Pith. sign in

REVIEW 4 major objections 5 minor 78 references

SPEAR: Structured Pruning for Spiking Neural Networks via Synaptic Operation Estimation and Reinforcement Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read SPEAR claims that a linear regression predicting post-fine-tuning SynOps from pre-fine-tuning SynOps is enough to make RL-based structured pruning of spiking neural networks meet an explicit SynOps budget.

desk verdict A practical RL-based structured pruning method for SNNs with a genuinely useful SynOps estimator, but the LRE validation gap and missing seeds make the headline numbers fragile; worth serious review. read the letter →

arxiv 2507.02945 v1 pith:7WYO4SHZ submitted 2025-06-28 cs.NE cs.AIcs.LG

Hui Xie , Yuhe Liu , Shaoqi Yang , Jinyang Guo , Yufei Guo , Yuqing Ma , Jiaxin Chen , Jiaheng Liu
show 1 more author
Xianglong Liu
This is my paper
classification cs.NEcs.AIcs.LG MSC 68T07
keywords spikingneuralnetworksstructuredpruningsynapticoperationsSynOpsreinforcementlearninglinearregressiontarget-awarerewardneuromorphichardware
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

SPEAR is a reinforcement-learning framework for structured pruning of deep spiking neural networks (SNNs) that takes synaptic operations (SynOps) as a direct search constraint. The paper sets out to show that the main obstacle—SynOps changes irregularly and substantially after the standard fine-tuning step, so a network selected on pre-fine-tuning counts can end up off-budget—can be removed by a simple linear regression that predicts post-fine-tuning SynOps from pre-fine-tuning SynOps. The predicted count then enters a target-aware reward that penalizes only over-budget states, letting the RL agent learn pruning ratios without a hard, layer-by-layer budget calculation that SNNs cannot provide. If this works as reported, SNN compression can meet energy and storage budgets before the costly final fine-tuning, and the search stays cheap because SynOps estimates come from a fitted line rather than from repeated training.

What carries the argument

LRE (linear regression for SynOps estimation) is the object that carries the argument: a single affine map learned from a small sample of randomly pruned and fine-tuned networks, which converts the unpredictable post-fine-tuning SynOps into a quantity the RL agent can see in its state and reward. TAR (target-aware reward) is the companion mechanism: a reward term that is zero below the target and grows exponentially above it, so the hard SynOps constraint becomes a smooth optimization signal. The RL agent is DDPG, which outputs a continuous pruning ratio per layer, and the state includes the estimated remaining SynOps plus layer geometry and the previous action.

What would settle it

Take a trained SNN, fit LRE on randomly pruned policies, then evaluate the learned map on held-out pruning policies including the final policy produced by the agent after search; if the predicted post-fine-tuning SynOps deviates from the measured value by more than a small tolerance, or if a network selected to satisfy a 50% SynOps budget lands substantially off-budget after fine-tuning, the central claim fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that post-fine-tuning SynOps is linearly correlated with pre-fine-tuning SynOps for structured-pruned SNNs, and that this one fitted relation, $\text{SynOps}_{es} = W \cdot \text{SynOps}_{cur} + b$, is enough to keep an RL-based search within a specified SynOps ratio. The learned estimator feeds the searched network's state, and the target-aware reward $R_s = \text{Acc} + F(S_{es}, S_t)$ with $F = -\lambda \cdot \max(S_{es}/S_t - 1, 0)^\alpha$ applies a unilateral exponential penalty when the estimate exceeds the target; an analogous term handles parameter count. Reported results include VGG16 on CIFAR10 at 91.77% accuracy with 52.5% SynOps and 14.4% parameters, and ResNet18 on ImageNet at 54.66% accuracy with 72.9% SynOps, both with lower SynOps than the SCA-based comparison at comparable or better accuracy. Ablations attribute the budget adherence to LRE and the stability of the search to TAR.

Load-bearing premise

The load-bearing premise is that the linear relationship between SynOps before and after fine-tuning, learned from a small set of randomly pruned and fine-tuned networks, holds for every pruning policy the RL agent visits and for the final searched policy.

Editorial extensions

If this is right

  • SynOps can be used as a direct constraint during SNN pruning search, so the final fine-tuned network does not need to be re-measured and re-pruned to discover a budget violation.
  • The same RL search can target SynOps alone, parameter count alone, or both by selecting among $R_s$, $R_p$, and $R_{sp}$, so one framework covers energy-dominated and storage-dominated deployment scenarios.
  • The search is affordable because LRE needs only a small set of pruned and fine-tuned networks for fitting and uses a 500-sample proxy for dataset-level SynOps.
  • Reported results put the method ahead of SCA-based and NetworkSliming baselines on SynOps, parameters, and Top-1 accuracy simultaneously for configurations such as VGG16 on CIFAR10.
  • The compressed models show practical gains, including a 1.67x speedup and lower estimated energy at comparable accuracy for VGG16.

