How does the proposed method (COPR) extend the Module network?
Describe the model: what are objective function, hidden/observed variables, parameters.
Any room for improvement, in terms of modeling or learning?