Research map / Continual learning / Class-incremental
The Lie of the Average: How Class Incremental Learning Evaluation Deceives You?
Guannan Lai, Da-Wei Zhou, Xin Yang, Han-Jia Ye
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 |
|---|---|---|---|
| use EDGE (easy/median/hard sequences by inter-task similarity) instead of 3 random sequences | Hide-Prompt · CIFAR-100 (6 classes, 3 tasks) | Wasserstein distance: 3.89 -> 1.42 | Improved |
| use EDGE instead of random sampling | CODA-Prompt · ImageNet-R (6 classes, 3 tasks) | Wasserstein distance: 9.85 -> 2.37 | Improved |
| use EDGE instead of random sampling | EWC · CIFAR-100 (6 classes, 3 tasks) | estimated lower bound accuracy (true 12.50): 26.17 -> 12.50 | Improved |
| use EDGE instead of random sampling on stable models (EASE, RanPAC) | EASE, RanPAC · CIFAR-100 and ImageNet-R (6 classes) | JSD / Wasserstein: 0.00 vs 0.00 JSD on CIFAR-100 | No clear effect |
| use EDGE instead of random sampling | DER · CIFAR-100 (6 classes, 3 tasks) | Jensen-Shannon divergence: 0.29 -> 0.31 | Hurt |
| vary CLIP text encoder scale and backbone (ResNet vs ViT) in EDGE (ablation) | EDGE · Hide-Prompt and TagFex enumerable | distribution estimate closeness | No clear effect |
| add extreme sequences to the evaluation sample | random sampling evaluation · theory | required number of sequences L: 2x10^4 -> around 50 | Improved |
| evaluate on easier class sequence with higher inter-task similarity | CIL methods · CIFAR-100 and ImageNet-R enumerable | accuracy vs similarity | Improved |
Closest papers by meaning
- Navigating Semantic Drift in Task-Agnostic Class-Incremental Learning (ICML 2025 · similarity 0.88)
- Class Incremental Learning via Likelihood Ratio Based Task Prediction (ICLR 2024 · similarity 0.88)
- Long-Tail Class Incremental Learning via Independent Sub-prototype Construction (CVPR 2024 · similarity 0.87)
- Prediction Error-based Classification for Class-Incremental Learning (ICLR 2024 · similarity 0.87)
- Mixture of Noise for Pre-Trained Model-Based Class-Incremental Learning (NeurIPS 2025 · similarity 0.87)
- Defying Imbalanced Forgetting in Class Incremental Learning (AAAI 2024 · 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.
