Publications

Neural activity unfolds not as a set of independent signals, but as a coordinated dynamical process that can be described as a point moving through a high-dimensional state space. This perspective has contributed substantially to recent developments in systems neuroscience, especially through studies of directly recorded neuronal population activity, but remains comparatively underused in non-invasive human recordings such as EEG and MEG. In this primer, we present a neural trajectory analysis based on Principal Component Analysis (PCA) as an accessible route into extending the concepts of neural trajectories in state space to human electrophysiology. We explain how EEG/MEG activity can be organized into neural state representations, projected onto axes organized by decreasing covariance, truncated into low-dimensional representations, and visualized as trajectories that capture how distributed activity evolves over time. We show how trajectory geometry can be used to compare conditions and relate neural dynamics to behavior or clinical outcomes. An example of EEG motor execution and imagery, accompanied by a tutorial notebook, illustrates the workflow from conventional EEG summaries to trajectory-based analysis. We also clarify the limits of PCA, including its linear, variance-driven nature, and discuss EEG/MEG-specific challenges such as spatial mixing, preprocessing sensitivity, validation, and overinterpretation. Finally, we situate PCA within a broader family of state-space methods. By making neural trajectories practical and interpretable, this primer offers a guide for bringing the conceptual framework of population-level dynamics into mainstream human cognitive neuroscience.

Individuals adopt different encoding strategies to facilitate learning, yet few studies have examined the neurophysiological basis of these strategies across individuals. The present work addresses this gap by extending our previous findings on the direct relationship between cortical spectral power, measured via resting-state magnetoencephalography, and standard cognitive performance, to test whether resting-state neural features predict individual differences in encoding strategy preferences.

Our results highlight the complex interactions between endogenous brain oscillations, learning, and verbal encoding strategies assessed by the California Verbal Learning Test-Second Edition (CVLT-2). First, resting-state theta oscillations were significantly associated with verbal learning and subjective clustering strategies. Second, semantic clustering was facilitated by oscillatory patterns in the left sensory-motor regions.

Finally, serial and semantic clustering strategies showed opposite regression patterns, indicating a competitive interaction. Together, these findings provide insights into resting-state neural markers associated with diverse encoding strategies in verbal learning.

How neural activity across the ventral visual hierarchy supports face recognition is an open question. A long-standing debate asks whether face processing, particularly in fusiform cortex, relies on face-specific computations or representations shared with broader visual recognition. Here we combine source-resolved magnetoencephalography (MEG) with task-optimized neural networks as controlled computational models of visual experience. Rather than manipulating long-term expertise in human observers, we systematically vary learning objective on the model side—what the networks are trained to recognize—while holding architecture and loss function constant within model comparisons. We then ask which learned representational geometries best align with neural responses, where and when. We measured millisecond-resolved brain-model alignment across V1, lateral occipital cortex (LOC) and fusiform cortex while participants viewed familiar, unfamiliar and scrambled faces. The same stimuli were presented to seven CNN architectures trained for face-identity recognition (FR), object-category recognition (OR) or object categorization including a face category (Dual), alongside untrained controls. Familiarity produced a stage-dependent dissociation: in LOC, familiar faces showed earlier brain-model alignment than unfamiliar faces in the M170 range, an effect most consistent in models trained for face-recognition, whereas broader objectives produced more variable, architecture-dependent peak alignment latencies. In fusiform cortex they showed stronger alignment around the M200 range. This fusiform advantage was not uniquely associated with face-recognition training: dual- and object-trained models showed greater fusiform correspondence than face-trained models. Together, these findings support stage-dependent specialization, with training for face-identity recognition constraining intermediate-stage timing while later fusiform representations remain compatible with representational structure acquired through broader visual computations.

