Pith. sign in

REVIEW 4 major objections 4 minor 75 references

Nearly lossless ANN-to-SNN conversion at the first timestep via polarity multi-spike mapping

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A signed quantized activation plus a multi-spike signed neuron makes a one-timestep spiking network reproduce the quantized ANN exactly, giving ANN-identical accuracy at T=1 on CIFAR-10/100 and state-of-the-art one-step accuracy on ImageNet and event datasets.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection The T=1 lossless conversion is genuine and by construction, but the paper's own proof of it is invalid as written; fixable, and worth refereeing. the 4 major comments →

arxiv 2508.14520 v2 pith:OJEYYXFD submitted 2025-08-20 cs.NE

Quantization Meets Spikes: Nearly Lossless Conversion at the First Timestep via Polarity Multi-Spike Mapping

Hangming Zhang , Zheng Li , Chenxiang Ma , Huajin Tang , Long Cheng , Kay Chen Tan , Qiang Yu This is my paper
classification cs.NE
keywords ANN-to-SNN conversionspiking neural networksquantized activationpolarity spikesmulti-spike neuronsultra-low latencyinformation entropyneuromorphic computing
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 claims that an ANN can be converted into a spiking neural network that reproduces the ANN's own quantized activations exactly on the very first timestep, so the SNN's accuracy equals the ANN's accuracy. The conversion works by pairing a polarity-aware quantized activation (PQA) with an augmented integrate-and-fire (AIF) neuron that emits multiple positive or negative spikes in one timestep. The paper argues the quantization step is nearly lossless when its bounds are chosen so that the entropy of the quantized output matches the entropy of a standard normal input, and it reports identical or near-identical ANN/SNN accuracies on CIFAR-10/100, ImageNet (with a small ViT attention gap), and three event-based datasets.

Core claim

The central claim is that PMSM achieves nearly lossless ANN-to-SNN conversion at the first timestep: using PQA with entropy-guided bounds, an AIF neuron, and membrane initialization v(0)=ϑ/2+ε, the SNN reproduces the quantized ANN layer by layer at T=1, so the SNN accuracy equals the ANN accuracy. Table IV shows identical numbers, e.g., CIFAR-10 VGG-16 95.67 to 95.67, CIFAR-100 VGG-16 76.71 to 76.71, ResNet-20 93.78 to 93.78. The admitted exception is the ViT softmax path, where a SpikeZIP-TF-style attention approximation leaves an ImageNet gap (82.34 to 81.61).

What carries the argument

The machinery has two parts: (1) Polarity Quantized Activation (PQA), a quantization function that keeps negative activations and rounds to levels within [α,β], with (α,β) fixed by a grid search that seeks entropy ratio R≈1 relative to a standard normal input; and (2) the Augmented Integrate-and-Fire (AIF) neuron, which emits multiple positive or negative spikes in a single timestep, bounded by C_neg=α*L and C_pos=β*L so that spike counts reconstruct PQA outputs. The membrane potential initialization v(0)=ϑ/2+ε turns rounding into an identity at T=1.

Load-bearing premise

Post-batch-normalization activations are approximately standard normal in every layer, so a single grid search over (α,β) under that model fixes bounds that apply to all layers; if real layer distributions drift or are skewed, the entropy ratio R may not be near 1 and the 'lossless' claim could fail.

What would settle it

Measure the actual entropy ratio R = H_PQA / H_BN on the trained layer activations (not the Gaussian approximation) and check whether R is close to 1; alternatively, on a probe network with deliberately shifted BN affine parameters, verify whether T=1 SNN accuracy still equals ANN accuracy or degrades.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Ultra-low-latency SNN inference (T=1) becomes feasible without accuracy loss for CNNs and GCNs, removing a main barrier to real-time neuromorphic deployment.
  • ANN-to-SNN conversion can match or beat direct training on several benchmarks while using far fewer timesteps.
  • Energy estimates in the paper show more than two orders of magnitude reduction versus ANNs and over 55% reduction versus prior conversion methods.
  • The framework extends to ViT but retains a small gap tied to softmax attention approximation, isolating a concrete target for further work.
  • Multi-timestep operation remains stable because spike-count deviations are bounded and average out as T grows.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The entropy argument only needs an entropy ratio R≈1, not a full distributional match; an obvious stress test is to check whether the same (α*,β*) selection holds on layers whose post-BN activations are skewed or have shifted variance.
  • The T=1 identity is essentially a rounding-and-offset identity on a fixed quantization grid; the claim of 'nearly lossless' ultimately depends on the grid covering almost all activations, so one could extend the method with per-layer adaptive grid placement.
  • The reported identical ANN/SNN accuracies likely follow because the conversion becomes deterministic at T=1; comparing layer-wise logits on a probe network would reveal whether the match is exact or merely close.
  • The paper notes that object detection and semantic segmentation remain unexplored; a direct extension would apply PMSM to those tasks where latency matters.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes PMSM, an ANN-to-SNN conversion framework built on two components: a polarity quantized activation (PQA, Eq. 15) that preserves both positive and negative activations, and an augmented integrate-and-fire (AIF) neuron (Eq. 28) that emits multiple positive/negative spikes per timestep. An entropy analysis over assumed standard-normal activations is used to initialize the quantization threshold and bounds (Eqs. 16-27). With membrane initialization v(0)=θ/2+ε, the paper claims that the converted SNN reproduces the quantized ANN layer-by-layer at T=1, and reports state-of-the-art accuracy at T=1 across CNNs, GCNs, and a ViT-S transformer, together with large energy savings.

