On the Structure and Learning of Perceptual Representations in Deep Neural Networks