Description
I have been increasingly turning towards topics that sit at the intersection between computing, particularly artificial intelligence, and (the philosophy of) science. One such intersection is the automation of scientific reasoning and the epistemology of machine learning. Five papers deal with these topics. Votsis (2016) proposes an intuitive way to formalise the notion of ad hoc hypotheses for (among other things) the benefit of scientific automation. Votsis (2024a) argues that we need to augment existing models of analogical reasoning via a supplementary condition, namely the relevant uniformity criterion on concepts, with the aim of improving our understanding of good analogical reasoning and its computational implementation. Votsis (2024b) proposes a neuro-symbolic approach to automating scientific discovery, making use of both neural nets and automated theorem provers to identify interesting hypotheses. Votsis (2025) argues that there is considerable potential for methodological cross-pollination between conceptual change practices in 'classical' scientific representations and feature changes in machine learning representations. Norelli*, Votsis and Williamson (2026) make a case for the view that machine learning models in science establish bi-directional inferential relations between theory and data, contrary to the claims of Bogen and Woodard’s highly influential view.
Beyond this work, I have conducted research on automated theorem proving and related work on logical inference representation. This was the topic I pursued while visiting the Openproof Project, Center for the Study of Language and Information (which is subsumed under the Institute for Human-Centered AI), Stanford University during my last Sabbatical. I came up with a novel way of automating the construction of propositional logic proofs, whose outputs may potentially be easier to understand by humans than standard methods. My approach builds on the well-known resolution method, but avoids converting formulae into clausal form. I already implemented this idea in code (C#) - see video below where 146 arguments are correctly processed (109 valid and 37 invalid) in 44 seconds. I'm planning to release an app that allows users to employ this method to construct proofs. I foresee two papers coming out of this work: one that attempts to prove the soundness and completeness of the proposed system, and another that provides a graph-theoretic representation of its inferences.
My most recent work concerns an examination of Darwin's language in his correspondence, using natural language processing. In a contribution to a joint paper (currently under review), I employed word2vec and cluster analysis to investigate the role an addressee's profession (scientist vs. non-scientist) plays on Darwin's language. It turns out that moving across periods, there is generally stronger semantic similarity in the use of terms in letters addressed to non-scientists (vs. scientists), suggesting more terminological negotiation with scientists. It also turns out that, in letters addressed to scientsts, the most dramatic meaning shifts across those periods tended to be technical words like 'fossil'. This suggests that Darwin vexed over the precise meaning of such words, as they were crucial to his whole programme. The work on how Darwin's language evolves over time is ongoing and will be shared on this website at the earliest opportunity.
* Maria Federica Norelli is my PhD student (secondary supervision: Jon Williamson and Rogerio de Lemos).
Bibliography:
Norelli, M.F., Votsis, I. and Williamson, J. (2026) ‘Data, Phenomena, Models, and Theories’, Philosophy of Science, doi:10.1017/psa.2025.10161.
Votsis, I. (2016) ‘Ad hoc Hypotheses and the Monsters within’, in V. C. Müller (ed.), Fundamental Issues of Artificial
Intelligence (Synthese Library), Berlin: Springer, pp. 299-313.
Votsis, I. (2024a) ‘Modelling Analogical Reasoning: One-Size-Fits-All?’, Wittgenstein and AI (Volume I), Anthem Press.
Votsis, I. (2024b) ‘A Neuro-Symbolic Approach to the Logic of Scientific Discovery’ in E. Ippoliti, L. Magnani, and S. Arfini (eds.), Model-Based Reasoning, Abductive Cognition, Creativity, Inferences & Models in Science, Logic, Language, and Technology Series, Springer.
Votsis, I. (2025) ‘Concept and Feature Change in Scientific and Deep Neural Net Representations’, Cognitive Science Proceedings, vol. 47: 1654-1660.