“School of Cognitive Sciences”
Back to Papers HomeBack to Papers of School of Cognitive Sciences
| Paper IPM / Cognitive Sciences / 15903 |
|
||||
| Abstract: | |||||
|
Bayesian networks were efficiently applied for brain decoding along with connectivity information used in structure learning of Bayesian networks. The modified structure learning proposed expands the application of Bayesian networks in brain-decoding.
There are various structure learning methods which usually require large training samples. As the simplest method, tree structure learning commonly uses measures like correlation or mutual information to quantify dependencies in a network. Correlation captures the linear relations and mutual information detects linear and nonlinear dependencies while it needs more data to estimate the distribution of the variables. In current study, these measures were substituted with cross recurrence quantifiers which efficiently describes nonlinear behaviors of biological systems and effectively detect linear and nonlinear interactions without requiring long length of data. As novel applications, we used cross recurrence quantifiers for structure learning of Bayesian networks and functional connectivity analysis of fMRI data.
To compare the performance of cross recurrence quantifiers with conventional indices in learning the structure of Bayesian networks, we used them in retrieving known structures. Moreover, the new method was applied for extracting fMRI brain connections in Bayesian networks devised for decoding-reconstruction of hand written digits. Objective and subjective assessment of results showed that Bayesian networks which used correlation or mutual information yielded significantly inferior accuracy in comparison to Bayesian networks which employed cross recurrence quantifiers.
Therefore, cross recurrence quantifiers could be efficient measures in structure learning of Bayesian networks or functional connectivity analysis of fMRI data.
Download TeX format |
|||||
| back to top | |||||
