Learning Deep Representation with Energy-Based Self-Expressiveness for Subspace Clustering