Discrete breathers are localized nonlinear modes that appear in discrete lattice systems. They can be generated numerically with very precise techniques that also allow the study of their stability and mobility. The use of machine learning techniques permits learning and subsequent recognition of discrete modes through convolutional neural networks. The latter can learn and then differentiate discrete breathers from linearized phonon modes. The breathers are localized in space and time periodic solutions of nonlinear discrete lattices, while phonons are the linear collective oscillations of interacting atoms and molecules. Deep learning neural networks are seen to be able not only to distinguish breather from phonon modes but also to determine with high accuracy underlying nonlinear on-site potentials that generate breathers. These results mean that it is indeed possible to use neural networks and solve inverse nonlinear problems.