Research map / Architectures and efficiency / Quantization
Scheduling Weight Transitions for Quantization-Aware Training
Junghyup Lee, Jeimin Jeon, Dohyung Kim, Bumsub Ham
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 |
|---|---|---|---|
| schedule transition rate (SGDT) instead of learning rate (SGD) | MobileNetV2 · ImageNet val W/A 2/2 | top-1 accuracy: 46.9 -> 53.6 | Improved |
| schedule transition rate (SGDT) instead of LR | ResNet-18 · ImageNet val W/A 1/1 | top-1 accuracy: 55.3 -> 55.8 | Improved |
| schedule transition rate (AdamT) instead of LR (Adam) | ResNet-18 · ImageNet val W/A 2/2 | top-1 accuracy: 66.7 -> 67.2 | Improved |
| schedule transition rate (AdamWT) instead of AdamW | DeiT-S · ImageNet val W/A 2/2 | top-1 accuracy: 68.4 -> 71.8 | Improved |
| schedule transition rate (SGDT) instead of LR | ReActNet-18 · CIFAR-100 test W/A 1/1 | top-1 accuracy: 69.7 -> 72.2 | Improved |
| schedule transition rate (SGDT) instead of LR | RetinaNet ResNet-50 · MS COCO val W/A 4/4 | AP: 38.05 -> 38.36 | Improved |
| use SGDT with step decay scheduler instead of SGD | ResNet-20 · CIFAR-100 test W/A 1/1 (step decay) | top-1 accuracy: 45.8 -> 55.3 | Improved |
| TR scheduling gains grow at lower bit-width for MobileNetV2 | SGDT · ImageNet val MobileNetV2 | top-1 accuracy: 66.8 -> 66.9 (ResNet-18 2/2) vs 46.9 -> 53.6 (MobileNetV2 2/2) | Improved |
| raise final target TR from 0 to 1e-3 (ablation) | SGDT · CIFAR-100 test W/A 2/2 | top-1 accuracy: 65.61 -> 62.12 | Hurt |
| vary TR factor lambda from 4e-3 to 6e-3 versus extremes (ablation) | SGDT · CIFAR-100 test W/A 2/2 | top-1 accuracy: 1e-3: 62.5; 5e-3: 65.5; 7e-3: 63.1 | Mixed |
| use additive TALR update (Eq. 11) rather than multiplicative update (ablation) | SGDT · CIFAR-100 W/A 2/2 step scheduler | stability | Improved |
| apply TR scheduling to Adagrad | Adagrad · CIFAR-100 test W/A 2/2 | top-1 accuracy: 54.2 -> 61.1 | Improved |
Closest papers by meaning
- Compute-Optimal Quantization-Aware Training (ICLR 2026 · similarity 0.90)
- MagR: Weight Magnitude Reduction for Enhancing Post-Training Quantization (NeurIPS 2024 · similarity 0.87)
- HESTIA: A Hessian-Guided Differentiable Quantization-Aware Training Framework for Extremely Low-Bit LLMs (NeurIPS 2026 · similarity 0.87)
- LoQT: Low-Rank Adapters for Quantized Pretraining (NeurIPS 2024 · similarity 0.87)
- Optimal and Approximate Adaptive Stochastic Quantization (NeurIPS 2024 · similarity 0.87)
- BoA: Attention-aware Post-training Quantization without Backpropagation (ICML 2025 · 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.
