Dear Adeel:
After I solved the problem of functional connectivity constraints, I now run into the problem of convergence – the spectral DCM (for more than half of subjects) cannot converge within 128 iterations, even though I reduced the number of nodes from 36 to 27. As reported in your 2017 Network Neuroscience paper, all spectral DCMs were able to converge within 128 iterations. So I am wondering what possible reasons could be that cause very slow convergence. If a DCM does not converge, I cannot use the estimates, right? Should I increase the maximal iteration beyond 128?
As you suggested, I may try to split the graph into two subgraphs (hemispheres) and use the posterior estimates of subgraphs as the initial values for inverting a full graph. Will this help to converge faster (within 128 iterations)? I am also not sure how to set the initial values of the full graph with the posterior estimates of subgraphs. For example, if DCM_L stores the (posterior) estimates of the left graph (network) and DCM_R stores the (posterior) estimates of the right graph. How do I use the posterior estimates from DCM_L and DCM_R to set the initial values in the full graph (DCM_F)? I know the initial parameters are set in DCM.options.P, but not sure what DCM.options.P should look like. Could you please give me some example codes for this initialization?
Thank you very much!
Best regards,
Guoshi Li
|