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 $𝐲=\Gamma \left(𝐱\right)+𝐛$. Here, $𝐛$ is an additive noise and $\Gamma$ 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 ${\Gamma }_{\mathrm{M}}/\mathrm{EEG}:{ℝ}^N\mapsto {ℝ}^M$ and $𝐛$ is considered white and Gaussian, while in magnetic resonance imaging (MRI), $\Gamma$ is a complex linear mapping ${\Gamma }_{\mathrm{MRI}}:{ℂ}^N\mapsto {ℂ}^M$ and $𝐛$ is circular complex white Gaussian. Despite the linearity of ${\Gamma }_{\mathrm{M}}/\mathrm{EEG}$ and ${\Gamma }_{\mathrm{MRI}}$, estimating $𝐱$ 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  74, 123, 96, 79, 124, 108, 107, 94, 99, 95. 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  146, 145, 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  147, 106. A particular application has been on diffusion MRI (dMRI), where we have linked the dMRI signal with physiological tissue models of grey matter tissue  106. Still in MRI but in susceptibility weighted imaging, another approach  88 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 87, 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 57, 58, 69, 81. 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  77, 72, 115.

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  159, 142 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  115, 116, 75, 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 64, 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. 104, 126). Since the challenge is to train such models on limited and noisy data, we will extend our recent work  62 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  104 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  60 as well as in the probabilistic case  156. 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  105. 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  157 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  158. 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  66. 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  63 in high-dimensional settings which are specific to neuroimaging research. In sum, by leveraging recent advances in deductive database systems  66 and this novel DSL  158 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  110, as well as simulation-based approaches (knockoff inference  73 or conditional randomization tests  119). 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  133. 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  68 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  148 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 70. 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,  70 have proposed a randomization-based procedure known as $\lambda$-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  70, 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  150 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 61, 135. 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  140, 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  120 or advanced strategies such as channel, time or frequency masking or phase randomizations  114, 117. 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  85, 119. 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  137 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  137, 128 and time series processing  138, 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  56.

Recently the team has applied SSL to two large cohorts of clinical EEG data  65 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  80.

While rather small networks have been employed so far on EEG data  76, 149 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  130, 129 and adapted them to neurophysiological data 109, 93, 80, 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  144, 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  78, tissue microstructure with cognitive control  125, functional gradients on the cortical surface 91 to functional territory segregation 143.

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  155. 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  154, 83, 92 and reveals neurodegenerative processes  118, 101, 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  97, we demonstrated, using the Cam-CAN cohort ($n=650$), 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  111. 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  152 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 112. 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  86 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  89 or natural language processing 84, 90 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  84, 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 67, 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 136. They are at the root of heterogeneous data integration, such as multi-modal machine learning 64. 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  113. 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 EEG-based modeling of clinical endpoints

Neurological and psychiatric disorders can show complex neurological patterns. Diagnosis is often performed clinically (based on cerebral signs and behavioral symptoms), leading to important variability across doctors. In clinical neuroscience, predicting diagnosis from brain signals is therefore a common application. In the clinical context, EEG is an economically viable option that can be applied in a wide array of circumstances. In collaboration with the Salpêtrière Hospital and the Paris Brain Institute (ICM) we have developed and validated an approach for an EEG-based modeling of diagnosis for severely brain injured patients suffering from consciousness disorders (DoC)  98. Expert-defined features from consciousness studies were rigorously combined using random forest classification. Sensitivity analysis and benchmarks showed robustness across EEG-configurations (channels, time points), protocols (resting state vs evoked responses), label noise and differences between recording sites. When changes in the signal are more subtle than they are in DoC patients (average power turned out to be one of the strongest stand-alone features) more general approaches are needed.

Our future activities will focus on extending this line of research to other clinical populations and other endpoints. We have started a collaboration with the Institut Pasteur (GHFC team, T Bourgeron, R Delorme) and the University of Montreal (PPSP team, G Dumas), to characterize differences between normally developing children and children diagnosed with autism spectrum disorders. A wide array of EEG tasks will be used and endpoints (i.e. developmental timepoints) will go beyond the usually accurate diagnosis, focusing on symptom severity and social developmental scores. With the anesthesiology department at the Lariboisière hospital (A Mebazaa, E Gayat, F Vallée) and the cognitive neurology unit (C Paquet) we aim at developing EEG-based models of cognitive decline and dysfunction in two different settings. Postoperative cognitive decline is an important complication after general anesthesia and its antecedents must be better understood. As this might be an indicator for a latent neurodegenerative condition, we plan to use our EEG-based models of both Alzheimer's Disease and Lewy body dementia in which disease progression is an important change over time.

