BASC : Behavior-Aligned Quantization and Pruning for Low-Bit Spiking Neural Networks

Linliang Chen; Yan Zhong; Xin Liu; Sai Li; Wang Kang

cs.NEAug 12, 2026

BASC : Behavior-Aligned Quantization and Pruning for Low-Bit Spiking Neural Networks

Authors: Linliang Chen, Yan Zhong, Xin Liu, Sai Li, Wang Kang

Organizations: Hangzhou International Innovation Institute, Beihang University, Hangzhou 311115, China · School of Mathematical Sciences, Peking University, Beijing 100871, China · School of Electronic and Information Engineering; Beihang University, Beijing 100191, China · School of Integrated Circuit Science and Engineering, Beihang University, Beijing 100191, China

Abstract

Spiking Neural Networks (SNNs) encode information through binary spikes and compute in an event-driven manner, offering an energy-efficient paradigm for machine intelligence. However, high-performance SNNs incur substantial memory and timestep-wise computation costs that hinder deployment on resource-constrained devices. Quantization and pruning provide complementary routes to reducing these costs, yet both make their decisions with local criteria that overlook temporal task feedback in quantization and inter-channel dependencies in pruning. Consequently, optimizing either criterion can still yield suboptimal compression performance. We refer to this discrepancy as criterion-behavior mismatch and propose Behavior-Aligned SNN Compression (BASC), a unified framework with two lightweight modules. For quantization, the scale is applied to synaptic current at every timestep and therefore shifts spike timing. Temporal-Behavior Scale Correction (TSC) makes the scale learnable under a temporal loss, allowing firing behavior to inform scale optimization. For pruning, channel importance depends on how channels jointly drive the membrane potential across the firing threshold. Boundary-Level Inter-Channel Correction (BIC) uses channelwise importance scores for initial selection and inter-channel information to re-evaluate only channels near the pruning threshold. Extensive experiments on static and neuromorphic benchmarks show that lower-bit BASC models match or outperform higher-bit baselines and retain this accuracy advantage after structured pruning, while further reducing model storage and synaptic operations.

Figures & tables

Figure 1: Overall pipeline of the proposed joint compression framework.

Figure 2: Temporal-behavior scale correction. (a) A scale fixed from the weight distribution settles at a suboptimal point of the task loss, whereas a learnable scale is driven toward the task optimum. (b–c) Top-1 accuracy on CIFAR-10 and CIFAR-100 under three bit-width pairings.

Figure 3: Multiplier sweep around the learned scale ( $m{=}1.00$ ). (a) The reconstruction error is minimized at $m{=}0.90$ while accuracy peaks at $m{=}1.00$ : the two optima do not coincide. (b) The accuracy peak sits at a markedly lower clipping ratio than the error optimum does, showing that the task trades a coarser grid for less clipping.

Figure 4: Boundary-level inter-channel correction. (a) SVS-only scoring can miss redundancy and yield different keep–prune decisions from ICM-aware evaluation. (b) BIC re-evaluates only boundary channels while preserving confident keep and drop decisions. (c) SVS and ICM agree on most channels, while their disagreements concentrate near the decision boundary.

Scoring scheme	ICM evals	Rel. evals	Wall time (s)
SVS only	0	0	0.88
Full-channel CHIP-CI	3,712	1.00	2,995
Boundary-only (ours)	572	0.154	480.35

Table 1: Scoring cost on CIFAR-10/VGG-16 at 4-bit (

T{=}4

), measured on one RTX 5090 with 6 calibration batches of 256 samples from the same checkpoint. Restricting ICM to the boundary is

6.24\times

faster than scoring every channel.

