ci Mν Bζ(Ni) 1Ni PNi ni= 1`i fθ(T)(xni),yni

Still,webelievetheyarerelevant.3 Models invariant to gender, e.g., can be used for fair decision-making and, in a more data analysis context, to seek4 patterns maskedbygenderoridentify,through thedualvariables, thefeatures (phenotypes) moresensitivetogender.5 We illustrate such analyses in Section D of the appendices. We will include these discussions in the camera-ready. Nevertheless, they affect only theparametrization error 0,17 which is independent of the sample size (Def. Hence, 0 is an intrinsic limitation of the parametrized learning18 problem and sample sizes can always be estimated relative to this lower error bound.