Convergence of a gradient algorithm with penalty for training two-layer neural networks | IEEE Conference Publication | IEEE Xplore