Pith. sign in

REVIEW 3 major objections 4 minor 53 references

Low-Bit Data Processing Using Multiple-Output Spiking Neurons with Non-linear Reset Feedback

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · 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 multiple-output spiking neuron with a nonlinear, learnable reset can learn under unstable linear state dynamics—relaxing the strict stability constraints of deep state-space models—and match established SNN baseline

desk verdict Honest empirical study of a decoupled reset for MIMO spiking SSMs, but the headline claim about overcoming instability is not established because all unstable runs rely on state clipping. read the letter →

arxiv 2508.06292 v1 pith:VZQO6UAG submitted 2025-08-08 cs.LG

Sanja Karilanova , Subhrakanti Dey , Ayça Özçelikkale This is my paper
classification cs.LG MSC 68T07
keywords spikingneuralnetworksstate-spacemodelsresetmechanismlow-bitprocessingneuromorphiccomputingevent-basedsignalunstabledynamicssurrogategradients
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

This paper tries to establish that a spiking neuron built on a linear state-space transition remains trainable under unstable linear dynamics if it is equipped with a nonlinear, learnable reset. The proposed neuron separates spike generation, reset condition, and reset action: the reset triggers when the norm of a multi-output projection of the state crosses a learnable threshold, and the action scales every state variable by a learnable complex factor. On keyword spotting, event-based gesture recognition, and sequential MNIST, the reset lifts 1-bit spike communication from a collapsed no-reset baseline (e.g., 40.2% to 91.5% on MSWC, 82.9% to 96.3% on sMNIST) and reaches accuracy comparable to published SNN/SSM results. If true, this means deep state-space models need not enforce stable eigenvalues to learn, and low-bit spike communication can be combined with richer, even unstable, recurrent dynamics.

What carries the argument

The load-bearing object is the decoupled reset feedback: a hard reset triggered not by the spike itself but by a separate condition on the neuron's multi-output $y$, which then scales all state variables by a learnable complex factor $r_{\mathrm{scale}}$. The condition uses a Euclidean norm with a learnable bias $r_{\mathrm{bias}}$, and surrogate gradients are used to train through both the spike and reset Heaviside functions. This reset acts as a nonlinear feedback path that can bound states even when the diagonal transition matrix $A$ has eigenvalues with modulus greater than one—precisely the regime deep SSMs exclude by stable initialization and eigenvalue clipping.

What would settle it

Train the Unstable+Reset model on MSWC with state clipping removed (and no gradient clipping) using the paper's hyperparameters: if hidden states diverge or accuracy falls toward the 40% no-reset level, then the reset alone does not overcome instability. A complementary check: construct instability by setting all eigenvalues' moduli above 1 instead of multiplying every second eigenvalue by 1.5; if reset no longer stabilizes, the claim is specific to near-marginal dynamics.

Watch

Extended reading notes

Core claim

The central claim is that the nonlinearity of a state reset—not the linear dynamics—can be the stabilizing mechanism in a spiking SSM. The model is $v[t+1]=Av[t]+Bi[t]$, $y[t]=Cv[t]+c_{\mathrm{bias}}$, spike $s=f_\theta(y)$, and if $\frac{1}{n_{\mathrm{out}}}\lVert y[t]\rVert^2 + r_{\mathrm{bias}} \ge 1$ then $v \leftarrow r_{\mathrm{scale}} v$, with $A$ diagonal and complex, $r_{\mathrm{bias}}$ real, $r_{\mathrm{scale}}$ complex, and both learnable. With $A$ initialized so that half its eigenvalues have modulus above 1, the no-reset model with binary spikes fails (MSWC 40.2%, sMNIST 82.9%), while the reset model reaches 91.5% and 96.3% under the same 1-bit communication. On the same tasks t

Load-bearing premise

The load-bearing premise is that the reset mechanism—not the state clipping, gradient clipping, or hyperparameters transferred from the stable case—is what stabilizes training when the transition-matrix eigenvalues exceed 1.

