The recent extension of the Hubble diagram of Supernovae and quasars to redshifts much higher than 1 prompted a revived interest in non-parametric approaches to test cosmological models and to measure the expansion rate of the Universe. In particular, it is of great interest to infer model-independent constraints on the possible evolution of the dark energy component. Here we present a new method, based on a Neural Network Regression, to analyze the Hubble Diagram in a completely non-parametric, model-independent fashion. We first validate the method through simulated samples with the same redshift distribution as the real ones, and discuss the limitations related to the"inversion problem"for the distance-redshift relation. We then apply this new technique to the analysis of the Hubble diagram of Supernovae and quasars. We confirm that the data up to $z \sim 1-1.5$ are in agreement with a flat ${\Lambda}CDM$ model with ${\Omega}_M \sim 0.3$, while $\sim 5$-sigma deviations emerge at higher redshifts. A flat ${\Lambda}CDM$ model would still be compatible with the data with ${\Omega}_M>0.4$. Allowing for a generic evolution of the dark energy component, we find solutions suggesting an increasing value of ${\Omega}_M$ with the redshift, as predicted by interacting dark sector models.