Lune

CVPR Findings2026

Training-Free Uncertainty-guided Logit Adjustment for Few-Shot Class-Incremental Learning

Sungwon Woo, Dongjun Hwang, Shiwon Kim, Junsuk Choe, Jongho Nang

Abstract

Few-shot Class-Incremental Learning (FSCIL) aims to learn new classes from limited samples while preserving knowledge of previously learned base classes. Most existing FSCIL methods adopt an incremental-frozen strategy, which freezes the feature extractor after the base session to maintain a stable embedding space. However, this design inherently induces a group-wise disparity in predictive uncertainty-incremental classes exhibit consistently higher uncertainty than base classes-leading to prediction bias toward base classes. To address this, we systematically probe inherent uncertainty in FSCIL and introduce two uncertainty measures tailored for this setting: (1) a grouped margin uncertainty, which captures ambiguity near the decision boundary between base and incremental classes, and (2) a base-entropy uncertainty, which computes softmax entropy only over the subset of base classes to quantify confidence within the well-learned base region. We leverage these two complementary cues to perform uncertainty-guided logit adjustment without modifying existing training pipelines. Extensive experiments demonstrate that our method consistently improves incremental accuracy and mitigates prediction bias, while also enhancing comprehensive performance metrics such as aAcc, gAcc, and hAcc. Code is available at https://github.com/sungwon-woo/ULA FSCIL.

Ask Lune

Know this paper in minutes, not an afternoon.

Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.

Questions to start from

Lune can draw on

the full paper · related work

Get the answer

Free to start. No credit card required.

Lineage

  • Forward Compatible Few-Shot Class-Incremental Learning cites IL2M: Class Incremental Learning With Dual Memory
  • Forward Compatible Few-Shot Class-Incremental Learning cites Few-Shot Lifelong Learning
Builds onCited byLeft to right by year; the more cited, the higher.

Builds on19

  1. Variational Adversarial Active LearningSamarth Sinha, Sayna Ebrahimi, Trevor DarrellICCV 2019 · 662 citations
  2. IL2M: Class Incremental Learning With Dual MemoryEden Belouadah, Adrian PopescuICCV 2019 · 385 citations
  3. Forward Compatible Few-Shot Class-Incremental LearningDa-Wei Zhou, Fu-Yun Wang, Han-Jia Ye, Liang Ma et al.CVPR 2022 · 259 citations
  4. Few-Shot Class-Incremental Learning via Relation Knowledge DistillationSonglin Dong, Xiaopeng Hong, Xiaoyu Tao, Xinyuan Chang et al.AAAI 2021 · 215 citations
  5. Few-Shot Lifelong LearningPratik Mazumder, Pravendra Singh, Piyush RaiAAAI 2021 · 153 citations

Related papers

  1. Adaptive Decision Boundary for Few-Shot Class-Incremental LearningLinhao Li, Yongzhang Tan, Siyuan Yang, Hao Cheng et al.AAAI 2025 · 10 citations
  2. Few-Shot Class-Incremental Learning via Training-Free Prototype CalibrationQi-Wei Wang, Da-Wei Zhou, Yi-Kai Zhang, De-Chuan Zhan et al.NeurIPS 2023 · 140 citations
  3. Margin-Based Few-Shot Class-Incremental Learning with Class-Level Overfitting MitigationYixiong Zou, Shanghang Zhang, Yuhua Li, Ruixuan LiNeurIPS 2022 · 100 citations
  4. Few-Shot Class-Incremental Learning via Class-Aware Bilateral DistillationLinglan Zhao, Jing Lu, Yunlu Xu, Zhanzhan Cheng et al.CVPR 2023
  5. Learning with Fantasy: Semantic-Aware Virtual Contrastive Constraint for Few-Shot Class-Incremental LearningZeyin Song, Yifan Zhao, Yujun Shi, Peixi Peng et al.CVPR 2023
Dusk over the sea between two cliffs drawn in fine vertical lines