25 Sep 2019 (modified: 24 Dec 2019)ICLR 2020 Conference Blind SubmissionReaders: Everyone
  • Original Pdf: pdf
  • Abstract: Latent-variable models represent observed data by mapping a prior distribution over some latent space to an observed space. Often, the prior distribution is specified by the user to be very simple, effectively shifting the burden of a learning algorithm to the estimation of a highly non-linear likelihood function. This poses a problem for the calculation of a popular distance function, the geodesic between data points in the latent space, as this is often solved iteratively via numerical methods. These are less effective if the problem at hand is not well captured by first or second-order approximations. In this work, we propose less complex likelihood functions by allowing complex distributions and explicitly penalising the curvature of the decoder. This results in geodesics which are approximated well by the Euclidean distance in latent space, decreasing the runtime by a factor of 1,000 with little loss in accuracy.
15 Replies