Exploring the Latent Space of styleGAN representation of identity, gender and emotion in latent space
Although only recently introduced in 2014, Generative Adversarial Networks (GANs) have already undeniably shown their potential. However, there is still a lot of room for improvement regarding the understanding of the latent space. This thesis aims to improve this understanding by exploring the representation of iden- tity, gender and emotion in the structure of the latent space of styleGAN trained on two di erent face-datasets (i.e., FFHQ and CelebA-HQ). Latent representations of input images from the Radboud Faces Database (RaFD) were found, which were analyzed using Representational Distance Matrices (RDMs) and visualized using t-SNE. The results for the latent spaces of the two styleGAN networks were ana- lyzed both individually and comparatively, in order to determine the e ect of the training dataset on the structure of the latent space. The results of this thesis indicate that both networks cluster for identity and gender in their latent space, yet not for emotion. However, for all identities, distinct directions were found for the di erent emotions when considering the average face of an individual as the origin point. Lastly, the two networks both represent identity, gender and emotion similarly in their latent space, thus the training dataset does not seem to have a substantial e ect on the structure of the latent space for the investigated features.
Faculteit der Sociale Wetenschappen