ENH: decoding module 2017

[kingjr]

This aims at keeping track of the development related to the Google summer of code 2016 on decoding analyses.

We have quite a lot ongoing PRs, so I thought I would try to organize them here to keep track of the big picture. I'll edit the post along the issue along the way.

The aim is to make transformers that follow the sklearn API:

pipe = make_pipeline(
    CSP(sfreq=200, None, 30),
    TimeFrequency(),
    SlidingEstimator(make_pipeline(StandardScaler(), LogisticRegression())
)

score = cross_val_score(X=epochs.get_data(), y=epochs.event[:, 2])

For now, we're focusing on sklearn integration, not high level features (plotting, getcoefs etc).

• Spatial filters:

  • [x] FIX: XdawnTransformer: #3245
  • [x] FIX: Xdawn refactoring: #3425
  • [x] FIX: CSP renaming of epochs_data into X #3485
  • [x] ENH: CSP adding multiclass support see #3484, #3485
  • [ ] ENH: CSP uwedge multiclass + example https://github.com/mne-tools/mne-python/issues/3495
  • [x] ENH: Unsupervised spatial filter https://github.com/mne-tools/mne-python/pull/3447
  • [X] ENH: EMS: #3427, #3446
  • [X] ENH: CSP transform into power or time course or #3583 #3586
  • [x] ENH: Mixin class to plot patterns and filters? see https://github.com/mne-tools/mne-python/issues/3424
  • [ ] FIX: Move Xdawn to decoding?

• Freq - time/freq transformers:

  • [x] FIX: TemporalFilter: needs refactoring from mne.decoding.FilterEstimator to pass explicit args and not info + BUG in FilterEstimator: #3395, #3471, #3472
  • [X] FIX: PSDEstimator
  • [x] ENH: TimeFrequencyDecomposer #3488
  • [x] ENH: SPoC #4144
  • [ ] Time Frequency Decoding object based on covariances / CSP ...
  • [ ] Power Decoding Estimator based on covariances / CSP: #4138 (example, but need to be converted in object)

• Continuous signals (raw not epochs)

  • [x] FIX: Xdawn overlap: to be integrated with the RERP code #2332, also see https://github.com/mne-tools/mne-python/blob/master/mne/stats/tests/test_regression.py#L124, PR in #3563
  • [x] FIX: ReceptiveField: started in #3310, done in #3728
  • [ ] ~FIX: combine rERP and STRF: e.g. float delays instead of list, discrete vs continuous regressors etc~
  • [ ] ~Optimize ReceptiveField for continuous regression~

• Search lights

  • [x] ENH: SlidingEstimator and GeneralizingEstimator: #3381 #4103
  • [x] ENH: n-dimensional search light #3481
  • [X] FIX: Refactor TimeDecoding & GeneralizationAcrossTime #4103
  • [x] ENH: 'scoring' param in SearchLight-like object #3475, #3502, #3833 #4103
  • [ ] ENH: option to warm_start from one estimator to the next
  • [X] cross_val_multiscore #4103
  • [ ] window_size and step parameter in SlidingEstimator

• Preprocessing

  • [X] ENH: Scaler: scale channels over all time their corresponding time points to deal with issues related to mix channel types needs refactoring to comply to X y API
  • [x] ENH: Vectorizer: to pass from n-D X to 2D X to be reviewed in https://github.com/mne-tools/mne-python/pull/3409

• Examples:

  • [x] RSA #3923
  • [x] TimeFrequency Example #4115
  • [x] PSD example #4138
  • [x] get_coef for SlidingEstimator
  • [ ] Denoise tutorial
  • [ ] ~Frequency generalization #4019~
  • [ ] UnsupervisedSpatialFilter example

• Other:

  • [ ] ~partial_fit #3483, #3591~
  • [x] get_coef(pipeline) to retrieve and/or invert_transform linear coefficients if they exist

[larsoner]

@kingjr is this issue still the roadmap, or should we close?

[kingjr]

Still rough roadmap, although needs to be updated. Let's close it after the sprint IMO.

[kingjr]

I'm actually for keeping it up ;)

[jona-sassenhagen]

@kingjr , @agramfort 's idea is one which will result in rERP and the receptive field module staying separate forever.

• rERP becomes restricted to the original idea, allowing only sparse predictors; dense predictors are handled by receptive field module
• receptive field module should be extended slightly to simply also using sparse predictors (- XDawn and rERP should still be united)

[jona-sassenhagen]

I closed/crossed out a few that had been done since. 8 issues still open.

[jona-sassenhagen]

@kingjr any idea for an example for UnsupervisedSpatialFilter with the example data? The only thing I can think of is show that you can use an unregularized algorithm when p > n by reducing p in dimensionality via PCA, i.e., for the MEG data with many channels and few trials.

[kingjr]

How about PCA and then sliding with svm multiclass on the left/right audio/visual dataset ?

On Saturday, 4 August 2018, jona-sassenhagen notifications@github.com wrote:

@kingjr https://github.com/kingjr any idea for an example for UnsupervisedSpatialFilter with the example data? The only thing I can think of is show that you can use an unregularized algorithm when p > n by reducing p in dimensionality via PCA, i.e., for the MEG data with many channels and few trials.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/mne-tools/mne-python/issues/3442#issuecomment-410476362, or mute the thread https://github.com/notifications/unsubscribe-auth/AEp7DPMjiTgyOOWCHNAnB7cb8KjPACRaks5uNgcxgaJpZM4JRF5A .