Thèse Prédictions Temporelles Filtres Attentionnels Spectro-Temporels ou Codage Prédictif H/F

Université Paris-Saclay GS Life Sciences and Health

Établissement : Université Paris-Saclay GS Life Sciences and Health
École doctorale : Signalisations et Réseaux Intégratifs en Biologie
Laboratoire de recherche : Neuroimagerie Cognitive
Direction de la thèse : Sophie HERBST ORCID 0000000161613167
Début de la thèse : 2026-10-01
Date limite de candidature : 2026-04-15T23:59:59

Ce projet a pour objectif d'étudier les mécanismes neuronaux de la prédiction temporelle, en se basant sur le système auditif. L'enjeu est de résoudre le conflit théorique entre l'augmentation attentionnelle et la suppression prédictive des réponses sensorielles induites par l'anticipation temporelle. Pour ce faire, nous mettrons en oeuvre une approche de neuro-imagerie intégrative, combinant la résolution temporelle élevée de la magnétoencéphalographie (MEG) et les capacités de cartographie spatiale et par couche corticale de l'imagerie par résonance magnétique fonctionnelle (IRMf) à 7T. Cette méthodologie nous permettra de différencier l'effet d'amélioration attentionnelle de celui de la suppression prédictive et de vérifier si les signaux de prédiction temporelle s'inscrivent dans le schéma hiérarchique (top-down) et laminaire du codage prédictif.
En cartographiant les réseaux neuronaux mettant en oeuvre la prédiction temporelle, ce projet offrira une observation détaillée inédite de la façon dont le cerveau utilise les dynamiques temporelles des environnements sensoriels pour assurer un traitement efficace, permettant ainsi une meilleure intégration de la dimension temporelle dans notre compréhension de la fonction cérébrale globale.

In complex, dynamic sensory environments (e.g., a conversation or sports game), the brain must rapidly predict, anticipate, and select relevant sensory signals to meet behavioral goals. This requires representing both what will happen (sensory predictions) and when it will happen (temporal predictions). While sensory predictions have been widely studied, temporal predictions have received less consideration. This project addresses the urgent need to clarify how the brain represents and uses the temporal structure of sensory environments, which is crucial for understanding timing deficits observed in pathologies like Parkinson's or schizophrenia.
From a theoretical perspective, the question lies in framing timing within existing cognitive neuroscience models, specifically attention versus prediction. The attentional orienting framework suggests that perceptual benefits from timing arise from allocating a limited cognitive resource to moments in time, thereby enhancing perception [1]. This prioritization likely interacts with the priorization of sensory channels, forming a spectro-temporal filter [2]. However, a purely attentional framework typically fails to account for the learning of temporal contingencies and statistical structures inherent in dynamic environments. This learning aspect is captured by the predictive coding (PC) framework [3], which treats the brain as a Bayesian learner that adapts to minimize sensory processing energy. PC posits that prediction signals are conveyed top-down (e.g., from frontal areas), while prediction errors propagate bottom-up to refine predictions. PC offers a flexible, mathematically formalizable scheme for internalizing environmental statistics, but whether temporal predictions align with it remains unknown.
Predicting a sensory signal versus attending to it lead to contradicting theoretical assumptions: attention generally enhances sensory responses, while prediction typically reduces them [4].
Thus, a first step will be to clarify whether early sensory responses are enhanced or suppressed at temporally predicted moments. The existing literature on N100/M100 modulations is divided including my own work [5,6] likely because non-invasive imaging methods used in humans lack the spatio-temporal precision of invasive animal techniques.
To overcome this limit, we will examine the temporal dynamics of tonotopically defined auditory neural populations, leveraging the temporal precision of MEG and the spatial precision of high-resolution 7T fMRI. The direction of modulation (enhancement vs. suppression) in sensory evoked responses will be the primary tool to distinguish between predictive suppression and attentional enhancement.
A further critical feature of prediction signals is their cortical depth: true predictions occur in deep layers, while prediction errors or attentional modulation appear in superficial layers [7]. We will employ newly available layer-specific analyses of E/MEG data [8], informed by individual anatomical and functional 7T fMRI maps, to provide crucial insights into the propagation of temporal and sensory prediction signals.
Finally, in dynamic environments, prediction signals must be proactive - occurring before the predicted event - to encode its temporal unfolding. Such proactive signals are expected to be found in deep cortical layers. Beta band oscillations are considered the most likely candidates for these signals.
This project seeks to unify these opposing views by clarifying whether temporal predictions qualify as distinct prediction signals, or whether they modulate sensory prediction signals, akin to attention.

This project aims to provide an unprecedentedly detailed observation of how the brain uses the temporal structure of sensory environments to predict, anticipate, and filter relevant signals from noise. We will test if temporal predictions align with predictive coding principles, to distinguish local bottom-up prediction error signals as a key signature of statistical learning from top-down signals governing prediction and attentional orienting in time across brain regions. This project primarily targets fundamental research questions, but it also paves the way for more specific assessments of timing deficits in clinical conditions such as Parkinson's disease, Schizophrenia, and ADHD by providing clear neural indices of temporal statistical learning.

To achieve this, we will use an integrative neuroimaging approach combining 7T fMRI (for individual tonotopic maps and layer-specific cortical depth analysis) and M/EEG. Source reconstruction of the M/EEG data will allow investigation of neural dynamics from multiple distributed regions, including non-cortical structures like the cerebellum and the basal ganglia, at millisecond temporal resolution [9]. Temporally resolved decoding methods will be used to access the dynamics of these signals.

Throughout the project, the doctoral candidate will be tasked with putting in place the analyses pipelines for the high-resolution source reconstruction of MEG data. In a first step, this will be done using preliminary data that has already been recorded during an M2 internship of the candidate in the team. Then, a dedicated dataset (20 healthy adult human volunteers) will be collected at Neurospin, using both MEG and 7TfMRI. The PhD candidate will collect and analyse these data, with help from the PI, and a collaborator specialized in layer specific analyses of MEG data.