With Gaussian initialization scale σ_g=4 in the 2D image task, jointly optimizing a_j and b_j with the MLP using Adam does not improve performance relative to leaving the mapping fixed. · CiteArk
With Gaussian initialization scale σ_g=4 in the 2D image task, jointly optimizing a_j and b_j with the MLP using Adam does not improve performance relative to leaving the mapping fixed.