Reading between the lines

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

  • The linearity that LRE exploits is itself an empirical claim about SNN fine-tuning dynamics; a direct extension would test whether the slope and intercept transfer across architectures, neuron models, numbers of timesteps, and fine-tuning schedules.
  • If the regression generalizes, the estimator could replace the RL loop: a per-layer SynOps decomposition together with the affine map would turn a global SynOps target into per-layer channel budgets analytically, making the search unnecessary for simple settings.
  • Because LRE is fit on randomly pruned networks, its use inside RL extrapolates to non-random policies; holding out the final searched policy and partially pruned networks for validation would quantify how much extrapolation is actually happening.
  • TAR's soft penalty makes budget satisfaction probabilistic rather than guaranteed; a deployed system would likely add a hard clip or a verification pass on the final policy.
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 paper introduces SPEAR, a structured pruning framework for spiking neural networks (SNNs) that uses a reinforcement-learning (DDPG) agent to output per-layer pruning ratios. To incorporate a SynOps budget, the authors propose LRE, a linear regression that maps pre-finetuning SynOps to post-finetuning SynOps, and TAR, a target-aware reward that applies a soft penalty when the estimated SynOps or parameter count exceeds user-specified targets. The agent's state includes structural features and an LRE-based SynOps estimate. Experiments on CIFAR10/100, Tiny-ImageNet, ImageNet, and CIFAR10-DVS with VGG16, ResNet18, and a 5Conv+1FC network report that SPEAR achieves lower SynOps and parameter fractions than prior SNN pruning methods at comparable or higher accuracy. The paper also includes ablations on LRE, TAR, hyperparameters, and a comparison with a handcrafted policy.

Significance. The paper addresses a relevant and under-explored problem: directly using SynOps as a constraint in search-based structured pruning for SNNs, where post-finetuning SynOps drift makes hard budgets unreliable. The proposed LRE and TAR components are well motivated, and the experimental evaluation is broad in terms of datasets and architectures. The reported gains are large—for example, +10.11% top-1 accuracy over SCA-based on Tiny-ImageNet VGG16 with comparable parameters. The ablations qualitatively support the value of TAR and of using estimated post-finetuning SynOps rather than raw pre-finetuning values. However, the load-bearing LRE component is validated only in-sample, and the actual target budgets used in the main comparison are not reported, so the central claim of constraint satisfaction is not yet established. The absence of error bars or multiple-seed runs further weakens the evidence for a stochastic method. These issues are addressable with additional experiments, so the work has solid potential but requires substantial revision.

major comments (4)
  1. [Section 4.2, Eq. (3), Figs. 3-4] The LRE estimator is fitted on SynOps pairs from randomly generated pruning policies and is validated only in-sample, as shown in the Fig. 3 scatter plot and the Fig. 4 R2 values. It is then applied in the same form to (i) partially pruned states used as features in Eq. (7) (Ses) and (ii) the final policy's SynOps in the reward of Eq. (4), whose layer-correlated pruning ratios differ from the random policies used for fitting. The paper does not report held-out prediction error on unseen policies, partial-pruning states, or the final searched policy, so the claim that LRE 'accurately predicts the final SynOps' (Abstract) is not supported for the actual search distribution. Please add a held-out validation split of the random-fit data and report the LRE estimate versus the actual post-finetuning SynOps for the policy discovered by the trained agent.
  2. [Section 5.1, Table 1] The implementation details state that 'We adjust the target SynOps and #parameters to achieve similar SynOps/#parameters as the baseline methods for fair comparison.' Because the target budget is set after seeing the baseline numbers, the reported SynOps values in Table 1 are not demonstrations that a pre-specified constraint was met; they are the outcome of a search whose target was chosen to match the baselines. The paper should report the exact target SynOps and parameter ratios used for each row, together with the difference between the estimated SynOps (Ses) used during search and the actual post-finetuning SynOps. Without this information, the central claim of 'effectively compress SNN under specific SynOps constraint' is not directly testable.
  3. [Section 5.2, Tables 1-4] No error bars, confidence intervals, or multiple-seed runs are reported for any of the main results or ablations. Since both the RL exploration (truncated normal noise in Section 4.4) and the LRE fitting are stochastic, single-run numbers cannot establish that the reported accuracy/SynOps differences are reliable. Please report mean and standard deviation across at least three independent runs for the key tables, or explicitly state that all experiments used a fixed seed and demonstrate determinism.
  4. [Table 2, 'SPEAR (w/o LRE)'] The ablation 'SPEAR (w/o LRE)' demonstrates that using pre-finetuning SynOps as the constraint overshoots the budget, but it does not measure the error of the LRE itself on the final searched policy. To support the claim that LRE avoids constraint violation, the paper should report, for the final policy produced by the full SPEAR pipeline, the LRE estimate used during search versus the post-finetuning SynOps actually observed. This is the only direct test of whether LRE's prediction holds for the policy that matters.
minor comments (5)
  1. [Section 1, first paragraph] The phrase 'have attract many attention' should be 'have attracted much attention', and 'deployment SNNs' should be 'deployment of SNNs'.
  2. [Section 4.2, Eq. (3)] The notation 'SynOpses' (and 'SynOpscur') is inconsistent with the subscripted 'SynOps_es' used elsewhere; please unify the notation and use subscripts consistently.
  3. [Fig. 3 caption] The caption mentions 'Pearson Correlation Coefficient' but no coefficient values are reported in the figure or text; please include the values and state the number of sampled policies used to fit LRE.
  4. [Section 5.3, Table 4 caption] The caption says 'Sensitive analysis' and should read 'Sensitivity analysis'; the '[4]' in the caption appears to be a stray reference.
  5. [NeurIPS Checklist, items 2 and 7] The checklist states that limitations are discussed in Section 5.3 and Appendix D and that statistical significance/error bars are reported, but the main text contains no Limitations section and no error bars; the checklist answers should be corrected to match the actual content.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; LRE is a fitted surrogate with an extrapolation risk, not a tautology.

