Скачать или смотреть Veridical Data Science and Alignment in Medical AI

Bin Yu (UC Berkeley)
https://simons.berkeley.edu/talks/bin...
Alignment, Trust, Watermarking, and Copyright Issues in LLMs

Alignment and trust are crucial for the successful integration of AI in healthcare, a field involving diverse stakeholders such as medical personnel, patients, administrators, public health officials, and taxpayers, all of whom influence how these concepts are defined. This talk presents a series of collaborative medical case studies where AI algorithms progressively become, from transparency to more opaque thus with increasing difficulty of alignment assessment. These range from tree-based methods for trauma diagnosis, to LLM-based emergency department co-pilot, and mechanistic circuits for structured data extraction from pathology reports. They are guided by Veridical Data Science (VDS) principles—Predictability, Computability, and Stability (PCS)—for the goal of building trust and interpretability, enabling doctors to assess alignment. The talk concludes with a discussion on applying VDS to medical foundation models and next steps for evaluating AI algorithm alignment in healthcare.