This widening scope calls for a more general methodology as compared to our previous work on DoC. For example, in these conditions involving neurodegenerative problems, we have observed that both subtle and condition-specific spatial patterns matter more than strong and global amplitude changes. To approach these challenges we will draw on our latest M/EEG-methods that were recently developed for population-level modeling of brain health and brain aging  97. We found that frequency band-specific spatial patterns of M/EEG power spectra conveyed important information of cognitive function (memory and cognitive performance) that were not explained by MRI or fMRI. This was implemented by predicting from a filter-bank of frequency-band-specific source power and source connectivity features. Core challenges to enable clinical translation include lower SNR and absence of individual anatomical MRI scans needed for gold-standard source modeling. Through theoretical analysis, simulations and benchmarks we found  152, 151 that, in M/EEG sensor space, covariance matrices in combination with spatial filtering techniques and Riemannian embeddings provide good workarounds for absent anatomical MRI scans. This covariance-based approach allows to capture fine-grained spatial information related to power and connectivity without performing biophysics-based source localization. Moreover, Riemannian embeddings make predictive modeling from M/EEG covariance matrices more robust to noise, whereas their interpretability is more challenging than that of spatial filters, indicating a direction for further research. Another challenge is given by the limited numbers of labeled samples for supervised learning and EEG-devices with small channel numbers, such as monitoring or user-grade EEG with 2-4 electrodes for which random loss of electrodes can be frequent. In this context, we expect important enhancements from self-supervised learning approaches  65 and deep learning methods for data-augmentation for which we have obtained the first results on non-clinical data. In these settings, the previous elements from classical approaches such as Riemannian geometry or spatial filtering can be readily implemented alongside more involved computations and transformations.

4.3.2 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  116, 75 as well as our PySAP software  99 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  146, 145 to multiple acquisition setups and other downstream tasks (e.g. motion correction and correction of off-resonance artifacts related to $B_0$ inhomogeneities).

5.1.1 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

5.1.2 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

5.1.3 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:
  Alexandre Abraham, Alexandre Gramfort, Bertrand Thirion, Elvis Dohmatob, Fabian Pedregosa Izquierdo, Gael Varoquaux, Loic Esteve, Michael Eickenberg, Virgile Fritsch

5.1.4 Benchopt

• Keywords:
  Mathematical Optimization, Benchmarking, Reproducibility
• 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/1.3.0
• Contact:
  Thomas Moreau

5.1.5 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:
  Olivier Grisel
• Participants:
  Alexandre Gramfort, Bertrand Thirion, Gael Varoquaux, Loic Esteve, Olivier Grisel, Guillaume Lemaitre, Jeremie Du Boisberranger, Julien Jerphanion
• Partners:
  Boston Consulting Group - BCG, Microsoft, Axa, BNP Parisbas Cardif, Fujitsu, Dataiku, Assistance Publique - Hôpitaux de Paris, Nvidia

5.1.6 joblib

• Keywords:
  Parallel computing, Cache
• Functional Description:
  Facilitate parallel computing and caching in Python.
• URL:
  https://­joblib.­readthedocs.­io/­en/­latest/
• Contact:
  Gael Varoquaux

7.1.1 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 the context of the joint supervision of Guillaume Daval-Frérot's CIFRE PhD thesis dedicated to Deep learning for off-resonance artifact correction in MR image reconstruction in the specific application of susceptibility weighted imaging at 3 Tesla. On top of the PhD funding, this contract has generated 45k€ for Mind and was managed by CEA/NeuroSpin. G. Daval-Frérot's PhD defense was held on December, 16 2022. As this first collaboration was successful, we engaged strategic discussions with the Siemens-Healthineers headquarters (Erlangen, Germany) during Spring 2022 to pursue this partnership and eventually set up a new one in 2023 with again a CIFRE PhD thesis starting this Spring. The PhD candidate, M. Ahmed Boughdiri, has been recently selected. The focus of this PhD thesis will be self-supervised 3D MR image reconstruction still in the deep learning setting.

7.1.2 Facebook AI Research (FAIR)

There is currently a CIFRE PhD between FAIR and Mind (Alexandre Gramfort) to investigate the differences between deep learning models and the brain, especially considering NLP machine learning models. As the collaboration is led by Alexandre Gramfort in the team, it is therefore financially managed by Inria.

7.1.3 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.

8.1.1 Associate Teams in the framework of an Inria International Lab or in the framework of an Inria International Program

8.1.2 Inria associate team not involved in an IIL or an international program

8.2.1 Visits of international scientists

8.3.1 Horizon Europe

ANR DARLING

• Title:
  DARLING: Distributed adaptation and learning over graph signals
• Duration:
  2020 ->
• 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 ->
• 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${}^3$ 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.

KARAIB AI CHAIR

• Title:
  KARAIB: Knowledge And RepresentAtion Integration on the Brain
• Duration:
  2020 ->
• Coordinator:
  Bertrand Thirion
• Partners:
  • INRIA MIND, Gif-sur-Yvette, France
• Inria contact:
  Bertrand Thirion
• Summary:

  Cognitive science describes mental operations, and functional brain imaging provides a unique window into the brain systems that support these operations. A growing body of neuroimaging research has provided significant insight into the relations between psychological functions and brain activity. However, the aggregation of cognitive neuroscience results to obtain a systematic mapping between structure and function faces the roadblock that cognitive concepts are ill-defined and may not map cleanly onto the computational architecture of the brain.

