Exploring Visual Structures in Deep Representation Learning