Pith. sign in

REVIEW 4 major objections 5 minor 44 references

SDSNN: A Single-Timestep Spiking Neural Network with Self-Dropping Neuron and Bayesian Optimization

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 A single-timestep spiking network with a Self-Dropping neuron matches multi-timestep LIF accuracy on three benchmarks while using 21–56% less energy.

desk verdict A useful neuron variant and a three-stage search idea, but the 'single-timestep' claim is contradicted by the paper's own equations and the reported numbers are selected on the test set. read the letter →

arxiv 2508.10913 v1 pith:E2PY7TES submitted 2025-08-01 cs.NE cs.AI

Changqing Xu , Buxuan Song , Yi Liu , Xinfang Liao , Wenbin Zheng , Yintang Yang This is my paper
classification cs.NEcs.AI
keywords spikingneuralnetworkssingle-timestepinferenceSelf-DroppingNeuronBayesianoptimizationsurrogategradientlow-latencySNNenergy-efficient
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

Spiking neural networks usually trade one or more temporal simulation steps per layer for accuracy, and those steps multiply both latency and energy. This paper claims that a single timestep per layer is enough if the neuron's spike rule is changed: a Self-Dropping neuron fires only when its membrane potential has passed the threshold and starts to fall, and the threshold is lowered over timesteps to avoid dead neurons. A three-stage Bayesian optimization search chooses the per-layer timestep configuration, and training backpropagates only through the final timestep's surrogate gradient instead of through time. On Fashion-MNIST, CIFAR-10, and CIFAR-100 the reported results are 93.72%, 92.20%, and 69.45% accuracy with a single inference timestep, matching or exceeding multi-timestep LIF baselines while cutting energy by 56%, 21%, and 22%. The paper aims to make single-timestep SNNs a practical, low-energy alternative for edge inference.

What carries the argument

The Self-Dropping neuron is the load-bearing component. Its update is $u^{(t,n+1)} = \gamma u^{(t-1,n+1)} - V_{th}o^{(t-1,n+1)} + x^{(t,n)}$ with soft reset, and it spikes at time $t$ only if $u^{(t,n+1)} > V_{th}$ and $u^{(t,n+1)} < u^{(t-1,n+1)}$ (Eq. 7). The threshold follows $V_{th}=V_{th0}/t$ (Eq. 8), and the spike output is the integer ratio of potential to threshold, capped by a firing limit. The surrogate gradient in Eq. (10) lets a small band around the threshold pass gradients, and gradients are computed only for the final timestep, eliminating backpropagation through time. The same machinery includes the three-stage Bayesian optimization over layer-wise timestep combinations, which finds single-timestep inference modes.

What would settle it

Retrain the same networks on the same datasets with the Self-Dropping firing condition (Eq. 7) replaced by ordinary IF firing (spike whenever $u>V_{th}$), keeping the threshold schedule, the Bayesian timestep search, and the one-timestep surrogate gradient. If the IF variant matches or exceeds 93.72/92.20/69.45, the self-dropping mechanism is not the cause of the reported accuracies.

Watch

Extended reading notes

Core claim

The central claim is that the information lost by dropping to one timestep can be recovered at the neuron level, not by deeper temporal computation. The Self-Dropping (SD) neuron emits a spike only when the membrane potential $u$ exceeds the threshold $V_{th}$ and is below its previous value, i.e., when the accumulation has peaked and starts to decay; the spike amplitude is $\lfloor u/V_{th}\rfloor$ clipped by a maximum firing limit. The threshold is annealed as $V_{th}=V_{th0}/t$ to keep later timesteps active. Training uses the surrogate gradient of Eq. (10) only at the last timestep, so no temporal gradient chain is computed; the first timestep falls back to IF-style firing because there is no previous potential to compare. With layer-wise timesteps selected by a three-stage Bayesian search, the network classifies Fashion-MNIST at 93.72%, CIFAR-10 at 92.20%, and CIFAR-100 at 69.45% using one inference timestep, with reported energy reductions of 56%, 21%, and 22% against multi-timestep LIF networks.

