Strategies and techniques for deep learning on small data

• Ano de defesa: 2020
• Autor(a) principal: Pereira, Rafael Silva
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Laboratório Nacional de Computação Científica; Coordenação de Pós-Graduação e Aperfeiçoamento (COPGA); Brasil; LNCC; Programa de Pós-Graduação em Modelagem Computacional
• Palavras-chave em Português: Small data; Few shot learning; Deep learning; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: https://tede.lncc.br/handle/tede/320

The design of models enables the interpretation of complex problems. In computer Science, such models lead to the conception of algorithms and their implementation in computer systems, contributing to the problem solution. However, some problems are too complex to be described using an algorithmic approach. The introduction of machine learning methods aims to create models based directly on the collected data representing the observed phenomenon. While this approach led to great advances in many different fields, data driven methods often require a substantial amount of data in order to generalize its understanding of the modelled problem. In this work, we investigate the problem of small data for deep learning methods. We present strategies to minimize uncertainty on prediction by minimizing intra-class variation in classification tasks, constraining the solution space based on prior knowledge on the domain. Additionally, we discuss the few shot and zero shot scenarios, where we aim at training robust classifiers trough a fixed kernel function in order to create a model that generalizes for classes it was not trained upon. We present experiments for each of these and evaluate their properties on distinct datasets.