Dataset	Method	Arch.	Bit	T	Acc. (%)	Size (MB)
CIFAR-10	Deng et al. [TNNLS 2023]	7Conv2FC	32/3	8	90.19/87.59	62.16/5.84
	Wei et al. [ICLR 2025]	ResNet20	8/4/2	2	95.92/95.74/95.69	17.10/8.59/4.33
	BASC (Ours)	ResNet20	4/3/2	2	96.54/96.40/96.30	8.59/6.46/4.33
	Wei et al. [ICLR 2025]	VGG-16	8/4/2	4	93.33/93.28/93.01	14.77/7.42/3.74
	BASC (Ours)	VGG-16	4/3/2	4	93.76/93.77/93.41	7.42/5.58/3.74
CIFAR-100	Deng et al. [TNNLS 2023]	7Conv2FC	3/1	8	57.83/55.95	11.75/5.99

Table 2: Quantization-only top-1 accuracy and model size. Baseline rows are trained under our protocol at each bit width; other prior methods are quoted as published. Bold rows denote BASC.

Configuration	Acc. (%)	Compared to	$\boldsymbol{\Delta}$ Acc. (pp)
A. Fixed scale + SVS	74.68	–	–
B. TSC + SVS	75.94	A.	+1.26
C. Fixed scale + BIC	74.93	A.	+0.25
D. TSC + BIC	76.04	A.	+1.36

Table 3: Component and efficiency results. (a) Ablation of the two corrections on CIFAR-100/ResNet20 under the matched 4-bit,

T{=}2

setting. (b) Accuracy–energy comparison against the baseline under matched aggressive 4-bit compression.

Dataset	Method	Arch.	Bit	T	Acc. (%)	Size (MB)
CIFAR-10	Li et al. [ICML 2024]	VGG-16	32	4	91.67/90.26	17.32/5.68
	Wei et al. [ICLR 2025]	ResNet20	8/4/2	2	95.21/95.18/95.07	6.27/3.16/1.61
					93.92/93.77/93.53	3.92/1.98/1.02
	BASC (Ours)		4/3/2		95.61/95.53/95.40	3.16/2.38/1.61
					94.82/94.50/94.27	1.98/1.50/1.02
	Wei et al. [ICLR 2025]	VGG-16	8/4/2	4	91.61/91.61/91.22	4.28/2.16/1.10

Table 4: Top-1 accuracy and model size after quantization and structured pruning. Grouped rows list the evaluated pruning configurations for each backbone; bold rows denote BASC.

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Backbone	Downsample	Bit	T	Acc. (%)	Size (MB)
ImageNet	SEW ResNet	ReScaWConv+BN+SN	4	4	62.81	7.71
		Conv+BN+SN	4	4	63.40	8.31
		Conv+BN	4	4	63.26	8.31
	Spiking ResNet	Conv+BN+SN	4	4	61.66	8.31
		Conv+BN	4	4	62.02	8.31
		Padding	4	4	61.89	7.61

Appendix

Table 5: Architecture comparison on ImageNet. Size uses decimal MB under architecture-specific mixed-precision storage.

Code	Dataset	Backbone	Level	Params
V425	CIFAR-10	VGG-16	standard	4.25M
V142	CIFAR-10	VGG-16	aggressive	1.42M
V231	CIFAR-100	VGG-16	standard	2.31M
V168	CIFAR-100	VGG-16	aggressive	1.68M
V465	TinyImageNet	VGG-16	standard	4.65M
V343	TinyImageNet	VGG-16	aggressive	3.43M

Appendix

Table 6: Compression-level codes used throughout this document. Each code names the target parameter count of the QP-SNN configuration whose channel pruning ratios BASC reuses.

Backbone	Stage (channels)	Standard	Aggressive
VGG-16 (C10)	layers 1–9 (64–256)	0.45	0.49
VGG-16 (C10)	layers 11–16 (512)	0.51	0.80
VGG-16 (C100)	layers 1–9 (64–256)	0.45	0.45
VGG-16 (C100)	layers 11–16 (512)	0.70	0.78
VGG-16 (Tiny)	layers 1–9 (64–256)	0.45	0.45
VGG-16 (Tiny)	layers 11–16 (512)	0.51	0.62

Appendix

Table 7: Per-stage channel pruning ratios, reproduced from QP-SNN ( Wei et al. 2025 ) . The last convolutional layer of each backbone is left unpruned to preserve the classification head. For VGGSNN, the extreme level (D025) uses 0.82 and 0.93 for the two stages. Within ResNet20’s layer1, the two ratios apply to the first and second convolution of each block.

