Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Quantized Spike-driven Transformer

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A quantized spike-driven Transformer can reach 80.3% top-1 accuracy on ImageNet-1K with 4-bit weights and binary spikes while using a fraction of the parameters and power of prior spiking Transformers.

desk verdict Solid empirical step for low-bit spike-driven transformers; the theory is mostly packaging, but the ablations and ImageNet result are real. read the letter →

arxiv 2501.13492 v5 pith:U5LUD73W submitted 2025-01-23 cs.CV

Xuerui Qiu , Malu Zhang , Jieyuan Zhang , Wenjie Wei , Honglin Cao , Junsheng Guo , Rui-Jie Zhu , Yimeng Shan
show 2 more authors
Yang Yang Haizhou Li
This is my paper · ORCID
classification cs.CV
keywords spikingneuralnetworksquantizationspike-driventransformerself-attentioninformationentropyknowledgedistillationenergy-efficientinferenceImageNetclassification
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 sets out to prove that a spiking Transformer can be pushed to low-bit weights and binary spikes without the usual accuracy collapse, making spike-driven models viable on memory- and power-limited hardware. It builds a quantized spike-driven Transformer baseline and shows empirically that the collapse comes from the attention module: quantized spike-driven self-attention develops a bimodal score distribution that the paper calls spike information distortion. The remedy is a bi-level optimization that reshapes the attention distribution from inside the neuron and from outside through distillation to an ANN teacher. If the recipe holds, a 6.8M-parameter spiking model can reach 80.3% top-1 on ImageNet-1K, beating the 55.4M-parameter full-precision SNN baseline at roughly one eighth of the parameter count and one sixth of the power. This matters because it points to a path where energy-efficient neuromorphic inference does not require giving up Transformer-level accuracy.

What carries the argument

The load-bearing mechanism is the IE-LIF neuron in its two-regime design. During training it replaces the Heaviside step with a clipped multi-bit output $a^\ell[t] = \frac{1}{b}\mathrm{clip}\{v^\ell[t], 0, b\}$ (Eq. 9), so one timestep can carry levels $\{0, 0.25, 0.5, 0.75, 1\}$; during inference the same neuron emits binary spikes $s^\ell[t]$ over four virtual timesteps (Eq. 10), with $b=4$. A membrane-potential rectify function (MPRF) normalizes and shifts the pre-spike potential to avoid the right-tail bias of a zero-mean Gaussian. The upper-level mechanism is fine-grained distillation (FGD), which minimizes $\sum_{p\in\{q,k,v\}}\sum_{\ell,h}\|F^l_{p_A}-F^l_{p_S}\|$ over $\ell^2$-normalized attention matrices, pulling the quantized spike-driven self-attention (Q-SDSA) distribution toward its ANN counterpart. Together these implement the paper's bi-level reading of mutual-information maximization: IE-LIF maximizes $H(p_S)$ at the lower level, FGD minimizes $H(p_S|p_A)$ at the upper level.

What would settle it

Reproduce the ImageNet-1K training of the QSD-Transformer, then evaluate the same checkpoint twice: once with the training-time multi-bit outputs and once with the binary-spike four-timestep inference; if the binary-spike run falls substantially below the reported 80.3% top-1, the IE-LIF transfer premise fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the performance collapse seen when quantizing a spike-driven Transformer is not primarily a weight-quantization problem but a distribution problem in the attention module. Quantizing the spike-driven self-attention to 4-bit weights plus binary spikes makes the attention scores bimodal, whereas the ANN counterpart's scores are roughly normal; the paper calls this spike information distortion (SID). The proposed fix is a bi-level optimization that maximizes the SNN attention entropy at the neuron level and minimizes conditional entropy relative to an ANN teacher at the network level. Concretely, the information-enhanced LIF (IE-LIF) trains with multi-bit membrane outputs and switches to binary spikes at inference, the membrane-potential rectify function (MPRF) reshapes the pre-spike distribution, and fine-grained distillation (FGD) aligns the q/k/v attention matrices with an ANN Transformer. With this recipe, a 6.8M-parameter QSD-Transformer reaches 80.3% top-1 on ImageNet-1K, 0.6% above the 55.4M-parameter full-precision SD-Transformer v2-L baseline while using 87.6% fewer parameters and 83.4% less estimated power.

