Rethinking Importance Weighting for Deep Learning under Distribution Shift