Research map / Architectures and efficiency / Quantization
Learning Compression from Limited Unlabeled Data
Xiangyu He, Jian Cheng
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 BN re-normalization after 4-bit quantization (ablation) | ResNet-18 · ImageNet val 4-bit top-1 | top-1 accuracy gap: -13.21 -> -1.83 | Improved |
| add BN re-normalization after 4-bit quantization (ablation) | MobileNet · ImageNet val 4-bit top-1 | top-1 accuracy gap: -70.80 -> -9.75 | Improved |
| add BN re-normalization after 4-bit quantization (ablation) | ResNet-50 · ImageNet val 4-bit top-1 | top-1 accuracy gap: -7.16 -> -2.14 | Improved |
| add BN re-normalization after 4-bit quantization (ablation) | AlexNet · ImageNet val 4-bit top-1 | top-1 accuracy gap: -1.77 -> -0.39 | Improved |
| add BN re-normalization after 4-bit quantization (ablation) | AlexNet · ImageNet val 4-bit top-5 | top-5 accuracy gap: -1.30 -> -0.20 | Improved |
| add BN re-normalization after weight pruning (ablation) | MobileNet · ImageNet val 3x pruning | top-5 accuracy: +53.82% Top-5 improvement to 78.43% | Improved |
| add BN re-normalization after weight pruning (ablation) | ResNet-50 · ImageNet val 3x pruning | top-5 accuracy: +6.92% Top-5 improvement to 90.00% | Improved |
| add BN re-normalization to 3-bit weight AlexNet (ablation) | AlexNet · ImageNet val 3-8-bit | top-1 accuracy: +25.43% Top-1 to 50.69% | Improved |
| use 2^n quantizer instead of round-to-nearest | AlexNet 4-bit weights · ImageNet val top-1 | top-1 accuracy: 48.32 -> 58.66 | Improved |
| use 2^n quantizer instead of uniform | ResNet-18 4-bit weights · ImageNet val top-1 | top-1 accuracy: 50.18 -> 55.76 | Improved |
| use KL divergence metrics instead of l2 for weight quantization | AlexNet 4-bit 2^n · ImageNet val top-1 | top-1 accuracy: 58.90 -> 56.30 | Hurt |
| add feature-map fitting term to weight quantization objective | Quasi-Lloyd-Max · ImageNet val | accuracy | No clear effect |
Closest papers by meaning
- Transform Quantization for CNN Compression (TPAMI 2021 · similarity 0.89)
- Compression-aware Training of Deep Networks (NeurIPS 2017 · similarity 0.89)
- Model Preserving Compression for Neural Networks (NeurIPS 2022 · similarity 0.89)
- Holistic CNN Compression via Low-Rank Decomposition with Knowledge Transfer (TPAMI 2018 · similarity 0.89)
- Focused Quantization for Sparse CNNs (NeurIPS 2019 · similarity 0.89)
- Distribution-Aware Tensor Decomposition for Compression of Convolutional Neural Networks (NeurIPS 2025 · 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.
