dr_s comments on Deep learning models might be secretly (almost) linear