Reviews: Nonlinear random matrix theory for deep learning

This looks like a good theoretical contribution and an interesting direction in the theory of deep learning to me. In this paper, the authors compute the correlation properties (gram matrix) of the vector than went through some step of feedforward network with non linearities and random weights. Given the current interest in the theoretical description of neural nets, I think this is a paper that will be interesting for the NIPS audience and will be a welcome change between the hundreds of GAN posters. Some findings are particularly interesting.They applied their result to a memorization task and obtained an explicit characterizations of the training error. Also the fact that for some type of activation the eigenvalues of the data covariance matrix are constant suggests interesting directions for the future.