Load-bearing premise

The load-bearing premise is that a network trained with five-level multi-bit outputs at one timestep will keep its accuracy when switched to binary spikes over four virtual timesteps at inference; the paper supports this empirically in Table 6 but gives no proof that the two regimes are equivalent.

Editorial extensions

If this is right

  • The quantized model outperforms its full-precision baseline: 80.3% vs 79.7% top-1 on ImageNet-1K for the L-scale models, with 6.8M vs 55.4M parameters and 8.7mJ vs 52.4mJ estimated power.
  • The recipe transfers across tasks: COCO object detection reaches 57.0% mAP@0.5 (vs 51.2% for SD-Transformer v2-M), ADE20K segmentation reaches 40.5% mIoU, and CIFAR-10/CIFAR-100/CIFAR10-DVS transfer hits 98.4%/87.6%/89.8%.
  • The same IE-LIF + FGD treatment lifts a directly quantized Spikformer-8-384 from 64.1% to 75.5% on ImageNet-1K, so the fix is not tied to one architecture.
  • Performance degrades gracefully with bit width: 77.5% at 4-bit, 76.9% at 3-bit, 75.0% at 2-bit on the SD-Transformer v2-T backbone, suggesting ultra-low-bit spiking Transformers are usable.

Reading between the lines

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

  • Editorial inference: if the IE-LIF two-regime trick holds beyond Transformers, multi-bit training with binary-spike inference could become a general recipe for quantized SNNs, including convolutional backbones, where the same bimodal-distribution issue may appear in attention-like or channel-gating modules.
  • Editorial inference: the mutual-information framing gives a testable handle: the accuracy gain from FGD should track the reduction in KL divergence between Q-SDSA and ANN attention distributions, so that divergence could be monitored during training as an early indicator of transfer quality.
  • Editorial inference: the reported power savings use MAC/AC energy constants and omit memory access, which the paper itself flags; actual deployed savings on neuromorphic chips will depend on data movement and could be smaller, so the power figure is a theoretical estimate pending hardware measurement.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The manuscript proposes QSD-Transformer, a spike-driven vision Transformer with low-bit quantized weights and binary spike activities. It identifies a 'spike information distortion' (SID) problem in quantized spike-driven self-attention (Q-SDSA) and proposes a bi-level optimization remedy: at the lower level, an information-enhanced LIF (IE-LIF) neuron with a membrane-potential rectify function (MPRF) is used to increase the information entropy of attention scores; at the upper level, a fine-grained distillation (FGD) scheme aligns the student SNN's attention distributions with an ANN teacher's. Experiments report 80.3% top-1 accuracy on ImageNet-1K with 6.8M parameters and 8.7mJ power, as well as results on object detection, semantic segmentation, and transfer learning.

Significance. If the headline result is correct, this is a meaningful advance: a 4-bit-weight, 1-bit-spike transformer that beats a full-precision SNN baseline at a fraction of the parameter and energy budget. The empirical ablations are internally consistent, the method transfers across several tasks, and the reported gains over SD-Transformer v2 are substantial. However, the central claim rests on a multi-bit-to-binary conversion that is not cleanly specified (Eq. 10), on a comparison to a much larger baseline model that conflates architecture scaling with quantization, and on a theoretical framing that is not actually established. The significance is therefore conditional on resolving these issues.

