Thèse Mécanismes Cérébraux de la Formation de Consensus dans les Réseaux Sociaux H/F - Doctorat.Gouv.Fr

Établissement : Université Claude Bernard Lyon 1 École doctorale : NSCo - Neurosciences et Cognition Laboratoire de recherche : ISC-MJ - INSTITUT DES SCIENCES COGNITIVES MARC JEANNEROD Direction de la thèse : Jean-claude DREHER ORCID 0000000221571529 Début de la thèse : 2026-10-01 Date limite de candidature : 2026-05-31T23:59:59 Ce projet de thèse vise à identifier les mécanismes computationnels et neuronaux qui sous-tendent la formation de consensus dans les réseaux sociaux. Dans un contexte où les interactions humaines sont largement médiées par des plateformes numériques, comprendre comment les individus intègrent des informations privées et sociales pour converger vers des décisions collectives constitue un enjeu central en neurosciences computationnelles, économie comportementale et sciences des réseaux.
Deux cadres théoriques dominent l'étude de l'apprentissage social : les modèles de DeGroot, fondés sur une moyenne pondérée des opinions des voisins, et les modèles d'apprentissage par renforcement, reposant sur des mises à jour séquentielles guidées par des erreurs de prédiction. Si ces modèles décrivent efficacement la propagation d'information dans les réseaux, leurs implémentations cérébrales restent peu comprises dans des contextes d'interactions sociales en temps réel.
Le projet propose de combler cette lacune en combinant modélisation computationnelle, manipulation expérimentale de réseaux sociaux et hyperscanning en fNIRS, permettant l'enregistrement simultané de l'activité cérébrale de plusieurs participants dans des contextes écologiquement valides. L'objectif est d'examiner comment les dynamiques d'interaction interindividuelles se traduisent en synchronisation neuronale inter-cerveaux (INS), notamment entre le cortex préfrontal dorsolatéral (DLPFC) et la jonction temporo-pariétale (TPJ), régions impliquées dans le contrôle cognitif et la cognition sociale.
Sur le plan méthodologique, des groupes de 6 participants (n=480) seront engagés dans une tâche de décision dans un contexte incertain, implémentée en temps réel dans des réseaux dont la topologie est systématiquement manipulée. Chaque essai comprendra une phase d'information privée (signal probabiliste bruité), une décision initiale avec estimation de sa confiance, une phase d'interaction sociale où les décisions des voisins sont révélées de manière temporellement contrôlée, puis une décision révisée. Des incitations monétaires permettront de moduler le poids relatif de l'exactitude individuelle et de la coordination collective.
Les données comportementales seront analysées à l'aide de modèles computationnels (DeGroot pondéré, apprentissage par renforcement), ajustés par maximum de vraisemblance ou approches bayésiennes hiérarchiques, et comparés via des critères d'information (AIC, BIC). Les données fNIRS seront prétraitées (filtrage, correction des artefacts, normalisation) puis analysées en termes de cohérence temporelle inter-cerveaux, à la fois au niveau global et en fonction de la structure du réseau. Des approches basées sur les modèles permettront de relier paramètres computationnels et dynamiques neuronales. Des analyses statistiques multi-niveaux testeront les liens entre synchronisation neuronale, poids social attribué aux voisins et émergence du consensus.
La thèse se déroulera sur trois ans : développement du protocole, pré-registration et études pilotes (année 1), collecte des données et analyses intermédiaires (année 2), analyses avancées et rédaction (année 3). Il s'appuie sur des collaborations internationales et des infrastructures d'hyperscanning existantes, garantissant sa faisabilité.
Ce travail apportera une contribution majeure à la compréhension mécanistique de la cognition collective, en reliant apprentissage social, structure des réseaux et synchronie entre cerveaux (n>2). Il présente également des implications importantes pour l'étude de la diffusion de l'information, de l'influence sociale et de la coordination dans les sociétés contemporaines.
Understanding how individuals coordinate their beliefs and actions within social networks is a central challenge across computational neuroscience, behavioral economics, and network science. In modern societies, where large-scale interactions are mediated by digital platforms, collective phenomena such as consensus formation, polarization, and misinformation diffusion emerge from repeated local interactions structured by network topology. These processes are inherently dynamic and distributed, requiring an integrative approach that bridges individual cognition and group-level outcomes.
Two major computational frameworks have been proposed to account for belief updating in social environments. The DeGroot model describes social learning as a process of iterative averaging of neighbors' opinions, potentially weighted by structural features such as degree centrality (DeGroot, 1974; Jackson, 2008). In contrast, reinforcement learning models describe belief updating as a sequential, prediction-error-driven process that occurs asynchronously, either because individuals receive information at different times or due to cognitive constraints on aggregating information from multiple sources (Jiang et al., 2023; Sutton & Barto, 2018). As a result, a large body of work on gossip algorithms has examined consensus formation under asynchronous communication. Recent work from our group suggests that DeGroot-like mechanisms provide a robust account of consensus formation in networked environments, particularly when individuals have access to information about network structure (Benistant et al., 2025). However, these computational models account for belief updating at the behavioral level, leaving unresolved how such processes are implemented in the brain during real-time interaction.
At the brain level, an emerging body of research emphasizes the importance of inter-brain dynamics in social interaction. Hyperscanning studies have demonstrated that coordinated behavior is associated with inter-brain neural synchrony (INS), particularly in frontal and temporoparietal regions implicated in cognitive control and social cognition (Cui et al., 2012; Babiloni & Astolfi, 2014; Hasson et al., 2012). Functional Near-Infrared Spectroscopy (fNIRS) is especially well suited for this line of research, as it enables simultaneous recording of multiple participants in ecologically valid settings. Despite these advances, no study has combined formal computational modeling, controlled network interactions, and hyperscanning neuroimaging. This project addresses this gap by developing a unified framework to study how individual-level computations give rise to coordinated group behavior and shared neural dynamics. This project contributes to an emerging international research effort integrating computational modeling, network science, and social neuroscience. By linking formal models of belief updating with neural dynamics measured during real-time interaction, it should provide a mechanistic account of collective cognition that is directly relevant to global challenges related to information transmission and social influence.
The general objective of this project is to identify the computational and neural mechanisms underlying coordination and consensus formation in social networks. More specifically, the project aims to determine how individuals integrate private and social information under coordination constraints, and how these processes are reflected in inter-brain neural dynamics. Based on previous analyses of 2 large behavioral datasets in social networks and simulations from our group, we hypothesize that belief updating in networked environments will be better explained by Weighted DeGroot models (Benistant et al., 2025). At the neural level, it is expected that inter-brain synchrony in the dorsolateral prefrontal cortex (DLPFC) will be associated with individual error processing (Jiang et al., 2023), whereas synchrony in the temporoparietal junction (TPJ) will reflect the integration of socially weighted information. More importantly, it is predicted that the strength of inter-brain synchrony will be associated with the emergence of group-level coordination and consensus. The Phd project will employ a real-time networked coordination task adapted for fNIRS hyperscanning. This paradigm allows simultaneous investigation of behavioral dynamics, computational mechanisms, and inter-brain neural coupling within controlled social networks. Participants will be tested in groups of six and embedded in predefined network structures (total n=480) (based on Zhang et al., 2023). On each trial, they will perform a binary decision task under uncertainty, requiring them to infer the true state of the environment based on both private signals and social information obtained from their neighbors. Each trial will begin with a private information phase, during which participants receive a noisy signal about the true state (e.g., a probabilistic cue). This ensures heterogeneity in initial beliefs and prevents trivial convergence. Participants will then make an initial decision, accompanied by a confidence estimate, providing a baseline measure of individual belief. In the subsequent social interaction phase, participants will observe the decisions of their direct neighbors within the network. The structure of the network will be visible and remain constant within each block, allowing participants to take into account the relative position and connectivity of others. Social information will be presented in a temporally controlled manner, enabling precise alignment with neural recordings. Finally, participants will make a revised decision, reflecting the integration of private and social information. This updated response constitutes the primary behavioral measure of interest.
A key feature of the task is the introduction of coordination incentives. Participants' payoffs will depend not only on the accuracy of their decisions but also on the extent to which their choices align with those of others. By manipulating the relative importance of these components across experimental blocks, the design allows investigation of how individuals balance truth-seeking and coordination pressures, a central aspect of real-world social decision-making. Network topology will be systematically manipulated by varying the distribution of degree centrality across nodes. This enables testing whether individuals assign greater weight to highly connected neighbors, as predicted by Weighted DeGroot models. Neural activity will be recorded simultaneously from all participants using fNIRS hyperscanning, with optodes placed over frontal and temporoparietal regions. The primary neural measure will be inter-brain neural synchrony, quantified using time-resolved coherence methods. Synchrony will be analyzed both globally and in relation to network connections, allowing investigation of how neural alignment reflects social structure. Behavioral data will be analyzed using computational modeling, comparing DeGroot and reinforcement learning frameworks. Neural data will be analyzed using both event-related and model-based approaches, linking computational parameters to brain activity and inter-brain synchrony. Particular attention will be paid to the relationship between neural synchrony and group-level coordination outcomes.
The project is fully feasible within a three-year PhD framework. fNIRS hyperscanning provides a practical and well-established method for studying group interactions, and the experimental paradigm builds on validated designs from previous research from the group of Dr Yina Ma (Zhang et al., 2023). Collaboration with her lab will allow data collection which requires large sample sizes, and the ability to do multiple fNIRS (6 participants simultaneously). In parallel, we are planning to acquire multiple fNIRS at the ISCMJ to perform related group decision making/collective intelligence experiments. The first year will be dedicated to ethics writing and approval, literature review, methodological training, task development, and pilot testing. The second year will focus on data collection and initial analyses, including computational modeling and preliminary neural analyses. The third year will be devoted to advanced analyses, integration of results, and thesis writing.

Le candidat devrait être familier avec au moins une technique de neuroimagerie (EEG, fNIRS, IRMf) et sera à l'aise dans les approches de modélisation du comportement, comme l'apprentissage par renforcement ou les modèles de propagation d'information dans les réseaux. Un profil de master en neurosciences computationnelle ou ingénieur est attendu.