Text this: Batch gradient based smoothing L 2/3 regularization for training pi-sigma higher-order networks