Mutual Information Collapse Explains Disentanglement Failure in ${\hat{I}}^2$-VAEs

Minh Vu, Xiaoliang Wan, Shuangqing Wei

The ${\hat{I}}^2$-VAE is a foundational framework for unsupervised disentanglement, using ${\hat{I}}^2$ to regulate the trade-off between latent factorization and reconstruction fidelity. Empiricall…