Research map / Architectures and efficiency / Quantization
Up or Down? Adaptive Rounding for Post-Training Quantization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use AdaRound instead of rounding-to-nearest for 4-bit weights | ResNet18 · ImageNet validation | top-1 accuracy: 23.99 -> 68.71±0.06 | Helped |
| Use AdaRound instead of OMSE+opt for 4-bit weights | ResNet50 · ImageNet validation | top-1 accuracy: 74.67 -> 75.23±0.04 | Helped |
| Use AdaRound instead of OMSE+opt for 4-bit weights | InceptionV3 · ImageNet validation | top-1 accuracy: 73.66 -> 75.76±0.09 | Helped |
| Use AdaRound instead of OMSE+opt for 4-bit weights | MobileNetV2 · ImageNet validation | top-1 accuracy: 71.2 -> 69.78±0.05 | Hurt |
| Add 8-bit activation quantization to AdaRound (ablation) | ResNet18 · ImageNet validation | top-1 accuracy: 68.71±0.06 -> 68.55±0.01 | No clear effect |
| Add 8-bit activation quantization to AdaRound (ablation) | DeepLabV3+ (MobileNetV2 backend) · Pascal VOC | mIOU: 70.89±0.33 -> 70.86±0.37 | No clear effect |
| Replace rounding-to-nearest with task-loss Hessian rounding (ablation) | ResNet18, first layer · ImageNet validation | top-1 accuracy: 52.29 -> 68.62±0.17 | Helped |
| Approximate task-loss Hessian optimization with local MSE loss (ablation) | ResNet18, first layer · ImageNet validation | top-1 accuracy: 52.29 -> 69.39±0.04 | Helped |
| Add continuous relaxation to local MSE optimization (ablation) | ResNet18, all layers · ImageNet validation | top-1 accuracy: 65.83±0.14 -> 66.56±0.12 | Helped |
| Use explicit regularization rather than sigmoid temperature annealing (ablation) | ResNet18 · ImageNet validation, all layers | top-1 accuracy: 65.22±0.67 -> 66.25±0.15 | Helped |
| Use rectified sigmoid rather than sigmoid with explicit regularization (ablation) | ResNet18 · ImageNet validation, all layers | top-1 accuracy: 66.25±0.15 -> 66.56±0.12 | Helped |
| Add asymmetric reconstruction to layer-wise optimization (ablation) | ResNet18 · ImageNet validation, all layers | top-1 accuracy: 66.56±0.12 -> 68.37±0.07 | Helped |
Closest papers by meaning
- FlexRound: Learnable Rounding based on Element-wise Division for Post-Training Quantization (ICML 2023 · similarity 0.92)
- Model-Preserving Adaptive Rounding (ICML 2026 · similarity 0.90)
- Relaxed Quantization for Discretized Neural Networks (ICLR 2019 · similarity 0.89)
- LogART: Pushing the Limit of Efficient Logarithmic Post-Training Quantization (ICLR 2026 · similarity 0.89)
- MagR: Weight Magnitude Reduction for Enhancing Post-Training Quantization (NeurIPS 2024 · similarity 0.88)
- A2Q+: Improving Accumulator-Aware Weight Quantization (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.