Research map / Continual learning / Class-incremental
Class-Incremental Learning With Strong Pre-Trained Models
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 |
|---|---|---|---|
| Use feature augmentation and score fusion rather than learning-based routing | ResNet10 CIL · ImageNet disjoint-CIL, 800 base/40 novel | Acc_avg: 65.57 -> 70.06 | Helped |
| Use feature augmentation and score fusion rather than learning-based routing | ResNet10 CIL · ImageNet overlapping-CIL, style-changing split | Acc_avg: 61.09 -> 72.56 | Helped |
| Max-pool overlapping-class logits instead of average-pooling them (ablation) | Logit concatenation · ImageNet overlapping-CIL, style-changing split | Acc_ovlp: 40.40 -> 69.60 | Helped |
| Train layer3 as well as layer4 in stage I (ablation) | Score fusion · ImageNet domain-changing disjoint-CIL | Acc_avg: 67.93 -> 69.18 | Helped |
| Branch only at the classifier layer instead of adapting the last backbone block (ablation) | Score fusion · ImageNet disjoint-CIL, 800 base/40 novel | Acc_avg: 70.06 -> 64.40 | Hurt |
| Replace score fusion with BiC calibration after feature augmentation (ablation) | Feature augmentation (FA) · ImageNet disjoint-CIL, 800 base/40 novel | Acc_avg: 70.06 -> 58.93 | Hurt |
| Replace score fusion with weight alignment after feature augmentation (ablation) | Feature augmentation (FA) · ImageNet disjoint-CIL, 800 base/40 novel | Acc_avg: 70.06 -> 49.25 | Hurt |
| Fine-tune concatenated classifier weights rather than preserve the original classifier weights (ablation) | Logit concatenation · ImageNet disjoint-CIL, 800 base/40 novel | Acc_avg: 70.06 -> 59.95 | Hurt |
| Retrain concatenated classifier weights from scratch rather than preserve the original classifier weights (ablation) | Logit concatenation · ImageNet disjoint-CIL, 800 base/40 novel | Acc_avg: 70.06 -> 56.43 | Hurt |
| Optimize the operating point for Acc_avg rather than Acc_all (ablation) | Score fusion · ImageNet disjoint-CIL, 800 base/40 novel | Acc_all: 63.24 -> 58.90 | Hurt |
| Optimize the operating point for Acc_avg rather than Acc_all (ablation) | Score fusion · ImageNet disjoint-CIL, 800 base/40 novel | Acc_avg: 58.22 -> 70.06 | Helped |
| Fine-tune more backbone layers when base and novel domains differ (ablation) | Score fusion · ImageNet domain-changing disjoint-CIL | Acc_avg: 66.55 -> 69.18 | Helped |
Closest papers by meaning
- On the Power of Statistics in Class-Incremental Learning with Pretrained Models (ICML 2026 · similarity 0.93)
- A Model or 603 Exemplars: Towards Memory-Efficient Class-Incremental Learning (ICLR 2023 · similarity 0.90)
- Mimicking the Oracle: An Initial Phase Decorrelation Approach for Class Incremental Learning (CVPR 2022 · similarity 0.90)
- Cross-Class Feature Augmentation for Class Incremental Learning (AAAI 2024 · similarity 0.89)
- Prediction Error-based Classification for Class-Incremental Learning (ICLR 2024 · similarity 0.88)
- Gradient Reweighting: Towards Imbalanced Class-Incremental Learning (CVPR 2024 · similarity 0.88)
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.