2025Activity​​ reportProject-TeamMIND

Participants:​​​‌

P. Ciuciu, A. Gramfort,‌ T. Moreau, D. Wassermann‌​‌

Inverse problems are ubiquitous​​ in observational science. This​​​‌ necessitates the reconstruction of‌ a signal/image of interest,‌​‌ or more generally a​​ vector of parameters, from​​​‌ remote observations that are‌ possibly noisy and scarce.‌​‌ The link between the​​ parameters of interest and​​​‌ the observations is physics,‌ and is commonly well‌​‌ understood. Yet, the recovery​​ of parameters is challenging​​​‌ as the problem is‌ often ill-posed due to‌​‌ the ill-conditioning of the​​ forward model. Machine learning​​​‌ is now more frequently‌ used to address such‌​‌ problems, using likelihood-free inference​​ (LFI) to inverse nonlinear​​​‌ systems, or prior learning‌ using bi-level optimization and‌​‌ reinforcement learning to guide​​ the way to collect​​​‌ observations.

3.1.1 From linear‌ inverse problems to simulation‌​‌ based inference

Solving‌​‌ an inverse problem consists​​ in estimating the unobserved​​​‌ parameters at the origin‌ of some measurements. Typical‌​‌ examples are image denoising​​ or image deconvolution, where,​​​‌ given noisy or low‌ resolution data, the objective‌​‌ is to obtain an​​ underlying high-quality image. Inverse​​​‌ problems are pervasive in‌ experimental sciences such as‌​‌ physics, biology or neuroscience.​​ The common problem across​​​‌ these fields is that‌ the measurements are noisy‌​‌ and generally incomplete.

Mathematically​​ speaking, these inverse problem​​​‌ can be formulated as‌ estimating $𝐱$ from $\mathrm{𝐲‌‌}=\Gamma \left(\mathrm{𝐱}\right)+𝐛$.​​​‌ Here, $𝐛$ is an‌ additive noise and $\mathrm{\Gamma ‌‌}$ is a (generally non-injective)​​ mapping to a lower-dimensional​​​‌ space. For example, in‌ magneto- and electroenchephalography (M/EEG),‌​‌ $\Gamma$ is a real​​ linear mapping ${\Gamma }_{\mathrm{M‌}}/\mathrm{EEG}:{\mathrm{ℝ‌}}^N\mapsto {ℝ}^{\mathrm{M‌‌}}$ and $𝐛$ is considered​​ white and Gaussian, while​​​‌ in magnetic resonance imaging‌ (MRI), $\Gamma$ is a‌​‌ complex linear mapping ${\mathrm{\Gamma }}_{\mathrm{MRI}}:{ℂ}^{\mathrm{N‌}}\mapsto {ℂ}^M$ and‌ $𝐛$ is circular complex‌​‌ white Gaussian. Despite the​​ linearity of ${\Gamma }_{\mathrm{M‌}}/\mathrm{EEG}$ and ${\mathrm{\Gamma ‌}}_{\mathrm{MRI}}$, estimating $\mathrm{𝐱‌‌}$ is a challenging task​​​‌ when the measurements are​ incomplete, i.e., $M\ll ‌N$ and the problem​​ is ill-posed. This is​​​‌ often the case due​ to physical limitations on​‌ the measurement device (M/EEG)​​ or the acquisition time​​​‌ (MRI). Moreover, the linear​ Fourier operator ${\Gamma }_{\mathrm{MRI‌}}$ only reflects an ideal​​ acquisition process and part​​​‌ of the acquisition artifacts​ (e.g. B0 inhomogeneity) can​‌ be compensated by considering​​ nonlinear models at the​​​‌ cost of estimating additional​ parameters along with the​‌ MR image.

To tackle​​ these inverse problems, using​​​‌ adequate regularization will promote​ the right structure for​‌ the data to be​​ recovered. Over the last​​​‌ decade the members of​ Mind have proposed state-of-the-art​‌ models and efficient algorithms​​ based on sparsity assumptions​​​‌ 97, 148,​ 120, 102,​‌ 149, 132,​​ 131, 118,​​​‌ 122, 119.​ MNE is the reference​‌ software developed by the​​ team that implements these​​​‌ methods for MEG/EEG data​ while pysap-mri proposes solvers​‌ for MR image reconstruction.​​

The field is now​​​‌ progressing with novel approaches​ based on deep learning​‌ by either learning the​​ regularization from data in​​​‌ the context of MRI​ reconstruction 171, 169​‌, or by considering​​ nonlinear models grounded in​​​‌ the physics underlying the​ data. The team has​‌ started to explore this​​ direction using so-called Likelihood-Free​​​‌ Inference (LFI) techniques built​ on deep invertible networks​‌ 172, 130.​​ A particular application has​​​‌ been on diffusion MRI​ (dMRI), where we have​‌ linked the dMRI signal​​ with physiological tissue models​​​‌ of grey matter tissue​ 130. Still in​‌ MRI but in susceptibility​​ weighted imaging, another approach​​​‌ 112 has consisted in​ directly estimating the B0​‌ field map from non-Cartesian​​ k-space data to correct​​​‌ for off-resonance effects in​ non-Fourier operators ${\Gamma }_{\mathrm{MRI‌}}$. The Mind project​​ will continue along this​​​‌ direction studying nonlinear simulators​ of imaging data as​‌ building blocks. A key​​ aspect of the work​​​‌ proposed is to exploit​ knowledge on the physics​‌ of the data generation​​ mechanisms.

3.1.2 Bi-level optimization​​​‌

In​ recent years, bi-level optimization​‌ – minimizing over a​​ parameter which is itself​​​‌ the solution of another​ optimization problem – has​‌ raised great interest in​​ the machine learning community.​​​‌ Indeed, many methods in​ ML reduce to this​‌ bi-level framework, typically the​​ problem of hyper-parameter optimization.​​​‌

In most practical cases,​ hyper-parameter selection is done​‌ using cross-validation (CV), which​​ basically consists in splitting​​​‌ the whole dataset in​ training and validation sets.​‌ The parameters of the​​ method are computed by​​​‌ minimizing a loss function​ on the training set,​‌ and the hyper-parameters are​​ then set by minimizing​​​‌ the loss function on​ the validation set. This​‌ approach is a bi-level​​ optimization problem.

Other instances​​​‌ of such problems can​ be found in dictionary​‌ learning, robust training of​​ neural networks or the​​ use of implicit layers​​​‌ in deep learning. In‌ all these applications, the‌​‌ model or the latent​​ variables are learned by​​​‌ minimizing some loss while‌ the parameters or the‌​‌ dictionary are updated by​​ minimizing a second optimization​​​‌ problem depending on the‌ outcome of the first‌​‌ problem. While theoretical results​​ were produced in the​​​‌ early 70's 111,‌ there are still many‌​‌ challenges related to bi-level​​ optimization that need to​​​‌ be addressed to produce‌ methods that are both‌​‌ theoretically well grounded and​​ computationally efficient. Recently, the​​​‌ members of Mind have‌ published several works related‌​‌ to the subject 79​​, 80, 92​​​‌, 104. We‌ intend to pursue this‌​‌ effort in the following​​ directions.

3.1.3 Reinforcement learning​ for active k-space sampling​‌

Current under-sampling schemes​‌ in MRI allow for​​ shorter scan acquisition times,​​​‌ however at the cost​ of artifacts in various​‌ regions of the reconstructed​​ MR image. These artifacts​​​‌ arise due to uncertainties​ in some heavily under-sampled​‌ regions of the acquired​​ Fourier space (i.e. also​​​‌ called k-space). Modern reconstruction​ algorithms, with the use​‌ of strong priors, either​​ hand-crafted or learned, tend​​​‌ to reduce these uncertainties​ and behave as if​‌ the acquisition is fixed.​​

To go beyond the​​​‌ state of the art,​ we argue that there​‌ is a need to​​ jointly learn an algorithm​​​‌ that designs the optimal​ under-sampling pattern in k-space​‌ as well as the​​ reconstruction network.

As it​​​‌ can be summarized to​ learning a sequential decision​‌ algorithm, we will rely​​ on reinforcement learning (RL)​​​‌ to build up optimal​ k-space sampling patterns while​‌ enforcing physical constraints on​​ the MRI sequence, as​​​‌ originally proposed in 100​, 95, 140​‌.

The k-space acquisition​​ can be modeled by​​​‌ a sampling policy and​ the rewards for the​‌ joint network are based​​ on reconstructed image quality.​​​‌ Under this paradigm, after​ every fixed scan time,​‌ an instantaneous reconstruction can​​ be obtained and the​​​‌ Fourier space uncertainty maps​ analysed in depth. Based​‌ on this, the scan​​ can continue by actively​​​‌ sampling the k-space and​ enforcing denser samples in​‌ regions where uncertainty is​​ larger. In this way,​​​‌ the learned k-space trajectories​ may become more patient​‌ and organ specific. Further,​​ the trajectory can run​​​‌ and lead to instantaneous​ best results of reconstruction​‌ under a given variable​​ scan time budget. These​​​‌ aspects define one of​ the core directions we​‌ will investigate to produce​​ the next generation of​​​‌ state-of-the-art MR data sampling​ and image reconstruction algorithms.​‌ Recent contributions 185,​​ 166 only approach the​​​‌ problem in the Cartesian​ framework and hence perform​‌ 1D variable density sampling​​ along the phase encoding​​​‌ dimension. Given our expertise​ on non-Cartesian sampling in​‌ developing SPARKLING for both​​ for 2D and 3D​​​‌ MR imaging 140,​ 141, 98,​‌ we plan to extend​​ this framework to non-Cartesian​​​‌ acquisition setups while still​ remaining compatible with hardware​‌ constraints on the gradient​​ system. The access to​​​‌ various MRI scanners at​ CEA/NeuroSpin is necessary and​‌ an added advantage to​​ the success of the​​​‌ Mind team.

Participants:

B. Thirion, D.​​ Wassermann

Inferring the relationship​​​‌ between the physiological bases​ of the human brain​‌ and its cognitive functions​​ requires articulating different datasets​​​‌ in terms of their​ semantics and representation. Examples​‌ of these are spatio-temporal​​ brain images, tabular datasets,​​​‌ structured knowledge represented as​ ontologies, and probabilistic datasets.​‌ Developing a formalism that​​ can integrate all these​​ modalities requires constructing a​​​‌ framework able to represent‌ and efficiently perform computations‌​‌ on high-dimensional datasets as​​ well as to combine​​​‌ hybrid data representations in‌ deterministic and probabilistic settings.‌​‌ We will take on​​ two main angles to​​​‌ achieve this task: on‌ one hand, the automated‌​‌ inference of cross-dataset features,​​ or coordinated representations and​​​‌ on the other hand,‌ the use of probabilistic‌​‌ logic for knowledge representation​​ and inference. The probabilistic​​​‌ knowledge representation part is‌ now well advanced with‌​‌ the Neurolang project. It​​ is yet a long-term​​​‌ endeavor. The learning of‌ coordinated representations is less‌​‌ advanced.

3.2.1 Learning coordinated​​ representations

Inference is the​​​‌ pathway that leads from‌ data to knowledge. One‌​‌ crucial aspect is that​​ in the context of​​​‌ neuroscience, data comes in‌ different forms: Full texts,‌​‌ images and tables. Annotations​​ may be full texts​​​‌ or simply tags associated‌ with observed images. One‌​‌ challenge is thus to​​ develop automated techniques that​​​‌ learn coordinated representations across‌ such heterogeneous data sources.‌​‌

This learning endeavor rests​​ on several key machine​​​‌ learning techniques: Compression, embeddings,‌ and multi-layer networks. Compression‌​‌ (sketching) consists in building​​ a reduced representation of​​​‌ some input that leads‌ from large sparse and‌​‌ complex representation to low-dimension​​ ones, while minimizing some​​​‌ distortion criterion. Embedding techniques‌ also create representations, but‌​‌ possibly bias them to​​ enhance some aspects of​​​‌ the data. It thus‌ incorporates prior information on‌​‌ data distribution or the​​ relevance of features. Finally,​​​‌ multi-layer networks create intermediate‌ representation of data that‌​‌ are suitable to achieve​​ a prediction goal. Such​​​‌ representations are rich enough‌ in particular in multi-task‌​‌ settings, where the outputs​​ of the network are​​​‌ multi-dimensional. Following 86,‌ we call such latent‌​‌ data models coordinated representations​​.

Deep learning is​​​‌ well suited to the‌ goal of learning intermediate‌​‌ representations. As an example,​​ we plan to develop​​​‌ a framework that coalesces‌ in one deep learning‌​‌ formulation, the task of​​ estimating brain structures, cognitive​​​‌ concepts, and their relationships.‌

Brain structures and cognitive‌​‌ concepts will appear as​​ intermediate representations responsible for​​​‌ linking brain activity to‌ observed behavior. However deep‌​‌ learning cannot be considered​​ as a standard means​​​‌ to understand coordinated representations,‌ due to the limited‌​‌ data available, their poor​​ signal-to-noise ratio (SNR) and​​​‌ their heterogeneity. Deep learning‌ needs instead to be‌​‌ adapted by injecting our​​ expertise on statistical structure​​​‌ of the data (see‌ e.g. 128, 151‌​‌). Since the challenge​​ is to train such​​​‌ models on limited and‌ noisy data, we will‌​‌ extend our recent work​​ 84 that has developed​​​‌ regularization schemes for deep-learning‌ models: it relies on‌​‌ structured stochastic regularizations (a.k.a.​​ structured dropout). Such approaches​​​‌ are efficient, powerful and‌ can be used in‌​‌ wide settings. We will​​ enhance them with more​​​‌ generic, cross-layer, grouping schemes.‌ Additionally, we will develop‌​‌ two strategies: i) aggregation​​​‌ of predictors for variance​ reduction and stability of​‌ the model 128 and​​ ii) data augmentation –​​​‌ i.e. learning to augment,​ based on unlabeled data​‌ – to improve the​​ fit with limited data.​​​‌ For this we will​ consider plausible generative mechanisms.​‌

3.2.2 Probabilistic Knowledge Representation​​

Neuroscientific​ data used to infer​‌ the relationships between physiology​​ of the human brain​​​‌ and its cognitive function​ goes well beyond text,​‌ image, and tables. Knowledge​​ graphs representing human knowledge,​​​‌ and the ability to​ encode reasoning strategies in​‌ neuroscience are also key​​ to effectively bridge current​​​‌ data-centric approaches and decades-old​ domain knowledge. A main​‌ challenge in performing inferences​​ combining demographic data-centric approaches,​​​‌ imaging measurements, and domain​ knowledge, is to be​‌ able to infer new​​ knowledge soundly and efficiently​​​‌ taking into account the​ noisy nature of demographic​‌ and imaging measurements, and​​ the common open-world assumption​​​‌ of ontologies and knowledge​ graphs. Such probabilistic hybrid​‌ logic approaches are known​​ to be, in general,​​​‌ intractable in the deterministic​ 82 as well as​‌ in the probabilistic case​​ 182. Nonetheless, there​​​‌ is an opportunity to​ be seized in identifying​‌ tractable segments of probabilistic​​ hybrid logic representations able​​​‌ to solve open neuroscientific​ questions.

