Research map / Continual learning / Class-incremental
FOSTER: Feature Boosting and Compression for Class-Incremental Learning
Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, De-Chuan Zhan
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 |
|---|---|---|---|
| replace weight alignment with logits alignment (ablation) | FOSTER · CIFAR-100 B50 5 steps | final accuracy: about 4% | Improved |
| add feature enhancement (ablation) | FOSTER · CIFAR-100 B50 5 steps | last-stage accuracy: more than 3% decline without it | Improved |
| add feature enhancement (ablation) | FOSTER · CIFAR-100 B50 5 steps (old classes, last step) | accuracy: more than 4 % | Improved |
| add feature enhancement (ablation) | FOSTER · CIFAR-100 B50 5 steps (new classes, last step) | accuracy | No clear effect |
| use balanced distillation instead of normal KD (ablation) | FOSTER compression · CIFAR-100 B50 5 steps | incremental accuracy | Improved |
| vary beta from 0.93 to 0.999 (ablation) | FOSTER · CIFAR-100 B50 5 steps | accuracy: 0.93, 0.95, 0.97, 0.99, 0.995, 0.999 | No clear effect |
| increase exemplars per class from 5 to 200 (ablation) | FOSTER · CIFAR-100 B50 5 steps | last-step accuracy: 53.53% -> 71.4% | Improved |
| use only 10 exemplars per class (ablation) | FOSTER · CIFAR-100 B50 5 steps | last-round accuracy: more than 60% | Improved |
| initialize O by fine-tuning instead of all zero (ablation) | FOSTER boosting · CIFAR-100 B50 5 steps | accuracy | No clear effect |
| add bias learning to all-zero O (AZB) (ablation) | FOSTER boosting · CIFAR-100 B50 5 steps | accuracy | No clear effect |
| compress dual-branch model to single backbone | FOSTER B4 · CIFAR-100 and ImageNet-100 | incremental accuracy | No clear effect |
| FOSTER vs DER | FOSTER · ImageNet-1000 (10 steps) | average accuracy: 66.73% to 68.34% (+1.61%) | Improved |
Later papers that compared with this method
15 later accepted papers report FOSTER as a baseline in their results tables; 14 of them report a better number on most of the results they share with it.
- Forget Forgetting: Continual Learning in a World of Abundant Memory (ICLR 2026 · better on 3 of 7 shared results)
- Temporal Imbalance of Positive and Negative Supervision in Class-Incremental Learning (CVPR 2026 · better on 3 of 3 shared results)
- Plug-and-Play Compositionality for Boosting Continual Learning with Foundation Models (ICLR 2026 · better on 1 of 1 shared result)
- IncreFA: Breaking the Static Wall of Generative Model Attribution (CVPR 2026 · better on 1 of 1 shared result)
- Beyond Sharpness: A Flatness Decomposition Framework for Efficient Continual Learning (AAAI 2026 · better on 1 of 1 shared result)
- C-Flat++: Towards a More Efficient and Powerful Framework for Continual Learning (IJCV 2026 · better on 1 of 1 shared result)
- Reducing Class-wise Confusion for Incremental Learning with Disentangled Manifolds (CVPR 2025 · better on 2 of 2 shared results)
- Expandable Subspace Ensemble for Pre-Trained Model-Based Class-Incremental Learning (CVPR 2024 · better on 4 of 4 shared results)
7 more not listed.
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- iCaRL — iCaRL: Incremental Classifier and Representation Learning (CVPR 2017) · better on 7 of 7 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Class-Incremental Learning by Knowledge Distillation With Adaptive Feature Consolidation (CVPR 2022 · similarity 0.88)
- End-to-End Incremental Learning (ECCV 2018 · similarity 0.86)
- Compress to One Point: Neural Collapse for Pre-Trained Model-Based Class-Incremental Learning (AAAI 2025 · similarity 0.86)
- DER: Dynamically Expandable Representation for Class Incremental Learning (CVPR 2021 · similarity 0.86)
- Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning (CVPR 2025 · similarity 0.85)
- Pre-trained Vision and Language Transformers Are Few-Shot Incremental Learners (CVPR 2024 · similarity 0.84)
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.