major comments (4)
  1. [§4.2, Eq. (10)] The equation equates the multi-bit training activation a_l[t] (range [0,1] for b=4) with the sum of T=4 binary spikes s_l[t] (range [0,4]). As written, no 1/b normalization, threshold adjustment, or weight rescaling appears in the forward model, so inference would operate at up to 4× the training activation scale. This is load-bearing for the claim that the 80.3% result is achieved with binary-spike inference. Please state the exact conversion (e.g., a=(1/b)Σs or an equivalent thresholded sum) and confirm that the implementation used in Tables 1 and 6 follows it.
  2. [§5, Table 1] The labeled comparison 'QSD-Transformer SD-Transformer v2-L' versus 'SD-Transformer v2-L' is not a controlled quantization comparison, because the parameter counts are 6.8M versus 55.4M. The 6.8M model is a different, smaller architecture (Appendix G, Table 9), not the same architecture with quantized weights. The claimed 87.58% parameter reduction and 83.40% power reduction therefore conflate architecture scaling with quantization. Please either evaluate the same architecture before and after quantization, or explicitly state that the comparison is architecture-plus-quantization.
  3. [§4.2–4.3, Proposition 1, Eq. (8), Eq. (13), Appendix B] The proof of Proposition 1 is a central-limit-theorem statement for a normalized sum of independent binomial variables; it does not establish the stated equality H(Σ pS[t]) = H(pA). The premise that q/k/v attention scores follow independent binomial distributions is assumed rather than derived. Similarly, the move from Eq. (11) to the claim that IE-LIF maximizes H(pS) is not formally justified: Eq. (11) is not written as a properly normalized probability mass function, and MPRF is introduced without a proof that the rectified discrete distribution attains the bound of Proposition 2. The bi-level formulation in Eq. (8) is therefore motivational rather than derived. Please weaken these formal claims or supply the missing steps.
  4. [§5, Table 6 and Appendix G] The ablation intended to validate the multi-bit-to-binary transfer does not isolate the exact protocol used for the headline result. The table varies b and training T but does not state the inference-side b/T used for each row, and it appears to be run on a smaller model (the 77.5% values match the Table 5 / Table 1-T scale) rather than the 6.8M model behind the 80.3% result. A clean comparison is needed on the 6.8M model: training with b=4,T=1 and then evaluating both with the multi-bit activation and with the binary-spike inference protocol, so the transfer loss can be measured.
minor comments (5)
  1. [Abstract vs. Appendix F] The abstract states that code is available at a GitHub URL, while Appendix F says codes and models will be made available only after review. Please reconcile these two statements.
  2. [Table 1, notation] The 'Bits 4-1' column does not reflect that training activations are multi-bit; please clarify in the table header or caption that the 1-bit activity applies at inference and that training uses multi-bit IE-LIF outputs.
  3. [Appendix D, Eqs. (25)–(26)] Please specify whether µ(pmem) and σ(pmem) are running statistics computed during training; as written, σ depends on the input S, so the folding of MPRF into weights is not a constant affine transformation unless these statistics are fixed.
  4. [Eq. (11)] The entropy expression uses Dirac deltas in a discrete sum; please rewrite it as a proper discrete probability mass function (e.g., p_k = G(k/b) with normalization) to make the formula formally correct.
  5. [Figure 2] The labels in Figure 2 appear garbled in the provided manuscript, making the SID diagnostic plot difficult to read; please provide a clean version.

Circularity Check

1 steps flagged · score 6.0 of 10

The central multi-bit-to-binary spike conversion is asserted in Eq. (10) as an unnormalized equality, so the claimed spike-driven inference reduces to a definition rather than a derived equivalence.

  1. self definitional [Section 4.2, Eq. (10), read together with Eq. (9)]
    "Moreover, the output xℓ+1[t] of each layer in the SNN is represented as: xℓ+1[t] =wℓ · aℓ[t] =wℓ · Σ_{t=1}^T sℓ[t], (10) where aℓ[t] represents multi-bit spikes during training with one timestep and is denoted as {0, 0.25, 0.5, 0.75, 1}^{T=1}, while sℓ[t] represents binary spikes during inference and is extended to 4 virtual timesteps denoted as {0, 1}^{T=4}, with a maximum integer value b set to 4 in this paper."

    Eq. (9) defines the training output as a = (1/b) clip(v,0,b), so with b=4 the multi-bit activity a lies in {0, 0.25, ..., 1}. Eq. (10) then asserts x = w·a = w·Σ_{t=1}^{T} s[t], where inference spikes s are in {0,1} over T=4 virtual timesteps, so the sum ranges over {0,1,2,3,4}. The equality is therefore false as written unless an undocumented 1/4 normalization or threshold/weight rescaling is inserted. Because the paper's central claim is that IE-LIF converts multi-bit training to binary-spike inference while preserving performance, the conversion is not an empirical finding or a derived limit; it is a definitional identification of two quantities with different ranges and scales. The reported accuracy under this rule is an artifact of that identification.

full rationale

Apart from Eq. (10), the paper's ablations and benchmark results are empirical outcomes rather than fitted quantities, and the self-citations to prior spike-driven Transformer work are not load-bearing. The information-entropy propositions are stated mathematical results with explicit assumptions, and the FGD loss is a concrete training objective. However, the key mechanism for reconciling multi-bit training with binary-spike inference is not derived: Eq. (10) identifies the multi-bit activation a (range [0,1] with b=4) with the sum of four binary spikes (range [0,4]) with no normalization or threshold rescaling shown. The claimed equivalence between training and inference is therefore carried by the definitional equation itself, making the central spike-driven conversion circular as written.

