We need to formalize the mutual information objective, the computational graph, as written initially, will be subject to change and experimentation. We will also need accurate approximations to our MI since the calculation will not generally be tractable. Please coordinate with @blattnem.