Convex Approximation of Probabilistic Reachable Sets from Small Samples Using Self-supervised Neural Networks