Research map / Continual learning / Class-incremental
DER: Dynamically Expandable Representation for 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 |
|---|---|---|---|
| Expand the representation with a new feature extractor while freezing previously learned extractors (ablation) | DER · CIFAR100-B0, 10 steps | average accuracy: 61.84 -> 73.26 | Helped |
| Add the auxiliary loss for novel-class features (ablation) | DER with expandable representation · CIFAR100-B0, 10 steps | average accuracy: +2.10% | Helped |
| Remove channel-level mask-based pruning (ablation) | DER · CIFAR100-B0, 10 steps | average accuracy: 74.64\colorpm{0.28} -> 75.36\colorpm{0.36}; 4.96 -> 61.6 million parameters | Mixed |
| Remove channel-level mask-based pruning (ablation) | DER · CIFAR100-B0, 50 steps | average accuracy: 72.05\colorpm{0.55} -> 72.41\colorpm{0.36}; 10.15 -> 285.6 million parameters | Mixed |
| Remove channel-level mask-based pruning (ablation) | DER · CIFAR100-B50, 10 steps | average accuracy: 72.45\colorpm{0.76} -> 72.81\colorpm{0.88}; 8.79 -> 67.2 million parameters | Mixed |
| Remove channel-level mask-based pruning (ablation) | DER · ImageNet100-B0 | top-1 average accuracy: 76.12 -> 77.18; 7.67 -> 61.6 million parameters | Mixed |
| Remove channel-level mask-based pruning (ablation) | DER · ImageNet1000-B0 | top-1 average accuracy: 66.73 -> 68.84; 14.52 -> 61.6 million parameters | Mixed |
| Add the auxiliary loss (ablation) | DER · CIFAR100-B0, 10 steps | last accuracy: 63.07 -> 65.34 | Helped |
| Use expandable representation instead of the non-expanded baseline (ablation) | DER · CIFAR100-B0, 10 steps | last accuracy: 40.81 -> 63.07 | Helped |
| Increase the auxiliary-loss coefficient from 1 to 10 (ablation) | DER · CIFAR100-B0, 10 steps | average accuracy: 74.64\pm{0.28} -> 73.54\pm{0.22} | Hurt |
Closest papers by meaning
- Self-Sustaining Representation Expansion for Non-Exemplar Class-Incremental Learning (CVPR 2022 · similarity 0.91)
- Class-Incremental Learning With Strong Pre-Trained Models (CVPR 2022 · similarity 0.89)
- On the Power of Statistics in Class-Incremental Learning with Pretrained Models (ICML 2026 · similarity 0.88)
- Compress to One Point: Neural Collapse for Pre-Trained Model-Based Class-Incremental Learning (AAAI 2025 · similarity 0.87)
- Harnessing Neural Unit Dynamics for Effective and Scalable Class-Incremental Learning (ICML 2024 · similarity 0.87)
- Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning (CVPR 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.