From Prediction to Understanding: A Philosophical Framework for Machine Learning Research

Machine learning has emerged as a dominant paradigm for extracting knowledge from data across scientific disciplines, from protein folding prediction to climate modeling and natural language processing. Yet despite its remarkable empirical successes, the epistemic foundations of ML research remain surprisingly under-examined. This paper argues that addressing fundamental questions about what ML models can know, how they can explain, and when they can be trusted requires explicit engagement with philosophical traditions examining the nature of prediction, explanation, discovery, understanding, and decision-making. Drawing on recent work at the intersection of philosophy of science and computational methods, and aligning with the 2026 conference theme "Philosophical Unification of Diverse Forms of Inquiry," I propose a unified framework that situates ML research within a broader set of epistemic aims than predictive accuracy alone. The framework distinguishes five distinct epistemic functions—Prediction, Explanation, Discovery, Understanding, and Decision-making (the P.E.D.U.D. framework)—and examines how different ML methodologies, from deep learning to causal inference to interpretable modeling, serve these aims in complementary or conflicting ways. I argue that contemporary ML research operates predominantly within what can be called a "culture of prediction," wherein predictive performance on held-out benchmarks has become the primary, and often exclusive, metric of success. This culture has obscured other valuable epistemic functions, leading to an incomplete and potentially misleading vision of what ML can contribute to science and society. The dissociation between predictive accuracy and explanatory depth—exemplified by black-box neural networks that predict superbly yet provide little understanding—poses a central philosophical challenge: what is the value of prediction without understanding, and when is predictive success sufficient for practical decision-making? The paper concludes by outlining concrete directions for integrating philosophical reflection into ML research practice, including reforms to problem formulation, evaluation metrics, model reporting standards, and training curricula. Such integration, I contend, is not merely supplementary to technical ML research but is essential for addressing persistent challenges related to generalization under distribution shift, model opacity, fairness, and trustworthy deployment in high-stakes domains.