Aprendizado ativo profundo com comitê de modelos

Abstract

Manual annotation of large-scale datasets remains one of the main bottlenecks in computer vision tasks for training deep learning models. In this work, we propose and evaluate a committee of active learning strategies that combines uncertainty-based sample selection criteria (Least Confidence, Margin Sampling, and Entropy) with a diversity-based approach (k-means). For experimental validation, we employ three datasets with different levels of complexity, MNIST, FashionMNIST, and Parasitos, and adopt metrics such as accuracy, known-class coverage, and percentage of corrected samples. The results indicate that the committee achieves accuracies comparable to the best individual strategies while identifying all classes more rapidly, thus reducing the overall annotation effort. We conclude that committee-based deep active learning provides an effective balance between exploration and diversity, and further opens avenues for future investigations involving more challenging datasets, heterogeneous architectures, and domain adaptation scenarios.