Back

RecDis-SNN: Rectifying Membrane Potential Distribution for Directly Training Spiking Neural Networks

Yufei GuoXinyi TongYuanpei ChenLiwen ZhangXiaode LiuZhe MaXuhui Huang

article2022CVPR97 citations

Proposes a novel distribution loss that penalizes membrane potential shifts and reduces quantization errors during training, enabling direct optimization of deeper and more accurate spiking neural networks with fewer timesteps and zero inference overhead.

Listen

Modern artificial intelligence requires substantial computing power and energy, creating severe bottlenecks for edge devices and latency-critical systems. Brain-inspired Spiking Neural Networks offer an energy-efficient alternative by communicating through sparse, binary spike events rather than high-precision continuous values. However, directly training deep spiking models using gradient descent remains difficult because continuous neural voltages, known as membrane potentials, shift undesirably during information processing. These distribution shifts trigger three primary failures: degeneration, where neurons emit identical signals and destroy information flow; saturation, where gradients drop to zero and halt optimization; and gradient mismatch, where mathematical approximations used during back-propagation accumulate substantial estimation errors.

The article develops and evaluates a targeted training framework called RecDis-SNN to rectify these internal voltage distributions. Rather than adding complex model parameters or runtime normalization layers, the approach introduces an explicit regularization loss function during training to penalize degeneration, saturation, and gradient mismatch simultaneously.

The researchers assessed the method through comprehensive classification experiments across standard visual benchmarks, including CIFAR-10, CIFAR-100, and ImageNet, as well as a neuromorphic event-stream dataset, DVS-CIFAR10. They tested multiple deep network backbones, such as CIFARNet, VGG-16, ResNet-19, and ResNet-34, analyzing classification accuracy, latency requirements across discrete timesteps, training computational overhead, and internal quantization error.

The evaluation revealed several critical findings. First, the proposed regularizer prevented optimization breakdown and allowed randomly initialized networks to converge successfully, boosting standalone accuracy on CIFAR-10 by up to 3.11 percentage points over unregularized baselines. Second, the framework achieved state-of-the-art accuracy across all evaluated benchmarks while reducing latency; on CIFAR-10, it attained 95.55% accuracy within only 6 timesteps, and on ImageNet, it achieved 67.33% accuracy. Third, on neuromorphic event-stream data, the approach demonstrated substantial gains, outperforming prior state-of-the-art models on DVS-CIFAR10 by 4.62 to 6.80 percentage points across matching architectures. Fourth, the loss function reduced average internal quantization error by roughly 15% to 26% compared to threshold-dependent batch normalization because it naturally shapes internal voltages into a bimodal distribution. Finally, these improvements incurred only a modest training time overhead of under 17% and introduced zero extra computational operations during operational deployment.

These results establish that internal voltage rectification resolves fundamental optimization barriers in deep spiking models. By addressing gradient estimation mismatch and quantization error in addition to standard variance shifts, the framework enables high-accuracy neuromorphic models that require fewer time steps to process information. This significantly lowers runtime energy consumption and computational latency without introducing inference penalties.

Engineering teams deploying neural networks to energy-constrained hardware should consider integrating membrane potential distribution loss directly into their training pipelines. The method can serve as a standalone regularizer or be combined with existing batch normalization techniques for further performance gains. For future research, teams should explore applying this distribution control framework to other complex vision and temporal tasks beyond standard image classification.

The findings are supported by consistent multi-trial experiments across varied benchmark datasets and architectures. Nevertheless, potential users should note that the mathematical loss formulation assumes internal neuron voltages approximate a Gaussian distribution, and experimental evaluations were centered on visual classification tasks. Organizations should conduct targeted validation on their specific domain-specific tasks and neuromorphic hardware platforms prior to full-scale production deployment.

China Aerospace Science and Industry CorporationX LABenergy-efficientinferenceoptimizationRecDis-SNNdirectly training spiking neural networksmembrane potential distribution
Guo et al (2022).pdf
Cover for RecDis-SNN: Rectifying Membrane Potential Distribution for Directly Training Spiking Neural Networks

Abstract

The brain-inspired and event-driven Spiking Neural Network (SNN) aiming at mimicking the synaptic activity of biological neurons has received increasing attention. It transmits binary spike signals between network units when the membrane potential exceeds the firing threshold. This bio-mimetic mechanism of SNN appears energy-efficient with its power sparsity and asynchronous operations on spike events. Unfortunately, with the propagation of binary spikes, the distribution of membrane potential will shift, leading to degeneration, saturation, and gradient mismatch problems, which would be disadvantageous to the network optimization and convergence. Such undesired shifts would prevent the SNN from performing well and going deep. To tackle these problems, we attempt to rectify the membrane potential distribution (MPD) by designing a novel distribution loss, MPD-Loss, which can explicitly penalize the undesired shifts without introducing any additional operations in the inference phase. Moreover, the proposed method can also mitigate the quantization error in SNNs, which is usually ignored in other works. Experimental results demonstrate that the proposed method can directly train a deeper, larger, and better-performing SNN within fewer timesteps.

