Pith. sign in

REVIEW 3 major objections 6 minor 45 references

Optimization of Low-Latency Spiking Neural Networks Utilizing Historical Dynamics of Refractory Periods

T0 review · 3 major / 6 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 This paper claims that a refractory period adjusted dynamically from membrane-potential derivative and spiking history makes low-latency LIF SNNs more accurate, sparser, and more noise-robust than vanilla SNNs or equivalent ANNs.

desk verdict The HDRP neuron idea is promising and the experiments are broad, but the core normalization derivation has a sign reversal that, as printed, inverts the refractory mechanism. read the letter →

arxiv 2507.02960 v1 pith:RNJ73WAP submitted 2025-06-30 cs.NE cs.AI

Liying Tao , Zonglin Yang , Delong Shang This is my paper · ORCID
classification cs.NEcs.AI
keywords historicaldynamicrefractoryperiodspikingneuralnetworkslow-latencyinferenceLIFneuronmembranepotentialderivativenoiserobustnessneuromorphicdatasetssurrogategradienttraining
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 tries to show that the refractory period, the brief suppression of firing after a spike, can be made into a powerful and cheap mechanism for low-latency spiking neural networks. It proposes the Historical Dynamic Refractory Period (HDRP) neuron, which estimates how long a neuron should stay suppressed from the current derivative of its membrane potential and from how suppressed it was at past time steps. With this per-neuron, per-time-step refractory period and a threshold-dependent inhibition kernel, the neuron replaces the standard LIF neuron without breaking the binary spike interface. The reported results are that HDRP-SNNs surpass or match state-of-the-art SNNs on CIFAR-10, CIFAR-100, ImageNet, and CIFAR10-DVS while firing less, and that they withstand added Gaussian noise better than both standard SNNs and ANNs with the same architecture. If true, low-latency SNNs can be improved simply by swapping in a smarter neuron model, with no change to the training pipeline or hardware assumptions.

What carries the argument

The mechanism is the HDRP-LIF neuron: a LIF neuron whose refractory period is a state variable updated each time step, plus a refractory kernel. The update computes the membrane-potential derivative U'(l,t) from the input current and previous potential, normalizes it by its theoretical min/max (which the paper derives from weight bounds) to get k(U'), and combines it with the previous refractory period and a spike indicator. The refractory kernel g(tau_ref) = U_th tanh(A tau_ref) with learnable A subtracts inhibition from the synaptic current, effectively raising the effective threshold during the refractory period. These pieces retain binary spike transmission and are trained end-to-end with spatio-temporal backpropagation.

What would settle it

Run a trained HDRP-SNN on a large test set and record the empirical minimum and maximum of the membrane potential derivative across timesteps. If either falls outside the bounds predicted by Equations (9)-(10), or if the ordering of the two bounds is reversed as the printed derivation suggests, then the normalization k(U') is input-dependent and the central refractory update mechanism is not what is described.

Watch

Extended reading notes

Core claim

The central claim is that over-activation, not information loss, is what limits low-latency SNNs, and that a refractory period derived from local history fixes it. The HDRP model updates the refractory period as tau_ref = k(U')(tau_ref_past + S tau_refL) + S tau_ref0, where k(U') is a Min-Max normalized membrane-potential derivative, S is the spike indicator, and the previous refractory period implicitly encodes the neuron's spike history. A threshold-dependent refractory kernel g(tau_ref) = U_th tanh(A tau_ref) with learnable A subtracts inhibition from the synaptic current, suppressing noise spikes while preserving strong inputs. With surrogate gradient training via spatio-temporal backpropagation, this drop-in LIF replacement yields the reported accuracies on static and neuromorphic datasets.

Load-bearing premise

The method's normalization step assumes that the membrane potential derivative has fixed min/max bounds depending only on the network weights, so the refractory period update cannot be distorted by unusual inputs.

Editorial extensions

If this is right

  • At T=6, HDRP-SNN reaches 96.70% on CIFAR-10 and 80.91% on CIFAR-100 with ResNet-18, exceeding the same-architecture ANN by 0.21% and 0.38%.
  • On CIFAR10-DVS, HDRP-SNN sets 87.00% accuracy, 1.1% above the previous best, and on ImageNet it reaches 70.45%, nearly matching the 70.54% state of the art.
  • Across CIFAR-10, CIFAR-100, and ImageNet, HDRP-LIF reduces AC operations by roughly 15-18% relative to LIF-SNN while keeping energy within 1.2x of LIF and far below ANN.
  • Under Gaussian noise on CIFAR-10, HDRP-SNN keeps higher accuracy than LIF-SNN and ANN at every tested noise level, with improvements up to 2.89% over SNN and 8.29% over ANN.
  • Ablations show that absolute or fixed-decay refractory periods degrade accuracy, while the HDRP's adaptive decay plus historical state improves it, supporting the two-component mechanism.

