Learning Regularization Parameters of Radial Basis Functions in Embedded Likelihoods Space

Neural networks with radial basis activation functions are typically trained in two different phases: the first consists in the construction of the hidden layer, while the second consists in finding the output layer weights. Constructing the hidden layer involves defining the number of units in it, as well as their centers and widths. The training process of the output layer can be done using least squares methods, usually setting a regularization term. This work proposes an approach for building the whole network using information extracted directly from the projected training data in the space formed by the likelihoods functions. One can, then, train RBF networks for pattern classification with minimal external intervention.