Título: Espaço de Pesos com Geometria Riemanniana para Aceleração da Convergência de Perceptron Multicamada
Autores: Medeiros, Luciano Frontino de; Silva, Hamilton Pereira da
Resumo: The convergence of a multilayer perceptron toward a global minimum error in the backpropagation algorithm can be accelerated by introducing the representation of weight search space in accordance with the Riemann geometry, which consider curved spaces. The trajectory of the search vector at the training phase is modified by a ”force” originated from the curvature of weight space. The main idea is inspired at the General Relativity theory, and the search space (now a manifold) must be controlled by parameters like spatial density and scale factor, modifying the gradient descent in the backpropagation algorithm.
Palavras-chave: Neural Networks; Riemann Geometry; Curved Spaces; Tensorial Calculus; General Relativity; Backpropagation
Código DOI: 10.21528/CBRN2005-103
Artigo em PDF: CBRN2005_103.pdf
Arquivo BibTex: CBRN2005_103.bib