Learning Predictable and Discriminative Attributes for Visual Recognition