how do we increase linearity when most layers are applied with nonlinear activations?