FAIR_XAI framework reveals bias in multimodal models for wellbeing assessment

Researchers have developed FAIR_XAI, a framework to improve the fairness of multimodal foundation models used in wellbeing assessment. The study evaluated Phi3.5-Vision and Qwen2-VL on datasets like E-DAIC and AFAR-BSFT, finding performance variations and demographic biases, with Qwen2-VL showing gender disparities and Phi-3.5-Vision exhibiting racial bias. While explainability interventions showed mixed results, sometimes improving procedural consistency but not guaranteeing equitable outcomes, the work emphasizes the need to jointly optimize accuracy, demographic parity, and generalization. AI

IMPACT Highlights the challenges in achieving equitable outcomes with multimodal models in sensitive applications like wellbeing assessment.

RANK_REASON This is a research paper detailing a new framework and its evaluation on existing models.

Read on arXiv cs.LG →

• AFAR-BSFT
• arXiv
• E-DAIC
• FAIR_XAI
• Phi3.5-Vision
• Qwen2-VL

• paper
• safety

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