we should say that this can also be naively considered we have a nice exercise in optimml, which shows that this is not necessarly convex. this has also sometimes in the old days discussed as an option for ANNs, and downsides have been discussed there