Skip to main content
. Author manuscript; available in PMC: 2021 Jan 17.
Published in final edited form as: Comput Vis ECCV. 2020 Oct 7;12357:365–381. doi: 10.1007/978-3-030-58610-2_22

Fig.2: Quantitative Results on CelebA.

Fig.2:

The target attribute is the label attractiveness present in the CelebA dataset and the protected attribute is gender. (left) FairALM has a stable training profile in comparison to naive ℓ2 penalty. (right) FairALM attains a lower DEO measure and improves the test set errors (ERR).