Semantic Transformation-based Data Augmentation for Few-Shot Learning

Meihong Pan, Hongyi Xin, Hong-Bin Shen

Keywords: few-shot learning, data augmentation, semantic feature transformation, sample bias

TL;DR: We propose a semantic transformation based data augmentation approach by transferring samples from base dataset to the novel tasks in an encoder-decoder paradigm to alleviate the data scarce problem.

Abstract: Few-shot learning (FSL) as a data-scarce method, aims to recognize instances of unseen classes solely based on very few examples. However, the model can easily become overfitted due to the biased distribution formed with extremely limited training data. This paper presents a task specific data augmentation approach by transferring samples from base dataset to the novel tasks in an encoder-decoder paradigm, which guarantees generating semantically meaningful features. Specifically, the feature transfer process is carried out in semantic space. We further impose a compactness constraint to the generated features with the prototypes working as the reference points, which ensures the generated features distribute around the class centers. Moreover, we incorporate the cluster centers of the query set with the prototypes of the support set to reduce the bias of the class centers. With the supervision of the compactness loss, the model is encouraged to generate discriminative features with high inter-class dispersion and intra-class compactness. Extensive experiments show that our method outperforms the state-of-the-arts on four benchmarks, namely MiniImageNet, TieredImageNet, CUB and CIFAR-FS.

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.

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: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning