Artificial Intelligence for Studying Communication in Wild Animals – WildCom-AI

The WildCom-AI methodology is based on a coupling of computer vision and experimental psychology in the service of behavioral ecology. The objective is to utilize AI not only as a stimulus generator but also as a computational model of the primate visual system.

1. Stimulus Synthesis and Latent Space Manipulation. To isolate signaling traits (particularly sex), we employed state-of-the-art generative models such as StyleGAN3. Unlike previous versions, StyleGAN3 ensures spatial continuity and the absence of 'aliasing' or 'artifacts' during manipulations. We projected real images of mandrills into the latent space. The method then consisted of identifying latent directions (vectors) corresponding to key biological variables. By moving images along these axes, we generated 'edited' stimuli where only a single target trait varies, allowing for the perfect isolation of the signal while maintaining the subject's identity and ecological realism.

2. Modeling Perception and Classification (CNN Approach). Moving beyond simple classification, we used neural networks as proxies for the primate visual system. The work of Tieo et al. relies on CNN (Convolutional Neural Network) architectures to model the perception of social signals. These models were trained to categorize traits from raw images. Analyzing the activation maps of these networks allows for the identification of critical morphological zones (e.g., nasal ridges, colored areas) that carry social information. In a specific study regarding senescence (Renoult et al., 2025), Vision Transformers (ViT) were deployed for their superior predictive capacity. However, this use remained strictly predictive (technical performance) and was not used to model biological perception mechanisms, unlike the CNNs.

3. AI and Information Processing. The project also explores the computational foundations of visual preferences. Following the approach developed in Dibot et al. (2023), we analyzed the response of the internal layers of deep neural networks. The method involves calculating the sparsity (parsimony) of neural activations during exposure to a signal. We tested whether signals considered 'attractive' or 'efficient' induce more parsimonious activation patterns (fewer neurons activated, but more intensely). This approach links the mathematical complexity of image processing to the adaptive value of the social signal.

4. Experimental and Behavioral Validation. Finally, the validity of the AI models was tested against biological reality through behavioral trials. The synthetic stimuli produced by StyleGAN3 were presented to primates in a controlled environment. Eye-fixation duration was measured to quantify social interest or attractiveness. The final method consisted of correlating the observed behavioral preferences with AI-derived metrics (CNN classification scores and parsimony indices).

The WildCom-AI project has led to significant breakthroughs, both in evolutionary theory and in the technological tools provided to the scientific community.

1. The WildCom Database

One of the most tangible pillars of the project is the creation of a world-first photographic database. We have collected and standardized a dataset of 100,000 photographs of mandrills monitored in their natural habitat. This database was the subject of a dedicated Data Paper and is now available through Open Access. We believe it represents a major resource for future research in visual ecology and computer vision applied to wildlife.

2. Sparsity and Beauty

The research conducted by Dibot et al. (2023) has successfully bridged the gap between computational neuroscience and sexual selection. We demonstrated a robust statistical link between the sparsity of neural activations in deep neural networks and visual preference. The more "efficiently" a signal is processed by the visual system, the more attractive it is perceived to be. In a comparative study (Tieo et al., 2024), we also showed that sparsity is a better predictor of beauty than statistical typicality (the degree to which a face is "average" or representative of the population). This result suggests that signal evolution tends toward optimizing information-processing efficiency rather than simple conformity to a mean.

3. Social Signaling in Mandrills

The article published in iScience (Sonia Tieo et al. 2023) shed new light on the perception of femininity/masculinity signals in non-human primates. In contrast to human models, where certain forms of facial femininity are often associated with social or sexual benefits, our results show that in mandrills, the link between morphology and social benefits favors highly masculine individuals. The study demonstrates that facial "masculinity" is an honest signal of rank and reproductive success, shifting social preferences toward exaggerated traits—the opposite of trends observed in many studies on human perception.

4. Experimental Validation via Generative AI

The project also validated an innovative methodology for stimuli creation using StyleGAN3, enabling a transition from observation to experimentation. We published a method for generating synthetic stimuli to experimentally test the correlations observed in the field. Behavioral tests (looking-time assays) confirmed that primates process these synthetic images as they would real conspecifics. While these tests did not show massive preference differences regarding 'more masculine' traits, we demonstrated a crucial technical contribution: the manipulation of facial orientation.

The Multidimensional Nature of Visual Attractiveness

While our research has established the sparsity of neural activations as a robust predictor of preference, this parameter does not exhaust the complexity of the concept of "beauty" or biological attractiveness. Sparsity primarily models coding efficiency, but it does not account for the deeper semantic or hedonic value of signals. A signal can be processed efficiently by the visual system without necessarily triggering a positive behavioral response (e.g., a threat signal may be highly parsimonious yet unattractive). It is therefore necessary to develop models that integrate other dimensions of perception to complement the efficiency metric.

Modeling Selective Attention

The use of CNNs has allowed us to understand the global processing of traits, but these models struggle to simulate the dynamic prioritization of information by the gaze. The transition to Vision Transformers (ViT) represents a key bottleneck to overcome. Unlike CNNs, ViT attention mechanisms allow for the mathematical modeling of how an individual focuses cognitive resources on specific facial details (e.g., the brightness of a color or the texture of a nasal ridge). The objective will be to integrate attention maps to predict which areas of the stimulus capture the gaze first during behavioral tests.

Quantifying Perceived Complexity

Another identified bottleneck concerns the balance between efficiency (sparsity) and visual complexity. In cognitive psychology, it is well known that attractiveness often follows an inverted U-curve relative to complexity: a stimulus that is too simple is ignored, while one that is too complex is rejected. Future work must leverage AI to quantify the perceived complexity of communication signals in mandrills. Modeling should distinguish between "structural complexity" (morphology) and "processing complexity" (cognitive cost). We will need to develop hybrid metrics combining image entropy and processing depth within neural networks.

Stimulus Dynamism and "Liveliness"

The transition from static to dynamic stimuli remains a major technical challenge. Although generative AI via StyleGAN3 has enabled the manipulation of facial orientation to make portraits appear more "alive," full control over facial micro-gestures (lip movements, eye blinks) remains difficult to automate without losing biological signal integrity. Exploring Video GANs or temporal diffusion models will be crucial for modeling the perception of signals in motion—an essential dimension of real-world social communication.

In progress

Studying the brain’s processing of information is necessary to better understand how communication signals influence behaviours. However, the methods classically used in brain studies can be applied only to a limited number of model species. Artificial intelligence (AI) offers novel perspectives to model these neural processes and thus to investigate them outside laboratories. In this project, we will use AI to explore the links between information processing in the brain and complex behaviours expressed by non-human primates in natural environments. We will study the influence of face perception on socio-sexual behaviours in mandrills (Mandrillus sphinx), a primate from Central Africa, relying on a long-term research program on a wild population of more than 300 mandrills launched in Gabon in 2012.
We will first study the links between facial resemblance and kin recognition. We will model the perception of facial resemblance using AI applied to our database of 19,000 pictures portraying individuals of the studied population. We will explore the mechanisms allowing mandrills to evaluate their resemblance to other individuals of their social group, the link between parent-offspring facial resemblance and parental investment, how resemblance among adults is used to limit inbreeding, and to what extent fortuitous facial resemblance generates behavioural biases towards unrelated individuals, as a by-product of kin selection.
Then, we will study how the efficiency of information processing in the brain influences socio-sexual behaviours. The cognitive sciences have recently revealed that a communication signal that is processed efficiently is evaluated positively and triggers attractive behaviours in receivers. These results come from studies carried out on humans in laboratories, but their implications for other species in natural environments remain unknown. Here, for the first time in the cognitive sciences, we will use AI to quantify the efficiency of information processing. We will study the processing of facial information in mandrills, testing the hypothesis that a face that is processed efficiently is visually more attractive, and thus increases reproductive success, social integration and parental care.
Our scientific hypotheses will be tested with the wild population of mandrills using not only correlation analyses, but also using experiments with mandrills from a large captive population. For the first time in behavioural science, we will leverage state-of-the-art generative algorithms of AI to synthesise complex communication signals (i.e. mandrill faces) while controlling their variation. We will generate artificial portraits modifying resemblance to another face and controlling the efficiency of their processing. Then, we will analyse how resemblance and processing efficiency influence behaviours by displaying the modified portraits to captive mandrills in large-scale bioassays.
This project aims to explore new applications of AI (i.e. modelling the processing of information in the brain and generating stimuli) for ecology and evolutionary biology. While the vast majority of studies at the interface between AI and ecology aim at characterising the diversity of organisms and their interactions, this project will shed new light on the processes determining these interactions.