Learning nonlinear differentiable models for signals and systems: with applications

• Ano de defesa: 2020
• Autor(a) principal: Antônio Horta Ribeiro
• Orientador(a): Luis Antonio Aguirre
• Banca de defesa: Eduardo Mazoni Andrade Marçal Mendes, Frederico Gadelha Guimarães, Guilherme de Alencar Barreto, Leandro dos Santos Coelho, Maarten Schoukens
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal de Minas Gerais
• Programa de Pós-Graduação: Programa de Pós-Graduação em Engenharia Elétrica
• Departamento: ENG - DEPARTAMENTO DE ENGENHARIA ELÉTRICA; ENG - DEPARTAMENTO DE ENGENHARIA ELETRÔNICA
• País: Brasil
• Palavras-chave em Português: Nonlinear systems; Differentiable models; Deep learning; System identification; Machine learning
• Link de acesso: http://hdl.handle.net/1843/33922

Building empirical models from data is of fundamental importance in engineering, and pushing the boundaries of current engineering technology requires us to model and understand nonlinear systems. In this thesis, nonlinear differentiable models and its applications are studied. This class of models has gained traction in machine learning tasks with the introduction of deep learning. Indeed, deep models of stacked differentiable components have recently achieved super-human performance on several tasks, including computer games, image classification, and medical diagnosis. The application of nonlinear differentiable models is studied for modeling signals and systems both for engineering and machine learning applications. One central question is the role of recurrence and the pros and cons of recurrent and feedforward models. The question is approached from more than one angle: 1) by studying the effect of recurrence in neural networks in terms of robustness to noise, computational cost, and convergence; 2) by analyzing the smoothness of the cost function in nonlinear system identification problems and its relation to the model internal dynamics – and proposing the use of a technique called multiple shooting for improving the cost-function smoothness; and, 3) by investigating the interplay between the internal dynamics, the attractors and the expressiveness of the model in deep recurrent neural networks. The more applied part of the thesis consists of the use of deep neural networks to solve complex tasks and to model nonlinear behavior from real data. Data from the Telehealth Center of Minas Gerais is used to train a deep neural network capable of identifying abnormalities in the electrocardiogram exam with performance superior to the medical residents in the studied scenario. Also, a deep neural network is used for modeling an electronic oscillator and an F-16 aircraft using data from ground vibration experiments, obtaining competitive results in both cases.