  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. Our aim here is to develop multi-modal machine learning techniques to bridge these data sources.

  • Aim 1 develops representation techniques for noisy data to couple brain data with descriptions of behavior or diseases, in order to extract semantic structure.
  • Aim 2 challenges these representations to provide explanations to the observed relationships, based on two frameworks: i) a statistical analysis framework; ii) integration into a domain-specific language.
  • Aim 3 outputs readily-usable products for neuroimaging: atlases and ontologies and focuses on implementation, with contributions to neuroimaging web-based data sharing tools.site.

BrAIN AI CHAIR

• Title:
  BrAIN: Bridging Artificial Intelligence and Neuroscience
• Duration:
  2020 ->
• Coordinator:
  Alexandre Gramfort
• Partners:
  • INRIA MIND, Gif-sur-Yvette, France
• Inria contact:
  Alexandre Gramfort
• Summary:

  The BrAIN project investigates learning tasks from multivariate EEG and MEG time series. In clinical or cognitive neuroscience, electromagnetic signals emitted by synchronously firing neurons are collected by electroencephalography (EEG) or magnetoencephalography (MEG). Such data, typically sampled at millisecond resolution, are routinely used for clinical applications such as anesthesia monitoring, sleep medicine, epilepsy or disorders of consciousness. Low cost EEG devices are also becoming commodities with hardware startups such as DREEM in France or InteraXon in Canada that have collected hundred of thousands of neural recordings. The field of neuroscience urgently needs algorithms that can learn from such large and poorly labeled datasets. The general objectives of BrAIN is to develop ML algorithms that can learn with weak or no supervision on neural time series. It requires contributions to self-supervised learning, domain adaptation and data augmentation techniques, exploiting the known underlying physical mechanisms that govern the data generating process of neurophysiological signals.

  The BrAIN project is organized around four objectives:

  • Learn with no-supervision on noisy and complex multivariate signals
  • Learn end-to-end predictive systems from limited data exploiting physical constraints
  • Learn from data coming from many different source domains
  • Develop high-quality software tools that can reach clinical research

9.1.1 Scientific events: organisation

9.1.2 Scientific events: selection

• P. Ciuciu
  Member (Representative of the IEEE Signal Processing Society) of the steering committee of the IEEE International Symposium on Biomedical Imaging. I participated in selecting the bids for 2022 (Kolkata, IN) 2023 (Cartagena de Indias , CO) and 2024 (Athens, GR).

9.1.3 Journal

9.1.4 Invited talks

• P. Ciuciu

   

  • June 2022: LIONS lab at EPFL (Lausanne, CH)
  • July 2022: Plenary speaker at the 13th FORTH scientific retreat (Heraklion, GR)
• D. Wassermann

   

  • Johns Hopkins University, USA
  • Stanford University, USA
• T. Moreau

   

  • March 2022: LJK seminar at UGA (Grenoble, France)
  • June 2022: CSD seminar at ENS Paris (Paris, France)
  • Invited talk at Curves and Surfaces (Arcachon, France)
  • July 2022: invited talk for AI4SIP workshop at Institut Pascal (Orsay, France)
• B.Thirion

   

  • March 2022: SousSoucis Days, Institut Mathématique de Toulouse
  • May 2022: HBP WP1 conference, Paris
  • November 2022: ML for Life Science, Montpellier
  • December 2022: plenary talk at MAIN conference, Montreal
• A.Gramfort

   

  • March 2022: Invited talk, UCL ML seminar, London
  • June 2022: Brain and Spine Institute (ICM) symposium “Computational and mathematical approaches for neuroscience”, Paris, France
  • July 2022: invited talk for AI4SIP workshop at Institut Pascal (Orsay, France)
  • August 2022: Invited talk Biomag conference, Birmingham, UK

9.1.5 Scientific expertise

• P. Ciuciu
  : European expert reviewer for the European Innovation Council Accelerator actions (main track: AI and health).
• A. Gramfort
  : External reviewer for the European Research Council (ERC).

9.1.6 Research administration

• P. Ciuciu

   

  • Member of the Board of Directors at NeuroSpin (CEA).
  • 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
  • Member of ENS Paris-Saclay Scientific Council
  • Member of Inria Commission évaluation
• A. Gramfort

   

  • Membre du comité des programmes de l'Institut de Henry Poincaré (IHP)
  • Membre du pôle B de l'école doctorale de l'Université Paris-Saclay
  • Responsable du Center for Data Science (CDS) de l'institut DataIA en charge des data challenges
  • Représentant Inria au sein du comité opérationnel du cendre d'IA Hi!Paris d'IP Paris

9.2.1 Supervision

• P. Ciuciu

   

