erm

This paper resolves whether Empirical Risk Minimization is relatively smart for binary classification, explores semi-supervised learning with quadratic blowup in unlabeled data, and proves intractability results for certain learners.