Research map / Architectures and efficiency / CNN design
NORM: Knowledge Distillation via N-to-One Representation Matching
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use NORM instead of individually training the student | ResNet18 · ImageNet, ResNet34→ResNet18 | top-1 accuracy: 70.13% to 72.14% | Helped |
| Use NORM instead of individually training the student | MobileNet · ImageNet, ResNet50→MobileNet | top-1 accuracy: 4.63% gain | Helped |
| Expand student features to N=8 rather than N=1 (ablation) | NORM · CIFAR-100, ResNet32x4→ResNet8x4 | top-1 accuracy: 3.02% to N=1 and 3.62% to the baseline student model | Helped |
| Match the teacher with all 8 expanded student feature segments instead of 1 segment (ablation) | NORM · CIFAR-100, ResNet32x4→ResNet8x4 | top-1 accuracy: 0.98% | Helped |
| Add vanilla logits-based KD loss | NORM · CIFAR-100 | top-1 accuracy: 0.46% extra gain in average | Helped |
| Add contrastive KD loss | NORM · CIFAR-100 | top-1 accuracy: 0.18% extra gain in average | Helped |
| Use two matching layers instead of only the default last convolutional layer (ablation) | NORM · ImageNet, ResNet34→ResNet18 | top-1 accuracy | Hurt |
| Add one plain linear FT module without a residual connection (ablation) | ResNet18 trained individually · ImageNet | top-1 accuracy: 0.93% | Hurt |
| Contract teacher features instead of expanding student features (ablation) | NORM · ImageNet, ResNet34→ResNet18 | top-1 accuracy: 0.60% and 0.94% | Hurt |
| Replace the l2-norm loss with l1-norm (ablation) | NORM · CIFAR-100, ResNet32x4→ResNet8x4 | top-1 accuracy: 75.10 vs. 76.49 | Hurt |
| Replace the l2-norm loss with MMD (ablation) | NORM · CIFAR-100, ResNet32x4→ResNet8x4 | top-1 accuracy: 76.62; 0.13% improvement to l2-norm | Helped |
| Replace the learnable ensemble layer with fixed averaging (ablation) | NORM · ImageNet, ResNet34→ResNet18 | top-1 accuracy: 72.14 to 71.91 | Hurt |
Closest papers by meaning
- Knowledge distillation via softmax regression representation learning (ICLR 2021 · similarity 0.89)
- Structural Knowledge Distillation for Object Detection (NeurIPS 2022 · similarity 0.89)
- You Don’t Need Aligned Representations: Knowledge Distillation via Random Prototype Spaces (NeurIPS 2026 · similarity 0.89)
- SigLino: Efficient Multi-Teacher Distillation for Agglomerative Vision Foundation Models (CVPR 2026 · similarity 0.88)
- ScaleKD: Strong Vision Transformers Could Be Excellent Teachers (NeurIPS 2024 · similarity 0.88)
- DFD: Distilling the Feature Disparity Differently for Detectors (ICML 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.