The discriminability of feature representation is the key to open-set face
recognition. Previous methods rely on the learnable weights of the
classification layer that represent the identities. However, the evaluation
process learns no identity representation and drops the classifier from
training. This inconsistency could confuse the feature encoder in unde