Maximum-EntropyAdversarialDataAugmentation forImprovedGeneralizationandRobustness: SupplementaryMaterial

LetZ bearandomvariable(continuousordiscrete), and Y be a random variable with a finite set of outcomesY. Following the guidance of [3], we first assume a family of probability densities over the latent variablesθparameterized byψ,i.e.,q(θ|ψ). We use SGD for both minimization and maximization.