FRITES - Framework for Information Theoretical analysis of Electrophysiological data and Statistics

[Andrea Brovelli]

Dear all,

We would like to draw your attention to an ongoing project and software
development based on information theory and statistical tools for the
analysis of neurophysiological data (MEG and SEEG /LFP so far):

https://github.com/brainets/frites

Happy to discuss further and share ideas.

Best,

Andrea and Etienne

[Danylo Ulianych]

great project, thanks for sharing.

I was browsing the source files, in particular, to find the differences between your MI estimation and the original GCMI. Original Robin's function notations like 'gcmi_cc' and 'mi_gg' always require cognitive processing to understand which function to use, and, sadly, I see that you followed his notations. Also, if 'gcmi_1d.py' is a copy of Robin's GCMI, which I remember works with multi-dimensional arrays, then what 'gcmi_nd.py' does? The documentation says that the latter is for n-dimensional arrays, but surely it does not imply that the original Robin's MI estimates are only for 1-d arrays.

Maybe, your end users won't look into the core of your package, but that's exactly what interests me. If I knew which function of yours to use, I'd give it a try compare with other libraries I have been using so far for MI estimation https://github.com/dizcza/entropy-estimators.

I'm also unsure whether your core solution can work with any type of probability distributions or only Gaussian?

Danylo

[Andrea Brovelli]

Dear Danylo,

The code gcmi_nd.py is indeed a generalisation of Robin's code gcmi_1d.py which is multivariate, but does not allow to analyse multiple dimensions in parallel. This speeds up computations a lot, especially if you need to do statistics (e.g., permutations tests). The latest matlab version of Robin's toolbox however implements this feature as well.

Concerning the core functions, I suggest you to look into https://github.com/brainets/frites/blob/master/frites/core/gcmi_nd.py

For example, you can test and compare wit other methods the following functions

1) Multi-dimentional MI between two Gaussian variables mi_nd_gg

2) Multi-dimentional conditional MI between three Gaussian variables cmi_nd_ggg

3) And their associated Gaussian-copula versions e.g., gcmi_nd_cc

Concerning your last question, the input data are (potentially multivariate) random variables, rather than probability distributions.

PS1: I added Etienne Combrisson to the discussion, who is the main developper and contributor of FRITES.

PS2: we are currently writing the paper that explains the methods implemented, but for a clear overview of the theory behind te toos you can ready Ince et al., HBM 2017

Cheers,

Andrea

Le 05-Nov-20 à 3:34 PM, Danylo Ulianych a écrit :

--
You received this message because you are subscribed to the Google Groups "Neural Ensemble" group.
To unsubscribe from this group and stop receiving emails from it, send an email to neuralensembl...@googlegroups.com.
To view this discussion on the web, visit https://groups.google.com/d/msgid/neuralensemble/61441917-d2fd-4e27-8b68-3b9b014187dfn%40googlegroups.com.