Reading between the lines

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

  • Because the refractory update is local and cheap, the same derivative-plus-history rule could be grafted onto other spiking neuron variants and likely improve their noise robustness too; the paper only tests LIF.
  • The noise-suppression property suggests HDRP might also confer adversarial robustness, a different threat model than the Gaussian noise tested here, which could be checked with standard adversarial attacks.
  • The reported firing-rate reduction suggests HDRP could act as a regularizer in ANN-to-SNN conversion or knowledge distillation, potentially letting converted networks operate at fewer timesteps.
  • One could test whether the learned kernel parameter A varies systematically with layer depth or dataset noise, which would reveal whether the model learns a homeostatic gain schedule rather than a single global inhibition level.
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

3 major / 6 minor

Summary. The paper proposes a Historical Dynamic Refractory Period (HDRP) model for low-latency spiking neural networks, in which each neuron's refractory period is updated from the membrane potential derivative and its own past refractory state, together with a threshold-dependent refractory kernel that subtracts inhibition from the synaptic current. The authors report state-of-the-art accuracy on CIFAR10 (96.70%), CIFAR100 (80.91%), ImageNet (70.45%), and CIFAR10-DVS (87.00%) at low timesteps, reduced AC operation counts, and improved robustness to Gaussian noise relative to both ANNs and standard SNNs. The central claim is that HDRP-LIF is a drop-in replacement for LIF that reduces redundant spikes while improving accuracy and noise resistance.

Significance. The empirical study is broad: four datasets, two architectures, an ablation of refractory-period components, and a noise-robustness comparison provide a strong testbed for the proposed neuron model. If the mechanism were implemented exactly as described, the consistent accuracy gains and the reported noise robustness would be a useful contribution to low-latency SNN design, and the drop-in nature of the neuron would make it easy to adopt. The paper also explicitly reports energy estimates and firing-related operation counts, which is valuable. However, the theoretical derivation of the normalization that drives the refractory update contains multiple internal inconsistencies, and the energy table is not consistent with the stated operation-count formulas; these issues currently prevent the empirical results from being attributed to the mechanism as written.

