Mutual Information Collapse Explains Disentanglement Failure in $\beta$-VAEs

A new paper identifies a critical failure in β-VAEs, a foundational AI model for learning independent data features. The research proves that strong regularization causes 'mutual information collapse,' destroying the model's ability to learn meaningful representations. To fix this, the authors propose the λβ-VAE, which decouples regularization pressure. Experiments on dSprites and Shapes3D show the new model stabilizes performance across a much wider range of settings.