Speedup of Training Deep Neural Networks in the Streaming Approach Using Genetic Algorithms with an Application of Drift Detection

Piotr Duda, Mateusz Wojtulewicz, Robert Nowicki, Leszek Rutkowski

Abstract: The paper presents a novel approach to training neural networks employing genetic algorithms. This novel method utilizes mini-batches as the fundamental components of the population, fed to specialized crossover and mutation operators. An innovative algorithm tracks network performance fluctuations to prevent overfitting and ensure comprehensive data utilization. Additionally, the paper investigates the impact of various parameters, including population size, mini-batch size, and mutation rate, on the algorithm’s performance. Finally, the proposed method is compared against the traditional epoch-based training method, demonstrating its superior performance.