Semi-Supervised Models with Deep Architectures and Applications

In many real-life applications, ranging from data mining to machine perception, labeled data are very hard and expensive to attain as it requires expert human efforts. Therefore in many cases one often encounters a large number of unlabeled data whereas the labeled data are rare. Semi-Supervised Learning (SSL) is a framework in machine learning that aims at learning from both labeled and unlabeled data points and thus reducing the human (expert) labor. This type of machine learning paradigm has gained significant interest in recent years in areas where labeled information is limited, such as computer vision, text mining, information retrieval and medical image analysis among others. The research objectives that will be investigated in this proposal are summarized as follows: developing advanced cutting-edge models with deep architecture in the semi-supervised framework, bridging the gap between existing deep kernel and neural network based models and discovering new synergies, incorporating prior-knowledge and side information into the analysis from multiple sources of information, learning robust and maximally predictive features in noisy and imbalanced environments.