Research map / Continual learning / Class-incremental
Pseudo Informative Episode Construction for Few-Shot Class-Incremental Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PIEC, a meta-learning framework for few-shot class-incremental learning that builds pseudo incremental sessions from distribution-level mixing of base-class Gaussians, filtered by novel-novel and base-novel diversity criteria.
Problem
Few-shot class-incremental learning (FSCIL) where existing meta-learning methods build pseudo novel classes from base classes, which fail to mimic the feature distribution of real novel classes.
Approach
Generate pseudo novel classes by pairwise mixing of base-class Gaussian distributions in the novel-class feature space, then select informative ones via novel-novel dissimilarity and KL-divergence information gain over base classes to build pseudo incremental sessions for episodic training with FEAT-style metric inference.
Key results
- On CUB200 10-way 5-shot, PIEC reaches 59.47% final accuracy with PD 16.41%, versus LIMIT's PD of 18.48% and CEC's 23.57%.
- Relative PD improvements over meta-learning methods on miniImageNet (21.04% PD) and CIFAR100 (21.36% PD) are reported against CEC, MetaFSCIL and LIMIT.
- Ablation: removing both diversity criteria increases CUB200 PD by 2.04 points (18.45% vs 16.41%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add novel-novel diversity criteria for selecting pseudo novel classes (ablation) | PIEC · CUB200 | performance dropping (PD): 16.41% vs. 17.43% | Helped |
| add base-novel diversity criteria for selecting pseudo novel classes (ablation) | PIEC · CUB200 | performance dropping (PD): 16.41% vs. 17.77% | Helped |
| add both diversity criteria instead of randomly selecting pseudo novel classes (ablation) | PIEC · CUB200 | performance dropping (PD): 18.45% to 16.41% | Helped |
| vary the Beta-distribution parameter β for pairwise mixing (ablation) | PIEC · CUB200 | performance dropping (PD) | Mixed |
| vary the pseudo novel class selection ratio r (ablation) | PIEC · CUB200 | performance dropping (PD) | Mixed |
| increase the number of pseudo incremental sessions beyond 2 (ablation) | PIEC · CUB200 | accuracy | Helped |
Limitations
- Evaluated only on three small classification benchmarks with fixed session splits.
- Sensitive to hyperparameters such as beta and selection ratio r (performance varies in ablations).
- Does not address the more realistic setting where new samples may belong to previously seen classes.
Experimental setup
Datasets: CUB200 miniImageNet CIFAR100
Benchmarks: CUB200 miniImageNet CIFAR100
Metrics: accuracy per session performance drop (PD)
Baselines: Finetune-CNN iCaRL EEIL NCM TOPIC SF-FSCIL ERDFR FACT DSN SoftNet WaRP MCNet
Builds on: FEAT CEC LIMIT MetaFSCIL
Closest papers by meaning
- Few-Shot Incremental Learning With Continually Evolved Classifiers (CVPR 2021 · similarity 0.91)
- MetaFSCIL: A Meta-Learning Approach for Few-Shot Class Incremental Learning (CVPR 2022 · similarity 0.90)
- Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning (CVPR 2025 · similarity 0.88)
- M2SD:Multiple Mixing Self-Distillation for Few-Shot Class-Incremental Learning (AAAI 2024 · similarity 0.87)
- Adaptive Decision Boundary for Few-Shot Class-Incremental Learning (AAAI 2025 · similarity 0.87)
- Pre-trained Vision and Language Transformers Are Few-Shot Incremental Learners (CVPR 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.