Controle PID em Redes Neurais Convolucionais

Abstract

This paper investigates the integration of Proportional-IntegralDerivative (PID) controllers into Convolutional Neural Networks (CNNs) to enhance image classification performance. We explore different approaches, including adding a layer with a PID kernel, replacing the first convolutional layer with a PID kernel, and applying PID kernels to all convolutional layers. The results obtained indicate that introducing PID controllers into CNNs can significantly influence the network’s performance, depending on the configuration and parameters used. While some configurations showed promising results, such as an increase in learning rate, others led to a degradation in performance. The results suggest that combining CNNs and PID’s is a promising area for future research, with the potential to develop more robust and adaptable machine learning models.