MULTI-VIEW ROBUST REPRESENTATION LEARNING

With the widespread deployment of sensors and the Internet-of-Things, multi-view data have become more common and publicly available. For example, a self-driving car uses radar, lidar, and camera sensors to collect real-time 3D information to drive safely on the road; disease diagnosis models utilize multiple modalities of neuroimage data, clinical scores, and genetics measurements for disease prediction; object detection techniques prefer object images from different views for high-fidelity recognition. The presence of multiple information sources provides an opportunity of learning better representations to improve performance by analyzing multiple views simultaneously and also poses great challenges for the existing data representation algorithms. First, different views tend to be treated as different domains from different distributions due to the view discrepancy. Second, they often require large-scale labeled data to sufficiently learn such representations, which significantly hinders their adaption into unsupervised learning tasks, and limits their applications into critical domains where obtaining massive labeled data is prohibitively expensive. To enable learning on those domains, this dissertation focuses on robust representation learning-based algorithms to alleviate the view discrepancy of the multi-view data in an unsupervised manner. Specially, we explore two scenarios upon data association for robust representation learning of multi-view data: First, the samples across different views have a sample-wise association in multi-view data, falling in the multi-view clustering scenario; Second, the samples across different views have a class-wise association, falling in the unsupervised domain adaption scenario, where the discriminant knowledge (representations) of views with labeled data samples are transferred to the views with unlabeled data samples