Research map / Architectures and efficiency / Quantization
LS-ViT: Least-Squares Hessian Based Block Reconstruction for Low-Bit Post-Training Quantization of Vision Transformers
Hyunha Hwang, Xuan Truong Nguyen, Hyuk-Jae Lee
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 |
|---|---|---|---|
| replace averaging-based diagonal Hessian (FIMA-Q-D) with least-squares diagonal Hessian (ablation) | Hessian-based block reconstruction · ImageNet (W3/A3), ViT-S | Top-1 accuracy (%): 60.02 -> 63.25 | Improved |
| replace averaging-based diagonal with least-squares diagonal on Swin-S (ablation) | Hessian-based block reconstruction · ImageNet (W3/A3), Swin-S | Top-1 accuracy (%): 75.08 -> 77.29 | Improved |
| drop the division by dz_i and use g_i alone instead of g_i/dz_i (ablation) | Sample-wise diagonal Hessian · ImageNet (W3/A3), ViT-S | Top-1 accuracy (%): 23.59 -> 58.16 | Improved |
| use averaging-based FIMA-Q-D instead of sample-wise g_i (ablation) | Diagonal Hessian approximation · ImageNet (W3/A3), Swin models | Top-1 accuracy (%): +0.68%p average over g_i; Swin-S 76.02 -> 75.08 | Mixed |
| reduce FIMA-Q Hessian estimation from 15 to 1 full-model computation (ablation) | FIMA-Q · ImageNet (W3/A3), Swin-S | Top-1 accuracy (%): 77.26 -> 75.85 | Hurt |
| use LS-ViT with a single full-model computation instead of FIMA-Q with 15 | Hessian estimation · ImageNet (W3/A3), Swin-S | Top-1 accuracy (%): 77.26 -> 77.39 | Improved |
| replace FIMA-Q with LS-ViT at W2/A3 | Swin-S · ImageNet (W2/A3), Swin-S | Top-1 accuracy (%): 59.57 -> 65.77 | Improved |
| replace FIMA-Q with LS-ViT at W3/A3 | seven ViT/DeiT/Swin models · ImageNet (W3/A3) | Top-1 accuracy (%): average +0.17%p | No clear effect |
| replace FIMA-Q with LS-ViT at W2/A4 | seven ViT/DeiT/Swin models · ImageNet (W2/A4) | Top-1 accuracy (%): average +1.46%p | Improved |
| replace FIMA-Q with LS-ViT on COCO detection | Mask R-CNN Swin-T · COCO (W4/A4) | APb: 38.7 -> 38.9 | Improved |
| use LS-ViT instead of FIMA-Q for quantization training | Quantization time · Swin-B training time | time (min): 480 -> 180 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 3 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
- APHQ-ViT — APHQ-ViT: Post-Training Quantization with Average Perturbation Hessian Based Reconstruction for Vision Transformers (CVPR 2025) · better on 3 of 3 shared results
- QDrop — QDrop: Randomly Dropping Quantization for Extremely Low-bit Post-Training Quantization (ICLR 2022) · better on 0 of 1 shared result
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
- APHQ-ViT: Post-Training Quantization with Average Perturbation Hessian Based Reconstruction for Vision Transformers (CVPR 2025 · similarity 0.92)
- FIMA-Q: Post-Training Quantization for Vision Transformers by Fisher Information Matrix Approximation (CVPR 2025 · similarity 0.91)
- GradQ-ViT: Robust and Efficient Gradient Quantization for Vision Transformers (AAAI 2025 · similarity 0.90)
- RPFQ-ViT: Rotated Phase-Frame Quantization for Extremely Low-Bit Weights in Vision Transformers (NeurIPS 2026 · similarity 0.89)
- AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers (AAAI 2025 · similarity 0.89)
- Bi-ViT: Pushing the Limit of Vision Transformer Quantization (AAAI 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.
