Is number of features selected by lasso bounded by number of samples in the dataset?

Question

I have been working on a project with fMRI data. It has a training set of $300 \times 21764$. I am working in python and have used lasso on it for feature selection.

For different values of the hyperparameter it sometimes selects more than $300$ features, is that possible for lasso? Or does sci-kit not use the standard lasso?

Answer 1

It should not be possible for lasso, or anything with L1 penalty. Check this thread for example: If p > n, the lasso selects at most n variables As far as I know, scikit learn doesn't use anything nonstandard with lasso, you might get better answer on a scikit learn mailing list/github apge. Maybe it's a bug.