Topics

energy-efficientinferenceoptimizationspiking neural networksnetwork architecturerecurrent neural networksfeedforward neural networksquantizers

Keywords

RecDis-SNNdirectly training spiking neural networksmembrane potential distributionneuromorphic datasetsspike feature mapdeep SNNsspike activity functionneuron model

Table of Contents

  • 1. Introduction
  • 2. Related Work
  • 3. Materials and Methodology
  • 3.1. Spiking Neural Networks
  • 3.2. Rectifying Membrane Potential Distribution
  • 3.3. Analysis and Discussion
  • 4. Experiment
  • 4.1. Ablation Study for MPD-Loss
  • 4.2. Ablation Study for Training Cost
  • 4.3. Study for Quantization Error Reduction
  • 4.4. Comparison with The Normalization
  • 4.5. Comparisons with Other Methods
  • 5. Conclusion
  • References

Knowls

  1. Knowl 1 — Undesired Membrane Potential Distribution Shifts in Directly Trained SNNs

    definition

    Let U∈RB×T×W×HU \in \mathbb{R}^{B \times T \times W \times H} denote the 4D tensor of neuron membrane potentials across batch size BB, timesteps TT, feature map width WW, and feature map height HH for a specific channel in a given layer of a Spiking Neural Network (SNN). Let U(q)U_{(q)} denote the qq-quantile (0≤q≤10 \le q \le 1) of the sorted elements of UU, and let VthV_{th} denote the neuron firing threshold (typically Vth=0.5V_{th} = 0.5). When training SNNs directly using surrogate gradient descent with a standard rectangular surrogate derivative active in the unit interval [0,1][0, 1], three distinct distribution shifts degrade network optimization:

    1. Degeneration: U(0)≥VthU_{(0)} \ge V_{th} or U(1)≤VthU_{(1)} \le V_{th}. All neuron potentials across the channel are either entirely above or entirely below the firing threshold, causing neurons to emit homogeneous binary spike trains (all 1s or all 0s), which destroys feature representations and prevents information flow.
    2. Saturation: U(0)≥1U_{(0)} \ge 1 or U(1)≤0U_{(1)} \le 0. Membrane potentials fall completely outside the surrogate gradient non-zero support interval [0,1][0, 1], forcing backpropagated surrogate gradients to zero and halting parameter updates.
    3. Gradient Mismatch: 0≤U(0)≤10 \le U_{(0)} \le 1 and 0≤U(1)≤10 \le U_{(1)} \le 1. Membrane potentials fall entirely within [0,1][0, 1]. In this regime, surrogate gradient approximation is applied across all activations, magnifying the cumulative discrepancy between the surrogate continuous derivative and the true non-differentiable Dirac delta spike derivative.

  2. Knowl 2 — Differentiable Membrane Potential Distribution Loss (MPD-Loss) Formulation

    equation

    To explicitly penalize degeneration, saturation, and gradient mismatch during backpropagation, the membrane potentials UU of a channel are modeled as a Gaussian distribution N(μ,σ2)\mathcal{N}(\mu, \sigma^2), where μ\mu and σ\sigma are the sample mean and standard deviation of UU. Under this assumption, the lower and upper ϵ\epsilon-quantiles are parameterized as U(ϵ)≈μ−kϵσU_{(\epsilon)} \approx \mu - k_\epsilon \sigma and U(1−ϵ)≈μ+kϵσU_{(1-\epsilon)} \approx \mu + k_\epsilon \sigma, where kϵk_\epsilon is a constant quantile scaling factor. The individual differentiable distribution loss terms are defined using the hinge operator (x)+=max⁡(x,0)(x)_+ = \max(x, 0) as:

    LD=[(μ−kϵ,Dσ−Vth)+]2+[(Vth−μ−kϵ,Dσ)+]2\mathcal{L}_D = [(\mu - k_{\epsilon, D}\sigma - V_{th})_+]^2 + [(V_{th} - \mu - k_{\epsilon, D}\sigma)_+]^2

    LS=[(μ−kϵ,Sσ−1)+]2+[−μ−kϵ,Sσ)+]2\mathcal{L}_S = [(\mu - k_{\epsilon, S}\sigma - 1)_+]^2 + [-\mu - k_{\epsilon, S}\sigma)_+]^2

    LM=[max⁡(μ−kϵ,Mσ,1−μ−kϵ,Mσ)+]2\mathcal{L}_M = [\max(\mu - k_{\epsilon, M}\sigma, 1 - \mu - k_{\epsilon, M}\sigma)_+]^2

    where Vth=0.5V_{th} = 0.5 is the firing threshold. The hyperparameters penalizing degeneration, saturation, and gradient mismatch are set to kϵ,D=1.0k_{\epsilon, D} = 1.0, kϵ,S=0.25k_{\epsilon, S} = 0.25, and kϵ,M=1.25k_{\epsilon, M} = 1.25, respectively.