Load-bearing premise

The load-bearing premise is that the hand-designed Self-Dropping rule—spike only when the potential exceeds the threshold and is lower than the previous timestep—together with the one-timestep surrogate gradient, trains networks reliably; the paper reports no ablation or gradient analysis showing this rule is what produces the accuracy.

Editorial extensions

If this is right

  • Inference in a single timestep per layer removes the temporal loop, so latency is close to that of an ANN while the spikes keep event-driven, addition-only compute.
  • Training memory drops from $O(T)$ to $O(1)$ in the temporal dimension; the paper reports 0.91 GB peak memory on CIFAR-10 versus 3.00 GB for BPTT.
  • Layer-wise Bayesian-selected timesteps beat any single fixed timestep (on CIFAR-10, 92.20% versus a best fixed-step 90.45%).
  • The threshold-annealing schedule and spike-suppression rule together keep firing rates comparable to LIF (e.g., 13.66% vs 13.56% on CIFAR-10), so the energy saving comes from removing timesteps rather than from sparser firing.
  • Reported energy savings of 56%, 21%, and 22% on the three benchmarks follow directly if the analytical energy model is accurate.

Reading between the lines

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

  • A transferable hypothesis the paper does not test: the falling-potential trigger may act as a temporal-derivative detector, so the same single-timestep recipe could transfer to other neurons that encode local change rather than absolute level.
  • The threshold schedule $V_{th0}/t$ is a fixed annealing curve; a natural extension is to make the per-layer thresholds learnable or Bayesian-searched, which could improve the accuracy/energy frontier further.
  • The energy numbers are analytical estimates; measuring actual MAC/spike counts on neuromorphic hardware would test whether the 21–56% savings survive implementation overheads.
  • A cheap ablation would settle the mechanism's contribution: replace the SD rule with standard IF firing while keeping the Bayesian timestep search and one-timestep surrogate gradient; if accuracy stays, the self-dropping condition is not the source of the gains.
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 proposes SDSNN, a spiking neural network built around a new 'Self-Dropping' (SD) neuron whose firing condition depends on a decrease of membrane potential, a time-decreasing threshold Vth0/t, a maximum firing limit, and a surrogate gradient truncated to a single backward time step. A three-stage Bayesian optimization procedure selects per-layer time-step vectors T, and the authors report accuracy and energy-consumption gains on Fashion-MNIST, CIFAR-10, and CIFAR-100, claiming classification 'using only single-timestep spikes' with energy reductions of 56%, 21%, and 22%.

Significance. If the claims were correct, the work would address a real bottleneck in SNN deployment: latency and energy from multi-timestep forward computation. The paper also compares against single-timestep SNN baselines and reports a useful energy-estimation methodology. However, the central claim that inference uses single-timestep spikes is contradicted by the paper's own equations and reported configurations, and the reported accuracies are obtained by optimizing on the test set. These issues are load-bearing rather than cosmetic, so the significance as stated cannot be credited.