Assumptions & free parameters 3 free parameters · 6 assumptions · 1 invented entities

The central claim rests on several hand-chosen or learned parameters (MPRF scale/shift, IE-LIF bit width b, distillation weight lambda), standard domain assumptions about membrane potential distributions and energy costs, and an ad hoc assumption that L2 similarity matching implements entropy minimization. The only invented conceptual entity is SID, which is defined by the paper itself and not externally verified.

free parameters (3)
  • MPRF scale and shift (gamma, alpha) = learned per layer during training
    Eq. 12, Section 4.2. These parameters control the rectification that aligns Q-SDSA membrane potentials to a standard normal; they are fitted on the training set, not derived.
  • IE-LIF maximum integer value b = 4
    Section 4.2. Hand-chosen to equal the inference timestep; Table 6 shows accuracy varies with b (77.5 at b=4 vs 67.6 at b=1).
  • Distillation loss weight lambda = 2
    Eq. 14. Chosen by hand; no ablation on lambda is provided.
assumptions (6)
  • standard math Central Limit Theorem: sums of independent random variables converge to a Gaussian distribution.
    Used in Proposition 1 (Appendix B) to claim SNN attention scores become normally distributed as T approaches infinity.
  • domain assumption Membrane potentials in Q-SDSA approximately follow N(0, sigma).
    Section 4.2, citing Guo et al. 2022a,b; this assumption justifies the MPRF and the entropy calculation in Eq. 11.
  • ad hoc to paper Minimizing L2 distance between normalized q/k/v outer products reduces conditional entropy H(pS|pA).
    Eq. 13: the paper states the optimal solution is pS=pA, but does not prove that the chosen Frobenius loss on similarity matrices minimizes mutual information or conditional entropy.
  • domain assumption The AC energy model, where low-bit spike-driven operations are counted as additions rather than multiply-accumulates, is a valid proxy for hardware power.
    Appendix E: energy estimates rely on 4.6pJ MAC and 0.9pJ AC on 45nm (Horowitz 2014), a standard assumption in SNN literature, but not verified on real hardware.
  • domain assumption The straight-through estimator (STE) and surrogate gradients provide valid update directions for quantization and spiking non-differentiabilities.
    Section 3 and Appendix A: STE for quantization and the rectangle surrogate for spike firing are standard practice, though not theoretically guaranteed.
  • ad hoc to paper The ANN teacher's attention distributions are a suitable target for the student SNN.
    Section 4.3: the teacher is defined as 'the same ANN Transformer', but its exact architecture, training, and accuracy are not specified, and the optimality of this target is assumed.
invented entities (1)
  • Spike Information Distortion (SID)
    purpose: Explains why quantizing spike-driven self-attention causes performance degradation: the bimodal distribution limits representational capacity.
    Introduced in Section 4.1, supported only by the paper's own histogram analysis (Fig. 2b) and ablations; no independent falsifiable prediction outside this system.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantized Spike-driven Transformer." pith.science (2026). https://pith.science/paper/U5LUD73W

@misc{pith2026250113492,
  author       = {Pith},
  title        = {Pith review of: Quantized Spike-driven Transformer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U5LUD73W}},
  note         = {Machine review of arXiv:2501.13492}
}
abstract

Spiking neural networks are emerging as a promising energy-efficient alternative to traditional artificial neural networks due to their spike-driven paradigm. However, recent research in the SNN domain has mainly focused on enhancing accuracy by designing large-scale Transformer structures, which typically rely on substantial computational resources, limiting their deployment on resource-constrained devices. To overcome this challenge, we propose a quantized spike-driven Transformer baseline (QSD-Transformer), which achieves reduced resource demands by utilizing a low bit-width parameter. Regrettably, the QSD-Transformer often suffers from severe performance degradation. In this paper, we first conduct empirical analysis and find that the bimodal distribution of quantized spike-driven self-attention (Q-SDSA) leads to spike information distortion (SID) during quantization, causing significant performance degradation. To mitigate this issue, we take inspiration from mutual information entropy and propose a bi-level optimization strategy to rectify the information distribution in Q-SDSA. Specifically, at the lower level, we introduce an information-enhanced LIF to rectify the information distribution in Q-SDSA. At the upper level, we propose a fine-grained distillation scheme for the QSD-Transformer to align the distribution in Q-SDSA with that in the counterpart ANN. By integrating the bi-level optimization strategy, the QSD-Transformer can attain enhanced energy efficiency without sacrificing its high-performance advantage. For instance, when compared to the prior SNN benchmark on ImageNet, the QSD-Transformer achieves 80.3% top-1 accuracy, accompanied by significant reductions of 6.0$\times$ and 8.1$\times$ in power consumption and model size, respectively. Code is available at https://github.com/bollossom/QSD-Transformer.

