2023Activity 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 $𝐲=\Gamma \left(𝐱\right)+𝐛$. 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  81, 131, 103, 86, 132, 116, 115, 101, 106, 102. 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  155, 153, 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  156, 114. A particular application has been on diffusion MRI (dMRI), where we have linked the dMRI signal with physiological tissue models of grey matter tissue  114. Still in MRI but in susceptibility weighted imaging, another approach  95 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 94, 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 64, 65, 76, 88. 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  84, 79, 123.

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  170, 150 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  123, 124, 82, 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 71, 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. 112, 134). Since the challenge is to train such models on limited and noisy data, we will extend our recent work  69 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  112 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  67 as well as in the probabilistic case  167. 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  113. 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  168 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  169. 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  73. 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  70 in high-dimensional settings which are specific to neuroimaging research. In sum, by leveraging recent advances in deductive database systems  73 and this novel DSL  169 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  118, as well as simulation-based approaches (knockoff inference  80 or conditional randomization tests  127). 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  141. 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  75 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  157 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 77. 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,  77 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  77, 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  161 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 68, 143. 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  148, 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  128 or advanced strategies such as channel, time or frequency masking or phase randomizations  122, 125. 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  92, 127. 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  145 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  145, 136 and time series processing  146, 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  63.

Recently the team has applied SSL to two large cohorts of clinical EEG data  72 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  87.

While rather small networks have been employed so far on EEG data  83, 160 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  138, 137 and adapted them to neurophysiological data 117, 100, 87, 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  152, 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  85, tissue microstructure with cognitive control  133, functional gradients on the cortical surface 98 to functional territory segregation 151.

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  166. 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  165, 90, 99 and reveals neurodegenerative processes  126, 109, 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  104, 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  119. 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  163 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 120. 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  93 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  96 or natural language processing 91, 97 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  91, 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 74, 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 144. They are at the root of heterogeneous data integration, such as multi-modal machine learning 71. 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  121. 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)  105. 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  104. 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  163, 162 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  72 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  124, 82 as well as our PySAP software  106 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  155, 153 to multiple acquisition setups and other downstream tasks (e.g. motion correction and correction of off-resonance artifacts related to $B_0$ inhomogeneities).

7.1.1 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

7.1.2 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

7.1.3 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

7.1.4 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/1.5.1
• News of the Year:

  We organized in July 2023 a benchmarking sprint with benchopt with more than 40 participants.

  During this sprint, we developed new benchmarks, gathered feedback, and saw which benchmarking tools were most needed. Most of the participants had never worked with benchopt. Their goal was to create new benchmarks using benchopt, ranging from optimization benchmarks – e.g. training a neural network with pytorch or jax on image classification – to more complex workflows – e.g. comparing different fMRI alignment or domain adaptation methods. The participants were all able to produce benchmarks within a few hours! Note that an important aspect of this benchmarking initiative is that the goal is not to re-code the various methods but to aggregate them from the different sources they are available from.

  Take aways from this sprint have been discussed in a blog post accessible here: https://notes.inria.fr/s/OuVJYf8bM#

• Publication:
  hal-03830604
• Contact:
  Thomas Moreau
• Participants:
  Thomas Moreau, Alexandre Gramfort, Mathurin Massias, Badr Moufad

7.1.5 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:
  Axa, BNP Parisbas Cardif, Dataiku, Nvidia, Chanel

7.1.6 joblib

• Keywords:
  Parallel computing, Cache
• Functional Description:
  Facilitate parallel computing and caching in Python.
• News of the Year:

  We organized a dev sprint for joblib in February 2023, where we performed many maintenance changes and refactored a big part of the code necessary to provide asynchronous results collection in the library. This novel result collection paradigm was merged in March 2023 and released in the library.

  See more info in this thread: https://x.com/tomamoral/status/1648436595677331462

• URL:
  https://­joblib.­readthedocs.­io/­en/­latest/
• Contact:
  Gael Varoquaux

7.1.7 MRI-NUFFT

• Keywords:
  MRI, Off-the-grid, Preconditioner, Brain MRI
• 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.

• URL:
  https://­mind-inria.­github.­io/­mri-nufft/
• Contact:
  Philippe Ciuciu

7.1.8 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

7.1.9 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

9.1.1 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. Finally, a new CIFRE PhD thesis will start in January 2024 where Mrs. Asma Tanabene will be working on self-supervised 3D MR image reconstruction using deep learning models.

