april 29, 2021
Man's attraction to geometric shapes appears to be as old as mankind itself - zigzag engravings have been found dating back over 500,000 years! Several studies suggest that humans share a high level of understanding of the abstract properties of geometric shapes. A collaboration of researchers from CEA-Joliot (NeuroSpin department), Collège de France, CNRS and Université Paris 8 has devised an empirical test for the validity of a stronger hypothesis, using the simplest possible shapes: that this affinity for geometry is a human exception and does not exist in other primates. This result is also an interesting challenge for object recognition neural networks, which do not capture human behavior.
In their work, recently published in the journal PNAS(Proceedings of the National Academy of Sciences of America), Mathias Sablé-Meyer and Prof. Stanislas Dehaene of the Collège de France designed a set of 11 quadrilaterals of varying regularity (number of right angles, parallel sides, symmetries). For each of these shapes, they constructed four alternative versions by applying an identical transformation, then used these shapes in an intruder detection task. On each trial, six shapes were presented: five identical to within one rotation or dilation, and one different.
605 French adults, who took part in an online experiment shared on Twitter, demonstrated a geometric regularity effect: they responded more quickly and made fewer mistakes as the regularity of the shapes increased. 28 kindergarten pupils, as well as 156 first-grade pupils, took part in the same study, which took place on a tablet at school. In these children, the team was able to demonstrate the presence of the geometric regularity effect. The same effect was found in 22 Himbas adults who did not attend school. The Himbas, a pastoral people from northern Namibia with a limited vocabulary for geometry, were able to take part in this study thanks to the collaboration of Serge Caparos.
In parallel, with the collaboration of Joël Fagot (Laboratoire de Psychologie Cognitive, CNRS/Aix-Marseille Université), the team trained 26 baboons to detect an intruder among images. These baboons succeeded in learning the task, but even the 11 baboons persevering enough to achieve performances comparable to those of kindergarten pupils showed no geometric regularity effect. For example, they did not learn any faster to recognize the square among its deviants than any other shape among its deviants. Even after more than 8,000 trials, they were wrong one time out of two, whatever the geometric shape presented.
To model the baboons' behavior, the study's authors used convolution neural networks as models of the brain's object perception mechanisms. This modeling correctly predicts the behavior of all baboons, but the model needs to be enriched with symbolic information (number of right angles, parallel sides, etc.) to explain human behavior.
These results reveal a new signature of human singularity, more elementary than language or mathematical ability, and present a new challenge for non-symbolic models of human perception.
