Known Bugs and Solutions
Bug (unsolved): mexTrainDL_Memory does not perform as well as mexTrainDL when param.lambda is very small.
mexTrainDL uses LARS to solve l1-decomposition problems. mexTrainDL_Memory uses coordinate descent.
When the regularization parameter is very small, coordinate descent can be very slow and so mexTrainDL_Memory becomes unusable.
This is a particular case where mexTrainDL should be used instead.
Bug (unsolved): When the norm of my input vectors is very big or very small, mexLasso
does not return the correct solution.
This is a numerical stability problem that will be solved in the next release. In the meantime, you should rescale your data.
Bug (unsolved): the path returned by mexLasso for small values of lambda is sometimes incorrect.
The regularization path of the Lasso is sometimes hard to follow because of numerical instabilities. This happens in pathological cases.