yongjiezhu / nICAmeg

Code for "Nonlinear ICA for MEG/EEG data analysis"

Unsupervised representation learning of spontaneous MEG data with nonlinear ICA

This repository contains code to show how to train nonlinear ICA with resting-state MEG and then be used to dowenstream classification tasks on label-limited data sets, which is presented in Unsupervised representation learning of spontaneous MEG data with nonlinear ICA, published at NeuroImage 2023. Work done by Yongjie Zhu (U. of Helsinki), Tiina Parviainen (U. of Jyväskylä), Erkka Heinilä (U. of Jyväskylä), Lauri Parkkonen (Aalto) and Aapo Hyvärinen (U. of Helsinki).

Dependencies

This project was tested with the following versions:

• python 3.8
• PyTorch 1.13.1
• mne 1.3.0
• scikit-learn 1.2.1
• pickle 4.0

Usage

The well-trained models are saved under ‎results/. You can download it and use it to your specific data (e.g., for classification). You can also train the model with your data by type ‎python nICA_trainning.py --data_path "with your path" to train the NICA model with your data.
e.g. trainning NICA(TCL) model with CamCan:

python nICA_trainning.py --data_path /camcan/meg_rest -m tcl

Type ‎python nICA_trainning.py --help to learn about hyperparameters.

For the downstream tasks, type ‎python nICA_downstreams.py --data_path "with your path" to perform specific downstream tasks with trained NICA models. e.g.:

python nICA_downstreams.py --data_path /camcan/meg_passive --path results/

Refer to utils/vis.py for visualizing the neural networks with occlusion sensitivity analysis. You can easily modify some parameters in ‎utils/preprocessing.py with your styles to preprocess the MEG data. ‎utils/featsextraction.py contains functions for feature extraction using trained models. The bash job can also be used to run multiple experiments in parallel (e.g. with different random seeds) on a SLURM-equipped server.

Reference

If you find this code helpful/inspiring for your research, we would be grateful if you cite the following:

@article{zhu2023unsupervised,
  title={Unsupervised representation learning of spontaneous MEG data with Nonlinear ICA},
  author={Zhu, Yongjie and Parviainen, Tiina and Heinilä, Erkka and Parkkonen, Lauri and Hyvärinen, Aapo},
  journal={NeuroImage},
  volume = {274},
  pages={120142},
  year={2023},
  publisher={Elsevier}
}

@inproceedings{hyvarinen2016unsupervised,
  title={Unsupervised feature extraction by time-contrastive learning and nonlinear ica},
  author={Hyvärinen, Aapo and Morioka, Hiroshi},
  booktitle={Advances in Neural Information Processing Systems (NIPS2016)},
  volume={29},
  address={Barcelona, Spain},
  year={2016}
}

@inproceedings{morioka2021independent,
  title={Independent innovation analysis for nonlinear vector autoregressive process},
  author={Morioka, Hiroshi and Hälvä, Hermanni and Hyvärinen, Aapo},
  booktitle={Proc.\ Artificial Intelligence and Statistics (AISTATS2021)},
  pages={1549--1557},
  year={2021},
  organization={PMLR}
}

License

A full copy of the license can be found here.