Significance. If the results hold, the paper makes a strong practical contribution: exact T=1 conversion for CNN/GCN models would remove the latency-accuracy trade-off that has constrained conversion-based SNNs, and the energy reductions are notable. The information-entropy perspective on quantization is a useful framing, and the experimental coverage is broad (six datasets, four architectures). However, the central theoretical proof in Section IV.C compares the SNN output to the wrong reference operation, and the headline entropy ratio in Eq. (13) is arithmetically incorrect. These are local but load-bearing issues; they do not necessarily invalidate the empirical construction, but they must be corrected before the paper's theoretical claims can be accepted.

major comments (4)
  1. [Section IV.C, Eqs. (29)-(30)] The proof of zero conversion error at T=1 uses the wrong reference. PQA (Eq. 15) targets round-to-nearest, ⌊xL/ϑ⌉, while Eq. (29) defines the error relative to floor(z/ϑ). With v(0)=ϑ/2+ε, the AIF output is clip(floor(z/ϑ+1/2+ε/ϑ),...), which is a round-to-nearest operation, not floor(z/ϑ). The cancellation in Eq. (30) is therefore false for the stated baseline: e.g., z/ϑ=0.7 gives SNN count 1 but floor(z/ϑ)=0. A corrected proof should define the target as round(z/ϑ); then the cancellation is straightforward. As written, the central 'negligible conversion error' theorem is unsupported.
  2. [Section III.B, Eq. (13)] The numerical claim that H_ReLU ≈ 0.69 H_BN is wrong. Computing from the paper's own expression: H_ReLU = 0.5 log 2 + 0.25 log(2π) + 0.25 ≈ 1.0561, while H_BN = 0.5 log(2πe) ≈ 1.4189, giving a ratio ≈ 0.744, not 0.69. Thus Eq. (14) and the associated statement that conventional QA retains about 69% of BN entropy are quantitatively incorrect. The qualitative conclusion (entropy loss) survives, but the number must be corrected.
  3. [Section IV.A, Eqs. (16)-(18), (27)] The entropy-guided initialization assumes x ∼ N(0,1) for every layer and ignores the affine parameters in BN, which shift and scale activations away from the standard normal. The selected (α*,β*) are then frozen during training, but the paper provides no distributional measurements, no ablation against simpler heuristics, and no reported values of (α*,β*) or grid-search resolution. This leaves the causal role of the entropy criterion unvalidated: the accuracy improvements could come from the signed quantized activation and the AIF neuron alone. The authors should provide per-layer entropy-ratio measurements and at least one ablation (e.g., fixed symmetric bounds vs. entropy-selected bounds) to support the mechanism claimed.
  4. [Section IV.C, Eqs. (32)-(34)] The multi-timestep error derivation also appears to drop the half-integer offset. Eq. (33) contains a floor of z/ϑ+1/2+o(ε), but Eq. (34) simplifies the difference against floor(z/ϑ) as if the 1/2 term vanishes. For fractional parts of z/ϑ near 1/2, this introduces an O(1) offset that is not captured by the average Δ_i term. The T>1 analysis should be restated with the correct round/floor baseline and the half-integer handling made explicit.
minor comments (4)
  1. [Throughout] The notation is easy to confuse: ⌊·⌋ (floor) and ⌊·⌉ (round) are both used. Eqs. (29)-(30) use floor while Eq. (15) uses round; please use distinct, consistently defined symbols and restate the target operation.
  2. [Section V.A] No concrete values of (α*,β*) or the grid-search configuration are given, and the resolution of the R≈1 search is not described. Adding a short appendix with the selected bounds and the resulting R values would make the method reproducible.
  3. [Section V.B, Table V] For ViT-S, the ANN accuracy of the proposed method is lower than SpikeZIP-TF's ANN accuracy on CIFAR-10/100; a brief explanation of whether this is due to PQA fine-tuning or training budget would help the comparison.
  4. [Section IV.A] The sentence 'Since x ∼ N(0,1), k approximately follows a normal distribution' conflates the discrete index k with a continuous approximation; the subsequent CDF expression is understandable but should be phrased as an approximation with its error quantified or at least acknowledged.

Circularity Check

2 steps flagged

T=1 losslessness is definitional: the AIF neuron is PQA by construction, and the stated proof is algebraically wrong; the 'entropy-lossless' condition is a fitted constraint, not a measured result.

specific steps
  1. self definitional [Section IV.B Eq.(28); Section IV.A Eq.(15); Section IV.C Eq.(30)]
    ""The key idea is to directly align spike generation with the discrete quantization behavior of PQA, thereby ensuring that spike counts faithfully reproduce the quantized activations at the first timestep" (Sec. IV.B). AIF: s_l=clip(⌊m_l/ϑ^l_SNN⌋, Cneg, Cpos), Cneg=α*L, Cpos=β*L. PQA: y_l=ϑ_l clip(1/L ⌊x_l L/ϑ_l⌉, α, β)."

    With ϑ_SNN=ϑ/L (Eq. 5) and v(0)=ϑ_SNN/2+ε (Eq. 30), the AIF output at T=1 is s_l=clip(⌊x_l L/ϑ+1/2⌋, αL, βL) = clip(round(x_l L/ϑ), αL, βL), which is L times PQA's normalized output. Thus 'nearly lossless at the first timestep' is encoded in the neuron definition, not derived. Moreover, Eq. (30) is invalid as written: it subtracts ⌊z_l/ϑ⌋ while the SNN computes ⌊z_l/ϑ+1/2⌋; the cancellation is false whenever frac(z_l/ϑ) ≥ 1/2 (e.g., z_l=0.7ϑ). The claimed proof does not establish the result; the equality holds only by definitional construction.

