Information Transfer in Multitask Learning, Data Augmentation, and Beyond

A hallmark of human intelligence is that we continue to learn new information and then extrapolate the learned onto tasks domains (see, e.g., Thrun Pratt (1998)). While this a fairly intuitive observation, formulating such ideas has proved be challenging research problem continues inspire studies. Recently, there been increasing interest in AI/ML about building models generalize across tasks, even when they have some form distribution shifts. How can ground solid framework develop principled methods for better practice? This talk will present my recent works addressing question. My involve three parts: revisiting multitask learning from lens deep theory, designing robust transfer, algorithmic implications data augmentation.