Simple robust neural network

The classification problem and applying simple neural networks for solving it are considered. The robust modification of the error backpropagation algorithm that is used for training neural networks is proposed. The proclaim that allows building the proposed modification with the Huber loss-function is proved. In order to study the properties of the obtained neural network, a number of computational experiments has been carried out. The different values of outliers' fraction, noise level, and training and test samples size have been considered. The result analysis shows that the proposed modification can significantly increase classification accuracy and learning rate of a neural network when working with noisy data.