Cross-Domain Few-Shot Classification via Adversarial Task Augmentation

• DOI: 10.24963/ijcai.2021/149
• Publication Date: 2021-08-11T07:00:49Z
• Authors: Haoqing Wang; Zhi-Hong Deng
• Keywords: Inductive bias; Robustness

Few-shot classification aims to recognize unseen classes with few labeled samples from each class. Many meta-learning models for few-shot elaborately design various task-shared inductive bias (meta-knowledge) solve such tasks, and achieve impressive performance. However, when there exists the domain shift between training tasks test obtained fails generalize across domains, which degrades performance of models. In this work, we aim improve robustness through task augmentation. Concretely, consider worst-case problem around source distribution, propose adversarial augmentation method can generate bias-adaptive 'challenging' tasks. Our be used as a simple plug-and-play module models, their cross-domain generalization capability. We conduct extensive experiments under setting, using nine datasets: mini-ImageNet, CUB, Cars, Places, Plantae, CropDiseases, EuroSAT, ISIC ChestX. Experimental results show that our effectively shift, outperforms existing works. code is available at https://github.com/Haoqing-Wang/CDFSL-ATA.