MetaAdapter: Leveraging Meta-Learning for Expandable Representation in Few-Shot Class Incremental Learning

Lingyu Wu, Wenhao Yang, Lijun Zhang

Keywords: few-shot class incremental learning, meta-learning, feature representation, residual adapter

Abstract: Few-shot class incremental learning (FSCIL) aims to enable models to learn new tasks from few labeled samples while retaining knowledge of previously ones. This scenario typically involves an offline base session with sufficient data for pre-training, followed by online incremental sessions where new classes are learned from limited samples. Existing methods either rely on a frozen feature extractor or meta-testing simulation to address overfitting issues in online sessions. However, they primarily learn feature representations using only the base session data, which significantly compromises the model's plasticity in feature representations. To enhance plasticity and reduce overfitting, we propose the MetaAdapter framework, which makes use of meta-learning for expandable representation. During the base session, we expand the network with pre-trained weights by inserting parallel adapters and employ meta-learning to encode generalizable knowledge into these modules. Then, the backbone is further trained on abundant data from the base classes to acquire fundamental classification ability. In each online session, the adapters are first initialized with parameters from meta-training, and subsequently tuned to adapt to the new classes. Leveraging meta-learning to produce initial adapters, MetaAdapter enables the feature extractor to effectively adapt to few-shot new classes, thus improving the generalization of the model. Experimental results on the mini-ImageNet, CUB200, and CIFAR100 datasets demonstrate that our proposed framework achieves the state-of-the-art performance.

Primary Area: transfer learning, meta learning, and lifelong learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 9612