The ambition of the 2025 PNPL competition (Landau et al., 2025) was to launch a multi-year curriculum for non-invasive speech decoding. Designed to progress from foundational tasks toward the linguistic complexity required for a practical brain-computer interface (BCI), it set the stage with speech detection and phoneme classification tasks. Winning submissions reached F1-macro scores of 95.6% and 73.6% on the respective tasks (Elvers et al., 2026), highly significant advances. This success was built on the LibriBrain dataset (Özdogan et al., 2025), the largest within-subject MEG dataset recorded at the time with ~50 hours of data for one subject. However, while within-subject scale drives strong decoding performance, a practical BCI must generalise to new users from minutes of data, not hours. The 2026 PNPL competition responds to this challenge with LibriBrain100 (Mantegna et al., 2026), an extended LibriBrain dataset with 32 additional subjects (~40 minutes each) plus even more within-subject data (~80 hours). Advancing the curriculum of tasks to focus on word classification, two complementary tracks are presented in this competition: the Deep track targets within-subject word classification at scale, aiming at the best possible performance; the Broad track targets cross-subject generalisation, progressively reducing the amount of subject-specific fine-tuning data from ~40 to ~20 to ~10 minutes, a duration that falls within a clinically feasible range and brings us a step closer to a non-invasive BCI capable of restoring communication to people living with profound paralysis.

Foundation models are beginning to reshape brain-signal analysis by moving the field beyond task-specific decoding pipelines toward reusable models pretrained on broad neural datasets. Magnetoencephalography (MEG) is a compelling but still underdeveloped target for this shift: it captures human cortical dynamics at millisecond resolution while offering stronger spatial interpretability than EEG, making it especially valuable for source-resolved studies of perception, language, cognition, and clinical brain function. Yet MEG foundation models remain at an early stage, with only a small number of MEG-specific and MEG-inclusive multi-modal models, modest pretraining corpora, and emerging but still limited benchmarks. This perspective lays down the basic concepts needed to understand MEG foundation models and provides a didactic overview of the field's key design choices, including tokenization, sensor- versus source-space representations, sensor-geometry encoding, backbone architectures, self-supervised objectives, and pretraining data. We then offer a roadmap for future development, organized around native MEG pretraining, adaptation of EEG foundation models, transfer from generic time-series models, and multi-modal integration with EEG, fMRI, MRI, behaviour, and stimulus features. We highlight the need for coordinated infrastructure, including diverse and reusable MEG datasets, rigorous evaluation across subjects, sites, tasks, and clinical settings, and responsible data-sharing practices that address consent, privacy, access, and governance.

We present a retrospective on the 2025 PNPL Competition, an open machine-learning benchmark for decoding speech from non-invasive brain recordings, which ran from June to September 2025 and culminated in a workshop at NeurIPS that December. The competition used the LibriBrain dataset—over 50 hours of within-subject MEG with standard data splits—and challenged participants to tackle two tasks: Speech Detection (binary frame-level classification) and Phoneme Classification (39-class ARPAbet prediction). Across 155 registered teams and 6,041 submissions, top F1-macro scores reached 95.6% for Speech Detection and 73.6% for Phoneme Classification, exceeding prior baselines of 68% and 44%, respectively, by wide margins. High-performing submissions converged on a common architectural template—temporal CNN or TCN backbones with task-specific recurrent or attention-based heads—whilst task-specific strategies such as temporal smoothing and ensemble voting proved critical for competitive performance. Hidden track evaluations suggest that main-track Speech Detection models generalised poorly to balanced class distributions. This could imply a reliance on class priors rather than robust neural decoding, highlighting a useful lesson: evaluation distributions should reflect intended deployment conditions. We further reflect on design choices, organisational lessons, and plans for future competitions in this series.

