Dataset structure as prior information for parameter-free regularization of extreme learning machines

This paper proposes a novel regularization approach for Extreme Learning Machines. Regularization is performed using a priori spatial information expressed by an affinity matrix. We show that the use of this type of a priori information is similar to perform Tikhonov regularization. Furthermore, if a parameter free affinity matrix is used, like the cosine similarity matrix, regularization is performed without any need for parameter tuning. Experiments are performed using classification problems to validate the proposed approach.