Duke AI Health Roundup Explores Machine Comprehension and Algorithmic Bias
The latest Duke AI Health Friday Roundup highlights crucial developments in natural language processing, AI image generation, and the historical roots of racial bias in photography. The summary also points to significant public health updates regarding anxiety screening and pulse oximeter accuracy.
The latest Duke AI Health Friday Roundup highlights significant developments in artificial intelligence and data science, focusing heavily on natural language understanding. A new arXiv preprint by Choudhury and colleagues investigates whether deep learning models truly comprehend text or simply score well on benchmarks for the wrong reasons by examining coreference resolution and comparison skills. Additionally, the roundup notes that the AI image generator DALL-E now allows users to edit photographs of real people, enabling changes to specific features like clothing and hairstyles.
The roundup also delves into the critical issue of algorithmic bias, tracing the history of racial prejudice in photography. A field review by Nettrice R. Gaskins explains how mid-twentieth-century color film chemistry inherently erased the facial features of Black individuals, a problematic legacy that continues to influence modern image classifiers and facial recognition technology. This historical context provides vital background for understanding current disparities in how AI systems process and interpret human faces.
Beyond artificial intelligence, the summary covers important updates in clinical research and public health policy. The FDA closely examines pulse oximeter performance discrepancies across different skin tones, while the US Preventive Services Task Force (USPSTF) issues a new recommendation for widespread anxiety screening. The roundup also mentions emerging research into wearable sensors designed to measure tumor regression, showcasing the continued integration of innovative technology into patient care and clinical monitoring.