The LIIS (Laboratory of Intelligent Interactive Systems) and LPO (Operational Research Laboratory) from the Federal University of Pará (UFPA) made their mark at Computer on the Beach, an event held in Itajaí (SC), by presenting the work “OptiCore: Scalable Greedy Coreset Optimization Method for Efficient Deep Learning”.
Undergraduate student André Vinicius Neves Alves represented LIIS and LPO at the event, bringing to the academic and professional community an innovative proposal for dataset optimization in Deep Learning training. The research was developed in collaboration with the following authors:
André Vinicius Neves Alves – Federal University of Pará (UFPA)
Adriano Madureira dos Santos – Federal University of Pará (UFPA)
Lyanh Vinicios Lopes Pinto – Federal University of Pará (UFPA)
Flávio Rafael Trindade Moura – Federal University of Pará (UFPA)
Marcos César da Rocha Seruffo – Federal University of Pará (UFPA)
Saulo William da Silva Costa – Federal University of Pará (UFPA)
Vitor Barbosa – Federal University of Pará (UFPA)
Walter dos S. Oliveira Júnior – Federal University of Pará (INTECELERI/UFPA)
The presented project proposes the OptiCore method, which employs a Greedy Coreset-based approach to reduce up to 90% of the size of large datasets without compromising the quality of the trained model. This innovation enables a drastic reduction in computational cost and is particularly relevant for applications in resource-constrained environments.
In addition to the methodological proposal, the research introduced a new metric, the NRC (Normalized Relative Cost), which allows for a balanced evaluation of the computational efficiency of models trained with optimized datasets.