We present Coord2Region, an open-source Python package that streamlines coordinate-based neuroimaging workflows by automatically mapping 3D brain coordinates (e.g., MNI or Talairach) to anatomical regions across multiple atlases. The package links mapped coordinates to meta-analytic resources via the Neuroimaging Meta-Analysis Research Environment (NiMARE) , providing direct integration with Neurosynth and NeuroQuery. This directly connects coordinates and regions to the broader neuroimaging literature. In addition to atlas-based labeling and literature retrieval, Coord2Region offers an optional large language model (LLM) functionality that generates text summaries of linked studies and illustrative images of queried regions. These AI-assisted features are intended to support interpretation and exploration, while remaining clearly complementary to peer-reviewed literature and established neuroimaging tools. Coord2Region provides a unified pipeline with a robust command-line interface, flexible dataset management, and provider-agnostic LLM utilities, and it supports both single-coordinate and high-throughput batch queries with nearest-region fallback for volume and surface atlases. Furthermore, Coord2Region includes a web interface for interactive configuration (via JSON Schema forms) and cloud execution (via Hugging Face), enabling users to build YAML configurations and run analyses in-browser without local installation. Together, these capabilities lower friction, reduce manual errors, and improve reproducibility in coordinate-centric neuroimaging workflows, promoting more robust and transparent research practices.

Non-ordinary states of consciousness (NOC) provide an opportunity to experience highly intense, unique, and perceptually rich subjective states. The neural mechanisms supporting these experiences remain poorly understood. This study examined brain activity associated with a self-induced, substance-free NOC known as Auto-Induced Cognitive Trance (AICT). Twenty-seven trained participants underwent high-density electroencephalography (EEG) recordings during rest and AICT. We analyzed the aperiodic component of the power spectrum (1/f), Lempel-Ziv complexity, and sample entropy from five-minute signal segments. A machine learning approach was used to classify rest and AICT, identify discriminative features, and localize their sources. We also compared EEG metrics across conditions and assessed whether baseline activity predicted the magnitude of change during AICT. Classification analyses revealed condition-specific differences in spectral exponents, complexity, and entropy. The aperiodic component showed the strongest discriminative power, followed by entropy and complexity. Source localization highlighted frontal regions, the posterior cingulate cortex, and the left parietal cortex as key contributors to the AICT state. Baseline neural activity in frontal and parietal regions predicted individual variability in the transition from rest to AICT. These findings indicate that AICT engages brain regions implicated in rich subjective experiences and provide mechanistic insights into how self-induced trance states influence neural functioning.

Objective. Magnetoencephalography (MEG) is a cutting-edge neuroimaging technique that measures the intricate brain dynamics underlying cognitive processes with an unparalleled combination of high temporal and spatial precision. While MEG data analytics have traditionally relied on advanced signal processing and mathematical and statistical tools, the recent surge in artificial intelligence has led to the growing use of machine learning (ML) methods for MEG data classification. An emerging trend in this field is the use of artificial neural networks (ANNs) to address various MEG-related tasks. This review aims to provide a comprehensive overview of the state of the art in this area. Approach. This topical review included studies that applied ANNs to MEG data. Studies were sourced from PubMed, Google Scholar, arXiv, and bioRxiv using targeted search queries. The included studies were categorized into three groups: 'Classification', 'Modeling', and 'Other'. Key findings and trends were summarized to provide a comprehensive assessment of the field. Main results. We identified 119 relevant studies, with 70 focused on 'Classification', 16 on 'Modeling', and 33 in the 'Other' category. 'Classification' studies addressed tasks such as brain decoding, clinical diagnostics, and brain-computer interfaces implementations, often achieving high predictive accuracy. 'Modeling' studies explored the alignment between ANN activations and brain processes, offering insights into the neural representations captured by these networks. The 'Other' category demonstrated innovative uses of ANNs for artifact correction, preprocessing, and neural source localization. Significance. By establishing a detailed portrait of the current state of the field, this review highlights the strengths and current limitations of ANNs in MEG research. It also provides practical recommendations for future work, offering a helpful reference for seasoned researchers and newcomers interested in using ANNs to explore the complex dynamics of the human brain with MEG.