Editorial extensions

If this is right

  • Training SSM-style spiking networks no longer requires the transition-matrix eigenvalues to stay inside the unit circle; unstable initialization becomes viable when a reset path exists.
  • One-bit spike communication suffices for near-continuous accuracy on several temporal tasks, with the reset adding only $O(h)$ operations per layer against $O(h^2)$ for the synaptic weights.
  • Decoupling the reset condition from spike generation is a viable design choice: it can improve accuracy and reach 80% accuracy earlier in the sequence without hurting convergence.
  • Multiple output channels per neuron are not redundant; removing a single channel drops accuracy substantially, so the multi-output structure contributes information rather than just parameter count.
  • In the examined configurations the reset mechanism reduces spike rate or leaves it unchanged, which supports low-power neuromorphic deployment.

Reading between the lines

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

  • If the reset is the stabilizer, the same decoupling idea could be applied to continuous-valued recurrent or SSM layers: a learnable norm-gated state scaling may allow training with eigenvalues above 1 without spiking, a direct testable extension beyond the paper's 1-bit setting.
  • The norm-based reset with learnable bias resembles an energy threshold; one could test whether making the threshold adaptive over time or per-channel improves long-sequence tasks where stable SSMs suffer from fast decay.
  • The paper's instability construction is mild: only half the eigenvalues are pushed to modulus roughly 1.35. A stronger test—all eigenvalues unstable, or a time-varying $A$—would reveal whether the reset generalizes or only rescues near-marginal dynamics.
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 / 4 minor

Summary. The paper proposes a multiple-output spiking neuron that combines a linear, diagonal state-space transition (complex-valued A) with a nonlinear reset mechanism. The reset condition is a learned norm threshold on the neuron output, and the reset action is a learned complex scaling of the state (Section IV-B, Eqs. (6)-(9)). The model is evaluated on MSWC keyword spotting, DVS-Gesture, and sequential MNIST under four configurations: Stable/Unstable x Reset/NoReset, with non-signed spikes, signed spikes, and GELU activations. The authors report that reset substantially improves accuracy in unstable configurations on MSWC and sMNIST, and they conclude in the abstract and Section VII that reset can overcome instability and go beyond the strict stability enforcement used in deep SSM models.

Significance. If the central claim were fully established, the paper would offer a meaningful bridge between SNN reset mechanisms and deep SSM stability practice: a spiking neuron whose linear part is allowed to be unstable, with divergence controlled by a nonlinear reset, would expand the expressivity of SSM-based sequence models under low-bit communication. The paper is thorough in reporting model equations, parameter counts, MAC complexity, initialization details, hyperparameters, and ablations over architecture dimensions. These strengths make the empirical protocol unusually transparent. However, the key instability claim is currently not fully supported by the experiments, because the unstable configurations always include a hard state-magnitude clip, and because the only dataset showing a systematic reset benefit under instability is MSWC/sMNIST, while DVS-Gesture shows no benefit or a mean degradation. The contribution is promising but needs a scoped claim or additional experiments.

major comments (3)
  1. [§VIII-A1, §VI-B] The central claim (abstract, §VII) that reset can 'overcome instability and enable learning' when the linear dynamics are unstable is not tested without additional stabilization. In §VIII-A1, all unstable runs clip the state as |v_k|<1000, and §VI-B explicitly states that without clipping the computation cannot be performed due to diverging states and exploding gradients. Thus the Unstable+Reset configuration is never evaluated as the sole stabilizer; the reported gaps (e.g., MSWC non-signed 91.5 vs 40.2; sMNIST 96.3 vs 82.9) show that reset plus clipping beats clipping alone, not that reset alone prevents divergence. The state clip is itself an externally imposed stability bound, so the claim of going 'beyond strictly enforced stability' outruns the experimental protocol. Please either remove/scope this claim or provide an experiment without state clipping (or with a clipping bound that
  2. [§VI-B, Table IIb] The instability-reset benefit does not hold on DVS-Gesture. Under non-signed spikes, Unstable+Reset achieves 69.0±7.0% versus 75.4±14.0% for Unstable+NoReset; the signed and GELU variants are statistically tied. The text states that the differences are within standard-deviation limits, but the mean direction is opposite to the claimed effect on this dataset. Since the abstract and Section VII make a general claim ('can overcome instability'), a dataset where reset numerically degrades unstable learning must be addressed, for example by unstable-specific hyperparameter optimization or by an analysis of why the reset mechanism fails there. As it stands, the evidence supports a dataset-dependent benefit, not the stated general claim.
  3. [§VIII-A2] Hyperparameter transfer weakens the instability comparison. For sMNIST and DVS-Gesture, hyperparameters (including reset-specific ones) were tuned on Stable+NoReset/GELU and Stable+Reset/GELU, respectively, and then used unchanged for the unstable variants. The Unstable+Reset and Unstable+NoReset configurations are therefore compared under hyperparameters that were never optimized for the unstable regime. The MSWC/sMNIST gaps are large and encouraging, but the sensitivity of the central claim to the specific instability construction (multiplying every second eigenvalue of the S4D-Lin initialization by 1.5) and to hyperparameter choices remains unknown. At minimum, the paper should validate that the unstable configurations are not disadvantaged by the transfer, or vary the instability multiplier/clipping threshold to show that the reset benefit is not an artifact of this particular protoc
minor comments (4)
  1. [§V-3] Typos: 'Figure!4b' appears in two places; the exclamation mark should be a period or comma.
  2. [§VIII-A1] The notation 'f (˙)' is malformed; likely 'f(·)' is intended.
  3. [References] Reference [1] has an awkward space in 'V . Sze'; please fix spacing.
  4. [§VI-G, Table IV] The channel-drop experiment shows a very large degradation when a single output channel is removed (e.g., MSWC drops from 95.0 to ~48). The conclusion that all channels are 'similarly important' is reasonable, but the magnitude deserves a sentence of interpretation, since it also suggests high redundancy at the neuron level.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the central claim is empirical and does not reduce to its inputs, though the state-clipping protocol confounds the instability claim.

