Reviews: Batch-Instance Normalization for Adaptively Style-Invariant Neural Networks

A nice analysis is performed on the relative weighting of IN and BN in different tasks. First it is shown that the interpolation weights follow a bi-modal distribution. Thus features either use IN to become invariant or BN to retain input variation across images. Moreover, while for object recognition the majority of features use BN, the fraction of IN usage increases substantially for style transfer as expected from the requirements of the task. Nevertheless, I have a couple of concerns/questions.