Abordagens evolutivas para otimização de redes neurais convolucionais baseadas em algoritmos genéticos

• Ano de defesa: 2021
• Autor(a) principal: Mendes, Raphael de Lima
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Uberlândia; Brasil; Programa de Pós-graduação em Ciência da Computação
• Palavras-chave em Português: Aprendizagem de Máquina; Redes Neurais Convolucionais; Algoritmos Genéticos; Otimização; Neuroevolução; Machine Learning; Convolutional Neural Networks; Genetic Algorithms; Neuroevolution; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO; Ciência da Computação; Inteligência Artificial; Redes neurais (Computação)
• Link de acesso: https://repositorio.ufu.br/handle/123456789/34044; http://doi.org/10.14393/ufu.di.2021.694

Convolutional Neural Networks (CNN) are considered the state-of-the-art in computer vision applications. However, building machine learning models such method still requires the parameter and hyperparameter tuning, and considerably large training datasets. Genetic Algorithms (GA) can be a promising technique to optimize various aspects of CNNs. In this work, two approaches to optimize CNN using GA are proposed: gaCNN, for architectural optimization and MLTLGA, for transfer learning optimization. In the gaCNN, a new individual codification strategy for activation functions is proposed alongside new mutation operators. The proposed method outperformed 9 out 13 methods evaluated in classification accuracy. In the MLTLGA, a new initialization operation is presented that outperformed by at least 2% the other transfer learning methods evaluated. Therefore, the two methods are promising in the study of optimization of CNN.