  2. fitted input called prediction [Section IV.A Eq.(27) and Fig. 1(B)]
    ""Find (α∗, β∗) s.t. R(α, β; ϑ0, L)≈1, ϑ0=L." ... "there exist regions where the entropy ratio R=H_PQA/H_BN≈1, indicating lossless information preservation.""

    The R≈1 condition is imposed by grid-searching (α,β) under an assumed N(0,1) activation model (Eqs. 7–14), then the selected bounds are fixed and reused as AIF spike limits. Calling this 'lossless information preservation' restates a fitted constraint as a discovered property; no measurement of real post-BN distributions, no ablation, and no reported (α*,β*) values are given to show the condition holds outside the assumed Gaussian model. This is a fitted input, not an independent prediction.

full rationale

Central claim 'nearly lossless ANN-to-SNN conversion at the first timestep' is true only relative to the PQA-trained ANN, and it is true by construction: the AIF neuron is defined as clip(round(xL/ϑ)) scaled, which is exactly the PQA function (Eq. 15 vs Eq. 28). The headline is therefore a definitional identity. The paper's own error analysis in Eq. (30) is additionally algebraically incorrect because it cancels a floor against a round, so the proof is not independent support for the claim. The entropy-guided initialization does not cure this: R≈1 is fit by parameter search under a Gaussian assumption, so 'lossless information preservation' is a constraint rather than a measurement on real activations. On the positive side, there is no load-bearing self-citation chain: prior work by the authors on augmented/polarity spikes is cited as background, not as a uniqueness theorem, and the empirical comparisons against external baselines are independent. The admitted ViT softmax gap further shows the method is not universally lossless. Overall, the conversion-losslessness claim reduces to the definition of the neuron, giving partial circularity; the entropy analysis and experiments retain independent content, so the score is 6 rather than 8.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 2 invented entities

The framework pulls three things from prior work: the half-threshold exact-conversion trick (QCFS), multi-spike and polarity firing (the authors' own prior work), and signed quantization (symmetric-threshold ReLU, signed neurons). What the paper adds is a signed 8-level activation trained to near-ANN accuracy and an entropy-based rule for choosing its bounds. The load-bearing free choices are (α*, β*), fit by grid search under a Gaussian assumption with values never reported, and the initialization ϑ0=L. The theoretical claims assume standard-normal activations and mix differential with discrete entropy, so the 'lossless' language is stronger than the derivation supports.

free parameters (4)
  • (α*, β*) quantization bounds, reused as AIF spike-count limits = not reported in text
    Selected by grid search over α in [-1,0], β in (0,1] to satisfy entropy ratio R≈1 under an assumed N(0,1) activation distribution (Eq. 27, Fig. 1B). Frozen during training; reused as C_neg=α*L, C_pos=β*L in the AIF neuron (Eq. 28). The paper never states the selected values.
  • ϑ0 = L initialization of the learnable quantization threshold = ϑ0 = 8 with L = 8
    Chosen from the intermediate-regime condition ϑ=cL+ε with c=1 'for analytical simplicity' (Section IV.A, Eq. 26). ϑ is learnable afterward, so this is an initialization, but the entropy analysis only holds in the chosen regime.
  • L (number of quantization levels) = 8
    Fixed following QCFS 'for fair comparison'; the paper's own analysis motivates L approaching infinity (Eq. 11), so the finite L=8 choice is taken from prior work, not from the theory.
  • membrane potential init v(0) = ϑ_SNN/2 + ε = 0.5·ϑ_SNN plus infinitesimal tie-break
    Inherited from QCFS; required for the T=1 exactness of round-to-nearest reproduction (Eq. 30). It is a declared design choice rather than a fitted number, but it is load-bearing for the central claim.
axioms (6)
  • domain assumption Post-BN activations are approximately standard normal, x~N(0,1), in every layer
    Used to compute H_BN (Eqs. 7-9), the ReLU upper bound (Eqs. 12-14), the PQA bin probabilities (Eqs. 16-18), and the R≈1 grid search (Eq. 27). BN's learnable scale/shift (γ, β) make per-layer distributions non-standard; no measurement is provided.
  • domain assumption The differential entropy of a standard normal is the upper bound of information capacity of BN outputs
    Section III.B, Eqs. 7-9. Mixes differential entropy (continuous x) with discrete entropy of quantized outputs (Eqs. 10-14) without a bin-width term, so the 0.69/0.744 ratio is not a valid capacity comparison.
  • standard math Round-to-nearest bin probabilities can be computed from the normal CDF via Φ((k±1/2)·ϑ/L)
    Eqs. 16-18. Standard approximation, valid only under the Gaussian assumption; note Eqs. 16-17 use an inconsistent binning convention versus Eq. 18.
  • standard math Soft-reset IF dynamics with constant input yield bounded membrane potential and rate convergence
    Section IV.C, Eqs. 31-34. Standard result used for the T>1 error bound; the Δi in {-1,0,1} argument gives a bound, but the claimed convergence of the mean to zero is asserted, not proven, and is false against the paper's own error reference.
  • domain assumption At T=1 every neuron holds v(0)=ϑ_SNN/2+ε and the input pathway is scaled exactly like the ANN pre-activation (BN folding, residual adds, first-layer scaling)
    Required for the exact-match results of Table IV; the paper does not specify BN folding details, first-layer input scaling, or residual handling, and ships no code to verify them.
  • domain assumption The SpikeZIP-TF attention scheme approximates softmax well enough at T=1
    Admitted in Section V.B as the reason for the ViT gap (ImageNet 82.34 to 81.61); loaded from cited work, not re-derived.
invented entities (2)
  • AIF neuron (augmented integrate-and-fire with polarity multi-spike firing, Eq. 28) independent evidence
    purpose: Emits multiple positive and negative spikes in a single timestep, with counts clipped to [α*L, β*L], so that spike counts equal PQA quantization indices at T=1
    A new neuron model, but a direct combination of the authors' prior augmented-spike and double-threshold polarity designs. Its empirical handle is the T=1 exact-match accuracy, which is traceable to its arithmetic.
  • PQA (polarity quantized activation, Eq. 15) independent evidence
    purpose: Quantizes both positive and negative activations into L levels within detected bounds [α, β]
    New activation function in this paper; related to symmetric-threshold ReLU and signed neurons. Its effectiveness is measured by the ANN accuracy in Tables IV-VII, but no ablation isolates its contribution from the AIF neuron.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantization Meets Spikes: Nearly Lossless Conversion at the First Timestep via Polarity Multi-Spike Mapping." pith.science (2026). https://pith.science/paper/OJEYYXFD