9.1.2 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.

9.1.3 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.

9.2.1 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.

Participants: Demian Wassermann, Bertrand Thirion, Philippe Ciuciu, Thomas Moreau, Chaithya Giliyar Radhakrishna.

———————————–

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

10.2.1 Visits of international scientists

10.3.1 H2020 projects

CEA BlueSky project

• Title:
  Brain & Computers
• Duration:
  2023 ->
• 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${}^3$ 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 -> 2024
• 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 -> 2024
• 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

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, Lyon, 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

11.1.1 Scientific events: organisation

• T. Moreau
  2019-2023: Co-organizer of the Séminaire Palaisien, a monthly seminar in the Saclay ecosystem around statistics and machine learning (40 people).

11.1.2 Journal

11.1.3 Invited talks

• P. Ciuciu

   

  • August 2023: Keynote speaker at the 2023 GRETSI conference, Grenoble, France.
  • October 2023: Invited speaker at the Minnesota Workshop on high and ultra-high field imaging (Minneapolis, MI, USA).
  • October 2023: Invited speaker at the 2023 AMU Biomedical Imaging school in Cargese (Corsica, France).
• B.Thirion

   

  • March 2023: Data science Institute, Oxford: "How to remove or Control Confounds in predictive models, with application to brain biomarkers"
  • April 2023: Conference at CYCERON imaging center, CEA/CNRS, Caen, France
  • April 2023: Conference at the Brain Matters symposium "Synthesizing brain images from the neuroscientific literature"
  • June 2023: HBP WP2 final conference, 2 presentations; Barcelona, Spain
  • June 2023: One health day, Université Paris Saclay "Dealing with the data deluge: identifying important variables from multimodal health data".
• T. Moreau

   

  • June 2023: Invited speaker at the SIAM OP 2023 conference, Seattle, USA.
  • August 2023: Invited speaker at the ICIAM 2023 conference, Tokyo, Japan.
• D.Wassermann

   

  • October 2023: Neurolang Tutorial at Cogbases: Boosting Specificity in Cognitive Neuroimaging
  • October 2023: Keynote at Computational Diffusion MRI Workshop: How do we link diffusion MRI with brain structure and cognition? Breaching the gap between classical neuroanatomy and machine learning.
  • May 2023: Neural 2023: Specificity in Cognitive Neuroimaging: Pushing the Envelope in Meta-Analyses by Harnessing Rich Probabilistic Logical Models
  • March 2023: Neurosynth 2.0 group meeting: Specificity in Cognitive Neuroimaging: Pushing the Envelope in Meta-Analyses by Harnessing Rich Probabilistic Logical Models

11.1.4 Scientific expertise

• P. Ciuciu

  : 

  • European expert reviewer for the European Innovation Council Accelerator actions (main track: AI and health).
  • External reviewer for the European Research Council (ERC) Synergy grant.

11.1.5 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 Telecom Sud Paris Scientific Council
  • Member of Inria Commission évaluation

11.2.1 Teaching

• P. Ciuciu

   

  • Tutorial presenter at the 2023 IEEE ISBI: Computational MRI in the Deep Learning Era.
  • Instructor at the at Harmonic and Multifractal analysis summer school.
  • October 2023: Invited speaker at the 2023 AMU Biomedical Imaging school in Cargese (Corsica, France).
  • Lecturer at the Institut d'Optique Graduate School (3rd year, Signal & Images major).
  • Lecturer at the M2 ATSI (CentraleSupelec, ENS Paris-Saclay): Medical imaging course.
• D. Wassermann

   

  • Master MVA (École Polytechnique, École Normale Superiore, Centrale Supelec): Graphical Models
  • Master in Biomedical Engineering (Université Paris Descartes): Quantification in NeuroImaging.
• T. Moreau

   

  • Master Data Science (IPP/UP Saclay): Datacamp.
  • Executive Master Big Data (Université Dauphine): Introduction to Python.
• B.Thirion

   

  • MVA Master (École Polytechnique, École Normale Superiore, Centrale Supelec): Brain Function ; 12h
  • NeuroEngineering Master (UPSaclay): fMRI data analysi; 2h
  • Instructor at the HBP/Ebrains conference, March 2023: tutorial on fMRI data analysis on Ebrains ; 2h
  • Instructor at ESMRMB conference, Oct 2023, Basel. Two courses taught: "When does the use of statistics make sense ?" and "Help ! My sample is too small."

