An investigation on deep active learning

• Ano de defesa: 2024
• Autor(a) principal: Jacomassi, Renan Cesar
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Active learning; Aprendizado ativo; Aprendizado ativo profundo; Classificação de imagens de plâncton; Deep active learning; Plankton image classification
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/45/45134/tde-17072024-223436/

Active Learning is an approach in the field of machine learning where learning algorithms have the autonomy to choose which instance can provide the most relevant information. In this way, even with a reduced amount of data, it is possible to achieve a classifier with similar performance to what would be obtained using all available data. This technique is extremely important in cases where the cost of annotation is prohibitive, especially in situations that require a specialist. Additionally, in recent years, with the advancement of convolutional neural networks, studies on the application of Deep Active Learning are emerging. In the applications, the neural network can be incrementally trained with the most informative examples. Given this context, the main objective of this work is to investigate the application of Deep Active Learning to understand its viability and challenges. For this, two plankton image datasets resembling real-world situations were used. We empirically demonstrated that Deep Active Learning can reduce the annotation effort by up to 50%, and we showed how two parameters can significantly increase the performance and reduce the training time of the model.