  3. Knowl 3 — RecDis-SNN Optimization Objective and Layerwise Regularization

    model/method

    In RecDis-SNN, the composite membrane potential distribution loss LMPDb\mathcal{L}_{MPD}^b for a mini-batch bb is computed across all hidden layers l∈{1,…,L−1}l \in \{1, \dots, L-1\} and all corresponding channels c∈{1,…,Cl}c \in \{1, \dots, C_l\} (excluding the non-spiking output layer, which only accumulates synaptic inputs without leakage or spiking threshold):

    LMPDb=1L−1∑l=1L−1(1Cl∑c=1Cl(LDb,l,c+LSb,l,c+LMb,l,c))\mathcal{L}_{MPD}^b = \frac{1}{L-1} \sum_{l=1}^{L-1} \left( \frac{1}{C_l} \sum_{c=1}^{C_l} \left( \mathcal{L}_D^{b,l,c} + \mathcal{L}_S^{b,l,c} + \mathcal{L}_M^{b,l,c} \right) \right)

    where LDb,l,c\mathcal{L}_D^{b,l,c}, LSb,l,c\mathcal{L}_S^{b,l,c}, and LMb,l,c\mathcal{L}_M^{b,l,c} are the degeneration, saturation, and gradient mismatch losses for batch bb, layer ll, and channel cc.

    The total training loss LCE−MPDb\mathcal{L}_{CE-MPD}^b combines temporal mean cross-entropy loss LCEb\mathcal{L}_{CE}^b and the MPD regularization term:

    LCE−MPDb=LCEb+λLMPDb\mathcal{L}_{CE-MPD}^b = \mathcal{L}_{CE}^b + \lambda \mathcal{L}_{MPD}^b

    with balance coefficient λ=2\lambda = 2. During network inference, LMPD\mathcal{L}_{MPD} is omitted entirely, maintaining standard SNN spike inference without adding auxiliary normalization layers or inference computational overhead.

  4. Knowl 4 — SNN Classification Training with MPD-Loss

    algorithm

    The training procedure for an LL-layer SNN with MPD-Loss over one training epoch is structured as follows:

    Input: SNN with LL layers, Timestep TT, Firing threshold Vth=0.5V_{th}=0.5, Trade-off weight λ=2\lambda=2, Hyperparameters kϵ,D=1.0,kϵ,S=0.25,kϵ,M=1.25k_{\epsilon, D}=1.0, k_{\epsilon, S}=0.25, k_{\epsilon, M}=1.25, Training iterations ItrainI_{train}, Validation iterations IvalI_{val}
    Output: Trained SNN parameters
    for iteration i=1,…,Itraini = 1, \dots, I_{train} do
        Get mini-batch training input and ground-truth labels YiY^i
        Forward propagate over TT timesteps to obtain network outputs Oi(t)O^i(t) and channel membrane potentials Ui,l,cU^{i,l,c}
        Compute classification cross-entropy loss: LCE=1T∑t=1TLCE(Oi(t),Yi)\mathcal{L}_{CE} = \frac{1}{T} \sum_{t=1}^T \mathcal{L}_{CE}(O^i(t), Y^i)
        for layer l=1,…,L−1l = 1, \dots, L-1 do
            for channel c=1,…,Clc = 1, \dots, C_l do
                Compute mean μ\mu and standard deviation σ\sigma of Ui,l,cU^{i,l,c}
                Compute LDi,l,c=[(μ−kϵ,Dσ−Vth)+]2+[(Vth−μ−kϵ,Dσ)+]2\mathcal{L}_D^{i,l,c} = [(\mu - k_{\epsilon, D}\sigma - V_{th})_+]^2 + [(V_{th} - \mu - k_{\epsilon, D}\sigma)_+]^2
                Compute LSi,l,c=[(μ−kϵ,Sσ−1)+]2+[−μ−kϵ,Sσ)+]2\mathcal{L}_S^{i,l,c} = [(\mu - k_{\epsilon, S}\sigma - 1)_+]^2 + [-\mu - k_{\epsilon, S}\sigma)_+]^2
                Compute LMi,l,c=[max⁡(μ−kϵ,Mσ,1−μ−kϵ,Mσ)+]2\mathcal{L}_M^{i,l,c} = [\max(\mu - k_{\epsilon, M}\sigma, 1 - \mu - k_{\epsilon, M}\sigma)_+]^2
            end for
        end for
        Compute LMPD=1L−1∑l=1L−11Cl∑c=1Cl(LDi,l,c+LSi,l,c+LMi,l,c)\mathcal{L}_{MPD} = \frac{1}{L-1} \sum_{l=1}^{L-1} \frac{1}{C_l} \sum_{c=1}^{C_l} (\mathcal{L}_D^{i,l,c} + \mathcal{L}_S^{i,l,c} + \mathcal{L}_M^{i,l,c})
        Compute total objective LCE−MPD=LCE+λLMPD\mathcal{L}_{CE-MPD} = \mathcal{L}_{CE} + \lambda \mathcal{L}_{MPD}
        Backpropagate gradients using surrogate gradient do/du=10<u<1do/du = \mathbf{1}_{0 < u < 1} and update network parameters
    end for
    for validation iteration i=1,…,Ivali = 1, \dots, I_{val} do
        Compute temporal mean output Omeani=1T∑t=1TOi(t)O_{mean}^i = \frac{1}{T} \sum_{t=1}^T O^i(t)
        Assign predicted class label arg⁡max⁡(Omeani)\arg\max(O_{mean}^i) for evaluation
    end for

