Reviews: Improving Variational Autoencoders with Inverse Autoregressive Flow

The idea is interesting, particularly in its approach to enable parallel computation on GPUs by whitening the data. This seems like a practical approach to choosing the family of transformations in the normalizing flows framework. Its novelty is low based on whitening, and being a simple extension to just one of many approaches for building expressive variational families. The experiments are lacking in comparison to the slew of recent approaches to expressive approximating families for variational inference. It mostly compares to non-variational inference based approaches and traditional approaches (e.g., diagonal Gaussians). For example, how does it compare to planar or radial flows as in the original paper?