Cut-Based Graph Learning Networks to Discover Compositional Structure of Sequential Video Data.