The paper is well-written paper and analyzes how signals propagate in random neural networks. It does so by analyzing mean and variance of activations and gradients, given random inputs and weights. The technical contributions are okay, and the analysis leads to new insights on the use of batch/layer normalization.