Research map / Label-efficient and robust learning / Knowledge distillation
Decoupled Knowledge Distillation
Check 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 |
|---|---|---|---|
| Replace classical KD with DKD | ResNet20 distilled from ResNet56 · CIFAR-100 validation | top-1 accuracy: 70.66 -> 71.97 | Helped |
| Replace classical KD with DKD | ShuffleNet-V1 distilled from ResNet32×4 · CIFAR-100 validation | top-1 accuracy: 74.07 -> 76.45 | Helped |
| Replace classical KD with DKD | ResNet-18 distilled from ResNet-34 · ImageNet validation | top-1 accuracy: 70.66 -> 71.70 | Helped |
| Combine DKD with ReviewKD | Faster R-CNN-FPN with ResNet-18 student · MS-COCO val2017 | AP: 36.75 -> 37.01 | Helped |
| Add NCKD without TCKD (ablation) | ResNet8×4 distilled from ResNet32×4 · CIFAR-100 validation | top-1 accuracy: 72.50 -> 74.26 | Helped |
| Add TCKD without NCKD (ablation) | ResNet8×4 distilled from ResNet32×4 · CIFAR-100 validation | top-1 accuracy: 72.50 -> 68.63 | Hurt |
| Add TCKD to NCKD (ablation) | ResNet8×4 distilled from ResNet32×4 using AutoAugment · CIFAR-100 validation | top-1 accuracy: 73.82 -> 75.33 | Helped |
| Add TCKD to NCKD (ablation) | ResNet8×4 distilled from ResNet32×4 · CIFAR-100 validation, noisy ratio 0.1 | top-1 accuracy: 70.99 -> 70.96 | No clear effect |
| Add TCKD to NCKD (ablation) | ResNet8×4 distilled from ResNet32×4 · CIFAR-100 validation, noisy ratio 0.2–0.3 | top-1 accuracy: 67.55 -> 68.03; 64.62 -> 65.26 | Helped |
| Add TCKD to NCKD (ablation) | ResNet-18 distilled from ResNet-34 · ImageNet validation | top-1 accuracy: 70.71 -> 71.03 | Helped |
| Replace confidence-coupled NCKD weight with fixed β=8.0 (ablation) | DKD on ResNet8×4 distilled from ResNet32×4 · CIFAR-100 validation | top-1 accuracy: 73.63 -> 76.32 | Helped |
| Set α=1.0 rather than remove TCKD with α=0.0 (ablation) | DKD on ResNet8×4 distilled from ResNet32×4, β=8.0 · CIFAR-100 validation | top-1 accuracy: 75.30 -> 76.32 | Helped |
Closest papers by meaning
- Scaled Decoupled Distillation (CVPR 2024 · similarity 0.92)
- VkD: Improving Knowledge Distillation using Orthogonal Projections (CVPR 2024 · similarity 0.90)
- DeepKD: A Deeply Decoupled and Denoised Knowledge Distillation Trainer (NeurIPS 2025 · similarity 0.90)
- Wasserstein Distance Rivals Kullback-Leibler Divergence for Knowledge Distillation (NeurIPS 2024 · similarity 0.89)
- CrossKD: Cross-Head Knowledge Distillation for Object Detection (CVPR 2024 · similarity 0.89)
- DFD: Distilling the Feature Disparity Differently for Detectors (ICML 2024 · similarity 0.89)
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.