Semantic Technologies Committee

Abstract: AI has significantly improved the convenience of daily life, but its potential in scientific discovery remains underexplored. Across scientific domains, progress is often bottlenecked by data that are either too small and narrow to capture real-world diversity or too biased to convey the faithful context needed for grounded representation. Moreover, scientific predictions must be explainable and verifiable: without interpretability, models are hard to trust, diagnose, and use to generate actionable hypotheses. In this talk, I use biodiversity and ecology as a motivating domain to illustrate how data-centric and interpretable AI can help address these challenges. First, I will present how I scale and create data to train biodiversity foundation models that encapsulate scientifically meaningful representations. Next, I will discuss methods that make scientific AI more interpretable, enabling explanations and validation of learned representations for trustworthy scientific use. Finally, I will outline future directions for advancing data-driven scientific discovery of previously unknown correlations between the environment and scientific concepts.