Research Area:  Machine Learning

Abstract:

Recent advances in deep learning lead to breakthroughs in many machine learning tasks. Due to the data-driven nature of deep learning, the training procedure often requires large amounts of manually annotated data, which is often unavailable. One-shot learning aims to categorize the new classes unseen in the training set, given only one example of each new class. Can we transfer knowledge learned by one shot learning from one domain to another? In this paper, we formulate the problem of domain adaption in one-shot image classification, where the training data and test data come from similar but different distribution. We propose a domain adaption framework based on adversarial networks. This framework is generalized for situations where the source and target domain have different labels. We use a policy network, inspired by human learning behaviors, to effectively select samples from the source domain in the training process. This sampling strategy can further improve the domain adaption performance. We investigate our approach in one-shot image classification tasks on different settings and achieve better results than previous methods.

Keywords:  
Domain Adaption
Machine Learning
Deep Learning
One-Shot Learning

Author(s) Name:  Nanqing Dong and Eric P. Xing

Publisher name:  ECML PKDD

DOI:  10.1007/978-3-030-10925-7_35

Volume Information:  

Paper Link:   http://www.ecmlpkdd2018.org/wp-content/uploads/2018/09/81.pdf