Propagação de alvos em redes neurais artificiais

• Ano de defesa: 2019
• Autor(a) principal: Farias, Tiago de Souza
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Santa Maria; Brasil; Física; UFSM; Programa de Pós-Graduação em Física; Centro de Ciências Naturais e Exatas
• Palavras-chave em Português: Rede neural; Treinamento; Propagação de alvos; Inicialização; Hiperparâmetros; Emaranhamento; Neural network; Training; Target propagation; Initialization; Hyperparameters; Entanglement; CNPQ::CIENCIAS EXATAS E DA TERRA::FISICA
• Link de acesso: http://repositorio.ufsm.br/handle/1/16956

Artificial neural network is a type of artificial intelligence focued on the study of learning systems with data. The training of neural networks requires a method that evolves their internal parameters, which are responsible for solving some given problem. Target propagation is a training method where it evaluates the best neural activity to accomplish an objective. We developed a variation of target propagation in which the ideal activity is obtained from a gradient of an objective function. Random weights initialization is a technique to set the random values in neural networks before training. We present an initialization scheme that considers the non-linear effects from neurons and the distribution of data. Hyperparameters are values that regulate evolving parameters. These values are, in general, obtained heuristically, wasting computational resources. We show a method to obtain the hyperparameters without the need of searches algorithms. Quantum neural networks are a type of artificial intelligence where it harness quantum phenomena for computation power. Inspired by a theory of entanglement in the biological brain, we developed a quantum correlation technique among neurons that can improve performance. Within the problem of image classification, the results show that the four techniques can improve the neural network performance and, under certain conditions, lower the computational cost.