Abstract- In this talk, we will discuss some tools for assessing numerically the information contained in the layers of a Feedforward Neural Network. Each layer output can be seen as a signal whose entropy can be evaluated. Necessary conditions for the entropy flow to be decreasing are provided. Mutual information is used to describe quantitatively the information conveyed by one layer about another layer. The invariance property and data processing inequalities regarding the mutual information are discussed and used in applications to Deep Neural Networks.