  5. Knowl 5 — Quantization Error Reduction via Bimodal Membrane Potential Distribution

    theoretical result

    In SNNs, real-valued membrane potentials uu are quantized into discrete spikes o∈{0,1}o \in \{0, 1\} according to whether uu exceeds the firing threshold Vth=0.5V_{th} = 0.5. The neuron quantization error is defined as (u−o)2(u - o)^2, where o=0o = 0 when u<Vthu < V_{th} and o=1o = 1 when u≥Vthu \ge V_{th}.

    Standard batch normalization techniques (such as threshold-dependent batch normalization, tdBN) constrain the pre-activations into a unimodal distribution centered within the unit interval [0,1][0, 1], which forces large proportions of membrane potentials to reside midway between 0 and 1, maximizing quantization loss.

    In contrast, the gradient mismatch loss LM=[max⁡(U(ϵ),1−U(1−ϵ))+]2\mathcal{L}_M = [\max(U_{(\epsilon)}, 1 - U_{(1-\epsilon)})_+]^2 penalizes potentials that remain tightly concentrated within [0,1][0, 1] and pulls the distribution tails outward. This regularizes the membrane potential distribution across training epochs from an initial unimodal distribution into a bimodal distribution with peaks concentrated near 0 and 1. By clustering membrane potentials closer to the discrete binary values o∈{0,1}o \in \{0, 1\}, the bimodal distribution directly minimizes the quantization error (u−o)2(u - o)^2 while preserving representational capacity.

  6. Knowl 6 — Ablation Study of MPD-Loss Components

    data/table

    The ablation experiments on CIFAR-10 evaluate the standalone and combined effects of degeneration loss LD\mathcal{L}_D, saturation loss LS\mathcal{L}_S, and gradient mismatch loss LM\mathcal{L}_M using CIFARNet and ResNet-19 backbones at T=4T = 4 timesteps:

    Architecture Method Top-1 Accuracy
    CIFARNet None 89.83%
    CIFARNet w/ LD\mathcal{L}_D 91.29%
    CIFARNet w/ LS\mathcal{L}_S 91.01%
    CIFARNet w/ LM\mathcal{L}_M 91.72%
    CIFARNet w/ MPD-Loss 92.08%
    ResNet-19 None 91.23%
    ResNet-19 w/ LD\mathcal{L}_D 93.84%
    ResNet-19 w/ LS\mathcal{L}_S 93.83%
    ResNet-19 w/ LM\mathcal{L}_M 94.04%
    ResNet-19 w/ MPD-Loss 94.34%

    Vanilla models initialized randomly without distribution losses suffer from layer-by-layer spike rate decay toward zero, stalling training convergence. Incorporating any single component (LD\mathcal{L}_D, LS\mathcal{L}_S, or LM\mathcal{L}_M) prevents spike death and improves classification accuracy by 1.18%–1.89% on CIFARNet and 2.60%–2.81% on ResNet-19. Combining all three components into MPD-Loss achieves superimposing performance gains, reaching 92.08% on CIFARNet and 94.34% on ResNet-19.

  7. Knowl 7 — Quantization Error and Computational Overhead Comparisons

    empirical result

    Average quantization error (u−o)2(u - o)^2 measured at the final layer of the first residual block in ResNet-19 (T=4T=4) demonstrates that MPD-Loss consistently reduces quantization error compared to threshold-dependent batch normalization (tdBN):

    • CIFAR-10: 0.66 (MPD-Loss) vs. 0.78 (tdBN)
    • CIFAR-100: 0.64 (MPD-Loss) vs. 0.77 (tdBN)
    • CIFAR10-DVS: 0.61 (MPD-Loss) vs. 0.83 (tdBN)

    Training wall-clock time measured on a single NVIDIA RTX 2080Ti (batch size 64, 1000 epochs, T=4T=4) shows:

    • CIFARNet: Vanilla baseline takes 2524.82 min; tdBN takes 2753.65 min (+9.1%); MPD-Loss takes 2948.88 min (+16.8%).
    • ResNet-19: Vanilla baseline takes 7685.88 min; tdBN takes 8099.43 min (+5.4%); MPD-Loss takes 8648.85 min (+12.5%).

