Um estudo de redes profundas aplicadas a dados de fMRI

• Ano de defesa: 2016
• Autor(a) principal: Santos, Jeferson de Souza [UNIFESP]
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• dARK ID: ark:/48912/001300001r2gz
• Idioma: por
• Instituição de defesa: Universidade Federal de São Paulo (UNIFESP)
• Palavras-chave em Português: machine learning; aprendizado de máquina
• Link de acesso: https://sucupira.capes.gov.br/sucupira/public/consultas/coleta/trabalhoConclusao/viewTrabalhoConclusao.jsf?popup=true&id_trabalho=3697699; https://repositorio.unifesp.br/handle/11600/47216

Machine learning has been spreading to many applications, among them, supporting medical diagnosis. Artificial Neural Networks (ANNs) are in the list of used techniques for fulfilling these tasks. With the breaking of learning limit to the ANNs with bigger architectures in 2006, emerged a big area inside machine learning, Deep Learning. This architectures stood out in competitions and in real problem applications. One of its characteristics is flexibility, its architecture can be molded and that way generate new models with great problem abstraction capacity. The use of this architectures in medical diagnosis, more specifically in neuroscience with its functional Magnetic Resonance Imaging (fMRI), appeared in 2014. Nonetheless, nowadays there are few works related to this area, the potential union of deep learning with fMRI supporting medical decision making can bring great social benefits. As an example of application, deep learning can use fMRI data to detect chemical dependent. This thesis aims to use models of deep learning to solve the fMRI data classification problem originated from an addiction control group. Deep Learning architectures used in the study are known as Deep Belief Networks and Deep Neural Networks. In addition to the classification problem, this work contributes with a new approach of Deep Learning for a limited amount of data and a new way of building the results evaluation model based on the quantization of the data.