Medicine is inherently multimodal and demands an interpretation of diverse data sets, including text, imaging, and genomics. Current AI models in biomedical applications, although effective in their tasks, are usually unimodal and narrowly focused, which limits their practical utility. To address this challenge, the authors introduce Med-PaLM M, a generalist biomedical AI system that has the capability to flexibly interpret and integrate various types of multimodal biomedical data.
Med-PaLM M is a transformative AI system that surpasses specialist models in performance across a wide array of tasks. These tasks include medical question answering, image interpretation, radiology report generation, and genomic variant calling. Importantly, Med-PaLM M demonstrates emergent zero-shot medical reasoning, which means the system can generate reasonable solutions for new tasks without explicit prior training. This groundbreaking development is a significant milestone towards the creation of comprehensive AI solutions for the complex, multimodal demands of medicine.
Publication date: July 26, 2023
Project Page: Not Provided
Paper: https://arxiv.org/pdf/2307.14334.pdf
