Adding input noise to increase the generalization of neural networks is a bad idea

摘要

In the literature and in commercial software most of the Neural Networks (NN) models are trained using a simple Output Error (OE) cost function. This OE approach may lead to severe bias errors on the predicted output when noisy input data is used. This paper proposes a solution to this problem if input noise cannot be avoided, using the Errors-In-Variables (EIV) approach that is currently used in system identification. In this paper interpolation is suggested as a better alternative when input noise can be avoided.