Research map / Architectures and efficiency / Quantization
Explicit Model Size Control and Relaxation via Smooth Regularization for Mixed-Precision Quantization
Vladimir Chikin, Kirill Solodskikh, Irina Zhelavskaya
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 smooth bounded regularizers to bit-width tuning (2-bit) (ablation) | bit-width tuning · CIFAR-10 ResNet-20 2-bit | Top-1 accuracy (%): 65.94 -> 77.05 | Improved |
| add smooth regularizers to bit-widths and scales tuning (2-bit) (ablation) | bit-widths and scales tuning · CIFAR-10 ResNet-20 2-bit | Top-1 accuracy (%): 79.20 -> 86.85 | Improved |
| add smooth regularizers to all-parameter tuning (2-bit) (ablation) | full QAT · CIFAR-10 ResNet-20 2-bit | Top-1 accuracy (%): 91.27 -> 91.57 | Improved |
| add smooth regularizers to bit-width tuning (3-bit) (ablation) | bit-width tuning · CIFAR-10 ResNet-20 3-bit | Top-1 accuracy (%): 89.70 -> 90.13 | Improved |
| add smooth regularizers to all-parameter tuning (3-bit) (ablation) | full QAT · CIFAR-10 ResNet-20 3-bit | Top-1 accuracy (%): 91.69 -> 91.89 | Improved |
| also tune scales, not only bit-widths (2-bit, no regularizers) (ablation) | bit-width tuning · CIFAR-10 ResNet-20 2-bit | Top-1 accuracy (%): 65.94 -> 79.20 | Improved |
| also tune all model weights (2-bit, no regularizers) (ablation) | bit-widths and scales tuning · CIFAR-10 ResNet-20 2-bit | Top-1 accuracy (%): 79.20 -> 91.27 | Improved |
| bit-width tuning instead of BN-tuned fixed PTQ (2-bit) (ablation) | post-training quantization · CIFAR-10 ResNet-20 2-bit | Top-1 accuracy (%): 64.34 -> 65.94 | Improved |
| add smooth regularizers to full method, activations 8x compression (ablation) | Ours · CIFAR-10 ResNet-20 (AC 8) | Top-1 accuracy (%): 91.55 -> 91.62 | No clear effect |
| add smooth regularizers, activations not quantized (ablation) | Ours · CIFAR-10 ResNet-20 (AC 1) | Top-1 accuracy (%): 91.75 -> 91.97 | Improved |
| add smooth regularizers, activations not quantized (ablation) | Ours · CIFAR-10 ResNet-20 (AC 1) | model compression ratio: 16.19 -> 14.73 | Hurt |
| learned layer for 4-bit vs fixed alternative layer choice (ablation) | Ours · Vimeo-90K ESPCN | PSNR (dB): 31.10 -> 30.74 | Improved |
Closest papers by meaning
- FracBits: Mixed Precision Quantization via Fractional Bit-Widths (AAAI 2021 · similarity 0.91)
- Towards Mixed-Precision Quantization of Neural Networks via Constrained Optimization (ICCV 2021 · similarity 0.91)
- Mixed Precision DNNs: All you need is a good parametrization (ICLR 2020 · similarity 0.90)
- Retraining-Free Model Quantization via One-Shot Weight-Coupling Learning (CVPR 2024 · similarity 0.88)
- Learning from Loss Landscape: Generalizable Mixed-Precision Quantization via Adaptive Sharpness-Aware Gradient Aligning (ICML 2025 · similarity 0.88)
- LOTION: Smoothing the Optimization Landscape for Quantized Training (NeurIPS 2026 · similarity 0.87)
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.
