Variational Autoencoders (VAEs) have become a cornerstone of modern representation learning, with the β-VAE variant introducing a tunable weight β to balance the trade-off between reconstruction fidelity and latent space disentanglement. However, the choice of β has a strong impact on the quality of the learned embeddings, and in practice it is often determined by manual tuning or trial-and-error, which is inefficient and may yield suboptimal results. This problem is particularly pronounced in biomedical applications, where latent representations are expected to capture meaningful biological variability while maintaining sufficient reconstruction accuracy.

This thesis will explore one principled strategy for determining and adjusting the β weight, either by (a) deriving upper and lower bounds for reconstruction loss and KL-divergence to estimate meaningful β ranges, or by (b) implementing an adaptive adjustment scheme inspired by ControlVAE (Shao et al., 2020), which dynamically tunes β during training. The chosen approach will serve as the main focus of the thesis, aiming to provide a systematic alternative to manual hyperparameter selection.

The implementation will be carried out within the group’s autoencoder framework AUTOENCODIX (https://github.com/jan-forest/autoencodix), which provides a modular infrastructure for training and evaluating autoencoder architectures. The developed method will be benchmarked on biomedical data modalities and integration tasks, assessing both technical performance (e.g. reconstruction error, latent disentanglement, and convergence behavior) and biological relevance of the learned embeddings.

This work will contribute to making β-VAEs more robust and user-friendly for biomedical applications, reducing reliance on manual tuning and enabling more principled exploration of latent representations in complex datasets.

Student profile: Master Mathematics/Physics/Data Science/Bioinformatik/Informatik. Thesis requires either prior knowledge of probability measure theory for probabilistic representation learning (VAEs) OR prior experience in PyTorch implementation of neural networks, ideally autoencoders.