Neuroprotective interventions for dementia with Lewy bodies (DLB) and Parkinson’s disease (PD) are still in their early days. Clinical trials are expected to target idiopathic rapid eye movement sleep behavior disorder (iRBD), their strongest predictor. However, the presentation and progression of symptoms within this population show significant heterogeneity. We used machine learning (ML) to identify the clinical markers that are best at distinguishing iRBD patients (n=156) who developed DLB (n=26) from those who developed PD (n=34) at a mean follow-up of 4.37 years. Our model classified subsequent conversion to DLB versus PD with 0.80 accuracy, with mild cognitive impairment as best predictive feature. Cognitive tests of executive functions and verbal learning also played a major role in classifying other related pathological trajectories. These findings support the use of ML with clinical markers in iRBD, paving the way for a more targeted selection of participants in future neuroprotective trials of synucleinopathies.

In the past decade, there has been a surge of research at the intersection of neuroscience and artificial intelligence (AI) aimed at advancing our understanding of both artificial and natural cognition. Growing evidence suggests that biological and artificial neural networks trained on similar tasks can exhibit striking functional parallels. Driven by the imperative to model the brain in order to decipher its underlying mechanisms, artificial neural networks (ANNs)—originally inspired by its architecture and functions—have been proposed as effective models of various brain systems. Convolutional Neural Networks (CNNs) trained on object recognition have demonstrated their ability to approximate the human visual system’s processing hierarchy and internal representations. In the context of face perception, neuroscience findings highlight a specialized neural system; yet whether familiar and unfamiliar faces are processed by the same mechanisms or via distinct pathways remains debated. Although numerous studies have compared AI-based face models to human behavior or fMRI data, questions persist about how closely these models capture the temporal dynamics of human face processing. This thesis first reviews current knowledge of the human visual system, focusing on the dedicated face recognition circuitry, and then introduces foundational concepts in AI, including the modeling of face perception with CNNs. The core work compares seven CNN architectures against source-localized magnetoencephalography (MEG) data to probe the neural signatures of face recognition and familiarity over time. These networks were optimized for different tasks—face recognition, object recognition, or both—allowing us to assess how task-specific representations capture the brain’s face processing in distinct ways. Our findings show that FaceNet aligns particularly well with occipital and fusiform regions implicated in face perception, while certain other deep architectures (e.g., ResNet) also achieve comparable levels of neural alignment. In the occipital region, the M170 component associated with familiarity occurs earlier (around 160ms) for familiar faces and later for unfamiliar ones (approximately 180ms), suggesting that novel identities demand more prolonged processing. We additionally observe strong CNN–MEG similarities in theta and gamma frequency bands, with earlier peaks (M170–M200) for familiar stimuli and a shift toward M400 for unfamiliar faces. Comparing multiple training objectives confirms the training task could have an impact on the temporal alignment with brain data. Finally, the discussion addresses potential limitations of CNNs as models of the brain, while highlighting their promise in shedding light on the neural mechanisms underlying face recognition. The insights gained from this work may guide the development of more robust models of face perception for both AI and computational neuroscience.

The advance of speech decoding from non-invasive brain data holds the potential for profound societal impact. Among its most promising applications is the restoration of communication to paralysed individuals affected by speech deficits such as dysarthria, without the need for high-risk surgical interventions. The ultimate aim of the 2025 PNPL competition is to produce the conditions for an 'ImageNet moment' or breakthrough in non-invasive neural decoding, by harnessing the collective power of the machine learning community. To facilitate this vision we present the largest within-subject MEG dataset recorded to date (LibriBrain) together with a user-friendly Python library (pnpl) for easy data access and integration with deep learning frameworks. For the competition we define two foundational tasks (i.e. Speech Detection and Phoneme Classification from brain data), complete with standardised data splits and evaluation metrics, illustrative benchmark models, online tutorial code, a community discussion board, and public leaderboard for submissions. To promote accessibility and participation the competition features a Standard track that emphasises algorithmic innovation, as well as an Extended track that is expected to reward larger-scale computing, accelerating progress toward a non-invasive brain-computer interface for speech.