major comments (3)
  1. [§3.1, Eq. (2) and §4.1, Eqs. (9)–(10)] The discretized LIF recurrence is not the explicit Euler discretization of Eq. (1). With a unit step, Eq. (1) gives U^{t+1} = (1 - 1/τm) U^t + (1/τm) Urest + (Rm/τm) I^t, not U^{t+1} = (1/τm) U^t + Urest + Rm I^t. This error propagates: Eq. (9) uses a geometric sum with ratio (1 - 1/τm), whereas the recurrence in Eq. (2) would imply ratio 1/τm; and Eq. (10) swaps the max and min labels, since from U' = -(U - Urest)/τm + I the maximum derivative occurs at minimum U and maximum I. As a result, the Min-Max normalization k in Eq. (12) can have a negative denominator or inverted monotonicity. Because Eq. (15) applies k multiplicatively to the historical refractory period, the refractory dynamics implemented may be the opposite of the mechanism described in Section 4.2. These issues are load-bearing for the central claim, and the authors should correct the derivation and verify the monotonicity of k, either analytically or through code-level inspection.
  2. [§4.5, Eq. (18) and Table 2] Equation (18) defines MACs as a sum over synaptic fan-outs that depends only on the architecture and timesteps, not on the neuron model. Table 2 nonetheless reports exactly double the MACs for HDRP-LIF relative to LIF on the same architectures (e.g., 16.51 M vs 8.25 M MACs for CIFAR10 ResNet18 at T=6). Adding the refractory kernel in Eq. (14) does not change the number of synaptic connections, and the extra neuron-internal computations (U', H, τ_ref) are not captured by Eq. (18). This inconsistency undermines the energy-efficiency comparison and the claim that HDRP preserves the low-energy advantage. Please explain how the MAC values in Table 2 were computed and how they relate to Eq. (18).
  3. [§4.4, Eq. (16)] The provided backpropagation recurrence for δτ_ref omits the dependence of the refractory period on its own past value through the chain U' → I → g(τ_ref). Specifically, τ_ref(t+1) depends on U'(t+1), U'(t+1) depends on I(t+1), and I(t+1) depends on τ_ref(t) through Eq. (14); this path contributes terms involving ∂k/∂U' and ∂g/∂τ_ref that are not present in Eq. (16). If in practice the model is trained with automatic differentiation, the printed equations are an incomplete description of the learning rule; if the printed equations are actually used, they may be incorrect. Please clarify which is the case.
minor comments (6)
  1. [§2] There is a grammatical error: 'we proposes the HDRP model' should be 'we propose the HDRP model'.
  2. [§3.2, Eq. (3)] Equation (3) and its surrounding text define γ as a surrogate-function smoothing parameter, but γ does not appear in Eq. (3); it first appears in Eq. (17). Please introduce γ where the loss is defined or move its definition to Section 4.4.
  3. [§4.1, Eqs. (6)–(8)] The constants C1 and C2 from the weight-norm bound in Eq. (6) are not used in the subsequent derivation of I_min/I_max or the U' bounds; please clarify their role or remove them to avoid an unused assumption.
  4. [§5.3 and Fig. 3] No error bars, standard deviations, or number of independent runs are reported for the noise-robustness experiments. Given the performance fluctuations mentioned for CIFAR10-DVS, a statement of experimental variance is needed.
  5. [Table 1 footnote] The ANN baseline is described as 'self-implemented' with identical structures and hyperparameters; this is stated only in a footnote. Consider moving this to the main text so the comparison is not missed by readers.
  6. [§5.4, Table 3] The CIFAR10 improvement of HDRP over the LIF baseline is 0.62% (96.08% to 96.70%). Please report the standard deviation across runs or a statistical significance measure to confirm that this gain is not within run-to-run variance.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the HDRP mechanism is a new recurrent state evaluated against external benchmarks, and no derived result reduces by construction to a fitted parameter or self-citation.

full rationale

The paper's central claims—accuracy on CIFAR10/100, ImageNet, CIFAR10-DVS, reduced spike counts, and noise robustness—are supported by external benchmark experiments and not by a chain that re-inserts the claimed results as inputs. The HDRP refractory period in Eq. (11) and Eq. (15) is defined as a recurrent state from the membrane-potential derivative and previous refractory value; it is not fitted to the reported accuracies. The learnable parameter A in Eq. (13) is trained with STBP, not solved to reproduce Table 1. The bounds on U' in Section 4.1 are derived, not assumed from the target results, and while the derivation contains apparent mathematical errors (e.g., Eq. (10) swapping min/max labels), an incorrect derivation is a correctness issue, not circularity. No load-bearing self-citation appears: the reference list contains no works by the same authors, and cited prior work is used only for comparison or as ablation baselines. The only mild concern—selecting the best ablation configuration from the same test set—is a methodological weakness rather than a circular reduction of the paper's derivation chain. Accordingly, no specific circular step can be quoted from the paper.

Assumptions & free parameters 3 free parameters · 5 assumptions · 1 invented entities

The HDRP mechanism introduces three tunable quantities (tau_ref0, tau_refL, and learnable A) and a recurrent state variable. The bounds used to normalize the membrane potential derivative are asserted rather than cleanly derived. No new physical entities are proposed; the historical refractory state is a computational construct.

