Dropout as a bayesian approximation: Representing model uncertainty in deep learning