Novel maximum-margin training algorithms for supervised neural networks