Mitigating bias in training on biased datasets is an important open problem.
Several techniques have been proposed, however the typical evaluation regime is
very limited, considering very narrow data conditions. For instance, the effect
of target class imbalance and stereotyping is und