Learning Latent Geometries
Prof Søren Hauberg
Latent variable models, such as the variational autoencoder, suffer from the identifiability problem: there is no unique configuration of the latent variables. This is problematic as latent variables are often inspected, e.g. through visualization, to gain insights into the data generating process. The lack of identifiability then raise the risk of
misinterpreting the data as conclusions may be drawn from arbitrary latent instantiations.
In this project you will investigate a geometric solution to the
identifiability problem that amounts to endowing the latent space with a particular Riemannian metric. You will learn latent representations and compute geodesics accordingly.
(*) Latent Space Oddity: on the Curvature of Deep Generative Models Georgios Arvanitidis, Lars Kai Hansen and Søren Hauberg. In International Conference on Learning Representations (ICLR), 2018.
(*) Only Bayes should learn a manifold (on the estimation of
differential geometric structure from data)
Project timezone: C