Research map / Architectures and efficiency / Quantization
Data-Free Quantization Through Weight Equalization and Bias Correction
Markus Nagel, Mart van Baalen, Tijmen Blankevoort, Max Welling
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 cross-layer equalization (ablation) | MobileNetV2 INT8 per-tensor · ImageNet val | top-1 accuracy (%): 0.11% -> 69.91% | Improved |
| add high-bias absorption on top of equalization (ablation) | MobileNetV2 INT8 per-tensor · ImageNet val | top-1 accuracy (%): 69.91% -> 70.92% | Improved |
| add bias correction alone (ablation) | MobileNetV2 INT8 per-tensor · ImageNet val | top-1 accuracy (%): 0.12% -> 52.02% | Improved |
| add bias correction to weight clipping at 15 (ablation) | MobileNetV2 INT8 per-tensor · ImageNet val | top-1 accuracy (%): 2.55% -> 70.43% | Improved |
| add bias correction to weight clipping at 15 (full precision) (ablation) | MobileNetV2 FP32 · ImageNet val | top-1 accuracy (%): 67.06% -> 71.15% | Improved |
| add bias correction to equalization plus bias absorption (ablation) | MobileNetV2 INT8 per-tensor · ImageNet val | top-1 accuracy (%): 70.92% -> 71.19% | Improved |
| replace ReLU6 with ReLU (ablation) | MobileNetV2 · ImageNet val | top-1 accuracy FP32 (%): 71.72% -> 71.70% | No clear effect |
| use empirical instead of analytic bias correction (ablation) | DFQ MobileNetV2 · ImageNet val | top-1 accuracy (%): 71.19% -> 71.15% | No clear effect |
| add 1 epoch quantization-aware fine-tuning after DFQ | MobileNetV2 · ImageNet val | top-1 accuracy (%): 71.19% -> 71.42% | Improved |
| use asymmetric instead of symmetric quantization with DFQ (ablation) | DFQ MobileNetV1 · ImageNet val | top-1 accuracy (%): 70.32% -> 70.51% | No clear effect |
| add cross-layer equalization to per-channel quantization (ablation) | MobileNetV2 per-channel · ImageNet val | top-1 accuracy (%): 70.65% -> 70.93% | Improved |
| add bias correction to CLE+BA with per-channel quantization (ablation) | MobileNetV2 per-channel · ImageNet val | top-1 accuracy (%): 71.03% -> 71.33% | Improved |
Closest papers by meaning
- Scalable methods for 8-bit training of neural networks (NeurIPS 2018 · similarity 0.90)
- Toward INT4 Fixed-Point Training via Exploring Quantization Error for Gradients (ECCV 2024 · similarity 0.90)
- F8Net: Fixed-Point 8-bit Only Multiplication for Network Quantization (ICLR 2022 · similarity 0.90)
- Quantization Networks (CVPR 2019 · similarity 0.90)
- Optimization-Based Post-Training Quantization With Bit-Split and Stitching (TPAMI 2022 · similarity 0.89)
- Efficient Multi-bit Quantization Network Training via Weight Bias Correction and Bit-wise Coreset Sampling (NeurIPS 2025 · 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.
