Minds and Machines Special Issue: Machine Learning: Prediction Without Explanation?

Abstract

design drugs to turn proteins off (or on)? How can we design proteins to perform new functions? Hence, there is a sense in which AlphaFold2's remarkable prediction comes without an explanation.This is an important sense, to be sure: AlphaFold2 does not explain how protein folding works. It seems to have somehow learned to bypass the step of explicitly modeling the biological mechanisms leading to the folded protein. Or maybe, there is an image of this mechanism somehow contained in the activation patterns of the nonlinear functions making up AlphaFold2. But this brings us to yet another sense in which the prediction comes without an explanation: AlphaFold2's own functioning in many ways needs an explanation itself.The implications of this want of explanation in the face of successful predictions are far reaching. For instance, the trustworthiness of ML algorithms in society is a big issue, and it depends, for the most part, on the ability to explain their functioning. Ethical issues like these generally profit from philosophers' inputs, as evidenced by the numerous projects and niches on the ethics and societal impact of Artificial Intelligence that we can see originate today. 1 However, for the sake of using ML in science, the other need for explanationthat concerned with an understanding of the mechanisms whose outcome is so successfully predicted by the ML algorithm-certainly obtains a special relevance as well. For, assuming that it remains difficult to understand and explain ML predictions, but that the scientific use of these methods keeps increasing over time, the question arises whether this changes the aims of science from explanation to 'mere' prediction.These and further issues were first explored, by the editors of this issue, in a workshop organized by P. Grünke and R. Hillerbrand at the Karlsruhe Institute of Technology. This was done as part of a project called The impact of computer simulations and machine learning on the epistemic status of LHC Data, in which F. J. Boge is also involved as a postdoctoral researcher. Said project, in turn, is part of an interdisciplinary research unit between physics, philosophy, history and social science, called The Epistemology of the Large Hadron Collider and co-funded by the German Research foundation (DFG) and the Austrian Science Fund (FWF).Much in the spirit of the research unit, the resulting workshop was an interdisciplinary effort, as it involved, next to philosophers, also scholars from the earth sciences (see Boge & Poznic, 2021). Given the fruitfulness of this workshop, the present Special Issue was created as a follow-up publication, even though the contributions to both largely differ.The essays collected in this Special Issue represent a broad spectrum of perspectives on the issue of explanation in the context of ML, as used in science and beyond. Below, we offer a brief summary of their core theses for the reader's orientation.1 Two of us (P. Grünke & R. Hillerbrand) have, for instance, as members of the AI Ethics Impact Group (AIEIG), particip...

Harnessing Machine Learning for Prediction of Postoperative Pulmonary Complications: Retrospective Cohort Design

“…Additionally, the retrospective nature of the study introduces potential sources of bias and limitations in data collection. Finally, ML models predict outcomes; however, they may not provide causal explanations or elucidate the underlying mechanisms of the predicted outcomes [36]. Although our study identified the risk factors associated with PPCs, it may not definitively establish causality or provide deep insights into the biological or physiological pathways involved when considering the present performance of our models.…”

The Explanatory Role of Machine Learning in Molecular Biology

“…Philosophers of science and scientists have taken up the central question of this debate by addressing the impact that the introduction of AI and machine learning is likely to have on the general character of scientific research (e.g. Pietsch, 2015;Canali, 2016;Coveney et al, 2016;Boon, 2020;Creel, 2020;Ourmazd, 2020;Boge and Poznic, 2021;López-Rubio and Ratti, 2021;Boge et al, 2022;Krenn et al, 2022;Duede, 2023;Andrews, 2023). The areas of genetics and molecular biology, which over the last few decades have become highly "data-centric" (Leonelli, 2016), seem particular prone to making the shift from a theory-or hypothesis-driven mode towards purely data-driven modes of research.…”

Explainable Artificial Intelligence in Hydrology: Interpreting Black-Box Snowmelt-Driven Streamflow Predictions in an Arid Andean Basin of North-Central Chile

“…That is, the general idea that indicates the existence of an inverse relationship between the interpretability of black-box models and the degree of precision reached by their predictions, although this relationship is not as simple as has been argued [28][29][30]. Recent studies on the challenge posed by the prediction/interpretability trade-off [28,29,31] and its explicit incorporation into hydrological modeling show that this is a hot topic and therefore transcendental in the discussion on the role that AI/ML/DL has in current hydrology and will continue to have in an increasingly digitized world [32][33][34].…”