Learning Cortico-Muscular Dependence through Orthonormal Decomposition of Density Ratios

Overview

Welcome to the GitHub repository for our paper Learning Cortico-Muscular Dependence through Orthonormal Decomposition of Density Ratios. Here, you will find datasets, models, and codes for training and visualization as described in the paper. 🧠💪

Contents

The repository is structured with notebook files for training and visualization, organized as follows:

• 1.1. Preprocessing ✅
• 2.1 Main Experiment - Inter Subject Experiments ✅
• 2.2 Main Experiment - Cross Subject Experiments ✅
• 3.1 Visualization - Span of Eigenfunctions ✅
• 3.2 Visualization - Temporal Channel Dependence ✅
• 4.1 Baseline - CMC ✅
• 4.2 Baseline - MI ✅

Getting Started

1. Preprocessing

Downloadable Files:

• Raw Data: https://doi.org/10.1093/gigascience/giaa098
• Filtered EEG/EMG Numpy Arrays: https://drive.google.com/drive/folders/1e0nqKcb7MoAJQd4aaEAfbvKjzURMVkyH?usp=sharing
• Trained Models: https://drive.google.com/file/d/1G-GdNQ6B38Eh2d4tTczLaZ5_NALxVa-o/view?usp=sharing, https://drive.google.com/file/d/1Y06Quuf1wYhHyZ6NnKDSTymN5PCmfR6u/view?usp=sharing

Processing Steps:

1. Download the necessary raw data files: eeg_train_trial.pt, emg_filtered_train_trial.pt, label_train_trial.pt, sub_train_trial.pt, and update the paths accordingly.
2. Execute the blocks in Preprocessing.ipynb to save filtered signals (outliers removed) and primary movement labels as numpy arrays.

2. Running Experiments

Inter Subject Experiments:

1. After preprocessing, execute the Inter Subject Experiments notebook section by section.
2. This will store the trained models and learning curves locally.
3. Note: Inter Subject is used for visualizations and must be run before visualization.

Cross Subject Experiments:

1. Running the Cross Subject Experiments notebook will save the highest accuracies during training.

3. Visualization

Span of Eigenfunctions:

1. Load the trained models and extract training and testing set features for both EEG and EMG data.
2. Use EEG features for visualizations.

Produced Figures:

• Projection of all samples, color-coded by subject index and movement types.
• Specific projections of individual subjects and their movements.
• Approximated density ratios values.

Temporal-Level and Channel-Level Dependencies:

1. Use the saved model from Inter Subject Experiments (2.1).
2. Produce temporal-level and channel-level dependencies figures.

4. Baseline:

• Contains codes for CMC and MINE.

Happy Experimenting!