major comments (4)
  1. [Abstract; Methodology, Eqs. (6)-(8); Overall Framework; Table 3] The assertion 'using only single-timestep spikes' is not supported by the mechanism. Eq. (6) is a recurrence over t, Eq. (7) fires a spike only when u(t,n+1) is below the previous membrane potential, and the footnote to Table 4 concedes that the first time step uses the IF rule because no previous value exists. The Overall Framework states that the input 'circulates for tl time steps, and the information from the tl-th time step is subsequently propagated.' Table 3 lists per-layer vectors such as [1,3,1,1], [1,2,2,2,2,1], and [1,2,2,1,2,1], meaning the forward pass performs multiple internal updates per layer. Thus the reported accuracies and energy reductions are not for single-timestep inference; the only single-step aspect is the backward gradient truncation. This invalidates the headline comparison against multi-timestep LIF networks.
  2. [Methodology, Three-Stage Time Step Bayesian Search Method; Table 3; Table 4] The experimental protocol selects time-step configurations using test accuracy. Stage 2 'evaluates the objective function (test accuracy)' and Stage 3 'optimizes recognition accuracy in the final target epoch,' and the same test accuracies are then reported in Tables 3 and 4 as the method's results. This is a form of test-time model selection, so the final numbers are not unbiased estimates of generalization performance. The energy-consumption numbers are also computed only for these selected configurations, compounding the selection effect. A nested or validation-based selection procedure is required.
  3. [Performance Analysis; Table 3] The energy comparisons are not apples-to-apples. The paper states that 'our single-timestep network involves only the convolution calculation of a single layer of spike information,' but Table 3 shows per-layer time-step vectors whose total number of internal updates is at least 6, 10, and 9 for the three datasets. The energy figures compare SD against LIF with a fixed time step per layer, but the SD numbers must count all t_l recurrent updates in each spiking layer, not just one forward step per layer. Without a per-layer and per-timestep energy model that reflects the reported T vectors, the claimed 56%, 21%, and 22% reductions are not credible.
  4. [Self-Dropping Neuron Model Design, Eqs. (7)-(8); Table 3, footnote] The SD neuron's behavior depends on several hand-designed and unvalidated choices: the threshold-lowering rule Vth = Vth0/t (Eq. 8), the maximum firing limit, and the use of IF firing at the first time step. No ablation or gradient analysis shows that these choices, rather than the network architecture or the Bayesian search, are responsible for the reported accuracies. Given that the first-step IF patch is already an acknowledged exception to the SD rule, the paper should at least verify that the SD rule itself is what drives performance.
minor comments (5)
  1. [Related Work, Eqs. (1)-(2)] Eq. (1) and Eq. (2) are identical; one of them appears intended to define the spike output s[t] via the Heaviside function rather than repeat the membrane update.
  2. [Table 2] The text describes three training stages, but Table 2 lists only Epoch s1 and Epoch s2; the value for Epoch s3 (the final target epoch) is missing.
  3. [Table 6] The memory comparison uses ResNet-18, but Table 1 defines a custom convolutional architecture; the discrepancy should be explained or the same architecture should be used throughout.
  4. [Table 1 footnote] 'V oting' should be 'Voting' in the table footnote.
  5. [Performance Analysis] The sentence 'our single-timestep network involves only the convolution calculation of a single layer of spike information' is vague and appears to conflict with the per-layer time-step vectors in Table 3; the wording should be clarified.

Circularity Check

2 steps flagged · score 7.0 of 10

Reported accuracies are the optimized test-accuracy objective from Bayesian search, and 'single-timestep' is defined as one-step backprop while forward inference runs multiple steps per layer.

  1. fitted input called prediction [Methodology, 'Three-Stage Time Step Bayesian Search Method' (Stages 2-3); Results, Table 3]
    "The objective function f (X) that we attempt to maximize is the accuracy of spiking neural networks. ... In each iteration, Bayesian optimization selects the next evaluation point based on the current surrogate model, evaluates the objective function (test accuracy), and updates the surrogate model. ... Table 3 presents a comparison of accuracy between the proposed method and other methods on the FashionMNIST, CIFAR-10, and CIFAR-100 datasets."

    The time-step vector T is the hyperparameter being searched, and its objective is the network's test accuracy. Stages 2 and 3 evaluate T by computing test accuracy after training, and the same test accuracy is then reported as the final result in Table 3. The reported accuracies (93.72%, 92.20%, 69.45%) are therefore the maximum values of the very objective used to choose T; they are selected values, not independent predictions. Comparisons against fixed-T baselines are also made after selecting T on the same test set, so the apparent advantage is forced by the selection procedure rather than by an out-of-sample evaluation. The energy savings are computed only at these test-selected T values and inherit the same selection dependence.

  2. self definitional [Abstract; Methodology 'Self-Dropping Neuron Model Design', Backpropagation item; Eq. (6)-(8); 'Overall Framework'; Table 4 footnote]
    "We employ a single-time-step surrogate gradient method, where gradients are calculated only on the spikes in the final time step of the forward propagation. ... For the l-th spiking layer, the input circulates for tl time steps, and the information from the tl-th time step is subsequently propagated backward. ... It is worth noting that the spike firing mechanism of the SD neuron relies on comparing the membrane potential value from the previous time step."