11.2.2 Supervision

• P. Ciuciu

   

  • C. Giliyar Radhakrishna, PhD 2020-2023
  • M. Dumeur (with M. Palva, Aalto Univ), PhD in cotutelle (4y), 2020-2024
  • Z. Amor (with A. Vignaud, CEA) PhD 2020-2024
  • P.-A. Comby (with A. Vignaud, CEA), PhD 2021-2024
  • S. Brosset, (with Z. Saghi, CEA) PhD 2022-2025
  • D. Nuñez-Fernandez (with A. Vignaud, CEA), PhD 2023-2026
  • M. Marzulli, M1 BME Paris, (June - Sep 2023)
  • L. Oudjman, work-study professional master's degree at EpiTech (Oct 2023 - Jul 2025)
• B. Thirion

   

  • Alexandre Pasquiou, PhD 2020-2023 (with C.Pallier)
  • Alexis Thual, PhD 2020-2024 (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)
  • Nicolas Salvy, PhD 2023-2026 (with H.Talbot)
• D. Wassermann

   

  • Gaston Zanitti, PhD 2020-2023
  • Louis Rouillard, PhD 2021-2024
  • Raphael Meudec, PhD 2021-2024 (with B. Thirion)
  • Alexandre Le Bris, PhD 2022-2025
  • Gabriela Gomez Jimenez, PhD 2023-2026 (with J. Valette CEA)
• T. Moreau

   

  • B. Malézieux (with M. Kowalski, UP Saclay), PhD 2020-2023
  • C. Allain (with A. Gramfort, Meta), PhD 2021-2024
  • M. Dagréou (with S. Vaiter, Université Cote d'Azur and P. Ablin, Apple), PhD 2021-2024
  • F. Michel (with M. Kowalski, UP Saclay), PhD 2022-2025
  • V. Loison (with J. Cartailler, APHP), PhD 2023-

11.2.3 Juries

• P. Ciuciu
  : External reviewer of the PhD thesis defended by Clément Poiret (Univ Paris-Cité), Paris, France.
• P. Ciuciu
  : PhD committee president of the PhD thesis defended by Pierre Barbot (Univ Paris-Saclay), L2S, Gif-sur-Yvette, France.
• B. Thirion
  External reviewer of the PhD thesis defended by Lucrezia Carboni (Grenoble Université), October 2023
• B. Thirion
  External reviewer of the PhD thesis defended by Etrit Haxholli  (Université Nice Côte d'Azur), October 2023
• B. Thirion
  PhD committee president for the PhD thesis defended by Adrien Pavao (Université Paris-Saclay), in Dec 2023.
• B. Thirion
  PhD committee president for the PhD thesis defended by Sydnez Krystal (Université Paris-Saclay), in Dec 2023.

11.3.1 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.

11.3.2 Education

• P.Ciuciu
  Talk given at the 3rd Doctoral Students day organized by DataIA for PhD students in AI/Data Science at the Paris-Saclay University, December 7.

