Technical Program

Paper Detail

Paper: PS-2B.27
Session: Poster Session 2B
Location: Symphony/Overture
Session Time: Friday, September 7, 19:30 - 21:30
Presentation Time:Friday, September 7, 19:30 - 21:30
Presentation: Poster
Publication: 2018 Conference on Cognitive Computational Neuroscience, 5-8 September 2018, Philadelphia, Pennsylvania
Paper Title: A mathematical model of real-world object shape predicts human perceptual judgments
Manuscript:  Click here to view manuscript
Authors: Caterina Magri, Andrew Marantan, L Mahadevan, Talia Konkle, Harvard University, United States
Abstract: How do we represent the shape of different real-world objects? Modern approaches to explore this question with deep neural networks are highly efficient but as of yet not clearly interpretable. In this paper, we examined the Normalized Contour Curvature model (NCC), which represents a shape as an interpretable histogram over curvature values (from very concave, to straight, to very convex). To explore the shape-space produced by this model, we submitted the feature profile of thousands of objects to a principal component analysis, revealing that 4 axes summarized the space. To compare this model to behavior we tested both perceived curvature (E1) and overall shape similarity (E2). Behavioral judgments of the perceived curvature of an object were well predicted by this model, largely isolated to the second PC loading. Behavioral measures of overall shape similarity were also predicted reasonably well using the four PCs, and approached the performance of deep neural networks (E3). The success of this model implies that perceptual shape-space can be summarized with a relatively small number of dimensions through an interpretable feature space, where the major axes are meaningfully related to perceived shape and curvature of inanimate objects.