Abstract:
Increasingly capable AI is likely to be widely transformative within the next decade. Healthcare may be the area where this impact is most profound, as foundation models have the potential to increase quality and access to care and accelerate biomedical scientific discovery. Existing models have limitations that prevent their uptake in real-world workflows; limited reliability, interactivity, and alignment with human values are all crucial problems in this high-stakes setting.
I will discuss three recent works from my team that aim to measure and mitigate these limitations: Med-PaLM, Med-PaLM 2, and Med-PaLM M. Med-PaLM (published in Nature) was the first AI system to surpass the passing grade on US Medical License Exam (USMLE) style questions. Med-PaLM generates accurate, helpful long-form answers to consumer health questions, as judged by panels of physicians and users. Med-PaLM 2 further improves performance and reliability, achieving an expert-level score on exam questions. Its answers are preferred over physician answers across several clinically-relevant axes. Med-PaLM M is a multimodal foundation model that can perform a wide variety of biomedical tasks at or near state-of-the-art performance.
I will discuss the implications of these advances and outline a path towards improved health for billions, including open challenges along the way.
Bio:
Karan Singhal is a Staff Research Engineer at Google Research leading teams working on biomedical AI, foundation models, and representation learning. Karan's recent work includes Med-PaLM, a series of medical large language models featured in The Scientific American, Wall Street Journal, The Economist, and others. Karan is particularly motivated by advancing AI safety in the medical setting to create more reliable, steerable systems that could improve the health of billions. Prior to Med-PaLM, Karan researched robust and private representation learning, deploying novel algorithms to hundreds of millions of users. Karan's work has been published in Nature, NeurIPS, ICLR, and other venues. He received his M.S. and B.S. in Computer Science from Stanford University, where he initiated and was the main instructor for the "AI for Social Good" course.
14 окт 2024