  • Dr Z. Ramzi (with J.-L. Starch, CEA), PhD 2019-2022
  • Dr G. Daval-Frérot (with A. Vignaud, CEA), PhD 2019-2022
  • C. Giliyar Radhakrishna, PhD 2020-2023 (defense planned in April 2023)
  • M. Dumeur (with M. Palva, Aalto Univ), PhD in cotutelle (4y), 2020-2024
  • Z. Amor (with A. Vignaud, CEA) PhD 2019-2023
  • A. Waguet (with T. Druet, CEA) PhD 2019-2023
  • P.-A. Comby (with A. Vignaud, CEA), PhD 2021-2024
  • Maja Pantic (with A. Vignaud, CEA), PhD resigned in Dec 2022 for personal reasons after obtaining an UDOPIA scholarship in June 2022.
  • Serge Brosset, (with Z. Saghi, CEA) PhD 2022-2025
  • W. Omezzine, M1 INSA Toulouse, (June - Sep 2022)
  • Maja Pantic M2, Univ Paris-Saclay (Apr - Sep 2022)
• D. Wassermann

   

  • L. Rouillard, PhD 2021-2023
  • M. Jallais, PhD 2019-2022
  • G. Zanitti, PhD 2020-2023
  • C. Fang (with J-R Li), PhD 2020-2023
  • R. Meudec (with B. Thirion), PhD 2021-2023
  • A. Le Bris, M2 Telecom ParisTech, 2022
• T. Moreau

   

  • B. Malézieux, PhD 2020-2023 (with M. Kowalski)
  • C. Allain, PhD 2021-2024 (with A. Gramfort)
  • M. Dagréou, PhD 2021-2024 (with S. Vaiter and P. Ablin)
  • F. Michel, PhD 2022-2025 (with M. Kowalski)
  • E. Lai, 3A Ecole Polytechnique, 2022 (with A. Gramfort and J.-C. Pesquet)
  • M. Nargeot, L2 Université de Bordeaux, 2022
• B. Thirion

   

  • Alexandre Pasquiou, PhD 2020-2023 (with C.Pallier)
  • Alexis Thual, PhD 2020-2023 (with S. Dehaene)
  • Ahmad Chamma, PhD 2021-2023 (with D.Engemann)
  • Raphael Meudec, PhD 2021-2024 (with D.Wassermann)
  • Thomas Chapalain, PhD 2021-2024 (with E.Eger)
  • Alexandre Balin, PhD 2021-2024 (with P.Neuvial)
• A. Gramfort

   

  • Charlotte Caucheteux, PhD 2020-2023 (with J.R. King)
  • Omar Chehab, PhD 2020-2023 (with A. Hyvarinen)
  • Cédric Allain, PhD 2021-2024 (with T.Moreau)
  • Apolline Mellot, PhD 2021-2024 (with D.Engemann)
  • Julia Linhart, PhD 2021-2024 (with P.Rogrigues)
  • Theo Gnassounou, PhD 2022-2025 (with R.Flamary)
  • Ambroise Heurtebise, PhD 2022-2025 (with P.Ablin)

9.2.2 Juries

• P. Ciuciu
  : External reviewer of the PhD thesis defended by Thomas Sanchez (EPFL, Lausanne, CH) in April 2022.
• D. Wassermann
  : Examinator for the PhD thesis defended by J. Ringard at École Polytechnique.
• T. Moreau
  : Examinator for the PhD thesis defended by L. Dragoni at Université Côte d'Azure (Nice, France) in May 2022.
• B. Thirion
  External reviewer of the PhD thesis defended by Frederico Bertoni (Sorbonne Université), in Febuary 2022
• A. Gramfort
  External reviewer of the PhD thesis defended by Martin Grignard (Université de Liège, Belgique), in September 2022
• A. Gramfort
  External reviewer of the PhD thesis defended by Jean-Yves Franceschi (Sorbonne Université), in Febuary 2022
• A. Gramfort
  External reviewer of the PhD thesis defended by Ali Hashemi  (TU Berlin), in September 2022
• A. Gramfort
  External reviewer of the HdR defended by Levgen Redko (Université St Etienne), in July 2022
• A. Gramfort
  Examinator for the PhD thesis defended by Theo Desbordes (Sorbonne Université), in Dec 2022.
• A. Gramfort
  Examinator for the PhD thesis defended by Antoine Collas (Université Paris-Saclay), in Nov 2022.

9.3.1 Internal or external Inria responsibilities

• D. Wassermann
  COERLE Scientific representative for Inria Saclay Île-de-France; Representative at the Graduate School in Computer Science of Université Paris-Saclay for Inria Saclay Île-de-France.
• T. Moreau
  President of the Inria Saclay CUMI; Representative for Inria Saclay in the commission for the development of the national computational resources.
• A. Gramfort
  Membre de la comission d'évaluation technologique (CDT) du centre de Saclay
• A. Gramfort
  Representant de l'Inria Saclay dans le comité opérationnel de Hi!Paris

