Filipp Schmidt^1,2 (), Emily A-Izzeddin¹, Henning Tiedemann¹, Christian Houborg¹, Roland W. Fleming^1,2; ¹Justus Liebig University Giessen, ²Center for Mind, Brain and Behavior, Universities of Marburg, Giessen, and Darmstadt

Visually classifying objects into superordinate classes, such as “plant” or "animal," presents a significant computational challenge as radically different items, like "jellyfish" and "flamingo," must be grouped together. Here, we probed observers’ visual intuitions about the key features that unite members of superordinate classes by asking them to generate novel items. Sixteen participants drew new (i.e., original, unfamiliar) members of nine specified classes (animal, building, clothing, furniture, household appliance, musical instrument, plant, tool, and vehicle) as well as for a general “object” class. Another 16 participants assigned each drawing to one of the nine class labels (or suggested a new class), yielding average accuracy of 69% (compared to 10% chance and 89% for familiar control drawings). Virtually all drawings were grouped into one of the nine specified classes, including most created for the general “object” class. Performance was particularly high for unfamiliar animals, plants, buildings and vehicles. They also rated the typicality of each drawing on a 10-point scale, yielding a mean typicality of 4.9 for unfamiliar drawings, vs 8.2 for familiar control drawings. Our results suggest people can generate novel, unfamiliar drawings that capture key features of superordinate classes. To identify the visual features driving these classifications, we asked another group of 35 participants to classify the drawings, as well as mark and label the drawings’ defining “parts”. The most frequent labels were consistent with signature features of each class (e.g. “leg” and “eye” for animals), while both the overlap of labels between classes—and consistency of participants’ labels—was related to classification performance. Together, our results suggest that observers learn key features shared by highly diverse members of superordinate classes, and can ‘remix’ these to create new examples—a process that can be uniquely probed through analysis of creative drawings.

Acknowledgements: Supported by German Research Foundation (222641018–SFB/TRR 135 TP C1); the European Research Council (ERC) Advanced Grant ‘‘STUFF’’ (ERC-2022-AdG-101098225); and Research Cluster ‘‘The Adaptive Mind’’ funded by the Hessian Ministry for Higher Education, Research, Science and the Arts.