Discriminative Regularization for Latent Variable Models with Applications to Electrocardiography

Authors: Andrew Miller, Ziad Obermeyer, John Cunningham, Sendhil Mullainathan

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

Reproducibility Variable Result LLM Response
Research Type Experimental We measure our approach on synthetic and real data with statistical summaries and an experiment carried out by a physician.
Researcher Affiliation Academia Andrew C. Miller 1 Ziad Obermeyer 2 John P. Cunningham 3 Sendhil Mullainathan 4 1Data Science Institute, Columbia University, New York, NY, USA. 2School of Public Health, UC Berkeley, Berkeley, CA, USA. 3Department of Statistics, Columbia University, New York, NY, USA. 4Booth School of Business, University of Chicago, Chicago, IL, USA.
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Code available at https://github.com/andymiller/DR-VAE.
Open Datasets No The paper mentions using a dataset of clinical EKGs but does not provide specific access information like a link, DOI, repository name, or formal citation with authors and year for a publicly available dataset.
Dataset Splits Yes We split our cohort by patients into a training/validation development set (75% of patients) and a testing set (25% of patients) no patients overlap in these two groups.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments.
Software Dependencies No The paper mentions using Adam optimizer but does not specify version numbers for any software dependencies.
Experiment Setup Yes Each discriminative model has two hidden layers of size 100 and a Re LU nonlinearity. Discriminative models are trained with dropout (p = .5) and stochastic gradient optimization with Adam (Kingma & Ba, 2014), starting with the learning rate set to .001 and halved every 25 epochs. We save the model with the best performance on the validation set. Throughout our experiments we compare a standard VAE to the DR-VAE with values of β = [1,5,10,100] (note that a DR-VAE with β = 0 corresponds to a standard VAE). All deep generative models have one hidden layer with 500 units and a Re LU nonlinearity. We also train generative models with gradient-based stochastic optimization using Adam (Kingma & Ba, 2014), with an initial learning rate of .001 that is halved every 25 epochs.