Interpretação de emoções com machine learning:
Tipo de produção
Trabalho de Conclusão de Curso
Data de publicação
2024-12-04
Periódico
Editor
Texto completo na Scopus
Citações na Scopus
Autores
Souza, Guilherme Reis Queiros de
Oliveira, Gustavo Miranda de
Silveira, Luiz Henrique
Paula, João Victor da Silva
Andrade, Vinícius Cristiano Nagatomo de
Orientadores
Pimentel, Fagner de Assis Moura
Resumo
Emoções são uma das principais características que nos tornam humanos. Algo abstrato, deduzido e catalogado por nós mesmos, que dita toda manifestação da percepção humana no mundo. Um dos grandes desafios atuais da ciência é conseguir, apropriadamente, interceptar e interpretar de forma coesa esses sinais neurais que constituem as emoções bem como explorar as diferentes possibilidades que tais descobertas trazem consigo. Tendo isso em vista, foi realizado neste trabalho a interpretação de sinais neurológicos pré-classificados em um sistema de aprendizado de máquina e a comparação entre as diferentes técnicas utilizadas como SVM, MLP e classificador Random Forest, assim como o levantamento de uma discussão sobre os diferentes impactos e aplicações deste sistema.Os métodos utilizados neste trabalho são formados pela interpretação de sinais neurais através de inteligência artificial e detecção de padrões, desenvolvendo novas técnicas e melhorando as já existentes metodologias, estratégias e resultados. A proveniência dos sinais neurais se dá pelo uso de bases de dados públicas que contemplam detalhes de atividade cerebral de voluntários enquanto estímulos específicos eram exibidos, capturados através de EEG. Estes e outros métodos e materiais são mais detalhadamente descritos nas próximas sessões deste trabalho. Ao utilizar o algoritmo Random Forest Classifier em dados separados em conjuntos de duas emoções tratados utilizando a transformada de Fourier, foi possível obter um índice de acerto de 88,89% na melhor classificação, com outras duplas seguindo em 80% e 77%, conferindo uma melhora no desempenho com relação a outros trabalhos semelhantes. Algumas emoções também tiveram pior desempenho, chegando a bater apenas 22% de acerto, o que é abordado neste trabalho como um possível indicador da semelhança de certas emoções, observada em matrizes de confusão. No final, foi incluída uma demonstração de aplicação da saída deste algoritmo em um jogo simples, a fim de mostrar a capacidade de integração do programa e ilustrar possíveis impactos deste estudo em áreas diversas.
Emotions are one of the main characteristics that defines us as humans. Something abstract, deduced and catalogued by ourselves, which dictates every manifestation of human perception in the world. One of the great challenges of science today is to be able to appropriately intercept and interpret in a cohesive way these neural signals that constitute emotions, as well as to explore the different possibilities that such discoveries bring with them. With this in mind, this work carried out the interpretation of pre-classified neurological signals in a machine learning system and a comparison between the different techniques used such as SVM, MLP and Random Forest classifier, as well as the raising of a discussion on the different impacts and applications of this system. The methods used in this work are formed by the interpretation of neural signals through artificial intelligence and pattern detection, developing new techniques and improving existing methodologies, strategies and results. The source of the neural signals is given by the use of public databases that include details of brain activity of volunteers while specific stimuli were displayed, captured through EEG. These and other methods and materials are described in more detail in the next sections of this work. By using the Random Forest Classifier algorithm on data separated into sets of two emotions treated using the Fourier transform, it was possible to obtain an accuracy rate of 88.89% in the best classification, with other pairs following at 80% and 77%, showing an improvement in performance in relation to other similar works. Some emotions also had worse performance, reaching only 22% accuracy, which is addressed in this work as a possible indicator of the similarity of certain emotions observed in confusion matrices. At the end, a demonstration of the application of the algorithm’s output in a simple game was included, in order to show the program’s integration capacity and illustrate possible impacts of this study in different areas.
Emotions are one of the main characteristics that defines us as humans. Something abstract, deduced and catalogued by ourselves, which dictates every manifestation of human perception in the world. One of the great challenges of science today is to be able to appropriately intercept and interpret in a cohesive way these neural signals that constitute emotions, as well as to explore the different possibilities that such discoveries bring with them. With this in mind, this work carried out the interpretation of pre-classified neurological signals in a machine learning system and a comparison between the different techniques used such as SVM, MLP and Random Forest classifier, as well as the raising of a discussion on the different impacts and applications of this system. The methods used in this work are formed by the interpretation of neural signals through artificial intelligence and pattern detection, developing new techniques and improving existing methodologies, strategies and results. The source of the neural signals is given by the use of public databases that include details of brain activity of volunteers while specific stimuli were displayed, captured through EEG. These and other methods and materials are described in more detail in the next sections of this work. By using the Random Forest Classifier algorithm on data separated into sets of two emotions treated using the Fourier transform, it was possible to obtain an accuracy rate of 88.89% in the best classification, with other pairs following at 80% and 77%, showing an improvement in performance in relation to other similar works. Some emotions also had worse performance, reaching only 22% accuracy, which is addressed in this work as a possible indicator of the similarity of certain emotions observed in confusion matrices. At the end, a demonstration of the application of the algorithm’s output in a simple game was included, in order to show the program’s integration capacity and illustrate possible impacts of this study in different areas.
Citação
Palavras-chave
aprendizado de máquina; inteligência artificial; sinais neurais; machine learning; artificial intelligence; neural signals