full rationale

The paper's central claim is empirical: the proposed reset feedback improves learning when the diagonal state-transition eigenvalues are destabilized. No derivation is presented that would make the conclusion equivalent to an input by construction. The reset parameters (rscale, rbias) are learned from data, not fitted to the reported test accuracies, and the unstable configurations are created by a fixed eigenvalue multiplier (VIII-A1), not by optimizing for the desired outcome. The self-citations [15] and [29] are disclosed and non-load-bearing: [15] is the prior no-reset model that the paper explicitly generalizes, and [29] is one of several external baselines. The main validity concern is that Unstable+NoReset and Unstable+Reset both use hard state clipping |v_k|<1000 and gradient clipping to 10^5 (VIII-A1), so reset is never tested as the sole stabilizer; Section VI-B also admits that without clipping the computation cannot be performed. This is a confound in the experimental protocol, not a circularity: the comparison is not defined in terms of the quantity being predicted, and no fitted parameter is renamed as a prediction. Therefore the circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 1 invented entities

All ordinary network weights are trained on the benchmark tasks and are not listed here. The entries above are the hand-chosen constants and background assumptions that the instability-reset conclusion specifically depends on: the instability construction, the state clipping safeguard, the reset parameter initializations, and the validity of cross-paper comparisons.

free parameters (3)
  • eigenvalue instability multiplier = 1.5
    Applied to every second eigenvalue of the S4D-Lin initialization to produce 'unstable' dynamics; the instability results depend on this hand-chosen factor.
  • state clipping bound = |v_k| < 1000
    Used for all unstable configurations to prevent divergence; the reset mechanism is never tested without this clipping, so it is a co-stabilizer.
  • rscale and rbias initializations = rscale: 0.1/0.8/0.5 per dataset; rbias: 0
    Selected by hyperparameter optimization for reset models; the reset benefit is partly tuned per dataset.
assumptions (4)
  • domain assumption Discrete-time linear SSM state transition with complex diagonal A is a valid model for temporal data.
    Invoked throughout Section IV; follows from the S4/S4D deep SSM literature and is adopted without independent validation here.
  • domain assumption Surrogate gradients of the Heaviside functions provide a usable training signal for both the spiking condition and the reset condition.
    Section V-4 states that BPTT with surrogate gradients is used; no convergence guarantee is given, and the reset Heaviside is non-differentiable.
  • domain assumption Cross-paper accuracy comparisons in Table II are meaningful without re-running the cited baselines.
    RadLIF, PLIF, SpikingSSM, S5, and other baselines were trained under different protocols, hardware, and hyperparameters; the paper assumes these numbers are directly comparable.
  • domain assumption S4D-Lin initialization with bilinear discretization produces eigenvalue moduli near 0.9.
    Section VIII-A1 relies on this to construct unstable dynamics via the 1.5 multiplier; this is taken from [10] rather than verified here.
invented entities (1)
  • Multiple-output SSM-based spiking neuron with nonlinear reset feedback
    purpose: Low-bit temporal processing; decouples spike output, reset condition, and reset action; aims to stabilize learning when the linear transition matrix is unstable.
    The proposed architecture itself is evaluated only in this paper's experiments; it makes no external falsifiable prediction and has no theoretical guarantee beyond the reported benchmark runs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Low-Bit Data Processing Using Multiple-Output Spiking Neurons with Non-linear Reset Feedback." pith.science (2026). https://pith.science/paper/VZQO6UAG

