Research map / Continual learning / Class-incremental
Balancing between Forgetting and Acquisition in Incremental Subpopulation Learning
Mingfu Liang, Jiahuan Zhou, Wei Wei, Ying Wu
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 cross-entropy with the margin-enforce loss when finetuning the last layer (ablation) | last-layer finetuning · Entity-13 (13 steps, uneven) | average top-1 accuracy on all subclasses (%): 55.23 -> 60.38 | Improved |
| add Stage-2 with a random alpha_t to the margin loss (ablation) | margin-loss classifier · Entity-13 (13 steps, uneven) | average top-1 accuracy on all subclasses (%): 60.38 -> 71.59 | Improved |
| add Stage-2 with fixed alpha_t = 1 to the margin loss (ablation) | margin-loss classifier · Entity-13 (13 steps, uneven) | average top-1 accuracy on all subclasses (%): 60.38 -> 75.24 | Improved |
| search alpha_t with the acquisition-minus-forgetting objective instead of fixed alpha_t = 1 (ablation) | two-stage method · Entity-13 (13 steps, uneven) | average top-1 accuracy on all subclasses (%): 75.24 -> 78.75 | Improved |
| search alpha_t with the objective instead of random alpha_t (ablation) | two-stage method · Entity-13 (13 steps, uneven) | average top-1 accuracy on all subclasses (%): 71.59 -> 78.75 | Improved |
| add Stage-2 alpha search to the margin loss: forgetting (ablation) | two-stage method · Entity-13 (13 steps, uneven) | average forgetting F13 (%): 29.48 -> 3.35 | Improved |
| use the full two-stage method instead of PASS | two-stage method vs PASS · Entity-13 (13 steps, uneven) | average top-1 accuracy on all subclasses (%): 52.49 -> 78.75 | Improved |
| use the full two-stage method instead of PASS | two-stage method vs PASS · Entity-30 (15 steps, uneven) | average top-1 accuracy on all subclasses (%): 39.75 -> 71.69 | Improved |
| use the full two-stage method instead of PASS | two-stage method vs PASS · Entity-30 (8 steps, uneven) | average top-1 accuracy on all subclasses (%): 54.99 -> 72.14 | Improved |
| use the full two-stage method instead of PASS: unseen accuracy, 4 steps | two-stage method vs PASS · Entity-30 (4 steps, even) | accuracy on unseen subclasses (%): 64.50 -> 64.73 | No clear effect |
| use the full two-stage method instead of PASS: unseen accuracy, 5 steps | two-stage method vs PASS · Entity-13 (5 steps, even) | accuracy on unseen subclasses (%): 73.12 -> 72.02 | Hurt |
| use the full two-stage method instead of PASS: forgetting, 15 steps | two-stage method vs PASS · Entity-30 (15 steps, uneven) | average forgetting F15 (%): 58.27 -> 3.48 | Improved |
Closest papers by meaning
- Gradient Reweighting: Towards Imbalanced Class-Incremental Learning (CVPR 2024 · similarity 0.87)
- Long-Tail Class Incremental Learning via Independent Sub-prototype Construction (CVPR 2024 · similarity 0.87)
- Make Domain Shift a Catastrophic Forgetting Alleviator in Class-Incremental Learning (AAAI 2025 · similarity 0.86)
- Defying Imbalanced Forgetting in Class Incremental Learning (AAAI 2024 · similarity 0.86)
- AnaCP: Toward Upper-Bound Continual Learning via Analytic Contrastive Projection (NeurIPS 2025 · similarity 0.86)
- GACL: Exemplar-Free Generalized Analytic Continual Learning (NeurIPS 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.
