Rao-Blackwellized Stochastic Gradients for Discrete Distributions

Authors: Runjing Liu, Jeffrey Regier, Nilesh Tripuraneni, Michael Jordan, Jon Mcauliffe

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In Section 5, we exhibit the benefits of our procedure on synthetic data, a semi-supervised classification problem, and a pixel attention task.
Researcher Affiliation Collaboration Runjing Liu 1 Jeffrey Regier 2 Nilesh Tripuraneni 2 Michael I. Jordan 1 2 Jon Mc Auliffe 1 3 1Department of Statistics, University of California, Berkeley 2Department of Electrical Engineering and Computer Sciences, University of California, Berkeley 3The Voleon Group.
Pseudocode No The paper describes procedures using mathematical formulations and textual explanations but does not include any clearly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code No The paper does not contain any explicit statements or links indicating that the source code for the described methodology is publicly available.
Open Datasets Yes We work with the MNIST dataset (Lecun et al., 1998).
Dataset Splits Yes We used 50 000 MNIST digits in the training set, 10 000 digits in the validation set, and 10 000 digits in the test set.
Hardware Specification Yes Training was run on a p3.2xlarge instance on Amazon Web Services.
Software Dependencies No The paper mentions various methods and models (e.g., REINFORCE, Gumbel-softmax, Adam) but does not provide specific software names along with their version numbers required for reproduction.
Experiment Setup Yes We initialized the optimization with K-means. Figure 3 shows that Rao-Blackwellization improves the convergence rate, with faster convergence when more categories are summed.