full rationale

SPEAR's central derivation is not circular. LRE (Eq. 3) fits W and b to SynOps measured before and after fine-tuning on randomly sampled pruning policies; the fitted line is then used to estimate post-fine-tuning SynOps for policies visited by the RL agent. The final reported SynOps are measured after the search and fine-tuning, not defined by Eq. 3. The paper's own ablation (SPEAR w/o LRE, Table 2) shows that using raw pre-fine-tuning SynOps misses the target, and Table 4 shows the soft TAR penalty does not force the outcome (e.g., lambda=0.1 yields 73.18% SynOps), so neither the estimator nor the reward defines the reported result by construction. No load-bearing self-citation or imported uniqueness theorem appears; comparisons are against external baselines. The in-sample R2 evidence for LRE and the extrapolation to final and partial policies is a legitimate statistical validity concern, but it is about generalization, not circular equivalence.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The central claim rests on an empirical linearity assumption (LRE), a set of hand-chosen reward and sampling hyperparameters, and standard RL/pruning machinery. The only genuinely new content is the observation of linear correlation and its use inside an RL loop; everything else is either standard or tuned to the experiments.

free parameters (8)
  • LRE scale W = not reported
    Fitted scalar slope in Eq. (3) mapping pre-finetuning SynOps to post-finetuning SynOps; fit on a small set of randomly pruned and fine-tuned networks per dataset/architecture.
  • LRE bias b = not reported
    Fitted scalar intercept in Eq. (3); same fit as W.
  • TAR coefficient lambda = 1.0 (default)
    Balances accuracy reward and SynOps penalty in Eq. (4); sensitivity analyzed in Table 4.
  • TAR exponent alpha = 1.2 (default)
    Sets penalty growth rate in Eq. (4); sensitivity analyzed in Table 4.
  • SynOps sampling tolerance = 1% error, 500 samples
    Stopping criterion for the sample-based SynOps proxy in Appendix C; empirically chosen.
  • Target SynOps ratio = 0.4-0.8 depending on experiment
    Chosen per experiment to match baseline compression levels (Section 5.1); directly determines the constraint the search optimizes against.
  • Target parameter ratio = 0.2-0.6 depending on experiment
    Chosen per experiment to match baseline parameter counts (Section 5.1).
  • RL noise schedule = std 0.5, decay 0.98 per episode
    Exploration noise parameters for DDPG (Section 4.4); set by hand.
assumptions (5)
  • domain assumption Linear relationship between pre- and post-finetuning SynOps
    Observed empirically in Fig. 3 for a few random pruning policies, assumed to hold for all policies visited during search (Section 4.2).
  • domain assumption L1-norm criterion is a valid base pruner
    The LRE training data is generated by pruning with the L1-norm criterion [32], and the RL agent's actions are executed through the same criterion (Section 4.2, Algorithm 1).
  • domain assumption SynOps is the dominant energy metric on neuromorphic hardware
    Adopted from prior work [13, 48] and used as the primary constraint; the paper does not model other hardware costs.
  • domain assumption Pre-finetuning validation accuracy is a suitable reward proxy for post-finetuning accuracy
    The TAR reward uses the accuracy of the pruned but un-finetuned model (Algorithm 1), while reported accuracies are after 210 epochs of finetuning; no analysis links the two (Section 4.3).
  • standard math Standard DDPG convergence assumptions
    The RL search relies on DDPG [36] without modification; no convergence guarantees are stated (Section 4.4).