@misc{pith2026250806292,
  author       = {Pith},
  title        = {Pith review of: Low-Bit Data Processing Using Multiple-Output Spiking Neurons with Non-linear Reset Feedback},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VZQO6UAG}},
  note         = {Machine review of arXiv:2508.06292}
}
read the original abstract

Neuromorphic computing is an emerging technology enabling low-latency and energy-efficient signal processing. A key algorithmic tool in neuromorphic computing is spiking neural networks (SNNs). SNNs are biologically inspired neural networks which utilize stateful neurons, and provide low-bit data processing by encoding and decoding information using spikes. Similar to SNNs, deep state-space models (SSMs) utilize stateful building blocks. However, deep SSMs, which recently achieved competitive performance in various temporal modeling tasks, are typically designed with high-precision activation functions and no reset mechanisms. To bridge the gains offered by SNNs and the recent deep SSM models, we propose a novel multiple-output spiking neuron model that combines a linear, general SSM state transition with a non-linear feedback mechanism through reset. Compared to the existing neuron models for SNNs, our proposed model clearly conceptualizes the differences between the spiking function, the reset condition and the reset action. The experimental results on various tasks, i.e., a keyword spotting task, an event-based vision task and a sequential pattern recognition task, show that our proposed model achieves performance comparable to existing benchmarks in the SNN literature. Our results illustrate how the proposed reset mechanism can overcome instability and enable learning even when the linear part of neuron dynamics is unstable, allowing us to go beyond the strictly enforced stability of linear dynamics in recent deep SSM models.

Figures

Figures reproduced from arXiv: 2508.06292 by the authors.

Figure 1
Figure 1. Illustration of dynamics of a neuron with multiple [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Visualization of the reset condition region, represented [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. An illustration of a network with two hidden layers [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Visualization of activation function and their surrogate [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Test classification accuracy over training epochs for [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Test classification accuracy over time-steps of the input [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 47 canonical work pages

  1. [1]

    Data centers on wheels: Emissions from computing onboard autonomous vehicles,

    S. Sudhakar, V . Sze, and S. Karaman, “Data centers on wheels: Emissions from computing onboard autonomous vehicles,” IEEE Micro, vol. 43, no. 1, pp. 29–39, 2023

    work page 2023

  2. [2]

    Low-power neuromorphic hardware for signal processing appli- cations: A review of architectural and system-level design approaches,

    B. Rajendran, A. Sebastian, M. Schmuker, N. Srinivasa, and E. Elefthe- riou, “Low-power neuromorphic hardware for signal processing appli- cations: A review of architectural and system-level design approaches,” IEEE Signal Processing Magazine , vol. 36, no. 6, pp. 97–110, 2019

    work page 2019

  3. [3]

    Advancing neuromorphic computing with Loihi: A survey of results and outlook,

    M. Davies et al., “Advancing neuromorphic computing with Loihi: A survey of results and outlook,” Proceedings of the IEEE , vol. 109, no. 5, pp. 911–934, 2021

    work page 2021

  4. [4]

    Neurobench: A framework for benchmarking neuromorphic computing algorithms and systems,

    J. Yik et al., “Neurobench: A framework for benchmarking neuromorphic computing algorithms and systems,” 2024

    work page 2024

  5. [5]

    Gerstner and W

    W. Gerstner and W. M. Kistler, Spiking Neuron Models: Single Neurons, Populations, Plasticity. Cambridge University Press, 2002

    2002

  6. [6]

    Advancing spatio-temporal processing in spiking neural networks through adaptation,

    M. Baronig, R. Ferrand, S. Sabathiel, and R. Legenstein, “Advancing spatio-temporal processing in spiking neural networks through adaptation,”

    work page

  7. [7]