A noticeable opportunity​‌ to incorporate all statistical​​ evidence gathered from noisy​​​‌ data into a usable​ knowledge base is to​‌ formalize the inferred relationships​​ into probabilistic symbolic representations​​​‌ 129. These representations​ are much better suited​‌ to simultaneously handle data​​ across topologies and logic​​​‌ systems, implementing inferential algorithms​ avoiding the brittleness of​‌ deterministic logic as well​​ as causal probabilistic reasoning.​​​‌

A typical application of​ such heterogeneous data processing​‌ is meta-analytic applications which​​ combine neuroimaging data with​​​‌ results found in the​ scientific literature. Current tools​‌ to perform this task​​ are NeuroSynth or Neuroquery​​​‌ (developed by the team).​ However, knowledge inferred by​‌ such tools is tremendously​​ limited by the expressive​​​‌ power of the language​ used to query the​‌ data. Current meta-analytic tools​​ are able to express​​​‌ queries relating test makers,​ article annotations, and their​‌ relationship with reported brain​​ activations, support propositional logic​​​‌ only. Propositional logic requires​ the user to explicitly​‌ express every desired term​​ with their characteristics and​​​‌ their relationships. Our goal​ is to extend the​‌ inference capabilities of such​​ applications by leveraging current​​​‌ advances in probabilistic logic​ languages and embedding them​‌ in the Neurolang language.​​ Neurolang enables the encoding​​​‌ of complex knowledge in​ terms of more expressive​‌ queries. Neurolang queries first-order​​ logic segment, FO${}^{\neg ‌\exists }$, with a​ tractable probabilistic extension allowing​‌ for high-dimensional and large​​ dataset computations. Such segment​​​‌ of first order logic​ enables formalising questions such​‌ as “what brain areas​​ are most likely reported​​​‌ active in a study​ specifically when terms related​‌ to consciousness are mentioned​​ in such study”, hence​​ being able to infer,​​​‌ amongst other tasks, specificity‌ and causality 183 of‌​‌ diverse neuroscience phenomena. To​​ disseminate our results allowing​​​‌ complex expressive searches of‌ massively aggregated diverse data,‌​‌ we will leverage Neurolang​​. The latter produces​​​‌ a domain-specific language (DSL)‌ for human neuroscience research,‌​‌ while being able to​​ combine imaging data, anatomical​​​‌ descriptions and ontologies. Three‌ main characteristics of the‌​‌ DSL are key to​​ fulfilling this goal: First,​​​‌ it represents neuroimaging-derived information‌ and spatial relationships in‌​‌ a syntax close to​​ natural language used by​​​‌ neuroscientists 184. Second,‌ through a back-end belonging‌​‌ to the Datalog${}^{\pm }$ family, it allows querying​​​‌ ontologies with the same‌ expressive power as current‌​‌ standards SPARQL and OWL​​ 88. Finally, we​​​‌ will extend Neurolang to‌ a probabilistic language able‌​‌ to express graphical models​​ allowing the implementation of​​​‌ a wide variety of‌ causal inference and machine‌​‌ learning algorithms 85 in​​ high-dimensional settings which are​​​‌ specific to neuroimaging research.‌ In sum, by leveraging‌​‌ recent advances in deductive​​ database systems 88 and​​​‌ this novel DSL 184‌ we will provide a‌​‌ more flexible tool to​​ express and infer knowledge​​​‌ on brain structure-function relationships.‌

Participants:

A. Gramfort, T.​​​‌ Moreau, B. Thirion, D.‌ Wassermann

Statistics is the‌​‌ natural pathway from data​​ to knowledge. Using statistics​​​‌ on brain imaging data‌ involves dealing with high-dimensional‌​‌ data that can induce​​ intensive computation and low​​​‌ statistical power. Besides, statistical‌ models on large-scale data‌​‌ also need to take​​ potential confounding effects and​​​‌ heterogeneity into account. To‌ address these questions the‌​‌ Mind team will employ​​ causal modeling and post-selection​​​‌ inference. Conditional and post-hoc‌ inference are rather short-term‌​‌ perspectives, while the potential​​ of causal inference stands​​​‌ as a longer-term endeavor.‌

3.3.1 Conditional inference in‌​‌ high dimension

Conditional inference consists of​​​‌ assessing the importance of‌ a certain feature in‌​‌ a predictive model, while​​ taking into account the​​​‌ information carried by alternative‌ features. One motivation for‌​‌ using this inference scheme​​ is that brain regions​​​‌ that sustain behavior and‌ cognition are strongly interacting.‌​‌ Taking these interactions into​​ account is critical to​​​‌ avoid confusing correlation with‌ causation in brain/behavior analysis.‌​‌

Technical difficulties come when​​ the set of explanatory​​​‌ features $𝐗$ becomes extremely‌ large as frequently met‌​‌ in neuroimaging: Conditioning on​​ many variables (or equivalently,​​​‌ high dimensional variables) is‌ computationally costly and statistically‌​‌ inefficient. The main solutions​​ to date are based​​​‌ either on linear model‌ debiasing 134, as‌​‌ well as simulation-based approaches​​ (knockoff inference 96 or​​​‌ conditional randomization tests 144‌). Importantly the latter‌​‌ involves simulating data with​​ statistical characteristics described explicitly​​​‌ (in a parametric family)‌ or implicitly (by samples).‌​‌ There remain two gaps​​ to bridge for these​​​‌ methods: i) The computational‌ gap, as the algorithmic‌​‌ complexity of these approaches​​​‌ is typically cubic in​ the number of samples,​‌ unless more efficient generative​​ mechanisms are available; ii)​​​‌ the power gap, related​ to the limited number​‌ of available samples. The​​ best solution thus far​​​‌ consists of associating these​ inference procedures with dimension​‌ reduction procedures 157.​​ The next step is​​​‌ adaptation to more general​ settings: Conditional inference has​‌ been formulated in the​​ linear framework, where it​​​‌ boils down to controlling​ that the corresponding coefficient​‌ is non-zero, hence it​​ has to be generalized​​​‌ to nonlinear models: Non-parametric​ models like random forests,​‌ then possibly deep networks.​​

3.3.2 Post-selection inference on​​​‌ image data

Large-scale​​​‌ statistical testing is pervasive​ in many scientific fields,​‌ where high-dimensional datasets are​​ collected and compared with​​​‌ an outcome of interest.​ In such high-dimensional contexts,​‌ false discovery rate (FDR)​​ control 90 is attractive​​​‌ because it yields reasonable​ power, while providing an​‌ explicit and interpretable control​​ on false positives. Yet​​​‌ the FDR rate is​ the expectation of the​‌ FDP. Controlling the FDR​​ does not mean that​​​‌ the FDP is controlled,​ a distinction that is​‌ most often ignored by​​ practitioners. For the sake​​​‌ of scientific reproducibility, there​ is a need for​‌ methods controlling the FDP.​​

Such an approach has​​​‌ been developed in the​ context of neuroimaging, namely​‌ the all-resolution inference framework​​ 173 based on classical​​​‌ multiple correction error control​ bounds. Yet, the empirical​‌ behavior of this method​​ remains to be assessed.​​​‌ Moreover, it has been​ clearly established that the​‌ procedure is over-conservative in​​ some settings 93.​​​‌ Indeed, it relies on​ the Simes statistical bound,​‌ that is not adaptive​​ to the specific type​​​‌ of dependence for a​ particular data set. To​‌ bypass these limitations, 93​​ have proposed a randomization-based​​​‌ procedure known as $\mathrm{\lambda }$-calibration, which yields tighter​‌ mathematical bounds that are​​ adapted to the dependency​​​‌ observed in the dataset​ at hand. It rests​‌ on a non-parametric (permutation-based)​​ estimation of the null​​​‌ distribution, leading to tight​ and valid inference under​‌ general assumptions.

In this​​ research axis, we propose​​​‌ to fix some of​ the open issues with​‌ the approach described in​​ 93, namely the​​​‌ choice of a template​ family to calibrate the​‌ error distribution in the​​ permutation procedure. We hope​​​‌ to propose a practical​ choice for this family​‌ to avoid putting the​​ burden of choice on​​​‌ practitioners.

We will characterize​ by simulations and theoretical​‌ arguments the behavior of​​ these error control procedures​​​‌ and develop efficient computational​ methods for the use​‌ of these tools in​​ brain imaging analysis.

3.3.3​​​‌ Causal inference for population​ analysis

Modern​​​‌ health datasets present population​ characteristics with many variables​‌ and in multiple modalities.​​ They can ground prediction​​ and understanding of individual​​​‌ outcomes, using machine learning‌ techniques. Still, heterogeneous variables‌​‌ have complex relationships, making​​ it hard to tease​​​‌ apart each factor in‌ an outcome of interest.‌​‌ Potential outcome theory 175​​ provides a valuable framework​​​‌ to evaluate the impact‌ of treatment (interventions). Treatment‌​‌ effects can be heterogeneous.​​ In particular, interactions between​​​‌ background and treatment variables‌ have to be considered.‌​‌

The statistical behavior (consistency​​ and efficiency) under non-parametric​​​‌ models is actively investigated‌ 83, 159.‌​‌ However, their behavior in​​ high-dimensional settings, when both​​​‌ the number of features‌ and the number of‌​‌ samples are large, is​​ still poorly understood. Our​​​‌ objective is thus to‌ extend the theory and‌​‌ algorithms of causal inference​​ to noisy high-dimensional settings,​​​‌ where the noise level‌ implies that effects sizes‌​‌ are proportionally small, and​​ classic methods often become​​​‌ inefficient and potentially inaccurate‌ due to overfitting. More‌​‌ specifically, we plan to​​ explore the following directions.​​​‌

Participants:

P. Ciuciu,​‌ A. Gramfort, T. Moreau,​​ D. Wassermann, B.Thirion

The​​​‌ brain is a dynamic​ system. A core task​‌ in neuroscience is to​​ extract the temporal structures​​​‌ in the recorded signals​ as a means to​‌ linking them to cognitive​​ processes or to specific​​​‌ neurological conditions. This calls​ for machine learning methods​‌ that are designed to​​ handle multivariate signals, possibly​​​‌ mapped to some spatial​ coordinate system (e.g. like​‌ in fMRI).

3.4.1 Injecting​​ structural priors with Physics-informed​​​‌ data augmentation

Data​‌ augmentation consists of virtually​​ increasing dataset size during​​​‌ learning by applying random,​ yet plausible, transformations to​‌ the input data. In​​ computer vision, this means​​​‌ altering data by applying​ symmetries, rotations, geometric deformations​‌ etc. While such strategies​​ are reasonable for natural​​​‌ or medical images 164​, it is still​‌ unclear how neural or​​ BOLD signals can be​​​‌ augmented in order to​ improve prediction performance and​‌ robustness.

Some purely data​​ driven strategies have been​​​‌ proposed to augment EEG​ data using spectral transforms​‌ 145 or advanced strategies​​ such as channel, time​​​‌ or frequency masking or​ phase randomizations 139,​‌ 142. Although dozens​​ of transformations have been​​​‌ considered in the literature​ to augment EEG signals,​‌ it is now apparent​​ that different augmentation strategies​​​‌ should be applied to​ the data as a​‌ function of the prediction​​ task to be handled.​​​‌ For example when considering​ sleep stage classification or​‌ BCI applications, the spatial​​ sampling of electrodes and​​​‌ the duration of signals​ varies considerably, with the​‌ consequence being that different​​ augmentation parameters and even​​​‌ transformations need to be​ employed.

In this line​‌ of work we will​​ develop algorithms that can​​​‌ quickly identify the relevant​ augmentation techniques, building for​‌ example on 108,​​ 144. The aim​​​‌ is to provide a​ system that can automatically​‌ learn invariance within a​​ class and across subjects​​​‌ in order to maximize​ the prediction performance on​‌ unseen data. The methodology​​ developed will be relevant​​​‌ beyond neuroscience as long​ as a family of​‌ physics-informed transformations is available​​ for prediction tasks at​​​‌ hand.

3.4.2 Learning structural​ priors with self-supervised learning​‌

Self-supervised learning​​ (SSL) is a recently​​​‌ developed area of research​ that provides a compelling​‌ approach for exploiting large​​ unlabeled datasets. With SSL,​​​‌ the structure of the​ data is used to​‌ turn an unsupervised learning​​ problem into a supervised​​​‌ one, called a “pretext​ task”, such as solving​‌ Jigsaw puzzles from images​​ 161 or learning how​​ to color gray-scaled images.​​​‌ The representation learned on‌ the pretext task can‌​‌ then be reused for​​ unsupervised data exploration or​​​‌ on a supervised downstream‌ task, with the potential‌​‌ to greatly reduce the​​ number of labeled examples​​​‌ required to train a‌ good predictive model.

In‌​‌ fields like computer vision​​ 161, 153 and​​​‌ time series processing 162‌, SSL has shown‌​‌ great promise in terms​​ of prediction performance but​​​‌ also in ease of‌ use. Indeed, SSL simplifies‌​‌ model selection and evaluation​​ as it relies on​​​‌ prediction scores and cross-validation,‌ contrarily to unsupervised learning‌​‌ methods like ICA 78​​.

Recently the team​​​‌ has applied SSL to‌ two large cohorts of‌​‌ clinical EEG data 87​​ revealing insights on the​​​‌ data without any human‌ supervision. However many challenges‌​‌ remain. For example in​​ Mind, we aim​​​‌ to explore novel SSL‌ strategies applicable to electrophysiology‌​‌ as well as to​​ haemodynamic signals measured with​​​‌ fMRI. As such, our‌ goal is to expand‌​‌ the recent multivariate method​​ we have introduced in​​​‌ the field for the‌ blind deconvolution of BOLD‌​‌ signals in both task-related​​ and resting-state experiments 103​​​‌.

While rather small‌ networks have been employed‌​‌ so far on EEG​​ data 99, 174​​​‌ due to limited sets‌ of annotations, the use‌​‌ of SSL tasks opens​​ the possibility to work​​​‌ with much larger labeled‌ datasets, and therefore many‌​‌ more overparametrized models. We​​ aim to explore these​​​‌ directions, hoping to reach‌ a state where pre-trained‌​‌ models could be available​​ for EEG or MEG​​​‌ signals as is presently‌ the case for images‌​‌ or for natural language​​ processing (NLP) tasks.

3.4.3​​​‌ Revealing spatio-temporal structures with‌ convolutional sparse coding and‌​‌ driven point processes

The convolutional‌ sparse linear model is‌​‌ one established unsupervised learning​​ framework designed for signals.​​​‌ Using algorithms known as‌ convolutional sparse coding (CSC),‌​‌ this framework allows for​​ the learning of shift-invariant​​​‌ patterns to sparsely reconstruct‌ a time series. These‌​‌ patterns, also called atoms,​​ correspond to recurrent structures​​​‌ present in the data.‌ While some of our‌​‌ recent advances have improved​​ the computational tractability of​​​‌ these methods 155,‌ 154 and adapted them‌​‌ to neurophysiological data 133​​, 117, 103​​​‌, there are still‌ many shortcomings that make‌​‌ them unpractical for applications​​ beyond denoising.

4.1.1 Unveiling Cognition Through​ Population Modeling

Linking the​‌ human brain's structure and​​ function with cognitive abilities​​​‌ has been a research​ epicenter for the past​‌ 40 years. The sophistication​​ of brain mapping machinery​​​‌ such as MRI, EEG​ and MEG, has produced​‌ a treasure trove of​​ data. Nonetheless, the effect​​​‌ size of the phenomena​ leading to understanding cognition​‌ is often drowned out​​ by noise and inter-individual​​​‌ variability. A main goal​ of Mind is to​‌ simultaneously harness the power​​ of large-scale general purpose​​​‌ datasets, such as the​ Human Connectome Project (HCP)​‌ and the Adolescent Brain​​ Cognitive Development Study (ABCD),​​​‌ as well as small​ scale high precision ones,​‌ such as the Individual​​ Brain Charting (IBC) dataset​​​‌ 168, to understand​ the link between the​‌ human brain's architecture and​​ function, and cognition. Parietal​​​‌'s expertise has already​ been demonstrated in this​‌ field. Examples of this​​ include using diffusion MRI​​​‌ (dMRI) to link the​ brain's macrostructure with language​‌ comprehension 101, tissue​​ microstructure with cognitive control​​​‌ 150, functional gradients​ on the cortical surface​‌ 115 to functional territory​​ segregation 167.

