Research map / Label-efficient and robust learning / Knowledge distillation
Regularizing Class-Wise Predictions via Self-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 |
|---|---|---|---|
| match predictions across different samples with the same label using CS-KD | ResNet-18 and DenseNet-121 · CIFAR-100, TinyImageNet, CUB-200-2011, Stanford Dogs, MIT67 | top-1 error (%): 24.71 -> 21.99; 43.53 -> 41.62; 46.00 -> 33.28; 36.29 -> 30.85; 44.75 -> 40.45 (ResNet-18) | Helped |
| use CS-KD instead of cross-entropy | ResNet-50, ResNet-101, ResNeXt-101-32x4d · ImageNet | top-1 error (%): 24.0 -> 23.6; 22.4 -> 22.0; 21.6 -> 21.2 | Helped |
| add CS-KD to Mixup (ablation) | ResNet-18 with Mixup · CUB-200-2011 | top-1 error (%): 37.09% to 30.71% | Helped |
| add CS-KD to knowledge distillation (ablation) | ResNet-10 student with KD · CUB-200-2011 | top-1 error (%): 39.32% to 34.47% | Helped |
| add sample-wise consistency regularization to CS-KD (ablation) | CS-KD-E · CIFAR-100 | ECE (%): 5.17% to 4.69% | Helped |
| add sample-wise consistency regularization to CS-KD (ablation) | CS-KD-E · CIFAR-100 | top-5 error (%): 5.69 ± 0.03 -> 5.93 ± 0.06 | Hurt |
| add sample-wise consistency regularization to CS-KD (ablation) | CS-KD-E · TinyImageNet | ECE (%): 7.26 ± 0.93 -> 3.79 ± 0.35 | Helped |
| vary the class-wise loss weight and temperature (ablation) | CS-KD · CIFAR-100 | top-1 error (%): 21.66 to 37.50 | Mixed |
| increase the class-wise loss weight to extreme values (ablation) | CS-KD with T=4 · CIFAR-100 | top-1 error (%): 21.66 -> 32.25; 21.66 -> 35.45 | Hurt |
| replace CS-KD with other output regularizers | ResNet-18 · CUB-200-2011 | top-1 error (%): 46.00 -> 33.28 | Helped |
| use CS-KD rather than cross-entropy | ResNet-18 · CIFAR-100 | ECE (%): 15.45% to 5.17% | Helped |
| use CS-KD rather than cross-entropy | ResNet-18 · CIFAR-100 | top-5 error (%): 6.91% to 5.69% | Helped |
Closest papers by meaning
- Self-Distillation From the Last Mini-Batch for Consistency Regularization (CVPR 2022 · similarity 0.89)
- Debiased Distillation for Consistency Regularization (AAAI 2025 · similarity 0.88)
- Data-Distortion Guided Self-Distillation for Deep Neural Networks (AAAI 2019 · similarity 0.88)
- Dataset Distillation via Knowledge Distillation: Towards Efficient Self-Supervised Pre-training of Deep Networks (ICLR 2025 · similarity 0.87)
- Boost Self-Supervised Dataset Distillation via Parameterization, Predefined Augmentation, and Approximation (ICLR 2025 · similarity 0.86)
- Training Deep Neural Networks with Virtual Smoothing Classes (AAAI 2025 · similarity 0.86)
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.