    The abstract and contributions call the network 'single-timestep' and claim the results use 'only single-timestep spikes,' but the method defines 'single-timestep' operationally as a surrogate-gradient computation restricted to the last forward step. In the forward pass, Eq. (6) is a recurrence over t and Eq. (7) fires only when u(t) < u(t-1), so multiple time steps are required by construction; the framework states each layer circulates input for t_l steps, and Table 3 lists vectors such as [1,2,1,4], [1,2,2,2,2,1], and [1,2,2,1,2,1]. The first time step cannot use the SD rule at all because no previous potential exists, so the IF rule is substituted.

full rationale

The central empirical claim is circular in the fitted-input sense: the time-step vector T is chosen by Bayesian optimization whose objective function is the network's test accuracy, and the same test accuracy is then reported as the method's headline result. Stage 2 and Stage 3 both 'evaluate the objective function (test accuracy)' during the search; the numbers in Table 3 are the best values of that search, not independent test-set evaluations. The 'single-timestep' claim is also self-definitional: the method section defines 'single-timestep' as computing a surrogate gradient at only the last forward time step, while Eq. (6)-(7) require a recurrent membrane-potential comparison and the framework has each layer run t_l steps; Table 3 and the Table 4 footnote confirm multi-step forward operation and an IF fallback at the first step. The paper's self-citations (Xu et al. 2022/2023/2024, Pei et al. 2023) are contextual and are not used to justify the core mechanism, and no uniqueness theorem is imported from them, so those do not add circularity. The neuron equations and surrogate gradient are otherwise specified in the text; the circularity is thus concentrated in the evaluation protocol and the labeling of 'single-timestep,' which together materially construct the headline claims.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The central claim rests on several hand-set parameters (tau, Vth0, surrogate width, max firing limit) and on the BO-selected time-step vectors, which are fitted to test accuracy. The neuron rule itself is an ad hoc construction with an unanalyzed first-step exception. No independent evidence is given for the SD neuron outside the reported in-paper results.

free parameters (5)
  • Membrane time constant tau = 0.25
    Manually set in Table 2; controls potential decay in Eq. (6) and directly affects spike timing.
  • Initial threshold Vth0 = 0.5
    Manually set in Table 2; Eq. (8) divides it by t, so all later thresholds and hence spike rates depend on this value.
  • Surrogate gradient width a = 1
    In Eq. (10), gradient mask width; set by hand and not ablated.
  • Per-layer time-step vector T = Fashion-MNIST [1,3,1,1]; CIFAR-10 [1,2,2,2,2,1]; CIFAR-100 [1,2,2,1,2,1] (Stage 2)
    Chosen by Bayesian optimization using test accuracy as the objective; these are fitted model-selection parameters.
  • Maximum firing limit = Not stated
    Mentioned in Activation Phase and Eq. (7) comparison, but the value is never given.
assumptions (4)
  • domain assumption The surrogate gradient h(u) in Eq. (10) with width 1 provides a useful training signal for the non-differentiable SD spike rule.
    The paper relies on this in the Backpropagation step, but gives no gradient analysis or ablation.
  • ad hoc to paper Lowering the threshold as Vth = Vth0 / t prevents dead neurons.
    Stated in Methodology before Eq. (8) without evidence; if false, late time steps produce no spikes and the BO search is meaningless.
  • ad hoc to paper The first time step of the SD neuron can be treated with IF firing because no previous membrane potential exists.
    Noted only in the Table 4 footnote; the discontinuity between IF and SD behavior is unanalyzed.
  • domain assumption Max pooling after spiking layers preserves spike information and improves temporal tasks.
    Cited to Fang et al. 2021; used in the architecture but not central to the SD neuron claim.
invented entities (1)
  • Self-Dropping neuron
    purpose: Spike generation triggered when membrane potential exceeds threshold and then begins to decay; intended to increase information per spike.
    The neuron rule is introduced and evaluated only inside this paper's own experiments; no external benchmark or mechanistic prediction supports it beyond the reported accuracies.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SDSNN: A Single-Timestep Spiking Neural Network with Self-Dropping Neuron and Bayesian Optimization." pith.science (2026). https://pith.science/paper/E2PY7TES

