pith. machine review for the scientific record. sign in

arxiv: 2604.25688 · v1 · submitted 2026-04-28 · 💻 cs.CV

Recognition: unknown

QB-LIF: Learnable-Scale Quantized Burst Neurons for Efficient SNNs

Dewei Bai , Hongxiang Peng , Jiajun Mei , Yang Ren , Hong Qu , Dawen Xia , Zhang Yi
Authors on Pith no claims yet

Pith reviewed 2026-05-07 16:55 UTC · model grok-4.3

classification 💻 cs.CV
keywords spiking neural networksquantized burst neuronslearnable scalesurrogate gradientneuromorphic computinglow-latency inferenceevent-driven vision
0
0 comments X

The pith

A learnable quantization scale for burst spiking lets each SNN layer adapt its resolution to membrane-potential statistics while folding into weights for accumulate-only inference.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that burst spiking can be recast as saturated uniform quantization of membrane potentials with a trainable scale parameter instead of fixed thresholds. This lets every layer tune its own spiking levels to match the statistics of its internal voltages, increasing information throughput beyond binary spikes. The change matters because deep SNNs lose performance when restricted to short simulation horizons on low-power hardware. The method preserves hardware efficiency by absorbing the scale into synaptic weights after training and uses a rectified-linear surrogate gradient with exponential tails to keep gradients flowing through the discrete burst levels.

Core claim

The Quantized Burst-LIF neuron reformulates burst spiking as saturated uniform quantization of membrane potentials with a learnable scale. Each layer trains its scale to fit its membrane-potential distribution. An absorbable scale strategy multiplies the learned value into synaptic weights at inference so the neuron still emits integer burst counts using only additions. The ReLSG-ET surrogate gradient, a rectified linear function with exponential tails, supplies nonzero gradients across quantization steps to support backpropagation through the discrete multi-level space. On static image and event-driven vision benchmarks the resulting networks exceed the accuracy of both binary and fixed-bur

What carries the argument

The QB-LIF neuron, which treats burst spiking as uniform quantization of membrane potential with a trainable scale that is absorbed into synaptic weights at inference.

If this is right

  • Each layer autonomously adapts its spiking resolution to local membrane-potential statistics.
  • Synaptic operations remain strictly accumulate-only after scale absorption.
  • Stable gradient flow is maintained across multiple burst levels during training.
  • Accuracy exceeds binary and fixed-burst SNNs on CIFAR-10, CIFAR-100, ImageNet, CIFAR10-DVS and DVS128-Gesture at ultra-low latency.
  • Neuromorphic hardware compatibility is preserved.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The absorbable-scale idea could be reused for other trainable parameters in SNNs to simplify mapping to standard digital hardware.
  • The surrogate-gradient design may apply to training other discrete multi-level spiking mechanisms.
  • Per-layer adaptation of discretization granularity offers a route to deeper SNNs without proportional increases in simulation time.

Load-bearing premise

The learned quantization scale can be absorbed into synaptic weights without changing the effective spiking behavior or final accuracy.

What would settle it

Train a QB-LIF network, then run inference once with the scale folded into weights and once without; a clear accuracy drop in the folded case on CIFAR-10 or ImageNet at the same latency would falsify the absorption claim.

Figures

Figures reproduced from arXiv: 2604.25688 by Dawen Xia, Dewei Bai, Hong Qu, Hongxiang Peng, Jiajun Mei, Yang Ren, Zhang Yi.

