Uma estrutura para execução de redes neurais evolutivas na GPU

• Ano de defesa: 2019
• Autor(a) principal: Mandoju, Jorge Rama Krsna
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Rio de Janeiro; Brasil; Instituto Alberto Luiz Coimbra de Pós-Graduação e Pesquisa de Engenharia; Programa de Pós-Graduação em Engenharia de Sistemas e Computação; UFRJ
• Palavras-chave em Português: Redes neurais evolutivas; GPU; CUDA; CNPQ::ENGENHARIAS
• Link de acesso: http://hdl.handle.net/11422/14045

[EN] In neuroevolution, neural networks are trained using evolutionary algorithms instead of the gradient descent method. One of the advantages over the gradient descent method is that it makes it possible not only to define the value of the weights of a neural network, but also its structure. In the optimization of evolutionary neural networks with the same weight, all neural networks of a population are evaluated to verify what is the value of the fitness function that each neural network will possess and with this value, to verify which are the neural networks that will move on to the next generation. The GPU (Graphic Processor Unit) is widely used in neural network training due to its high parallelism capability [1]. However, because their architecture is different from a common processor, some algorithms need to be executed differently to take advantage of the increased performance that the architecture can provide. In this work an architecture is created that is able to reduce the training time of evolutionary neural networks by joining the weights of all population by layer making each layer represent the weights of the entire population. In this way it is possible to vectorize the evaluation functions of neural networks. In training to classify the MNIST dataset, this structure has achieved a performance gain of up to 64% in neural networks MLP and a speedup of 20 in fitness calculation.