Figures

Figures reproduced from arXiv: 2501.13492 by the authors.

Figure 1
Figure 1. Overview of the QSD-Transformer. (a) Proposed quantized spike-driven self-attention (Q [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. (a) Accuracy of quantizing different modules in the SD-Transformer v2 and its same ANN [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. (a) Comparative results of accuracy, power, and parameters on ImageNet. (b) Comparison [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visualization of results on COCO dataset. Our QSD-Transformer excels in the target [PITH_FULL_IMAGE:figures/full_fig_p024_4.png]
Figure 5
Figure 5. Figure 5: Visualization of results on ADE20K dataset. Our QSD-Transformer excels in the segmenta [PITH_FULL_IMAGE:figures/full_fig_p025_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Efficient Logit-based Knowledge Distillation of Deep Spiking Neural Networks for Full-Range Timestep Deployment

    cs.LG 2025-01 conditional novelty 5.0 of 10

    Temporal-wise logits distillation with ensemble self-distillation lets one SNN, trained at T=6, be deployed at T=1 through T=6 without retraining and with competitive accuracy.

Reference graph

Works this paper leans on

34 extracted references · 21 canonical work pages · cited by 1 Pith paper

  1. [1]

    Estimating or propagating gradients through stochastic neurons for conditional computation

    Yoshua Bengio, Nicholas L´eonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432,

    arXiv

  2. [2]

    For a random variable x ∼ N(µ, σ), the information entropy x reaches its maximum value H(x) = 1 2 log 2πeσ 2(x) and is observed to increase with the expansion of variance σ. Proof. For a continuous random variable x obeying a normal distribution, its probability density function p(x) is given by: p(x) = 1 (2πσ 2)1/2 exp ( − (x − µ)2 2σ2 ) , (20) Consequen...

    work page 2025

  3. [7]

    Distilling the knowledge in a neural network

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531,

    arXiv

  4. [8]

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

    Mark Horowitz. 1.1 computing’s energy problem (and what we can do about it). In 2014 IEEE International Solid-State Circuits Conference Digest of Technical Papers (ISSCC), pp. 10–14,

    work page 2014

  5. [9]

    For the three different model scales—1.8M, 3.8M and 6.8M parameters—we allocated 24, 28 and 36 hours of training time, respectively

    We conducted training on eight 40GB A100 GPUs. For the three different model scales—1.8M, 3.8M and 6.8M parameters—we allocated 24, 28 and 36 hours of training time, respectively. Table 8: Hyper-parameters for image classification on ImageNet-1K and CIFAR10/100. Hyper-parameter ImageNet CIFAR10/10 Timestep (Training/Inference) 1/4 1/4 Epochs 300 100 Resol...

    work page 2025

  6. [12]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision– ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pp. 740–755. Springer,

    2014

  7. [13]

    Spikebert: A language spikformer trained with two-stage knowledge distillation from bert

    Changze Lv, Tianlong Li, Jianhan Xu, Chenxi Gu, Zixuan Ling, Cenyuan Zhang, Xiaoqing Zheng, and Xuanjing Huang. Spikebert: A language spikformer trained with two-stage knowledge distillation from bert. arXiv preprint arXiv:2308.15122,

    arXiv

  8. [15]

    Efficient 3d recognition with event-driven spike sparse convolution

    Xuerui Qiu, Man Yao, Jieyuan Zhang, Yuhong Chou, Ning Qiao, Shibo Zhou, Bo Xu, and Guoqi Li. Efficient 3d recognition with event-driven spike sparse convolution. arXiv preprint arXiv:2412.07360, 2024a. Xuerui Qiu, Rui-Jie Zhu, Yuhong Chou, Zhaorui Wang, Liang-Jian Deng, and Guoqi Li. Gated attention coding for training high-performance and efficient spiki...

    arXiv

Show all 34 references
  1. [17]

    Spikingresformer: Bridging resnet and vision transformer in spiking neural networks

