Research map / Architectures and efficiency / Quantization
Deep Transferring Quantization
Zheng Xie, Zhiquan Wen, Jing Liu, Zhiqiang Liu, Xixian Wu, Mingkui Tan
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 AFA loss (attentive feature alignment) to CE (ablation) | DTQ (5-bit MobileNetV2) · Caltech 256-30 | Top-1 accuracy (%): 74.9 -> 79.7 | Improved |
| add KL loss on top of CE + AFA (ablation) | DTQ (5-bit MobileNetV2) · Caltech 256-30 | Top-1 accuracy (%): 79.7 -> 80.1 | Improved |
| add attentive transfer module (attention) to feature alignment (ablation) | DTQ (4-bit MobileNetV2) · Food-101 | Top-1 accuracy (%): 77.1 -> 78.4 | Improved |
| add attentive transfer module (attention) to feature alignment (ablation) | DTQ (4-bit MobileNetV2) · Stanford Dogs 120 | Top-1 accuracy (%): 75.2 -> 76.1 | Improved |
| add attentive transfer module (attention) to feature alignment (ablation) | DTQ (4-bit MobileNetV2) · CUB-200-2011 | Top-1 accuracy (%): 71.6 -> 72.1 | Improved |
| transfer and quantize simultaneously (one-stage DTQ) instead of two-stage DT then L2-Q (ablation) | DTQ (4-bit MobileNetV2) · Stanford Dogs 120 | Top-1 accuracy (%): 71.5 -> 76.1 | Improved |
| add a transfer-only first stage before DTQ (two-stage DT->DTQ) (ablation) | DTQ (4-bit MobileNetV2) · Food-101 | Top-1 accuracy (%): 78.4 -> 79.9 | Improved |
| add a transfer-only first stage before DTQ (two-stage DT->DTQ) (ablation) | DTQ (4-bit MobileNetV2) · Caltech 256-60 | Top-1 accuracy (%): 79.9 -> 80.9 | Improved |
| use DTQ instead of DELTA-Q (learnable rather than fixed channel attention, plus KL) | MobileNetV2 4-bit · Caltech 256-30 | Top-1 accuracy (%): 74.9 -> 75.9 | Improved |
| use DTQ instead of DELTA-Q | ResNet-50 5-bit · CUB-200-2011 | Top-1 accuracy (%): 80.9 -> 80.3 | Hurt |
| use DTQ instead of DELTA-Q | ResNet-50 4-bit · CUB-200-2011 | Top-1 accuracy (%): 76.4 -> 77.8 | Improved |
| use DTQ instead of DELTA-Q | 4-bit MobileNetV2 with 10 training samples per class · Caltech 256-10 | Top-1 accuracy (%): +2.6% | Improved |
Closest papers by meaning
- Vector Quantized Image-to-Image Translation (ECCV 2022 · similarity 0.85)
- Learning Generalizable Mixed-Precision Quantization via Attribution Imitation (IJCV 2024 · similarity 0.84)
- Zero-Shot Quantization via Weight-Space Arithmetic (NeurIPS 2026 · similarity 0.82)
- Finite Scalar Quantization: VQ-VAE Made Simple (ICLR 2024 · similarity 0.82)
- Quantization without Tears (CVPR 2025 · similarity 0.81)
- ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization (NeurIPS 2025 · similarity 0.80)
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.
