Decision-Making with Auto-Encoding Variational Bayes

Authors: Romain Lopez, Pierre Boyeau, Nir Yosef, Michael Jordan, Jeffrey Regier

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In addition to toy examples, we present a full-fledged case study of single-cell RNA sequencing. In this challenging instance of multiple hypothesis testing, our proposed approach surpasses the current state of the art.
Researcher Affiliation Academia Romain Lopez1, Pierre Boyeau1, Nir Yosef1,2,3, Michael I. Jordan1,4, and Jeffrey Regier5 1 Department of Electrical Engineering and Computer Sciences, University of California, Berkeley 2 Chan-Zuckerberg Biohub, San Francisco 3 Ragon Institute of MGH, MIT and Harvard 4 Department of Statistics, University of California, Berkeley 5 Department of Statistics, University of Michigan
Pseudocode No The paper does not contain any clearly labeled pseudocode blocks or algorithms.
Open Source Code Yes Our code is available at http://github.com/Pierre Boyeau/decision-making-vaes
Open Datasets Yes We consider the MNIST dataset, which includes features x for each of the images of handwritten digits and a label c. [...] We split the MNIST dataset evenly between training and test datasets. For the labels 0 to 8, we use a total of 1,500 labeled examples.
Dataset Splits No The paper mentions splitting the MNIST dataset into training and test sets but does not specify the exact percentages or counts for these splits, nor does it explicitly detail a separate validation set split or how these splits can be reproduced.
Hardware Specification Yes In the p PCA experiment, training a single VAE takes 12 seconds, while step one and two of our method together take 53 seconds (on a machine with a single NVIDIA Ge Force RTX 2070 GPU).
Software Dependencies No The paper mentions software components like "neural network" and the "Gumbel-softmax trick", but it does not specify any programming languages, libraries, or solvers with version numbers (e.g., Python 3.x, PyTorch 1.x).
Experiment Setup Yes Unless stated otherwise, we use 30 particles per iteration for training the models (as in [19]), 10,000 samples for reporting log-likelihood proxies, and 200 particles for making decisions. All results are averaged across five random initializations of the neural network weights. [...] AIS is computationally intensive, so we used 500 steps and 100 samples from the prior to keep the runtime manageable.