@misc{pith2026250814520,
  author       = {Pith},
  title        = {Pith review of: Quantization Meets Spikes: Nearly Lossless Conversion at the First Timestep via Polarity Multi-Spike Mapping},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OJEYYXFD}},
  note         = {Machine review of arXiv:2508.14520}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Spiking neural networks (SNNs) offer advantages in computational efficiency via event-driven computing, compared to traditional artificial neural networks (ANNs). While direct training methods tackle the challenge of non-differentiable activation mechanisms in SNNs, they often suffer from high computational and energy costs during training. As a result, ANN-to-SNN conversion approach remains a valuable and practical alternative. These conversion-based methods aim to leverage the discrete output produced by the quantization layer to obtain SNNs with low latency. Although the theoretical minimum latency is one timestep, existing conversion methods have struggled to realize such ultra-low latency without accuracy loss. Moreover, current quantization approaches often discard negative-value information following batch normalization and are highly sensitive to the hyperparameter configuration, leading to degraded performance. In this work, we, for the first time, analyze the information loss introduced by quantization layers through the lens of information entropy. Building on our analysis, we introduce polarity multi-spike mapping (PMSM) framework and a hyperparameter initialization strategy tailored for the quantization layer. Our method achieves nearly lossless ANN-to-SNN conversion at the extremity, i.e., the first timestep, while also leveraging the temporal dynamics of SNNs across multiple timesteps to maintain stable performance on complex tasks. Extensive experiments on six image and neuromorphic datasets consistently demonstrate that PMSM achieves nearly lossless accuracy at the first timestep. Remarkably, despite operating under ultra-low-latency constraints, PMSM surpasses state-of-the-art direct training methods on multiple benchmarks.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

75 extracted references · 63 canonical work pages · 2 internal anchors

  1. [1]

    An introduction to probabilistic spiking neural networks: Probabilistic models, learning rules, and applications,

    H. Jang, O. Simeone, B. Gardner, and A. Gruning, “An introduction to probabilistic spiking neural networks: Probabilistic models, learning rules, and applications,”IEEE Signal Processing Magazine, vol. 36, no. 6, pp. 64–77, 2019

    work page 2019

  2. [2]

    A low-cost and high-speed hardware implementation of spiking neural network,

    G. Zhang, B. Li, J. Wu, R. Wang, Y . Lan, L. Sun, S. Lei, H. Li, and Y . Chen, “A low-cost and high-speed hardware implementation of spiking neural network,”Neurocomputing, vol. 382, pp. 106–115, 2020

    work page 2020

  3. [3]

    A hybrid neural coding approach for pattern recognition with spiking neural networks,

    X. Chen, Q. Yang, J. Wu, H. Li, and K. C. Tan, “A hybrid neural coding approach for pattern recognition with spiking neural networks,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 5, pp. 3064–3078, 2024

    work page 2024

  4. [4]

    Spikeprop: backprop- agation for networks of spiking neurons

    S. M. Bohte, J. N. Kok, and J. A. La Poutr ´e, “Spikeprop: backprop- agation for networks of spiking neurons.” inProc. Eur. Symp. Artif. Neural Netw. Comput. Intell. Mach. Learn., vol. 48. Bruges, 2000, pp. 419–424

    work page 2000

  5. [5]

    Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based opti- mization to spiking neural networks,

    E. O. Neftci, H. Mostafa, and F. Zenke, “Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based opti- mization to spiking neural networks,”IEEE Signal Processing Magazine, vol. 36, no. 6, pp. 51–63, 2019

    2019

  6. [6]

    Scaling spike-driven transformer with efficient spike firing approximation training,

    M. Yao, X. Qiu, T. Hu, J. Hu, Y . Chou, K. Tian, J. Liao, L. Leng, B. Xu, and G. Li, “Scaling spike-driven transformer with efficient spike firing approximation training,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 47, no. 4, pp. 2973–2990, 2025

    work page 2025

  7. [7]

    Graph spiking attention network: Sparsity, efficiency and robustness,

    B. Wang, B. Jiang, J. Tang, L. Bai, and B. Luo, “Graph spiking attention network: Sparsity, efficiency and robustness,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 47, no. 11, pp. 10 862– 10 869, 2025

    work page 2025

  8. [8]

    Fast- classifying, high-accuracy spiking deep networks through weight and threshold balancing,

