This is the project page for Bayes-Factor-VAE: Hierarchical Bayesian Deep Auto-Encoder Models forFactor Disentanglement. The work was accepted by ICCV 2019 Oral. [Paper Link][Youtube Link].

We propose a family of novel hierarchical Bayesian deepauto-encoder models capable of identifying disentangled fac-tors of variability in data. While many recent attempts at fac-tor disentanglement have focused on sophisticated learningobjectives within the VAE framework, their choice of a stan-dard normal as the latent factor prior is both suboptimal anddetrimental to performance. Our key observation is that thedisentangled latent variables responsible for major sourcesof variability, therelevant factors, can be more appropriatelymodeled using long-tail distributions. The typical Gaussianpriors are, on the other hand, better suited for modeling ofnuisance factors. Motivated by this, we extend the VAE to ahierarchical Bayesian model by introducing hyper-priors onthe variances of Gaussian latent priors, mimicking an infinitemixture, while maintaining tractable learning and inferenceof the traditional VAEs. This analysis signifies the impor-tance of partitioning and treating in a different manner thelatent dimensions corresponding to relevant factors and nui-sances. Our proposed models, dubbed Bayes-Factor-VAEs,are shown to outperform existing methods both quantita-tively and qualitatively in terms of latent disentanglementacross several challenging benchmark tasks.

(Comming Soon)