International journals

• 8 articleM.Majd Abdallah, V.Valentin Iovene, G.Gaston Zanitti and D.Demian Wassermann. Meta-Analysis of the Functional Neuroimaging Literature with Probabilistic Logic Programming.Scientific Reports2022
  HALDOIback to textback to textback to text
• 9 articleM.Majd Abdallah, G.Gaston Zanitti, V.Valentin Iovene and D.Demian Wassermann. Functional gradients in the human lateral prefrontal cortex revealed by a comprehensive coordinate-based meta-analysis.eLife2022
  HALDOIback to text
• 10 articleJ.Joseph Benzakoun, M.-A.Marc-Antoine Deslys, L.Laurence Legrand, G.Ghazi Hmeydia, G.Guillaume Turc, W. B.Wagih Ben Hassen, S.Sylvain Charron, C.Clément Debacker, O.Olivier Naggara, J.-C.Jean-Claude Baron, B.Bertrand Thirion and C.Catherine Oppenheim. Synthetic FLAIR as a Substitute for FLAIR Sequence in Acute Ischemic Stroke.RadiologyJanuary 2022, 211394
  HALDOI
• 11 articleA.Alexandre Blain, B.Bertrand Thirion and P.Pierre Neuvial. Notip: Non-parametric True Discovery Proportion control for brain imaging.NeuroImage2602022, 119492
  HALDOI
• 12 articleC.Charlotte Caucheteux, A.Alexandre Gramfort and J.-R.Jean-Rémi King. Deep language algorithms predict semantic comprehension from brain activity.Scientific ReportsSeptember 2022
  HALDOIback to text
• 13 articleH.Hyesang Chang, L.Lang Chen, Y.Yuan Zhang, Y.Ye Xie, C.Carlo de Los Angeles, E.Emma Adair, G.Gaston Zanitti, D.Demian Wassermann, M.Miriam Rosenberg-Lee and V.Vinod Menon. Foundational number sense training gains are predicted by hippocampal–parietal circuits.Journal of Neuroscience42192022, JN-RM-1005-21
  HALDOIback to text
• 14 articleO.Omar Chehab, A.Alexandre Défossez, L.Loiseau Jean-Christophe, A.Alexandre Gramfort and J.-R.Jean-Rémi King. Deep Recurrent Encoder: an end-to-end network to model magnetoencephalography at scale.Neurons, Behavior, Data analysis, and TheorySeptember 2022
  HALDOI
• 15 articleD.Darya Chyzhyk, G.Gaël Varoquaux, M.Michael Milham and B.Bertrand Thirion. How to remove or control confounds in predictive models, with applications to brain biomarkers.GigaScience11March 2022
  HALDOI
• 16 articleG.Guillaume Daval-Frérot, A.Aurélien Massire, B.Boris Mailhé, M.Mariappan Nadar, A.Alexandre Vignaud and P.Philippe Ciuciu. Iterative static $$B0 field map estimation for off-resonance correction in non-Cartesian susceptibility weighted imaging.Magnetic Resonance in Medicine884June 2022, 1592-1607
  HALDOIback to text
• 17 articleA.Arnaud Delorme, R.Robert Oostenveld, F.Francois Tadel, A.Alexandre Gramfort, S.Srikantan Nagarajan and V.Vladimir Litvak. Editorial: From Raw MEG/EEG to Publication: How to Perform MEG/EEG Group Analysis With Free Academic Software.Frontiers in Neuroscience16April 2022
  HALDOI
• 18 articleC.Chaithya Giliyar Radhakrishna, P.Pierre Weiss, G.Guillaume Daval-Frérot, A.Aurélien Massire, A.Alexandre Vignaud and P.Philippe Ciuciu. Optimizing full 3D SPARKLING trajectories for high-resolution T2*-weighted Magnetic Resonance Imaging.IEEE Transactions on Medical ImagingAugust 2022
  HAL
• 19 articleB.Bruno Hebling Vieira, F.Franziskus Liem, K.Kamalaker Dadi, D.Denis Engemann, A.Alexandre Gramfort, P.Pierre Bellec, R. C.Richard Cameron Craddock, J.Jessica Damoiseaux, C.Christopher Steele, T.Tal Yarkoni, N.Nicolas Langer, D.Daniel Margulies and G.Gaël Varoquaux. Predicting future cognitive decline from non-brain and multimodal brain imaging data in healthy and pathological aging.Neurobiology of Aging118October 2022, 55-65
  HALDOI
