A Bayesian approach for extreme learning machine-based subspace learning

In this paper, we describe a supervised subspace learning method that combines Extreme Learning methods and Bayesian learning. We approach the standard Extreme Learning Machine algorithm from a probabilistic point of view. Subsequently and we devise a method for the calculation of the network target vectors for Extreme Learning Machine-based neural network training that is based on a Bayesian model exploiting both the labeling information available for the training data and geometric class information in the feature space determined by the network´s hidden layer outputs. We combine the derived subspace learning method with Nearest Neighbor-based classification and compare its performance with that of the standard ELM approach and other standard methods.