Abstract

Convolutional Neural Networks (CNNs) have provided several real-world solutions to computer vision problems. However, for any given task, their performance heavily depends on the choice of its architecture. In most cases, the structural hyperparameters are optimized manually by researches in a time-consuming trial and error approach. We propose that by using metaheuristics in collaboration with machine learning, we can achieve a good trade-off between computational complexity and learning performance. To this end, we evaluate the performance of a genetic algorithm to optimize the depth, the number of filters, and the kernel size of CNNs for two different tasks of image classification. We achieve a better compromise between computational cost and final accuracy than the state of the art proposals, proving the usefulness of the collaboration of genetic algorithms and convolutional neural networks for image classification tasks.