In this contribution we compare deep learning techniques with traditional methods on the analysis of lattice Monte Carlo data. Specifically we investigate the reconstruction of the spectral function from the Euclidean 2-point correlation function measured on the lattice. The two quantities are related by an integral transformation and the inversion of this transform is made difficult by the noise and the relatively low number of data points. When an analytical form of the dependence is assumed, it still poses a challenge for traditional methods, because the coefficients must be strictly positive thus the fitting is a constrained minimization problem. We compare the performance of chi square and maximum entropy methods as well as deep neural networks on the problem.