Strong Novelty Regained: High-Impact Outcomes of Machine Learning for Science
Published in Synthese, forthcoming, 2025
A general class of presupposition arguments holds that the background knowledge and theory required to design, develop, and interpret a machine learning (ML) system imply a strong upper limit to ML’s impact on science. I consider two proposals for how to assess the scientific impact of ML predictions, and I argue that while these accounts prioritize conceptual change, the presuppositions they take to be disqualifying for strong novelty are too restrictive. I characterize a general form of their arguments I call the Concept-free Design Argument: that strong novelty is curtailed by utilizing prior conceptualizations of target phenomena in model design. However, I argue that if ML design choices (such as ground-truth labels for supervised ML and inductive biases) are based on prior conceptualizations of phenomena, it need not impede conceptual change. Furthermore, while their accounts focus narrowly on conceptual change, a variety of learning outcomes also contribute to strong scientific change. Thus, I present a variety of types of strong novelty from philosophy of creativity, epistemology, and philosophy of science that paint a more varied picture of how ML advances science. One of these is a form of local theory-independent learning from data that signals an aim to substantially revise existing theory, but it is not easily undermined by prior assumptions about target phenomena. Furthermore, generating surprise, reducing utility blindness, and eliminating deep ignorance also indicate high impact to scientific knowledge or research direction. I illustrate these types of strong novelty with several cases of scientific discovery with algorithms. My taxonomy clarifies several desiderata for machine-based exploration and should inform choices in designing for scientific change.