free parameters (3)
  • tau_ref0 (minimum refractory period) = not reported
    Hand-set scalar that sets the minimum refractory period; enters Eq. (11)/(15) and affects spike suppression.
  • tau_refL (refractory period range) = ablation table uses max length 6 time steps
    Controls the adjustable range of the refractory period; enters Eq. (11)/(15).
  • A (learnable refractory kernel scale) = learned, not reported
    Free parameter in g(tau_ref)=Uth*tanh(A*tau_ref), trained with STBP; controls inhibition strength relative to threshold.
assumptions (5)
  • domain assumption Membrane potential derivative magnitude indicates degree of neuronal over-activation
    Section 4.2 states 'The larger the derivative ... the higher the degree of neuronal over-activation.' No proof or biological calibration is given.
  • domain assumption Input current bounds I_min and I_max are computed solely from positive and negative weights and are independent of the input signal
    Section 4.1 uses this to bound the membrane potential derivative; the stated bound ignores the dependence of membrane potential and current on the input trajectory.
  • ad hoc to paper The recurrence in Eq. (15) keeps tau_ref within (tau_ref0, tau_refmax)
    The paper asserts this range but provides no clamping or proof; for k(U') near 1 and repeated spikes, tau_ref can grow beyond tau_refmax.
  • domain assumption Min-Max normalization using U'_min and U'_max is numerically stable and avoids gradient vanishing
    Asserted in Section 4.2; the derivation of the bounds is inconsistent, so this is unverified.
  • standard math Surrogate gradient approximation in Eq. (17) is a valid replacement for the spike derivative
    STBP training relies on this standard approximation; the paper does not prove its validity beyond prior work.
invented entities (1)
  • Historical refractory period state tau_ref^(l,t)
    purpose: A recurrent internal state used to modulate refractory duration and inhibition strength
    Introduced in Eq. (11); it is a computational construct whose dynamics are defined by the paper, with no independent physiological measurement provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Optimization of Low-Latency Spiking Neural Networks Utilizing Historical Dynamics of Refractory Periods." pith.science (2026). https://pith.science/paper/RNJ73WAP

@misc{pith2026250702960,
  author       = {Pith},
  title        = {Pith review of: Optimization of Low-Latency Spiking Neural Networks Utilizing Historical Dynamics of Refractory Periods},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RNJ73WAP}},
  note         = {Machine review of arXiv:2507.02960}
}
read the original abstract

The refractory period controls neuron spike firing rate, crucial for network stability and noise resistance. With advancements in spiking neural network (SNN) training methods, low-latency SNN applications have expanded. In low-latency SNNs, shorter simulation steps render traditional refractory mechanisms, which rely on empirical distributions or spike firing rates, less effective. However, omitting the refractory period amplifies the risk of neuron over-activation and reduces the system's robustness to noise. To address this challenge, we propose a historical dynamic refractory period (HDRP) model that leverages membrane potential derivative with historical refractory periods to estimate an initial refractory period and dynamically adjust its duration. Additionally, we propose a threshold-dependent refractory kernel to mitigate excessive neuron state accumulation. Our approach retains the binary characteristics of SNNs while enhancing both noise resistance and overall performance. Experimental results show that HDRP-SNN significantly reduces redundant spikes compared to traditional SNNs, and achieves state-of-the-art (SOTA) accuracy both on static datasets and neuromorphic datasets. Moreover, HDRP-SNN outperforms artificial neural networks (ANNs) and traditional SNNs in noise resistance, highlighting the crucial role of the HDRP mechanism in enhancing the performance of low-latency SNNs.

Figures

Figures reproduced from arXiv: 2507.02960 by the authors.

Figure 1
Figure 1. HDRP-SNN architecture. The internal state updates of the HDRP-LIF model are performed locally, where the refrac￾tory time constant τ (l,t) ref is modulated by both the previous time step τ (l,t−1) ref and the derivative of the current membrane potential. Additionally, the refractory kernel Inhibits the synaptic current at the next time step, effectively regulating the firing rate of the neu￾rons. The HDRP-LIF can be… view at source ↗
Figure 2
Figure 2. Examples of CIFAR10 Images Under Different Levels [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Accuracy Comparison of Different Models Under Varying Gaussian Noise Levels. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 41 canonical work pages

  1. [1]

    A spiking central pattern generator for the control of a simulated lamprey robot running on spin- naker and loihi neuromorphic boards

    Emmanouil Angelidis, Emanuel Buchholz, Jonathan Arre- guit, Alexis Roug´e, Terrence Stewart, Axel von Arnim, Alois Knoll, and Auke Ijspeert. A spiking central pattern generator for the control of a simulated lamprey robot running on spin- naker and loihi neuromorphic boards. Neuromorphic Com- puting and Engineering, 1(1):014005, 2021. 2

    work page 2021

  2. [2]

    Refractoriness enhances temporal cod- ing by auditory nerve fibers

    Michael Avissar, John H Wittig, James C Saunders, and Thomas D Parsons. Refractoriness enhances temporal cod- ing by auditory nerve fibers. Journal of Neuroscience , 33 (18):7681–7690, 2013. 2

    work page 2013

  3. [3]

    A solution to the learning dilemma for recurrent net- works of spiking neurons

    Guillaume Bellec, Franz Scherr, Anand Subramoney, Elias Hajek, Darjan Salaj, Robert Legenstein, and Wolfgang Maass. A solution to the learning dilemma for recurrent net- works of spiking neurons. Nature communications, 11(1): 3625, 2020. 2

    work page 2020

  4. [4]

