. 2024 International Joint Conference on Neural Networks (IJCNN) , pages=. 2024 , organization=

    2024

  3. [11]

    arXiv preprint arXiv:2409.15298 , year=

    Sorbet: A neuromorphic hardware-compatible transformer-based spiking language model , author=. arXiv preprint arXiv:2409.15298 , year=

  4. [12]

    arXiv preprint arXiv:2209.15425 , year=

    Spikformer: When spiking neural network meets transformer , author=. arXiv preprint arXiv:2209.15425 , year=

    arXiv

  5. [13]

    Advances in neural information processing systems , volume=

    Spike-driven transformer , author=. Advances in neural information processing systems , volume=

  6. [14]

    Advances in Neural Information Processing Systems , volume=

    Qkformer: Hierarchical spiking transformer using qk attention , author=. Advances in Neural Information Processing Systems , volume=

  7. [15]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Spiking transformer with spatial-temporal attention , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  8. [16]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Rethinking spiking self-attention mechanism: implementing a-XNOR similarity calculation in spiking transformers , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  9. [17]

    Nature machine intelligence , volume=

    Fast and energy-efficient neuromorphic deep learning with first-spike times , author=. Nature machine intelligence , volume=. 2021 , publisher=

    2021

  10. [18]

    2020 57th ACM/IEEE design automation conference (DAC) , pages=

    T2FSNN: Deep spiking neural networks with time-to-first-spike coding , author=. 2020 57th ACM/IEEE design automation conference (DAC) , pages=. 2020 , organization=

    2020

  11. [19]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Parallel training time-to-first-spike spiking neural networks , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  12. [20]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Temporal-coded spiking neural networks with dynamic firing threshold: Learning with event-driven backpropagation , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  13. [21]

    Forty-second International Conference on Machine Learning , year=

    Ttfsformer: a ttfs-based lossless conversion of spiking transformer , author=. Forty-second International Conference on Machine Learning , year=

  14. [22]

    arXiv preprint arXiv:2509.18968 , year=

    Otters: An Energy-Efficient SpikingTransformer via Optical Time-to-First-Spike Encoding , author=. arXiv preprint arXiv:2509.18968 , year=

  15. [23]

    2014 , publisher=

    Neuronal dynamics: From single neurons to networks and models of cognition , author=. 2014 , publisher=

    2014

  16. [24]

    International conference on machine learning , pages=

    Training data-efficient image transformers & distillation through attention , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  17. [25]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Pyramid vision transformer: A versatile backbone for dense prediction without convolutions , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  18. [26]

    arXiv preprint arXiv:2106.08254 , year=

    Beit: Bert pre-training of image transformers , author=. arXiv preprint arXiv:2106.08254 , year=

    arXiv

  19. [27]

    arXiv preprint arXiv:2305.05954 , year=

    Enhancing the performance of transformer-based spiking neural networks by SNN-optimized downsampling with precise gradient backpropagation , author=. arXiv preprint arXiv:2305.05954 , year=

    arXiv

  20. [28]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Masked spiking transformer , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  21. [29]

    arXiv preprint arXiv:2406.03470 , year=

    Spikezip-tf: Conversion is all you need for transformer-based snn , author=. arXiv preprint arXiv:2406.03470 , year=

    arXiv

  22. [30]

    Advances in Neural Information Processing Systems , volume=

    Spiking neural networks need high-frequency information , author=. Advances in Neural Information Processing Systems , volume=

  23. [31]

    Transformer dissection: An unified understanding for transformer’s attention via the lens of kernel , author=. Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNL...

    2019

  24. [32]

    arXiv preprint arXiv:2009.14794 , year=

    Rethinking attention with performers , author=. arXiv preprint arXiv:2009.14794 , year=

    2009 arXiv

  25. [33]

    Advances in Neural Information Processing Systems , volume=

    Skyformer: Remodel self-attention with gaussian kernel and nystr " om method , author=. Advances in Neural Information Processing Systems , volume=

  26. [34]

    arXiv preprint arXiv:2202.08791 , year=

    cosformer: Rethinking softmax in attention , author=. arXiv preprint arXiv:2202.08791 , year=

    arXiv

  27. [35]

    arXiv preprint arXiv:2604.20368 , year=

    LaplacianFormer: Rethinking Linear Attention with Laplacian Kernel , author=. arXiv preprint arXiv:2604.20368 , year=

    arXiv

  28. [36]

    arXiv preprint arXiv:2507.20096 , year=

    EcoTransformer: Attention without multiplication , author=. arXiv preprint arXiv:2507.20096 , year=

    arXiv

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.