Elastic Net Inverse Solutions

✉️ Mailing List Archive (read-only)
1

Hi All,

Has anyone done an inverse solution using Elastic Nets? Any pointers for
setting it up using as much pre-existing, well-tested code as possible?
E.g. I see that sklearn has Elastic Net functionality -- any best practices
for using sklearn linear models with the mne-python data structures?

Thanks,
Emily
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.nmr.mgh.harvard.edu/pipermail/mne_analysis/attachments/20170919/833e4f19/attachment.html

2

Hi Emily

Has anyone done an inverse solution using Elastic Nets?

you do have sparse solvers with L1 penalties in MNE. We call them MxNE or
TF-MxNE and are used with the (tf_)mixed_norm functions.

See:

https://martinos.org/mne/stable/auto_examples/inverse/
plot_mixed_norm_inverse.html

https://martinos.org/mne/stable/auto_examples/inverse/
plot_time_frequency_mixed_norm_inverse.html

such solvers take into account the spatio-temporal structure of the M/EEG
data.

See the referenced papers.

If you want to use with own solver and eventually use the sklearn sparse
solvers (that will only work with 1 time instant) you can start with this
example:

http://martinos.org/mne/dev/auto_examples/inverse/plot_
custom_inverse_solver.html

Let me know if you have any questions.

I've been working with these for some years now.

HTH
Alex

Any pointers for setting it up using as much pre-existing, well-tested
code as possible? E.g. I see that sklearn has Elastic Net functionality --
any best practices for using sklearn linear models with the mne-python data
structures?

Thanks,
Emily

_______________________________________________
Mne_analysis mailing list
Mne_analysis at nmr.mgh.harvard.edu
Mne_analysis Info Page

The information in this e-mail is intended only for the person to whom it
is
addressed. If you believe this e-mail was sent to you in error and the
e-mail
contains patient information, please contact the Partners Compliance
HelpLine at
MyComplianceReport.com: Compliance and Ethics Reporting . If the e-mail was sent to you in
error
but does not contain patient information, please contact the sender and
properly
dispose of the e-mail.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.nmr.mgh.harvard.edu/pipermail/mne_analysis/attachments/20170919/7c9a0c92/attachment.html

3

Hi Alex,

I was drawn to Elastic Nets because I'm dealing with a dataset that is
unlikely to be strictly sparse spatially, and I expect a large-ish region
to have quite highly correlated activity. My understanding and experience
with L1 regularization is that it chooses a small subset of these
correlated sources, rather than selecting a set of them (I woudl like the
solution to capture all of the correlated active sources). L2
regularization, of course, has the opposite problem, allowing all of the
sources to be nonzero.

In an ideal world I would like an inverse solution that can have a large
contiguous region of highly correlated active sources, and the rest of the
brain as zero or close to zero. I should be able to do this with elastic
nets by tuning the two regularization parameters. Is there a configuration
of mixed norm solvers that can do it?

I'll deal with the time dimension later, although I'm open to the idea of
doing the whole thing in the frequency domain, since I have a long stretch
of stationary data and the interesting effects are all in one frequency
band.

Thanks,
Emily

4

hi,

I was drawn to Elastic Nets because I'm dealing with a dataset that is

unlikely to be strictly sparse spatially, and I expect a large-ish region
to have quite highly correlated activity. My understanding and experience
with L1 regularization is that it chooses a small subset of these
correlated sources, rather than selecting a set of them (I woudl like the
solution to capture all of the correlated active sources). L2
regularization, of course, has the opposite problem, allowing all of the
sources to be nonzero.

this is correct for ENet vs L1 but my reaction is that it then boils down
to the problem of interpretation of an activation foci. ENet will not give
you the extent of the course and it's just a "proxy" towards a region of
confidence / uncertainty around the localized focal dipolar foci the L1
solver will give you.

In an ideal world I would like an inverse solution that can have a large
contiguous region of highly correlated active sources, and the rest of the
brain as zero or close to zero.

careful not to interpret this as source extension.

I should be able to do this with elastic nets by tuning the two
regularization parameters. Is there a configuration of mixed norm solvers
that can do it?

we never implemented E-Net with MxNE in MNE sorry.

I'll deal with the time dimension later, although I'm open to the idea of

doing the whole thing in the frequency domain, since I have a long stretch
of stationary data and the interesting effects are all in one frequency
band.

if it's stationary you can use MxNE after filtering the data in the band of
interest.

HTH
Alex
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.nmr.mgh.harvard.edu/pipermail/mne_analysis/attachments/20170919/79ea078e/attachment.html

5

Thanks, Alex! I'll keep thinking carefully about my options.

Emily

6

OK, one more question:

My ultimate goal *is* to estimate the spatial extent of the sources, and
I'm thinking very carefully about ways to approach the problem. Elastic
nets by themselves won't do it, but I'm hoping that I can make progress by
(1) choosing the tuning parameters in a data-driven way and (2) rigorously
describing the uncertainty in the resulting estimates.

Are you aware of other/better ways of doing spatial extent estimation?

Thanks,
Emily

7

Hi Emily,

I think for the spatial extents estimation, you can choose to use an L1 norm of total variation (TV) term defined using cortex spatial information.

Please check the papers from Dr. Lei Ding "Reconstructing cortical current density by exploring sparseness in the transform domain" or you can check "s-SMOOTH: Sparsity and Smoothness Enhanced EEG Brain Tomography" which has a higher order of smoothness definition.

Best regards,

Feng Liu

8

These papers look really interesting. Thanks!