Otimização do vision transformer no CIFAR-100: avaliação em ambientes computacionais distintos e estratégias de aprendizado

Abstract

The main objective of this work is to study and analyze, through experiments, the functioning of the Vision Transformer (ViT) model, a deep learning architecture originally proposed for computer vision tasks and, in this project, applied to the processing and classification of images from the CIFAR-100 dataset. In this context, the proposal seeks to reproduce and expand upon the experiments of a baseline study in order to evaluate strategies that sustain the best possible performance even under limited-resource scenarios. Thus, during the development of this work, different execution configurations and optimization techniques were explored, including dynamic learning rate policies, hyperparameter adjustments, and regularization methods, with the goal of improving accuracy, accelerating convergence, and enhancing the model’s generalization capacity. In doing so, this study contributes by systematizing choices that enable training the ViT effectively without the need for pre-training on massive datasets, bringing academic practice closer to the real conditions of many research contexts.