notesum.ai

Published at November 22

An Attention-based Framework for Fair Contrastive Learning

Released Date: November 22, 2024

Authors: Stefan K. Nielsen¹, Tan M. Nguyen²

Aff.: ¹FPT Software AI Center; ²Department of Mathematics, National University of Singapore

Arxiv: http://arxiv.org/abs/2411.14765v1

Refer to caption

Baseline Models
Model	Accuracy ( $\uparrow$ )	Bias Removal ( $\uparrow$ )
InfoNCE (Oord et al., 2018)	84.1 $\pm$ 1.8	$48.8\pm 4.5$
Fair-InfoNCE (Tsai et al. (2022))	$85.9\pm 0.4$	$64.9\pm 5.1$
CCLK (Tsai et al. (2022))	$86.4\pm 0.9$	$64.7\pm 3.9$
Attention-based Models
FARE (ours)	$85.7\pm 0.9$	$68.4\pm 4.3$
SparseFARE (ours)	$\textbf{86.4}\pm 1.3$	$\textbf{74.0}\pm 3.8$