Andrea Ivan Costantino¹ (), Esna Mualla Gunay¹, Emily Van Hove¹, Laura Van Hove¹, Felipe Fontana Vieira¹, Merim Bilalic², Hans Op de Beeck¹; ¹KU Leuven, ²Northumbria University

Exploring the neural and behavioral correlates of expertise offers a window to understand how cognitive and neural representations may change due to domain-specific training. Previous studies suggest that experts undergo a representational re-organization compared to non-experts, resulting in more linearly separable representations, particularly for task-relevant, high-level dimensions. Chess, with its rich history as a metaphor for human intelligence and strategic thought, serves as an ideal domain for probing expertise effects. Studies on chess expertise suggest expert players analyze board setups differently from novices, emphasizing piece relationships over visual traits. However, prior studies did not explore representational structure and information processing changes in expertise, and in what brain areas these changes may occur. Our work bridges this gap by employing computational, behavioural, and neuroimaging methodologies to uncover representational changes in expert biological and artificial systems. In this study, 40 participants (20 chess experts and 20 novices) performed a chess-related task during fMRI scanning. We applied multivariate decoding techniques (MVPA, RSA) to analyze representational changes in human brain activity and artificial deep neural networks (DNNs). We hypothesized that (i) experts would exhibit distinct neural patterns associated with chess-relevant features in high-level brain regions, reflecting their advanced understanding, and (ii) there would be alignment in representational and behavioral patterns between expert biological and artificial systems. Our findings reveal a striking similarity in information processing between humans and DNNs, highlighting representational and behavioral alignment among expert systems. Additionally, experts—both biological and artificial—show a representational re-organization, resulting in more linearly separable representations for task-relevant high-level features at later processing stages.