Representation Learning WIthout Representationalism: Towards a Pragmatist Philosophy of AI in Science

Applications of artificial intelligence (AI) for modelling in science put pressure on the dominant representationalist conception of scientific models. Representationalism holds that models must represent their targets to be epistemically useful. This thesis aims to reorient the discussion around how scientists use AI models as epistemic tools in practice without assuming a prior theory of model-based science in terms of representation. In Chapter 2, I disambiguated two conceptions of representation in philosophy of science and AI. I argue that these are epistemically distinct and that conflating the two creates difficulties in assessing the epistemic features of AI models in science. I then outline an artifactualist account of AI models in science that I develop further throughout the rest of the thesis: successful applications of AI models in science depend on skilled engagement with domain-general, computational tools to make a model adequate for more specific purposes. In Chapter 3, I examine the case of AlphaFold2 in structural biology as an example of the use of formal analogies in the transfer of model building techniques across disparate domains. Chapter 4 examines the evidence for neural scaling laws, which appears to affirm the idea that scaling up can lead to quantitative and qualitative improvements in model performance, displacing the need for human judgement concerning the design of model architectures. I argue that, in science, scaling up the size of a model cannot replace principled model design. Although neural scaling laws do play a role in guiding model design choices, attention to how this unfolds in practice reveals that modellers must nevertheless make theoretically informed choices about model architectures. Chapter 5 defends a pragmatist account of scientific understanding with AI models that is not reducible to or dependent on explanation. I argue that such pragmatic understanding comprises a set of implicit methodological principles that underlie model design-choices and secure their reliability. Past successes provide the grounds for the articulation of methods. These methods establish design-principles for future model-building. In Chapter 6, I develop an account of goal-directed neural network models in cognitive neuroscience. I demonstrate that the cognitive sciences turn out to be a notable exception in which scientists do in fact use AI models representationally.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2025

Author dc:creator

• Kieval, Phillip

Advisors dc:contributor.advisor

• Halina, Marta
• Chang, Hasok

• AI in science
• models in science
• philosophy of AI
• representation
• scientific understanding

Licence

• https://www.repository.cam.ac.uk/bitstreams/b7665811-3bc4-48e9-95f3-098d1f003559/download
• http://purl.org/NET/rdflicense/allrightsreserved

Language dc:language

eng

DOI dc:identifier.doi

https://doi.org/10.17863/CAM.123311

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/392666