    Xinyu Shi, Zecheng Hao, and Zhaofei Yu. Spikingresformer: Bridging resnet and vision transformer in spiking neural networks. arXiv preprint arXiv:2403.14302,

    arXiv

  2. [19]

    Training low-latency spiking neural network through knowledge distillation

    Sugahara Takuya, Renyuan Zhang, and Yasuhiko Nakashima. Training low-latency spiking neural network through knowledge distillation. In 2021 IEEE Symposium in Low-Power and High-Speed Chips, pp. 1–3. IEEE,

    2021

  3. [20]

    Training low-latency deep spiking neural networks with knowledge distillation and batch normalization through time

    Thi Diem Tran, Kien Trung Le, and An Luong Truong Nguyen. Training low-latency deep spiking neural networks with knowledge distillation and batch normalization through time. In 2022 5th International Conference on Computational Intelligence and Networks, pp. 01–06. IEEE,

    2022

  4. [21]

    Event-driven learning for spiking neural networks

    Wenjie Wei, Malu Zhang, Jilin Zhang, Ammar Belatreche, Jibin Wu, Zijing Xu, Xuerui Qiu, Hong Chen, Yang Yang, and Haizhou Li. Event-driven learning for spiking neural networks. arXiv preprint arXiv:2403.00270,

    arXiv

  5. [22]

    Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips

    Man Yao, JiaKui Hu, Tianxiang Hu, Yifan Xu, Zhaokun Zhou, Yonghong Tian, XU Bo, and Guoqi Li. Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. In The Twelfth International Conference on Learning R...

    2025 arXiv

  6. [23]

    Resnest: Split-attention networks

    Hang Zhang, Chongruo Wu, Zhongyue Zhang, Yi Zhu, Haibin Lin, Zhi Zhang, Yue Sun, Tong He, Jonas Mueller, R Manmatha, et al. Resnest: Split-attention networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 2736–2746, 2022a. Jiqing Zha...

    2022

  7. [24]

    Spikingformer: Spike-driven residual learning for transformer-based spiking neural network

    Chenlin Zhou, Liutao Yu, Zhaokun Zhou, Zhengyu Ma, Han Zhang, Huihui Zhou, and Yonghong Tian. Spikingformer: Spike-driven residual learning for transformer-based spiking neural network. arXiv preprint arXiv:2304.11954, 2023a. Chenlin Zhou, Han Zhang, Liutao Yu, Yumin Ye, Zhaok...

  8. [25]

    16 Published as a conference paper at ICLR 2025 Appendix A B ACKPROPAGATION PROCESS OF SPIKING NEURONS There exist two primary methods of training high-performance SNNs. One way is to discretize ANN into spike form through neuron equivalence (Li et al., 2021; Bu et al., 2022; ...

    2025

  9. [26]

    and apply surrogate gradient training. Then in this section, we introduce the training process of SNN gradient descent and the parameter update method of spatio-temporal backpropagation (STBP) (Wu et al., 2018; Xiao et al., 2022). SNNs’ parameters can be taught using gradient ...

    2018

  10. [29]

    Many performance analyses in the research of spiking neural networks (SNNs) (Panda et al., 2020; Qiu et al., 2024a; Shan et al.,

    indicates that, on a 45nm process hardware platform, the energy consumption for a single multiply-accumulate (MAC) operation is 4.6pJ (with 3.7pJ for multiplication and 0.9pJ for addition). Many performance analyses in the research of spiking neural networks (SNNs) (Panda et a...

    2020

  11. [30]

    also reference this data. E.1 C OMPARISION ON MHSA AND SDSA Given a float-point input sequence X ∈ RN ×D, the float-point Query (q), Key (k), and Value (v) in RN ×D are computed using three learnable linear matrices, where N is the token number, and D is the channel dimension....

    2020

  12. [31]

    Table 7: Theoretical FLOPs/SOPs of self-attention modules. Multi-head Self-attention (MHSA) Spike-driven Self-attention (SDSA) Function MHSA(q, k, v) =softmax qkT √ d v SDSA(qs, ks, ks) =SN s((qsks T)vs) q, k, v 3N D2 T · f r1 · 3 · F LConv f (q, k, v) 2N 2D T · f r2 · N D2 Sc...

    2024

  13. [32]

    CIFAR10-DVS consists of 10K event streams created by capturing CIFAR10 images using a DVS camera. In these experiments, we first loaded pre-trained ImageNet-1K checkpoints and replaced the final fully connected layer to match the number of classes in each dataset (e.g., replac...