    P. U. Diehl, D. Neil, J. Binas, M. Cook, S.-C. Liu, and M. Pfeiffer, “Fast- classifying, high-accuracy spiking deep networks through weight and threshold balancing,” in2015 International joint conference on neural networks. IEEE, 2015, pp. 1–8

    work page 2015

  9. [9]

    Spiking deep convolutional neural networks for energy-efficient object recognition,

    Y . Cao, Y . Chen, and D. Khosla, “Spiking deep convolutional neural networks for energy-efficient object recognition,”International Journal of Computer Vision, vol. 113, pp. 54–66, 2015

    work page 2015

  10. [10]

    Fast-snn: Fast spiking neural network by converting quantized ann,

    Y . Hu, Q. Zheng, X. Jiang, and G. Pan, “Fast-snn: Fast spiking neural network by converting quantized ann,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 12, pp. 14 546–14 562, 2023

    work page 2023

  11. [11]

    Spikeconverter: An efficient conversion framework zipping the gap between artificial neural networks and spiking neural networks,

    F. Liu, W. Zhao, Y . Chen, Z. Wang, and L. Jiang, “Spikeconverter: An efficient conversion framework zipping the gap between artificial neural networks and spiking neural networks,”Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, no. 2, pp. 1692–1701, 2022

    work page 2022

  12. [12]

    An efficient brain- inspired accelerator using a high-accuracy conversion algorithm for spiking deformable CNN,

    H. Zhang, S. Zhang, W. Mao, and Z. Wang, “An efficient brain- inspired accelerator using a high-accuracy conversion algorithm for spiking deformable CNN,”IEEE Transactions on Circuits and Systems II: Express Briefs, vol. 72, no. 1, pp. 288–292, 2025

    work page 2025

  13. [13]

    Toward energy-efficient spike-based deep reinforcement learning with temporal coding,

    M. Zhang, S. Wang, J. Wu, W. Wei, D. Zhang, Z. Zhou, S. Wang, F. Zhang, and Y . Yang, “Toward energy-efficient spike-based deep reinforcement learning with temporal coding,”IEEE Computational Intelligence Magazine, vol. 20, no. 2, pp. 45–57, 2025

    work page 2025

  14. [14]

    Spikezip-tf: Conversion is all you need for transformer-based SNN,

    K. You, Z. Xu, C. Nie, Z. Deng, Q. Guo, X. Wang, and Z. He, “Spikezip-tf: Conversion is all you need for transformer-based SNN,” in Proceedings of the 41st International Conference on Machine Learning, vol. 235, Jul. 2024, p. 57367–57383

    work page 2024

  15. [15]

    Optimal ANN- SNN conversion for high-accuracy and ultra-low-latency spiking neural networks,

    T. Bu, W. Fang, J. Ding, P. Dai, Z. Yu, and T. Huang, “Optimal ANN- SNN conversion for high-accuracy and ultra-low-latency spiking neural networks,” inInternational Conference on Learning Representations, 2022

    work page 2022

  16. [16]

    Cs- qcfs: Bridging the performance gap in ultra-low latency spiking neural networks,

    H. Yang, S. Yang, L. Zhang, H. Dou, F. Shen, and J. Zhao, “Cs- qcfs: Bridging the performance gap in ultra-low latency spiking neural networks,”Neural Networks, vol. 184, p. 107076, 2025

    work page 2025

  17. [17]

    Quantization framework for fast spiking neural networks,

    C. Li, L. Ma, and S. Furber, “Quantization framework for fast spiking neural networks,”Frontiers in Neuroscience, vol. 16, p. 918793, 2022

    work page 2022

  18. [18]

    Optimal conversion of conventional artificial neural networks to spiking neural networks,

    S. Deng and S. Gu, “Optimal conversion of conventional artificial neural networks to spiking neural networks,” inInternational Conference on Learning Representations, 2021

    work page 2021

  19. [19]

    A free lunch from ANN: Towards efficient, accurate spiking neural networks calibration,

    Y . Li, S. Deng, X. Dong, R. Gong, and S. Gu, “A free lunch from ANN: Towards efficient, accurate spiking neural networks calibration,” in Proceedings of the 38th International Conference on Machine Learning, vol. 139, 2021, pp. 6316–6325

    work page 2021

  20. [20]

    Converting Artificial Neural Networks to Spiking Neural Networks via Parameter Calibration

    Y . Li, S. Deng, X. Dong, and S. Gu, “Converting artificial neural networks to spiking neural networks via parameter calibration,”arXiv preprint arXiv:2205.10121, May 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  21. [21]

    Efficient learning with augmented spikes: A case study with image classification,

    S. Song, C. Ma, W. Sun, J. Xu, J. Dang, and Q. Yu, “Efficient learning with augmented spikes: A case study with image classification,”Neural Networks, vol. 142, p. 205–212, Oct. 2021

    work page 2021

  22. [22]

    Efficient converted spiking neural network for 3d and 2d classification,

    Y . Lan, Y . Zhang, X. Ma, Y . Qu, and Y . Fu, “Efficient converted spiking neural network for 3d and 2d classification,” in2023 IEEE/CVF International Conference on Computer Vision, Oct. 2023, p. 9177–9186. 16

    work page 2023

  23. [23]

    Lm-ht SNN: Enhancing the performance of SNN to ANN counterpart through learnable multi- hierarchical threshold model,

    Z. Hao, X. Shi, Y . Liu, Z. Yu, and T. Huang, “Lm-ht SNN: Enhancing the performance of SNN to ANN counterpart through learnable multi- hierarchical threshold model,” inThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    work page 2024

  24. [24]