Humans excel at face recognition, relying on a specialized system whose representation of familiar and unfamiliar faces have been debated. Some would argue that unfamiliar and familiar faces are processed in the same way, while others claim otherwise. Similarly, artificial neural networks (ANNs) have shown remarkable abilities, prompting discussions on their similarities to human face processing and their use as models of the brain. This study employs Convolutional Neural Networks (CNNs) and Magnetoencephalography (MEG) to explore the signatures of face recognition and familiarity, and investigate whether the face selective areas in the brain are specialized only for faces (domain-specific) or are developed for more general purposes (domain-general). Our findings reveal distinct brain responses: occipital areas distinguish faces from non-faces, while fusiform and inferior temporal areas engage in familiar face recognition. When training our ANNs on face recognition, we observe significant results supporting the idea of domain-specific brain regions for faces. Our findings shed new light on the temporal dynamics of familiar versus unfamiliar face processing in the visual cortex and more globally, highlight the potentials of combining ANNs and MEG to uncover the neural mechanisms that mediate face processing in humans.

Machine learning (ML) is increasingly used in cognitive, computational and clinical neuroscience. The reliable and efficient application of ML requires a sound understanding of its subtleties and limitations. Training ML models on datasets with imbalanced classes is a particularly common problem, and it can have severe consequences if not adequately addressed.

With the neuroscience ML user in mind, this paper provides a didactic assessment of the class imbalance problem and illustrates its impact through systematic manipulation of data imbalance ratios in (i) simulated data and (ii) brain data recorded with electroencephalography (EEG), magnetoencephalography (MEG) and functional magnetic resonance imaging (fMRI).

Our results illustrate how the widely-used Accuracy (Acc) metric, which measures the overall proportion of successful predictions, yields misleadingly high performances, as class imbalance increases. Because Acc weights the per-class ratios of correct predictions proportionally to class size, it largely disregards the performance on the minority class. A binary classification model that learns to systematically vote for the majority class will yield an artificially high decoding accuracy that directly reflects the imbalance between the two classes, rather than any genuine generalizable ability to discriminate between them.

We show that other evaluation metrics such as the Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC), and the less common Balanced Accuracy (BAcc) metric - defined as the arithmetic mean between sensitivity and specificity, provide more reliable performance evaluations for imbalanced data. Our findings also highlight the robustness of Random Forest (RF), and the benefits of using stratified cross-validation and hyperprameter optimization to tackle data imbalance.

Critically, for neuroscience ML applications that seek to minimize overall classification error, we recommend the routine use of BAcc, which in the specific case of balanced data is equivalent to using standard Acc, and readily extends to multi-class settings. Importantly, we present a list of recommendations for dealing with imbalanced data, as well as open-source code to allow the neuroscience community to replicate and extend our observations and explore alternative approaches to coping with imbalanced data.

Mounting evidence suggests that biological and artificial neural networks trained on similar tasks can exhibit remarkable functional similarities. In particular, Convolutional Neural Networks (CNNs) trained on object recognition have been shown to learn representations that model the processing hierarchy in the human visual system. But what about the specific case of face recognition? Would CNNs trained on face recognition learn representations that capture the neural dynamics of the neural circuits that mediate face recognition? Here, we investigated the putative similarities between the representations learned by three CNN architectures trained on face recognition (FaceNet, ResNet50 and CORnet-S) and brain patterns assessed with magnetoencephalography (MEG) recorded during a face recognition task. We found that the neuromagnetic brain signals (especially in visual areas) were correlated with activations in multiple CNN layers. However, these correlations were not as strong as those observed in the more general task of object recognition. This may suggest that CNNs trained on face recognition capture only a small portion of the complexity of the real brain patterns associated with face recognition. Our results contribute to an emerging stream of research that seeks to probe face recognition mechanisms by joint exploration of the associated neural representations in both brains and machines.