Biology subjects

Proix, T.

Publications and source records attributed to Proix, T..

9 recordsLinked to original sources

bioRxiv2026

Cortico-subcortical multi-head self-attention as a substrate for cognitive performance

The neocortex is central to mammalian cognition, yet a computational framework that is both biologically constrained and capable of performing complex cognitive tasks remains missing. Here we show that cortico-thalamic circuits are well suited to implement multi-head self- and cross-attention, the mechanism underlying the cognitive abilities of transformer networks. We propose that layer 2/3 pyramidal cells maintain a recurrent key-value memory, while layer 5 pyramidal cells decode the memory retrieved by an incoming query. The computation of keys, values and queries maps onto core and matrix thalamo-cortical projections, distributed across the micro- and macro-columns of a cortical area. One cortical area forms an attention head, and cortex a multi-head self-attention network. The same thalamo-cortical microcircuit also calculates sensory prediction errors guiding gradient-based synaptic plasticity. A reward-prediction error gates via basal ganglia the cortical output and the re-activation of hippocampal memories. The trained network aligns with human intracranial recordings during speech perception. Overall, the suggested cortico-subcortical attention circuit may represent a substrate for the cognitive capacity of mammals.

neuroscience↗

bioRxiv2026

Cortico-hippocampal dynamics of hierarchical syntactic planning in natural speech production

The human brain must rapidly construct hierarchical structures to organize complex sequential behavior, yet the neural dynamics supporting this process during natural behavior remain poorly understood. Spoken language provides a powerful model system for investigating this computation, requiring rapid transformation of conceptual intent into structured sequential output. Using rare intracranial stereo-electroencephalography (SEEG) recordings from patients producing extended spontaneous speech, we examined how syntactic planning unfolds over time using measures of constituency, dependency structure, and probabilistic syntactic categories. We identified a hierarchical planning architecture in which global sentence structure and core syntactic categories (nouns and verbs) were specified before more local planning operations. Neural representations of these categories emerged up to 1 s before articulation and persisted throughout the planning period, whereas optional modifiers, including adjectives and adverbs, were recruited only closer to speech onset. These observations support a model of hierarchical incremental planning in which abstract sentence structure precedes the incremental specification of individual sentence elements. While core syntactic categories engaged a broader fronto-temporo-parietal network than other word classes, syntactic-depth-related activity emerged in parallel across cortical regions and the hippocampus, suggesting that hippocampal relational representations contribute to sentence structure building. Together, these findings support a cortico-hippocampal model of speech production in which hierarchical sentence structure and core syntactic categories are planned before secondary syntactic elements are incrementally incorporated into the evolving sentence plan. These results provide a neural account of how abstract linguistic structure is transformed into fluent speech.

neuroscience↗

bioRxiv2026

Conjunctive coding of letter pairs emerges in segregated human medial temporal lobe neurons during working memory

Composing individual memory traces into unified representations is fundamental to encoding of structured relationships and flexible cognition. A central debate in neuroscience concerns the neural mechanisms of these compositions: are these compositions encoded through mixed selectivity, where the same neurons simultaneously represent individual and composed items, or through conjunctive coding within a distinct and specialized neuronal population? Here, we tested these competing hypotheses by recording 996 single units from the medial temporal lobe of 11 epilepsy patients performing a letter-string working memory task. We found neural encoding of both single letters and letter pairs during task related periods, with the latter peaking during the maintenance phase. Crucially, letter pairs were encoded by a sparse subset of neurons that did not show significant selectivity for their constituent letters. This suggests the existence of a segregated population of neurons recruited to compose items independently of single letter neural encoding. These findings provide evidence that the human medial temporal lobe separately encodes compositions within segregated neurons, offering direct neurophysiological support for conjunctive coding at the single-neuron level.

neuroscience↗

bioRxiv2025

Directed cortico-limbic dialogue in the human brain

How can one trace the brains orderly directed signals amid a tangle of nerve fibers? Because direct access to actual brain signaling is rare in humans, the precise wiring diagrams for cortico-limbic communication during sleep and wake remain essentially unmapped, hampering progress in neuroscience. Now, a unique neurosurgical window on the human brain allows for electrically mapping cortical connections at the hospital, but studies so far have relied on average signals, masking the dynamic nature of signal flow across brain regions. To causally estimate signaling dynamics, we repeatedly probed cortico-limbic networks with short-lived electrical pulses over days and assessed the variable fate of each transmitted signal on a single-trial basis. In the resulting openly available dataset, we characterized signaling probabilities and directionality across thousands local and long-range cortico-limbic connections over days. Challenging established views, we found that limbic structures send twice as many signals as they receive, in both wakefulness and sleep. Our findings provide a fundamental framework for causally interpreting signal flow in the brain and formulating therapeutic strategies for brain networks disorders.

neuroscience↗

bioRxiv2024

From spiking neuronal networks to interpretable dynamics: a diffusion-approximation framework

