Reviews: BRUNO: A Deep Recurrent Model for Exchangeable Data

This paper introduces an unsupervised approach to modeling exchangeable data. The proposed method learns an invertible mapping from a latent representation, distributed as correlated-but-exchangeable multivariate-t RVs, to an implicit data distribution that can be efficiently evaluated via recurrent neural networks. I found the paper interesting and well-written. Justification and evaluation of the method could, however, be much better. In particular, the authors do not provide good motivation for their choice of a multivariate t-distribution beyond the standard properties that 1) the posterior variance is data-dependent 2) it is heavier tailed compared to normal.