Research map / Architectures and efficiency / Quantization
SHARP-Q: Spectral Hessian Alignment and Rectification for Post-training Quantization
Menghao Lv, Huiqiong Wang, Li Sun, Mingli Song
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 Hessian-Aware Rectification (HAR) to MSE reconstruction (ablation) | MSE baseline · ImageNet (ResNet-18, W2A2) | Top-1 accuracy (%): 52.85 -> 54.18 | Improved |
| add HAR to MSE reconstruction (ablation) | MSE baseline · ImageNet (RegNetX-600MF, W2A2) | Top-1 accuracy (%): 40.06 -> 41.69 | Improved |
| add HAR to MSE reconstruction (ablation) | MSE baseline · ImageNet (ViT-S, W3A3) | Top-1 accuracy (%): 39.17 -> 40.51 | Improved |
| add HAR to MSE reconstruction (ablation) | MSE baseline · ImageNet (DeiT-T, W3A3) | Top-1 accuracy (%): 47.25 -> 48.63 | Improved |
| add Dynamic Fisher-Subspace Compensation (DFSC) on top of HAR (ablation) | SHARP-Q · ImageNet (ViT-S, W3A3) | Top-1 accuracy (%): 40.51 -> 66.93 | Improved |
| add DFSC on top of HAR (ablation) | SHARP-Q · ImageNet (DeiT-T, W3A3) | Top-1 accuracy (%): 48.63 -> 56.14 | Improved |
| add DFSC on top of HAR (ablation) | SHARP-Q · ImageNet (RegNetX-600MF, W2A2) | Top-1 accuracy (%): 41.69 -> 46.89 | Improved |
| add DFSC on top of HAR (ablation) | SHARP-Q · ImageNet (ResNet-18, W2A2) | Top-1 accuracy (%): 54.18 -> 56.74 | Improved |
| use SHARP-Q instead of CL-Calib | CNN PTQ · ImageNet (ResNet-18, W2A2) | Top-1 accuracy (%): 54.45 -> 56.74 | Improved |
| use SHARP-Q instead of CL-Calib | CNN PTQ · ImageNet (RegNetX-600MF, W2A2) | Top-1 accuracy (%): 42.19 -> 46.89 | Improved |
| use SHARP-Q instead of CL-Calib | CNN PTQ · ImageNet (ResNet-50, W4A4) | Top-1 accuracy (%): 75.38 -> 75.25 | No clear effect |
| use SHARP-Q instead of CL-Calib | CNN PTQ · ImageNet (MNasNet, W2A4) | Top-1 accuracy (%): 63.46 -> 63.44 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- FIMA-Q — FIMA-Q: Post-Training Quantization for Vision Transformers by Fisher Information Matrix Approximation (CVPR 2025) · better on 10 of 10 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- FIMA-Q: Post-Training Quantization for Vision Transformers by Fisher Information Matrix Approximation (CVPR 2025 · similarity 0.88)
- Zero-Shot Quantization via Weight-Space Arithmetic (NeurIPS 2026 · similarity 0.87)
- Not All Tasks Quantize Equally: Fisher-Guided Quantization for Visual Geometry Transformer (NeurIPS 2026 · similarity 0.87)
- PARQ: Piecewise-Affine Regularized Quantization (ICML 2025 · similarity 0.86)
- Learning from Loss Landscape: Generalizable Mixed-Precision Quantization via Adaptive Sharpness-Aware Gradient Aligning (ICML 2025 · similarity 0.86)
- APHQ-ViT: Post-Training Quantization with Average Perturbation Hessian Based Reconstruction for Vision Transformers (CVPR 2025 · similarity 0.86)
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.
