Research map / Continual learning / Class-incremental
Few-shot Class-Incremental Learning via Generative Co-Memory Regularization
Kexin Bao, Yong Li, Dan Zeng, Shiming Ge
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 co-memory regularization to incremental learning with ResNet12 (vs C-FSCIL) | ResNet12 FSCIL · CIFAR100 FSCIL | avg accuracy over sessions: 61.64 -> 67.65 | Improved |
| add co-memory regularization versus NC-FSCIL with ResNet12 | ResNet12 FSCIL · MiniImageNet FSCIL | avg accuracy over sessions: 67.82 -> 68.04 | Improved |
| add weight memory and representation memory together (beta 0.8) vs weight memory only (ablation) | GCMR ViT-B · MiniImageNet FSCIL | avg accuracy over sessions: 66.84 -> 70.26 | Improved |
| use beta 1.0 (representation memory only) vs beta 0.0 (ablation) | GCMR ViT-B · MiniImageNet FSCIL | avg accuracy over sessions: 66.84 -> 70.55 | Improved |
| combine memories with beta 0.7 vs weight memory only (ablation) | GCMR ViT-B · CIFAR100 FSCIL | avg accuracy over sessions: 68.48 -> 69.77 | Improved |
| replace ResNet18 with ViT-B encoder at beta 0.3 (ablation) | GCMR · MiniImageNet FSCIL | avg accuracy over sessions: 68.15 -> 68.15 | No clear effect |
| replace ResNet18 with ViT-B encoder at beta 0.3 (ablation) | GCMR · CIFAR100 FSCIL | avg accuracy over sessions: 68.26 -> 69.12 | Improved |
| use ResNet34 instead of ResNet18 encoder (ablation) | GCMR beta 0.3 · MiniImageNet FSCIL | avg accuracy over sessions: 68.15 -> 68.32 | Improved |
| add generative domain adaptation finetuning (ablation) | GCMR base session · MiniImageNet/CIFAR100/CUB200 | base session accuracy: +0.47 to +1.49 | Improved |
| add pretraining of the ViT encoder (ablation) | GCMR base session · MiniImageNet | base session accuracy: over 40% than that without pretraining | Improved |
| GCMR vs NC-FSCIL base session on CUB200 | GCMR · CUB200 FSCIL | base session accuracy: 80.45 -> 80.64 | Improved |
| large pretrained models (PriViLege) vs GCMR | PriViLege · CIFAR100 FSCIL | last session accuracy: 58.84 -> 86.06 | 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.
- OrCo — OrCo: Towards Better Generalization via Orthogonality and Contrast for Few-Shot Class-Incremental Learning (CVPR 2024) · better on 2 of 2 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
- Attraction Diminishing and Distributing for Few-Shot Class-Incremental Learning (CVPR 2025 · similarity 0.87)
- Subspace Regularizers for Few-Shot Class Incremental Learning (ICLR 2022 · similarity 0.85)
- Few-Shot Class-Incremental Learning via Class-Aware Bilateral Distillation (CVPR 2023 · similarity 0.85)
- Rethinking Few-shot Class-incremental Learning: Learning from Yourself (ECCV 2024 · similarity 0.85)
- Generative Multi-modal Models are Good Class Incremental Learners (CVPR 2024 · similarity 0.84)
- OrCo: Towards Better Generalization via Orthogonality and Contrast for Few-Shot Class-Incremental Learning (CVPR 2024 · similarity 0.83)
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.
