Dear Chou-Ching,
2010/4/16 林宙晴 <[log in to unmask]>:
>
>
> Dear Vladimir
>
> Thank you very much for your help.
>
>
>
> I am wondering how DCM deals with local and global minimum. At different
> multiples of 64 runs, sometimes A Hz and B Hz look different.
>
Yes, we encountered the problem of local maxima also here and we are
planning to improve the optimization scheme to deal with it better.
You should always make sure you get decent fits before you interpret
the parameters. From our experience if your optimization gets stuck in
a local maximum (in this case it usually converges after <30
iterations) you can restart the inversion and answer 'yes' to 'use
previous priors' and 'use previous posteriors'. Then the optimization
should find a better solution.
>
>
> Does the old version of spm_dcm_ind_data.m presented wrong data?
>
Yes, if you use ECDs the data that is computed is wrong in the old
version (this is a specific problem in DCM-IR so the users of other
DCMs shouldn't panic).
> So my results using the old file might be wrong?
>
Yes.
>
>
> The attached file is an example that I have obtained by using DCM.
>
> When I type DCM.B, it shows [4x4 double] [4x4 double], but I can’t see the
> values.
>
> When it is reloaded into DCM, dots of B just disappears.
>
>
Could you also send me the data (mat and dat)? I can't load the DCM
properly without it. You can use yousendit.com if it's too big.
Best,
Vladimir
>
> It seems those A Hz and B Hz data are stored, so that they can be re-plotted
> when the result DCM file is re-loaded.
>
>
> Sincerely,
>
> Chou-Ching
>
>
>
>>
>> 寄件人: Vladimir Litvak <[log in to unmask]>
>> 日期: 2010年4月15日 下午05時48分23秒格林威治標準時間-04:00
>> 收件人: "Chou-Ching K. Lin" <[log in to unmask]>
>> 副本: [log in to unmask]
>> 標題: Re: [SPM] Problems with DCM for M/EEG
>>
>> Dear Chou-Ching,
>>
>> On Thu, Apr 15, 2010 at 2:50 PM, Chou-Ching K. Lin <[log in to unmask]>
>> wrote:
>>
>> Hi, Dear SPM friends, now I am probing and running DCM for M/EEG (induced
>> response). I found some small problems.
>>
>> First, it stops at 64 rounds of calculation every time before convergence.
>> Is there any way to change this magic number, 64?
>>
>> 64 iterations should be sufficient to get close to the optimum even
>> with no convergence but if you'd like to increase this number you can
>> only do it by changing the code (spm_nlsi_N around line 278).
>>
>> Also we recently found a bug in how the data for DCM-IR was prepared.
>> Put the attached function in toolbox\dcm_meeg to fix it.
>>
>> Second, when I reload the saved DCM file, all the parameters are
>> recovered, but those dots for B in the lower box are missing (every time).
>>
>> This might be a GUI problem. I haven't seen it before. You can send me
>> one example DCM file and I'll check.
>>
>> Third, is there any way to examine the calculated parameters (A and B)?
>> When I open the DCM result file in Matlab command line, it shows it is a
>> matrix or structure. But it does not give me the values.
>>
>> The values are in the structure. I suggest that you put a breakpoint
>> in spm_dcm_ind_results when you do the results display, find the
>> relevant part if the code and see how what you are interested in is
>> extracted from the structure. Some things are not actually stored but
>> are computed on the fly from things that are stored. If you are just
>> interested in coupling matrices in Hz you can export them as images
>> using the option in the 'Results' part.
>>
>> Best,
>>
>> Vladimir
>>
>>
>
>
|