Emily Van Hove¹, Hans Stuyck¹, Andrea Ivan Costantino¹, Kobe Desender¹, Hans Op de Beeck¹; ¹KU Leuven

Visual representations are often studied as stable constructs. Yet in a constantly changing visual world, the flexibility of these representations is essential to reorganize stored knowledge, and crucial to learning, creativity, problem-solving, and generalization. Nevertheless, the mechanisms underlying this dynamic reorganization remain poorly understood. In this project we study insight problem-solving, a cognitive process that requires restructuring, as a prime candidate of profound representational change in the visual system. Insight was induced through a newly developed visual version of the Remote Associates Task, previously mainly used in semantic literature, in which participants searched for a non-trivial association between three images. Each trail required integrating the representations of all images to find the solution. Behavioral representational spaces were assessed before and after the insight task, through a similarity sorting task. Using tools widely adopted in vision science, including representational geometry, RSA/PCA, and deep artificial neural network models, allowed us to understand the multidimensional similarity space and how it restructures as a consequence of insight. Results from behavioral experiments (n = 197) show a clear restructuring of behavioral similarity spaces succeeding the visual insight task. Initial category- and scene- based organization, shifts towards a geometry of insight-induced associations assessed through a multimodal encoding model. Nevertheless, the structure of the initial representational space remains partially preserved, suggesting that insight allows integration of new associations atop preexisting ones. These results provide a behavioral demonstration of flexibility in visual representations and a reference for preliminary neuroimaging results on the spatial dynamics of representational changes.

Acknowledgements: FWO