• 20 articleC.Cecile Herate, P.Patricia Brochard, F.Florent de Vathaire, M.Michelle Ricoul, B.Bernadette Martins, L.Laurence Laurier, J.-R.Jean-Robert Deverre, B.Bertrand Thirion, L.Lucie Hertz-Pannier and L.Laure Sabatier. The effects of repeated brain MRI on chromosomal damage.European Radiology Experimental612March 2022
  HALDOI
• 21 articleM.Maëliss Jallais, P. L.Pedro Luiz Coelho Rodrigues, A.Alexandre Gramfort and D.Demian Wassermann. Inverting brain grey matter models with likelihood-free inference: a tool for trustable cytoarchitecture measurements.Journal of Machine Learning for Biomedical ImagingMay 2022, 1-27
  HALDOI
• 22 articleS.-L.Sook-Lei Liew, B.Bethany Lo, M.Miranda Donnelly, A.Artemis Zavaliangos-Petropulu, J.Jessica Jeong, G.Giuseppe Barisano, A.Alexandre Hutton, J.Julia Simon, J.Julia Juliano, A.Anisha Suri, Z.Zhizhuo Wang, A.Aisha Abdullah, J.Jun Kim, T.Tyler Ard, N.Nerisa Banaj, M.Michael Borich, L.Lara Boyd, A.Amy Brodtmann, C.Cathrin Buetefisch, L.Lei Cao, J.Jessica Cassidy, V.Valentina Ciullo, A.Adriana Conforto, S.Steven Cramer, R.Rosalia Dacosta-Aguayo, E.Ezequiel de la Rosa, M.Martin Domin, A.Adrienne Dula, W.Wuwei Feng, A.Alexandre Franco, F.Fatemeh Geranmayeh, A.Alexandre Gramfort, C.Chris Gregory, C.Colleen Hanlon, B.Brenton Hordacre, S.Steven Kautz, M. S.Mohamed Salah Khlif, H.Hosung Kim, J.Jan Kirschke, J.Jingchun Liu, M.Martin Lotze, B.Bradley Macintosh, M.Maria Mataró, F.Feroze Mohamed, J.Jan Nordvik, G.Gilsoon Park, A.Amy Pienta, F.Fabrizio Piras, S.Shane Redman, K.Kate Revill, M.Mauricio Reyes, A.Andrew Robertson, N. J.Na Jin Seo, S.Surjo Soekadar, G.Gianfranco Spalletta, A.Alison Sweet, M.Maria Telenczuk, G.Gregory Thielman, L.Lars Westlye, C.Carolee Winstein, G.George Wittenberg, K.Kristin Wong and C.Chunshui Yu. A large, curated, open-source stroke neuroimaging dataset to improve lesion segmentation algorithms.Scientific Data 91December 2022, 320
  HALDOI
• 23 articleR.Romuald Menuet, R.Raphael Meudec, J.Jérôme Dockès, G.Gaël Varoquaux and B.Bertrand Thirion. Comprehensive decoding mental processes from Web repositories of functional brain images.Scientific Reports127050December 2022
  HALDOI
• 24 articleS.Sofiane Mrah, M.Maxime Descoteaux, M.Michel Wager, A.Arnaud Boré, F.François Rheault, B.Bertrand Thirion and E.Emmanuel Mandonnet. Network-level prediction of set-shifting deterioration after lower-grade glioma resection.Journal of NeurosurgeryMarch 2022, 1-9
  HALDOI
• 25 articleZ.Zaccharie Ramzi, C.Chaithya Giliyar Radhakrishna, J.-L.Jean-Luc Starck and P.Philippe Ciuciu. NC-PDNet: a Density-Compensated Unrolled Network for 2D and 3D non-Cartesian MRI Reconstruction.IEEE Transactions on Medical ImagingJanuary 2022
  HALDOI
• 26 articleZ.Zaccharie Ramzi, K.Kevin Michalewicz, J.-L.Jean-Luc Starck, T.Thomas Moreau and P.Philippe Ciuciu. Wavelets in the deep learning era.Journal of Mathematical Imaging and VisionOctober 2022
  HALDOI
• 27 articleA.Alexander Rockhill, E.Eric Larson, B.Brittany Stedelin, A.Alessandra Mantovani, A.Ahmed Raslan, A.Alexandre Gramfort and N.Nicole Swann. Intracranial Electrode Location and Analysis in MNE-Python.Journal of Open Source Software770February 2022, 3897
  HALDOI
• 28 articleC.Cédric Rommel, J.Joseph Paillard, T.Thomas Moreau and A.Alexandre Gramfort. Data augmentation for learning predictive models on EEG: a systematic comparison.Journal of Neural EngineeringNovember 2022
  HALDOIback to text
• 29 articleG. E.Gaston E Zanitti, Y.Yamil Soto, V.Valentin Iovene, M.Maria Vanina Martinez, R. O.Ricardo O Rodriguez, G. I.Gerardo I Simari and D.Demian Wassermann. Scalable Query Answering under Uncertainty to Neuroscientific Ontological Knowledge: The NeuroLang Approach.Neuroinformatics2022
  HALDOIback to text

