Research map / Architectures and efficiency / Spiking networks
Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications
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 a post-training ANN-SNN conversion framework with inference-scale cost, using local threshold balancing (channel-wise) and delayed evaluation to convert pretrained ANNs into SNNs across classification, segmentation, detection, and video tasks.
Problem
ANN-SNN conversion typically requires quantized ANN retraining or fine-tuning, which adds substantial training cost; converted SNNs also need many time-steps and suffer from spike propagation delays.
Approach
Derives an upper bound on layer-wise conversion error and optimizes per-layer (channel-wise) firing thresholds with a local, gradient-based rule run on a small data sample during inference, plus a delayed evaluation strategy that starts averaging output spikes after the expected first-spike time.
Key results
- On ImageNet with ResNet-34, the converted SNN exceeds 71% accuracy within 64 time-steps and reaches 71.94% in 32 steps when applied to QCFS-pretrained models.
- Within 64 time-steps, the converted SNN achieves 27.3% mAP with RetinaNet-50 and 26.5% mAP with FCOS-50 on MS COCO 2017.
- Energy estimate: ResNet-34 SNN achieves 622 FPS/W at 90% of ANN performance versus 22 FPS/W for the original ANN (about 28x).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add local threshold balancing and delayed evaluation to LCP | LCP on ResNet-34 · ImageNet | accuracy: 50.21 -> 53.80 | Helped |
| Add local threshold balancing and delayed evaluation to ACP | ACP on ResNet-34 · ImageNet | accuracy: 64.54 -> 65.95 | Helped |
| Add local threshold balancing and delayed evaluation to QCFS | QCFS on ResNet-34 · ImageNet | accuracy: 64.43 -> 71.94 | Helped |
| Use channel-wise threshold balancing instead of layer-wise threshold balancing (ablation) | Proposed conversion framework · ImageNet | accuracy | Helped |
| Add delayed evaluation (ablation) | Proposed conversion framework with channel-wise threshold balancing · ImageNet | accuracy | Helped |
| Increase local threshold balancing iterations from 1000 to 5000 (ablation) | Proposed conversion framework on ResNet-34 · ImageNet | accuracy: approximately 1% -> around 0.2% | Helped |
| Set delayed evaluation using the estimated first-spike time (ablation) | Proposed conversion framework · ImageNet | accuracy | Helped |
Surprising findings
- Plugging the local threshold balancing and delayed evaluation into existing methods (LCP, ACP, QCFS) improves them by over 3% at T=32 on ImageNet.
Limitations
- Performance is somewhat weaker than training-involved SNN methods, a trade-off with training cost.
- Only convolutional SNNs are studied; transformer architectures and other complex layers are not yet handled.
Experimental setup
Datasets: ImageNet Pascal VOC 2012 SBD MS COCO 2017 Kinetics-400
Benchmarks: ImageNet Pascal VOC 2012 MS COCO 2017 Kinetics-400
Metrics: top-1 accuracy mIoU mAP FPS/W energy efficiency
Baselines: RMP LCP ACP QCFS Robust-Norm Han and Roy (RMP) Kim et al. spike-based detection
Builds on: QCFS Robust-Norm (RN) LCP/ACP delayed evaluation (Deng and Gu) membrane potential initialization
Closest papers by meaning
- Faster and Stronger: When ANN-SNN Conversion Meets Parallel Spiking Calculation (ICML 2025 · similarity 0.93)
- Efficient ANN-SNN Conversion with Error Compensation Learning (ICML 2025 · similarity 0.93)
- Adaptive Calibration: A Unified Conversion Framework of Spiking Neural Networks (AAAI 2025 · similarity 0.92)
- Temporal Misalignment in ANN-SNN Conversion and its Mitigation via Probabilistic Spiking Neurons (ICML 2025 · similarity 0.91)
- LM-HT SNN: Enhancing the Performance of SNN to ANN Counterpart through Learnable Multi-hierarchical Threshold Model (NeurIPS 2024 · similarity 0.90)
- QP-SNN: Quantized and Pruned Spiking Neural Networks (ICLR 2025 · similarity 0.90)
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.