Wednesday 12th October 2016
Ground Floor Seminar Room
25 Howland Street, London, W1T 4JG
Mixture Proportion Estimation for Weakly Supervised Learning
Weakly supervised learning has come to refer to a spectrum of problems
in machine learning, lying somewhere between supervised and unsupervised
learning, where partial or noisy label information is available in some
sense. In this talk I will overview many examples of weakly supervised
learning problems, and argue that several of them can be solved in terms
of a fundamental problem called mixture proportion estimation (MPE). MPE
is the problem of inferring the maximum proportion of one distribution
that is present in another, given random samples from each distribution.
I will discuss several approaches to MPE, including one based on the
kernel mean embedding, which exhibits state-of-the-art performance and
comes with theoretical guarantees.
Clayton Scott is Associate Professor in the Department of Electrical
Engineering and Computer Science, with a courtesy appointment in the
Department of Statistics, at the University of Michigan. He received his
undergraduate degree from Harvard University in mathematics, and his
master's and doctoral degrees in electrical engineering from Rice
University. His research interests include the theory, methods, and
applications of statistical machine learning.