Disentangled Latent Spaces for Synthetic Data

Sammanfattning

Generative AI is changing society and people's lives through its many useful use-cases. For instance, AI systems are now capable of generating images which can be used to train and validate Machine Learning models in data-scarce domains such as medicine. Current models are highly photo-realistic and can synthesize images with strong variation, but they lack control in the specific features that are present in the image. For example, it is desirable to generate images of drivers that exhibit drowsiness, as this is seldom present when naturally collecting data, but this is difficult to achieve in state-of-the-art AI systems.In this thesis, we address this challenge by developing methods that can synthesize data with high control over the features. More precisely, we build upon the StyleGAN architecture by training a proxy latent space that is disentangled with respect to features related to a person's identity, as this enables the editing of images that preserve the identity. Furthermore, our method is cost-effective since the training of the proxy latent space is based on self-supervised learning objectives.We empirically evaluate the training objectives by formulating metrics to assess the disentanglement and find that there is great potential in achieving an identity-disentangled latent space. Visual evaluations, such as image editing using principal components, qualitatively show the degree of disentanglement, and also highlight the many potential use cases of such a latent space. Lastly, we discuss the many downstream tasks that are possible with such a latent space and propose directions for future research, such as estimating data distributions and filling in underrepresented classes.