Discover and Mitigate Unknown Biases with Debiasing Alternate Networks

Deep image classifiers have been found to learn biases from datasets. Tomitigate the biases, most previous methods require labels of protectedattributes (e.g., age, skin tone) as full-supervision, which has twolimitations: 1) it is infeasible when the labels are unavailable; 2) they areincapable of mitigating unknown biases -- biases that humans do notpreconceive. To resolve those problems, we propose Debiasing Alternate Networks(DebiAN), which comprises two networks -- a Discoverer and a Classifier. Bytraining in an alternate manner, the discoverer tries to find multiple unknownbiases of the classifier without any annotations of biases, and the classifieraims at unlearning the biases identified by the discoverer. While previousworks evaluate debiasing results in terms of a single bias, we createMulti-Color MNIST dataset to better benchmark mitigation of multiple biases ina multi-bias setting, which not only reveals the problems in previous methodsbut also demonstrates the advantage of DebiAN in identifying and mitigatingmultiple biases simultaneously. We further conduct extensive experiments onreal-world datasets, showing that the discoverer in DebiAN can identify unknownbiases that may be hard to be found by humans. Regarding debiasing, DebiANachieves strong bias mitigation performance.