Aprendizagem semissupervisionada por meio de técnicas de Deep Learning e de Teoria da Informação

• Ano de defesa: 2021
• Autor(a) principal: Lima, Bruno Vicente Alves de
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Rio Grande do Norte; Brasil; UFRN; PROGRAMA DE PÓS-GRADUAÇÃO EM ENGENHARIA ELÉTRICA E DE COMPUTAÇÃO
• Palavras-chave em Português: Semissupervisionado; Deep Learning; Rotulação; Agrupamento; Classificação; Teoria da Informação
• Link de acesso: https://repositorio.ufrn.br/handle/123456789/33968

The expressive growth of modern data sets, combined with the difficulty of obtaining information about labels, has made semi-supervised learning one of the problems of practical importance in modern data analysis. In most cases, obtaining a dataset with enough examples to induce a classifier can be costly, as it is necessary to perform labeling of the data by an expert. Unlabeled data is easier to obtain but more difficult to analyze and interpret. In the semi-supervised learning problem, there is a database formed by a small labeled part and a larger unlabelled part, with two possible aspects: semi-supervised classification and semi-supervised clustering. With this, this work aims to apply models that use deep learning techniques in semi-supervised learning. Using a deep autoencoder, the data was transformed to feature space Z, and, from that, these data were clustered and labeled, with the help of the labeled data. Information Theory Learning techniques were applied to increase the robustness of the model proposed in this work. Experiments performed showed the proposed model efficiency in labeling and classifying data after training. It was also compared to other classic semi-supervised learning models, such as co-training, tri-training, STRED and SEEDED K-means, as well as other more recent works, showing the proposed model feasibility for the semi-supervised learning problem. Finally, the model was applied to a real problem in remote sensing problem and stream data classification.