Mind​​ project will continue this​​​‌ task by seizing our‌ core methodological developments, described‌​‌ in the previous section,​​ and our global collaborative​​​‌ network of cognitive scientists.‌

4.1.2 Imaging for health‌​‌ in the general population​​

Individual differences in brain​​​‌ function and cognition have‌ historically been investigated by‌​‌ studies carried out by​​ individual laboratories having access​​​‌ mainly to small sample‌ sizes. The growing availability‌​‌ of public large-scale data​​ of epidemiological dimensions curated​​​‌ by dedicated consortia (e.g.‌ UK Biobank) has enabled‌​‌ studying the relationship between​​ cognition and the brain​​​‌ with unparalleled granularity and‌ statistical power. These resources‌​‌ now allow researchers to​​ relate brain signals/images to​​​‌ rich descriptions of the‌ participants including behavioral and‌​‌ clinical assessments in addition​​ to social and lifestyle​​​‌ factors. Machine learning has‌ proven essential when modeling‌​‌ biomedical outcomes from the​​ large-scale and high-dimensional data​​​‌ brought by consortia and‌ biobanks. It is used‌​‌ to to build predictive​​ models of heterogenous biomedial​​​‌ outcomes (cognitive, social, clinical)‌ based on different neuroscientific‌​‌ modalities. Taken together, this​​ facilitates the study of​​​‌ lifestyle and health-related behavior‌ in the general population,‌​‌ potentially revealing risk factors​​ leading to biomarker discovery.​​​‌

Mind will greatly contribute‌ to this effort by‌​‌ focusing on population modeling​​ as a tool for​​​‌ enhancing the analysis of‌ clinical data and mental‌​‌ health.

4.1.3 Proxy measures​​ of brain health

Clinical​​​‌ datasets tend to be‌ small as sharing of‌​‌ data is not incentivized​​ or institutional and economic​​​‌ resources are missing. As‌ a consequence, the capacity‌​‌ of machine learning to​​ learn functions that relate​​​‌ complex-to-grasp biomedical outcomes to‌ heterogeneous data cannot be‌​‌ fully exploited. This has​​ stimulated growing interest in​​​‌ proxy measures of neurological‌ conditions derived from the‌​‌ general population, such as​​ individual biological aging. One​​​‌ counter-intuitive aspect of the‌ methodology is that measures‌​‌ of biological aging (e.g.​​ via brain imaging) can​​​‌ be obtained by focusing‌ on the age of‌​‌ a person, which is​​ known in advance and​​​‌ is, in itself not‌ interesting as a target.‌​‌ However, by predicting the​​ age, machine-learning can capture​​​‌ the relevant information about‌ aging. Based on a‌​‌ population of brain images,​​ it extracts the best​​​‌ guess for the age‌ of a person, indirectly‌​‌ positioning that person within​​ the population. Individual-specific prediction​​​‌ errors therefore reflect deviations‌ from what is statistically‌​‌ expected 181. The​​ brain of a person​​​‌ can look similar to‌ brains commonly seen in‌​‌ older (or younger) people.​​ The resulting brain-predicted age​​​‌ reflects physical and cognitive‌ impairment in adults 180‌​‌, 106, 116​​ and reveals neurodegenerative processes​​​‌ 143, 124,‌ which could be overlooked‌​‌ without using machine learning.​​

Mind will extend this​​​‌ line of research in‌ two directions: 1) Assessment‌​‌ of brain age using​​ EEG and non-brain data​​​‌ such as health-records and‌ 2) proxy measures of‌​‌ mental health beyond aging.​​

4.1.4 Studying brain age​​​‌ using electrophysiology

MRI is‌ not yet available in‌​‌ all clinical situations and​​ certain aspects of brain​​​‌ function are better understood‌ using electrophysiological modalities (M/EEG).‌​‌ Until recently, it was​​​‌ unclear if brain age​ can be meaningfully estimated​‌ from M/EEG. In a​​ recent study 121,​​​‌ we demonstrated, using the​ Cam-CAN cohort ($\mathrm{n‌}=650$), that​​ combining MRI and MEG​​​‌ enhanced detection of cognitive​ dysfunction. The proposed approach​‌ not only achieved integration​​ of brain signals from​​​‌ distinct modalities but explicitly​ handled the absence of​‌ MEG or MRI recordings,​​ adapting ideas from 136​​​‌. This is key​ for clinical translation where​‌ one cannot afford excluding​​ cases because one modality​​​‌ is missing. In the​ clinical setting, EEG is​‌ predominantly used (and not​​ MEG). Clinical recordings are​​​‌ far noisier than lab​ EEG and gold-standard source​‌ modeling with MRI is​​ rarely done outside the​​​‌ lab. Supported by theoretical​ analysis and simulations, we​‌ found through empirical benchmarks​​ 176 that Riemannian embeddings​​​‌ 1) capture individual head​ geometry 2) bring robustness​‌ to extreme noise and,​​ 3) enable good age​​​‌ prediction from clinical 20-channel​ EEG (n=1300) with performance​‌ close to 306-channel lab​​ MEG.

Mind will extend​​​‌ this line of research​ by translating EEG-based brain​‌ age measures into the​​ hospital setting and probe​​​‌ these in different patient​ populations in which ageing-related​‌ differences in brain structure​​ and function are part​​​‌ of the clinical picture,​ e.g., neurodevelopmental disorders, postoperative​‌ cognitive decline and dementia​​ (cf. MIND:subsec:MCE).

4.1.5 Proxy​​​‌ measures of mental health​ beyond brain aging

Quantitative​‌ measures of mental health​​ remain challenging despite substantial​​​‌ research efforts 137.​ Mental health, can only​‌ be probed indirectly through​​ psychological constructs, e.g. intelligence​​​‌ or anxiety gauged by​ valid and statistically relevant​‌ questionnaires or structured examinations​​ by a specialist. In​​​‌ practice, full neuropsychological evaluation​ is not an automated​‌ process but relies on​​ expert judgment to confront​​​‌ multiple responses and interpret​ them in the context​‌ of a larger environmental​​ context including the cultural​​​‌ background of the participant.​ Inspired by brain age,​‌ we set out to​​ build empirical measures of​​​‌ mental health 109 by​ predicting traditional and broadly​‌ used psychological constructs such​​ as fluid intelligence or​​​‌ neuroticism in the UK​ Biobank. Our results have​‌ shown that all proxies​​ captured the target constructs​​​‌ and were more useful​ than the original measures​‌ for characterizing real-world health​​ behavior (sleep, exercise, tobacco,​​​‌ alcohol consumption). In the​ long run, we anticipate​‌ that using proxies could​​ complement psychometric assessments by​​​‌ corroborating data and potentially​ providing more accurate data​‌ faster and more efficiently​​ for clinical populations.

Mind​​​‌ will expand this line​ of research by systematically​‌ searching for proxy measures​​ of physical and mental​​​‌ health derived from large​ clinical population using electronic​‌ health records or transcripts​​ from clinical interviews. We​​​‌ will propose a systematic​ causal analysis (treatment effect​‌ size and mediation) to​​ provide a clearer understanding​​​‌ of the relationships between​ the many variables that​‌ characterize mental health. We​​ will study more in​​​‌ detail the impact of​ general health markers on​‌ brain status, as this​​ may well fit much​​​‌ of the unexplained variance​ on brain health.

4.2.1 Problem statement​​

Cognitive science and psychiatry​​​‌ aim at describing mental‌ operations: cognition, emotion, perception‌​‌ and their dysfunction. As​​ an investigation device, they​​​‌ use functional brain imaging,‌ that provides a unique‌​‌ window to bridge these​​ mental concepts to the​​​‌ brain, neural firing and‌ wiring. Yet aggregating results‌​‌ from experiments probing brain​​ activity into a consistent​​​‌ description faces the roadblock‌ that cognitive concepts and‌​‌ brain pathologies are ill-defined​​. Separation between them​​​‌ is often blurry. In‌ addition, these concepts (a.k.a.‌​‌ psychological constructs) may​​ not correspond to actual​​​‌ brain structures or systems.‌ To tackle this challenge,‌​‌ we propose to leverage​​ rapidly increasing data sources:​​​‌ text and brain locations‌ described in neuroscientific publications,‌​‌ brain images and their​​ annotations taken from public​​​‌ data repositories, and several‌ reference datasets.

4.2.2 What‌​‌ machine learning can do​​ for neuroscience

Recent works​​​‌ in computer vision 113‌ or natural language processing‌​‌ 107, 114 have​​ tackled predictions on a​​​‌ large number of classes,‌ getting closer to open-ended‌​‌ knowledge. These approaches, that​​ rely on uncovering some​​​‌ form of relational structure‌ across these classes, in‌​‌ effect capture the semantics​​ of the domain 107​​​‌, including the similarity‌ structure of the relevant‌​‌ classes and the ambiguities​​ across classes or the​​​‌ multiple aspects of a‌ class. Broadly speaking, these‌​‌ contributions converge to the​​ concept of representation learning​​​‌ 89, i.e. estimating‌ latent factors that reformulate‌​‌ a learning problem into​​ a new set of​​​‌ input features or output‌ classes that are more‌​‌ natural for the data​​ and help further analysis.​​​‌ These new tools enable‌ extraction of knowledge, for‌​‌ instance ontology induction, with​​ statistical learning 160.​​​‌ They are at the‌ root of heterogeneous data‌​‌ integration, such as multi-modal​​ machine learning 86.​​​‌ The machine learning challenges‌ that we aim to‌​‌ tackle are three-fold:

• Existing​​ multi-modal machine learning techniques​​​‌ have been developed for‌ relatively abundant data, with‌​‌ overall high SNR: text,​​ natural images, videos, sound.​​​‌ These data are most‌ often non-ambiguous, while brain‌​‌ data typically are, due​​ to the low SNR​​​‌ per image and, more‌ crucially, poor annotation quality‌​‌. We propose to​​ tackle this by adapting​​​‌ machine learning solutions to‌ this low-SNR regime: introduction‌​‌ of priors, aggressive dimension​​ reduction, aggregation approaches and​​​‌ data augmentation to reduce‌ overfitting.
• Leveraging implicit supervisory‌​‌ signals: While data​​ sources contain lots of​​​‌ implicit information that could‌ be used as targets‌​‌ in supervised learning, there​​ is most often no​​​‌ obvious way to extract‌ it. We propose to‌​‌ tackle this by using​​ additional, ill- or not-annotated​​​‌ data, relying on self-supervision‌ methods.
• Model interpretability‌​‌: Our goal is​​ to provide clear assertions​​​‌ on the relationships between‌ brain structures and cognition:‌​‌ the inference should always​​ lead to an updated​​​‌ knowledge base, i.e. updated‌ relationships between concepts pertaining‌​‌ to neuroscience on one​​ hand, psychology on the​​​‌ other hand. Specifically, one‌ should be able to‌​‌ reason about the information​​ extracted within Mind.​​​‌ For this, we will‌ develop dedicated statistical, causal‌​‌ and formal (ontology-based) data​​​‌ analysis schemes.

Associating knowledge​ engineering with statistical learning​‌ to boost cognitive neuroimaging,​​ requires tackling the challenge​​​‌ of multimodal machine learning​ under noisy conditions with​‌ limited data. Doing so,​​ it will capture links​​​‌ between behavior and brain​ activity, and enable aggregating​‌ the information carried by​​ neuroimaging data to redefine​​​‌ and link concepts in​ psychology and psychiatry.

4.2.3​‌ Perspective taken: combine distributional​​ semantics with brain images​​​‌

In natural language processing​ (NLP), distributional semantics capture​‌ meanings of words using​​ similarities in the way​​​‌ they appear in their​ environment. We want to​‌ adapt these ideas to​​ learn data-driven organizations of​​​‌ psychological concepts. Importantly, applying​ these techniques solely to​‌ the psychology literature merely​​ captures the current status​​​‌ quo of the field.​ Including brain images is​‌ necessary to bring new​​ information.

To link observed​​​‌ cognition to brain activity,​ two typical statistical learning​‌ problems arise: encoding,​​ that seeks to describe​​​‌ brain activity from behavior;​ and decoding, that​‌ seeks the converse, predicting​​ behavior from brain activity​​​‌ 138. In addition,​ statistical modeling of each​‌ aspect of the data​​ on its own generates​​​‌ knowledge, typically spatial decompositions​ from resting-state data, and​‌ topic modeling on descriptions​​ of behavior. The research​​​‌ strategy followed in this​ proposal is to combine​‌ the different statistical learning​​ problems in a unified​​​‌ framework to extract core​ structures from the aggregation​‌ of neuroimaging data: on​​ one side brain structures,​​​‌ and on the other​ side semantic relationships and​‌ concepts in psychological sciences.​​

Mind will in particular​​​‌ publish automated functional meta-analyses​ to give a systematic​‌ assessment of the publicly​​ available data and question​​​‌ the limitations of the​ current conceptual framework of​‌ systems neuroscience as well​​ as of these resources.​​​‌

4.3.1 MRI-based modeling​​​‌ of clinical endpoints

Image​ based biomarkers can be​‌ objectively measured and are​​ a sign of normal​​​‌ or abnormal processes, of​ a condition or disease.​‌ Incorporating new potential imaging​​ biomarkers requires several steps,​​​‌ often in parallel and​ complementary to each other,​‌ to be undertaken for​​ translation into clinical practice.​​​‌ These can be divided​ into the following phases​‌ after identification: Development and​​ evaluation, validation, implementation, qualification,​​ and utilization. Our team​​​‌ aims to cross two‌ main translational gaps, that‌​‌ is, the translation from​​ patients first and then​​​‌ to practice. Our aim‌ through our current and‌​‌ active projects is to​​ ensure that potential biomarkers,​​​‌ like the clear delineation‌ of subterritories of the‌​‌ subthalamic nucleus (STN) in​​ pharmaco-resistant Parkinson's disease (PD)​​​‌ patients (i.e.candidates for implantation‌ of a deep brain‌​‌ stimulator) are `fit for​​ purpose' and associated with​​​‌ the clinical endpoint of‌ interest with the overarching‌​‌ goal being to demonstrated​​ efficacy and health impact.​​​‌ This process is key‌ to the translation into‌​‌ clinical practice and widespread​​ utilization.

Through the ANR​​​‌ VLFMRI grant we aim‌ to derive new MR‌​‌ imaging-based biomarkers related to​​ prematurity and abnormal neurodevelopment​​​‌ of hospitalized neonates at‌ low magnetic field (20‌​‌ mTesla). In this setup,​​ the objective is to​​​‌ perform an almost continuous‌ monitoring to detect early‌​‌ signs of adverse events​​ including ischemic stroke or​​​‌ encephalopathy (collaboration with Prof.‌ V. Biran, APHP Robert‌​‌ Debré Hospital). An additional​​ collaboration is already underway​​​‌ with the AP-HP Henri‌ Mondor Hospital (neuroradiologist Dr‌​‌ B. Bapst, doing part​​ of her PhD at​​​‌ NeuroSpin), to achieve high-resolution‌ susceptibility weighted imaging (600‌​‌ µisotropic) in a scan​​ time of 2m30s for​​​‌ an accurate delineation of‌ the STN in PD‌​‌ patients prior to surgical​​ planning. A database of​​​‌ 123 patients has already‌ been collected using both‌​‌ the standard SWI imaging​​ protocol and ours based​​​‌ on the SPARKLING technology.‌ This annotated database will‌​‌ be key to compare​​ the diagnosis power of​​​‌ our solution with that‌ of the current care,‌​‌ analyse to what extent​​ a higher image resolution​​​‌ is instrumental in providing‌ a more accurate clinical‌​‌ diagnostic, and finally make​​ our protocol more widely​​​‌ accepted in the clinical‌ practice.