Figure 1
Figure 1. Figure 1: Information capacity comparison between binary and burst spiking. view at source ↗
Figure 2
Figure 2. Figure 2: QB-LIF neuron dynamics. Multi-channel spike inputs are integrated into the membrane potential U[t]. A learnable quantized scale γ defines an adaptive threshold ladder {γ, 2γ, . . . }, enabling multi-level spike emission and precise soft reset based on the emitted burst. Menbrane Potential Adaptive Scale Ladder Quantization Scale Layer-wise Optimized Trainning Phase Trainning Phase Neuron Output: Neuron Out… view at source ↗
Figure 3
Figure 3. Figure 3: Adaptive activation and surrogate gradient mechanisms in quantized burst spiking neurons. view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of ReLSG-ET and traditional “ac view at source ↗
Figure 5
Figure 5. Figure 5: Burst-level activation distributions in representative layers. Empirical histograms of burst levels for view at source ↗
Figure 6
Figure 6. Figure 6: Layer-wise statistics of adaptive burst spiking. (a) Learned quantized scale view at source ↗
read the original abstract

Binary spike coding enables sparse and event-driven computation in spiking neural networks (SNNs), yet its 1-bit-per-timestep representation fundamentally limits information throughput. This bottleneck becomes increasingly restrictive in deep architectures under short simulation horizons. We propose the Quantized Burst-LIF (QB-LIF) neuron, which reformulates burst spiking as a saturated uniform quantization of membrane potentials with a learnable scale. Instead of relying on predefined multi-threshold structures, QB-LIF treats the quantization scale as a trainable parameter, allowing each layer to autonomously adapt its spiking resolution to the underlying membrane-potential statistics. To preserve hardware efficiency, we introduce an absorbable scale strategy that folds the learned quantized scale into synaptic weights during inference, maintaining a strict accumulate-only (AC) execution paradigm. To enable stable optimization in the discrete multi-level space, we further design ReLSG-ET, a rectified-linear surrogate gradient with exponential tails that sustains gradient flow across burst intervals. Extensive experiments on static (CIFAR-10/100, ImageNet) and event-driven (CIFAR10-DVS, DVS128-Gesture) benchmarks demonstrate that QB-LIF consistently outperforms binary and fixed-burst SNNs, achieving higher accuracy under ultra-low latency while preserving neuromorphic compatibility.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes the Quantized Burst-LIF (QB-LIF) neuron, which reformulates burst spiking as saturated uniform quantization of membrane potentials using a learnable scale parameter per layer. It introduces an absorbable scale strategy that folds the learned scale into synaptic weights to enable strict accumulate-only (AC) inference without multi-threshold hardware, and designs the ReLSG-ET surrogate gradient (rectified linear with exponential tails) to support stable training in the discrete multi-level space. Experiments on CIFAR-10/100, ImageNet, CIFAR10-DVS, and DVS128-Gesture report consistent accuracy gains over binary and fixed-burst SNNs at ultra-low latency while preserving neuromorphic compatibility.

Significance. If the claimed equivalence between training and inference dynamics holds and the accuracy improvements prove robust, the work could meaningfully advance low-latency, high-information-density SNNs by allowing data-adaptive spiking resolution without hardware modifications. The absorbable-scale and ReLSG-ET ideas directly target the information-throughput bottleneck of binary spikes under short simulation horizons.

major comments (2)
  1. [§3.2] §3.2 (absorbable scale strategy): The central claim that folding the learned quantization scale s into synaptic weights W' = s·W preserves spiking behavior and accuracy under strict AC execution is not demonstrated. Because the membrane update is V(t) = leak·V(t-1) + W·input followed by quantization, scaling only W alters the magnitude of V before the quantizer. Without explicit rescaling of the leak coefficient, reset threshold, or saturation bounds, the discrete burst levels at inference will differ from those observed during training with the explicit scale, violating the claimed equivalence. A proof or numerical verification that the output spike counts remain invariant is required.
  2. [§5, Tables 1–4] Experiments (Tables 1–4 and §5): No error bars, standard deviations across runs, or statistical significance tests are reported for the accuracy gains. In addition, the ablation isolating the contribution of the learnable scale versus the ReLSG-ET surrogate is absent. These omissions make it impossible to determine whether the reported outperformance is load-bearing for the proposed method or sensitive to random seeds and hyper-parameters.
minor comments (2)
  1. [§3.3] The derivation or explicit functional form of the ReLSG-ET surrogate gradient is not provided; only its qualitative description appears. Adding the mathematical definition and a brief analysis of gradient magnitude across burst intervals would improve reproducibility.
  2. [§3.1] Notation for the quantization function and burst levels is introduced without a clear equation reference in the main text; readers must infer the exact saturation bounds from the abstract description.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and insightful comments on our manuscript. We address each major comment point by point below and outline the revisions we will make to strengthen the paper.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (absorbable scale strategy): The central claim that folding the learned quantization scale s into synaptic weights W' = s·W preserves spiking behavior and accuracy under strict AC execution is not demonstrated. Because the membrane update is V(t) = leak·V(t-1) + W·input followed by quantization, scaling only W alters the magnitude of V before the quantizer. Without explicit rescaling of the leak coefficient, reset threshold, or saturation bounds, the discrete burst levels at inference will differ from those observed during training with the explicit scale, violating the claimed equivalence. A proof or numerical verification that the output spike counts remain invariant is required.

