. Author manuscript; available in PMC: 2022 Jun 22.

Published in final edited form as: Adv Neural Inf Process Syst. 2021 Dec;34:14056–14068.

Table 3:

Fairness results on CelebA, We applied CCA on three different layers in Resnet-18 respeetively, See appendix for positions of conv 0, 1, 2, “Ours-conv[0,1]-conv[1,2]” means staeking features from different layers to form hypercolumn features Hariharan et al, [2015], which shows that our approach allows two networks to have different shape/size.

	Accuracy(%)	DEO(%)	DDP(%)

Unconstrained	76.3	22.3	4.8
Ours-conv0	76.5	17.4	1.4
Ours-conv1	77.7	15.3	3.2
Ours-conv2	75.9	22.0	2.8
Ours-conv[0,1]-conv[1,2]	76.0	22.1	3.9