Data-induced multiscale losses and efficient multirate gradient descent schemes