    The extra training cost of MPD-Loss does not exceed 17% relative to vanilla training, while adding zero inference latency since the distribution loss is deactivated during evaluation.

  8. Knowl 8 — Comparison and Combination of MPD-Loss with Batch Normalization

    data/table

    Comparison of directly trained SNNs on CIFAR-10 (T=4T=4) trained from scratch with random weight initialization under different regularization configurations:

    Architecture Method Top-1 Accuracy
    CIFARNet None —
    CIFARNet w/ tdBN 90.69%
    CIFARNet w/ MPD-Loss 92.08%
    CIFARNet w/ tdBN MPD-Loss 92.20%
    ResNet-19 None —
    ResNet-19 w/ tdBN 92.92%
    ResNet-19 w/ MPD-Loss 94.34%
    ResNet-19 w/ tdBN MPD-Loss 95.53%

    Without normalization or distribution loss ("None"), networks initialized from scratch fail to converge. MPD-Loss alone outperforms tdBN alone by +1.39% on CIFARNet (92.08% vs. 90.69%) and by +1.42% on ResNet-19 (94.34% vs. 92.92%). Furthermore, MPD-Loss is compatible with existing SNN normalization schemes: combining tdBN with MPD-Loss yields the highest performance, achieving 92.20% on CIFARNet and 95.53% on ResNet-19.

  9. Knowl 9 — Classification Accuracy on Static Vision Benchmarks (CIFAR-10, CIFAR-100, ImageNet)

    data/table

    Classification accuracy comparisons of RecDis-SNN against conversion (ANN2SNN), hybrid, and direct SNN training methods on CIFAR-10, CIFAR-100, and ImageNet (reported as mean ±\pm standard deviation across 3 runs):

    Dataset Method Type Architecture Timestep (TT) Accuracy
    CIFAR10 RMP-SNN ANN2SNN ResNet-20 2048 91.36%
    CIFAR10 ANN-SNN ANN2SNN CIFARNet 128 90.58%
    CIFAR10 Hybrid Training Hybrid ResNet-20 250 92.22%
    CIFAR10 STBP SNN direct CIFARNet 12 90.53%
    CIFAR10 Spike-basedBP SNN direct ResNet-11 100 90.95%
    CIFAR10 STBP-tdBN SNN direct ResNet-19 6 93.16%
    CIFAR10 PLIF SNN direct PLIFNet 8 93.50%
    CIFAR10 TSSL-BP SNN direct CIFARNet 5 91.41%
    CIFAR10 Diet-SNN SNN direct CIFARNet 5 91.59%
    CIFAR10 RecDis-SNN SNN direct CIFARNet 2 90.58% ±\pm 0.12
    CIFAR10 RecDis-SNN SNN direct CIFARNet 4 92.20% ±\pm 0.10
    CIFAR10 RecDis-SNN SNN direct CIFARNet 6 92.91% ±\pm 0.09
    CIFAR10 RecDis-SNN SNN direct ResNet-19 2 93.64% ±\pm 0.07
    CIFAR10 RecDis-SNN SNN direct ResNet-19 4 95.53% ±\pm 0.05
    CIFAR10 RecDis-SNN SNN direct ResNet-19 6 95.55% ±\pm 0.05
    CIFAR100 RMP-SNN ANN2SNN ResNet-20 2048 67.82%
    CIFAR100 ANN-SNN ANN2SNN VGG-16 128 70.47%
    CIFAR100 BinarySNN BNN2SNN VGG-15 62 63.20%
    CIFAR100 Hybrid Training Hybrid VGG-11 125 67.78%
    CIFAR100 Diet-SNN SNN direct VGG-16 5 69.67%
    CIFAR100 RecDis-SNN SNN direct VGG-16 5 69.88% ±\pm 0.08
    CIFAR100 RecDis-SNN SNN direct ResNet-19 4 74.10% ±\pm 0.13
    ImageNet RMP-SNN ANN2SNN ResNet-34 1024 66.61%
    ImageNet Hybrid Training Hybrid ResNet-34 250 61.48%
    ImageNet STBP-tdBN SNN direct ResNet-34 6 63.72%
    ImageNet PLIF SNN direct ResNet-34 7 67.04%
    ImageNet RecDis-SNN SNN direct ResNet-34 6 67.33% ±\pm 0.10

    RecDis-SNN (combined with tdBN) achieves top-1 accuracies of 95.55% on CIFAR-10 with ResNet-19 (T=6T=6), 74.10% on CIFAR-100 with ResNet-19 (T=4T=4), and 67.33% on ImageNet with ResNet-34 (T=6T=6), outperforming prior direct SNN methods while requiring substantially fewer timesteps.

