The Monge gap (unofficial implementation)

Project description

Original paper: https://arxiv.org/abs/2302.04953.

Project report and its presentation can be found in the docs folder.

Computational resources

All the models were trained using Colab, Kaggle or DataSphere environments with GPU configuration selected (T4 for Colab, P100 for Kaggle and V100 for DataSphere).

Possible improvements

• Experiment with MLP's hidden layers sizes.
• Use alternative Sinkhorn algorithm implementations (at least increase the number of iterations).
• For single-cell genomics dataset, test the method for all 34 treatments within 4i part.