how reviews work

0 comments
Cite this review

Pith. "Pith review of SPEAR: Structured Pruning for Spiking Neural Networks via Synaptic Operation Estimation and Reinforcement Learning." pith.science (2026). https://pith.science/paper/7WYO4SHZ

@misc{pith2026250702945,
  author       = {Pith},
  title        = {Pith review of: SPEAR: Structured Pruning for Spiking Neural Networks via Synaptic Operation Estimation and Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7WYO4SHZ}},
  note         = {Machine review of arXiv:2507.02945}
}
read the original abstract

While deep spiking neural networks (SNNs) demonstrate superior performance, their deployment on resource-constrained neuromorphic hardware still remains challenging. Network pruning offers a viable solution by reducing both parameters and synaptic operations (SynOps) to facilitate the edge deployment of SNNs, among which search-based pruning methods search for the SNNs structure after pruning. However, existing search-based methods fail to directly use SynOps as the constraint because it will dynamically change in the searching process, resulting in the final searched network violating the expected SynOps target. In this paper, we introduce a novel SNN pruning framework called SPEAR, which leverages reinforcement learning (RL) technique to directly use SynOps as the searching constraint. To avoid the violation of SynOps requirements, we first propose a SynOps prediction mechanism called LRE to accurately predict the final SynOps after search. Observing SynOps cannot be explicitly calculated and added to constrain the action in RL, we propose a novel reward called TAR to stabilize the searching. Extensive experiments show that our SPEAR framework can effectively compress SNN under specific SynOps constraint.

Figures

Figures reproduced from arXiv: 2507.02945 by the authors.

Figure 1
Figure 1. Characteristics of SynOps in finetuning and pruning [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of our SPEAR framework. where SynOpsj means the SynOps of the j-th sample, and N denotes number of samples in the dataset. For simplicity, the term SynOps mentioned in the following paper refers to the average SynOps, unless otherwise specified. 4 Methodology 4.1 Overview The overview of our Structured Pruning for SNNs via Synaptic Operation Estimation and Reinforce￾ment Learning (SPEAR) framework is shown … view at source ↗
Figure 3
Figure 3. Linear relationship of SynOps pre and post finetuning [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Results under different compression ratios with different constrains. is better). We observe that our linear LRE is better than nonlinear regression with smaller root mean square error and higher R2 . Analysis on different target constraints. To demonstrate the general…
Figure 6
Figure 6. Figure 6: SynOps Relative Error converges rapidly with sampling iteration (10 samples per iteration) [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: The pruning policy and SynOps distribution of each layer under different targets given by [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

78 extracted references · 57 canonical work pages

  1. [1]

    Spiking neural networks hardware implementations and challenges: A survey

    Maxence Bouvier, Alexandre Valentian, Thomas Mesquida, Francois Rummens, Marina Reyboz, Elisa Vianello, and Edith Beigne. Spiking neural networks hardware implementations and challenges: A survey. ACM Journal on Emerging Technologies in Computing Systems (JETC), 15(2):1–35, 2019

    work page 2019

  2. [2]

    Differentiable hierarchical and surrogate gradient search for spiking neural networks

    Kaiwei Che, Luziwei Leng, Kaixuan Zhang, Jianguo Zhang, Qinghu Meng, Jie Cheng, Qinghai Guo, and Jianxing Liao. Differentiable hierarchical and surrogate gradient search for spiking neural networks. Advances in Neural Information Processing Systems, 35:24975–24990, 2022

    work page 2022

  3. [3]

    Pruning of deep spiking neural networks through gradient rewiring

    Yanqi Chen, Zhaofei Yu, Wei Fang, Tiejun Huang, and Yonghong Tian. Pruning of deep spiking neural networks through gradient rewiring. In Zhi-Hua Zhou, editor, Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 1713–

    work page

  4. [4]

    State tran- sition of dendritic spines improves learning of sparse spiking neural networks

    Yanqi Chen, Zhaofei Yu, Wei Fang, Zhengyu Ma, Tiejun Huang, and Yonghong Tian. State tran- sition of dendritic spines improves learning of sparse spiking neural networks. In International Conference on Machine Learning, pages 3701–3715. PMLR, 2022

    work page 2022

  5. [5]

