Evolutionary Computation Paradigm to Determine Deep Neural Networks Architectures

Image classification is usually done using deep learning algorithms. Deep learning architectures are set deterministically. The aim of this paper is to propose an evolutionary computation paradigm that optimises a deep learning neural network's architecture. A set of chromosomes are randomly generated, after which selection, recombination, and mutation are applied. At each generation the fittest chromosomes are kept. The best chromosome from the last generation determines the deep learning architecture. We have tested our method on a second trimester fetal morphology database. The proposed model is statistically compared with DenseNet201 and ResNet50, proving its competitiveness.