CVPR Findings2026
AdaMeta: Adaptive Meta-Learning with Dynamic Task Relational Inference for Few-shot learning
Xingyu Yang, Yidan Ma, Hanzhang Qu, Jianfu Cao
Abstract
Few-shot learning (FSL) aims to enable models to generalize from only a handful of examples. A central challenge is adapting across diverse task streams whose relational structures are implicit and may shift over time. Most FSL methods treat each task as an isolated episode and lack mechanisms for modeling such dynamics, limiting knowledge accumulation and cross-domain generalization. We propose AdaMeta, an Adaptive Meta-Learning framework with Dynamic Task Relational Inference. AdaMeta builds a task-level relational graph that captures evolving inter-task dependencies and pairs it with a meta-knowledge distillation module that separates persistent meta-knowledge from transient task-specific signals, allowing transferable knowledge to be accumulated and reused across tasks. A regularized MAML-style online meta-optimizer then uses the inferred graph as a structural regularizer, balancing flexible adaptation to new task structures with training stability. We provide theoretical convergence guarantees under non-i.i.d. task conditions. Experiments on few-shot classification and cross-domain benchmarks show that AdaMeta consistently outperforms existing methods, achieving stateof-the-art (SOTA) performance.
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.
Lineage
- Rethinking Generalization in Few-Shot Classification cites MELR: Meta-Learning via Modeling Episode-Level Relationships for Few-Shot Learning
- BOIL: Towards Representation Change for Few-shot Learning cites Rapid Learning or Feature Reuse? Towards Understanding the Effectiveness of MAML
Builds on24
- Rapid Learning or Feature Reuse? Towards Understanding the Effectiveness of MAMLAniruddh Raghu, Maithra Raghu, Samy Bengio, Oriol VinyalsICLR 2020 · 736 citations
- Few-Shot Image Recognition With Knowledge TransferZhimao Peng, Zechao Li, Junge Zhang, Yan Li et al.ICCV 2019 · 230 citations
- BOIL: Towards Representation Change for Few-shot LearningJaehoon Oh, Hyungjun Yoo, ChangHwan Kim, Se-Young YunICLR 2021 · 185 citations
- MELR: Meta-Learning via Modeling Episode-Level Relationships for Few-Shot LearningNanyi Fei, Zhiwu Lu, Tao Xiang, Songfang HuangICLR 2021 · 121 citations
- Rethinking Generalization in Few-Shot ClassificationMarkus Hiller, Rongkai Ma, Mehrtash Harandi, Tom DrummondNeurIPS 2022 · 119 citations
Related papers
- Meta-Semantics Augmented Few-Shot Relational LearningHan Wu, Jie YinEMNLP 2025
- Automated Relational Meta-learningHuaxiu Yao, Xian Wu, Zhiqiang Tao, Yaliang Li et al.ICLR 2020 · 102 citations
- Bayesian Meta-Learning for the Few-Shot Setting via Deep KernelsMassimiliano Patacchiola, Jack Turner, Elliot J. Crowley, Michael F. P. O'Boyle et al.NeurIPS 2020 · 167 citations
- MoEMeta: Mixture-of-Experts Meta Learning for Few-Shot Relational LearningHan Wu, Jie YinNeurIPS 2025 · 1 citation
- MetaNorm: Learning to Normalize Few-Shot Batches Across DomainsYing-Jun Du, Xiantong Zhen, Ling Shao, Cees G. M. SnoekICLR 2021 · 26 citations