International journals

• 4 articleZ.Zaineb Amor, C.Caroline Le Ster, C.Chaithya Gr, G.Guillaume Daval-Frérot, N.Nicolas Boulant, F.Franck Mauconduit, B.Bertrand Thirion, P.Philippe Ciuciu and A.Alexandre Vignaud. Impact of B0 field imperfections correction on BOLD sensitivity in 3D‐SPARKLING fMRI data.Magnetic Resonance in MedicineDecember 2023HALDOIback to text
• 5 articleC.Charlotte Caucheteux, A.Alexandre Gramfort and J.-R.Jean-Rémi King. Evidence of a predictive coding hierarchy in the human brain listening to speech.Nature Human BehaviourMarch 2023HALDOI
• 6 articleG.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 SWI.Magnetic Resonance in Medicine904June 2023, 1431-1445HALDOIback to text
• 7 articleC.Chengran Fang, D.Demian Wassermann and J.-R.Jing-Rebecca Li. Fourier representation of the diffusion MRI signal using layer potentials.SIAM Journal on Applied Mathematics831February 2023, 99-121HALDOI
• 8 articleC.Chengran Fang, Z.Zheyi Yang, D.Demian Wassermann and J.-R.Jing-Rebecca Li. A simulation-driven supervised learning framework to estimate brain microstructure using diffusion MRI.Medical Image Analysis90December 2023, 102979HALDOI
• 9 articleC.Chaithya Giliyar Radhakrishna and P.Philippe Ciuciu. Jointly Learning Non-Cartesian k-Space Trajectories and Reconstruction Networks for 2D and 3D MR Imaging through Projection.Bioengineering102January 2023, 158HALDOIback to text
• 10 articleC.Chaithya Giliyar Radhakrishna, G.Guillaume Daval-Frérot, A.Aurélien Massire, A.Alexandre Vignaud and P.Philippe Ciuciu. Improving spreading projection algorithm for rapid k‐space sampling trajectories through minimized off‐resonance effects and gridding of low frequencies.Magnetic Resonance in Medicine903May 2023, 1069-1085HALDOI
• 11 articleA.-S.Anne-Sophie Hamy, J.Judith Abécassis, K.Keltouma Driouch, L.Lauren Darrigues, M.Mathias Vandenbogaert, C.Cecile Laurent, F.Francois Zaccarini, B.Benjamin Sadacca, M.Myriam Delomenie, E.Enora Laas, O.Odette Mariani, T.Thanh Lam, B.Beatriz Grandal, M.Marick Laé, I.Ivan Bieche, S.Sophie Vacher, J.-Y.Jean-Yves Pierga, E.Etienne Brain, C.Celine Vallot, J.Judicael Hotton, W.Wilfrid Richer, D.Dario Rocha, Z.Zakia Tariq, V.Veronique Becette, D.Didier Meseure, L.Laetitia Lesage, A.Anne Vincent-Salomon, N.Natalie Filmann, J.Jenny Furlanetto, S.Sibylle Loibl, E.Elise Dumas, J.Joshua Waterfall and F.Fabien Reyal. Evolution of synchronous female bilateral breast cancers and response to treatment.Nature MedicineMarch 2023HALDOI
• 12 articleA.Apolline Mellot, A.Antoine Collas, P. L.Pedro Luiz Coelho Rodrigues, D.Denis Engemann and A.Alexandre Gramfort. Harmonizing and aligning M/EEG datasets with covariance-based techniques to enhance predictive regression modeling.Neuroscience ImagingNovember 2023, 1-26HALDOI
• 13 articleA.Alexandre Pasquiou, Y.Yair Lakretz, B.Bertrand Thirion and C.Christophe Pallier. Information-Restricted Neural Language Models Reveal Different Brain Regions' Sensitivity to Semantics, Syntax and Context: LANGUAGE MODELS SHOW BRAIN SENSITIVITY TO SEMANTICS, SYNTAX AND CONTEXT.Neurobiology of Language44December 2023, 611-636HALDOI
• 14 articleL.Lindsey Power, C.Cédric Allain, T.Thomas Moreau, A.Alexandre Gramfort and T.Timothy Bardouille. Using convolutional dictionary learning to detect task-related neuromagnetic transients and ageing trends in a large open-access dataset.NeuroImage267February 2023, 119809HALDOI
• 15 articleH.Hugo Richard and B.Bertrand Thirion. FastSRM: A fast, memory efficient and identifiable implementation of the shared response model.Aperture NeuroOctober 2023HALDOI
• 16 articleL.Louis Rouillard, A.Alexandre Le Bris, T.Thomas Moreau and D.Demian Wassermann. PAVI: Plate-Amortized Variational Inference.Transactions on Machine Learning Research Journal2023HAL
• 17 articleD.David Sabbagh, J.Jérôme Cartailler, C.Cyril Touchard, J.Jona Joachim, A.Alexandre Mebazaa, F.Fabrice Vallée, É.Étienne Gayat, A.Alexandre Gramfort and D. A.Denis A Engemann. Repurposing electroencephalogram monitoring of general anaesthesia for building biomarkers of brain ageing: an exploratory study.British Journal of Anaesthesia open7September 2023, 100145HALDOI
• 18 articleG.Guillaume Staerman, E.Eric Adjakossa, P.Pavlo Mozharovskyi, V.Vera Hofer, J.Jayant Sen Gupta and S.Stéphan Clémençon. Functional anomaly detection: a benchmark study.International Journal of Data Science and Analytics161June 2023, 101-117HALDOI
• 19 articleB.Bertrand Thirion, H.Himanshu Aggarwal, A. F.Ana Fernanda Ponce, A. L.Ana Luísa Pinho and A.Alexis Thual. Should one go for individual or group-level brain parcellations ? A deep-phenotyping benchmark.Brain Structure and FunctionNovember 2023HALDOI
• 20 articleB.Bertrand Thirion. On the statistics of brain/behavior associations.Aperture NeuroMarch 2023HALDOI
• 21 articleS. H.Sheng H. Wang, M.Marco Fuscà, F.Felix Siebenhühner, V.Vladislav Myrov, G.Gabriele Arnulfo, L.Lino Nobili, J. M.J Matias Palva and S.Satu Palva. Brain criticality predicts individual levels of inter-areal synchronization in human electrophysiological data.Nature Communications144736August 2023HALDOI
• 22 articleN.Nitin Williams, S.Sheng Wang, G.Gabriele Arnulfo, L.Lino Nobili, S.Satu Palva and M. J.Matias J. Palva. Modules in connectomes of phase-synchronization comprise anatomically contiguous, functionally related regions.NeuroImage272May 2023, 120036HALDOI
• 23 articleZ.Zheyi Yang, C.Chengran Fang and J.-R.Jing-Rebecca Li. Incorporating interface permeability into the diffusion MRI signal representation while using impermeable Laplace eigenfunctions.Physics in Medicine and Biology6817August 2023, 175036HALDOI

