Date of Award

9-24-2019

Publication Type

Master Thesis

Degree Name

M.Sc.

Department

Computer Science

Keywords

evolutionary deep neural networks, Evolutionary GAN, Evolved GANs, Generative Adversarial Networks, Neuroevolution, Neuroevolutionary training

Supervisor

Kobti, Z.

Rights

info:eu-repo/semantics/openAccess

Abstract

Recent developments in Deep Learning are noteworthy when it comes to learning the probability distribution of points through neural networks, and one of the crucial parts for such progress is because of Generative Adversarial Networks (GANs). In GANs, two neural networks, Generator and Discriminator, compete amongst each other to learn the probability distribution of points in visual pictures. A lot of research has been conducted to overcome the challenges of GANs which include training instability, mode collapse and vanishing gradient. However, there was no significant proof found on whether modern techniques consistently outperform vanilla GANs, and it turns out that different advanced techniques distinctively perform on different datasets. In this thesis, we propose two neuroevolutionary training techniques for deep convolutional GANs. We evolve the deep GANs architecture in low data regime. Using Fréchet Inception Distance (FID) score as the fitness function, we select the best deep convolutional topography generated by the evolutionary algorithm. The parameters of the best-selected individuals are maintained throughout the generations, and we continue to train the population until individuals demonstrate convergence. We compare our approach with the Vanilla GANs, Deep Convolutional GANs and COEGAN. Our experiments show that an evolutionary algorithm-based training technique gives a lower FID score than those of benchmark models. A lower FID score results in better image quality and diversity in the generated images.

Share

COinS