Our key contribution‌​‌ in these projects is​​ to translate to the​​​‌ clinical realm both the‌ SPARKLING technology on the‌​‌ acquisition side 141,​​ 98 as well as​​​‌ our PySAP software 122‌ for MR image reconstruction.‌​‌ In this regard, the​​ recently accepted CEA postdoc​​​‌ funding should help us‌ move the technology to‌​‌ clinical 7T MR Systems​​ (Magnetom Terra Siemens-Healthineers) in​​​‌ the University hospital of‌ Poitiers through a nascent‌​‌ collaboration with Prof. Rémy​​ Guillevin. Their interest is​​​‌ to use the high-resolution‌ SPARKLING SWI protocol at‌​‌ 7T to better delineate​​ the anomalies along the​​​‌ central vein for the‌ diagnostic of multiple sclerosis‌​‌ as the number of​​ anomalies predicts the grade/severity​​​‌ of this inflammatory pathology.‌ On a longer perspective,‌​‌ we aim to generalize​​ the use of our​​​‌ recently DL networks for‌ MR image reconstruction 171‌​‌, 169 to multiple​​ acquisition setups and other​​​‌ downstream tasks (e.g. motion‌ correction and correction of‌​‌ off-resonance artifacts related to​​ $B_0$ inhomogeneities).

6.1.1 MNE​​​‌

• Name:
  MNE-Python
• Keywords:
  Neurosciences,​ EEG, MEG, Signal processing,​‌ Machine learning
• Functional Description:​​
  Open-source Python software for​​​‌ exploring, visualizing, and analyzing​ human neurophysiological data: MEG,​‌ EEG, sEEG, ECoG, and​​ more.
• Release Contributions:
  https://mne.tools/stable/whats_new.html​​
• URL:
  https://­mne.­tools/
• Contact:
  Alexandre​​​‌ Gramfort
• Partners:
  HARVARD Medical‌ School, New York University,‌​‌ University of Washington, CEA,​​ Aalto university, Telecom Paris,​​​‌ Boston University, UC Berkeley,‌ Macquarie University, University of‌​‌ Oregon, Aarhus University

6.1.2​​ NeuroLang

• Name:
  NeuroLang
• Keywords:​​​‌
  Neurosciences, Probabilistic Programming, Logic‌ programming
• Functional Description:
  NeuroLang‌​‌ is a probabilistic logic​​ programming system specialised in​​​‌ the analysis of neuroimaging‌ data, but not exclusively‌​‌ determined by it.
• Release​​ Contributions:
  https://neurolang.github.io/
• URL:
  https://­neurolang.­github.­io/­#​​​‌
• Contact:
  Demian Wassermann

6.1.3‌ Nilearn

• Name:
  NeuroImaging with‌​‌ scikit learn
• Keywords:
  Health,​​ Neuroimaging, Medical imaging
• Functional​​​‌ Description:
  NiLearn is the‌ neuroimaging library that adapts‌​‌ the concepts and tools​​ of scikit-learn to neuroimaging​​​‌ problems. As a pure‌ Python library, it depends‌​‌ on scikit-learn and nibabel,​​ the main Python library​​​‌ for neuroimaging I/O. It‌ is an open-source project,‌​‌ available under BSD license.​​ The two key components​​​‌ of NiLearn are i)‌ the analysis of functional‌​‌ connectivity (spatial decompositions and​​ covariance learning) and ii)​​​‌ the most common tools‌ for multivariate pattern analysis.‌​‌ A great deal of​​ efforts has been put​​​‌ on the efficiency of‌ the procedures both in‌​‌ terms of memory cost​​ and computation time.
• Release​​​‌ Contributions:

  HIGHLIGHTS - Updated‌ docs with a new‌​‌ theme using furo. -​​ permuted_ols and non_parametric_inference now​​​‌ support TFCE statistic. -‌ permuted_ols and non_parametric_inference now‌​‌ support cluster-level Family-wise error​​ correction. - save_glm_to_bids has​​​‌ been added, which writes‌ model outputs to disk‌​‌ according to BIDS convention.​​

  NEW - save_glm_to_bids has​​​‌ been added, which writes‌ model outputs to disk‌​‌ according to BIDS convention.​​ - permuted_ols and non_parametric_inference​​​‌ now support TFCE statistic.‌ - permuted_ols and non_parametric_inference‌​‌ now support cluster-level Family-wise​​ error correction. - Updated​​​‌ docs with a new‌ theme using furo.

  See‌​‌ all details in https://nilearn.github.io/stable/changes/whats_new.html​​

• URL:
  http://­nilearn.­github.­io/
• Contact:
  Bertrand​​​‌ Thirion
• Participants:
  Pierre-Louis Barbarant,‌ Remi Gau, Himanshu Aggarwal,‌​‌ Bertrand Thirion, Gael Varoquaux​​

6.1.4 Benchopt

• Keywords:
  Benchmarking,​​​‌ Machine learning, Optimization
• Functional‌ Description:

  BenchOpt is a‌​‌ package to simplify, make​​ more transparent and more​​​‌ reproducible the comparisons of‌ optimization algorithms. It is‌​‌ written in Python but​​ it is available with​​​‌ many programming languages. So‌ far it has been‌​‌ tested with Python, R,​​ Julia and compiled binaries​​​‌ written in C/C++ available‌ via a terminal command.‌​‌ If it can be​​ installed via conda, it​​​‌ should just work!

  BenchOpt‌ is used through a‌​‌ simple command line and​​ ultimately running and replicating​​​‌ an optimization benchmark should‌ be as easy a‌​‌ cloning a repo and​​ launching the computation with​​​‌ a single command line.‌ For now, BenchOpt features‌​‌ benchmarks for around 10​​ convex optimization problems and​​​‌ we are working on‌ expanding this to feature‌​‌ more complex optimization problems.​​ We are also developing​​​‌ a website to display‌ the benchmark results easily.‌​‌

• Release Contributions:
  https://github.com/benchopt/benchopt/releases/tag/1.8.0
• News​​ of the Year:
  Building​​​‌ on the momentum of‌ the 2024 sprint, Benchopt’s‌​‌ development in 2025 focused​​ on maturing its infrastructure​​​‌ for large-scale machine learning‌ and industrial-grade workflows. A‌​‌ key highlight was the​​ "supercharging" of the parallel​​​‌ backend, particularly through the‌ SLURM executor, which now‌​‌ supports solver-specific overrides (allowing,​​​‌ for instance, GPU allocation​ only for specific methods)​‌ and better management of​​ parallel job limits. Machine​​​‌ learning support was significantly​ bolstered by the introduction​‌ of the collect command​​ for gathering distributed results,​​​‌ improved serialization for complex​ parameter spaces, and the​‌ integration of high-profile benchmarks​​ like NanoGPT and SLOPE.​​​‌ Furthermore, the framework moved​ toward a more robust​‌ ecosystem with the release​​ of version 1.7.0, featuring​​​‌ enhanced HTML result pages​ with interactive tables, improved​‌ Windows compatibility, and the​​ ability for objectives to​​​‌ produce multiple entries simultaneously—solidifying​ Benchopt as a comprehensive​‌ tool for both classical​​ optimization and modern deep​​​‌ learning research.
• Publication:
  hal-03830604​
• Contact:
  Thomas Moreau
• Participants:​‌
  Thomas Moreau, Mathurin Massias,​​ Hippolyte Verninas, Melvine Nargeot,​​​‌ Jad Yehya

6.1.5 Scikit-learn​

• Keywords:
  Clustering, Classification, Regression,​‌ Machine learning
• Scientific Description:​​
  Scikit-learn is a Python​​​‌ module integrating classic machine​ learning algorithms in the​‌ tightly-knit scientific Python world.​​ It aims to provide​​​‌ simple and efficient solutions​ to learning problems, accessible​‌ to everybody and reusable​​ in various contexts: machine-learning​​​‌ as a versatile tool​ for science and engineering.​‌
• Functional Description:

  Scikit-learn can​​ be used as a​​​‌ middleware for prediction tasks.​ For example, many web​‌ startups adapt Scikitlearn to​​ predict buying behavior of​​​‌ users, provide product recommendations,​ detect trends or abusive​‌ behavior (fraud, spam). Scikit-learn​​ is used to extract​​​‌ the structure of complex​ data (text, images) and​‌ classify such data with​​ techniques relevant to the​​​‌ state of the art.​

  Easy to use, efficient​‌ and accessible to non​​ datascience experts, Scikit-learn is​​​‌ an increasingly popular machine​ learning library in Python.​‌ In a data exploration​​ step, the user can​​​‌ enter a few lines​ on an interactive (but​‌ non-graphical) interface and immediately​​ sees the results of​​​‌ his request. Scikitlearn is​ a prediction engine .​‌ Scikit-learn is developed in​​ open source, and available​​​‌ under the BSD license.​

• URL:
  http://­scikit-learn.­org
• Publications:
  hal-00650905​‌, hal-00856511, hal-01093971​​
• Contact:
  Gael Varoquaux
• Participants:​​​‌
  Thomas Moreau, Jerome Dockes,​ Alexandre Gramfort, Bertrand Thirion,​‌ Gael Varoquaux, Loic Esteve,​​ Olivier Grisel, Guillaume Lemaitre,​​​‌ Jeremie Du Boisberranger, Julien​ Jerphanion
• Partners:
  Axa, BNP​‌ Parisbas Cardif, Dataiku, Nvidia,​​ Chanel, Probabl

6.1.6 joblib​​​‌

• Keywords:
  Parallel computing, Cache​
• Functional Description:
  Facilitate parallel​‌ computing and caching in​​ Python.
• URL:
  https://­joblib.­readthedocs.­io/­en/­latest/
• Contact:​​​‌
  Thomas Moreau
• Participants:
  Thomas​ Moreau, Loic Esteve, Olivier​‌ Grisel, Gael Varoquaux, Yoann​​ Coudert–Osmont
• Partner:
  Probabl

6.1.7​​​‌ MRI-NUFFT

• Keywords:
  Brain MRI,​ NUFFT, Trajectory Generation
• Functional​‌ Description:

  MRI-NUFFT is a​​ python package that extends​​​‌ various NUFFT (Non-Uniform Fast​ Fourier Transform) python bindings​‌ used for MRI reconstruction.​​ It provides a unified​​​‌ interface with a large​ number of backends with​‌ implementations ranging from CPU​​ to GPU.

  In particular,​​​‌ it provides a unified​ interface for all the​‌ methods, with extra features​​ such as coil sensitivity,​​​‌ density compensated adjoint and​ off-resonance corrections (for static​‌ B0 inhomogeneities). Additionally, useful​​ IO tools like reading​​​‌ a k-space sampling trajectory​ and writing a binary​‌ file for run on​​ MR scanner is also​​​‌ offered. Finally, it helps​ algorithmically speed up MR​‌ image reconstruction algorithms through​​ fast ways to estimate​​ preconditioning weights, also known​​​‌ as density compensators for‌ a given sampling pattern.‌​‌

• Release Contributions:
  MRI-NUFFT now​​ provides a physical model​​​‌ of the MRI acquisition‌ processes, including multi-coil acquisition‌​‌ and static-field inhomogeneities. This​​ model is compatible with​​​‌ the NUFFT libraries, and‌ can be used to‌​‌ simulate the acquisition of​​ MRI data, or to​​​‌ reconstruct data from a‌ given set of measurements.‌​‌ MRI-NUFFT comes with a​​ wide variety of non-Cartesian​​​‌ trajectory generation routines that‌ have been gathered from‌​‌ the literature. It also​​ provides ways to extend​​​‌ existing trajectories and export‌ them to specific formats,‌​‌ for use in other​​ toolkits and on MRI​​​‌ hardware. Finally, MRI-NUFFT provides‌ automatic differentiation for all‌​‌ NUFFT backends, with respect​​ to both gradients and​​​‌ data (image or k-space).‌ This enables efficient backpropagation‌​‌ through NUFFT operators and​​ supports research on learned​​​‌ sampling model and image‌ reconstruction network.
• URL:
  https://­mind-inria.­github.­io/­mri-nufft/‌​‌
• Contact:
  Chaithya Giliyar Radhkrishna​​

6.1.8 SPARKLING

• Name:
  Spreading​​​‌ Projection Algorithm for Rapid‌ K-space sampLING
• Keywords:
  Brain‌​‌ MRI, MRI, Optimization
• Scientific​​ Description:
  This python package​​​‌ allows us to generate‌ "SPARKLING" curves as a‌​‌ new type of non-Cartesian​​ trajectories to perform a​​​‌ more efficient sampling in‌ 2D and 3D for‌​‌ anatomical imaging while using​​ the same number of​​​‌ samples for a limited‌ time budget. These segmented‌​‌ curves are obtained using​​ a projection method on​​​‌ measure sets which offers‌ three main advantages: i)‌​‌ generating segmented Non-Cartesian trajectories​​ along a chosen density,​​​‌ ii) meeting the hardware‌ constraints on the magnetic‌​‌ field gradients (magnitude, slew​​ rate), iii) performing a​​​‌ fast coverage of k-space.‌
• Functional Description:
  This python‌​‌ package implements "SPARKLING": an​​ optimization driven method to​​​‌ obtain hardware compliant sampling‌ curves that globally satisfy‌​‌ a user specified target​​ sampling density. The resulting​​​‌ non-cartesian sampling curves can‌ be used to efficiently‌​‌ undersample and speed up​​ acquisitions on an MR​​​‌ scanner. This method is‌ generic enough that it‌​‌ can be applied to​​ any of the imaging​​​‌ modalities in MR.
• Publications:‌
  hal-01577200v1, hal-02067080v1,‌​‌ hal-03090471v2, hal-04370475v1,​​ hal-02361265v1
• Contact:
  Chaithya Giliyar​​​‌ Radhkrishna

6.1.9 PySAP

• Name:‌
  Python Sparse data Analysis‌​‌ Package
• Keywords:
  Image reconstruction,​​ Image compression
• Functional Description:​​​‌

  The PySAP (Python Sparse‌ data Analysis Package, https://github.com/CEA-COSMIC/pysap)‌​‌ open-source image processing software​​ package has been developed​​​‌ for the 3 years‌ between the Compressed Sensing‌​‌ group at Iniria-CEA Parietal​​ team led by Philippe​​​‌ Ciuciu and the CosmoStat‌ team (CEA/IRFU) led by‌​‌ Jean-Luc Statck. It has​​ been developed for the​​​‌ COmpressed Sensing for Magnetic‌ resonance Imaging and Cosmology‌​‌ (COSMIC) project. This package​​ provides a set of​​​‌ flexible tools that can‌ be applied to a‌​‌ variety of compressed sensing​​ and image reconstruction problems​​​‌ in various research domains.‌ In particular, PySAP offers‌​‌ fast wavelet transforms and​​ a range of integrated​​​‌ optimisation algorithms. It also‌ offers a variety of‌​‌ plugins for specific application​​ domains: on top of​​​‌ Pysap-MRI and PySAP-astro plugins,‌ several complementary modules are‌​‌ now in development for​​ electron tomography and electron​​​‌ microscopy for CEA colleagues.‌ In October 2019, PySAP‌​‌ has been released on​​​‌ PyPi (https://pypi.org/project/python-pySAP/, currently version​ 0.0.3) and in conda​‌ (https://anaconda.org/agrigis/python-pysap).

  The Pysap-MRI has​​ been advertised through a​​​‌ specific abstract accepted to​ the next workshop of​‌ ISMRM on Data Sampling​​ & Image Reconstruction in​​​‌ late January 2020. It​ will be presented during​‌ a power pitch session​​ together wih an hands-on​​​‌ demo session using JuPyter​ notebooks.

