We thank the reviewers for the time and efforts spent reviewing our paper. We are glad the reviewers found the paper very interesting and forming a substantial and novel contribution to the community.$ We thank Assigned_Reviewer_1 for suggesting further improvements to the clarity of our submission. We've been in touch with Osband et al. (''Deep Exploration via Bootstrapped DQN'') regarding their results, and put a demo online demonstrating the results of our technique on the datasets they have used (in appendix A). We will add a clarification in our appendix pointing the reader to these results. We thank Assigned_Reviewer_2 for the kind words. We will update the appendix with a survey of recent work in the field since our submission (we will put this in the appendix due to the page limit in the main paper). As a side note, Korattikara et al. (''Bayesian Dark Knowledge'', [http://papers.nips.cc/paper/5965-bayesian-dark-knowledge.pdf]) cite an arXiv version of this submission. We thank Assigned_Reviewer_4 for the kind words and constructive criticism. We will clarify the terminology used in the paper ("variation ratios" for example). We chose to provide an extensive set of experiments in the main body of the paper since many deep learning practitioners place a high weight on empirical evidence. Answering the reviewer's questions: * In our setting we assume the p_i are given in advance (as is standard practice in the field). These could be optimised as variational parameters as well. * The time complexity at test time is that of the normal model multiplied by T - the number of samples used in MC integration. This number depends on dataset characteristics and from our experience can vary from 20 to 100. Most hardware architectures support parallel processing though, allowing us to perform multiple MC evaluations concurrently. We will clarify this in the paper. * "large standard deviations because of a single outliers" - we assume this is because of the relatively small datasets used, where different splits of the data can lead to very different results.