Rihuan Ke

From multi-tasking to nonlinear modelling: label-efficient learning algorithms for image analysis

Recent advances in deep learning techniques have contributed to powerful tools for solving image analysis tasks. The major success of these tools relies on supervised models that assume a large set of high-quality labels available for training. This assumption could be problematic in practice, since obtaining the required labels is often time-consuming and expensive. In this talk, we discuss new learning algorithms to overcome these restrictions through various strategies such as multi-tasking and nonlinear modelling. By taking advantage of the intrinsic structures in noisy/unlabelled data and incorporating mathematical and statistical knowledge relevant to the underlying tasks, we make the learning algorithms label-efficient.