Research map / Architectures and efficiency / Spiking networks
Generalized Threshold Optimization with Harmony Multi-Threshold Neurons for Accurate ANN-to-SNN Conversion
Wenhan Zhang, Zihan Huang, Tong Bu, Tiejun Huang, Zhaofei Yu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add communication mechanism to H-MT 4/Opt (ablation) | H-MT 4/Opt · ImageNet ViT-base (T=16) | accuracy (%): 69.214 -> 74.942 | Improved |
| add communication mechanism to H-MT 4/999 (ablation) | H-MT 4/999 · ImageNet ViT-base (T=16) | accuracy (%): 63.482 -> 73.878 | Improved |
| use generalized threshold optimization instead of 99.9% quantile, no communication (ablation) | H-MT 4 · ImageNet ViT-base (T=16) | accuracy (%): 63.482 -> 69.214 | Improved |
| use generalized threshold optimization instead of 99.9% quantile (ablation) | H-MT 4 · ImageNet ResNet34 (T=16) | accuracy (%): 72.85 -> 75.80 | Improved |
| use generalized threshold optimization instead of 99.9% quantile (ablation) | H-MT 4 · ImageNet VGG-16bn | accuracy (%): 72.19 -> 71.94 (T=4); 72.15 -> 72.92 (T=16) | Mixed |
| use generalized threshold optimization instead of 99.9% quantile (ablation) | H-MT 8 · ImageNet ResNet34 (T=2) | accuracy (%): 72.908 -> 76.276 | Improved |
| use generalized threshold optimization instead of 99.9% quantile (ablation) | H-MT 8 · ImageNet ResNet34 (T=2) | energy ratio: 0.5435 -> 0.2986 | Improved |
| double the 99.9% quantile threshold | H-MT 8/999 · ImageNet ResNet34 (T=2) | accuracy (%): 72.908 -> 75.702 | Improved |
| increase thresholds from n=3 to n=4 | H-MT /Opt · ImageNet ResNet34 (T=16) | accuracy (%): 73.91 -> 75.80 | Improved |
| increase time-steps from 2 to 16 | H-MT 4/Opt · ImageNet ResNet34 | accuracy (%): 71.73 -> 75.80 | Improved |
| replace identity mapping with H-MT (y=x) neurons | H-MT 4/Opt · ImageNet ResNet18 (T=16) | accuracy (%): 70.95 -> 67.61 | Hurt |
| increase thresholds from n=4 to n=8 at T=2 | H-MT /Opt · ImageNet ResNet34 | accuracy (%): 71.73 -> 76.276 | Improved |
Closest papers by meaning
- Efficient ANN-SNN Conversion with Error Compensation Learning (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)
- Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications (CVPR 2025 · similarity 0.89)
- Faster and Stronger: When ANN-SNN Conversion Meets Parallel Spiking Calculation (ICML 2025 · similarity 0.89)
- Temporal Misalignment in ANN-SNN Conversion and its Mitigation via Probabilistic Spiking Neurons (ICML 2025 · similarity 0.88)
- Adaptive Calibration: A Unified Conversion Framework of Spiking Neural Networks (AAAI 2025 · similarity 0.88)
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.
