Regularized deep learning with a non-convex penalty