Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]

Semi-Supervised Biomedical Translation With Cycle Wasserstein Regression GANs

Authors: Matthew McDermott, Tom Yan, Tristan Naumann, Nathan Hunt, Harini Suresh, Peter Szolovits, Marzyeh Ghassemi

AAAI 2018 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We first evaluate our method on synthetic experiments, demonstrating two primary advantages of the system: 1) distribution matching via the adversarial loss and 2) regularization towards invertible mappings via the cycle loss. We then show a regularization effect and improved performance when paired data is supplemented by additional unpaired data on two real biomedical regression tasks: estimating the physiological effect of medical treatments, and extrapolating gene expression (transcriptomics) signals.
Researcher Affiliation Collaboration Matthew B. A. Mc Dermott MIT, Cambridge, MA EMAIL Tom Yan MIT, Cambridge, MA EMAIL Tristan Naumann MIT, Cambridge, MA EMAIL Nathan Hunt MIT, Cambridge, MA EMAIL Harini Suresh MIT, Cambridge, MA EMAIL Peter Szolovits MIT, Cambridge, MA EMAIL Marzyeh Ghassemi MIT, Verily, Cambridge, MA EMAIL
Pseudocode No No pseudocode or algorithm blocks are explicitly presented in the paper.
Open Source Code Yes Code available at https://github.com/mmcdermott/CWR-GAN.
Open Datasets Yes We use data from the publicly available Multiparameter Intelligent Monitoring in Intensive Care (MIMIC-III v1.4) database (Johnson et al. 2016)... The L1000 developers have released a dataset of 100,000 full transcriptomes, split between the 978 landmark genes and those remaining, to the NCBI GEO database under series number GSE70138 (Broad Connectivity Map Team 2016).
Dataset Splits Yes Models were tuned, then evaluated via nested crossvalidation... Hyperparameters were chosen according to a grid search with a randomly sampled 15% validation set.
Hardware Specification No The paper does not provide specific hardware details (e.g., CPU/GPU models, memory) used for running its experiments. It only mentions general setups like "All networks in this work" or refers to previous WGAN works for similar settings.
Software Dependencies No All models were implemented in Tensorflow (Abadi et al. 2016). We use the Adam optimizer (Kingma and Ba 2014)... While software is mentioned, specific version numbers for TensorFlow or Adam are not provided, which are necessary for reproducible software dependencies.
Experiment Setup Yes All networks in this work use a Leaky Re Lu activation, with α = 0.3... Adam optimizer... with hyperparameters similar to those recommended in prior work (Gulrajani et al. 2017) (α = 0.00005, β1 = 0.5, β2 = 0.9) in the CWR-GAN for critics and generators... All regression and critic networks were 3layer, bidirectional regressors using leaky Re LU activations, dropout of 0.75, and L2 & L1 regularization of 1e 3... Loss multipliers were fixed independently of task at a multiplier of 10 for the regression component and 1 for both the adversarial and cycle reconstruction error losses. The gradient loss multiplier was set to 10, but if a critic appeared to suffer from gradient explosion during training, it was increased to 50. Models were trained for up to 9 consecutive critic epochs, stopping after 3 critic epochs that did not improve the adversarial loss, then 1 translator epoch.