    2023

  14. [33]

    We conducted training on four 40GB A100 GPUs for a duration of 26 hours

    We used the AdamW optimizer with an initial learning rate of 1e-4, and the learning rate was decayed polynomially with a power of 0.9. We conducted training on four 40GB A100 GPUs for a duration of 26 hours. G.3 ADE20K EXPERIMENTS The ADE20K semantic segmentation dataset compr...

    2025

  15. [1997]

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

    13 Published as a conference paper at ICLR 2025 Paul A Merolla, John V Arthur, Rodrigo Alvarez-Icaza, Andrew S Cassidy, Jun Sawada, Filipp Akopyan, Bryan L Jackson, Nabil Imam, Chen Guo, Yutaka Nakamura, et al. A million spiking- neuron integrated circuit with a scalable commu...

    2025

  16. [2012]

    1 − j2 2T + o j√ T σ 2!#T =

    of the random variable yS [t] be φyS [t](j). Then let the random variable η = yS [1]+yS [2]+yS [3]+...+yS [T ]√ tσ . Then the characteristic function of η is: φη = φyS [t]( j√ T σ ) · φyS [t]( j√ T σ ) . . . φyS [t]( j√ T σ ) = φyS [t]( j√ T σ ) T , (18) Then when SNN’s timest...

    2025

  17. [2014]

    JiaKui Hu, Man Yao, Xuerui Qiu, Yuhong Chou, Yuxuan Cai, Ning Qiao, Yonghong Tian, Bo Xu, and Guoqi Li

    doi: 10.1109/ISSCC.2014.6757323. JiaKui Hu, Man Yao, Xuerui Qiu, Yuhong Chou, Yuxuan Cai, Ning Qiao, Yonghong Tian, Bo Xu, and Guoqi Li. High-performance temporal reversible spiking neural networks with o(l) training memory and o(1) inference cost. arXiv preprint arXiv:2405.16...

    2014

  18. [2015]

    Reducing ann-snn conversion error through residual membrane potential

    Zecheng Hao, Tong Bu, Jianhao Ding, Tiejun Huang, and Zhaofei Yu. Reducing ann-snn conversion error through residual membrane potential. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pp. 11–21, 2023a. Zecheng Hao, Jianhao Ding, Tong Bu, Tiejun Hu...

    arXiv

  19. [2017]

    Spike calibration: Fast and accurate conversion of spiking neural network for object detection and segmentation

    Yang Li, Xiang He, Yiting Dong, Qingqun Kong, and Yi Zeng. Spike calibration: Fast and accurate conversion of spiking neural network for object detection and segmentation. arXiv preprint arXiv:2207.02702,

    arXiv

  20. [2018]

    Imagenet: A large-scale hier- archical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hier- archical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pp. 248–255. IEEE,

    2009

  21. [2019]

    Quantizing deep convolutional networks for efficient inference: A whitepaper

    Raghuraman Krishnamoorthi. Quantizing deep convolutional networks for efficient inference: A whitepaper. arXiv preprint arXiv:1806.08342,

    arXiv

  22. [2020]

    Advancing spiking neural networks towards multiscale spatiotemporal interaction learning

    Yimeng Shan, Malu Zhang, Rui-jie Zhu, Xuerui Qiu, Jason K Eshraghian, and Haicheng Qu. Advancing spiking neural networks towards multiscale spatiotemporal interaction learning. arXiv preprint arXiv:2405.13672,

    arXiv

  23. [2021]

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

    11 Published as a conference paper at ICLR 2025 Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for ...

    2025

  24. [2022]

    Mmdetection: Open mmlab detection toolbox and benchmark

    Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155,

    1906 arXiv

  25. [2023]

    Jianhao Ding, Zhaofei Yu, Yonghong Tian, and Tiejun Huang

    doi: 10.1109/TNNLS.2021.3109064. Jianhao Ding, Zhaofei Yu, Yonghong Tian, and Tiejun Huang. Optimal ann-snn conversion for fast and accurate inference in deep spiking neural networks. arXiv preprint arXiv:2105.11654,

    2021

  26. [2024]

    Slayer: spike layer error reassignment in time

    14 Published as a conference paper at ICLR 2025 Sumit Bam Shrestha and Garrick Orchard. Slayer: spike layer error reassignment in time. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, pp. 1419–1428,

    2025

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.