PAC-BayesLearningBoundsfor Sample-DependentPriors

Sample-dependent priors are also relevant in emerging scenarios such as adversarial training. A common practice here is to smooth a given classifier by injecting Gaussian noise into the inputs.