@misc{pith2026250810913,
  author       = {Pith},
  title        = {Pith review of: SDSNN: A Single-Timestep Spiking Neural Network with Self-Dropping Neuron and Bayesian Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E2PY7TES}},
  note         = {Machine review of arXiv:2508.10913}
}
read the original abstract

Spiking Neural Networks (SNNs), as an emerging biologically inspired computational model, demonstrate significant energy efficiency advantages due to their event-driven information processing mechanism. Compared to traditional Artificial Neural Networks (ANNs), SNNs transmit information through discrete spike signals, which substantially reduces computational energy consumption through their sparse encoding approach. However, the multi-timestep computation model significantly increases inference latency and energy, limiting the applicability of SNNs in edge computing scenarios. We propose a single-timestep SNN, which enhances accuracy and reduces computational energy consumption in a single timestep by optimizing spike generation and temporal parameters. We design a Self-Dropping Neuron mechanism, which enhances information-carrying capacity through dynamic threshold adjustment and selective spike suppression. Furthermore, we employ Bayesian optimization to globally search for time parameters and obtain an efficient inference mode with a single time step. Experimental results on the Fashion-MNIST, CIFAR-10, and CIFAR-100 datasets demonstrate that, compared to traditional multi-timestep SNNs employing the Leaky Integrate-and-Fire (LIF) model, our method achieves classification accuracies of 93.72%, 92.20%, and 69.45%, respectively, using only single-timestep spikes, while maintaining comparable or even superior accuracy. Additionally, it reduces energy consumption by 56%, 21%, and 22%, respectively.

Figures

Figures reproduced from arXiv: 2508.10913 by the authors.

Figure 1
Figure 1. Schematic diagram of spike generation in the Self [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Three-Stage Time Step Bayesian Optimization [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Schematic diagram of the overall network frame [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 34 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Bergstra, J.; and Bengio, Y. 2012. Random search for hyper-parameter optimization. The journal of machine learning research, 13(1): 281--305

    work page 2012

  4. [4]

    Bu, T.; Fang, W.; Ding, J.; Dai, P.; Yu, Z.; and Huang, T. 2023. Optimal ANN-SNN conversion for high-accuracy and ultra-low-latency spiking neural networks. arXiv preprint arXiv:2303.04347

    arXiv 2023

  5. [5]

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

    work page 2015

  6. [6]

    bnn-bn=?

    Chen, T.; Zhang, Z.; Ouyang, X.; Liu, Z.; Shen, Z.; and Wang, Z. 2021. " bnn-bn=?": Training binary neural networks without batch normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 4619--4629

    work page 2021

  7. [7]

    Cheng, X.; Hao, Y.; Xu, J.; and Xu, B. 2020. LISNN: Improving spiking neural networks with lateral interactions for robust object recognition. In IJCAI, 1519--1525. Yokohama

    work page 2020

  8. [8]

    S.; Rathi, N.; and Roy, K

    Chowdhury, S. S.; Rathi, N.; and Roy, K. 2021. One timestep is all you need: Training spiking neural networks with ultra low latency. arXiv preprint arXiv:2110.05929

    arXiv 2021

Show all 44 references
  1. [9]

    Deng, S.; and Gu, S. 2021. Optimal conversion of conventional artificial neural networks to spiking neural networks. arXiv preprint arXiv:2103.00476

    2021 arXiv

  2. [10]

    Fang, W.; Yu, Z.; Chen, Y.; Masquelier, T.; Huang, T.; and Tian, Y. 2021. Incorporating Learnable Membrane Time Constant To Enhance Learning of Spiking Neural Networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2661--2671

    2021

  3. [11]

    Han, B.; Srinivasan, G.; and Roy, K. 2020. Rmp-snn: Residual membrane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 13558--13567

    2020

  4. [12]

    Hu, Y.; Deng, L.; Wu, Y.; Yao, M.; and Li, G. 2024. Advancing spiking neural networks toward deep residual learning. IEEE transactions on neural networks and learning systems, 36(2): 2353--2367

    2024

  5. [13]

    Huang, Y.; Lin, X.; Ren, H.; Fu, H.; Zhou, Y.; Liu, Z.; Pan, B.; and Cheng, B. 2024. CLIF: Complementary leaky integrate-and-fire Neuron for spiking neural networks. arXiv preprint arXiv:2402.04663