    Synaptic learning with augmented spikes,

    Q. Yu, S. Song, C. Ma, L. Pan, and K. C. Tan, “Synaptic learning with augmented spikes,”IEEE Transactions on Neural Networks and Learning Systems, vol. 33, no. 3, pp. 1134–1146, 2022

    work page 2022

  25. [25]

    Construct- ing accurate and efficient deep spiking neural networks with double- threshold and augmented schemes,

    Q. Yu, C. Ma, S. Song, G. Zhang, J. Dang, and K. C. Tan, “Construct- ing accurate and efficient deep spiking neural networks with double- threshold and augmented schemes,”IEEE Transactions on Neural Net- works and Learning Systems, vol. 33, no. 4, p. 1714–1726, Apr. 2022

    work page 2022

  26. [26]

    Augmapping: Accurate and efficient inference with deep double-threshold spiking neural networks,

    C. Ma and Q. Yu, “Augmapping: Accurate and efficient inference with deep double-threshold spiking neural networks,” in2020 IEEE Symposium Series on Computational Intelligence, 2020, pp. 2002–2007

    work page 2020

  27. [27]

    Con- version of continuous-valued deep networks to efficient event-driven networks for image classification,

    B. Rueckauer, I.-A. Lungu, Y . Hu, M. Pfeiffer, and S.-C. Liu, “Con- version of continuous-valued deep networks to efficient event-driven networks for image classification,”Frontiers in Neuroscience, vol. 11, 2017

    work page 2017

  28. [28]

    Optimized potential initialization for low-latency spiking neural networks,

    T. Bu, J. Ding, Z. Yu, and T. Huang, “Optimized potential initialization for low-latency spiking neural networks,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 36, Jun. 2022, p. 11–20

    work page 2022

  29. [29]

    Reducing ANN-SNN conversion error through residual membrane potential,

    Z. Hao, T. Bu, J. Ding, T. Huang, and Z. Yu, “Reducing ANN-SNN conversion error through residual membrane potential,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 37, Jun. 2023, p. 11–21

    work page 2023

  30. [30]

    Bridging the gap between ANNs and SNNs by calibrating offset spikes,

    Z. Hao, J. Ding, T. Bu, T. Huang, and Z. Yu, “Bridging the gap between ANNs and SNNs by calibrating offset spikes,” inThe Eleventh International Conference on Learning Representations, Feb. 2023

    work page 2023

  31. [31]

    Optimal ANN-SNN conversion for fast and accurate inference in deep spiking neural networks,

    J. Ding, Z. Yu, Y . Tian, and T. Huang, “Optimal ANN-SNN conversion for fast and accurate inference in deep spiking neural networks,” in Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, Aug. 2021, p. 2328–2336

    work page 2021

  32. [32]

    A new ANN- SNN conversion method with high accuracy, low latency and good robustness,

    B. Wang, J. Cao, J. Chen, S. Feng, and Y . Wang, “A new ANN- SNN conversion method with high accuracy, low latency and good robustness,” inProceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, Aug. 2023, p. 3067–3075

    work page 2023

  33. [33]

    Symmetric-threshold ReLU for fast and nearly lossless ANN-SNN conversion,

    J. Han, Z. Wang, J. Shen, and H. Tang, “Symmetric-threshold ReLU for fast and nearly lossless ANN-SNN conversion,”Machine Intelligence Research, vol. 20, no. 3, p. 435–446, Jun. 2023

    work page 2023

  34. [34]

    A unified optimization framework of ANN-SNN conversion: Towards optimal mapping from activation values to firing rates,

    H. Jiang, S. Anumasa, G. D. Masi, H. Xiong, and B. Gu, “A unified optimization framework of ANN-SNN conversion: Towards optimal mapping from activation values to firing rates,” inProceedings of the 40th International Conference on Machine Learning, vol. 202, Jul. 2023, pp. 14 945–14 974

    work page 2023

  35. [35]

    Masked spiking transformer,

    Z. Wang, Y . Fang, J. Cao, Q. Zhang, Z. Wang, and R. Xu, “Masked spiking transformer,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 1761–1771

    work page 2023

  36. [36]

    Spikedattention: Training-free and fully spike-driven transformer-to-SNN conversion with winner-oriented spike shift for softmax operation,

    S. Hwang, S. Lee, D. Park, D. Lee, and J. Kung, “Spikedattention: Training-free and fully spike-driven transformer-to-SNN conversion with winner-oriented spike shift for softmax operation,” inThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    work page 2024

  37. [37]

    Rmp-SNN: Residual membrane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network,

    B. Han, G. Srinivasan, and K. Roy, “Rmp-SNN: Residual membrane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network,” in2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Jun. 2020, p. 13555–13564

    work page 2020

  38. [38]

    Efficient ANN-SNN conversion with error compensation learning,

    C. Liu, J. Shen, X. Ran, M. Xu, Q. Xu, Y . Xu, and G. Pan, “Efficient ANN-SNN conversion with error compensation learning,” inForty- second International Conference on Machine Learning, 2025

    work page 2025

  39. [39]

    Signed neuron with memory: Towards simple, accurate and high-efficient ANN-SNN conversion,

    Y . Wang, M. Zhang, Y . Chen, and H. Qu, “Signed neuron with memory: Towards simple, accurate and high-efficient ANN-SNN conversion,” inProceedings of the Thirty-First International Joint Conference on Artificial Intelligence, Jul. 2022, p. 2501–2508

    work page 2022

  40. [40]

    A universal ANN- to-SNN framework for achieving high accuracy and low latency deep spiking neural networks,

    Y . Wang, H. Liu, M. Zhang, X. Luo, and H. Qu, “A universal ANN- to-SNN framework for achieving high accuracy and low latency deep spiking neural networks,”Neural Networks, vol. 174, p. 106244, Jun. 2024

    work page 2024

  41. [41]

