m-guggenmos / megmvpa

MEG MVPA Tutorial

MEG MVPA Tutorial (Python & Matlab)

This tutorial accompanies the paper Multivariate pattern analysis for MEG: a comparison of dissimilarity measures.

Citation: Guggenmos, M., Sterzer, P., & Cichy, R. M. (2018). Multivariate pattern analysis for MEG: a comparison of dissimilarity measures. NeuroImage. DOI: 10.1016/j.neuroimage.2018.02.044

Python tutorial

Preparation

This tutorial is based on IPython/Jupyter Notebook files, which are linked below. In addition, the tutorial can be downloaded as a zip file, which includes the notebook files, additional code files and the example dataset used for this tutorial. To reduce computational costs, the dataset is for one participant only and includes only 9 of 92 experimental conditions.

Content of the zip file:

File	Description
cv.py	containing code for pseudo-trials/permutations/cross-validation
dissimilarity.py	containing a number of custom dissimilarity measures
weird.py	weighted robust distance classifier (WeiRD), see also here
python_decoding.ipynb	Notebook on Decoding
python_reliability.ipynb	Notebook on RDMs and Reliability
python_distance.ipynb	Notebook on Distance measures and cross-validation
data01_sess1.npy	data for subject 1, session 1
data01_sess2.npy	data for subject 1, session 2
labels01_sess1.npy	trial labels for subject 1, session 1
labels01_sess2.npy	trial labels for subject 1, session 2

In addition, the tutorial requires 4 established scientific python packages: numpy, scipy, scikit-learn, matplotlib

List of tutorials:

• Decoding
• RDMs and Reliability
• Distance measures and cross-validation

Matlab tutorial

Preparation

This tutorial is based on IPython/Jupyter Notebook files, which are linked below. In addition, the tutorial can be downloaded as a zip file, which includes the notebook files, additional code files and the example dataset used for this tutorial. To reduce computational costs, the dataset is for one participant only and includes only 9 of 92 experimental conditions.

Content of the zip file:

File	Description
cov1para.m	Shrinkage code (Ledoit & Wolf, 2004) for covariances
weirdtrain.m & weirdpredict.m	Weighted Robust Distance (WeiRD) classifier
gnbtrain.m & gnbpredict.m	Gaussian Naive Bayes (GNB) classifier
matlab_decoding.ipynb	Notebook on Decoding
matlab_reliability.ipynb	Notebook on RDMs and Reliability
matlab_distance.ipynb	Notebook on Distance measures and cross-validation
data01_sess1.mat	data for subject 1, session 1
data01_sess2.mat	data for subject 1, session 2
labels01_sess1.mat	trial labels for subject 1, session 1
labels01_sess2.mat	trial labels for subject 1, session 2

In addition, the tutorial assumes a working LIBSVM installation for Matlab.

List of tutorials:

• Decoding
• RDMs and Reliability
• Distance measures and cross-validation

megmvpa's Issues

Question about cross-validated correlation distance

Hello, thanks for making this code available! I'm working to replicate your distance notebook with some of our own data and in looking through the definition for the cross validated correlation I see that the denominator is sometimes the square root of covariances, which is not guaranteed to be positive (ie if train and test set have negative covariance, that calculation would fail). If regularize_var is set to true (which you have) this won't be an issue because of the max function, but I'm wondering if you have suggestions for the case when there is negative covariance between data splits? Thank you!