CS Events

Qualifying Exam

Bayes-Factor-VAE: Hierarchical Bayesian Deep Auto-Encoder Models for Factor Disentanglement

 

Download as iCal file

Wednesday, August 12, 2020, 11:00am - 12:00pm

 

Speaker: 

Yuting Wang

Location : Remote via Webex

Committee

Prof. Vladimir Pavlovic

Prof. Yongfeng Zhang

Prof. Konstantinos Michmizos

Prof. Martin Farach-Colton

Event Type: Qualifying Exam

Abstract: The success of machine learning techniques, including few-shot learning, transfer learning, and image synthesis, relies on how data is represented. Flagship approaches, such as variational autoencoders (VAEs), model the nonlinear generative relationship between data and latent factors in an unsupervised manner. However, such representations entangle and hide more or less the factors of variation inherent in the data. To tackle this limitation, we argue that the desired representation is supposed to satisfy two properties. First, its latent space is able to explain the variation of data in a way that the factors are disentangled. That is, when changing one factor while freezing other factors, the variation is observed in only one aspect of the data. Second, relevant factors that are statistically dependent on the data should be discerned from nuisance factors. Inspired by these properties, we extend VAEs to a hierarchical Bayesian model that introduces hyper-priors on the variances of Gaussian latent priors, thus mimicking an infinite mixture, while maintaining tractable learning and inference of traditional VAEs. Our key insight is that the prior distribution of disentangled latent factors should be learned from data instead of being fixed to normal Gaussian distribution as in VAEs. By doing so, we are also able to better model nuisance factors. Extensive analysis shows the importance of partitioning the latent dimensions corresponding to relevant factors and nuisances, and treating them in different ways. Our proposed model outperforms the state of the art both quantitatively and qualitatively in terms of latent disentanglement on several challenging benchmarks, such as Sprites, 3D-Face, Teapots, and Celeb-A datasets.

 

Webex: https://rutgers.webex.com/rutgers/j.php?MTID=m32ca8d4de4d2bfb4483af67f79601d2e