    Authors: We appreciate the referee's careful analysis of the absorbable scale strategy. Upon re-examination, we recognize that the manuscript presents the folding W' = s·W but does not explicitly derive or verify the invariance of spike counts when the leak factor, thresholds, and saturation bounds are left unscaled. In the revised version we will add a formal proof under the standard LIF assumptions used in the paper (leak = 1 and reset to zero after each burst emission) showing that the quantized output levels and resulting spike counts are identical before and after absorption. We will also include a small-scale numerical verification (a 2-layer network on a toy dataset) that confirms identical spike trains and accuracy at inference. revision: yes

  2. Referee: [§5, Tables 1–4] Experiments (Tables 1–4 and §5): No error bars, standard deviations across runs, or statistical significance tests are reported for the accuracy gains. In addition, the ablation isolating the contribution of the learnable scale versus the ReLSG-ET surrogate is absent. These omissions make it impossible to determine whether the reported outperformance is load-bearing for the proposed method or sensitive to random seeds and hyper-parameters.

    Authors: We agree that the absence of variability measures and targeted ablations limits the strength of the experimental claims. In the revision we will (i) rerun all reported experiments with five independent random seeds, reporting mean accuracy ± standard deviation in Tables 1–4 together with paired t-test p-values against the strongest baseline, and (ii) add a dedicated ablation subsection that isolates the learnable scale (fixed vs. trainable) while keeping the ReLSG-ET surrogate fixed, and conversely isolates the surrogate while keeping the scale fixed. These additions will clarify the individual contributions of each component. revision: yes

Circularity Check

0 steps flagged

No circularity: learnable scale and absorbable folding are independent training choices

full rationale

The abstract presents QB-LIF as introducing a trainable quantization scale for membrane potentials and a post-training folding into weights for AC inference. No equations, predictions, or results are shown to reduce by construction to fitted inputs or prior self-citations. The ReLSG-ET surrogate and benchmark gains are described as empirical outcomes of optimization, not tautological redefinitions. The derivation chain remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 0 axioms · 0 invented entities

The central claim rests on treating the quantization scale as a trainable parameter and on the absorbability of that scale into weights; both are introduced without upstream derivation.

free parameters (1)
  • quantization scale
    Per-layer trainable parameter that sets the step size of the saturated uniform quantization of membrane potential.

pith-pipeline@v0.9.0 · 5542 in / 1085 out tokens · 32754 ms · 2026-05-07T16:55:21.105606+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 2 canonical work pages · 1 internal anchor

  1. [1]

    Chinya, Y.Cao, S

    M.Davies, N.Srinivasa, T.-H.Lin, G. Chinya, Y.Cao, S. H. Choday, G. Dimou, P. Joshi, N. Imam, S. Jain, et al., Loihi: A neuromorphic manycore processor with on-chip learning, IEEE Micro 38 (1) (2018) 82– 99. 11

    2018

  2. [2]

    J. Pei, L. Deng, S. Song, M. Zhao, Y. Zhang, S. Wu, G. Wang, Z. Zou, Z. Wu, W. He, et al., Towards artificial general intelligence with hybrid Tianjic chip architecture, Nature 572 (7767) (2019) 106–111

    2019

  3. [3]

