Improving self-training under distribution shifts via anchored confidence with theoretical guarantees

Taejong Joo, Diego Klabjan

Self-training often falls short under distribution shifts due to an increased discrepancy between prediction confidence and actual accuracy. This typically necessitates computation…