Abstract
Despite the breakthroughs achieved by deep learning models in conventionalsupervised learning scenarios, their dependence on sufficient labeled trainingdata in each class prevents effective applications of these deep models insituations where labeled training instances for a subset of novel classes arevery sparse -- in the extreme case only one instance is available for eachclass. To tackle this natural and important challenge, one-shot learning, whichaims to exploit a set of well labeled base classes to build classifiers for thenew target classes that have only one observed instance per class, has recentlyreceived increasing attention from the research community. In this paper wepropose a novel end-to-end deep triplet ranking network to perform one-shotlearning. The proposed approach learns class universal image embeddings on thewell labeled base classes under a triplet ranking loss, such that the instancesfrom new classes can be categorized based on their similarity with the one-shotinstances in the learned embedding space. Moreover, our approach can naturallyincorporate the available one-shot instances from the new classes into theembedding learning process to improve the triplet ranking model. We conductexperiments on two popular datasets for one-shot learning. The results show theproposed approach achieves better performance than the state-of-the- artcomparison methods.