Receptive field estimation for Gaussian-based neural networks

A receptive field estimation technique is offered to improve the performance of Gaussian-based neural networks. In the standard approach a Gaussian function with the same receptive field (covariance matrix) is applied to every presentation of the data. In the proposed technique, every local Gaussian has a different covariance matrix. These matrices are found by the Gram-Schmidt orthogonalization process, as opposed to trial and error. The radial basis function (RBF) and probabilistic neural networks (PNNs) are used to show the effectiveness of this approach.<
>