Rethinking Importance Weighting for Deep Learning under Distribution Shift T ongtong Fang 1 Nan Lu1,2 Gang Niu