International peer-reviewed conferences

• 30 inproceedingsO.Omar Chehab, A.Alexandre Gramfort and A.Aapo Hyvärinen. The Optimal Noise in Noise-Contrastive Learning Is Not What You Think.UAI 2022 - 38th Conference on Uncertainty in Artificial IntelligenceEindhoven, NetherlandsAugust 2022
  HAL
• 31 inproceedingsC.Chaithya Giliyar Radhakrishna, Z.Zaccharie Ramzi and P.Philippe Ciuciu. Hybrid learning of Non-Cartesian k-space trajectory and MR image reconstruction networks.ISBI 2022 - IEEE 19th International Symposium on Biomedical ImagingKolkata, IndiaMarch 2022
  HAL
• 32 inproceedingsJ.Juliette Millet, C.Charlotte Caucheteux, P.Pierre Orhan, Y.Yves Boubenec, A.Alexandre Gramfort, C. C.Christophe C Pallier, E.Ewan Dunbar and J.-R.Jean-Rémi King. Toward a realistic model of speech processing in the brain with self-supervised learning.NeurIPS 2022 - 36th Conference on Neural Information Processing SystemsNew Orleans, United StatesNovember 2022
  HALback to text
• 33 inproceedingsT.Thomas Moreau, M.Mathurin Massias, A.Alexandre Gramfort, P.Pierre Ablin, P.-A.Pierre-Antoine Bannier, B.Benjamin Charlier, M.Mathieu Dagréou, T.Tom Dupré La Tour, G.Ghislain Durif, C. F.Cassio F Dantas, Q.Quentin Klopfenstein, J.Johan Larsson, E.En Lai, T.Tanguy Lefort, B.Benoit Malézieux, B.Badr Moufad, B. T.Binh T Nguyen, A.Alain Rakotomamonjy, Z.Zaccharie Ramzi, J.Joseph Salmon and S.Samuel Vaiter. Benchopt: Reproducible, efficient and collaborative optimization benchmarks.NeurIPS 2022 - 36th Conference on Neural Information Processing SystemsNew Orleans, United StatesNovember 2022
  HAL
• 34 inproceedingsB. T.Binh T. Nguyen, B.Bertrand Thirion and S.Sylvain Arlot. A Conditional Randomization Test for Sparse Logistic Regression in High-Dimension.NeurIPS 202235Advances in Neural Information Processing SystemsNew Orleans, United StatesNovember 2022
  HAL
• 35 inproceedingsK.Kumari Pooja, Z.Zaccharie Ramzi, C.Chaithya Giliyar Radhakrishna and P.Philippe Ciuciu. MC-PDNET: Deep unrolled neural network for multi-contrast mr image reconstruction from undersampled k-space data.IEEE International Symposium on Biomedical Imaging 2022Kolkata, IndiaMarch 2022
  HAL
• 36 inproceedingsZ.Zaccharie Ramzi, F.Florian Mannel, S.Shaojie Bai, J.-L.Jean-Luc Starck, P.Philippe Ciuciu and T.Thomas Moreau. SHINE: SHaring the INverse Estimate from the forward pass for bi-level optimization and implicit models.ICLR 2022 - International Conference on Learning RepresentationsVirtual, United StatesApril 2022
  HALback to text
• 37 inproceedingsC.Cédric Rommel, T.Thomas Moreau and A.Alexandre Gramfort. Deep invariant networks with differentiable augmentation layers.Thirty-sixth Conference on Neural Information Processing SystemsAdvances in neural information processing systemsNew Orleans, United StatesOctober 2022
  HALback to text

