# FairViT: Fair Vision Transformer via Adaptive Masking

**Authors:** Bowei Tian, Ruijie Du, Yanning Shen  
**arXiv:** [2407.14799](https://arxiv.org/abs/2407.14799)  
**Date:** 2024-07-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2407.14799v1/

## Key points

- Authors: Bowei Tian, Ruijie Du, Yanning Shen
- Title: FairViT: Fair Vision Transformer via Adaptive Masking
- Enhances accuracy and fairness in ViT models for real-world deployment
- Introduces unique distance loss function
- Utilizes adaptive fairness-aware masks on attention layers
- Demonstrates superior accuracy compared to alternative methods
- Maintains competitive computational efficiency
- Achieves commendable levels of fairness, mitigating biases in computer vision applications
- Represents a significant advancement towards equitable and effective ViT models for real-world deployment
- Prioritizes both accuracy and fairness simultaneously, setting a new standard for ethical considerations in computer vision research.

## Summary

, , , , 

In their paper titled "FairViT: Fair Vision Transformer via Adaptive Masking," authors Bowei Tian, Ruijie Du, and Yanning Shen introduce a novel framework that addresses the growing importance of fairness in computer vision models. The framework, named FairViT, focuses on enhancing both accuracy and fairness in ViT models for real-world deployment. This is achieved through the introduction of a unique distance loss function and the utilization of adaptive fairness-aware masks on attention layers that update alongside model parameters. Through extensive experiments, FairViT demonstrates superior accuracy compared to alternative methods while maintaining competitive computational efficiency. Importantly, the framework also achieves commendable levels of fairness, highlighting its potential for mitigating biases in computer vision applications. Overall, FairViT represents a significant advancement towards developing more equitable and effective ViT models for real-world deployment by prioritizing both accuracy and fairness simultaneously. This sets a new standard for ethical considerations in cutting-edge computer vision research.