Export Reviews, Discussions, Author Feedback and Meta-Reviews

Summary: This paper describes an approach to statistical model criticism using the kernel two-sample test maximum mean discrepancy. The idea behind model criticism is simply to assess the ability of a given model to explain the observed data, and more importantly, to determine in which regions of the space the data is most misinterpreted by the model. For this purpose, the witness function of the MMD test is employed. This function takes large absolute values where the predictive distribution of the model considered is most different from the distribution of the actual observed data. The benefits of the approach described are shown in experiments involving restricted Boltzmann machines, deep networks and Gaussian processes.