Improving Bayesian Inference in Deep Neural Networks with Variational Structured Dropout