Convolutional Neural Networks (CNN) have been extensively employed in the last years as an alternative to the traditional methods of feature extraction. Their methodology is based on the hierarchical processing of human brain, and they have obtained good results for image classification tasks, since they are invariant to changes in scale and rotation. However, a big issue of such approaches concern with their parameter configuration, which can reach dozens and, depending on the neural network size, hundreds of parameters. Such parameter configuration is essential to the good performance of a CNN, being a manual choice of them a problem with high combinatorial complexity. This project aims at modeling the problem of CNN parameter optimization as being an evolutionary optimization task, since just a few works have been conducted in this context. The present work has as main goal to study and also to apply a recent proposed evolutionary algorithm called Migrating Birds Optimization (MBO) for CNN parameter optimization. Additionally, the proposed approach will be evaluated in the context of human facial expression recognition.
News published in Agência FAPESP Newsletter about the scholarship: