Email updates

Keep up to date with the latest news and content from BMC Neuroscience and BioMed Central.

This article is part of the supplement: Abstracts from the Twenty Second Annual Computational Neuroscience Meeting: CNS*2013

Open Access Poster presentation

Inferring effective computational connectivity using incrementally conditioned multivariate transfer entropy

Joseph T Lizier12* and Mikail Rubinov34

Author Affiliations

1 CSIRO Information and Communications Technology Centre, Marsfield, NSW 2122, Australia

2 Max Planck Institute for Mathematics in the Sciences, 04103 Leipzig, Germany

3 Department of Psychiatry, University of Cambridge, Cambridge, UK

4 Churchill College, University of Cambridge, Cambridge, UK

For all author emails, please log on.

BMC Neuroscience 2013, 14(Suppl 1):P337  doi:10.1186/1471-2202-14-S1-P337

The electronic version of this article is the complete one and can be found online at: http://www.biomedcentral.com/1471-2202/14/S1/P337


Published:8 July 2013

© 2013 Lizier and Rubinov; licensee BioMed Central Ltd.

This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Poster presentation

Effective connectivity analysis is a popular approach in computational neuroscience, whereby one seeks to infer a network of directed edges between neural variables (e.g. voxels or regions in fMRI, or reconstructed source time series in MEG data) which can explain their observed time-series dynamics. This is an important approach in understanding brain function, contrasting with functional connectivity analysis in being directed and dynamic, and with structural connectivity analysis in not requiring interventions and in being task-modulated. In particular, effective connectivity analysis seeks to find a minimal circuit model that can reconstruct the activity patterns contained in the given data. Ideally, such inference would be: made using model-free techniques; capture non-linear, multivariate, directional relationships; handle small amounts of data, and be statistically robust.

The information-theoretic measure transfer entropy (TE) [1] (a non-linear Granger causality) is becoming widely used for this purpose [2]. However its use is generally focussed only on interactions between a single source and destination (inferring an edge where TE is statistically significant), or else attempting multivariate considerations by conditioning on all other variables in the system (only practically possible with small numbers of variables or linear interactions). We aim to extend TE-based effective network inference to multivariate techniques, specifically: capturing collective interactions where target outcomes are due to multiple source variables (synergies); eliminating spurious connections for correlated sources (redundancies); and avoiding combinatorial explosions in source groups evaluated. We aim to maximize inference of true interactions while minimizing inference of spurious interactions.

In this manner, we describe a new method [3] which addresses the above requirements in considering multivariate source interactions. For each node in the network, the method identifies the set of source nodes which provide the most statistically significant information regarding its dynamics, and are thus inferred as those source information nodes from which that destination is computed. This is done using incrementally conditioned TE, gradually building the set of source nodes for a destination conditioned on the previously identified sources.

For validation purposes, we apply our method to various synthetic models of dynamics on networks (e.g. linearly coupled Gaussian autoregressive time series in Figure 1), and demonstrate the utility of the method in revealing significant proportions of the underlying structural network given only short time-series of the network dynamics. The results show particular utility in comparison to other methods (i.e. Figure 1 shows our incremental method significantly outperforms conditioning on all other variables, even on this data set with linear relationships where the latter is possible). Following validation studies we will apply our technique to neuroimaging data sets.

thumbnailFigure 1. ROC curve for inference on discrete-time autoregressive time-series for 400 observations (100 nodes, 400 edges, random directed structure, edge weights 0.5, random noise variance 1).

References

  1. Schreiber T: Measuring information transfer.

    Phys Rev Lett 2000, 85:461-464. PubMed Abstract | Publisher Full Text OpenURL

  2. Vicente R, Wibral M, Lindner M, Pipa G: Transfer entropy-a model-free measure of effective connectivity for the neurosciences.

    J Comp Neurosci 2011, 30(1):45-67. PubMed Abstract | Publisher Full Text | PubMed Central Full Text OpenURL

  3. Lizier JT, Rubinov M: Multivariate construction of effective computational networks from observational data.

    Max Planck Institute for Mathematics in the Sciences Preprint 25/2012 2012. OpenURL