  10. Knowl 10 — Classification Accuracy on Neuromorphic DVS-CIFAR10 Benchmark

    data/table

    Top-1 accuracy on the event-stream neuromorphic benchmark DVS-CIFAR10 evaluated at T=10T = 10 timesteps over 3 trials:

    Method Architecture Accuracy
    STBP CIFARNet 60.50%
    STBP-tdBN ResNet-19 67.80%
    RecDis-SNN (ours) CIFARNet 67.30% ±\pm 0.05
    RecDis-SNN (ours) ResNet-19 72.42% ±\pm 0.06

    Under identical network architectures and inference timesteps (T=10T = 10), RecDis-SNN achieves 67.30% on CIFARNet (+6.80% over STBP) and 72.42% on ResNet-19 (+4.62% over STBP-tdBN).

Coverage note — None was omitted; all contributed formulations of MPD shifts, differentiable distribution loss derivations, training algorithm, quantization error analysis, and empirical evaluations across static and neuromorphic datasets were included.

References

  1. 1.J. Ba, J. Kiros, and G. Hinton. Layer normalization. arXiv, 07 2016. 3
  2. 2.D. Bahdanau, K. Cho, and Y. Bengio. Neural machine trans- lation by jointly learning to align and translate. arXiv, 2014. 1
  3. 3.P. Blouw, X. Choo, E. Hunsberger, and C. Eliasmith. Bench- marking keyword spotting efficiency on neuromorphic hard- ware. arXiv, 2018. 1
  4. 4.R. Bodo, L. Iulia-Alexandra, Y. Hu, P. Michael, and S. C. Liu. Conversion of continuous-valued deep networks to ef- ficient event-driven networks for image classification. Fron- tiers in Neuroscience, 11:682, 2017. 1
  5. 5.Y. Cao, Y. Chen, and D. Khosla. Spiking deep convolu- tional neural networks for energy-efficient object recogni- tion. International Journal of Computer Vision, 113(1):54– 66, 2015. 1
  6. 6.S. Darabi, M. Belbahri, M. Courbariaux, and V. P. Nia. Reg- ularized binary network training. arXiv, 2018. 3
  7. 7.M. Davies, N. Srinivasa, T. H. Lin, G. Chinya, P. Joshi, A. Lines, A. Wild, H. Wang, and et al. Loihi: A neuromor- phic manycore processor with on-chip learning. IEEE Micro, 38(1):82–99, 2018. 1
  8. 8.J. Deng, W. Dong, R. Socher, L. Li, K. Li, and F. Li. Ima- genet: a large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 06 2009. 6
  9. 9.S. Deng and S. Gu. Optimal conversion of conventional arti- ficial neural networks to spiking neural networks. arXiv, 02 2021. 1, 8
  10. 10.S. Deng, Y. Li, S. Zhang, and S. Gu. Temporal efficient training of spiking neural network via gradient re-weighting. arXiv, 2022. 1
  11. 11.P. Diehl, D. Neil, J. Binas, M. Cook, S. C. Liu, and M. Pfeif- fer. Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing. In 2015 Interna- tional Joint Conference on Neural Networks (IJCNN), pages 1–8, 07 2015. 1, 3
  12. 12.W. Fang, Z. Yu, Y. Chen, T. Huang, Timothée M., and Y. Tian. Deep residual learning in spiking neural networks. arXiv, 2021. 1
  13. 13.W. Fang, Z. yu, Y. Chen, T. Masquelier, T. Huang, and Y. Tian. Incorporating learnable membrane time constant to en- hance learning of spiking neural networks. arXiv, 08 2021. 1, 3, 4, 8
  14. 14.Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detec- tion and semantic segmentation. In 2014 IEEE Conference on Computer Vision and Pattern Recognition(CVPR), pages 580–587, 2014. 1
  15. 15.B. Han, G. Srinivasan, and K. Roy. Rmp-snn: Residual mem- brane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13555–13564, 2020. 1, 8
  16. 16.K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 770– 778, 06 2016. 1, 8
  17. 17.D. Huh and T. Sejnowski. Gradient descent for spiking neu- ral networks. arXiv, 06 2017. 1
  18. 18.T. Hwu, J. Isbell, N. Oros, and J. Krichmar. A self-driving robot using deep convolutional neural networks on neuro- morphic hardware. In 2017 International Joint Conference on Neural Networks (IJCNN), pages 635–641, 05 2017. 1
  19. 19.S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In 2015 Proceedings of the 32nd International Conference on Machine Learning (ICML), 02 2015. 3
  20. 20.Y. Jin, P. Li, and W. Zhang. Hybrid macro/micro level back- propagation for training deep spiking neural networks. In 2018 Proceedings of the 32nd International Conference on Neural Information Processing(NIPS), page 7005–7015, 09 2018. 1
  21. 21.M. Khan, D. Lester, L. Plana, A. Rast, X. Jin, E. Painkras, and S. B. Furber. Spinnaker: Mapping neural networks onto a massively-parallel chip multiprocessor. In 2008 Interna- tional Joint Conference on Neural Networks (IJCNN), pages 2849–2856, 2008. 1
  22. 22.A. Krizhevsky, V. Nair, and G. Hinton. Cifar-10 (canadian institute for advanced research). 2010. 6
  23. 23.C. Lee, S. Sarwar, P. Panda, G. Srinivasan, and K. Roy. En- abling spike-based backpropagation for training deep neural network architectures. Frontiers in Neuroscience, 14:119, 02 2020. 1, 8
  24. 24.J. H. Lee, T. Delbruck, and M. Pfeiffer. Training deep spik- ing neural networks using backpropagation. Frontiers in Neuroscience, 10(508):1662–4548, 2016. 1
  25. 25.H. Li, H. Liu, X. Ji, G. Li, and L. Shi. Cifar10-dvs: An event-stream dataset for object classification. Frontiers in Neuroscience, 11:309, 2017. 6
