notesum.ai

Published at October 22

FairLoRA: Unpacking Bias Mitigation in Vision Models with Fairness-Driven Low-Rank Adaptation

q-bio.QM

cs.AI

cs.LG

Released Date: October 22, 2024

Authors: Rohan Sukumaran¹, Aarash Feizi², Adriana Romero-Sorian², Golnoosh Farnadi³

Aff.: ¹DIRO - Université de Montréal, Mila - Quebec AI Institute, Quebec, Canada; ²McGill University, Mila - Quebec AI Institute, Quebec, Canada; ³McGill University, DIRO - Université de Montréal, Mila - Quebec AI Institute, Quebec, Canada

Arxiv: https://arxiv.org/abs/2410.17358v1

Model	Method	Accuracy (↑)	F1 Min (↑)	Recall Min (↑)	$\Delta$ F1 (↓)
CLiP	LoRA	97.35 ± 0.17	87.24 ± 0.25	83.74 ± 2.15	12.29 ± 0.16
	FairLoRA	97.58 ± 0.06	88.93 ± 0.48	86.51 ± 0.17	10.71 ± 0.48
	FFT	97.49 ± 0.19	88.26 ± 1.12	85.91 ± 0.47	11.24 ± 0.92
	FairFFT	97.57 ± 0.03	88.12 ± 0.42	85.64 ± 0.47	11.52 ± 0.42
DiNO	LoRA	94.38 ± 0.23	86.96 ± 0.91	82.54 ± 2.61	12.93 ± 0.73
	FairLoRA	94.53 ± 0.07	87.65 ± 0.83	83.88 ± 0.83	11.99 ± 0.83
	FFT	91.05 ± 0.84	83.08 ± 0.48	77.65 ± 2.00	14.77 ± 0.10
	FairFFT	91.63 ± 0.98	83.96 ± 1.00	78.39 ± 4.36	15.20 ± 0.59
ViT	LoRA	94.29 ± 0.07	86.76 ± 0.77	83.46 ± 1.67	12.84 ± 0.32
	FairLoRA	94.71 ± 0.08	87.09 ± 1.45	83.71 ± 2.14	12.81 ± 1.26
	FFT	94.39 ± 0.33	87.03 ± 0.32	83.45 ± 0.46	12.61 ± 0.04
	FairFFT	94.89 ± 0.27	87.44 ± 0.73	85.64 ± 0.94	12.05 ± 0.40