Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session C: Friday, August 15, 2:00 – 5:00 pm, de Brug & E‑Hall

Which Neural Networks best model the human perception of geometric shapes?

Maxence Pajot¹, Théo Morfoisse², Mathias Sablé-Meyer³, Yair Lakretz⁴, Stanislas Dehaene; ¹CEA, ²Université Paris Cité, ³University College London, University of London, ⁴Ecole Normale Supérieure de Paris

Presenter: Maxence Pajot

While impressive in many vision tasks, artificial neural networks are limited dealing with geometric shapes. In this study, we systematically evaluate the ability of Convolutional Neural Networks (CNNs) and Vision Transformers - varying in sizes and training datasets- to recognize and process geometric shapes. We compare the models’ internal representations to human data collected from an outlier detection task involving quadrilaterals. We find that networks trained on large scale datasets, develop shape representations that closely resemble those of humans in some tasks.

Topic Area: Visual Processing & Computational Vision

Extended Abstract: Full Text PDF