Skip to contentSkip to filters

We use cookies for analytics.

Artificial Intelligence for Science

Sašo Džeroski

Jozef Stefan Institute

Abstract

Sašo Džeroski surveys the scientific challenges that must be met for AI to transform science: explainable models and predictions, learning effectively from limited data, integrating data with prior domain knowledge, and supporting open, reproducible science by formalizing and sharing scientific knowledge. He presents explainable machine-learning methods, including multi-target prediction and relational learning, that yield accurate yet interpretable models and have been applied in environmental, life and materials science. For data-scarce problems he compares semi-supervised learning, which uses unlabeled data alongside labeled data, with foundation models whose representations learned from vast unlabeled data support downstream tasks with minimal supervision. He then describes automated scientific modelling that learns interpretable models of dynamical systems, such as process-based models and differential equations, from time series and domain knowledge, and closes with ontologies and semantic technologies for representing and annotating data, algorithms, models and experimental results in machine learning and optimization.

Topics