Humboldt-Universität zu Berlin - Mathematisch-Naturwissenschaftliche Fakultät - Institut für Informatik

Der Raum wird noch bekannt gegeben, eine online Teilnahme ist möglich.

Deep learning models are widely used for image classification. However, especially in the field of medical image classification, biases have a major impact on the quality of the results. Recent developments in generative models, after initial approaches in the field of generative adversarial networks, provide better and more controllable latent representations as used in variational autoencoders. These models make different demands on the data used, which is particularly evident in the difference between paired and unpaired datasets. Using easily interpretable images from CelebA and images used for skin cancer detection, this thesis investigates the effect of different generative models on latent representations with the aim of isolating and thus mitigating biases. A new method based on the recombination of features from different images is proposed to control bias in latent representations. This thesis highlights links between the most important generative models and the applications of neural networks in the field of image classification.