    Ternary spike: Learning ternary spikes for spiking neural networks,

    Y . Guo, Y . Chen, X. Liu, W. Peng, Y . Zhang, X. Huang, and Z. Ma, “Ternary spike: Learning ternary spikes for spiking neural networks,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, Mar. 2024, p. 12244–12252

    work page 2024

  42. [42]

    Rethinking spikes in spiking neural networks for performance enhancement,

    F. Zhou, M. Fu, Y . Gao, B. Wang, and Q. Yu, “Rethinking spikes in spiking neural networks for performance enhancement,” in2024 IEEE International Conference on Cybernetics and Intelligent Systems and IEEE International Conference on Robotics, Automation and Mecha- tronics, 2024, pp. 374–379

    work page 2024

  43. [43]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, G. Hintonet al., “Learning multiple layers of features from tiny images,” University of Toronto, Tech. Rep., 2009

    2009

  44. [44]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in2009 IEEE Conference on Computer Vision and Pattern Recognition, 2009, pp. 248–255

    2009

  45. [45]

    CIFAR10-DVS: An event-stream dataset for object classification,

    H. Li, H. Liu, X. Ji, G. Li, and L. Shi, “CIFAR10-DVS: An event-stream dataset for object classification,”Frontiers in Neuroscience, vol. 11, 2017

    work page 2017

  46. [46]

    Caltech 101,

    F.-F. Li, M. Andreeto, M. Ranzato, and P. Perona, “Caltech 101,” CaltechDATA: Pasadena, CA, USA, vol. 5, 2022

    work page 2022

  47. [47]

    Tactilesgnet: A spiking graph neural network for event-based tactile object recognition,

    F. Gu, W. Sng, T. Taunyazov, and H. Soh, “Tactilesgnet: A spiking graph neural network for event-based tactile object recognition,” in2020 IEEE/RSJ International Conference on Intelligent Robots and Systems. IEEE, 2020, pp. 9876–9882

    work page 2020

  48. [48]

    Rmp-loss: Regularizing membrane potential distribution for spiking neural networks,

    Y . Guo, X. Liu, Y . Chen, L. Zhang, W. Peng, Y . Zhang, X. Huang, and Z. Ma, “Rmp-loss: Regularizing membrane potential distribution for spiking neural networks,” in2023 IEEE/CVF International Conference on Computer Vision, Oct. 2023, p. 17345–17355

    work page 2023

  49. [49]

    Low latency conversion of artificial neural network models to rate-encoded spiking neural net- works,

    Z. Yan, K. Tang, J. Zhou, and W.-F. Wong, “Low latency conversion of artificial neural network models to rate-encoded spiking neural net- works,”IEEE Transactions on Neural Networks and Learning Systems, vol. 36, no. 8, pp. 14 107–14 118, 2025

    work page 2025

  50. [50]

    Training spiking neural networks with local tandem learning,

    Q. Yang, J. Wu, M. Zhang, Y . Chua, X. Wang, and H. Li, “Training spiking neural networks with local tandem learning,” inAdvances in Neural Information Processing Systems, vol. 35, 2022, p. 12662–12676

    work page 2022

  51. [51]

    Converting High-Performance and Low-Latency SNNs through Explicit Modelling of Residual Error in ANNs

    Z. Huang, J. Ding, Z. Pan, H. Li, Y . Fang, Z. Yu, and J. K. Liu, “Converting high-performance and low-latency SNNs through explicit modelling of residual error in ANNs,”arXiv preprint arXiv:2404.17456, Apr. 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  52. [52]

    Very deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,”arXiv preprint arXiv:1409.1556, 2014

    Pith/arXiv arXiv 2014

  53. [53]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

    2016

  54. [54]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gellyet al., “An image is worth 16x16 words: Transformers for image recognition at scale,”arXiv preprint arXiv:2010.11929, 2020

    Pith/arXiv arXiv 2010

  55. [55]

    Spikformer: When spiking neural network meets transformer,

    Z. Zhou, Y . Zhu, C. He, Y . Wang, S. Y AN, Y . Tian, and L. Yuan, “Spikformer: When spiking neural network meets transformer,” inThe Eleventh International Conference on Learning Representations, 2023

    work page 2023

  56. [56]

    Spike- driven transformer,

    M. Yao, J. Hu, Z. Zhou, L. Yuan, Y . Tian, B. Xu, and G. Li, “Spike- driven transformer,”Advances in neural information processing systems, vol. 36, pp. 64 043–64 058, 2023

    work page 2023

  57. [57]

    Going deeper with directly-trained larger spiking neural networks,

    H. Zheng, Y . Wu, L. Deng, Y . Hu, and G. Li, “Going deeper with directly-trained larger spiking neural networks,” inProceedings of the AAAI conference on artificial intelligence, vol. 35, no. 12, 2021, pp. 11 062–11 070

    2021

  58. [58]

    Sglformer: Spiking global-local-fusion transformer with high performance,

    H. Zhang, C. Zhou, L. Yu, L. Huang, Z. Ma, X. Fan, H. Zhou, and Y . Tian, “Sglformer: Spiking global-local-fusion transformer with high performance,”Frontiers in Neuroscience, vol. 18, 2024

    work page 2024

  59. [59]

    One-step spiking trans- former with a linear complexity,

    X. Song, A. Song, R. Xiao, and Y . Sun, “One-step spiking trans- former with a linear complexity,” inProceedings of the Thirty- ThirdInternational Joint Conference on Artificial Intelligence, Aug. 2024, p. 3142–3150

    work page 2024

  60. [60]

