Includes bibliographical references and index.

Machine learning encompasses a broad range of problems ranging from detecting objects in images, finding documents relevant to a given query, or predicting the next element in a sequence, among countless others. Traditional approaches to these problems operate by collecting large, labeled datasets for training, uncovering informative features, and mining complex patterns that explain the association between features and labels. Typically, labels are regarded as an underlying 'truth' that should be predicted as accurately as possible.