Daniel Hsu
Wednesday 27th July 2016
Time: 4.00pm
Ground Floor Seminar Room
25 Howland Street, London, W1T 4JG
Interactive machine learning via reductions to supervised learning
I'll describe new approaches for two interactive machine learning
problems that operate via reductions to supervised learning. The first
problem is contextual bandit learning, where the learner repeatedly
selects an action based on the current context, observing and collecting
the reward only for the selected action. The second problem is agnostic
active learning, where data for a classification task are initially
unlabeled, and each individual label must be explicitly requested. In
each of these problems, the decision-making of the learner is guided by
policies constructed using repeated calls to an oracle for solving
cost-sensitive classification problems.
[Joint work with Alekh Agarwal,
Tzu-Kuo Huang, Satyen Kale, John Langford, Lihong Li, and Rob Schapire.]