PIMS-CORDS SFU Operations Research Seminar: Nicholas Richardson

Throughout scientific and commercial domains, we are often interested in separating mixed signals into their component sources. Supervised deep learning is state-of-the-art when large and well-labeled datasets can be used. But in many applications, large scale collection and labelling can be too impraticable, expensive, or behind copyright laws. This talk will explore a number of applications from sediment analysis, genome sequencing, and audio source separation that fall into the scarce data category. We will see a few approaches I have used to model and solve these problems such as sparse feature models and tensor factorizations. These unsupervised learning techniques avoid a training phase and have the advantage of adapting to the specific example at hand.