Hybrid encodings for neuroevolution of convolutional neural networks
Gustavo-Adolfo Vargas-Hákim, Efrén Mezura‐Montes, Héctor‐Gabriel Acosta‐Mesa
Abstract
The Neuroevolution of Convolutional Neural Networks have yield into highly competitive results in the field of visual recognition in contemporary years. Some of the most recent advances in this field have been related to the design of neural encodings to represent these highly complex Deep Learning structures. Hybrid encodings have shown potential at distributing the representation of these networks into different sub-structures and thus improving the search. In this paper, we propose a compact hybrid encoding, which is used in an evolutionary framework called Deep Genetic Algorithm (DeepGA). We assess the performance of our simple representation against a hybrid encoding based on DenseBlocks, to evaluate how certain encodings might bias the search towards larger CNNs, in both single- and multi-objective scenarios. Our case study is the classification of lung conditions in chest X-ray images.