Conferences without proceedings

• 38 inproceedingsZ.Zaineb Amor, C.Chaithya Giliyar Radhakrishna, G.Guillaume Daval-Frérot, B.Bertrand Thirion, F.Franck Mauconduit, P.Philippe Ciuciu and A.Alexandre Vignaud. 3D-SPARKLING for functional MRI: A pilot study for retinotopic mapping at 7T.OHBM 2022Glasgow, United KingdomJune 2022
  HALback to text
• 39 inproceedingsZ.Zaineb Amor, C.Chaithya Giliyar Radhakrishna, G.Guillaume Daval-Frérot, B.Bertrand Thirion, F.Franck Mauconduit, P.Philippe Ciuciu and A.Alexandre Vignaud. Prospects of non-Cartesian 3D-SPARKLING encoding for functional MRI: A preliminary case study for retinotopic mapping.Joint Annual Meeting ISMRM-ESMRMB & ISMRT 31st Annual MeetingLondon, United KingdomMay 2022
  HAL
• 40 inproceedingsZ.Zaineb Amor, C.Chaithya Giliyar Radhakrishna, C.Caroline Le Ster, G.Guillaume Daval-Frérot, N.Nicolas Boulant, F.Franck Mauconduit, C.Christian Mirkes, P.Philippe Ciuciu and A.Alexandre Vignaud. B0 field distortions monitoring and correction for 3D non-Cartesian fMRI acquisitions using a field camera: Application to 3D-SPARKLING at 7T.Joint Annual Meeting ISMRM-ESMRMB & ISMRT 31st Annual MeetingLondon, United KingdomMay 2022
  HALback to text
• 41 inproceedingsM.Mathieu Dagréou, P.Pierre Ablin, S.Samuel Vaiter and T.Thomas Moreau. A framework for bilevel optimization that enables stochastic and global variance reduction algorithms.Advances in Neural Information Processing Systems (NeurIPS)New Orleans, United StatesNovember 2022
  HAL
• 42 inproceedingsC.Chaithya Giliyar Radhakrishna, G.Guillaume Daval-Frérot, A.Aurélien Massire, B.Boris Mailhé, M.Mariappan Nadar, A.Alexandre Vignaud and P.Philippe Ciuciu. MORE-SPARKLING: non-cartesian trajectories with minimized off-resonance effects.Joint Annual Meeting ISMRM-ESMRMB & ISMRT 31st Annual MeetingLondon, United KingdomMay 2022
  HALback to text
• 43 inproceedingsJ.Julia Linhart, P. L.Pedro Luiz Coelho Rodrigues, T.Thomas Moreau, G.Gilles Louppe and A.Alexandre Gramfort. Neural Posterior Estimation of hierarchical models in neuroscience.GRETSI 2022 - XXVIIIème Colloque Francophone de Traitement du Signal et des ImagesNancy, FranceSeptember 2022, 1-3
  HAL
• 44 inproceedingsJ.Julia Linhart, A.Alexandre Gramfort and P. L.Pedro Luiz Coelho Rodrigues. Validation Diagnostics for SBI algorithms based on Normalizing Flows.NeurIPS 2022 - the 36th conference on Neural Information Processing Systems - Machine Learning and the Physical Sciences workshopNew Orleans, United StatesNovember 2022, 1-7
  HAL
• 45 inproceedingsC.Cédric Rommel, T.Thomas Moreau, J.Joseph Paillard and A.Alexandre Gramfort. CADDA: Class-wise Automatic Differentiable Data Augmentation for EEG Signals.ICLR 2022 - International Conference on Learning RepresentationsVirtual, FranceApril 2022
  HALback to text
• 46 inproceedingsA.Alexis Thual, H.Huy Tran, T.Tatiana Zemskova, N.Nicolas Courty, R.Rémi Flamary, S.Stanislas Dehaene and B.Bertrand Thirion. Aligning individual brains with Fused Unbalanced Gromov-Wasserstein.NeurIPS 2022 - Conference on Neural Information Processing SystemsNew-Orlean, FranceNovember 2022
  HAL
• 47 inproceedingsG. E.Gastón E Zanitti, M.Majd Abdallah and D.Demian Wassermann. Towards Heterogeneous Data Integration: The NeuroLang Approach.OHBM 2022 - Organization for Human Brain MappingGlasgow, United KingdomSeptember 2022
  HAL

Reports & preprints

• 48 miscZ.Zaineb Amor, C.Chaithya Giliyar Radhakrishna, G.Guillaume Daval-Frérot, B.Bertrand Thirion, F.Franck Mauconduit, P.Philippe Ciuciu and A.Alexandre Vignaud. 3D-SPARKLING en IRMf : Application à la rétinotopie.December 2022
  HAL
• 49 miscZ.Zaineb Amor, C.Caroline Le Ster, C.Chaithya Giliyar Radhakrishna, G.Guillaume Daval-Frérot, N.Nicolas Boulant, B.Bertrand Thirion, F.Franck Mauconduit, C.Christian Mirkes, P.Philippe Ciuciu and A.Alexandre Vignaud. Impact de la correction des imperfections $B_0$ sur les données IRMf collectées avec 3D-SPARKLING.December 2022
  HAL
• 50 miscG.Guillaume Daval-Frérot, A.Aurélien Massire, B.Boris Mailhé, M.Mariappan Nadar, B.Blanche Bapst, A.Alain Luciani, A.Alexandre Vignaud and P.Philippe Ciuciu. Deep learning-assisted model-based off-resonance correction for non-Cartesian susceptibility weighted imaging.November 2022
  HALback to text
• 51 miscR.Renata Porciuncula Baptista, M.Mathieu Naudin, C.Chaithya Giliyar Radhakrishna, G.Guillaume Daval-Frérot, F.Franck Mauconduit, A.Alexa Haeger, S.Sandro Romanzetti, M.Marc Lapert, P.Philippe Ciuciu, C.Cécile Rabrait-Lerman, R.Remy Guillevin, A.Alexandre Vignaud and F.Fawzi Boumezbeur. Accelerated sodium MRI using undersampled 3D SPARKLING at 7T.January 2023
  HAL
• 52 miscL.Louis Rouillard, T.Thomas Moreau and D.Demian Wassermann. PAVI: Plate-Amortized Variational Inference.May 2022
  HAL
• 53 miscG.Guillaume Staerman, C.Cédric Allain, A.Alexandre Gramfort and T.Thomas Moreau. FaDIn: Fast Discretized Inference for Hawkes Processes with General Parametric Kernels.October 2022
  HAL

Other scientific publications

• 54 inproceedingsC.Chaithya Giliyar Radhakrishna and P.Philippe Ciuciu. Benchmarking learned non-Cartesian k-space trajectories and reconstruction networks.Joint Annual Meeting ISMRM-ESMRMB & ISMRT 31st Annual MeetingLondon, United KingdomMay 2022
  HAL