Abstract
The growing use of Deep Learning in various domains has amplified
the challenges of training models due to high computational costs
and the need for large volumes of data. To address these limitations,
this study presents the OptiCore method, a new dataset optimization
approach based on the Greedy Coreset technique. OptiCore
strategically reduces the size of datasets while preserving their
representativeness and diversity, integrating computational cost
analyses through the Relative Cost Normalized metric. This method
balances data efficiency and model performance, offering a scalable
solution for practical applications. The methodology is designed
for generalization and reproducibility, extending its usefulness to
different Deep Learnig contexts. In the case study, Deep Learning
models were applied for the classification of three-dimensional
shapes, with the ResNet-50 architecture showing the best results.
OptiCore reduced the dataset by up to 90%, maintaining competitive
accuracy while significantly reducing computational demands.