Examining and Combating Spurious Features under Distribution Shift

               Chunting Zhou 1 Xuezhe Ma 2 Paul Michel 1 Graham Neubig 1
                                           (a) ERM          (b) Group DRO [clean]
                                      (average=93.3% robust=83.4%)  (average=96.5% robust=95.8%)
              Abstract
    A central goal of machine learning is to learn
    robust representations that capture the causal rela-
    tionship between inputs features and output l ...