  26. 26.Y. Li, S. Deng, X. Dong, R. Gong, and S. Gu. A free lunch from ann: Towards efficient, accurate spiking neural net- works calibration. In 2021 Proceedings of the 38th Inter- national Conference on Machine Learning (ICML), volume 139 of Proceedings of Machine Learning Research, pages 6316–6325. PMLR, 18–24 Jul 2021. 1
  27. 27.Y. Li, X. Dong, and W. Wang. Additive powers-of-two quan- tization: An efficient non-uniform discretization for neural networks. arXiv, 2019. 3
  28. 28.Y. Li, R. Gong, X. Tan, Y. Yang, P. Hu, Q. Zhang, F. Yu, W. Wang, and S. Gu. Brecq: Pushing the limit of post-training quantization by block reconstruction. arXiv, 2021. 1
  29. 29.Y Li, Y Guo, S Zhang, S Deng, Y Hai, and S Gu. Differen- tiable spike: Rethinking gradient-descent for training spik- ing neural networks. 2021 Advances in Neural Information Processing Systems (NIPS), 34, 2021. 1
  30. 30.Y. Li, F. Zhu, R. Gong, M. Shen, X. Dong, F. Yu, S. Lu, and S. Gu. Mixmix: All you need for data-free compression are feature and data mixing. In 2021 IEEE International Confer- ence on Computer Vision (ICCV), pages 4410–4419, 2021. 1
  31. 31.W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C. Y. Fu, and A. C. Berg. Ssd: Single shot multibox detector. In 2016 European Conference on Computer Vision(ECCV), pages 21–37, 2016. 1
  32. 32.S. Lu and A. Sengupta. Exploring the connection between bi- nary and spiking neural networks. Frontiers in Neuroscience, 14:535, 2020. 8
  33. 33.P. Merolla, J. Arthur, R. Alvarez-Icaza, A. Cassidy, J. Sawada, F. Akopyan, and et al. A million spiking-neuron in- tegrated circuit with a scalable communication network and interface. Science (New York, N.Y.), 345(6197):668–673, 08 2014. 1
  34. 34.E. Neftci, C. Augustine, S. Paul, and G. Detorakis. Event- driven random backpropagation: Enabling neuromorphic deep learning machines. In 2017 IEEE International Sym- posium on Circuits and Systems (ISCAS), pages 1–4, 2017. 1
  35. 35.Ding R., Chin T., Liu Z., and Marculescu D. Regularizing activation distribution for training binarized deep networks, 2019. 3
  36. 36.M. Rastegari, V. Ordonez, J. Redmon, and A. Farhadi. Xnor- net: Imagenet classification using binary convolutional neu- ral networks. In 2016 European Conference on Computer Vision (ECCV), pages 525–542, 2016. 3
  37. 37.N. Rathi and K. Roy. Diet-snn: Direct input encoding with leakage and threshold optimization in deep spiking neural networks. arXiv, 08 2020. 1, 3, 4, 6, 8
  38. 38.N. Rathi, G. Srinivasan, P. Panda, and K. Roy. Enabling deep spiking neural networks with hybrid conversion and spike timing dependent backpropagation. arXiv, 05 2020. 1, 8
  39. 39.J. Redmon, S. Divvala, R. Girshick, and A. Farhadi. You only look once: Unified, real-time object detection. In 2016 IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), pages 779–788, 2016. 1
  40. 40.K. Roy, A. Jaiswal, and P. Panda. Towards spike-based ma- chine intelligence with neuromorphic computing. Nature, 575(7784):607–617, 2019. 1
  41. 41.A. Samadi, T. P. Lillicrap, and D. B. Tweed. Deep learning with dynamic spiking neurons and fixed feedback weights. Neural Computation, 29(3):578–602, 2017. 1
  42. 42.A. Sengupta, Y. Ye, R. Wang, C. Liu, and K. Roy. Going deeper in spiking neural networks: Vgg and residual archi- tectures. Frontiers in Neuroscience, 13, 2019. 1, 3
  43. 43.D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. V. D. Driessche, and et al. Mastering the game of go with deep neural networks and tree search. Nature, 529:484–489, 2016. 1
  44. 44.K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv, 2014. 1
  45. 45.C. Szegedy, W. Liu, Y. Jia, P. Sermanet, and et al.. Go- ing deeper with convolutions. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1– 9, 2015. 1
  46. 46.M. Volodymyr, K. Koray, S. David, A. A. Rusu, V. Joel, M. G. Bellemare, G. Alex, R. Martin, A. K. Fidjeland, O. Georg, and et al. Human-level control through deep rein- forcement learning. Nature, 518:529–533, 2015. 1
  47. 47.Y. Wu, L. Deng, G. Li, J. Zhu, Y. Xie, and L.P. Shi. Direct training for spiking neural networks: Faster, larger, better. 2019 Proceedings of the AAAI Conference on Artificial In- telligence(AAAI), 33:1311–1318, 07 2019. 3, 4, 6, 7, 8
  48. 48.Y. Wu and K. He. Group normalization. International Jour- nal of Computer Vision, 128:742–755, 03 2020. 3
  49. 49.Y. Wu, D. Lei, G. Li, J. Zhu, and L. Shi. Spatio-temporal backpropagation for training high-performance spiking neu- ral networks. Frontiers in Neuroscience, 12:331, 2018. 1
  50. 50.W.i Zhang and P. Li. Spike-train level backpropagation for training deep recurrent spiking neural networks. In 2019 Proceedings of the International Conference on Neural In- formation Processing(NIPS), 08 2019. 8
  51. 51.X. Zhang, H. Qin, Y. Ding, R. Gong, Q. Yan, R. Tao, Y. Li, F. Yu, and X. Liu. Diversifying sample generation for accurate data-free quantization. In 2021 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 15658–15667, 2021. 1
  52. 52.H. Zheng, Y. Wu, L. Deng, Y. Hu, and G. Li. Going deeper with directly-trained larger spiking neural networks. arXiv, 10 2020. 1, 3, 4, 5, 6, 7, 8