Bit	Samples	Channels	Agreement (%)	Retained overlap (%)	Jaccard (%)	Spearman $\rho$
2	1,536	3,712	95.26	95.32	91.06	0.9864
3	1,536	3,712	94.29	94.36	89.33	0.9902
4	1,536	3,712	95.80	95.85	92.03	0.9964

Appendix

Table 8: Global SVS–ICM agreement on CIFAR-10/VGG-16 (V425), aggregated over all 12 pruned layers. The same six fixed calibration batches are used for both scores.

Dataset / Backbone	Spearman	Agreement	Retained ov.	Bound. capture
CIFAR-10 / ResNet20	0.9908	95.70%	96.43%	64.10%
CIFAR-100 / ResNet20	0.9692	94.19%	95.18%	63.29%
DVS-CIFAR10 / VGGSNN	0.9614	89.64%	84.72%	49.14%

Appendix

Table 9: Cross-architecture SVS–ICM agreement, 4-bit setting.

$\rho$	Best Acc. (%)	Final Acc. (%)
0.00	91.94	91.60
0.01	91.85	91.40
0.03	91.90	91.70
0.05	91.59	91.40
0.10	91.90	91.54
0.20	91.99	91.75

Appendix

Table 10: Replacement-budget check on CIFAR-10/VGG-16/B4 (V425).

Score	Pairwise Spearman	Top- $k$ overlap (%)
SVS	0.9929	97.74
ICM	0.9932	97.68

Appendix

Table 11: Calibration-batch repeatability on CIFAR-10/VGG-16/B4.

Seed	TSC+SVS	TSC+BIC	Gain
0	75.94	76.04	+0.10
1	75.99	76.16	+0.17
2	76.00	76.24	+0.24
3	76.34	76.49	+0.15
Mean $\pm$ std	76.068 $\pm$ 0.184	76.233 $\pm$ 0.190	+0.165 $\pm$ 0.058

Appendix

Table 12: Paired multi-seed best accuracy, TSC+SVS vs. TSC+BIC. Sign consistency 4/4 positive.

Setting	$m_{\mathrm{error}}$	Acc. at $m_{\mathrm{error}}$	$m_{\mathrm{val}}$	Best Acc.	Acc. at $m{=}1$
CIFAR-10 / ResNet20 / B2	0.800	90.45	1.025	96.22	96.19
CIFAR-10 / ResNet20 / B3	0.900	95.90	0.975	96.25	96.12
CIFAR-10 / ResNet20 / B4	0.900	96.26	1.000	96.47	96.47
CIFAR-10 / VGG-16 / B2	0.800	58.11	1.000	93.41	93.41
CIFAR-10 / VGG-16 / B3	0.800	86.76	1.000	93.79	93.79
CIFAR-10 / VGG-16 / B4	0.900	93.27	1.000	93.75	93.75

Appendix

Table 13: Global learned-scale multiplier sweep across 19 checkpoints. Weights and learned scales are frozen; no retraining.