Modeling and interpreting the complex recurrent dynamics of neuronal spiking activity is essential to understanding how networks implement behavior and cognition. Nonlinear Hawkes process models can capture a large range of spiking dynamics, but remain difficult to interpret, due to their discontinuous and stochastic nature. To address this challenge, we introduce a novel framework based on a piecewise deterministic Markov process representation of the nonlinear Hawkes process (NH-PDMP) followed by a diffusion approximation. We analytically derive stability conditions and dynamical properties of the obtained diffusion processes for single-neuron and network models. We established the accuracy of the diffusion approximation framework by comparing it with exact continuous-time simulations of the original neuronal NH-PDMP models. Our framework offers an analytical and geometric account of the neuronal dynamics repertoire captured by nonlinear Hawkes process models, both for the canonical responses of single-neurons and neuronal-network dynamics, such as winner-take-all and traveling wave phenomena. Applied to human and nonhuman primate recordings of neuronal spiking activity during speech processing and motor tasks, respectively, our approach revealed that task features can be retrieved from the dynamical landscape of the fitted models. The combination of NH-PDMP representations and diffusion approximations thus provides a novel dynamical analysis framework to reveal single-neuron and neuronal-population dynamics directly from models fitted to spiking data.

neuroscience↗

bioRxiv2023

Low-dimensional neuronal population dynamics in anterior superior temporal gyrus reactivate phonetic representations during semantic processing

AbstractSpeech processing involves a hierarchy of cognitive layers from low-level phonetic to high-level seman-tic representations. However, interactions are not only feed-forward; Feedback mechanisms are crucial for real-time, accurate speech recognition. Yet how distant levels interface during speech processing remains unclear. Here, we analyzed intracortical recordings from 624 neurons across three human participants implanted with microelectrode arrays in the anterior superior temporal gyrus during an auditory semantic categorization task and natural speech perception. We identified distinct neural subspaces, or manifolds, for semantic and phonetic features, with a functional separation of the corresponding low-dimensional dynamics. We contrasted a bottom-up cumulative and a top-down predictive hypothesis on phonetic-semantic temporal alignment, and found phonetic alignment to word-level semantic representations, signaling top-down prediction. These effects were consistent across participants at the spiking level, and remained undetected in adjacent ECoG recordings. These findings demonstrate the reorganization of neuronal population dynamics supporting phonetic representations during semantic prediction.

neuroscience↗

bioRxiv2022

Interpreting dynamics of neural activity after dimensionality reduction

Dimensionality reduction methods are frequently used to analyze high-dimensional activity of cortical neuron populations during behavior. The resulting oscillatory trajectories that consistently emerge from this analysis have been interpreted as a signature of latent dynamical systems. Here, we show that these oscillatory trajectories necessarily result from applying dimensionality reduction methods on recordings that approximately exhibit continuous variation in time, regardless of whether or not the recorded system incorporates latent dynamics.

neuroscience↗

bioRxiv2021

Imagined speech can be decoded from low- and cross-frequency features in perceptual space

Reconstructing intended speech from neural activity using brain-computer interfaces (BCIs) holds great promises for people with severe speech production deficits. While decoding overt speech has progressed, decoding imagined speech have met limited success, mainly because the associated neural signals are weak and variable hence difficult to decode by learning algorithms. Using three electrocorticography datasets totalizing 1444 electrodes from 13 patients who performed overt and imagined speech production tasks, and based on recent theories of speech neural processing, we extracted consistent and specific neural features usable for future BCIs, and assessed their performance to discriminate speech items in articulatory, phonetic, vocalic, and semantic representation spaces. While high-frequency activity provided the best signal for overt speech, both low- and higher-frequency power and local cross-frequency contributed to successful imagined speech decoding, in particular in phonetic and vocalic, i.e. perceptual, spaces. These findings demonstrate that low-frequency power and cross-frequency dynamics contain key information for imagined speech decoding, and that exploring perceptual spaces offers a promising avenue for future imagined speech BCIs.

neuroscience↗

bioRxiv2019

Intracortical neural activity distal to seizure-onset-areas predicts human focal seizures

The apparent unpredictability of epileptic seizures has a major impact in the quality of life of people with pharmacologically resistant seizures. Here, we present initial results and a proof-of-concept of how focal seizures can be predicted early in advance based on intracortical signals recorded from small neocortical patches away from identified seizure onset areas. We show that machine learning algorithms can discriminate between interictal and preictal periods based on multiunit activity (i.e. thresholded action potential counts) and multi-frequency band local field potentials recorded via 4 X 4 mm2 microelectrode arrays. Microelectrode arrays were implanted in 5 patients undergoing neuromonitoring for resective surgery. Post-implant analysis revealed arrays were outside the seizure onset areas. Preictal periods were defined as the 1-hour period leading to a seizure. A 5-minute gap between the preictal period and the putative seizure onset was enforced to account for potential errors in the determination of actual seizure onset times. We used extreme gradient boosting and long short-term memory networks for prediction. Prediction accuracy based on the area under the receiver operating characteristic curves reached 90% for at least one feature type in each patient. Importantly, successful prediction could be achieved based exclusively on multiunit activity. This result indicates that preictal activity in the recorded neocortical patches involved not only subthreshold postsynaptic potentials, perhaps driven by the distal seizure onset areas, but also neuronal spiking in distal recurrent neocortical networks. Beyond the commonly identified seizure onset areas, our findings point to the engagement of large-scale neuronal networks in the neural dynamics building up toward a seizure. Our initial results obtained on currently available human intracortical recordings warrant new studies on larger datasets, and open new perspectives for seizure prediction and control by emphasizing the contribution of multiscale neural signals in large-scale neuronal networks.

neuroscience↗