    2024 arXiv

  6. [14]

    Kim, Y.; Li, Y.; Park, H.; Venkatesha, Y.; Hambitzer, A.; and Panda, P. 2023. Exploring temporal information dynamics in spiking neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, 8308--8316

    2023

  7. [15]

    Krizhevsky, A.; and Hinton, G. 2009. Learning multiple layers of features from tiny images. Technical report, University of Toronto

    2009

  8. [16]

    Lemaire, E.; Cordone, L.; Castagnetti, A.; Novac, P.-E.; Courtois, J.; and Miramond, B. 2023. An Analytical Estimation of Spiking Neural Networks Energy Efficiency. In Tanveer, M.; Agarwal, S.; Ozawa, S.; Ekbal, A.; and Jatowt, A., eds., Neural Information Processing, 574--587...

    2023

  9. [17]

    Li, Y.; and Zeng, Y. 2022. Efficient and accurate conversion of spiking neural network with burst spikes. arXiv preprint arXiv:2204.13271

    2022 arXiv

  10. [18]

    Lian, S.; Shen, J.; Wang, Z.; and Tang, H. 2024. IM-LIF: Improved neuronal dynamics with attention mechanism for direct training deep spiking neural network. IEEE Transactions on Emerging Topics in Computational Intelligence, 8(2): 2075--2085

    2024

  11. [19]

    Liu, F.; Zhao, W.; Chen, Y.; Wang, Z.; and Jiang, L. 2022. Spikeconverter: An efficient conversion framework zipping the gap between artificial neural networks and spiking neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, 1692--1701

    2022

  12. [20]

    Meng, Q.; Xiao, M.; Yan, S.; Wang, Y.; Lin, Z.; and Luo, Z.-Q. 2022. Training high-performance low-latency spiking neural networks by differentiation on spike representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 12444--12453

    2022

  13. [21]

    Meng, Q.; Xiao, M.; Yan, S.; Wang, Y.; Lin, Z.; and Luo, Z.-Q. 2023. Towards memory-and time-efficient backpropagation for training spiking neural networks. In Proceedings of the IEEE/CVF international conference on computer vision, 6166--6176

    2023

  14. [22]

    Mo c kus, J. 1975. On Bayesian methods for seeking the extremum. In Optimization Techniques IFIP Technical Conference Novosibirsk, July 1--7, 1974 6, 400--404. Springer

    1975

  15. [23]

    Paszke, A.; Gross, S.; Massa, F.; Lerer, A.; Bradbury, J.; Chanan, G.; Killeen, T.; Lin, Z.; Gimelshein, N.; Antiga, L.; et al. 2019. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32

    2019

  16. [24]

    Pei, Y.; Xu, C.; Wu, Z.; Liu, Y.; and Yang, Y. 2023. ALBSNN: ultra-low latency adaptive local binary spiking neural network with accuracy loss estimator. Frontiers in Neuroscience, 17: 1225871

    2023

  17. [25]

    Putra, R. V. W.; and Shafique, M. 2023. TopSpark: a timestep optimization methodology for energy-efficient spiking neural networks on autonomous mobile agents. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 3561--3567. IEEE

    2023

  18. [26]

    Rathi, N.; and Roy, K. 2021. Diet-snn: A low-latency spiking neural network with direct input encoding and leakage and threshold optimization. IEEE Transactions on Neural Networks and Learning Systems, 34(6): 3174--3182

    2021

  19. [27]

    Su, Q.; Chou, Y.; Hu, Y.; Li, J.; Mei, S.; Zhang, Z.; and Li, G. 2023. Deep directly-trained spiking neural networks for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 6555--6565

    2023

  20. [28]

    Suetake, K.; Ikegawa, S.-i.; Saiin, R.; and Sawada, Y. 2023. S3NN: Time step reduction of spiking surrogate gradients for training energy efficient single-step spiking neural networks. Neural Networks, 159: 208--219

    2023

  21. [29]

    Tang, K.; Yan, Z.; and Wong, W.-F. 2024. OneSpike: Ultra-low latency spiking neural networks. In 2024 International Joint Conference on Neural Networks (IJCNN), 1--8. IEEE

