Despite the huge success of deep neural networks (NNs), finding good mechanisms for quantifying their prediction uncertainty is still an open problem. It was recently shown, that using an ensemble of NNs trained with a proper scoring rule leads to results competitive to those of Bayesian NNs... This ensemble method can be understood as finite mixture model with uniform mixing weights. We build on this mixture model approach and increase its flexibility by replacing the fixed mixing weights by an adaptive, input-dependent distribution (specifying the probability of each component) represented by an NN, and by considering uncountably many mixture components. The resulting model can be seen as the continuous counterpart to mixture density networks and is therefore referred to as compound density network. We empirically show that the proposed model results in better uncertainty estimates and is more robust to adversarial examples than previous approaches. (read more)