Research map / Continual learning / Class-incremental
Maintaining Discrimination and Fairness in Class Incremental Learning
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 |
|---|---|---|---|
| Add knowledge distillation to cross-entropy training | CIFAR-100 class-incremental learning, 5 steps · CIFAR-100, 80 old classes and 20 new classes | old samples misclassified as other old classes: 1,333 -> 1,012 | Helped |
| Add knowledge distillation to cross-entropy training | CIFAR-100 class-incremental learning, 5 steps · CIFAR-100, 80 old classes and 20 new classes | old samples misclassified as new classes: 4,027 -> 4,314 | Hurt |
| Add weight aligning to CE + KD training (ablation) | CIFAR-100 class-incremental learning, 5 steps · CIFAR-100, average over incremental steps | top-1 accuracy (%): 58.2 -> 66.6 | Helped |
| Add weight aligning without KD (ablation) | CE training on CIFAR-100 · CIFAR-100, 5 incremental steps, average | top-1 accuracy (%): 55.3 -> 61.5 | Helped |
| Add KD without weight aligning (ablation) | CE training on CIFAR-100 · CIFAR-100, 5 incremental steps, average | top-1 accuracy (%): 55.3 -> 56.4 | Helped |
| Replace post-training weight aligning with a weight normalization layer during training (ablation) | CE + KD training on CIFAR-100 · CIFAR-100, 5 incremental steps, average | top-1 accuracy (%): 66.6 vs 58.2 | Hurt |
| Restrict fully connected layer weights to be positive (ablation) | Weight Aligning on ImageNet-100 · ImageNet-100, 10 incremental steps | top-5 accuracy (%) | Helped |
| Use 1-norm instead of 2-norm for weight alignment (ablation) | Weight Aligning on ImageNet-100 · ImageNet-100, 10 incremental steps | top-5 accuracy (%) | No clear effect |
| Include a bias term in the fully connected layer (ablation) | Weight Aligning on ImageNet-100 · ImageNet-100, 10 incremental steps | top-5 accuracy (%) | No clear effect |
| Use herding instead of random exemplar selection (ablation) | Weight Aligning on ImageNet-100 · ImageNet-100, 10 incremental steps | top-5 accuracy (%) | No clear effect |
| Add weight aligning to CE + KD training | Class-incremental learning · ImageNet-1000, last incremental step | top-5 accuracy (%): 73.2 -> 81.1 | Helped |
| Use the proposed method instead of BiC | Class-incremental learning · ImageNet-100, last incremental step | top-5 accuracy (%): 84.4 -> 84.1 | Hurt |
Closest papers by meaning
- Gradient Reweighting: Towards Imbalanced Class-Incremental Learning (CVPR 2024 · similarity 0.89)
- Class Incremental Learning with Multi-Teacher Distillation (CVPR 2024 · similarity 0.89)
- Defying Imbalanced Forgetting in Class Incremental Learning (AAAI 2024 · similarity 0.87)
- Continual Learning Using a Kernel-Based Method Over Foundation Models (AAAI 2025 · similarity 0.87)
- Maintaining Fairness in Logit-based Knowledge Distillation for Class-Incremental Learning (AAAI 2025 · similarity 0.87)
- Debiased Distillation for Consistency Regularization (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.