• Contact:
  Philippe Ciuciu​‌
• Partner:
  CEA

6.1.10 SNAKE​​

• Name:
  Simulator from neuro-activation​​​‌ to K-space Exploration
• Keywords:​
  FMRI, NUFFT
• Functional Description:​‌
  We propose a new,​​ modular, open-source, Python-based 3D+time​​​‌ fMRI data simulation software,​ SNAKE-fMRI, which stands for​‌ Simulator from Neurovascular coupling​​ to Acquisition of K-space​​​‌ data for Exploration of​ fMRI acquisition techniques. Unlike​‌ existing tools, the goal​​ here is to simulate​​​‌ the complete chain of​ fMRI data acquisition, from​‌ the spatio-temporal design of​​ evoked brain responses to​​​‌ various multi-coil k-space data​ 3D sampling strategies, with​‌ the possibility of extending​​ the forward acquisition model​​​‌ to various noise and​ artifact sources while remaining​‌ memory-efficient. By using this​​ in silico setup, we​​​‌ are thus able to​ provide realistic and reproducible​‌ ground truth for fMRI​​ reconstruction methods in 3D​​​‌ accelerated acquisition settings and​ explore the influence of​‌ critical parameters, such as​​ the acceleration factor and​​​‌ signal-to-noise ratio (SNR), on​ downstream tasks of image​‌ reconstruction and statistical analysis​​ of evoked brain activity.​​​‌ We present three scenarios​ of increasing complexity to​‌ showcase the flexibility, versatility,​​ and fidelity of SNAKE-fMRI:​​​‌ From a temporally-fixed full​ 3D Cartesian to various​‌ 3D non-Cartesian sampling patterns,​​ we can compare —​​​‌ with reproducibility guarantees —​ how experimental paradigms, acquisition​‌ strategies and reconstruction methods​​ contribute and interact together,​​​‌ affecting the downstream statistical​ analysis.
• URL:
  https://­paquiteau.­github.­io/­snake-fmri/
• Contact:​‌
  Philippe Ciuciu

8.1.1 Siemens Healthineers​​ & AI lab (Princeton,​​​‌ USA)

Since Fall 2019,‌ Philippe Ciuciu has actively‌​‌ collaborated with the Siemens-Healthineers​​ AI lab, led by​​​‌ Mariappan Nadar. In this‌ context, a new CIFRE‌​‌ PhD student, Mrs Asma​​ Tanabene, has joint the​​​‌ MIND team as a‌ CIFRE PhD student under‌​‌ their joint supervision to​​ work on 3D sclalable​​​‌ deep learning architecture for‌ high-resolution multicoil MR image‌​‌ reconstruction at 3 Tesla​​ and beyond. On top​​​‌ of the PhD funding,‌ this contract has generated‌​‌ 50k€ for Mind,​​ a grant that is​​​‌ hosted at CEA/NeuroSpin.

8.1.2‌ Saint Gobain Research (SGR)‌​‌

There is currently a​​ consulting contract between SGR​​​‌ and Mind (Thomas Moreau)‌ to provide an expertise‌​‌ in machine learning to​​ process temporal data, numerical​​​‌ optimization and scientific computing.‌ The expertise is provided‌​‌ one half-day per month,​​​‌ in SGR offices, and​ it consists in scientific​‌ discussion sessions on the​​ ML projects leaded by​​​‌ SGR data scientists.

8.1.3​ Apple

A research collaboration​‌ between Apple (Pierre Ablin)​​ and Mind (Thomas Moreau)​​​‌ has been established. The​ goal is to develop​‌ research ideas on bilevel​​ optimization, in particular for​​​‌ the problem of data​ curation for the training​‌ of large models with​​ massive and heterogeneous datasets.​​​‌ This collaboration is being​ funded by Apple (150k€)​‌ and allowed to hire​​ a post-doc (Baptiste Goujaud).​​​‌

8.2.1 Google

Mind​‌ (Thomas Moreau) received a​​ 30k€ donation from Google​​​‌ to support its open​ source activity around benchopt​‌. In particular, the​​ grant aims to support​​​‌ the organization of coding​ and benchmarking sprints around​‌ benchopt, the development of​​ visualization tools, and the​​​‌ benchmarking of bilevel solvers,​ in particular the ones​‌ using jaxopt.

9.1.1 Inria​ associate team not involved​‌ in an IIL or​​ an international program

9.2.1 Horizon Europe

CEA Audace at-risk research​​ program

• Title:
  BrainSync
• Duration:​​​‌
  11/2024 -> 04/2029
• Coordinator:​
  Philippe Ciuciu (CEA/DRF/JOLIOT/NEUROSPIN/MIND), Saclay​‌
• Partners:
  • CEA/DRF/JOLIOT/NEUROSPIN (BAOBAB, UNICOG),​​ Saclay
  • CEA/DRT/LETI/CLINATEC, Grenoble
  • CEA/DRT/LIST/DSCIN,​​​‌ Grenoble
  • Inria-CEA MIND, Palaiseau​
  • APHP, FHU NeuroVasc, Paris-Nord​‌
  • GHU Paris Neuroscience &​​ Psychiatry
  • CHU Grenoble-Alpes
  • LPNC,​​​‌ University Grenoble-Alphes & CNRS​
  • CHU St Etienne
• Inria​‌ contact:
  Philippe Ciuciu
• Summary:​​

  The project aims to​​ understand the learning mechanisms​​​‌ that enable the human‌ brain to adapt to‌​‌ cognitive demand via the​​ flexible recruitment of different​​​‌ regions and connections, thus‌ enabling exploration of an‌​‌ uncertain and changing environment.​​ Understanding these mechanisms in​​​‌ healthy subjects, coupled with‌ the creation of a‌​‌ high-resolution anatomical-functional digital brain​​ atlas of a cohort​​​‌ of post-acute stroke patients‌ (MOTIF-STROKE clinical trial), will‌​‌ enable us to propose​​ AI models predictive of​​​‌ upper limb motor recovery‌ in these patients, and‌​‌ therapeutic innovations (use of​​ neuroprostheses) for a small​​​‌ number of them in‌ a second clinical trial.‌​‌ Clinical evaluation of relearning​​ processes will be based​​​‌ on the use of‌ incremental and adaptive artificial‌​‌ intelligence (AI) algorithms linked​​ to neuroplasticity processes.

  This​​​‌ project has received a‌ funding of 5M€ for‌​‌ a period of 4.5​​ years starting in November​​​‌ 2024 with a go/no-go‌ positioned after a period‌​‌ of 30 months (04/30/2027).​​

CEA BlueSky project

• Title:​​​‌
  Brain & Computers
• Duration:‌
  2023 -> 2026
• Coordinator:‌​‌
  Philippe Ciuciu (CEA/DRF/JOLIOT/NEUROSPIN/MIND), Saclay​​
• Partners:
  • CEA/DRF/JOLIOT/NEUROSPIN (BAOBAB, UNICOG),​​​‌ Saclay
  • CEA/DRT/LETI/CLINATEC, Grenoble
  • CEA/DRT/LIST/DSCIN,‌ Grenoble
• Inria contact:
  Philippe‌​‌ Ciuciu
• Summary:

  Artificial Intelligence​​ (AI) is now capable​​​‌ of approaching human performance‌ in tasks such as‌​‌ visual recognition, classification (i.e.,​​ decision-making), and even textual​​​‌ or visual production (e.g.,‌ GPT-4). Understanding the human‌​‌ brain mechanisms of learning​​ and decision-making in an​​​‌ uncertain environment remains a‌ major scientific challenge in‌​‌ neuroscience. This understanding will​​ enable the development of​​​‌ AI architectures that replicate‌ brain circuits. Additionally, in‌​‌ clinical applications, it can​​ lead to a generational​​​‌ leap in the design‌ of neuroprostheses. These neuroprostheses‌​‌ hold great promise for​​ improving the quality of​​​‌ life for individuals affected‌ by spinal cord injuries‌​‌ (approximately 30% of cases).​​

  The two main objectives​​​‌ of this BlueSky project‌ complement each other. On‌​‌ one hand, in healthy​​ subjects, the goal is​​​‌ to design computational models‌ based on AI that‌​‌ encode these cerebral functions​​ to gain a more​​​‌ precise understanding of the‌ associated brain activity. On‌​‌ the other hand, the​​ objective is to enable​​​‌ a greater number of‌ patients to control neuroprostheses‌​‌ autonomously through AI, making​​ these medical devices more​​​‌ widely accepted as a‌ therapeutic solution for motor‌​‌ rehabilitation. Tackling such a​​ challenge is not without​​​‌ risks and requires expanding‌ knowledge in neuroscience, surpassing‌​‌ current technological and clinical​​ limits. It also calls​​​‌ for strong synergies among‌ the various CEA institutes‌​‌ involved (Joliot, LETI, and​​ LIST) and their academic​​​‌ (Inserm, Inria, Universities Paris-Saclay,‌ and Grenoble Alpes) and‌​‌ clinical partners (CHU Grenoble-Alpes).​​

  This project has received​​​‌ a funding of 1.5M€‌ for the 2023-2025 period.‌​‌

ANR DARLING

• Title:
  DARLING:​​ Distributed adaptation and learning​​​‌ over graph signals
• Duration:‌
  2020 -> 2025 (extended)‌​‌
• Coordinator:
  Cédric Richard (cedric.richard@unice.fr),Professor​​ 3IA Senior Chair in​​​‌ UCA
• Partners:
  • Université Côte‌ d'Azur Nice, France
  • CNRS,‌​‌ École Normale Supérieure, Lyon,​​ France
  • Gipsa-lab, UMR 5216,​​​‌ CNRS, UGA, Grenoble, France‌
  • CentraleSupélec, University of Paris-Saclay,‌​‌ Gif-sur-yvette, France
• Inria contact:​​
  Philippe Ciuciu
• Summary:
  The​​​‌ DARLING project will aim‌ to propose new adaptive‌​‌ learning methods, distributed and​​​‌ collaborative on large dynamic​ graphs in order to​‌ extract structured information of​​ the data flows generated​​​‌ and/or transiting at the​ nodes of these graphs.​‌ In order to obtain​​ performance guarantees, these methods​​​‌ will be systematically accompanied​ by an in-depth study​‌ of random matrix theory.​​ This powerful tool, never​​​‌ exploited so far in​ this context although perfectly​‌ suited for inference on​​ random graphs, will thereby​​​‌ provide even avenues for​ improvement. Finally, in addition​‌ to their evaluation on​​ public data sets, the​​​‌ methods will be compared​ with each other using​‌ two advanced imaging techniques​​ in which two of​​​‌ the partners are involved:​ radio astronomy with the​‌ giant SKA instrument (Obs.​​ Côte d'Azur) and MEG​​​‌ brain imaging (Inria MIND​ at NeuroSpin, CEA Saclay).​‌ Sheng Wang as a​​ postdoc in MIND and​​​‌ Merlin Dumeur as a​ MIND PhD student in​‌ co-tutelle with Matias Palva​​ from Aalto University, Finland​​​‌ are actually involved in​ the processing of MEG​‌ and S/EEG time series​​ on graphs, notably to​​​‌ analyze scale-free (i.e. critical​ and bistability) phenomena across​‌ these graphs and extract​​ potentially new biomarkers for​​​‌ characterizing the pathophysiology of​ epileptogenic zone (EZ) in​‌ drug resistant epilepsy.

ANR​​ VLFMRI

• Title:
  VLFMRI: Very​​​‌ low field MRI for​ babies
• Duration:
  2021 ->​‌ 2025
• Coordinator:
  Claude Fermon​​ (CEA Saclay, DRF/IRAMIS/SPECT)
• Partners:​​​‌
  • CEA/SHFJ/BIOMAPS, Orsay, France
  • CEA/NeuroSpin,​ Gif-sur-Yvette, France
  • APHP Robert​‌ Debré hospital, Paris, France​​
  • APHP Bicêtre hospital, Kremlin-Bicêtre,​​​‌ France
• Inria contact:
  Philippe​ Ciuciu
• Summary:
  VLFMRI aims​‌ at developing a very​​ low-field Magnetic Resonance Imaging​​​‌ (MRI) system as an​ alternative to conventional high-field​‌ MRI for continuous imaging​​ of premature newborns to​​​‌ detect hemorrhages or ischemia.​ This system is based​‌ on a combination of​​ a new generation of​​​‌ magnetic sensors based on​ spin electronics, optimized MR​‌ acquisition sequences (based on​​ the SPARKLING patent, Inria-CEA​​​‌ MIND team at NeuroSpin)​ and a open and​‌ compatible system with an​​ incubator that will allow​​​‌ to achieve an image​ resolution of 1mm${}^{\mathrm{3‌}}$ on a whole baby​​ body in a short​​​‌ scan time. This project​ is a partnership of​‌ three academic partners and​​ two hospital departments. The​​​‌ different stages of the​ project are the finalization​‌ of the hardware development​​ and software system, preclinical​​​‌ validation on small animals​ and clinical validation. Kumari​‌ Pooja has been hired​​ in January 2022 as​​​‌ research engineer in MIND​ to interact with the​‌ coordinator of this ANR​​ project, Claude Fermon and​​​‌ design new accelerated acquisition​ methods for verly low​‌ field MRI. Preliminary encouraging​​ results allow us to​​​‌ retrospectively accelerate MRI acquisition​ by a factor of​‌ 10 without degrading image​​ quality at 2mm isotropic​​​‌ resolution.

ANR MICBrainPres

• Title:​
  MicBrainPres: Distributed adaptation and​‌ learning over graph signals​​
• Duration:
  2023 -> 2026​​​‌
• Coordinator:
  Demian Wassermann
• Partners:​
  • Brain and Spine Institute,​‌ Paris, France
  • CNRS, Université​​ de Paris, Paris, France​​​‌
• Inria contact:
  Demian Wassermann​
• Summary:
  The main goal​‌ of this project is​​ to harness the latest​​​‌ advances on machine learning-based​ neuroimage processing technologies to​‌ improve function-preserving brain tumour​​ resection. Identifying eloquent brain​​ regions is fundamental to​​​‌ performing tumour resection while‌ preserving a maximum level‌​‌ of cognitive function. Despite​​ the sustained advance in​​​‌ predicting subject-level cognitive abilities‌ from neuroimaging data, current‌​‌ approaches lack sensitivity and​​ specificity in identifying eloquent​​​‌ brain regions. This hinders‌ neuroimaging's usefulness for pre-surgical‌​‌ planning as a tool​​ to predict the preservation​​​‌ of cognitive function after‌ tumour resection. In this‌​‌ project, we propose that​​ using subject-specific parcellations, derived​​​‌ from functional and diffusion‌ MRI through deep-learning technologies,‌​‌ will achieve the needed​​ sensitivity and specificity to​​​‌ locate eloquent areas pre-surgically‌ and to predict cortical‌​‌ reshaping after tumour resection.​​

ANR EBUL

• Title:
  EBUL:​​​‌ Event-based Unsupervised Learning for‌ Physiological Signals
• Duration:
  2023‌​‌ -> 2027
• Coordinator:
  Thomas​​ Moreau
• Partners:
  • INRIA MIND,​​​‌ Gif-sur-Yvette, France
• Inria contact:‌
  Thomas Moreau
• Summary:

  Sensor-based‌​‌ body monitoring is now​​ routine clinical care. The​​​‌ resulting records are called‌ physiological signals. While enormous‌​‌ quantities of signals are​​ collected every day, the​​​‌ cost and time necessary‌ to clean and annotate‌​‌ them is prohibitive to​​ constitute large labeled databases.​​​‌ When working with physiological‌ signals, extra sources of‌​‌ information are the events​​ surrounding the recordings. Events​​​‌ are external phenomena that‌ impact the signal and‌​‌ can correlate with the​​ prediction task considered. EBUL​​​‌ propose to develop novel‌ unsupervised learning techniques to‌​‌ process such records based​​ on the notion of​​​‌ events, and to apply‌ them to process general‌​‌ anesthesia records collected in​​ Paris hospital Lariboisière. The​​​‌ methodology of the project‌ relies on the development‌​‌ of novel point process​​ models adapted to capture​​​‌ the distribution of physiological‌ events, and their coupling‌​‌ with event detection algorithms.​​ This will provide novel​​​‌ signal representations based on‌ the distribution of events‌​‌ inside them, which are​​ simpler to analyze and​​​‌ fine tune to derive‌ predictive bio-markers.