    Assisting training of deep spiking neural networks with parameter initialization,

    J. Ding, J. Zhang, T. Huang, J. K. Liu, and Z. Yu, “Assisting training of deep spiking neural networks with parameter initialization,”IEEE Transactions on Neural Networks and Learning Systems, vol. 36, no. 8, pp. 15 015–15 028, 2025

    work page 2025

  61. [61]

    Spikingresformer: Bridging resnet and vision transformer in spiking neural networks,

    X. Shi, Z. Hao, and Z. Yu, “Spikingresformer: Bridging resnet and vision transformer in spiking neural networks,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024

    work page 2024

  62. [62]

    QKFormer: Hierarchical spiking transformer using q-k attention,

    C. Zhou, H. Zhang, Z. Zhou, L. Yu, L. Huang, X. Fan, L. Yuan, Z. Ma, H. Zhou, and Y . Tian, “QKFormer: Hierarchical spiking transformer using q-k attention,” inThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    work page 2024

  63. [63]

    Spiking transformer with spatial-temporal attention,

    D. Lee, Y . Li, Y . Kim, S. Xiao, and P. Panda, “Spiking transformer with spatial-temporal attention,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, June 2025, pp. 13 948– 13 958

    work page 2025

  64. [64]

    Stca- SNN: self-attention-based temporal-channel joint attention for spiking neural networks,

    X. Wu, Y . Song, Y . Zhou, Y . Jiang, Y . Bai, X. Li, and X. Yang, “Stca- SNN: self-attention-based temporal-channel joint attention for spiking neural networks,”Frontiers in Neuroscience, vol. V olume 17 - 2023, 2023

    work page 2023

  65. [65]

    Tactilegcn: A graph convolutional network for predicting grasp stability with tactile sensors,

    A. Garcia-Garcia, B. S. Zapata-Impata, S. Orts-Escolano, P. Gil, and J. Garcia-Rodriguez, “Tactilegcn: A graph convolutional network for predicting grasp stability with tactile sensors,” in2019 International Joint Conference on Neural Networks. IEEE, 2019, pp. 1–8. 17

    work page 2019

  66. [66]

    Slayer: Spike layer error reassignment in time,

    S. B. Shrestha and G. Orchard, “Slayer: Spike layer error reassignment in time,”Advances in neural information processing systems, vol. 31, 2018

    2018

  67. [67]

    Event-driven tactile sensing with dense spiking graph neural networks,

    F. Guo, F. Yu, M. Li, C. Chen, J. Yan, Y . Li, F. Gu, X. Long, and S. Guo, “Event-driven tactile sensing with dense spiking graph neural networks,” IEEE Transactions on Instrumentation and Measurement, 2025

    work page 2025

  68. [68]

    Recurrent bilinear optimization for binary neural networks,

    S. Xu, Y . Li, T. Wang, T. Ma, B. Zhang, P. Gao, Y . Qiao, J. L ¨u, and G. Guo, “Recurrent bilinear optimization for binary neural networks,” in European conference on computer vision. Springer, 2022, pp. 19–35

    work page 2022

  69. [69]

    Inductive representation learning on large graphs,

    W. Hamilton, Z. Ying, and J. Leskovec, “Inductive representation learning on large graphs,”Advances in neural information processing systems, vol. 30, 2017

    2017

  70. [70]

    Attention spiking neural networks,

    M. Yao, G. Zhao, H. Zhang, Y . Hu, L. Deng, Y . Tian, B. Xu, and G. Li, “Attention spiking neural networks,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 8, pp. 9393–9410, 2023

    2023

  71. [71]

    1.1 computing’s energy problem (and what we can do about it),

    M. Horowitz, “1.1 computing’s energy problem (and what we can do about it),” in2014 IEEE international solid-state circuits conference digest of technical papers. IEEE, 2014, pp. 10–14

    work page 2014

  72. [72]

    Toward high-accuracy and low-latency spiking neural networks with two-stage optimization,

    Z. Wang, Y . Zhang, S. Lian, X. Cui, R. Yan, and H. Tang, “Toward high-accuracy and low-latency spiking neural networks with two-stage optimization,”IEEE Transactions on Neural Networks and Learning Systems, vol. 36, no. 2, pp. 3189–3203, 2025

    work page 2025

  73. [73]

    A million spiking-neuron integrated circuit with a scalable communication network and interface,

    P. A. Merolla, J. V . Arthur, R. Alvarez-Icaza, A. S. Cassidy, J. Sawada, F. Akopyan, B. L. Jackson, N. Imam, C. Guo, Y . Nakamura, B. Brezzo, I. V o, S. K. Esser, R. Appuswamy, B. Taba, A. Amir, M. D. Flickner, W. P. Risk, R. Manohar, and D. S. Modha, “A million spiking-neuron integrated circuit with a scalable communication network and interface,” Scien...

    2014

  74. [74]

    The spinnaker project,

    S. B. Furber, F. Galluppi, S. Temple, and L. A. Plana, “The spinnaker project,”Proceedings of the IEEE, vol. 102, no. 5, pp. 652–665, 2014

    2014

  75. [75]

    Efficient neuromorphic signal processing with loihi 2,

    G. Orchard, E. P. Frady, D. B. D. Rubin, S. Sanborn, S. B. Shrestha, F. T. Sommer, and M. Davies, “Efficient neuromorphic signal processing with loihi 2,” in2021 IEEE Workshop on Signal Processing Systems, 2021, pp. 254–259

    work page 2021

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.