Abstract
Deep learning models often require much annotated data to obtain good performance. In real-world cases, collecting and annotating data is easy in some domains while hard in others. A practical way to tackle this problem is using label-rich datasets with large amounts of labeled data to help improve prediction performance on label-poor datasets with little annotated data. Cross-domain few-shot learning (CD-FSL) is one of such transfer learning settings. In this paper, we propose a feature transformation ensemble model with batch spectral regularization and label propagation for the CD-FSL challenge. Specifically, we proposes to construct an ensemble prediction model by performing multiple diverse feature transformations after a shared feature extraction network. On each branch prediction network of the model we use a batch spectral regularization term to suppress the singular values of the feature matrix during pre-training to improve the generalization ability of the model. The proposed model can then be fine tuned in the target domain to address few-shot classification. We also further apply label propagation and data augmentation to further mitigate the shortage of labeled data in target domains. Experiments are conducted on a number of CD-FSL benchmark tasks with four target domains and the results demonstrate the superiority of our proposed method.
Abstract (translated)
URL
https://arxiv.org/abs/2005.08463
