Estudo de Seleção de Instâncias em Aprendizado de Métrica para o problema de classificação com Aprendizado Incremental

• Autor(a) principal: LUIZ HENRIQUE DA SILVA SANTOS
• Data de Publicação: 2023
• Outros Autores: MATHEUS VYCTOR ARANDA ESPÍNDOLA
• Tipo de documento: Trabalho de conclusão de curso
• Idioma: por
• Título da fonte: Repositório Institucional da UFMS
• Texto Completo: https://repositorio.ufms.br/handle/123456789/6170
• topic: Aprendizado de Métrica; Seleção de Instâncias; Aprendizado de Máquina; Aprendizado Incremental; Aprendizado Profundo.; Ciências Exatas e da Terra

Incremental Learning deals with a continuous increase in the number of classes over time. The main problem in this context is "catastrophic forgetting," where models lose the ability to solve tasks they originally learned when trained for new tasks, along with the increasing need for computational resources as new training occurs. To address this problem, the proposal evaluates the use of Deep Metric Learning models along with Instance Selection techniques on the CIFAR-100 dataset. In the experimental evaluation, the embeddings were assessed using K-Nearest Neighbor for performance analysis. The EfficientNetV2 Small model was chosen as the backbone. To generate high-dimensional numerical representations reflecting the similarity between classes, a Deep Metric Learning model, Zero-shot, was trained. Subsequently, fine-tuning of the Zero-shot model was performed to learn new classes added to the original dataset. To ensure that the model did not forget the old classes, Instance Selection techniques were employed, such as Random Mutation Hill Climbing (RMHC) and Gaussian Mixture (GM). The goal was to select representative instances of the old classes and reduce noise. The obtained result achieved an accuracy of 78.80% for the original dataset (Task 1) and an accuracy of 82.00% for the dataset with new classes (Task 2). Compared to the Baseline model, there was a reduction of 9.33% for Task 1, for which the Proposed model was initially trained, and a 2.12% increase in accuracy for Task 2. Moreover, the Proposed model required 81.00% fewer instances for training compared to the Baseline model, and the number of epochs was reduced by 87.50%. Thus, the study proposes a fast training approach with a limited training set size that avoids "catastrophic forgetting" of old classes.