    Refractoriness and neu- ral precision

    Michael Berry and Markus Meister. Refractoriness and neu- ral precision. Advances in neural information processing sys- tems, 10, 1997. 3

    work page 1997

  5. [5]

    Abhiroop Bhattacharjee, Ruokai Yin, Abhishek Moitra, and Priyadarshini Panda. Are snns truly energy-efficient?—a Model LIF LIF no His 1 His with Fixed Decay Factor2 Ours Refractory Period Type - absolute relative relative relative Refractory Period Length - 1 2 adaptive, ≤ 6 adaptive, ≤ 6 adaptive, ≤ 6 Decay Type - linear exponential exponential exponenti...

    work page 2024

  6. [6]

    Rhythms of the Brain

    Gy ¨orgy Buzs´aki. Rhythms of the Brain . Oxford university press, 2006. 2

    work page 2006

  7. [7]

    Biophysics-inspired spike rate adaptation for computa- tionally efficient phenomenological nerve modeling

    Jacob de Nobel, Savine SM Martens, Jeroen J Briaire, Thomas HW B ¨ack, Anna V Kononova, and Johan HM Fri- jns. Biophysics-inspired spike rate adaptation for computa- tionally efficient phenomenological nerve modeling. Hear- ing Research, 447:109011, 2024. 2, 3

    work page 2024

  8. [8]

    Surrogate module learning: Reduce the gradient error accumulation in training spiking neural networks

    Shikuang Deng, Hao Lin, Yuhang Li, and Shi Gu. Surrogate module learning: Reduce the gradient error accumulation in training spiking neural networks. In International Confer- ence on Machine Learning, pages 7645–7657. PMLR, 2023. 2

    work page 2023

Show all 45 references
  1. [9]

    Parallel spiking neurons with high efficiency and abil- ity to learn long-term dependencies

    Wei Fang, Zhaofei Yu, Zhaokun Zhou, Ding Chen, Yanqi Chen, Zhengyu Ma, Timoth ´ee Masquelier, and Yonghong Tian. Parallel spiking neurons with high efficiency and abil- ity to learn long-term dependencies. Advances in Neural In- formation Processing Systems, 36, 2024. 6

    2024

  2. [10]

    Hosnn: Adversarially-robust home- ostatic spiking neural networks with adaptive firing thresh- olds

    Hejia Geng and Peng Li. Hosnn: Adversarially-robust home- ostatic spiking neural networks with adaptive firing thresh- olds. arXiv preprint arXiv:2308.10373, 2023. 2

    2023 arXiv

  3. [11]

    Spatiotemporal pattern formation in two-dimensional neural circuits: roles of refractoriness and noise

    Pulin Gong, STC Loi, Peter A Robinson, and CYJ Yang. Spatiotemporal pattern formation in two-dimensional neural circuits: roles of refractoriness and noise. Biological cyber- netics, 107:1–13, 2013. 3

    2013

  4. [12]

    Direct learning-based deep spiking neural networks: a review

    Yufei Guo, Xuhui Huang, and Zhe Ma. Direct learning-based deep spiking neural networks: a review. Frontiers in Neuro- science, 17:1209795, 2023. 1

    2023

  5. [13]

    Mem- brane potential batch normalization for spiking neural net- works

    Yufei Guo, Yuhan Zhang, Yuanpei Chen, Weihang Peng, Xi- aode Liu, Liwen Zhang, Xuhui Huang, and Zhe Ma. Mem- brane potential batch normalization for spiking neural net- works. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 19420–19430, 2023. 3

    2023

  6. [14]

    Take a shortcut back: Mitigating the gradient vanishing for training spiking neural networks

    Yufei Guo, Yuanpei Chen, Zecheng Hao, Weihang Peng, Zhou Jie, Yuhan Zhang, Xiaode Liu, and Zhe Ma. Take a shortcut back: Mitigating the gradient vanishing for training spiking neural networks. arXiv preprint arXiv:2401.04486,

    arXiv

  7. [15]

    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. 6

    2023

  8. [16]

    Balanced resonate-and-fire neurons.arXiv preprint arXiv:2402.14603, 2024

    Saya Higuchi, Sebastian Kairat, Sander M Bohte, and Sebas- tian Otte. Balanced resonate-and-fire neurons.arXiv preprint arXiv:2402.14603, 2024. 3, 8