    A Unified Framework for Soft Threshold Pruning

    Yanqi Chen, Zhengyu Ma, Wei Fang, Xiawu Zheng, Zhaofei Yu, and Yonghong Tian. A unified framework for soft threshold pruning. arXiv preprint arXiv:2302.13019, 2023

    work page Pith review arXiv 2023

  6. [6]

    Spatio-temporal pruning and quantization for low-latency spiking neural networks

    Sayeed Shafayet Chowdhury, Isha Garg, and Kaushik Roy. Spatio-temporal pruning and quantization for low-latency spiking neural networks. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1–9. IEEE, 2021

    work page 2021

  7. [7]

    Loihi: A neuromorphic manycore processor with on-chip learning

    Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with on-chip learning. Ieee Micro, 38(1):82–99, 2018

    2018

  8. [8]

    Imagenet: A large- scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large- scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009

    2009

Show all 78 references
  1. [9]

    Temporal efficient training of spiking neural network via gradient re-weighting

    Shikuang Deng, Yuhang Li, Shanghang Zhang, and Shi Gu. Temporal efficient training of spiking neural network via gradient re-weighting. arXiv preprint arXiv:2202.11946, 2022

    2022 arXiv

  2. [10]

    Ec-snn: splitting deep spiking neural networks for edge devices

    Xin Du Di Yu, Linshan Jiang, Wentao Tong, and Shuiguang Deng. Ec-snn: splitting deep spiking neural networks for edge devices. In Proceedings of the Thirty-ThirdInternational Joint Conference on Artificial Intelligence (2024). https://api. semanticscholar. org/CorpusID, volume...

    2024

  3. [11]

    Deep residual learning in spiking neural networks

    Wei Fang, Zhaofei Yu, Yanqing Chen, Tiejun Huang, Timothée Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. In Neural Information Processing Systems, 2021

    2021

  4. [12]

    Spikingjelly: An open-source machine learning infrastructure platform for spike-based intelligence

    Wei Fang, Yanqi Chen, Jianhao Ding, Zhaofei Yu, Timothée Masquelier, Ding Chen, Liwei Huang, Huihui Zhou, Guoqi Li, and Yonghong Tian. Spikingjelly: An open-source machine learning infrastructure platform for spike-based intelligence. Science Advances, 9(40):eadi1480, 2023

    2023

  5. [13]

    Large-scale neuromorphic computing systems

    Steve Furber. Large-scale neuromorphic computing systems. Journal of neural engineering, 13 (5):051001, 2016. 10

    2016

  6. [14]

    Jointly training and pruning cnns via learnable agent guidance and alignment

    Alireza Ganjdanesh, Shangqian Gao, and Heng Huang. Jointly training and pruning cnns via learnable agent guidance and alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16058–16069, 2024

    2024

  7. [15]

    A low effort approach to structured cnn design using pca

    Isha Garg, Priyadarshini Panda, and Kaushik Roy. A low effort approach to structured cnn design using pca. IEEE Access, 8:1347–1360, 2019

    2019

  8. [16]

    A survey on efficient convolutional neural networks and hardware acceleration

    Deepak Ghimire, Dayoung Kil, and Seong-heum Kim. A survey on efficient convolutional neural networks and hardware acceleration. Electronics, 11(6):945, 2022

    2022

  9. [17]

    Model compression using progressive channel pruning

    Jinyang Guo, Weichen Zhang, Wanli Ouyang, and Dong Xu. Model compression using progressive channel pruning. IEEE Transactions on Circuits and Systems for Video Technology, 31(3):1114–1124, 2020

    2020

  10. [18]

    Im-loss: information maximization loss for spiking neural networks

    Yufei Guo, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Yinglei Wang, Xuhui Huang, and Zhe Ma. Im-loss: information maximization loss for spiking neural networks. Advances in Neural Information Processing Systems, 35:156–166, 2022

    2022

  11. [19]

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

    Yufei Guo, Xiaode Liu, Yuanpei Chen, Liwen Zhang, Weihang Peng, Yuhan Zhang, Xuhui Huang, and Zhe Ma. Rmp-loss: Regularizing membrane potential distribution for spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 17391–17...

