Abstract
Deep learning models have advanced the research and application of artificial intelligence noticeably. One critical reason behind the success of deep neural networks can be attributed to the large amount of well-annotated training data. However, it is impossible for the training dataset to include all the categories that machines would encounter. In fact, the real-world data follow the long-tailed distribution and there are always certain novel categories containing only few samples. Therefore, it is important to enable machines to learn from limited training data, similar to human’s remarkable learning ability. In this thesis, we first discuss two application tasks where the ability of learning with limited data plays an important role. In person re-identification, the trained model is expected to generalize well given one or few query images of novel persons, with knowledge learned from training dataset including disjoint identities. We propose to use multi-level visual similarities, guided by spatial attention, to determine whether the input two person images are matched or not. By learning how to compare two images effectively, we get rid of the restriction of limited training data. In brain imaging signal analysis, the collected data vary for each subject and the labeling is also challenging due to privacy. Therefore, we aim at developing an unsupervised system for different subjects. The graph convolutional networks are used to exploit the brain connection patterns characterized by each subject, which helps the system adapt to subjects with limited unlabeled data. We then move on to a more fundamental problem, how to directly learn a classifier from few labeled samples. We cast this problem in the paradigm of meta learning. The key insight here is that rich structural information exists in the few shot tasks. We combine this insight with metric based and weight generation based meta learning approaches for few shot image classification. In this way, we build systems more adaptive to various novel tasks and consequently the state-of-the-art performance is improved significantly. Finally, we conclude the thesis and discuss some future works, which could extend our research into more challenging domains.