Fast neural networks learning techniques for signal compression