    2023

  12. [20]

    Adaptive sparse structure development with pruning and regeneration for spiking neural networks

    Bing Han, Feifei Zhao, Wenxuan Pan, and Yi Zeng. Adaptive sparse structure development with pruning and regeneration for spiking neural networks. Information Sciences, 689:121481, 2025

    2025

  13. [21]

    A progressive training framework for spiking neural networks with learnable multi-hierarchical model

    Zecheng Hao, Xinyu Shi, Zihan Huang, Tong Bu, Zhaofei Yu, and Tiejun Huang. A progressive training framework for spiking neural networks with learnable multi-hierarchical model. In The Twelfth International Conference on Learning Representations, 2023

    2023

  14. [22]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. pages 770–778, 2016

    2016

  15. [23]

    Structured pruning for deep convolutional neural networks: A survey

    Yang He and Lingao Xiao. Structured pruning for deep convolutional neural networks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(5):2900–2919, 2024. doi: 10.1109/TPAMI.2023.3334614

    2024

  16. [24]

    Amc: Automl for model compression and acceleration on mobile devices

    Yihui He, Ji Lin, Zhijian Liu, Hanrui Wang, Li-Jia Li, and Song Han. Amc: Automl for model compression and acceleration on mobile devices. In Proceedings of the European conference on computer vision (ECCV), pages 784–800, 2018

    2018

  17. [25]

    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), pages 10–14,

    2014

  18. [26]

    Spiking deep residual networks

    Yang-Zhi Hu, Huajin Tang, and Gang Pan. Spiking deep residual networks. IEEE Transactions on Neural Networks and Learning Systems, 34:5200–5205, 2021

    2021

  19. [27]

    Clif: Complementary leaky integrate-and-fire neuron for spiking neural networks

    Yulong Huang, Xiaopeng Lin, Hongwei Ren, Haotian Fu, Yue Zhou, Zunchang Liu, Biao Pan, and Bojun Cheng. Clif: Complementary leaky integrate-and-fire neuron for spiking neural networks. arXiv preprint arXiv:2402.04663, 2024

    2024 arXiv

  20. [28]

    Neural architecture search for spiking neural networks

    Youngeun Kim, Yuhang Li, Hyoungseob Park, Yeshwanth Venkatesha, and Priyadarshini Panda. Neural architecture search for spiking neural networks. In European conference on computer vision, pages 36–56. Springer, 2022

    2022

  21. [29]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009

    2009

  22. [30]

    Tiny imagenet visual recognition challenge

    Yann Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015

    2015

  23. [31]

    Spike2former: Efficient spiking transformer for high-performance image segmentation

    Zhenxin Lei, Man Yao, Jiakui Hu, Xinhao Luo, Yanye Lu, Bo Xu, and Guoqi Li. Spike2former: Efficient spiking transformer for high-performance image segmentation. arXiv preprint arXiv:2412.14587, 2024. 11

    2024 arXiv

  24. [32]

    Pruning filters for efficient convnets, 2017

    Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets, 2017. URL https://arxiv.org/abs/1608.08710

    2017 arXiv

  25. [33]

    Cifar10-dvs: an event- stream dataset for object classification

    Hongmin Li, Hanchao Liu, Xiangyang Ji, Guoqi Li, and Luping Shi. Cifar10-dvs: an event- stream dataset for object classification. Frontiers in neuroscience, 11:309, 2017

    2017

  26. [34]

    Efficient structure slimming for spiking neural networks

    Yaxin Li, Xuanye Fang, Yuyuan Gao, Dongdong Zhou, Jiangrong Shen, Jian K Liu, Gang Pan, and Qi Xu. Efficient structure slimming for spiking neural networks. IEEE Transactions on Artificial Intelligence, 2024

    2024

  27. [35]

    Towards efficient deep spiking neural networks construction with spiking activity based pruning

    Yaxin Li, Qi Xu, Jiangrong Shen, Hongming Xu, Long Chen, and Gang Pan. Towards efficient deep spiking neural networks construction with spiking activity based pruning. arXiv preprint arXiv:2406.01072, 2024

    2024 arXiv

  28. [36]

    Continuous control with deep reinforcement learning

    TP Lillicrap. Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971, 2015

    2015 arXiv

  29. [37]

    Darts: Differentiable architecture search

    Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. arXiv preprint arXiv:1806.09055, 2018

    2018 arXiv

  30. [38]

    Lite-snn: Designing lightweight and efficient spiking neural network through spatial-temporal compressive network search and joint optimization

    Qianhui Liu, Jiaqi Yan, Malu Zhang, Gang Pan, and Haizhou Li. Lite-snn: Designing lightweight and efficient spiking neural network through spatial-temporal compressive network search and joint optimization. arXiv preprint arXiv:2401.14652, 2024

    2024 arXiv

  31. [39]

    A survey on evolutionary neural architecture search

    Yuqiao Liu, Yanan Sun, Bing Xue, Mengjie Zhang, Gary G Yen, and Kay Chen Tan. A survey on evolutionary neural architecture search. IEEE transactions on neural networks and learning systems, 34(2):550–570, 2021

    2021

  32. [40]

