Our mission is to enable machine learning algorithms to take a central role in all aspects of scientific discovery and to understand how such a transformation will impact the scientific approach as a whole. In particular, we aim to drive new developments in machine learning, by identifying and solving overarching problems that are common to many scientific disciplines; to advance scientific application domains by creating a sustainable, transforming impact on science through machine learning; and to investigate long-range implications of the envisioned transformation of science through machine learning by studying possible consequences on the general scientific approach using methods from philosophy of science and research ethics.
We identified four research areas in which progress is urgently needed to advance the use of machine learning in science and fully realize its potential to open new perspectives for science:
Beyond prediction, towards understanding: Many machine learning algorithms are highly successful at prediction, but do not support scientists in achieving a deeper understanding of underlying scientific principles. We will design algorithms that support the scientific discovery processes by inferring complex scientific structures and causality, automatic model criticism, and integrating scientific theories and machine learning.
Managing uncertainty: Machine learning needs to improve the methods by which uncertainty can be quantified and handled in large and complex scientific models. This includes techniques to separate uncertainty about data, predictions, parameters and models. We will therefore develop tools to estimate and handle the uncertainty in complex scientific models and algorithms and exploit this information for experimental design.
Interface between algorithms and scientists: To make the use of machine learning in the scientific process more successful, we need to develop new techniques that allow scientists from any discipline to judge, control and interact more easily with all the stages of the machine learning process. We will seek to develop interpretable machine learning algorithms and tools that help us to interact with them and study their inherent biases throughout the whole scientific discovery pipeline.
Philosophy and ethics of machine learning in science: The fact that machine learning algorithms will play a central role in the process of scientific discovery challenges our traditional understanding of the scientific process and raises fundamental questions about concepts of scientific discovery and the role of the scientists. We will study these questions from the perspective of philosophy and ethics of science.