Efficient Learning of Energy-Based Latent Variable Models


Ben Rhodes


I propose to work on unsupervised density estimation for energy-based and/or latent variable models. Specifically, my goal is to develop methods that overcome key technical challenges faced when estimating the parameters of such models. At least initially, I intend to build on noise-contrastive estimation (NCE), which is a method for learning unnormalised models, and variational inference, which is a method for learning latent variable models.


Supervisors: Michael Gutmann and Chris Williams