    Learning efficient convolutional networks through network slimming

    Zhuang Liu, Jianguo Li, Zhiqiang Shen, Gao Huang, Shoumeng Yan, and Changshui Zhang. Learning efficient convolutional networks through network slimming. In Proceedings of the IEEE international conference on computer vision, pages 2736–2744, 2017

    2017

  33. [41]

    Integer-valued training and spike-driven inference spiking neural network for high-performance and energy-efficient object detection

    Xinhao Luo, Man Yao, Yuhong Chou, Bo Xu, and Guoqi Li. Integer-valued training and spike-driven inference spiking neural network for high-performance and energy-efficient object detection. In European Conference on Computer Vision, pages 253–272. Springer, 2025

    2025

  34. [42]

    Autosnn: Towards energy-efficient spiking neural networks

    Byunggook Na, Jisoo Mok, Seongsik Park, Dongjin Lee, Hyeokjun Choe, and Sungroh Yoon. Autosnn: Towards energy-efficient spiking neural networks. In International Conference on Machine Learning, pages 16253–16269. PMLR, 2022

    2022

  35. [43]

    Towards artificial general intelligence with hybrid tianjic chip architecture

    Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenzhi Wu, Wei He, et al. Towards artificial general intelligence with hybrid tianjic chip architecture. Nature, 572(7767):106–111, 2019

    2019

  36. [44]

    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, 2024

    2024 arXiv

  37. [45]

    Gated attention coding for training high-performance and efficient spiking neural networks

    Xuerui Qiu, Rui-Jie Zhu, Yuhong Chou, Zhaorui Wang, Liang-jian Deng, and Guoqi Li. Gated attention coding for training high-performance and efficient spiking neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 601–610, 2024

    2024

  38. [46]

    On the distribution of the correlation coefficient in small samples

    Paul R Rider. On the distribution of the correlation coefficient in small samples. Biometrika, 24 (3/4):382–403, 1932

    1932

  39. [47]

    Wang, Chiao Liu, and Kaushik Roy

    Abhronil Sengupta, Yuting Ye, Robert Y . Wang, Chiao Liu, and Kaushik Roy. Going deeper in spiking neural networks: Vgg and residual architectures. Frontiers in Neuroscience, 13, 2018

    2018

  40. [48]

    Towards energy efficient spiking neural networks: An unstructured pruning framework

    Xinyu Shi, Jianhao Ding, Zecheng Hao, and Zhaofei Yu. Towards energy efficient spiking neural networks: An unstructured pruning framework. In The Twelfth International Conference on Learning Representations, 2024

    2024

  41. [49]

    Very deep convolutional networks for large-scale image recognition

    Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. 2015. 12

    2015

  42. [50]

    Snn-bert: Training-efficient spiking neural networks for energy-efficient bert

    Qiaoyi Su, Shijie Mei, Xingrun Xing, Man Yao, Jiajun Zhang, Bo Xu, and Guoqi Li. Snn-bert: Training-efficient spiking neural networks for energy-efficient bert. Neural Networks, 180: 106630, 2024

    2024

  43. [51]

    Rl-pruner: Structured pruning using reinforcement learning for cnn compression and acceleration

    Boyao Wang and V olodymyr Kindratenko. Rl-pruner: Structured pruning using reinforcement learning for cnn compression and acceleration. arXiv preprint arXiv:2411.06463, 2024

    2024 arXiv

  44. [52]

    Efficient spiking neural network design via neural architecture search

    Jiaqi Yan, Qianhui Liu, Malu Zhang, Lang Feng, De Ma, Haizhou Li, and Gang Pan. Efficient spiking neural network design via neural architecture search. Neural Networks, 173:106172, 2024

    2024

  45. [53]

    Temporal-wise attention spiking neural networks for event streams classification

    Man Yao, Huanhuan Gao, Guangshe Zhao, Dingheng Wang, Yihan Lin, Zhaoxu Yang, and Guoqi Li. Temporal-wise attention spiking neural networks for event streams classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10221– 10230, 2021

    2021

  46. [54]

    Attention spiking neural networks

    Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. Attention spiking neural networks. IEEE transactions on pattern analysis and machine intelligence, 45(8):9393–9410, 2023

    2023

  47. [55]

    Spike- driven transformer

    Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. Spike- driven transformer. Advances in neural information processing systems, 36, 2024

    2024

  48. [56]

    Scaling spike-driven transformer with efficient spike firing approximation training

    Man Yao, Xuerui Qiu, Tianxiang Hu, Jiakui Hu, Yuhong Chou, Keyu Tian, Jianxing Liao, Luziwei Leng, Bo Xu, and Guoqi Li. Scaling spike-driven transformer with efficient spike firing approximation training. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  49. [57]

