Pro-t-VAE: Closed-Form Student's t Mixtures for Robust Prototype Discovery
The authors developed Pro-t-VAE, a robust deep generative framework for unsupervised image prototype discovery using a heavy-tailed Student's t mixture prior. They derived a closed-form training objective by treating precision as a latent variable, allowing the model to learn data-dependent trust levels and downweight outliers naturally. This method achieves superior robustness and accuracy compared to standard Gaussian or diffusion models.