    A surrogate gradient spiking baseline for speech command recognition,

    A. Bittar and P. N. Garner, “A surrogate gradient spiking baseline for speech command recognition,” Frontiers in Neuroscience, vol. 16, 2022

    work page 2022

  8. [8]

    Incorporating learnable membrane time constant to enhance learning of spiking neural networks,

    W. Fang, Z. Yu, Y . Chen, T. Masquelier, T. Huang, and Y . Tian, “Incorporating learnable membrane time constant to enhance learning of spiking neural networks,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , October 2021, pp. 2661–2671

    work page 2021

Show all 53 references
  1. [9]

    Efficiently modeling long sequences with structured state spaces,

    A. Gu, K. Goel, and C. Ré, “Efficiently modeling long sequences with structured state spaces,” 2022

    2022

  2. [10]

    On the parameterization and initialization of diagonal state space models,

    A. Gu, K. Goel, A. Gupta, and C. Ré, “On the parameterization and initialization of diagonal state space models,” Advances in Neural Information Processing Systems (NeurIPS) , vol. 35, pp. 35 971–35 983, 2022

    2022

  3. [11]

    Simplified state space layers for sequence modeling,

    J. T. Smith, A. Warrington, and S. Linderman, “Simplified state space layers for sequence modeling,” in The Eleventh Inter. Conf. on Learning Representations (ICLR), 2023

    2023

  4. [12]

    Multilingual spoken words corpus,

    M. Mazumder et al., “Multilingual spoken words corpus,” in Conf. on Neural Inf. Process. Systems (NeurIPS) , 2024

    2024

  5. [13]

    A low power, fully event-based gesture recognition system,

    A. Amir et al. , “A low power, fully event-based gesture recognition system,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , July 2017

    2017

  6. [14]

    A simple way to initialize recurrent networks of rectified linear units,

    Q. V . Le, N. Jaitly, and G. E. Hinton, “A simple way to initialize recurrent networks of rectified linear units,” arXiv preprint arXiv:1504.00941, 2015

    2015 arXiv

  7. [15]

    State-space model inspired multiple-input multiple-output spiking neurons,

    S. Karilanova, S. Dey, and A. Özçelikkale, “State-space model inspired multiple-input multiple-output spiking neurons,” in 2025 Neuro Inspired Computational Elements (NICE) , 2025

    2025

  8. [16]

    Perfect Recovery and Sensitivity Analysis of Time Encoded Bandlimited Signals,

    A. Lazar and L. Toth, “Perfect Recovery and Sensitivity Analysis of Time Encoded Bandlimited Signals,” IEEE Tran. on Circuits and Systems I, 2004

    2004

  9. [17]

    FRI-TEM: Time encoding sampling of finite-rate-of-innovation signals,

    H. Naaman, S. Mulleti, and Y . C. Eldar, “FRI-TEM: Time encoding sampling of finite-rate-of-innovation signals,” IEEE Transactions on Signal Processing, vol. 70, pp. 2267–2279, 2022

    2022

  10. [18]

    Bandlimited signal reconstruction from leaky integrate-and-fire encoding using POCS,

    N. T. Thao, D. Rzepka, and M. Mi ´skowicz, “Bandlimited signal reconstruction from leaky integrate-and-fire encoding using POCS,” IEEE Transactions on Signal Processing , vol. 71, pp. 1464–1479, 2023

    2023

  11. [19]

    Asynchrony increases efficiency: Time encoding of videos and low-rank signals,

    K. Adam, A. Scholefield, and M. Vetterli, “Asynchrony increases efficiency: Time encoding of videos and low-rank signals,” IEEE Transactions on Signal Processing , vol. 70, pp. 105–116, 2022

    2022

  12. [20]

    Scalable event-by-event processing of neuromorphic sensory signals with deep state-space models,

    M. Schöne, N. M. Sushma, J. Zhuge, C. Mayr, A. Subramoney, and D. Kappel, “Scalable event-by-event processing of neuromorphic sensory signals with deep state-space models,” IEEE/ACM International Conference on Neuromorphic Systems (ICONS) , 2024

    2024

  13. [21]

    S7: Selective and simplified state space layers for sequence modeling,

    T. Soydan, N. Zubi ´c, N. Messikommer, S. Mishra, and D. Scaramuzza, “S7: Selective and simplified state space layers for sequence modeling,” 2024