    Glif: A unified gated leaky integrate-and- fire neuron for spiking neural networks

    Xingting Yao, Fanrong Li, Zitao Mo, and Jian Cheng. Glif: A unified gated leaky integrate-and- fire neuron for spiking neural networks. Advances in Neural Information Processing Systems, 35:32160–32171, 2022

    2022

  50. [58]

    Workload-balanced pruning for sparse spiking neural networks

    Ruokai Yin, Youngeun Kim, Yuhang Li, Abhishek Moitra, Nitin Satpute, Anna Hambitzer, and Priyadarshini Panda. Workload-balanced pruning for sparse spiking neural networks. IEEE Transactions on Emerging Topics in Computational Intelligence, 2024

    2024

  51. [59]

    Going deeper with directly-trained larger spiking neural networks

    Hanle Zheng, Yujie Wu, Lei Deng, Yifan Hu, and Guoqi Li. Going deeper with directly-trained larger spiking neural networks. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 11062–11070, 2021

    2021

  52. [60]

    -” indicates results are not reported in original paper

    Chenlin Zhou, Han Zhang, Liutao Yu, Yumin Ye, Zhaokun Zhou, Liwei Huang, Zhengyu Ma, Xiaopeng Fan, Huihui Zhou, and Yonghong Tian. Direct training high-performance deep spiking neural networks: A review of theories and methods. arXiv preprint arXiv:2405.04289, 2024. 13 A More ...

    2024 arXiv

  53. [63]

    It can be found in 4 for method and 5.2 for experiments results

    Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? Answer: [Yes] Justification: The abstract appropriately describes the research contributions. It can be found in 4 for method and 5.2 for experimen...

  54. [64]

    Limitations

    Limitations Question: Does the paper discuss the limitations of the work performed by the authors? Answer: [Yes] Justification: We have discussed our limitations in 5.3 and Appendix D. For examples, the hyper-params in TAR should be limited to a reasonable range. Guidelines: •...

  55. [65]

    Guidelines: • The answer NA means that the paper does not include theoretical results

    Theory assumptions and proofs Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? 17 Answer: [NA] Justification: Our article reports an empirical finding - there is a linear relationship between the SynO...

  56. [66]

    And we have also described our algorithm in 4 with details

    Experimental result reproducibility Question: Does the paper fully disclose all the information needed to reproduce the main ex- perimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and da...

  57. [67]

    Additionally, we have uploaded our code

    Open access to data and code Question: Does the paper provide open access to the data and code, with sufficient instruc- tions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: We use publicly available d...

  58. [68]

    Guidelines: • The answer NA means that the paper does not include experiments

    Experimental setting/details Question: Does the paper specify all the training and test details (e.g., data splits, hyper- parameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? Answer: [Yes] Justification: We have thoroughly reported o...

  59. [69]

    4, both of which demonstrated significant statistical relevance/significance

    Experiment statistical significance Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? Answer: [Yes] Justification: We computed both the Mean Squared Error (MSE) and ...

  60. [70]

    Guidelines: • The answer NA means that the paper does not include experiments

    Experiments compute resources Question: For each experiment, does the paper provide sufficient information on the com- puter resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: We have reported it in ...

  61. [71]

    Code of ethics Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? Answer: [Yes] Justification: We sincerely confirm that the research conducted in our paper fully conforms to...

  62. [72]

    This work promises to bring significant advantages to Edge AI as introduced in Sec 1

    Broader impacts Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? Answer: [Yes] 20 Justification: Our research focus on the compression of Spiking Neural Networks (SNNs), aiming to realize efficient i...

  63. [73]

    Guidelines: • The answer NA means that the paper poses no such risks

    Safeguards Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? Answer: [NA] Justification: Our work focuse...

  64. [74]

    Licenses for existing assets Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? Answer: [Yes] Justification: In the paper, al...

  65. [75]

    Guidelines: • The answer NA means that the paper does not release new assets

    New assets Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? Answer: [Yes] Justification: Yes, any new assets introduced in the paper are well documented. Guidelines: • The answer NA means that the paper do...

  66. [76]

    Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects

    Crowdsourcing and research with human subjects Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? A...

  67. [77]

    Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals...

  68. [78]

    Answer: [NA] Justification: The core method development in this research does not involve LLMs as any important, original, or non-standard components

    Declaration of LLM usage Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the ...

    2025

  69. [1721]

    doi: 10.24963/ijcai.2021/236

    International Joint Conferences on Artificial Intelligence Organization, 8 2021. doi: 10.24963/ijcai.2021/236. Main Track

    2021 doi

  70. [2014]

    doi: 10.1109/ISSCC.2014.6757323

    2014

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.