    2024

  22. [30]

    Teeter, C.; Iyer, R.; Menon, V.; Gouwens, N.; Feng, D.; Berg, J.; Szafer, A.; Cain, N.; Zeng, H.; Hawrylycz, M.; et al. 2018. Generalized leaky integrate-and-fire models classify multiple neuron types. Nature communications, 9(1): 709

    2018

  23. [31]

    Wang, Y.; Zhang, M.; Chen, Y.; and Qu, H. 2022. Signed Neuron with Memory: Towards Simple, Accurate and High-Efficient ANN-SNN Conversion. In IJCAI, 2501--2508

    2022

  24. [32]

    Werbos, P. J. 2002. Backpropagation through time: what it does and how to do it. Proceedings of the IEEE, 78(10): 1550--1560

    2002

  25. [33]

    Wu, Y.; Deng, L.; Li, G.; Zhu, J.; Xie, Y.; and Shi, L. 2019. Direct training for spiking neural networks: Faster, larger, better. In Proceedings of the AAAI conference on artificial intelligence, volume 33, 1311--1318

    2019

  26. [34]

    Xiao, H.; Rasul, K.; and Vollgraf, R. 2017. Fashion-MNIST: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747

    2017 arXiv

  27. [35]

    Xu, C.; Liu, Y.; Chen, D.; and Yang, Y. 2022. Direct training via backpropagation for ultra-low-latency spiking neural networks with multi-threshold. Symmetry, 14(9): 1933

    2022

  28. [36]

    Xu, C.; Liu, Y.; and Yang, Y. 2023. Ultra-low latency spiking neural networks with spatio-temporal compression and synaptic convolutional block. Neurocomputing, 550: 126485

    2023

  29. [37]

    Xu, C.; Liu, Y.; and Yang, Y. 2024. STCSNN: High energy efficiency spike-train level spiking neural networks with spatio-temporal conversion. Neurocomputing, 607: 128364

    2024

  30. [38]

    Xu, C.; Zhang, W.; Liu, Y.; and Li, P. 2020. Boosting throughput and efficiency of hardware spiking neural accelerators using time compression supporting multiple spike codes. Frontiers in neuroscience, 14: 104

    2020

  31. [39]

    K.; Tang, H.; and Pan, G

    Xu, Q.; Li, Y.; Shen, J.; Liu, J. K.; Tang, H.; and Pan, G. 2023. Constructing deep spiking neural networks from artificial neural networks with knowledge distillation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 7886--7895

    2023

  32. [40]

    R.; Lei, T.; and Linares-Barranco, B

    Yang, S.; Gao, T.; Wang, J.; Deng, B.; Azghadi, M. R.; Lei, T.; and Linares-Barranco, B. 2022. SAM: a unified self-adaptive multicompartmental spiking neuron model for learning with working memory. Frontiers in neuroscience, 16: 850945

    2022

  33. [41]

    Yao, X.; Li, F.; Mo, Z.; and Cheng, J. 2022. Glif: A unified gated leaky integrate-and-fire neuron for spiking neural networks. Advances in Neural Information Processing Systems, 35: 32160--32171

    2022

  34. [42]

    Zhang, S.; Yang, Q.; Ma, C.; Wu, J.; Li, H.; and Tan, K. C. 2024. Tc-lif: A two-compartment spiking neuron model for long-term sequential modelling. In Proceedings of the AAAI conference on artificial intelligence, volume 38, 16838--16847

    2024

  35. [43]

    Zhang, W.; and Li, P. 2020. Temporal Spike Sequence Learning via Backpropagation for Deep Spiking Neural Networks. In Larochelle, H.; Ranzato, M.; Hadsell, R.; Balcan, M.; and Lin, H., eds., Advances in Neural Information Processing Systems, volume 33, 12022--12033. Curran Ass...

    2020

  36. [44]

    Zhong, X.; and Pan, H. 2022. A spike neural network model for lateral suppression of spike-timing-dependent plasticity with adaptive threshold. Applied Sciences, 12(12): 5980

    2022

Pith tools

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