Robustness and convergence of adaptive schemes in blind equalization and neural network training