    2024 arXiv

  9. [17]

    Ad- vancing spiking neural networks toward deep residual learn- ing

    Yifan Hu, Lei Deng, Yujie Wu, Man Yao, and Guoqi Li. Ad- vancing spiking neural networks toward deep residual learn- ing. IEEE Transactions on Neural Networks and Learning Systems, 2024. 6

    2024

  10. [18]

    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. 3

    2024 arXiv

  11. [19]

    Rethinking the role of normalization and resid- ual blocks for spiking neural networks

    Shin-ichi Ikegawa, Ryuji Saiin, Yoshihide Sawada, and Nao- take Natori. Rethinking the role of normalization and resid- ual blocks for spiking neural networks. Sensors, 22(8):2876,

  12. [20]

    How inhibition shapes cortical activity

    Jeffry S Isaacson and Massimo Scanziani. How inhibition shapes cortical activity. Neuron, 72(2):231–243, 2011. 2

    2011

  13. [21]

    Dynamical systems in neuroscience

    Eugene M Izhikevich. Dynamical systems in neuroscience . MIT press, 2007. 2

    2007

  14. [22]

    Tab: Temporal accumulated batch normal- ization in spiking neural networks

    Haiyan Jiang, Vincent Zoonekynd, Giulia De Masi, Bin Gu, and Huan Xiong. Tab: Temporal accumulated batch normal- ization in spiking neural networks. In The Twelfth Interna- tional Conference on Learning Representations. 6

  15. [23]

    Biophysics of computation: information pro- cessing in single neurons

    Christof Koch. Biophysics of computation: information pro- cessing in single neurons. Oxford university press, 2004. 2

    2004

  16. [24]

    Bsnn: Towards faster and better conversion of artificial neural networks to spiking neural networks with bistable neurons

    Yang Li, Dongcheng Zhao, and Yi Zeng. Bsnn: Towards faster and better conversion of artificial neural networks to spiking neural networks with bistable neurons. Frontiers in neuroscience, 16:991851, 2022. 2

    2022

  17. [25]

    Spiking mode-based neu- ral networks

    Zhanghan Lin and Haiping Huang. Spiking mode-based neu- ral networks. Physical Review E, 110(2):024306, 2024. 3

    2024

  18. [26]

    Information dynam- ics of in silico eeg brain waves: Insights into oscillations and functions

    Gustavo Menesse and Joaqu ´ın J Torres. Information dynam- ics of in silico eeg brain waves: Insights into oscillations and functions. PLOS Computational Biology , 20(9):e1012369,

  19. [27]

    Training high- performance low-latency spiking neural networks by dif- ferentiation on spike representation

    Qingyan Meng, Mingqing Xiao, Shen Yan, Yisen Wang, Zhouchen Lin, and Zhi-Quan Luo. Training high- performance low-latency spiking neural networks by dif- ferentiation on spike representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition,...

    2022

  20. [28]

    Towards memory-and time-efficient backpropagation for training spiking neural networks

    Qingyan Meng, Mingqing Xiao, Shen Yan, Yisen Wang, Zhouchen Lin, and Zhi-Quan Luo. Towards memory-and time-efficient backpropagation for training spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6166–6176, 2023. 2, 3

    2023

  21. [29]

    A model for single neuron ac- tivity with refractory effects and spike rate estimation tech- niques

    Scott Monk and Harry Leib. A model for single neuron ac- tivity with refractory effects and spike rate estimation tech- niques. IEEE Transactions on Neural Systems and Rehabili- tation Engineering, 25(4):306–322, 2016. 2

    2016

  22. [30]

    Direct training of snn using local zeroth order method

    Bhaskar Mukhoty, Velibor Bojkovic, William de Vazelhes, Xiaohan Zhao, Giulia De Masi, Huan Xiong, and Bin Gu. Direct training of snn using local zeroth order method. Advances in Neural Information Processing Systems , 36: 18994–19014, 2023. 6

    2023

  23. [31]

    Mean-reverting neuronal model based on two alternating patterns

    Nikita Ratanov. Mean-reverting neuronal model based on two alternating patterns. BioSystems, 196:104190, 2020. 2

    2020

  24. [32]

    Exploring neuromorphic computing based on spiking neural networks: Algorithms to hardware

