The goal of this project is to better understand deep learning by drawing on insights from philosophy of science. Deep learning algorithms have emerged in recent years at the forefront of machine learning. They are a new, powerful technology with many applications in both science and everyday life, ranging from data analysis in particle physics to mastering the game of GO. While deep learning algorithms are successfully applied in many areas, they are not yet well understood: There is, first, a lack of "theory of deep learning", which means that many mathematical properties of these models are not known; second, there is a lack of "interpretability", which means that humans do not understand how these models achieve their goals. Computer scientists are well aware of both lacunae and have called for more research to fill them. These problems have even been acknowledged in the political sphere: The EU’s "General Data Protection Regulation" postulates a "right to explanation" for automatic decision-making. But it is far from clear what this means.
To contribute to a better understanding of deep learning, this project carries out interdisciplinary research at the borderline between computer science and philosophy. Specifically, we analyze theoretical work on understanding in deep learning, such as the Information Bottleneck method, and we connect them to the philosophical debates on (scientific and mathematical) explanation and understanding. We then use this work to clarify the ongoing debate on interpretability and deep learning theory in computer science. Thus, by moving back and forth between computer science and philosophy, we establish a firm connection between deep learning and philosophy of science, and we thereby contribute to a better understanding of this important new technology in society.