Seminar: Application of MDL theory in human connectomics

Academic Report of Professor Zhenghui Hu

Title:Application of MDL theory in human connectomics

Speaker: Zhenghui Hu (zhejiang university)

Time: 10 a.m. Wednesday. September 11th, 2013

Address:B502

Summary:

 In year 2008, David and his colleagues from INSERM of France, compared the effectiveness of the existing two models in neuron connectomics- the Dynamic Causality Model(DCM) and the Granger causality model(GCM), by simultaneously collecting iEEG and fMRI signals. The results showed that none of these methods could map out correct neural network. Their work was published on open-source  PLoS Biology, and commented  by Firston, the creater of DCM model,  in year 2010 PLoS Biology. The interaction led to numerous discussion and debates on the two models in the field of Neuroimaging, which were all then published on NeuroImage as a special edition.

By re-analyzing the David’s simultaneous EEG and fMRI data, we have found the result from MDL model perfectly matches the result from iEEG analysis, which serves as a ground truth. Meanwhile, the conventional F-test failed to show the consistency.  Our result improved the advantage of MDL theory in human connectomics, and showed how improtant the model selection is in neuron circuitry study.  

In addition, we have compared the effectiveness of the linear and nonlinear network models. We have found that with the increase of noisy, higher-ordered node and crossed information were gradually overcome by the noise signal. Actually, under the most fMRI noice level, the basic linear model was able to satisfy the requirements for quantitative marks in neuron circuits. This method applies to other medical imaging data as well.