Multiple Instance Learning for Computer Aided Diagnosis

by Glenn Fung, Murat Dundar, Balaji Krishnapuram, R. Bharat Rao

This paper talks about MIL poblems, addressing them by a convex hull based algorithm. The key idea comes from a former proposed relaxation (in another paper). In MIL, the basic assumption is that each positive bag contains at least one positive sample. The relaxation of this assumption is that in the convex hull of each positive bag, there exists a point that could be correctly classified.

With this relaxation, the author proposed a general form for MIL, with which many known algorithms can be formulated, e.g. SVM and Fisher discriminant criterion for MIL. On a whole this is an application paper.