Setting	$E_{\mathrm{native}}$	PrefixCE	Prefix margin	Acc. (%)	$\Delta$ Acc. [95% CI]
C100/VGG-16/T4	0.3183 $\rightarrow$ 0.4133	2.203 $\rightarrow$ 1.089	0.025 $\rightarrow$ 2.373	48.78 $\rightarrow$ 72.53	+23.75 [22.73, 24.76]
C100/ResNet20/T2	0.3001 $\rightarrow$ 0.3869	2.477 $\rightarrow$ 0.916	$-$ 0.449 $\rightarrow$ 5.277	45.06 $\rightarrow$ 79.51	+34.45 [33.42, 35.48]
DVS-C10/VGGSNN/T10	0.3090 $\rightarrow$ 0.3729	1.009 $\rightarrow$ 0.678	2.518 $\rightarrow$ 3.950	74.90 $\rightarrow$ 83.90	+9.00 [6.90, 11.10]

Appendix

Table 14: Fixed-weight scale-path endpoints for three B2 checkpoints. Each arrow runs from the per-group reconstruction optimum

\boldsymbol{\alpha}_{E}

to the learned TSC scale

\boldsymbol{\alpha}_{\mathrm{TSC}}

. Accuracy intervals use 20,000 paired sample-level bootstrap draws. C100 and DVS-C10 denote CIFAR-100 and DVS-CIFAR10.

Setting	$t_{\mathrm{under}}/T$	$N_{\mathrm{miss}}/N_{\mathrm{spur}}$ (M)	$L_{\mathrm{under}}/L$	$A_{\mathrm{TET}}$
C100/VGG-16/T4	4/4	446.2 / 382.6	9/12 (75.0%)	0.8315
C100/ResNet20/T2	2/2	1748.0 / 1257.9	16/18 (88.9%)	0.6617
DVS-C10/VGGSNN/T10	10/10	172.4 / 126.3	6/7 (85.7%)	0.7843

Appendix

Table 15: Activity and gradient evidence along the fixed-weight correction from

\boldsymbol{\alpha}_{E}

\boldsymbol{\alpha}_{\mathrm{TSC}}

. Here

t_{\mathrm{under}}

counts under-firing timesteps,

N_{\mathrm{miss}}/N_{\mathrm{spur}}

gives missed/spurious crossings in millions, and

L_{\mathrm{under}}

counts affected layers with lower activity at

\boldsymbol{\alpha}_{E}

. Crossing counts compare the same samples and neuron-time elements at the two endpoints. PrefixCE improves and

A_{t}

is positive at every timestep for all three settings.

Bit	Base LR	Scale LR	Best Acc. (%)
4	$2\times 10^{-3}$	$2.5\times 10^{-4}$	61.96
4	$1\times 10^{-3}$	$5.0\times 10^{-4}$	62.39
2	$2\times 10^{-3}$	$2.5\times 10^{-4}$	59.88
2	$1\times 10^{-3}$	$5.0\times 10^{-4}$	60.78

Appendix

Table 16: ImageNet/ResNet18 optimization-recipe check. Both learning rates vary together.

Seed	Layers	Mean shift	Min–max shift	$m_{\mathrm{error}}<1$
0	12	12.08%	5.00–40.00%	Yes
1	12	8.75%	5.00–15.00%	Yes
2	12	8.33%	5.00–20.00%	Yes

Appendix

Table 17: Layer-wise scale mismatch on CIFAR-100/VGG-16/B4. The shift is measured relative to the learned scale

m{=}1

Bit	Learnable scale	Fixed scale	Gain
2	72.87	69.26	$+3.61$
3	72.96	65.11	$+7.85$
4	73.23	59.93	$+13.30$

Appendix

Table 18: Learnable versus fixed quantization scale on CIFAR-100/VGG-16. Top-1 accuracy (%); runs are matched within each bit width.