  The EBUL‌​‌ project is organized around​​ 3 objectives:

  • Develop novel​​​‌ point process models for‌ physiological signals
  • Learn joint‌​‌ models for signals and​​ events
  • Develop high-quality models​​​‌ for general anesthesia that‌ can reach clinical research‌​‌

ANR BenchArk

• Title:
  BenchArk:​​ An efficient and robust​​​‌ benchmarking suite for AI‌
• Duration:
  2024 -> 2028‌​‌
• Coordinator:
  Thomas Moreau, Mathurin​​ Massias, Joseph Salmon
• Partners:​​​‌
  • INRIA MIND, Gif-sur-Yvette, France‌
  • INRIA OCKHAM, Lyon, France‌​‌
  • Université de Montpellier, Montpellier,​​ France
• Inria contact:
  Thomas​​​‌ Moreau
• Summary:
  Numerical evaluation‌ of novel methods, a.k.a.‌​‌ benchmarking, is a pillar​​ of the scientific method​​​‌ in machine learning. However,‌ due to practical and‌​‌ statistical obstacles, the reproducibility​​ of published results is​​​‌ currently insufficient: many details‌ can invalidate numerical comparisons,‌​‌ from insufficient uncertainty quantification​​ to improper methodology. In​​​‌ 2022, the benchopt initiative‌ provided an open source‌​‌ Python package together with​​ a framework to seamlessly​​​‌ run, reuse, share and‌ publish benchmarks in numerical‌​‌ optimization. In this project,​​ we aim at bringing​​​‌ benchopt to the whole‌ machine learning community, making‌​‌ it a new standard​​ in benchmarking by empowering​​​‌ researchers and practitioners with‌ efficient and valid benchmarking‌​‌ methods. Our goal is​​ to ensure reproducibility and​​​‌ consistency in model evaluation.‌ We will federate the‌​‌ machine learning community to​​​‌ develop informative and statistically​ valid benchmarks, while providing​‌ methods to reduce identified​​ hurdles in implementing such​​​‌ practices. The results of​ the project will be​‌ integrated in the open​​ source benchopt library.

Brain​​​‌ Health Trajectories

• Title:
  Brain​ Health Trajectories
• Duration:
  2023​‌ -> 2028
• Coordinator:
  Viktor​​ Jirsa, INS Marseille
• Partners:​​​‌
  • INSERM Délégation Provence-Alpes-Côte d'Azur​ et Corse
  • CNRS IDF​‌ Sud (Gif)
  • Université d'Aix-Marseille​​
  • CHU de Grenoble
  • CEA,​​​‌ DRF, Joliot, Neurospin
  • INSTITUT​ NATIONAL DE RECHERCHE EN​‌ INFORMATIQUE ET AUTOMATIQUE
• Inria​​ contact:
  Bertrand Thirion
• Summary:​​​‌

  This project is part​ of PEPR Santé Numérique.​‌ The Brain Health Trajectories​​ project is part of​​​‌ the PEPR Santé numérique​ program, a major research​‌ initiative of the French​​ government as part of​​​‌ the France 2030 investment​ plan.

  The project prioritizes​‌ the early identification of​​ individuals at risk of​​​‌ developing a disease, so​ they can receive appropriate​‌ treatment before the disease​​ develops or at least​​​‌ adapt their lifestyle to​ delay the onset of​‌ the pathology and slow​​ down the progression of​​​‌ impairment. It aims to​ develop tools to characterize​‌ individuals' brain health and​​ lay the foundation for​​​‌ a platform for screening,​ decision-making within the population,​‌ and prognostic monitoring of​​ treatment efficacy.

ANR VITE​​​‌

• Title:
  VITE: Explainable AI​ through Variable Importance Tests​‌
• Duration:
  2024 -> 2027​​
• Coordinator:
  Bertrand Thirion
• Partners:​​​‌
  • INRIA MIND, Gif-sur-Yvette, France​
  • Université de Toulouse, Toulouse,​‌ France
  • Université de Montpellier,​​ Montpellier, France
• Inria contact:​​​‌
  Bertrand Thirion
• Summary:
  The​ VITE project aims to​‌ develop new statistical methods​​ to interpret complex machine​​​‌ learning models, particularly in​ high-dimensional settings such as​‌ neuroimaging and genomics. The​​ focus will be on​​​‌ creating variable importance tests​ that can provide insights​‌ into the contributions of​​ individual features to model​​​‌ predictions. By leveraging recent​ advances in statistical theory​‌ and computational techniques, the​​ project seeks to enhance​​​‌ the interpretability of AI​ models, making them more​‌ transparent and trustworthy for​​ applications in healthcare and​​​‌ other critical domains.

ANR​ SPEAKOUT

• Title:
  SPEAKOUT: Statistical​‌ and machine learning methods​​ for decoding speech from​​​‌ brain signals
• Duration:
  2025​ -> 2028
• Coordinator:
  Anne-Lise​‌ Giraud
• Partners:
  • INRIA MIND,​​ Gif-sur-Yvette, France
  • Institut de​​​‌ l'audition, Paris, France
  • institut​ du cerveau, Paris, France​‌
• Inria contact:
  Bertrand Thirion​​
• Summary:
  The SPEAKOUT project​​​‌ aims to develop advanced​ statistical and machine learning​‌ methods to decode speech​​ from brain signals. By​​​‌ analyzing neural activity associated​ with speech production and​‌ perception, the project seeks​​ to create models that​​​‌ can accurately interpret and​ reconstruct spoken language from​‌ brain data. This research​​ has the potential to​​​‌ revolutionize communication for individuals​ with speech impairments, enabling​‌ them to express themselves​​ through brain-computer interfaces. The​​​‌ project will leverage cutting-edge​ techniques in signal processing,​‌ deep learning, and neuroscience​​ to achieve its goals.​​​‌

10.1.1​‌ Scientific events: organisation

• Thomas​​ Moreau:
  Co-organized the NeurIPS​​​‌ 2025 (San Diego) workshop​ on Foundation Models for​‌ Time Series, and​​ the Open Source Conference​​ In academIa on open​​​‌ source development and maintenance‌ in 2025. He was‌​‌ also a member of​​ the organizing committee of​​​‌ two international workshops at‌ CIRM, Learning and Optimization‌​‌ in Luminy 2022 and​​ 2024, and the​​​‌ co-founder and co-organizer of‌ the Séminaire Palaisien,‌​‌ a monthly seminar gathering​​ around 35 researchers from​​​‌ the Saclay area working‌ on statistics and machine‌​‌ learning since 2019.
• Bertrand​​ Thirion:
  Co-organized the CogBases​​​‌ Workshop in 2025,‌ the NeuroDecoder workshop,‌​‌ the Explanability for high-dimensional​​ statistics.

10.1.2 Journal​​​‌

10.1.3‌​‌ Invited talks

• B.Thirion

   

  • April​​ 2025: Invited talk at​​​‌ Bilkent University, Ankara, Turkey.‌
  • March 2025: Invited talk‌​‌ at the deep phenotyping​​ workshop, Paris, France.
  • January​​​‌ 2025: Invited talk at‌ the Académie des Sciences,‌​‌ Paris, France.
  • June 2025:​​ Invited talk at Sanofi,​​​‌ Gentilly, France.
  • April 2025:‌ Invited talk at Université‌​‌ Grenoble Alpes, Grenoble, France.​​
  • December 2025: Invited talk​​​‌ at the Morocco Academy‌ of Sciences, Rabat, Morocco.‌​‌
  • March 2025: presentation of​​ Nilearn development at OSCII​​​‌ workshop, Aussois, France.
  • June‌ 2025: presentation at ASIC‌​‌ workshop, Saint Gervais, France.​​
• Philippe Ciuciu:

   

  • Invited speaker​​​‌ for the 50th anniversary‌ of IRISA (Rennes, FR,‌​‌ 12/2025, .www)
  • FORTH​​ WS on Comput. Intell.​​​‌ in Imag. Inv. Prob.‌ (Heraklion, GR, 09/2025, .www‌​‌)
  • WS on Inv.​​ Prob. and AI in​​​‌ Medicine (Bath, UK, 07/2025,‌ .www)
• Thomas Moreau:‌​‌

   

  • November 2025 – Joint​​ invited talk for the​​​‌ Open Science Days and‌ journée de la recherche‌​‌ frugale, Grenoble
  • September 2025​​ – Inverse Problem and​​​‌ Imaging, CIRM, Luminy
  • June‌ 2025 – Plenary talk‌​‌ for Journée de la​​ donnée, AMDAC Ministère de​​​‌ la santé, Paris
• Demian‌ Wassermann:

   

  • 2025: Neurofunctional Imaging‌​‌ Group – Université de​​ Bordeaux 2025: Neuroscience Through​​​‌ the Lens of Machine‌ Learning
  • 2025: UCL Department‌​‌ of Computer Science (UK):​​ Linking Neuroimaging and Cognition:​​​‌ from Automated Meta Analyses‌ to Subject-Specific Prediction Through‌​‌ Probabilistic Modelling

10.1.4 Scientific​​ expertise

• B.Thirion

   

  • Expert for​​​‌ the European Research Council‌ (ERC) in the panel‌​‌ PE6
  • Member of the​​ Inria-Académie des Sciences joint​​​‌ committee for yearly prizes‌

10.1.5 Research administration

• P.‌​‌ Ciuciu

   

  • Member of the​​​‌ Board of Directors at​ NeuroSpin (CEA).
  • Member of​‌ the team leaders committee​​ at Inria Saclay Ile-de-France​​​‌
  • Member of the steering​ committee of the CEA​‌ cross-disciplinary research program on​​ numerical simulation and AI.​​​‌
• B. Thirion

   

  • Délégué Scientifique​ of Inria Saclay Center​‌ until September 2025.
  • Member​​ of ENS Paris-Saclay Scientific​​​‌ Council
  • Member of Telecom​ Sud Paris Scientific Council​‌
  • Member of Inria Commission​​ évaluation
• D. Wassermann

   

  • Deputy​​​‌ Director of the DataIA​ Institute (Université Paris Saclay)​‌

10.2.1 Supervision

• P.​​ Ciuciu

   

  • Merlin Dumeur (with​​​‌ M. Palva, Aalto Univ),​ PhD in cotutelle (4y),​‌ 2020-2025 (defense in Feb​​ 2025)
  • Pierre-Antoine Comby (with​​​‌ A. Vignaud, CEA), PhD​ 2021-2025 (defense in March​‌ 2025)
  • Serge Brosset, (with​​ Z. Saghi, CEA) PhD​​​‌ 2022-2025
  • Asma Tanabene (with​ C. Giliyar-Radhakrishna), PhD 2024-2027​‌
  • Caini Pan (w. A.​​ Vignaud & C. Giliyar-Radhakrishna),​​​‌ PhD 2024-2027
  • Caini Pan,​ M2, Telecom ParisTech, (Apr​‌ - Oct 2024, with​​ C. Giliyar-Radhakrishna)
• B. Thirion​​​‌

   

  • Thomas Chapalain, PhD 2021-2025​ (with E.Eger, CEA)
  • Nicolas​‌ Salvy, PhD 2023-2026 (with​​ H.Talbot, CentraleSupelec)
  • Joseph Paillard,​​​‌ 2024-2027 (with D.Engemann, Roche)​
  • Angel Reyero Lobo, 2024-2027​‌ (with P.Neuvial, CNRS)
  • Sonia​​ Mazelet, 2024-2027 (with R.Flamary,​​​‌ IPParis)
  • Pierre-Louis Barbarant, 2024-2027​ (with F.Meyniel, CEA)
  • Fernanda​‌ Ponce, 2024-2027 (with D.Wassermann,​​ Inria)
• D. Wassermann

   

  • Alexandre​​​‌ Le Bris, PhD 2022-2025​
  • Gabriela Gomez Jimenez, PhD​‌ 2023-2026 (with J. Valette​​ CEA)
• T. Moreau

   

  • J.​​​‌ Perdereau (with F. Vallée,​ APHP), PhD 2022-2025
  • V.​‌ Loison (with J. Cartailler,​​ APHP), PhD 2023-2026
  • C.​​​‌ Eve (with G. Varoquaux,​ Inria), PhD 2024-2027

10.2.2​‌ Juries

• B. Thirion

   

  • reviewer​​ for Mathieu Gilson's habilitation​​​‌ defense on September 25th,​ Marseille, France
  • Examiner for​‌ Julie Carlier's PhD defense​​ on Nov 26th, Paris,​​​‌ France
  • Examiner for Ulysse​ Klatzmann's PhD defense on​‌ Nov 26th, Paris, France​​
  • Reviewer for 's PhD​​​‌ defense on December 15th,​ Sophia Antipolis, France
  • Examiner​‌ for Shizhe Wu's public​​ defense on Jan 24th,​​​‌ Geneva, Switzerland

10.3.1 Participation in Live​‌ events

• February 2025
  Inria​​ interviewed Demian Wassermann in​​​‌ French to discuss how​ we will soon be​‌ able to predict thoughts​​.
• March 2025
  Demian​​​‌ Wassermann gave talks at​ different Parisian high-schools within​‌ the Chiche! program.
• March​​ 10, 2025
  During Brain​​​‌ Week, Anne Legall, a​ science journalist at France​‌ Info, interviewed Philippe Ciuciu​​ about the BrainSync research​​​‌ project; see Le billet​ Sciences.
• June 2025​‌
  An article was published​​ on the Inria website​​ about the BrainSync project​​​‌ led by Philipe Ciuciu,‌ titled: Et si l’IA‌​‌ aidait le cerveau à​​ réapprendre après un AVC​​​‌ ?
• September 2025
  An‌ interview with Philippe Ciuciu‌​‌ was published in the​​ monthly specialist newspaper "Pharmaceutiques"​​​‌ in French, discussing how‌ AI can help improve‌​‌ the care of stroke​​ patients.
• October 2025
  As​​​‌ deputy director of the‌ DataIA institute, Demian Wassermann‌​‌ was interviewed in Big​​ Data & AI Exhibition​​​‌ Paris to discuss how‌ AI contributes to decoding‌​‌ the brain and opening​​ the path to precision​​​‌ medicine.
• October 2025
  Bertrand‌ Thirion was invited by‌​‌ Café-éco des clubs Rotary​​ from the Aube region​​​‌ in the eastern part‌ of France.
• November 2025‌​‌
  Demian Wassermann was invited​​ to discuss AI and​​​‌ sovereignty at VivaTech 2025‌.

