Code accompanying the paper Shared Independent Component Analysis for Multi-subject Neuroimaging

Move into the ShICA directory cd ShICA

Install ShICA pip install -e .

Move into the experiments directory cd experiments

Run the bash script to produce results (should take approximately 3 minutes on a modern laptop) bash run_all.bash

Move into the plotting directory cd plotting

Run the bash script to produce figures from the results bash plot_all.bash

Figures are available in the figures directory.

Performances on Gaussian sources:

Performances on non Gaussian sources:

Performances when some sources are Gaussian and some non-Gaussian:

Note The current implementation uses only 10 seeds and 4 different number of samples in the curves so that computation time is low even on a laptop. In order to obtain exactly the same curves as in the paper you should modify the files rotation.py, full_nongaussian.py and semigaussian.py in the experiments directory so that

num_points = 20
seeds = np.arange(40)
ns = np.logspace(2, 5, num_points)

We give the code to run experiments on timesegment matching.

Move into the data directory

cd experiments/data

Launch the download script (Runtime 34m6.751s)

bash download_data.sh

Mask the data (Runtime 15m27.104s)

python mask_data.py

Move into the experiments directory

cd experiments

Run the experiment on masked data (Runtime 17m39.520s)

python timesegment_matching.py

This runs the experiment with n_components = 5 and benchmark ShiCA-J and ShICA-ML with SRM as the dimension reduction method.

https://hugorichard.github.io/ShICA/index.html

If you use this code in your project, please cite:

@inproceedings{NEURIPS2021_fb508ef0,
 author = {Richard, Hugo and Ablin, Pierre and Thirion, Bertrand and Gramfort, Alexandre and Hyvarinen, Aapo},
 booktitle = {Advances in Neural Information Processing Systems},
 editor = {M. Ranzato and A. Beygelzimer and Y. Dauphin and P.S. Liang and J. Wortman Vaughan},
 pages = {29962--29971},
 publisher = {Curran Associates, Inc.},
 title = {Shared Independent Component Analysis for Multi-Subject Neuroimaging},
 url = {https://proceedings.neurips.cc/paper/2021/file/fb508ef074ee78a0e58c68be06d8a2eb-Paper.pdf},
 volume = {34},
 year = {2021}
}