Bit	Setting	QP Top-1	BASC Top-1	$\Delta$ Acc	QP SOP (M)	BASC SOP (M)	$\Delta$ Energy
2	D146	79.1	79.1	0.0	279.801	269.886	$-$ 3.43%
2	D090	80.0	78.8	$-$ 1.2	230.228	218.547	$-$ 4.58%
2	D025	73.5	73.6	+0.1	44.678	42.814	$-$ 2.73%
3	D146	79.0	79.7	+0.7	280.579	270.407	$-$ 3.50%
3	D090	79.5	77.0	$-$ 2.5	229.530	222.302	$-$ 3.11%
3	D025	73.4	76.9	+3.5	44.489	46.024	+2.26%

Appendix

Table 19: DVS-CIFAR10/VGGSNN same-bit accuracy–energy matrix. Accuracy values are full-test measurements of the checkpoints instrumented for operation counting, and therefore differ from the best-epoch accuracies reported in the main pruning table; this matrix is used only for checkpoint-level accuracy–energy comparison.

Table 20: Additional efficiency comparisons.

Schedule	Best (%)	Final (%)
One-shot (ours)	73.65	73.24
Progressive A: fixed final plan	72.95	72.33
Progressive B: stage-wise re-scoring	72.78	72.06

Appendix

Table 21: Pruning schedules on CIFAR-100/ResNet20 under matched final compression.

Prune / regrow	Epochs	Best (%)
SCA / instant BN- $\gamma$	233	89.40
SCA / EMA BN- $\gamma$	233	90.69
SVS / EMA BN- $\gamma$	233	90.60

Appendix

Table 22: Exploratory dynamic-regrowth signal comparison on CIFAR-10/VGG-16 under a matched channel budget.

Spiking Neural Networks

Spiking neural networks process information through discrete, time-dependent neural events, aiming for efficient computation and biologically inspired learning.

See all 336 papers

Mar 16, 2026

Spiking Layer-Adaptive Magnitude-based Pruning

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.

Junqiao Wang, Zhehang Ye, Yuqi Ouyang

Sichuan University-Pittsburgh Institute, Sichuan University, Chengdu, 610207, China · College of Computer Science, Sichuan University, Chengdu, 610065, China

Spiking Neural Networks Neural Network Pruning Efficient Neural Network Inference

Ask PDF

Apr 28, 2026

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

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.

Dewei Bai, Hongxiang Peng, Jiajun Mei +4

Efficient Neural Network Inference Spiking Neural Networks Activation Quantization

Ask PDF

Oct 16, 2026

SpikeQ: Spike-Guided Post-Training Quantization for Spiking Neural Networks

On neuromorphic chips, spiking neural networks (SNNs) consume far less energy than artificial neural networks (ANNs), but the limited memory of these chips requires low-bit synaptic weights and membrane potentials. Existing post-training quantization (PTQ) methods provide them without retraining and, like those for ANNs, minimize numerical errors such as the mean squared error (MSE). In an SNN, however, a quantization error reaches later layers only through the spikes it flips. We find that errors of equal MSE can cause very different accuracy losses, whereas errors that flip the same fraction of spikes cause the same loss. We therefore propose SpikeQ, a spike-guided PTQ framework. For weights, we select the level of each weight by simulating the downstream neurons and minimizing their spike errors. For membrane potentials, most existing methods quantize the state after the reset, whereas neuromorphic chips compare the threshold with the membrane potential stored at a fixed bit width. We target this pre-fire quantization and build a reset-aware grid that places the threshold on a rounding boundary, so rounding never flips the spike at the current time step. We evaluate SpikeQ on seven benchmarks of image classification, semantic segmentation, and object detection. With 2-bit pre-fire membrane potentials, every existing grid collapses, whereas SpikeQ retains most of the full-precision accuracy. With both at 4 bits, the models stay within 2.59 points of full precision, shrink their memory footprint 7.8--7.9 times, and use 6.4--6.9 times less energy.

Kai Sun, Peibo Duan, Zhanglu Yan +3

Monash University · National University of Singapore · Northeastern University

Spiking Neural Networks Post-Training Quantization Neuromorphic Computing Activation Quantization

Ask PDF

Abstract

Spiking Neural Networks

Explore similar work

Spiking Layer-Adaptive Magnitude-based Pruning

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

SpikeQ: Spike-Guided Post-Training Quantization for Spiking Neural Networks