Research map / Continual learning / Class-incremental
Few-Shot Class-Incremental Learning by Sampling Multi-Phase Tasks
Da-Wei Zhou, Han-Jia Ye, Liang Ma, Di Xie, Shiliang Pu, 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 |
|---|---|---|---|
| add prototype-initialized classifier (ablation) | fine-tuning baseline · CIFAR100 (FSCIL) | final accuracy: 2.65 -> 46.98 | Improved |
| add meta-calibration module (transformer) (ablation) | Prototype classifier · CIFAR100 (FSCIL) | final accuracy: 46.98 -> 48.30 | Improved |
| add one-phase fake-incremental meta-training (ablation) | Prototype + Calibration · CIFAR100 (FSCIL) | final accuracy: 48.30 -> 50.01 | Improved |
| use two-phase instead of one-phase fake tasks (ablation) | LIMIT · CIFAR100 (FSCIL) | final accuracy: 50.01 -> 51.23 | Improved |
| use LIMIT fake-incremental task sampling instead of CEC PIL (ablation) | CEC · CUB200 (FSCIL) | final accuracy: 52.28 -> 53.01 | Improved |
| use LIMIT fake-incremental task sampling instead of CEC PIL (ablation) | LIMIT · CUB200 (FSCIL) | final accuracy: 56.37 -> 57.41 | Improved |
| use LIMIT fake-incremental task sampling instead of CEC PIL (ablation) | LIMIT · CIFAR100 (FSCIL) | final accuracy: 51.03 -> 51.23 | Improved |
| use AutoAug instead of BasicAug (ablation) | LIMIT · CIFAR100 (FSCIL) | final accuracy: 49.92 -> 51.23 | Improved |
| use AutoAug instead of BasicAug (ablation) | LIMIT · CUB200 (FSCIL) | final accuracy: 58.45 -> 57.41 | Hurt |
| use AutoAug instead of BasicAug (ablation) | CEC · CUB200 (FSCIL) | final accuracy: 52.28 -> 51.38 | Hurt |
| use AutoAug instead of BasicAug (ablation) | CEC · miniImageNet (FSCIL) | final accuracy: 47.63 -> 47.97 | Improved |
| replace CEC with LIMIT (multi-phase fake-task sampling plus transformer calibration) | CEC · CUB200 (FSCIL) | incremental-class accuracy: 33.9 -> 41.8 | Improved |
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 3 of 3 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
- Few-Shot Class-Incremental Learning via Training-Free Prototype Calibration (NeurIPS 2023 · similarity 0.94)
- Few-Shot Incremental Learning With Continually Evolved Classifiers (CVPR 2021 · similarity 0.92)
- Few-Shot Class-Incremental Learning via Class-Aware Bilateral Distillation (CVPR 2023 · similarity 0.92)
- Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning (CVPR 2025 · similarity 0.91)
- Pseudo Informative Episode Construction for Few-Shot Class-Incremental Learning (AAAI 2025 · similarity 0.90)
- Adaptive Decision Boundary for Few-Shot Class-Incremental Learning (AAAI 2025 · similarity 0.89)
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.
