Machine Learning Methods for Computational Sustainability
Sustainable development has attracted much attention as ecosystems are facing more and more pressure from human activities. Machine learning can play an important role in ecology study as a large amount of data is being collected from ecosystems. In this talk, I will focus on three learning challenges in ecology study: data interpretation, model fitting, and data- intensive policy making. Specifically, my collaborators and I have investigated three typical learning problems: bird song classification, bird migration modeling, and reserve design.
First, we model the problem of bird song classification as a weakly-supervised learning problem; develop a probabilistic classification model; and theoretically analyze when we can learn good classifiers in this setting. Second, we model bird migration with a probabilistic graphical model at the population level; propose an approximation to significantly improve the scalability of the model; and theoretically show that the approximation is asymptotically correct. Third, we propose a novel formulation of transductive classification to study the reserve design problem; and adapt existing optimization algorithms to handle large-scale problems. Our learning techniques can also be used in other applications which involve learning from sensor data or decision-making with the learned models.
Bio: Liping Liu is a Ph.D. candidate advised by Thomas Dietterich at Oregon State University. His research mainly focuses on machine learning problems motivated by the study of ecosystems and sustainability. His research has developed a variety of learning techniques to address a wide spectrum of learning problems in the field of computational sustainability. Working together with teams at Cornell and UMass, he aims to answer important ecological questions and provide advice for the protection of ecosystems. He also has industry research experience at IBM T.J.Watson and Alibaba.