Research map / Architectures and efficiency / Spiking networks
QP-SNN: Quantized and Pruned Spiking Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes QP-SNN, combining uniform weight quantization and structured kernel pruning for SNNs, with a weight rescaling strategy (ReScaW) to improve bit-width utilization and an SVS singular-value-based pruning criterion for more input-robust kernel removal.
Problem
SNNs are too large for resource-limited edge devices; existing joint quantization and pruning methods for SNNs either use unstructured pruning (hardware-unfriendly) or suffer large accuracy drops, and vanilla uniform quantization wastes bit-width while the SCA pruning criterion is input-dependent.
Approach
Rescale weights by a scale coefficient (1-norm mean) before uniform quantization so the integer grid is fully used (ReScaW), and score convolutional kernels by the count of significant singular values of their time-averaged spike matrix (SVS), which is more stable across inputs than the membrane-potential-based SCA score.
Key results
- On CIFAR-10 with ResNet20 at 8/4/2 bits and 2 timesteps, QP-SNN reaches 95.12/95.41/95.06% accuracy at 6.27/3.16/1.61 MB.
- Ablation on CIFAR-100 ResNet20 (1.20 MB): ReScaW plus SVS raises accuracy from 69.16% baseline to 73.89% (+4.73%).
- On Spikingformer CIFAR-100, QP-SNN reports 76.94% accuracy with 8.28x smaller model size versus the 79.09% uncompressed baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace vanilla quantization with ReScaW quantization (ablation) | quantized SNN without pruning · CIFAR-100 | accuracy: +0.63% | Helped |
| replace baseline quantization with ReScaW (ablation) | QP-SNN baseline · CIFAR-100 | accuracy: +4.24% | Helped |
| replace baseline pruning criterion with SVS (ablation) | QP-SNN baseline · CIFAR-100 | accuracy: +4.16% | Helped |
| add ReScaW and SVS to the baseline (ablation) | QP-SNN baseline · CIFAR-100 | accuracy: +4.73% | Helped |
| set ReScaW γ to the 1-norm mean instead of max absolute value (ablation) | quantized SNN without pruning · CIFAR-100 | accuracy: 77.85% vs 79.16% | No clear effect |
| set ReScaW γ to the percentile value instead of the 1-norm mean (ablation) | quantized SNN without pruning · CIFAR-100 | accuracy: 77.8% vs 79.16% | No clear effect |
| apply ReScaW and SVS while pruning before quantization (ablation) | PQ-SNN baseline · CIFAR-100 | accuracy: +1.83% | Helped |
| apply ReScaW and SVS while quantizing before pruning (ablation) | QP-SNN baseline · CIFAR-100 | accuracy: +4.46% | Helped |
| quantize before pruning rather than prune before quantization (ablation) | QP-SNN · CIFAR-100 | accuracy: surpassing PQ-SNN by 1.39% | Helped |
| quantize and prune Spikingformer with QP-SNN | Spikingformer-4-384 · CIFAR-100 | accuracy: -2.15 | Hurt |
| apply QP-SNN quantization and pruning | YOLO-v3 with ResNet10 · SSDD | mAP@0.5: +0.3 | Helped |
| apply QP-SNN quantization and pruning | YOLO-v3 with ResNet10 · NWPU VHR-10 | mAP@0.5: -3.21 | Hurt |
Surprising findings
- Vanilla uniform quantization under 8 bits uses under half of the assigned integer grid (14% to 32% in several layers), so many weights collapse onto the same integer levels.
- The SCA importance score has low robustness across inputs (min average cosine similarity about 0.87 to 0.95), which the authors argue undermines reliable kernel selection.
Limitations
- Pruning rates were manually selected without rigorous search, as stated by the authors.
- Accuracy drops relative to uncompressed SNNs remain (e.g., 2.44% on CIFAR-10 at 9.61% connections), and object detection shows a 3.21% mAP drop on NWPU VHR-10.
- Comparisons with prior methods use differing architectures, timesteps and training schemes, complicating direct comparison.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 TinyImageNet ImageNet-1k DVS-CIFAR10 SSDD NWPU VHR-10
Benchmarks: CIFAR-10 CIFAR-100 TinyImageNet ImageNet-1k DVS-CIFAR10 SSDD NWPU VHR-10
Metrics: top-1 accuracy model size (MB) SOPs power (mJ) mAP@0.5
Baselines: SCA-based structured pruning (ICML24) ICLR24 SNN compression method TNNLS21 quantized SNN IJCNN21 hybrid SNN WACV21 TinyImageNet SNN
Builds on: spiking channel activity (SCA) pruning STBP straight-through estimator
Closest papers by meaning
- Are Conventional SNNs Really Efficient? A Perspective from Network Quantization (CVPR 2024 · similarity 0.92)
- S$^2$NN: Sub-bit Spiking Neural Networks (NeurIPS 2025 · similarity 0.91)
- Quantized Spike-driven Transformer (ICLR 2025 · similarity 0.90)
- Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications (CVPR 2025 · similarity 0.90)
- Improving the Sparse Structure Learning of Spiking Neural Networks from the View of Compression Efficiency (ICLR 2025 · similarity 0.90)
- EnOF-SNN: Training Accurate Spiking Neural Networks via Enhancing the Output Feature (NeurIPS 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.