Research map / Label-efficient and robust learning / Knowledge distillation
Distilling Knowledge by Mimicking Features
Guohua Wang, Yifan Ge, Jianxin Wu
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 LSH loss to the l2 feature mimicking loss (ablation) | feature mimicking · CIFAR-100 (ResNet50 -> MobileNetV2) | top-1 accuracy (%): 65.73 -> 68.99 | Improved |
| add LSH loss to the l2 loss (ablation) | feature mimicking · CIFAR-100 (vgg13 -> vgg8) | top-1 accuracy (%): 72.33 -> 73.68 | Improved |
| replace KD logit loss with l2 feature mimicking (ablation) | student training · CIFAR-100 (WRN-40-2 -> WRN-16-2) | top-1 accuracy (%): 74.92 -> 75.53 | Improved |
| add linear embedding layer (2FC) for feature space alignment (ablation) | l2 feature mimicking · VOC2007 (ResNet34 -> ResNet18) | mAP (%): 88.75 -> 89.98 | Improved |
| use student pretrained by KD with 1FC (ablation) | l2 feature mimicking · VOC2007 (ResNet34 -> ResNet18) | mAP (%): 88.75 -> 90.89 | Improved |
| add 2FC to student already pretrained by KD (ablation) | l2 feature mimicking · VOC2007 (ResNet34 -> ResNet18) | mAP (%): 90.89 -> 90.77 | Hurt |
| two-stage finetuning (align embedding first) (ablation) | feature mimicking · VOC2007 (ResNet34 -> ResNet18) | mAP (%): 89.98 -> 90.59 | Improved |
| remove linear embedding layer (1FC) when dims match (ablation) | Ours · CIFAR-100 (resnet32x4 -> resnet8x4) | top-1 accuracy (%): 73.40 -> 76.75 | Hurt |
| bias init 0 vs median (ablation) | LSH module · CIFAR-100 (similar arch) | top-1 accuracy (%): 75.62 -> 76.04 | No clear effect |
| increase beta from 1 to 7 (ablation) | Ours · CIFAR-100 (WRN-40-2 -> WRN-16-2) | top-1 accuracy (%): 75.33 -> 76.34 | Mixed |
| combine Ours with SSKD | SSKD · CIFAR-100 (vgg13 -> vgg8) | top-1 accuracy (%): 74.90 -> 75.19 | Improved |
| add LSH loss to l2 loss on ImageNet (ablation) | Ours · ImageNet (ResNet34 -> ResNet18) | top-1 error (%): 28.61 -> 28.28 | Improved |
Closest papers by meaning
- Improving Knowledge Distillation via Regularizing Feature Direction and Norm (ECCV 2024 · similarity 0.91)
- Knowledge Distillation for Object Detection via Rank Mimicking and Prediction-Guided Feature Imitation (AAAI 2022 · similarity 0.90)
- CrossKD: Cross-Head Knowledge Distillation for Object Detection (CVPR 2024 · similarity 0.88)
- All You Need in Knowledge Distillation Is a Tailored Coordinate System (AAAI 2025 · similarity 0.88)
- Neural Collapse Inspired Knowledge Distillation (AAAI 2025 · similarity 0.88)
- Less or More From Teacher: Exploiting Trilateral Geometry For Knowledge Distillation (ICLR 2024 · 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.
