Estimating the electrophysiological activity at the origin of electroencephalography (EEG) measurements is an ill-posed inverse problem. Several methods solve this problem by imposing different priors on the solution. Machine learning could allow to learn the inverse function directly from the data and thusmake the choice of one of the multiple solutions of the inverse problem more reliable. This work is based on simulations of electrophysiologic data containing single or multiple extended sources, using the SEREEGA simulation toolbox [1]. These data are used to train a one-dimensional convolutional network (1D-CNN) and to compare the results of this learning approach to those obtained by a recurrent long short term memory (LSTM) network from the literature, and by minimum norm energy [2] (MNE) and standardized low resolution brain electromagnetic tomography [3] (sLORETA) methods. These results on simulated data are encouraging about the potential contribution of learning-based methods to the problem of spatio-temporal EEG sourceimaging. Additional work still needs to be done in order to also evaluate the ability of the network to generalize to real data.