Feature scaling is applied is that gradient descent converges much faster with feature scaling than without it. It's also important to apply feature scaling if regularization is used as part of the loss function. so that coefficients are penalized appropriately.