Semi-supervised learning by multi-manifold separation
Thursday, October 30, 2008 - 10:55am - 11:45am
Semi-supervised learning on a single manifold has been the subject of intense study. We consider the setting of multiple manifolds, in which it is assumed that the target function is smooth within each manifold, yet the manifolds can intersect and partly overlap. We discuss our recent work to separate these manifolds from unlabeled data, and perform a 'mild' form of semi-supervised learning which is hopefully robust to the model assumption.