Export Reviews, Discussions, Author Feedback and Meta-Reviews

Summary This paper proposes a maximum penalized likelihood version of the GP-LVM (equation after eq. 4). The penalty term added to the GP-LVM log likelihood is the log joint probability density function of the inputs under a Coulomb repulsive process. Clarity: good Originality: good, to the best of my knowledge Significance: medium (early days, still far from an easy to reimplement approach) Details The optimization depends critically on initialization, and the authors propose a heuristic that relies on using a similarity preserving traditional embedding, as well as a way to initialize the GPs hyperparameters. Obtaining posterior uncertainty is pretty tedious and there isn't a good solution. The authors propose a heuristic.