Citation

MLA
Guo, Y., et al. “RecDis-SNN: Rectifying Membrane Potential Distribution for Directly Training Spiking Neural Networks”. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 326–35, https://doi.org/10.1109/CVPR52688.2022.00042.
APA
Guo, Y., Tong, X., Chen, Y., Zhang, L., Liu, X., Ma, Z., & Huang, X. (2022). RecDis-SNN: Rectifying Membrane Potential Distribution for Directly Training Spiking Neural Networks. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 326–335. https://doi.org/10.1109/CVPR52688.2022.00042
Chicago
Guo, Y., X. Tong, Y. Chen, et al. 2022. “RecDis-SNN: Rectifying Membrane Potential Distribution for Directly Training Spiking Neural Networks”. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 326–35. https://doi.org/10.1109/CVPR52688.2022.00042.
Harvard
Guo, Y. et al. (2022) “RecDis-SNN: Rectifying Membrane Potential Distribution for Directly Training Spiking Neural Networks”, 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, pp. 326–335. Available at: https://doi.org/10.1109/CVPR52688.2022.00042.
Vancouver
1. Guo Y, Tong X, Chen Y, Zhang L, Liu X, Ma Z, Huang X (2022) RecDis-SNN: Rectifying Membrane Potential Distribution for Directly Training Spiking Neural Networks. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, pp 326–335

BibTeX

@inproceedings{Guo_2022, title={RecDis-SNN: Rectifying Membrane Potential Distribution for Directly Training Spiking Neural Networks}, url={http://dx.doi.org/10.1109/CVPR52688.2022.00042}, DOI={10.1109/cvpr52688.2022.00042}, booktitle={2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, publisher={IEEE}, author={Guo, Yufei and Tong, Xinyi and Chen, Yuanpei and Zhang, Liwen and Liu, Xiaode and Ma, Zhe and Huang, Xuhui}, year={2022}, month=June, pages={326–335} }
Metadata:Crossref

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF
License: IEEE