NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:4018
Title:GRU-ODE-Bayes: Continuous Modeling of Sporadically-Observed Time Series


		
This paper proposes a method for modeling irregularly sampled multivariate time-series using a novel continuous-time version of a gated recurrent unit (GRU-ODE) to parameterize the gradient of the latent state combined with a probabilistic model for accommodating discontinuties in the latent state due to encountering new observations (GRU-Bayes). The problem the authors are solving is important and the method proposed in novel. The author response addressed a number of reviewer questions, but some important items remained. The most import is that several of the reviewers would like to see a number of ablations to the model to provide better evidence for where the improved performance is deriving from. Several reviewers mentioned wanting to see results when the GRU-Bayes part of the model is removed and only the GRU-ODE part of the model is used. This will help to quantify the importance of including the GRU-Bayes portion of the model. Several reviewers also asked for a broader review of the related literature to better situate the work. Nevertheless, the consensus opinion of the reviewers is that the paper is over the bar for acceptance. The authors should carefully take the items mentioned in this meta review into account when updating the manuscript, as well as the suggestions provided in the individual reviews.