 |
 |
At 2010-12-02 19:42:41，"Vladimir Litvak" <[log in to unmask]> wrote:
Dear Haoran,
What I meant is that the generative model for MEG/EEG data looks something like:y(t)= L*f(t,u,θ)+εwhere y is your sensor data f is the function generating predicted source data, based on the inputs and involving a neural mass model (see http://www.fil.ion.ucl.ac.uk/spm/doc/papers/od_dcm_erp.pdf) and L is the so called lead field matrix or a set of weights mapping sources to sensors. Now, you could invert this model in two ways.You could minimizenorm(pinv(L)*y(t) - f(t,u,θ))i.e. first compute source data and then model it, which is what you implied in your original question. But this is not what SPM does. SPM minimizes:norm(y(t) - L* f(t,u,θ))i.e. the predicted source data is multiplied by the lead field or 'projected to the sensors' and then compared to the sensor data. So there is no 'observed source data' computed in DCM at any point there is only predicted source data, predicted sensor data and observed sensor data. The model evidence is composed of the error term above and some other terms that have to do with complexity, again you can find the details in the papers.Best,Vladimir
2010/12/2 飞鸟 <[log in to unmask]>:
> Dear Vladimir,
> Thank you very much for your zealous help, I have got
> a further understanding about my first question, but not absolutely, so I
> will try to verify my thoughts.
> By the way, after reading your last letter, I was always thinking what the
> first sentence meant. I think there may be somewhere wrong when I study the
> basis of the DCM. Could you give me a more detailed explannation about the
> first sentence you wrote. Or you can tell me if I get a wrong recognition.
> 1. What is the model data? Where is it from? Is it the computation result
> of the foward model (y=g(x,θ)+ε) ?
> 2.Does the word "projects" mean "through the lead field"?
> 3.After the procedure of "projects model data to the sensors and
> compares it to the sensor data", and then will we get the log-evidence of
> the current model?
> Best wishes to you!
>
> Haoran Li (Ms.)
> Brain Imaging Lab
> Research Center for Learning Science,
> Southeast University,China
> Your relpy:"DCM for ERP/ERF does not extract source data at all, but
> projects
> model data to the sensors and compares it to the sensor data. The IMG
> option represents each area by multiple sources already, but the
> activities of these sources are related in a particular
> way.Effectively it is similar to extracting a principal component
> which is what you seem to suggest."