A PAC-Bayesian Analysis of Dropouts

Intuitively, a neural network that is robust to dropout perturbations should have better generalization properties - it should perform better on novel inputs. Stochastic model perturbation is the fu