    K. Roy, A. Jaiswal, P. Panda, Towards spike-based machine intelligence with neuromorphic computing, Nature 575 (7784) (2019) 607–617

    2019

  4. [4]

    Y. Guo, Y. Chen, X. Liu, W. Peng, Y. Zhang, X. Huang, Z. Ma, Ternary Spike: Learning ternary spikes for spiking neural networks, in: Proc. of AAAI, Vol. 38, 2024, pp. 12244–12252

    2024

  5. [5]

    L. Feng, Q. Liu, H. Tang, D. Ma, G. Pan, Multi-level Firing with Spiking DS-ResNet: Enabling better and deeper directly-trained spiking neural networks, in: Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence (IJCAI), 2022

    2022

  6. [6]

    X. Luo, M. Yao, Y. Chou, B. Xu, G. Li, Integer- Valued Training and Spike-Driven Inference Spiking Neural Network for High-Performance and Energy- Efficient Object Detection, in: European Conference on Computer Vision (ECCV), Springer, 2024

    2024

  7. [7]

    S. Deng, S. Gu, Optimal Conversion of Conventional Artificial Neural Networks to Spiking Neural Net- works, in: Proc. of ICLR, 2021

    2021

  8. [8]

    T. Bu, W. Fang, J. Ding, P. DAI, Z. Yu, T. Huang, Optimal ANN-SNN Conversion for High-accuracy and Ultra-low-latency Spiking Neural Networks, in: Proc. of ICLR, 2022

    2022

  9. [9]

    Y. Li, S. Deng, X. Dong, R. Gong, S. Gu, A Free Lunch from ANN: Towards efficient, accurate spiking neural networks calibration, in: Proc. of ICML, 2021, pp. 6316–6325

    2021

  10. [10]

    E. O. Neftci, H. Mostafa, F. Zenke, Surrogate gra- dient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks, IEEE Signal Processing Magazine 36 (6) (2019) 51–63

    2019

  11. [11]

    C. Lee, S. S. Sarwar, P. Panda, G. Srinivasan, K. Roy, Enabling spike-based backpropagation for training deep neural network architectures, Frontiers in Neu- roscience 14 (2020) 497482

    2020

  12. [12]

    Y. Wu, L. Deng, G. Li, J. Zhu, L. Shi, Spatio-temporal backpropagation for training high- performance spiking neural networks, Frontiers in neuroscience 12 (2018) 331

    2018

  13. [13]

    X. Wang, M. Zhong, H. Cheng, J. Xie, Y. Zhou, J. Ren, M. Liu, Spikegoogle: Spiking neural net- works with GoogLeNet-like inception module, CAAI Transactions on Intelligence Technology 7 (3) (2022) 492–502.doi:10.1049/cit2.12082

    work page doi:10.1049/cit2.12082 2022

  14. [14]

    Y.Guo, Y.Chen, L.Zhang, Y.Wang, X.Liu, X.Tong, Y. Ou, X. Huang, Z. Ma, Reducing information loss for spiking neural networks, in: European Conference on Computer Vision, Springer, 2022, pp. 36–52

    2022

  15. [15]

    Y. Guo, Y. Chen, L. Zhang, X. Liu, Y. Wang, X. Huang, Z. Ma, IM-loss: information maximization loss for spiking neural networks, Advances in Neural Information Processing Systems 35 (2022) 156–166

    2022

  16. [16]

    Y. Guo, X. Tong, Y. Chen, L. Zhang, X. Liu, Z. Ma, X. Huang, RecDis-SNN: Rectifying membrane poten- tial distribution for directly training spiking neural networks, in: Proc. of CVPR, 2022, pp. 326–335

    2022

  17. [17]

    M. Yao, X. Qiu, T. Hu, J. Hu, Y. Chou, K. Tian, J. Liao, L. Leng, B. Xu, G. Li, Scaling Spike-driven Transformer with Efficient Spike Firing Approxima- tion Training, IEEE Transactions on Pattern Analysis and Machine Intelligence (2025)

    2025

  18. [18]

