Aplicação de máquinas de aprendizado extremo ao problema de aprendizado ativo

Detalhes bibliográficos
Ano de defesa: 2015
Autor(a) principal: Euler Guimarães Horta
Orientador(a): Não Informado pela instituição
Banca de defesa: Não Informado pela instituição
Tipo de documento: Tese
Tipo de acesso: Acesso aberto
Idioma: por
Instituição de defesa: Universidade Federal de Minas Gerais
UFMG
Programa de Pós-Graduação: Não Informado pela instituição
Departamento: Não Informado pela instituição
País: Não Informado pela instituição
Palavras-chave em Português:
Link de acesso: http://hdl.handle.net/1843/BUBD-A4BK3A
Resumo: The main objective of Active Learning is to choose only the most informative patterns to be labeled and learned. In Active Learning scenario a selection strategy is used to analyze a non-labeled pattern and to decide whether its label should be queried to a specialist. Usually, this labeling process has a high cost, which motivates the study of strategies that minimize the number of necessary labels for learning. Traditional Active Learning approaches make some unrealistic considerations about the data, such as requiring linear separability or that the data distribution should be uniform. Furthermore, traditional approaches require fine-tuning parameters, which implies that some labels should be reserved for this purpose, increasing the costs. In this thesis we present two Active Learning strategies that make no considerations about the data distribution and that do not require fine-tuning parameters. The proposed algorithms are based on Extreme Learning Machines (ELM) with a Hebbian Perceptron with normalized weights in the output layer. Our strategies decide whether a pattern should be labeled using a simple convergence test. This test was obtained by adapting the Perceptron Convergence Theorem. The proposed methods allow online learning, they are practical and fast, and they are able to obtain a good solution in terms of neural complexity and generalization capability. The experimental results show that our models have similar performance to regularized ELMs and SVMs with ELM kernel. However, the proposed models learn a fewer number of labeled patterns without any computationally expensive optimization process and without fine-tuning parameters.