    2024

  14. [22]

    Quamba: A post-training quantization recipe for selective state space models,

    H.-Y . Chiang, C.-C. Chang, N. Frumkin, K.-C. Wu, and D. Marculescu, “Quamba: A post-training quantization recipe for selective state space models,” 2024

    2024

  15. [23]

    Q-s5: Towards quantized state space models,

    S. Abreu, J. E. Pedersen, K. M. Heckel, and A. Pierro, “Q-s5: Towards quantized state space models,” 2024

    2024

  16. [24]

    A diagonal structured state space model on Loihi 2 for efficient streaming sequence processing,

    S. M. Meyer et al., “A diagonal structured state space model on Loihi 2 for efficient streaming sequence processing,” 2024

    2024

  17. [25]

    Rethinking spiking neural networks as state space models,

    M. Bal and A. Sengupta, “Rethinking spiking neural networks as state space models,” 2024

    2024

  18. [26]

    Learning long sequences in spiking neural networks,

    M.-I. Stan and O. Rhodes, “Learning long sequences in spiking neural networks,” Scientific Reports, vol. 14, no. 1, p. 21957, 2024

    2024

  19. [27]

    Spikingssms: Learning long sequences with sparse and parallel spiking state space models,

    S. Shen et al., “Spikingssms: Learning long sequences with sparse and parallel spiking state space models,” 2024

    2024

  20. [28]

    Structured state space model dynamics and parametrization for spiking neural networks,

    M. Fabre, L. Dudchenko, and E. Neftci, “Structured state space model dynamics and parametrization for spiking neural networks,” 2025. [Online]. Available: https://arxiv.org/abs/2506.06374

    2025 arXiv

  21. [29]

    Zero-shot temporal resolution domain adaptation for spiking neural networks,

    S. Karilanova, M. Fabre, E. Neftci, and A. Özçelikkale, “Zero-shot temporal resolution domain adaptation for spiking neural networks,”

  22. [30]

    Loihi: A neuromorphic manycore processor with on-chip learning,

    M. Davies et al. , “Loihi: A neuromorphic manycore processor with on-chip learning,” IEEE Micro, vol. 38, no. 1, pp. 82–99, 2018

    2018

  23. [31]

    Available: https://arxiv.org/abs/2411.04760

    [Online]. Available: https://arxiv.org/abs/2411.04760

  24. [32]

    A quantitative description of membrane current and its application to conduction and excitation in nerve,

    A. Hodgkin and A. Huxley, “A quantitative description of membrane current and its application to conduction and excitation in nerve,” Journal of Physiology, vol. 117, pp. 500–544, 1952

    1952

  25. [33]

    Simple model of spiking neurons,

    E. M. Izhikevich, “Simple model of spiking neurons,” IEEE Transactions on neural networks , vol. 14, no. 6, pp. 1569–1572, 2003

    2003

  26. [34]

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

    E. O. Neftci, H. Mostafa, and F. Zenke, “Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks,” IEEE Signal Process. Magazine, vol. 36, pp. 51–63, 2019

    2019

  27. [35]

    Gajic, Linear Dynamic Systems and Signals

    Z. Gajic, Linear Dynamic Systems and Signals . Prentice Hall - Pearson Education Inc., 2003

    2003

  28. [36]

    Self-adapting spiking neural p systems with refractory period and propagation delay,

    Y . Zhao, Y . Liu, X. Liu, M. Sun, F. Qi, and Y . Zheng, “Self-adapting spiking neural p systems with refractory period and propagation delay,” Information Sciences, vol. 589, pp. 80–93, 2022

    2022

  29. [37]

    Superspike: Supervised learning in multilayer spiking neural networks,

    F. Zenke and S. Ganguli, “Superspike: Supervised learning in multilayer spiking neural networks,” Neural Computation , vol. 30, no. 6, p. 1514–1541, jun 2018

    2018

  30. [38]

    Diagonal state spaces are as effective as structured state spaces,

    A. Gupta, A. Gu, and J. Berant, “Diagonal state spaces are as effective as structured state spaces,” in Proceedings of the 36th International Conference on Neural Information Processing Systems , 2022

    2022

  31. [39]

    Leaky integrate- and-fire neuron with a refractory period mechanism for invariant spikes,

