Research map / Label-efficient and robust learning / Knowledge distillation
Streamlined Knowledge Distillation
Hyeon-Jin Jeong, Han-Jin Lee, Seok-Hwan Choi
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 |
|---|---|---|---|
| apply Euclidean Batch Normalization to logits before Gramian (ablation) | SKD · CIFAR-100 (ResNet32x4 -> ResNet8x4) | top-1 accuracy: 76.61 -> 77.28 | Improved |
| replace L2-norm with Mahalanobis distance direction loss (ablation) | SKD · CIFAR-100 (ResNet32x4 -> ResNet8x4) | top-1 accuracy: 77.28 -> 77.76 | Improved |
| add Tikhonov regularization and Cholesky decomposition (ablation) | SKD · CIFAR-100 (ResNet32x4 -> ResNet8x4) | top-1 accuracy: 77.76 -> 78.33 | Improved |
| add SKD loss to FitNet | FitNet · CIFAR-100 (ResNet32x4 -> ResNet8x4) | top-1 accuracy: 75.26 -> 77.01 | Improved |
| add SKD loss to RKD | RKD · CIFAR-100 (VGG13 -> VGG8) | top-1 accuracy: 71.61 -> 75.16 | Improved |
| SKD vs best prior logit-based on ResNet32x4 -> ResNet8x4 | SKD · CIFAR-100 (ResNet32x4 -> ResNet8x4) | top-1 accuracy: 76.64 -> 78.33 | Improved |
| SKD vs MLKD on WRN-40-2 -> WRN-16-2 | SKD · CIFAR-100 (WRN-40-2 -> WRN-16-2) | top-1 accuracy: 76.83 -> 76.60 | Hurt |
| SKD vs ReviewKD on ResNet101 -> ResNet18 detection | SKD · MS-COCO detection (ResNet101 -> ResNet18) | AP: 36.75 -> 36.74 | No clear effect |
| SKD vs MLKD on ImageNet R50 -> MobileNetV1 | SKD · ImageNet (R50 -> MV1) | top-1 accuracy: 71.08 -> 71.53 | Improved |
| SKD vs DeiT distillation for ViT-Tiny student | SKD · ImageNet (RegNetY-16GF -> ViT-Tiny) | top-1 accuracy: 68.03 -> 68.68 | Improved |
Closest papers by meaning
- Harmonizing knowledge Transfer in Neural Network with Unified Distillation (ECCV 2024 · similarity 0.91)
- Scaled Decoupled Distillation (CVPR 2024 · similarity 0.90)
- Wasserstein Distance Rivals Kullback-Leibler Divergence for Knowledge Distillation (NeurIPS 2024 · similarity 0.88)
- Maintaining Fairness in Logit-based Knowledge Distillation for Class-Incremental Learning (AAAI 2025 · similarity 0.87)
- Enhancing Logits Distillation with Plug&Play Kendall's $\tau$ Ranking Loss (ICML 2025 · similarity 0.87)
- DA-KD: Difficulty-Aware Knowledge Distillation for Efficient Large Language Models (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.