    L. Fan, H. Shen, X. Lian, Y. Li, M. Yao, G. Li, D. Hu, A multisynaptic spiking neuron for simulta- neously encoding spatiotemporal dynamics, Nature Communications 16 (1) (2025) 7155

    2025

  19. [19]

    S. Park, S. Kim, H. Choe, S. Yoon, Fast and efficient information transmission with burst spikes in deep spiking neural networks, in: Proc. of the 56th Annual Design Automation Conference, 2019, pp. 1–6

    2019

  20. [20]

    Y. Li, Y. Zeng, Efficient and Accurate Conversion of Spiking Neural Network with Burst Spikes, in: Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, International Joint Conferences on Artificial Intelligence Organiza- tion, Vienna, Austria, 2022, pp. 2485–2491

    2022

  21. [21]

    Y. Lan, Y. Zhang, X. Ma, Y. Qu, Y. Fu, Efficient converted spiking neural network for 3d and 2d clas- sification, in: Proc. of ICCV, 2023, pp. 9211–9220

    2023

  22. [22]

    Bellec, D

    G. Bellec, D. Salaj, A. Subramoney, R. Legenstein, W. Maass, Long short-term memory and learning- to-learn in networks of spiking neurons, Advances in neural information processing systems 31 (2018)

    2018

  23. [23]

    W. Fang, Z. Yu, Y. Chen, T. Huang, T. Masquelier, Y. Tian, Deep residual learning in spiking neural networks, in: NeurIPS, 2021

    2021

  24. [24]

    T. M. Cover, Elements of Information Theory, 1999

    1999

  25. [25]

    B. Han, G. Srinivasan, K. Roy, RMP-SNN: Residual membrane potential neuron for enabling deeper high- accuracy and low-latency spiking neural network, in: Proc. of CVPR, 2020, pp. 13558–13567

    2020

  26. [26]

    Y. Li, Y. Zeng, Efficient and Accurate Conversion of Spiking Neural Network with Burst Spikes, in: Proc. of IJCAI, 2022. 12

    2022

  27. [27]

    X. Yao, F. Li, Z. Mo, J. Cheng, GLIF: A Unified Gated Leaky Integrate-and-Fire Neuron for Spiking Neural Networks, in: Proc. of NeurIPS, 2022

    2022

  28. [28]

    S. Deng, Y. Li, S. Zhang, S. Gu, Temporal Efficient Training of Spiking Neural Network via Gradient Re-weighting, in: Proc. of ICLR, 2022

    2022

  29. [29]

    Zhang, K

    T. Zhang, K. Yu, J. Zhang, H. Wang, DA-LIF: Dual Adaptive Leaky Integrate-and-Fire Model for Deep Spiking Neural Networks, in: Proc. of ICASSP, 2025, pp. 1–5

    2025

  30. [30]

    K. Yu, T. Zhang, H. Wang, Q. Xu, FSTA- SNN:Frequency-Based Spatial-Temporal Attention Module for Spiking Neural Networks, Vol. 39, 2025, pp. 22227–22235

    2025

  31. [31]

    Zhang, K

    T. Zhang, K. Yu, X. Zhong, H. Wang, Q. Xu, Q. Zhang, STAA-SNN: Spatial-Temporal Attention Aggregator for Spiking Neural Networks, in: Proc. of CVPR, 2025, pp. 13959–13969

    2025

  32. [32]

    Krizhevsky, V

    A. Krizhevsky, V. Nair, G. Hinton, CIFAR-10 Dataset, Canadian Institute for Advanced Research (2009)

    2009

  33. [33]

    J. Deng, W. Dong, R. Socher, L. Li, K. Li, L. Fei- Fei, ImageNet: A Large-Scale Hierarchical Image Database, in: Proc. of CVPR, 2009, pp. 248–255

    2009

  34. [34]

    H. Li, H. Liu, X. Ji, G. Li, L. Shi, CIFAR10-DVS: An Event-Stream Dataset for Object Classification, Frontiers in Neuroscience 11 (2017)

    2017

  35. [35]

