Controlling Dependence in Implicit Generative Models via Spread Mutual Information

Mutual information (MI) provides an objective for suppressing or encouraging statistical dependence in implicit generative models. However, direct MI evaluation is challenging in implicit models due to typically intractable densities.