International peer-reviewed conferences

• 24 inproceedingsZ.Zaineb Amor, P.-A.Pierre-Antoine Comby, C.Caroline Le Ster, A.Alexandre Vignaud and P.Philippe Ciuciu. Non-Cartesian non-Fourier fmri imaging for high-resolution retinotopic mapping at 7 Tesla.CAMSAP 2023 - IEEE International Workshop on Computational Advances in Multi-Sensor Adaptive ProcessingLos Suenos, Costa RicaDecember 2023HAL
• 25 inproceedingsA.Alexandre Blain, B.Bertrand Thirion, O.Olivier Grisel and P.Pierre Neuvial. False Discovery Proportion control for aggregated Knockoffs.Proceedings of Advances in Neural Information Processing Systems 36 (NeurIPS 2023)NeurIPS 2023 – 37th Conference on Neural Information Processing SystemsNew Orleans, United States2023HALDOI
• 26 inproceedingsC.Clément Bonet, B.Benoît Malézieux, A.Alain Rakotomamonjy, L.Lucas Drumetz, T.Thomas Moreau, M.Matthieu Kowalski and N.Nicolas Courty. Sliced-Wasserstein on Symmetric Positive Definite Matrices for M/EEG Signals.ICML 2023 - Fortieth International Conference on Machine LearningHonololu, United States2023, 1-29HAL
• 27 inproceedingsA.Ahmad Chamma, D.Denis Engemann and B.Bertrand Thirion. Statistically Valid Variable Importance Assessment through Conditional Permutations.NeurIPS 2023 - 37th Conference on Neural Information Processing SystemsNew Orleans (Louisiana), United StatesDecember 2023HAL
• 28 inproceedingsA.Ahmad Chamma, B.Bertrand Thirion and D.Denis Engemann. Variable importance in high-dimensional settings requires grouping.AAAI 2024 - The 38th Annual AAAI Conference on Artificial IntelligenceVancouver, CanadaDecember 2023HAL
• 29 inproceedingsA.Antoine Collas, T.Titouan Vayer, R.Rémi Flamary and A.Arnaud Breloy. Entropic Wasserstein component analysis.IEEE International Workshop on Machine Learning for Signal Processing (MLSP)Rome, ItalySeptember 2023HAL
• 30 inproceedingsP.-A.Pierre-Antoine Comby, Z.Zaineb Amor, A.Alexandre Vignaud and P.Philippe Ciuciu. Denoising of FMRI volumes using local low rank methods.ISBI 2023 - International Symposium on Biomedical ImagingCarthagena de India, ColombiaIEEEApril 2023, 1-5HALDOI
• 31 inproceedingsT.Theo Gnassounou, R.Rémi Flamary and A.Alexandre Gramfort. Convolutional Monge Mapping Normalization for learning on sleep data.NeurIPS 2023 - 37th Conference on Neural Information Processing SystemsNouvelle Orléans, United StatesNovember 2023HAL
• 32 inproceedingsA.Ambroise Heurtebise, P.Pierre Ablin and A.Alexandre Gramfort. MultiView Independent Component Analysis with Delays.MLSP 2023 - 33rd IEEE International Workshop on Machine Learning for Signal ProcessingRome, ItalySeptember 2023HAL
• 33 inproceedingsJ.Julia Linhart, A.Alexandre Gramfort and P. L.Pedro L C Rodrigues. L-C2ST: Local Diagnostics for Posterior Approximations in Simulation-Based Inference.NeurIPS 2023 - 37th Conference on Neural Information Processing SystemsNew Orleans (LA), United States2023, 1-27HALDOI
• 34 inproceedingsG.Guillaume Staerman, C.Cédric Allain, A.Alexandre Gramfort and T.Thomas Moreau. FaDIn: Fast Discretized Inference for Hawkes Processes with General Parametric Kernels.PMLRInternational Conference on Machine Learning202Honololu, Hawaii, United StatesJuly 2023, 32575-32597HALback to text
• 35 inproceedingsM. S.Maria Sayu Yamamoto, A.Apolline Mellot, S.Sylvain Chevallier and F.Fabien Lotte. Novel SPD matrix representations considering cross-frequency coupling for EEG classification using Riemannian geometry.EUSIPCO 2023 - The 31st European Signal Processing ConferenceHelsinki, FinlandSeptember 2023HAL