    A. Amir, B. Taba, D. Berg, T. Melano, J. McKinstry, C. Di Nolfo, T. Nayak, A. Andreopoulos, G. Garreau, M. Mendoza, J. Kusnitz, M. Debole, S. Esser, T. Del- bruck, M. Flickner, D. Modha, A Low Power, Fully Event-Based Gesture Recognition System, in: Proc. of CVPR, 2017, pp. 7388–7397

    2017

  36. [36]

    Zheng, Y

    H. Zheng, Y. Wu, L. Deng, Y. Hu, G. Li, Going Deeper with Directly-Trained Larger Spiking Neural Networks, in: Proc.ofAAAI,Vol.35, 2021, pp.11062– 11070

    2021

  37. [37]

    Rathi, K

    N. Rathi, K. Roy, DIET-SNN: A low-latency spiking neural network with direct input encoding and leak- age and threshold optimization, IEEE Transactions on Neural Networks and Learning Systems 34 (6) (2023) 3174–3182

    2023

  38. [38]

    E. D. Cubuk, B. Zoph, D. Mane, V. Vasudevan, Q. V. Le, Autoaugment: Learning augmentation strategies from data, in: Proc. of CVPR, 2019, pp. 113–123

    2019

  39. [39]

    Improved Regularization of Convolutional Neural Networks with Cutout

    T. DeVries, G. W. Taylor, Improved regularization of convolutional neural networks with cutout, arXiv preprint arXiv:1708.04552 (2017)

    work page internal anchor Pith review arXiv 2017

  40. [40]

    Rathi, G

    N. Rathi, G. Srinivasan, P. Panda, K. Roy, Enabling Deep Spiking Neural Networks with Hybrid Conver- sion and Spike Timing Dependent Backpropagation, in: Proc. of ICLR, 2020

    2020

  41. [41]

    Z. Wang, Y. Fang, J. Cao, H. Ren, R. Xu, Adaptive Calibration: A Unified Conversion Framework of Spiking Neural Networks, Proc. of AAAI 39 (2) (2025) 1583–1591

    2025

  42. [42]

    Huang, Z

    Y.Guo, L.Zhang, Y.Chen, X.Tong, X.Liu, Y.Wang, X. Huang, Z. Ma, Real Spike: Learning real-valued spikes for spiking neural networks, in: Proc. of ECCV, 2022, pp. 52–68

    2022

  43. [43]

    Y. Guo, X. Liu, Y. Chen, L. Zhang, W. Peng, Y. Zhang, X. Huang, Z. Ma, RMP-Loss: Regular- izing Membrane Potential Distribution for Spiking Neural Networks, in: Proc. of ICCV, 2023

    2023

  44. [44]

    Y. Guo, Y. Zhang, Y. Chen, W. Peng, X. Liu, L. Zhang, X. Huang, Z. Ma, Membrane Potential Batch Normalization for Spiking Neural Networks, in: Proc. of ICCV, 2023

    2023

  45. [45]

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

    2024

  46. [46]

    W. Fang, Z. Yu, Y. Chen, T. Masquelier, T. Huang, Y. Tian, Incorporating Learnable Membrane Time Constant to Enhance Learning of Spiking Neural Networks, in: Proc. of ICCV, 2021, pp. 2661–2671

    2021

  47. [47]

    M. Yao, G. Zhao, H. Zhang, Y. Hu, L. Deng, Y. Tian, B. Xu, G. Li, Attention Spiking Neural Networks, IEEE Transactions on Pattern Analysis and Machine Intelligence 45 (8) (2023) 9393–9410

    2023

  48. [48]

    M. Yao, J. Hu, G. Zhao, Y. Wang, Z. Zhang, B. Xu, G. Li, Inherent Redundancy in Spiking Neural Net- works, in: Proc. of ICCV, 2023, pp. 16924–16934

    2023

  49. [49]

    Y. Hu, H. Tang, G. Pan, Spiking deep residual net- works, IEEE Transactions on Neural Networks and Learning Systems 34 (8) (2021) 5200–5205. 13

    2021