演講資訊

摘要：

Pre-trained representations from large volumes of unlabelled data, known as foundation models, are rapidly emerging as a key AI technology in scientific research. I will review some of the recent applications of these models in astrophysics, highlight their advantages, and some of the potential challenges. I will also describe our recent work looking at recovering calibrated uncertainties from such models and how we can use related metrics to understand the diversity of astrophysical sources in our data, identify varying types of distributional shift - and preserve the potential for discovery. Data is central to both the development and exploitation of large AI models, and I will conclude with a future outlook that focuses on how data usage and access may evolve as AI tools become increasingly mainstream in science.