Exploring Visual Structures In Deep Representation Learning