Research map / Architectures and efficiency / Quantization
NoisyQuant: Noisy Bias-Enhanced Post-Training Activation Quantization for Vision Transformers
Yijiang Liu, Huanrui Yang, Zhen Dong, Kurt Keutzer, Li Du, Shanghang Zhang
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 fixed uniform noisy bias before activation quantization | EasyQuant on ViT-S · ImageNet-2012, W6A6 | top-1 accuracy: 75.13 -> 76.86 | Improved |
| Add fixed uniform noisy bias before activation quantization | EasyQuant on DeiT-S · ImageNet-2012, W6A6 | top-1 accuracy: 75.27 -> 76.37 | Improved |
| Add fixed uniform noisy bias before activation quantization | PTQ4ViT on DeiT-S · ImageNet-2012, W6A6 | top-1 accuracy: 76.28 -> 77.43 | Improved |
| Add fixed uniform noisy bias before activation quantization | PTQ4ViT on ViT-S · ImageNet-2012, W6A6 | top-1 accuracy: 78.63 -> 78.65 | No clear effect |
| Add fixed uniform noisy bias before activation quantization | PTQ4ViT on Swin-B · ImageNet-2012, W6A6 | top-1 accuracy: 84.01 -> 84.68 | Improved |
| Add noisy bias to each layer type separately (ablation) | EasyQuant on DeiT-S · ImageNet-2012, W6A6 | top-1 accuracy: 75.27 -> 75.38 (qkv); 75.27 -> 75.45 (proj); 75.27 -> 75.33 (fc1); 75.27 -> 76.21 (fc2) | Improved |
| Add noisy bias to qkv, proj, fc1, and fc2 layers (ablation) | EasyQuant on DeiT-S · ImageNet-2012, W6A6 | top-1 accuracy: 75.27 -> 76.37 | Improved |
| Add noisy bias to each layer type separately (ablation) | EasyQuant on Swin-T · ImageNet-2012, W6A6 | top-1 accuracy: 79.51 -> 79.53 (qkv); 79.51 -> 79.56 (proj); 79.51 -> 79.52 (fc1); 79.51 -> 79.80 (fc2) | Improved |
| Add noisy bias to qkv, proj, fc1, and fc2 layers (ablation) | EasyQuant on Swin-T · ImageNet-2012, W6A6 | top-1 accuracy: 79.51 -> 80.01 | Improved |
| Use calibration size 32 rather than 1024 (ablation) | NoisyQuant · ImageNet-2012, W6A6 | top-1 accuracy: 76.86 -> 76.81 (ViT-S); 76.37 -> 76.30 (DeiT-S); 80.01 -> 79.97 (Swin-T) | No clear effect |
| Use calibration size 128 rather than 1024 (ablation) | NoisyQuant · ImageNet-2012, W6A6 | top-1 accuracy: 76.86 -> 76.87 (ViT-S); 76.37 -> 76.47 (DeiT-S); 80.01 -> 80.13 (Swin-T) | No clear effect |
| Store noisy bias and denoising bias as INT16 instead of FP32 (ablation) | NoisyQuant · ImageNet-2012 | top-1 accuracy | 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.
- PTQ4ViT — PTQ4ViT: Post-Training Quantization for Vision Transformers with Twin Uniform Quantization (ECCV 2022) · better on 4 of 4 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
- AIQViT: Architecture-Informed Post-Training Quantization for Vision Transformers (AAAI 2025 · similarity 0.91)
- ERQ: Error Reduction for Post-Training Quantization of Vision Transformers (ICML 2024 · similarity 0.90)
- Amortized-Precision Quantization for Early-Exit Vision Transformers (NeurIPS 2026 · similarity 0.89)
- FrameQuant: Flexible Low-Bit Quantization for Transformers (ICML 2024 · similarity 0.89)
- APHQ-ViT: Post-Training Quantization with Average Perturbation Hessian Based Reconstruction for Vision Transformers (CVPR 2025 · similarity 0.88)
- Outlier-aware Slicing for Post-Training Quantization in Vision Transformer (ICML 2024 · 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.