International journals

• 12 article​‌H.Himanshu Aggarwal,​​ A. F.Ana Fernanda​​​‌ Ponce, S.Swetha​ Shankar, Y.Yasmin​‌ Mzayek, A.Agnés​​ Pérez-Millan, A. L.​​​‌Ana Luísa Pinho,​ A.Alexis Thual and​‌ B.Bertrand Thirion.​​ Subject fingerprinting and task​​​‌ classification rely on distinct​ functional connectivity features.​‌Brain Structure and Function​​2311January 2026​​​‌, 12HALDOI​
• 13 articleJ.Jan​‌ Boelts, M.Michael​​ Deistler, M.Manuel​​​‌ Gloeckler, Á.Álvaro​ Tejero-Cantero, J.-M.Jan-Matthis​‌ Lueckmann, G.Guy​​ Moss, P.Peter​​​‌ Steinbach, T.Thomas​ Moreau, F.Fabio​‌ Muratore, J.Julia​​ Linhart, C.Conor​​​‌ Durkan, J.Julius​ Vetter, B. K.​‌Benjamin Kurt Miller,​​ M.Maternus Herold,​​​‌ A.Abolfazl Ziaeemehr,​ M.Matthijs Pals,​‌ T.Theo Gruner,​​ S.Sebastian Bischoff,​​​‌ N.Nastya Krouglova,​ R.Richard Gao,​‌ J. K.Janne K​​ Lappalainen, B.Bálint​​​‌ Mucsányi, F.Felix​ Pei, A.Auguste​‌ Schulz, Z.Zinovia​​ Stefanidi, P. L.​​​‌Pedro Luiz Coelho Rodrigues​, C.Cornelius Schröder​‌, F. A.Faried​​ Abu Zaid, J.​​​‌Jonas Beck, J.​Jaivardhan Kapoor, D.​‌ S.David S Greenberg​​, P. J.Pedro​​​‌ J Gonçalves and J.​ H.Jakob H Macke​‌. sbi reloaded: a​​ toolkit for simulation-based inference​​​‌ workflows.Journal of​ Open Source Software10​‌108April 2025,​​ 7754HALDOI
• 14​​​‌ articleT.Thomas Chapalain​, B.Bertrand Thirion​‌ and E.Evelyn Eger​​. Encoding of Numerosity​​ With Robustness to Object​​​‌ and Scene Identity in‌ Biologically Inspired Object Recognition‌​‌ Networks.Neural Computation​​August 2025, 1-36​​​‌HALDOI
• 15 article‌P.-A.Pierre-Antoine Comby,‌​‌ G.Guillaume Daval-Frérot,​​ C.Caini Pan,​​​‌ A.Asma Tanabene,‌ L.Léna Oudjman,‌​‌ M.Matteo Cencini,​​ P.Philippe Ciuciu and​​​‌ L.Léna Oudjman.‌ MRI-NUFFT: Doing non-Cartesian MRI‌​‌ has never been easier​​.Journal of Open​​​‌ Source Software10108‌April 2025, 7743‌​‌HALDOIback to​​ text
• 16 articleS.​​​‌Samuel Davenport, B.‌Bertrand Thirion and P.‌​‌Pierre Neuvial. FDP​​ control in mass-univariate linear​​​‌ models using the residual‌ bootstrap.Electronic Journal‌​‌ of Statistics 19March​​ 2025, 1313 -​​​‌ 1336HALDOI
• 17‌ articleR.Reuben Dorent‌​‌, N.Nazim Haouchine​​, A.Alexandra Golby​​​‌, S.Sarah Frisken‌, T.Tina Kapur‌​‌ and W.William Wells​​ Iii. Unified Cross-Modal​​​‌ Medical Image Synthesis with‌ Hierarchical Mixture of Product-of-Experts‌​‌.IEEE Transactions on​​ Pattern Analysis and Machine​​​‌ IntelligenceOctober 2025,‌ Online ahead of print‌​‌In press. HALDOI​​
• 18 articleM.Merlin​​​‌ Dumeur, J. M.‌J. Matias Palva and‌​‌ P.Philippe Ciuciu.​​ Outlier detection and removal​​​‌ in multifractal analysis of‌ electrophysiological brain signals.‌​‌EURASIP Journal on Advances​​ in Signal Processing2025​​​‌August 2025, 35‌HALDOI
• 19 article‌​‌I.Ingrid Garberis,​​ V.Valentin Gaury,​​​‌ C.Charlie Saillard,‌ D.Damien Drubay,‌​‌ K.Kevin Elgui,​​ B.Benoit Schmauch,​​​‌ A.Alexandre Jaeger,‌ L.Loïc Herpin,‌​‌ J.J. Linhart,​​ M.M. Sapateiro,​​​‌ F.F. Bernigole,‌ A.Aurélie Kamoun,‌​‌ A.Alexandre Filiot,​​ O.Oussama Tchita,​​​‌ R.Rémy Dubois,‌ M.M. Auffret,‌​‌ L.Lionel Guillou,​​ I.I. Bousaid,​​​‌ M.Mikael Azoulay,‌ J.Jerome Lemonnier,‌​‌ M.Meriem Sefta,​​ S.Sibille Everhard,​​​‌ A.Alexandre Sarrazin,‌ J. F.Jean François‌​‌ Reboud, F.Fabien​​ Brulport, J.Jocelyn​​​‌ Dachary, B.Barbara‌ Pistilli, S.Suzette‌​‌ Delaloge, P.Pierre​​ Courtiol, F.Fabrice​​​‌ André, V.Victor‌ Aubert and M.Magali‌​‌ Lacroix-Triki. Deep learning​​ assessment of metastatic relapse​​​‌ risk from digitized breast‌ cancer histological slides.‌​‌Nature Communications161​​July 2025, 5876​​​‌HALDOI
• 20 article‌Y.Yanis Lalou,‌​‌ T.Theo Gnassounou,​​ A.Antoine Collas,​​​‌ A.Antoine de Mathelin‌, O.Oleksii Kachaiev‌​‌, A.Ambroise Odonnat​​, T.Thomas Moreau​​​‌, A.Alexandre Gramfort‌ and R.Rémi Flamary‌​‌. SKADA-Bench: Benchmarking Unsupervised​​ Domain Adaptation Methods with​​​‌ Realistic Validation On Diverse‌ Modalities.Transactions on‌​‌ Machine Learning Research Journal​​July 2025, 1-48​​​‌HAL
• 21 articleR.‌Ruizhe Liu, H.‌​‌Hyesang Chang, D.​​Dawlat El-Said, D.​​​‌Demian Wassermann, Y.‌Yuan Zhang and V.‌​‌Vinod Menon. Brain-wide​​ decoding of numbers and​​​‌ letters: Converging evidence from‌ multivariate fMRI analysis and‌​‌ probabilistic meta-analysis.Cortex​​​‌189August 2025,​ 256-274HALDOIback​‌ to text
• 22 article​​M.Morgane Marzulli,​​​‌ A.Alexandre Bleuzé,​ J.Joe Saad,​‌ F.Felix Martel,​​ P.Philippe Ciuciu,​​​‌ T.Tetiana Aksenova and​ L.Lucas Struber.​‌ Classifying mental motor tasks​​ from chronic ECoG-BCI recordings​​​‌ using phase-amplitude coupling features​.Frontiers in Human​‌ Neuroscience19March 2025​​, 1521491HALDOI​​​‌
• 23 articleL.Lino​ Nobili, S. A.​‌Steve Alex Gibbs,​​ G.Gaia Burlando,​​​‌ G.Gaia Patrone,​ G.Giuseppe de Venuto​‌, S. H.Sheng​​ H Wang, G.​​​‌Gabriele Arnulfo, M.​Michele Terzaghi, S.​‌Stefano Francione and I.​​Ivana Sartori. Sleep-related​​​‌ epilepsy through the lens​ of stereo-EEG: Clinical and​‌ research update.Clinical​​ Neurophysiology177September 2025​​​‌, 2110811HALDOI​
• 24 articleA.Antoine​‌ Pelcat, A.Alice​​ Le Berre, W.​​​‌Wagih Ben Hassen,​ C.Clément Debacker,​‌ S.Sylvain Charron,​​ B.Bertrand Thirion,​​​‌ L.Laurence Legrand,​ G.Guillaume Turc,​‌ C.Catherine Oppenheim and​​ J.Joseph Benzakoun.​​​‌ Generative T2*-weighted images as​ a substitute for true​‌ T2*-weighted images on brain​​ MRI in patients with​​​‌ acute stroke.Diagnostic​ and Interventional ImagingMarch​‌ 2025HALDOI
• 25​​ articleJ.Jade Perdereau​​​‌, J.Jona Joachim​, F.Fabrice Vallée​‌, J.Jerome Cartailler​​ and T.Thomas Moreau​​​‌. Harnessing operating room​ signals to estimate mean​‌ arterial pressure with AnesthNet​​.Scientific Reports15​​​‌1September 2025,​ 33988HALDOIback​‌ to text
• 26 article​​M.Monica Roascio,​​​‌ S.Sheng Wang,​ V.Vladislav Myrov,​‌ F.Felix Siebenhühner,​​ R.Rosella Trò,​​​‌ P.Pietro Mattioli,​ F.Francesco Famà,​‌ S.Silvia Morbelli,​​ M.Matteo Pardini,​​​‌ B.Beatrice Orso,​ J. M.J. Matias​‌ Palva, D.Dario​​ Arnaldi and G.Gabriele​​​‌ Arnulfo. Unveiling Cortical​ Criticality Changes along the​‌ Prodromal to the Overt​​ Continuum of Alpha-Synucleinopathy.​​​‌Journal of Neuroscience45​31July 2025,​‌ e1871242025HALDOI
• 27​​ articleV.Victoria Shevchenko​​​‌, R. A.R​ Austin Benn, R.​‌Robert Scholz, W.​​Wei Wei, C.​​​‌Carla Pallavicini, U.​Ulysse Klatzmann, F.​‌Francesco Alberti, T.​​ D.Theodore D Satterthwaite​​​‌, D.Demian Wassermann​, P.-L.Pierre-Louis Bazin​‌ and D. S.Daniel​​ S Margulies. A​​​‌ comparative machine learning study​ of schizophrenia biomarkers derived​‌ from functional connectivity.​​Scientific Reports152849​​​‌January 2025HALDOI​
• 28 articleJ.Julián​‌ Tachella, M.Matthieu​​ Terris, S.Samuel​​​‌ Hurault, A.Andrew​ Wang, L.Leo​‌ Davy, J.Jérémy​​ Scanvic, V.Victor​​​‌ Sechaud, R.Romain​ Vo, T.Thomas​‌ Moreau, T.Thomas​​ Davies, D.Dongdong​​​‌ Chen, N.Nils​ Laurent, B.Brayan​‌ Monroy, J.Jonathan​​ Dong, Z.Zhiyuan​​​‌ Hu, M.-H.Minh-Hai​ Nguyen, F.Florian​‌ Sarron, P.Pierre​​ Weiss, P.Paul​​ Escande, M.Mathurin​​​‌ Massias, T.Thibaut‌ Modrzyk, B.Brett‌​‌ Levac, T. I.​​Tobías I Liaudat,​​​‌ M.Maxime Song,‌ J.Johannes Hertrich,‌​‌ S.Sebastian Neumayer and​​ G.Georg Schramm.​​​‌ DeepInverse: A Python package‌ for solving imaging inverse‌​‌ problems with deep learning​​.Journal of Open​​​‌ Source Software10115‌November 2025, 8923‌​‌HALDOI
• 29 article​​A.Alexis Thual,​​​‌ Y.Yohann Benchetrit,‌ F.Félix Geilert,‌​‌ J.Jérémy Rapin,​​ I.Iurii Makarov,​​​‌ S.Stanislas Dehaene,‌ B.Bertrand Thirion,‌​‌ H.Hubert Banville and​​ J.-R.Jean-Rémi King.​​​‌ Sample-efficient decoding of visual‌ stimuli from fMRI through‌​‌ inter-individual functional alignment.​​Transactions on Machine Learning​​​‌ Research JournalMay 2025‌HAL

International peer-reviewed conferences‌​‌

• 30 inproceedingsJ.Judith​​ Abécassis, H.Houssam​​​‌ Zenati, S.Sami‌ Boumaïza, J.Julie‌​‌ Josse and B.Bertrand​​ Thirion. CO11.2 -​​​‌ Explorer les fonctions cognitives‌ dans UK Biobank avec‌​‌ une analyse de médiation​​ causale.EPICLIN 2025​​​‌ - Conférence francophone d’EPIdémiologie‌ CLINique73Bordeaux, France‌​‌May 2025, 203025​​HALDOI
• 31 inproceedings​​​‌A.Alexandre Blain,‌ B.Bertrand Thirion and‌​‌ P.Pierre Neuvial.​​ False Coverage Proportion Control​​​‌ for Conformal Prediction.‌Proceedings of the 42‌​‌ nd International Conference on​​ Machine Learning, Vancouver, Canada.​​​‌ PMLR 267, 2025ICML‌ 2025 - 42 nd‌​‌ International Conference on Machine​​ LearningVancouver (BC), Canada​​​‌June 2025HAL
• 32‌ inproceedingsP.-A.Pierre-Antoine Comby‌​‌, M.Matthieu Terris​​, A.Alexandre Vignaud​​​‌ and P.Philippe Ciuciu‌. Plug-and-Play reconstruction for‌​‌ 3D non-cartesian fMRI data​​.33rd European Signal​​​‌ Processing Conference33rd European‌ Signal Processing ConferencePalermo,‌​‌ ItalySeptember 2025HAL​​
• 33 inproceedingsP.-A.Pierre-Antoine​​​‌ Comby, A.Alexandre‌ Vignaud and P.Philippe‌​‌ Ciuciu. SNAKE-fMRI: A​​ modular fMRI simulator from​​​‌ the space-time domain to‌ k-space data and back‌​‌.Proceedings of the​​ 2024 ISMRM & ISMRT​​​‌ Annual MeetingISMRM 2024‌ - ISMRM & ISMRT‌​‌ Annual MeetingSingapour, Singapore​​May 2025HAL
• 34​​​‌ inproceedingsR.Reuben Dorent‌, P.Polina Golland‌​‌ and W. M.William​​ M Wells Iii.​​​‌ Connecting Jensen-Shannon and Kullback-Leibler‌ Divergences: A New Bound‌​‌ for Representation Learning.​​NeurIPS 2025 - The​​​‌ Thirty-Ninth Annual Conference on‌ Neural Information Processing Systems‌​‌NeurIPS 2025 - 39th​​ Annual Conference on Neural​​​‌ Information Processing SystemsSan‌ Diego, United StatesDecember‌​‌ 2025HAL
• 35 inproceedings​​Q.Quentin Ferdinand,​​​‌ R.Rémi Souriau,‌ L.Lucas Struber,‌​‌ H.Henri Lorach,​​ P.Philippe Ciuciu,​​​‌ M.Marina Reyboz and‌ T.Tetiana Aksenova.‌​‌ ECoG-Based Movement Classification and​​ Limbs 3D Translation Prediction​​​‌ : a Deep Learning‌ Study.2025 International‌​‌ Joint Conference on Neural​​ Networks (IJCNN)Rome, France​​​‌IEEEJune 2025,‌ 1-10HALDOI
• 36‌​‌ inproceedingsP.Paul Krzakala​​, G.Gabriel Melo​​​‌, C.Charlotte Laclau‌, F.Florence d'Alché-Buc‌​‌ and R.Rémi Flamary​​. The quest for​​​‌ the GRAph Level autoEncoder‌ (GRALE).NeurIPS 2025‌​‌ - Thirty-Ninth Annual Conference​​​‌ on Neural Information Processing​ SystemsSan Diego, United​‌ States2025HAL
• 37​​ inproceedingsV.Virginie Loison​​​‌, G.Guillaume Staerman​ and T.Thomas Moreau​‌. UNHaP: Unmixing Noise​​ from Hawkes Processes.​​​‌International Conference on Artificial​ Intelligence and StatisticsAISTATS​‌ 2025 - 28th International​​ Conference on Artificial Intelligence​​​‌ and StatisticsPhuket, Thailand​May 2025HALback​‌ to text
• 38 inproceedings​​S.Sonia Mazelet,​​​‌ R.Rémi Flamary and​ B.Bertrand Thirion.​‌ Unsupervised Learning for Optimal​​ Transport plan prediction between​​​‌ unbalanced graphs.NeurIPS​ 2025 - 39th Annual​‌ Conference on Neural Information​​ Processing SystemsSan Diego​​​‌ (California), United States2025​HAL
• 39 inproceedingsJ.​‌Joseph Paillard, A.​​Antoine Collas, D.​​​‌ A.Denis A Engemann​ and B.Bertrand Thirion​‌. Hierarchical Variable Importance​​ with Statistical Control for​​​‌ Medical Data-Based Prediction.​Information Processing in Medical​‌ Imaging 29th International Conference,​​ IPMI 2025, Kos, Greece​​​‌IPMI 2025 - Information​ Processing in Medical Imaging​‌LNCS-15830Lecture Notes in​​ Computer ScienceKos Island,​​​‌ GreeceSpringer Nature Switzerland​August 2025, 79-93​‌HALDOI
• 40 inproceedings​​J.Joseph Paillard,​​​‌ A. R.Angel Reyero​ Lobo, V.Vitaliy​‌ Kolodyazhniy, B.Bertrand​​ Thirion and D.Denis​​​‌ Engemann. Measuring Variable​ Importance in Heterogeneous Treatment​‌ Effects with Confidence.​​Measuring Variable Importance in​​​‌ Heterogeneous Treatment Effects with​ ConfidenceICML 2025 -​‌ 42nd International Conference on​​ Machine Learning42Vancouver,​​​‌ Canada2025HALDOI​
• 41 inproceedingsH.Hugues​‌ Roy, R.Reuben​​ Dorent and N.Ninon​​​‌ Burgos. Unsupervised anomaly​ detection using Bayesian flow​‌ networks: application to brain​​ FDG pet in the​​​‌ context of Alzheimer’s disease​.Lecture Notes in​‌ Computer Science : Deep​​ Generative ModelsDGM4MICCAI 2025​​​‌ - Deep Generative Models​16128Lecture Notes in​‌ Computer ScienceDaejong, South​​ KoreaSpringer Nature Switzerland​​​‌September 2026, 254-264​HALDOI
• 42 inproceedings​‌E.Emilia Siviero,​​ G.Guillaume Staerman,​​​‌ S.Stéphan Clémençon and​ T.Thomas Moreau.​‌ Numerically Efficient Parametric Inference​​ for Learning Space-Time Hawkes​​​‌ Processes.Proceedings of​ the 12th IEEE International​‌ Conference on Data Science​​ and Advanced Analytics (DSAA)​​​‌Birmingham, United KingdomOctober​ 2025HAL
• 43 inproceedings​‌M.Matthieu Terris,​​ S.Samuel Hurault,​​​‌ M.Maxime Song and​ J.Julián Tachella.​‌ Reconstruct Anything Model: a​​ lightweight foundation model for​​​‌ computational imaging.ICLR​ 2026 - Fourteenth International​‌ Conference on Learning Representations​​Rio de Janeiro (BR),​​​‌ BrazilApril 2026HAL​
• 44 inproceedingsH.Herwig​‌ Wendt, P.Patrice​​ Abry, P.Philippe​​​‌ Ciuciu, M.Merlin​ Dumeur, S.Stéphane​‌ Jaffard, W. B.​​Wejdene Ben Nasr and​​​‌ G.Guillaume Saës.​ Analyse multifractale construite sur​‌ les weak scaling exponents​​.GRETSI 2025 -​​​‌ XXXe Symposium Signal and​ Image ProcessingStrasbourg, France​‌August 2025HAL
• 45​​ inproceedingsH.Houssam Zenati​​​‌, J.Judith Abécassis​, J.Julie Josse​‌ and B.Bertrand Thirion​​. Double Debiased Machine​​​‌ Learning for Mediation Analysis​ with Continuous Treatments.​‌Proceedings of Machine Learning​​ ResearchAISTATS - 28th​​ International Conference on Artificial​​​‌ Intelligence and StatisticsPMLR-‌Mai Khao, ThailandMay‌​‌ 2025HAL

