From our experiments so far, about 90% of the training time is for generating the models in R. This will be faster in the next Zorro update that can run several R sessions in parallel, using multiple CPU cores.

Using covariance for eliminating features is a good idea.