Conferences without proceedings

• 36 inproceedingsZ.Zaineb Amor, C.Caroline Le Ster, C.Chaithya Giliyar Radhakrishna, G.Guillaume Daval-Frérot, B.Bertrand Thirion, N.Nicolas Boulant, F.Franck Mauconduit, M.Mirkes Christian, P.Philippe Ciuciu and A.Alexandre Vignaud. Impact of ${\mathrm{B}}_0$ field imperfections correction on BOLD sensitivity in 3D-SPARKLING fMRI data.ISMRM & ISMRT 2023 - Annual Meeting & Exhibition - Annual Meeting of the International Society for Magnetic Resonance in MedicineToronto, CanadaJune 2023HAL
• 37 inproceedingsA.Alexandre Pasquiou, Y.Yair Lakretz, B.Bertrand Thirion and C.Christophe Pallier. Probing brain context-sensitivity with masked-attention generation.CCN 2023 - Conference on Cognitive Computational NeuroscienceOxford, United Kingdom2023HALDOI

Scientific book chapters

• 38 inbookF.Florent Bouchard, A.Arnaud Breloy, A.Antoine Collas, A.Alexandre Renaux and G.Guillaume Ginolhac. The Fisher-Rao geometry of CES distributions.Elliptical Distributions in Signal ProcessingOctober 2023HAL

Doctoral dissertations and habilitation theses

• 39 thesisC.Charlotte Caucheteux. Language representations in deep learning algorithms and the brain.Université Paris-SaclayMay 2023HAL
• 40 thesisC.Chengran Fang. Neuron modeling, Bloch-Torrey equation, and their application to brain microstructure estimation using diffusion MRI.Université Paris-SaclayFebruary 2023HAL
• 41 thesisC.Chaithya Giliyar Radhakrishna. Designing and learning non-Cartesian k-space sampling trajectories for accelerated 3D MRI.Université Paris-SaclayApril 2023HAL
• 42 thesisB.Benoît Malézieux. A study of unrolled algorithms for dictionary learning and inverse problems, and contributions to M/EEG signal processing.Université Paris-SaclaySeptember 2023HAL
• 43 thesisA.Alexandre Pasquiou. Deciphering the neural bases of language comprehension using latent linguistic representations.Université Paris-SaclayJune 2023HAL
• 44 thesisG. E.Gaston Ezequiel Zanitti. Development of a probabilistic domain-specific language for brain connectivity including heterogeneous knowledge representation.Université Paris-SaclayMarch 2023HAL

