Facial expression recognition models show significant bias, study finds

A new study published on arXiv examines bias in facial expression recognition (FER) datasets and models, finding that all four common datasets analyzed exhibit significant demographic bias, particularly concerning race. The research also evaluated seven deep learning models, revealing that while transformer-based models like ViT and CLIP often achieve high accuracy, they also display the most bias. Conversely, residual-based CNN architectures such as ResNet and XceptionNet showed lower bias levels. The findings underscore that high accuracy does not equate to fairness and that addressing bias requires a joint approach to both datasets and models. AI

IMPACT Highlights the critical need to address demographic bias in AI datasets and models to ensure fairness, even at the expense of peak accuracy.

RANK_REASON The cluster is based on an academic paper detailing research findings on bias in AI models and datasets. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• AffectNet
• CEPrompt
• ExpW
• Fer2013
• MobileNet
• Mohammad Mehdi Hosseini
• POSTER
• RAF-DB
• ResNet
• ViT
• XceptionNet

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →