Adding noise during training as a method to increase resilience of neural network solution of inverse problems: Test on the data of magnetotelluric sounding problem

Abstract

In their previous studies, the authors proposed to use the approach associated with adding noise to the training set when training multilayer perceptron type neural networks to solve inverse problems. For a model inverse problem it was shown that this allows increasing the resilience of neural network solution to noise in the input data with different distributions and various intensity of noise. In the present study, the observed effect was confirmed on the data of the problem of magnetotelluric sounding. Also, maximum noise resilience (maximum quality of the solution) is generally achieved when the level of the noise in the training data set coincides with the level of noise during network application (in the test dataset). Thus, increasing noise resilience of a network when noise is added during its training is associated with the fundamental properties of multilayer perceptron neural networks and not with the properties of the data. So this method can be used solving other multi-parameter inverse problems.