Conferences without​​ proceedings

• 46 inproceedingsT.​​​‌Tiago Assis, I.‌Ines Machado, B.‌​‌Benjamin Zwick, N.​​Nuno Garcia and R.​​​‌Reuben Dorent. Deep‌ Biomechanically-Guided Interpolation for Keypoint-Based‌​‌ Brain Shift Registration.​​First International Workshop, COLAS​​​‌ 2025, Held in Conjunction‌ with MICCAI 2025, Daejeon,‌​‌ South Korea, September 23,​​ 2025, ProceedingsCOLAS 2025​​​‌ 2025 - First International‌ Workshop COLlaborative Intelligence and‌​‌ Autonomy in Image-guided Surgery​​ with MICCAI 2025LNCS-16298​​​‌Daejeon, South Korea2025‌HAL
• 47 inproceedingsM.‌​‌Maxime Bertrait, C.​​Chaithya Giliyar Radhakrishna and​​​‌ P.Philippe Ciuciu.‌ gGRAPPA: A Flexible, GPU-Accelerated‌​‌ Python Package for Fast​​ and Efficient generalized GRAPPA​​​‌ Reconstruction.ISMRM &‌ ISMRT 2025 - Annual‌​‌ Meeting & ExhibitionHonololu,​​ Hawaii, United StatesMay​​​‌ 2025HALback to‌ text
• 48 inproceedingsA.‌​‌Antoine Collas, C.​​Ce Ju, N.​​​‌Nicolas Salvy and B.‌Bertrand Thirion. Riemannian‌​‌ flow matching for brain​​ connectivity matrices via pullback​​​‌ geometry.NeurIPS 2025‌ - Thirty-Ninth Annual Conference‌​‌ on Neural Information Processing​​ SystemsSan Diego, United​​​‌ States2025HAL
• 49‌ inproceedingsP.-A.Pierre-Antoine Comby‌​‌, G.Guillaume Daval-Frérot​​, C.Chaithya Gr​​​‌, A.Alexandre Vignaud‌ and P.Philippe Ciuciu‌​‌. MRI-NUFFT: An open​​ source Python package to​​​‌ make non-Cartesian MR Imaging‌ easier.Proceedings of‌​‌ the 2025 ISMRM &​​ ISMRT Annual Meeting &​​​‌ ExhibitionISMRM & ISMRT‌ 2025 - Annual Meeting‌​‌ & ExhibitionHonolulu (Hawai),​​ United StatesMay 2025​​​‌HALDOI
• 50 inproceedings‌P.-A.Pierre-Antoine Comby,‌​‌ B.Benjamin Lapostolle,​​ M.Matthieu Terris and​​​‌ P.Philippe Ciuciu.‌ Robust plug-and-play methods for‌​‌ highly accelerated non-Cartesian MRI​​ reconstruction.2025 IEEE​​​‌ 22nd International Symposium on‌ Biomedical Imaging (ISBI)Houston,‌​‌ FranceIEEEApril 2025​​, 1-5HALDOI​​​‌
• 51 inproceedingsC.Chaithya‌ Giliyar Radhakrishna, A.‌​‌Alexandre Vignaud, M.​​Maxime Bertrait, A.​​​‌Aurélien Massire, M.‌Michel Bottlaender and P.‌​‌Philippe Ciuciu. Bringing​​ GRAPPA to non-Cartesian MRI​​​‌ through SPARKLING: An application‌ to MPRAGE anatomical MRI‌​‌.ISMRM & ISMRT​​ 2025 - Annual Meeting​​​‌ & ExhibitionHonololu, Hawaii,‌ United StatesMay 2025‌​‌HAL
• 52 inproceedingsL.​​Louis Jalouzot, A.​​​‌Alexis Thual, Y.‌Yair Lakretz, C.‌​‌Christophe Pallier and B.​​Bertrand Thirion. Optimizing​​​‌ fMRI Data Acquisition for‌ Decoding Natural Speech with‌​‌ Limited Participants.NeurIPS​​ Workshop 2025 : Foundation​​​‌ Models for the Brain‌ and Body.San Diego‌​‌ (CA), United StatesDecember​​ 2025HAL
• 53 inproceedings​​​‌A.Asma Tanabene,‌ C.Chaithya Giliyar Radhakrishna‌​‌, A.Aurélien Massire​​, M. S.Mariappan​​​‌ S. Nadar and P.‌Philippe Ciuciu. Benchmarking‌​‌ 3D multi-coil NC-PDNET MRI​​ reconstruction.ISMRM &​​​‌ ISMRT 2025 - Annual‌ Meeting & ExhibitionHonololu,‌​‌ Hawaii, United StatesMay​​ 2025HAL

Scientific book​​​‌ chapters

• 54 inbookB.‌Bertrand Thirion. Machine‌​‌ learning for NeuroImaging data​​ analysis.Encyclopedia of​​​‌ the Human BrainElsevier‌2025, 580-588HAL‌​‌DOI

Doctoral dissertations and​​​‌ habilitation theses

• 55 thesis​T.Thomas Chapalain.​‌ Investigating the representation of​​ numerosity in humans and​​​‌ convolutional neural networks using​ high-variability photorealistic stimuli.​‌Université Paris-SaclayMarch 2025​​HAL
• 56 thesisM.​​​‌Merlin Dumeur. Multifractal​ analysis for studying criticality​‌ in neural dynamics.​​Universite Paris-Saclay; Aalto University​​​‌May 2025HAL
• 57​ thesisT.Theo Gnassounou​‌. Multi-source domain adaptation​​ for learning on biosignals​​​‌.Université Paris-SaclayNovember​ 2025HAL

Reports &​‌ preprints

• 58 miscJ.​​Judith Abécassis, H.​​​‌Houssam Zenati, S.​Sami Boumaïza, J.​‌Julie Josse and B.​​Bertrand Thirion. Causal​​​‌ mediation analysis with one​ or multiple mediators: a​‌ comparative study.May​​ 2025HAL
• 59 misc​​​‌P.Patrice Abry,​ P.Phipippe Ciuciu,​‌ M.Merlin Dumeur,​​ S.Stéphane Jaffard and​​​‌ G.Guillaume Saës.​ Multifractal analysis based on​‌ the weak scaling exponent​​ and applications to MEG​​​‌ recordings in neuroscience.​March 2025HAL
• 60​‌ miscA.Alexandre Blain​​, A.Angel Reyero​​​‌ Lobo, J.Julia​ Linhart, B.Bertrand​‌ Thirion and P.Pierre​​ Neuvial. When Knockoffs​​​‌ fail: diagnosing and fixing​ non-exchangeability of Knockoffs.​‌April 2025HAL
• 61​​ miscG.Gaia Burlando​​​‌, C.Chiara Belforte​, F.Felix Siebenhühner​‌, L.Luca Di​​ Tullio, L.Lorenzo​​​‌ Chiarella, V.Vladislav​ Myrov, F.Frédéric​‌ Zubler, M.Monica​​ Roascio, F.Francesco​​​‌ Cardinale, S.Satu​ Palva, J. M.​‌J Matias Palva,​​ L.Lino Nobili,​​​‌ G.Gabriele Arnulfo and​ S. H.Sheng H​‌ Wang. Sleep-modulated cross-frequency​​ coupling between δ phase​​​‌ and β–γ bistability: a​ system-level modulation of epileptic​‌ activity.March 2025​​HALDOI
• 62 misc​​​‌A.Antoine Collas,​ C.Ce Ju,​‌ N.Nicolas Salvy and​​ B.Bertrand Thirion.​​​‌ Riemannian Flow Matching for​ Brain Connectivity Matrices via​‌ Pullback Geometry.2025​​HAL
• 63 miscM.​​​‌Michael Deistler, J.​Jan Boelts, P.​‌Peter Steinbach, G.​​Guy Moss, T.​​​‌Thomas Moreau, M.​Manuel Gloeckler, P.​‌ L.Pedro Luiz Coelho​​ Rodrigues, J.Julia​​​‌ Linhart, J. K.​Janne K. Lappalainen,​‌ B. K.Benjamin Kurt​​ Miller, P. J.​​​‌Pedro J. Gonçalves,​ J.-M.Jan-Matthis Lueckmann,​‌ C.Cornelius Schröder and​​ J. H.Jakob H.​​​‌ Macke. Simulation-Based Inference:​ A Practical Guide.​‌August 2025HAL
• 64​​ miscM.Merlin Dumeur​​​‌, P.Philippe Ciuciu​, W.Wenya Liu​‌, M.Maria Vesterinen​​, P.Paula Partanen​​​‌, A.Alexandra Andersson​, S.Samanta Knapič​‌, S.Satu Palva​​ and J. M.J​​​‌ Matias Palva. Characterization​ of multifractality in oscillatory​‌ dynamics.November 2025​​HAL
• 65 miscM.​​​‌Merlin Dumeur, S.​ H.Sheng H Wang​‌, J. M.J​​ Matias Palva and P.​​​‌Philippe Ciuciu. Multifractality​ in critical neural field​‌ dynamics.November 2025​​HAL
• 66 miscC.​​​‌Cedric Foucault, T.​Tiffany Bounmy, S.​‌Sébastien Demortain, B.​​Bertrand Thirion, E.​​Evelyn Eger and F.​​​‌Florent Meyniel. A‌ non-monotonic code for event‌​‌ probability in the human​​ brain.2025HAL​​​‌DOI
• 67 miscF.‌Fateme Ghayem, R.‌​‌Raphael Meudec, J.​​Jérôme Dockès, B.​​​‌Bertrand Thirion and D.‌Demian Wassermann. NeuroConText:‌​‌ Contrastive learning for neuroscience​​ meta-analysis with rich text​​​‌ representation.May 2025‌HALDOIback to‌​‌ text
• 68 miscT.​​Theo Gnassounou, A.​​​‌Antoine Collas, R.‌Rémi Flamary and A.‌​‌Alexandre Gramfort. PSDNORM:​​ TEST-TIME TEMPORAL NORMALIZATION FOR​​​‌ DEEP LEARNING IN SLEEP‌ STAGING.May 2025‌​‌HALDOI
• 69 misc​​B.Baptiste Goujaud,​​​‌ A.Adrien Taylor and‌ A.Aymeric Dieuleveut.‌​‌ Open Problem: Two Riddles​​ in Heavy-Ball Dynamics.​​​‌February 2025HAL
• 70‌ miscJ.Jona Joachim‌​‌, T.Thibaut Chamoux​​, J.Jade Perdereau​​​‌, V.Valentin Iovène‌, T.Thomas Moreau‌​‌, J.Jérôme Cartailler​​ and F.Fabrice Vallée​​​‌. Integrating Heterogeneous Perioperative‌ Data Sources: Platform Design‌​‌ and Implementation.November​​ 2025HALback to​​​‌ text
• 71 miscR.‌Raphael Meudec, J.‌​‌Jérôme Dockès, F.​​Fateme Ghayem, D.​​​‌Demian Wassermann and B.‌Bertrand Thirion. Peaks2Image:‌​‌ reconstructing fMRI maps from​​ stereotactic coordinates to enhance​​​‌ cognitive meta-analysis.August‌ 2025HAL
• 72 misc‌​‌N.Nicolas Salvy,​​ H.Hugues Talbot and​​​‌ B.Bertrand Thirion.‌ Enhanced generative model evaluation‌​‌ with Clipped Density and​​ Coverage.July 2025​​​‌HAL
• 73 miscS.‌ H.Sheng H Wang‌​‌, P.Paul Ferrari​​, G.Gabriele Arnulfo​​​‌, M.Morgane Marzulli‌, D.David Degras‌​‌, V.Vladislav Myrov​​, S.Satu Palva​​​‌, L.Lino Nobili‌, P.Philippe Ciuciu‌​‌ and M.Matias Palva​​. Toward unified biomarkers​​​‌ for focal epilepsy.‌April 2025HALDOI‌​‌
• 74 miscJ.Jad​​ Yehya, M.Mansour​​​‌ Benbakoura, C.Cédric‌ Allain, B.Benoît‌​‌ Malezieux, M.Matthieu​​ Kowalski and T.Thomas​​​‌ Moreau. RoseCDL: Robust‌ and scalable convolutional dictionary‌​‌ learning for rare-event detection​​.2025HALDOI​​​‌

Other scientific publications

• 75‌ inproceedingsP.-L.Pierre-Louis Barbarant‌​‌, F.Florent Meyniel​​ and B.Bertrand Thirion​​​‌. Functional templates in‌ fMRI: Building accurate and‌​‌ interpretable group-level decoders.​​CCN 2025 - 8th​​​‌ Annual Conference on Cognitive‌ Computational NeuroscienceAmsterdam, Netherlands‌​‌August 2025HAL
• 76​​ inproceedingsV.Victoria Shevchenko​​​‌, C. A.Carlo‌ Alberto Barbano, M.‌​‌Malo Renaudin, R.​​Robert Scholz, R.​​​‌ A.R. Austin Benn‌, W.Wei Wei‌​‌, C.Carla Pallavicini​​, U.Ulysse Klatzmann​​​‌, F.Francesco Alberti‌, P. L.Pierre‌​‌ Louis Bazin, T.​​ D.Theodore D Satterthwaite​​​‌, D. S.Daniel‌ S. Margulies and D.‌​‌Demian Wassermann. Reconstructing​​ Functional Connectomes and Predicting​​​‌ Behavior Using Contrastive Learning‌.OHBM 2025 -‌​‌ Organization for Human Brain​​ MappingBrisbane (AU), Australia​​​‌June 2025HAL