[论文解读] Spiking Layer-Adaptive Magnitude-based Pruning
SLAMP 将面向时间的输出失真建模引入层自适应幅值剪枝,能够在保持准确性与时间保真度的前提下进行更积极的剪枝。
Spiking Neural Networks (SNNs) provide energy-efficient computation but their deployment is constrained by dense connectivity and high spiking operation costs. Existing magnitude-based pruning strategies, when naively applied to SNNs, fail to account for temporal accumulation, non-uniform timestep contributions, and membrane stability, often leading to severe performance degradation. This paper proposes Spiking Layer-Adaptive Magnitude-based Pruning (SLAMP), a theory-guided pruning framework that generalizes layer-adaptive magnitude pruning to temporal SNNs by explicitly controlling worst-case output distortion across layers and timesteps. SLAMP formulates sparsity allocation as a temporal distortion-constrained optimization problem, yielding time-aware layer importance scores that reduce to conventional layer-adaptive pruning in single-timestep limit. An efficient two-stage procedure is derived, combining temporal score estimation, global sparsity allocation, and magnitude pruning with retraining for stability recovery. Experiments on CIFAR10, CIFAR100, and the event-based CIFAR10-DVS datasets demonstrate that SLAMP achieves substantial connectivity and spiking operation reductions while preserving accuracy, enabling efficient and deployable SNN inference.
研究动机与目标
- 通过在不牺牲时间保真度的前提下 reducing connectivity 与 spike operations 实现 SNN 的能效部署
- 将 Layer-Adaptive Magnitude-based Pruning (LAMP) 扩展到时序 SNN,考虑时间累积效应
- 提出一个理论框架,在剪枝下对最坏情况的时序输出失真进行界定
- 开发一个高效的两阶段剪枝与再训练流程,并具备稳定性恢复能力
提出的方法
- 提出一个在 T 个时间步内聚合贡献的时序失真约束剪枝目标函数(Eq. 3)
- 定义一个低失真剪枝预算约束以界定累积的输出扰动(Eq. 4)
- 计算时间感知的层重要性分数 R^i,反映累积的前突触活动(Eq. 3)
- 在失真约束下推导剪枝掩码 M^i 与被剪枝权重 f6W^i(Eq. 5 与 6)
- 提供一个迭代的剪枝–再训练循环:训练收敛后按时间分数剪枝,再细调以恢复膜稳定性(Algorithm 1)
- 证明在单时间步极限下,R^i 收敛为标准 LAMP(Eq. 8)
实验结果
研究问题
- RQ1如何将剪枝公式化以显式控制 SNNs 的时序失真?
- RQ2时间感知的层重要性分数是否能够实现全局稀疏、内存高效的 SNN 而不损失时序保真?
- RQ3时序剪枝对静态与事件驱动数据集的准确性、连通性和尖峰操作次数有何影响?
- RQ4两阶段剪枝-微调工作流是否能够在强剪枝下产出稳定、可部署的 SNN?
主要发现
| 数据集 | 剪枝方法 | 体系结构 | T | Acc. Diff(%) | Top-1 Acc.(%) | Conn.(%) | Param. (M) | SOPs (M) |
|---|---|---|---|---|---|---|---|---|
| CIFAR10 | ADMM | 7 Conv, 2 FC | 8 | -0.13 | 90.19 | 25.03 | 15.54 | - |
| CIFAR10 | Grad R | 6 Conv, 2 FC | 8 | -0.30 | 92.54 | 36.72 | 10.43 | - |
| CIFAR10 | ESLSNN | ResNet19 | 2 | -1.70 | 91.09 | 50.00 | 6.30 | 180.56 |
| CIFAR10 | STDS | 6 Conv, 2 FC | 8 | -0.35 | 92.49 | 11.33 | 1.71 | 147.22 |
| CIFAR10 | UPR | 6 Conv, 2 FC | 8 | -0.79 | 92.05 | 1.16 | 9.56 | 16.47 |
| CIFAR10 | SLAMP (Ours) | ResNet19 | 2 | +1.23 | 92.42 | 40.00 | 4.47 | 112.35 |
| CIFAR100 | ESLSNN | ResNet19 | 2 | -0.99 | 73.48 | 50.00 | 6.32 | 186.25 |
| CIFAR100 | UPR | SEW ResNet18 | 4 | -4.75 | 69.41 | 2.48 | - | 6.79 |
| CIFAR100 | SLAMP (Ours) | ResNet19 | 2 | +0.73 | 69.85 | 25.00 | 4.47 | 115.72 |
| CIFAR10-DVS | ESLSNN | VGGSNN | 10 | -0.28 | 78.30 | 10.00 | 0.92 | 129.64 |
| CIFAR10-DVS | STDS | VGGSNN | 10 | -2.60 | 79.80 | 4.67 | 0.24 | 38.85 |
| CIFAR10-DVS | UPR | VGGSNN | 10 | -0.50 | 78.30 | 0.77 | 1.81 | 6.75 |
| CIFAR10-DVS | SLAMP (Ours) | VGGSNN | 5 | +0.09 | 78.21 | 40.00 | 3.60 | 105.45 |
- SLAMP 在 CIFAR10、CIFAR100 与 CIFAR10-DVS 上实现了显著的连通性和尖峰操作减少。
- 在 CIFAR10 上,剪枝到 40% 连通性时,SLAMP 的 Top-1 准确率提升了 +1.23%。
- 在 CIFAR100 上,在 25% 连通性下达到 69.85% 的 Top-1 时保持精度并实现 0.73% 的增益(在 69.85% Top-1)。
- 在 CIFAR10-DVS 上,SLAMP 以 40% 连通性保持事件驱动的准确性(+0.09% Top-1)。
- 剪枝后进行微调显著提升准确性并降低膜方差,尤其是在低连通性时。
- 该方法在最终轮次将连通性降低到 10% 或更低,同时保持时序保真度。
更好的研究,从现在开始
从论文设计到论文写作,大幅缩短您的研究时间。
无需绑定信用卡
本解读由 AI 生成,并经人工编辑审核。