Intel Labs Mitigates AI Bias in Foundational Multimodal Models by 20 Percent
Intel Labs Mitigates AI Bias in Foundational Multimodal Models by 20 Percent
4 PagesIntel Labs reduced AI bias in foundational multimodal models by up to 20% using a novel approach based on social counterfactuals. Their Cognitive AI team built a large dataset of synthetic images that vary by race, gender, and physical traits across 260 occupations. Trained using Intel® Gaudi® 2 AI accelerators and 3rd Gen Intel® Xeon® processors, the models were analyzed using Retrieval-Augmented Generation (RAG) and filtered for quality and safety. The open-source dataset and findings aim to minimize bias in AI outputs and improve fairness across AI applications.
