We develop a framework to represent natural language attributes in a learned embedding space. This allows us to leverage open-ended user feedback to improve recommendations.

We use recently proposed robustness curves to show that point-wise measures for adversarial robustness do not capture important global properties that are essential to reliably compare the robustness of different classifiers.

We investigate realistic conditions for when unlabeled data improves upon the minimax learning rate of a supervised learning problem and demonstrate examples where these conditions are met.

We formalize two interpretations of the all-relevant problem and propose a polynomial method to approximate one of them for the important hypothesis class of linear classifiers, which also enables a distinction between weakly relevant features.