    H. M. Lehmann, J. Hille, C. Grassmann, and V . Issakov, “Leaky integrate- and-fire neuron with a refractory period mechanism for invariant spikes,” in 2022 17th Conference on Ph.D Research in Microelectronics and Electronics (PRIME), 2022, pp. 365–368

    2022

  32. [40]

    Spike-driven transformer,

    M. Yao et al., “Spike-driven transformer,” Advances in neural information processing systems, vol. 36, pp. 64 043–64 058, 2023

    2023

  33. [41]

    Batch normalization: accelerating deep network training by reducing internal covariate shift,

    S. Ioffe and C. Szegedy, “Batch normalization: accelerating deep network training by reducing internal covariate shift,” in Proceeding. of the International Conference on Machine Learning (ICML) , vol. 37, 2015, p. 448–456

    2015

  34. [42]

    Very deep convolutional neural networks for raw waveforms,

    W. Dai, C. Dai, S. Qu, J. Li, and S. Das, “Very deep convolutional neural networks for raw waveforms,” in 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2017, pp. 421–425

    2017

  35. [43]

    The mnist database of handwritten digits,

    Y . LeCun, “The mnist database of handwritten digits,” 2017. [Online]. Available: http://yann.lecun.com/exdb/mnist/

    2017

  36. [44]

    Going deeper with directly-trained larger spiking neural networks,

    H. Zheng, Y . Wu, L. Deng, Y . Hu, and G. Li, “Going deeper with directly-trained larger spiking neural networks,” Proceedings of the AAAI Conference on Artificial Intelligence , vol. 35, no. 12, pp. 11 062–11 070, May 2021. 15

    2021

  37. [45]

    Synaptic plasticity dynamics for deep continuous local learning (DECOLLE),

    J. Kaiser, H. Mostafa, and E. Neftci, “Synaptic plasticity dynamics for deep continuous local learning (DECOLLE),” Frontiers in Neuroscience, vol. V olume 14 - 2020, 2020

    2020

  38. [46]

    An adaptive threshold neuron for recurrent spiking neural networks with nanodevice hardware implementation,

    A. Shaban, S. S. Bezugam, and M. Suri, “An adaptive threshold neuron for recurrent spiking neural networks with nanodevice hardware implementation,” Nature Communications, vol. 12, no. 1, p. 4234, 2021

    2021

  39. [47]

    The role of temporal hierarchy in spiking neural networks,

    F. Moro, P. V . Aceituno, L. Kriener, and M. Payvand, “The role of temporal hierarchy in spiking neural networks,” arXiv preprint arXiv:2407.18838, 2024

    2024 arXiv

  40. [48]

    Speech2spikes: Efficient audio encoding pipeline for real-time neuro- morphic systems,

    K. M. Stewart, T. Shea, N. Pacik-Nelson, E. Gallo, and A. Danielescu, “Speech2spikes: Efficient audio encoding pipeline for real-time neuro- morphic systems,” in Proc. of the Annual Neuro-Inspired Computational Elements Conf. (NICE) , 2023, p. 71–78

    2023

  41. [49]

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

    P. A. Merolla et al., “A million spiking-neuron integrated circuit with a scalable communication network and interface,” Science, vol. 345, no. 6197, pp. 668–673, 2014

    2014

  42. [50]

    Efficient recurrent architectures through activity sparsity and sparse back-propagation through time,

    A. Subramoney, K. K. Nazeer, M. Schöne, C. Mayr, and D. Kappel, “Efficient recurrent architectures through activity sparsity and sparse back-propagation through time,” 2023

    2023

  43. [51]

    Tonic: event-based datasets and transformations

    G. Lenz, K. Chaney, S. B. Shrestha, O. Oubari, S. Picaud, and G. Zarrella, “Tonic: event-based datasets and transformations.” 2021

    2021

  44. [53]

    Temporal binary representation for event-based action recognition,

    S. U. Innocenti, F. Becattini, F. Pernici, and A. Del Bimbo, “Temporal binary representation for event-based action recognition,” in 2020 25th International Conference on Pattern Recognition (ICPR) . IEEE, 2021, pp. 10 426–10 432

    2020

  45. [2024]

    Available: https://arxiv.org/abs/2408.07517

    [Online]. Available: https://arxiv.org/abs/2408.07517

    arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.