Functional Connectivity via Total Correlation: Analytical results in Visual Areas
arXiv:2208.05770 · doi:10.1016/j.neucom.2023.127143
Abstract
Recent studies invoke the superiority of the multivariate Total Correlation concept over the conventional pairwise measures of functional connectivity in biological networks. Those seminal works certainly show that empirical measures of Total Correlation lead to connectivity patterns that differ from what is obtained using the most popular measure, linear correlation, or its higher order and nonlinear alternative Mutual Information. However, they do not provide analytical results that explain the differences beyond the obvious multivariate versus bivariate definitions. Moreover, the accuracy of the empirical estimators could not be addressed directly because no controlled scenario with known analytical result was provided either. This point is critical because empirical estimation of information theory measures is always challenging. As opposed to previous empirical approaches, in this work we present analytical results to prove the advantages of Total Correlation over Mutual Information to describe the functional connectivity. In particular, we do it in neural networks for early vision (retina-LGN-cortex) which are realistic but simple enough to get analytical results. The presented analytical setting is also useful to check empirical estimates of Total Correlation. Therefore, once certain estimate can be trusted, one can explore the behavior with natural signals where the analytical results (that assume Gaussian signals), may not be valid. In this regard, as applications (a) we explore the effect of connectivity and feedback in the analytical retina-LGN-cortex network with natural images, and (b) we assess the functional connectivity in visual areas V1-V2-V3-V4 from actual fMRI recordings.
arXiv comment31 pages, 14 figures, Accepted to Neurocomputing Journal
References in corpus (14)
- Measuring Information Transfer
- Estimating Mutual Information
- Networks beyond pairwise interactions: structure and dynamics
- Granger causality and transfer entropy are equivalent for Gaussian variables
- BOLD5000: A public fMRI dataset of 5000 images
- Iterative Gaussianization: from ICA to Random Rotations
- Information Theoretical Estimators Toolbox
- Derivatives and Inverse of Cascaded Linear+Nonlinear Neural Models
- Nonlinearities and Adaptation of Color Vision from Sequential Principal Curves Analysis
- Functional Connectome of the Human Brain with Total Correlation
- PerceptNet: A Human Visual System Inspired Neural Network for Estimating Perceptual Distance
- Information Flow in Biological Networks for Color Vision
- On the Suitable Domain for SVM Training in Image Coding
- Estimating the contribution of early and late noise in vision from psychophysical data