Hi all,
I am trying to use the MNE mixed norms solver to get the locations of
robust sources in a typical lexical decision task in adult subjects. An
issue I am having is how to set the regularization parameter (alpha) which
ranges from 0-100(high). Although the mne-python docstring has no
recommended default value, in the tutorials 55 is used.
At that value and e.g., 25, mxne_inverse throws No "active dipoles found.
alpha is too big." exception.
I am assuming alpha is or is related to lambda in formulation 3 in Gramfort
et al., 2012, and thus determines the best fitting solution? What's not
clear is how to set alpha across datasets to get around these
non-convergence cases?
I am trying to use the MNE mixed norms solver to get the locations of robust
sources in a typical lexical decision task in adult subjects. An issue I am
having is how to set the regularization parameter (alpha) which ranges from
0-100(high). Although the mne-python docstring has no recommended default
value, in the tutorials 55 is used.
first you should use n_mxne_iter > 1. Typically 10. It makes the
results less sensitive
to the alpha parameter. It makes it also possible to inspect the
residuals to see
how much data you have actually explained with the estimated sources.
At that value and e.g., 25, mxne_inverse throws No "active dipoles found.
alpha is too big." exception.
the more you have strong sources, the higher can be the alpha.
You should reduce the parameter here.
I am assuming alpha is or is related to lambda in formulation 3 in Gramfort
et al., 2012, and thus determines the best fitting solution? What's not
clear is how to set alpha across datasets to get around these
non-convergence cases?
yes it's the lambda with some scaling by lambda_max if you use n_mxne_iter=1.
If you use n_mxne_iter > 1, it corresponds to this extension:
note that we are still exploring options to make it easier to set this
regularization
parameter such as: