RBF neural networks design with graph based structural information from dominating sets
Carregando...
Data
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Federal de Minas Gerais
Descrição
Tipo
Artigo de periódico
Título alternativo
Primeiro orientador
Membros da banca
Resumo
The definition of an appropriate number of Radial Basis Functions and their parameters in Radial Basis Function networks is a non-trivial task. The fitting of its parameters has direct implications on model performance and generalization. Techniques such as cross-validation associated with error metrics, which frequently rely on iterative optimization, have been used to address this problem, requiring significant computational effort and uncertain solutions. We propose a method for determining the number and parameters of Radial Basis Functions based on the Dominating Set of the Gabriel graph, which represents the structure of the input data, such that no exterior/prior parameter estimation or heuristic methods are necessary.
Abstract
Assunto
Redes neurais (Computação)
Palavras-chave
Many methods have been developed for locating centers of Radial Basis Function (RBF) networks, usually aiming at identifying anchor points within the input data space that either result in a reliable estimation of class likelihoods or spatially separate class representations. Some existing techniques attempt to choose samples that are particularly representative of the input data by partition methods, e.g. clustering and others are based on input data structure analysis for the identification of marginal representative samples. A well-known example of a margin-based method are Support Vector Machines (SVM) with radial kernels, in which although all samples are considered to compute the kernel only the support vectors have an observable influence on the final classification. In both approaches, the intention is to identify training samples that are representative of data generator functions or class boundaries. The resulting model is a mixture of radial functions, usually Gaussian, which compute class likelihoods based on the entire data set or samples located on the separation margin.
Citação
Departamento
Curso
Endereço externo
https://link.springer.com/article/10.1007/s11063-022-11062-7