Scientists from the USA conducted a study that demonstrated similarities in geometric perception between humans, rhesus macaques, and annubis baboons. The experiment involved preschoolers as well as adult members of the Himba and Cimane tribes. It was found that rhesus macaques and annubis baboons perceive geometric shapes approximately the same way as humans.
Participants were asked to select the geometric figure that most closely matched the sample. The CORnet neural network was used to analyze the data, simulating how the primate brain processes visual information. Using this model, researchers were able to calculate the significance of simple, complex, and symbolic representations in shape perception.
It was previously noted that humans and non-human primates are capable of geometric perception. However, one experiment with guinea baboons revealed a difference: humans were better at finding the single differing figure among identical quadrilaterals if the outlier possessed a greater regularity, such as symmetry, right angles, or parallel lines. This tendency was observed in preschoolers and uneducated adults from the Himba tribe but was absent in the baboons, who solved the task regardless of the figures' regularity.
The research team, led by Jessica Cantlon from Carnegie Mellon University, conducted a follow-up experiment with children aged three to six, adults from the Himba and Cimane tribes, as well as four rhesus macaques and four annubis baboons. Participants were shown a figure and then asked to choose a similar option from two. Analysis using the CORnet neural network and logistic regression allowed researchers to determine which features participants relied upon.
Two levels were used to assess figure similarity: the CORnet neural network, which models the visual cortex (primary visual cortex V1 for simple features like lines and spots, and inferior temporal cortex IT for whole object perception), and a symbolic model that accounts for specific characteristics such as side equality, angles, symmetry, and the presence of parallel sides.
The results showed that children, monkeys, and adults from the Cimane tribe primarily rely on complex and symbolic features. For monkeys, the weight of the IT layer was 0.76; for preschoolers, it was 2.95; and for adults, it was 3.77. The weight of the symbolic model was 0.62 for monkeys, 0.89 for children, and 0.51 for adults. Furthermore, adults from the Cimane tribe showed less dependence on symbolic representations than macaques and baboons, preferring to perceive shapes holistically.
When figures were rotated by 20 degrees, there was a change in primate behavior: the weight of the symbolic model increased in monkeys from 0.62 to 1.4, and in children from 0.89 to 2.32. Additionally, the orientation towards the whole form significantly increased in monkeys, as the weight of the IT layer rose from 0.76 to 3.85. This indicates the ability of non-human primates to form representations similar to symbolic ones.
The authors also found a correlation between the difference degree indicators calculated through the IT layer, independent of rotation, and the symbolic model (r = 0.71). This led to the hypothesis that human geometric thinking is not necessarily purely symbolic, but rather based on a holistic image that does not depend on the angle of rotation. According to the scientists, the differences between species lie in the better development of this general mechanism in humans.
The article concludes by mentioning a separate biological study conducted using an underwater drone, sonar, and video cameras on a mesophotic coral reef off the coast of Benin, which was previously considered dead. A living coral community was discovered, consisting mainly of soft octocorals, associated with isolated hard substrate areas, which aligns more with the concept of a coral garden than a reef.
Neurobiologists from Belgium and Germany have successfully created realistic animated models of macaques. These avatars attracted the same intense attention from live macaques as real monkeys shown in videos.
During the research, the investigators avoided using physical markers or manually labeling every frame. Instead, they utilized data collected from eight cameras. Labeling was performed only on select frames, after which a neural network was trained to predict the location of key points on the avatar's body, allowing for the reconstruction of its movement trajectory.
The study was published in the scientific journal PLoS Biology.
It was found that as the avatar's realism decreased, the macaques' attention to it also diminished. This observation indicates the manifestation of the so-called 'uncanny valley' effect. Previously, scientists had created animated facial avatars to study perception in primates, which allowed them to overcome this effect: macaques paid almost as much attention to realistic facial avatars as to live conspecifics, whereas avatars without fur were ignored.
Studying the perception of moving bodies in primates is complicated by several factors. Motion capture technology using physical markers is inapplicable because macaques tend to dislodge them during grooming. Furthermore, manual frame-by-frame labeling would require a colossal amount of time, involving the placement of dozens of markers across hundreds of frames just to create a short animation.
Martin Giese, the leader of the group from the University Hospital Tübingen and the Catholic University of Leuven, managed to bypass these obstacles by developing a moving avatar of the rhesus macaque (Macaca mulatta). For this, a system of eight cameras was set up around the macaque cage, recording 2.5 hours of video material. Short clips demonstrating the walking or turning of an individual, as well as a two-minute segment showing the grooming of two macaques, were selected from this material. Each camera operated at a frequency of 80 frames per second.
Instead of full manual labeling, a person marked the position of 42 joints, averaging two frames per second from two cameras. Then, the program projected these marked points onto the frames captured by the remaining six cameras. The Anipose algorithm used spatio-temporal triangulation to convert the resulting two-dimensional coordinates into three-dimensional ones. A neural network was responsible for predicting the position in frames that were not manually labeled. To ensure animation realism, the algorithm guaranteed the constancy of distances between joints and the smoothness of movements, preventing abrupt changes in coordinates between frames. This three-dimensional data was mapped onto a ready 3D model, resulting in the creation of a realistic animated avatar.
The correspondence between the avatar's movement trajectory and that of the real macaque from the video clips showed an error within 18–29 millimeters. Subsequently, the scientists demonstrated video recordings containing both real macaques and their avatars to eight male rhesus macaques. To assess the reaction specifically to the body, rather than the face, the faces in the videos and avatars were intentionally blurred. The macaques fixed their gaze on the silhouette of the real monkey and the avatar the same number of times if the avatar possessed realism, including fur and textures. However, attention was drawn less to the wire model of the monkey. Conversely, the macaques paid twice as little attention to the gray avatar, which lacked color, fur, and textures. The authors concluded that such an avatar causes discomfort in macaques due to the 'uncanny valley' effect.
The creation of the first animated macaque body avatar allowed the authors to demonstrate that the mechanisms for processing different social signals—both facial and bodily—in rhesus macaques likely operate on similar principles, as the 'uncanny valley' effect is observed in both cases. In the future, such a controlled avatar could become a tool for studying primate reactions to the moving bodies of their conspecifics. Nevertheless, the researchers noted that tracking the movements of two macaques simultaneously using eight cameras proved less accurate due to frequent point overlaps.