    Nitin Rathi, Indranil Chakraborty, Adarsh Kosta, Abhronil Sengupta, Aayush Ankit, Priyadarshini Panda, and Kaushik Roy. Exploring neuromorphic computing based on spiking neural networks: Algorithms to hardware. ACM Computing Surveys, 55(12):1–49, 2023. 2

    2023

  25. [33]

    Effi- cient spiking neural networks with sparse selective activation for continual learning

    Jiangrong Shen, Wenyao Ni, Qi Xu, and Huajin Tang. Effi- cient spiking neural networks with sparse selective activation for continual learning. In Proceedings of the AAAI Confer- ence on Artificial Intelligence, pages 611–619, 2024. 3

    2024

  26. [34]

    Modeling eluci- dates how refractory period can provide profound nonlinear gain control to graded potential neurons

    Zhuoyi Song, Yu Zhou, and Mikko Juusola. Modeling eluci- dates how refractory period can provide profound nonlinear gain control to graded potential neurons. Physiological re- ports, 5(11):e13306, 2017. 2, 3

    2017

  27. [35]

    Addressing the speed-accuracy simulation trade-off for adaptive spiking neurons

    Luke Taylor, Andrew King, and Nicol S Harper. Addressing the speed-accuracy simulation trade-off for adaptive spiking neurons. Advances in Neural Information Processing Sys- tems, 36, 2024. 2

    2024

  28. [36]

    Significant anisotropic neuronal refractory period plasticity

    Roni Vardi, Yael Tugendhaft, Shira Sardi, and Ido Kanter. Significant anisotropic neuronal refractory period plasticity. Europhysics Letters, 134(6):60007, 2021. 2

    2021

  29. [37]

    Ssf: Accelerating training of spiking neural networks with stabilized spiking flow

    Jingtao Wang, Zengjie Song, Yuxi Wang, Jun Xiao, Yuran Yang, Shuqi Mei, and Zhaoxiang Zhang. Ssf: Accelerating training of spiking neural networks with stabilized spiking flow. In Proceedings of the IEEE/CVF International Confer- ence on Computer Vision, pages 5982–5991, 2023. 3

    2023

  30. [38]

    Spatio-temporal backpropagation for training high- performance spiking neural networks

    Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, and Luping Shi. Spatio-temporal backpropagation for training high- performance spiking neural networks. Frontiers in neuro- science, 12:331, 2018. 5

    2018

  31. [39]

    Biologically inspired structure learning with reverse knowledge distillation for spiking neu- ral networks

    Qi Xu, Yaxin Li, Xuanye Fang, Jiangrong Shen, Jian K Liu, Huajin Tang, and Gang Pan. Biologically inspired structure learning with reverse knowledge distillation for spiking neu- ral networks. arXiv preprint arXiv:2304.09500, 2023. 2

    2023 arXiv

  32. [40]

    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 Process- ing Systems, 35:32160–32171, 2022. 3

    2022

  33. [41]

    A review of spiking neuromorphic hardware commu- nication systems

    Aaron R Young, Mark E Dean, James S Plank, and Garrett S Rose. A review of spiking neuromorphic hardware commu- nication systems. IEEE Access, 7:135606–135620, 2019. 2

    2019

  34. [42]

    Advancing training efficiency of deep spiking neu- ral networks through rate-based backpropagation

    Chengting Yu, Lei Liu, Gaoang Wang, Erping Li, and Aili Wang. Advancing training efficiency of deep spiking neu- ral networks through rate-based backpropagation. arXiv preprint arXiv:2410.11488, 2024. 3

    2024 arXiv

  35. [43]

    Cutmix: Regu- larization strategy to train strong classifiers with localizable features

    Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regu- larization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international con- ference on computer vision, pages 6023–6032, 2019. 6

    2019

  36. [44]

    Self-evolutionary neuron model for fast-response spiking neural networks

    Anguo Zhang, Ying Han, Yuzhen Niu, Yueming Gao, Zhizhang Chen, and Kai Zhao. Self-evolutionary neuron model for fast-response spiking neural networks. IEEE Transactions on Cognitive and Developmental Systems , 14 (4):1766–1777, 2021. 3

    2021

  37. [45]

    Retina-like visual image reconstruction via spiking neural model

    Lin Zhu, Siwei Dong, Jianing Li, Tiejun Huang, and Yonghong Tian. Retina-like visual image reconstruction via spiking neural model. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 1438–1446, 2020. 3

    2020

Pith tools

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