Reports & preprints

• 45 miscZ.Zaineb Amor, P.Philippe Ciuciu, C.Chaithya Giliyar Radhakrishna, G.Guillaume Daval-Frérot, F.Franck Mauconduit, B.Bertrand Thirion and A.Alexandre Vignaud. Non-cartesian 3D-SPARKLING vs cartesian 3D-EPI encoding schemes for functional magnetic resonance Imaging at 7 Tesla.December 2023HAL
• 46 reportF.Frédéric Blanqui, A.Anne Canteaut, H.Hidde de Jong, S.Sébastien Imperiale, N.Nathalie Mitton, G.Guillaume Pallez, X.Xavier Pennec, X.Xavier Rival and B.Bertrand Thirion. Recommandations sur les « éditeurs de la zone grise ».InriaJanuary 2023, 1-3HAL
• 47 reportF.Frédéric Blanqui, A.Anne Canteaut, H. D.Hidde De Jong, S.Sébastien Imperiale, N.Nathalie Mitton, G.Guillaume Pallez, X.Xavier Pennec, X.Xavier Rival and B.Bertrand Thirion. Recommendations on "Grey-Zone Publishers": Recommendations from the Inria Evaluation Committee, translated from https://hal.inria.fr/hal-04001505.InriaJanuary 2023, 1-3HAL
• 48 miscM.Mathieu Dagréou, T.Thomas Moreau, S.Samuel Vaiter and P.Pierre Ablin. A lower bound and a near-optimal algorithm for bilevel empirical risk minimization.2023HALDOI
• 49 miscH.Ha Nguyen, J.Javier Albayay, R.Richard Höchenberger, S.Surabhi Bhutani, S.Sanne Boesveldt, N. A.Niko A. Busch, I.Ilja Croijmans, K. W.Keiland W. Cooper, J. H.Jasper H. B. de Groot, M. C.Michael C. Farruggia, A. W.Alexander W. Fjaeldstad, J. E.John E. Hayes, T.Thomas Hummel, P. V.Paule V. Joseph, T. K.Tatiana K. Laktionova, T.Thierry Thomas-Danguin, M. G.Maria G. Veldhuizen, V. V.Vera V. Voznessenskaya, V.Valentina Parma, M. Y.M. Yanina Pepino and K.Kathrin Ohla. Covid-19 affects taste independently of smell: results from a combined chemosensory home test and online survey from a global cohort (N=10,953).January 2023HALDOI
• 50 miscA. L.Ana Luísa Pinho, H.Hugo Richard, M.Michael Eickenberg, A.Alexis Amadon, E.Elvis Dohmatob, I.Isabelle Denghien, J. J.Juan Jesús Torre, S.Swetha Shankar, H.Himanshu Aggarwal, A. F.Ana Fernanda Ponce, A.Alexis Thual, T.Thomas Chapalain, C.Chantal Ginisty, S.Séverine Becuwe-Desmidt, S.Séverine Roger, Y.Yann Lecomte, V.Valérie Berland, L.Laurence Laurier, V.Véronique Joly-Testault, G.Gaëlle Médiouni-Cloarec, C.Christine Doublé, B.Bernadette Martins, G.Gaël Varoquaux, S.Stanislas Dehaene, L.Lucie Hertz-Pannier and B.Bertrand Thirion. Individual Brain Charting third release, probing brain activity during Movie Watching and Retinotopic Mapping.November 2023HAL
• 51 miscR. A.Russell A. Poldrack, C. J.Christopher J. Markiewicz, S.Stefan Appelhoff, Y. K.Yoni K. Ashar, T.Tibor Auer, S.Sylvain Baillet, S.Shashank Bansal, L.Leandro Beltrachini, C. G.Christian G. Benar, G.Giacomo Bertazzoli, S.Suyash Bhogawar, R. W.Ross W. Blair, M.Marta Bortoletto, M.Mathieu Boudreau, T. L.Teon L. Brooks, V. D.Vince D. Calhoun, F. M.Filippo Maria Castelli, P.Patricia Clement, A. L.Alexander L Cohen, J.Julien Cohen-Adad, S.Sasha d'Ambrosio, G.Gilles de Hollander, M.María de la Iglesia-Vayá, A.Alejandro de la Vega, A.Arnaud Delorme, O.Orrin Devinsky, D.Dejan Draschkow, E. P.Eugene Paul Duff, E.Elizabeth Dupre, E.Eric Earl, O.Oscar Esteban, F. W.Franklin W. Feingold, G.Guillaume Flandin, A.Anthony Galassi, G.Giuseppe Gallitto, M.Melanie Ganz, R.Rémi Gau, J.James Gholam, S. S.Satrajit S. Ghosh, A.Alessio Giacomel, A. G.Ashley G Gillman, P.Padraig Gleeson, S.Samuel Guay, T.Tal Yarkoni, K. J.Krzysztof J. Gorgolewski, A.Alexandre Gramfort, G.Giacomo Guidali, Y. O.Yaroslav O. Halchenko, D. A.Daniel A. Handwerker, N.Nell Hardcastle, P.Peer Herholz, D.Dora Hermes, C. J.Christopher J. Honey, R. B.Robert B. Innis, H.-I.Horea-Ioan Ioanas, A.Andrew Jahn, A.Agah Karakuzu, D. B.David B. Keator, G.Gregory Kiar, B.Balint Kincses, A. R.Angela R. Laird, J. C.Jonathan C. Lau, A.Alberto Lazari, J. H.Jon Haitz Legarreta, A.Adam Li, X.Xiangrui Li, B. C.Bradley C. Love, H.Hanzhang Lu, G.Giacomo Mazzamuto, S.Sajjad Torabian, B.Bradley Voytek, J.Julius Welzel, M.Martin Wilson, C.Camille Maumet, S. L.Steven L. Meisler, M.Mark Mikkelsen, H.Henk Mutsaerts, T. E.Thomas E. Nichols, A.Aki Nikolaidis, G.Gustav Nilsonne, G.Guiomar Niso, M.Martin Norgaard, T. W.Thomas W Okell, R.Robert Oostenveld, E.Eduard Ort, P. J.Patrick J. Park, M.Mateusz Pawlik, C. R.Cyril R. Pernet, F.Franco Pestilli, J.Jan Petr, C.Christophe Phillips, J.-B.Jean-Baptiste Poline, L.Luca Pollonini, P. R.Pradeep Reddy Raamana, P.Petra Ritter, G.Gaia Rizzo, K. A.Kay A. Robbins, A. P.Alexander P. Rockhill, C.Christine Rogers, A.Ariel Rokem, C.Chris Rorden, J. M.Jose Manuel Saborit-Torres, A. M.Alexandre M Routier, R. E.Robert E. Smith, T.Tamas Spisak, J.Julia Sprenger, N. C.Nicole C. Swann, M.Martin Szinte, S.Sylvain Takerkart, B.Bertrand Thirion, A. G.Adam G. Thomas, T.Taylor Salo, M.Michael Schirner and G.Gaël Varoquaux. The Past, Present, and Future of the Brain Imaging Data Structure (BIDS).September 2023HAL
• 52 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 2023HAL
• 53 miscH.Hugo Richard, E.Etienne Boursier and V.Vianney Perchet. Constant or logarithmic regret in asynchronous multiplayer bandits.May 2023HAL
• 54 miscM.M Roascio, S.Sh Wang, V.V Myrov, F.F Siebenhühner, R.R. Tro, P.P. Mattioli, F.F. Famà, S.S. Morbelli, M.M. Pardini, J.Jm Palva, D.D Arnaldi and G.G. Arnulfo. Altered Brain Dynamics in idiopathic REM sleep behavior disorder: Implications for a continuum from prodromal to overt alpha-synucleinopathies.May 2023HALDOI
• 55 miscM.Matthieu Terris and T.Thomas Moreau. Meta-Prior: Meta learning for Adaptive Inverse Problem Solvers.November 2023HAL
• 56 miscM.Matthieu Terris, T.Thomas Moreau, N.Nelly Pustelnik and J.Julian Tachella. Equivariant plug-and-play image reconstruction.December 2023HAL
• 57 miscA.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.2022HALDOI
• 58 miscS.Sheng Wang, G.Gabriele Arnulfo, L.Lino Nobili, V.Vladislav Myrov, P.Paul Ferrari, P.Philippe Ciuciu, S.Satu Palva and J. M.J Matias Palva. Neuronal Synchrony and Critical Bistability: Mechanistic Biomarkers for Localizing the Epileptogenic Network.May 2023HALDOI
• 59 miscH.-T.Hao-Ting Wang, S. L.Steven L Meiseler, H.Hanad Sharmarke, N.Natasha Clarke, N.Nicolas Gensollen, C. J.Christopher J Markiewicz, F.François Paugam, B.Bertrand Thirion and P.Pierre Bellec. Continuous Evaluation of Denoising Strategies in Resting-State fMRI Connectivity Using fMRIPrep and Nilearn.April 2023HALDOI

Other scientific publications

• 60 inproceedingsS. H.Sheng H. Wang, P.Paul Ferrari, A.Angel Hernandez, G.Gabriele Arnulfo, L.Lino Nobili, S.Satu Palva, P.Philippe Ciuciu and J. M.J Matias Palva. Improving epileptogenic zone (EZ) classification by combining brain criticality and connectivity features.International Society for the Advancement of Clinical MagnetoencephalographyOsaka (JP), JapanMay 2023HAL
• 61 inproceedingsS. H.Sheng H Wang, M.Morgane Marzulli, P.Paul Ferrari, G.Gabriele Arnulfo, L.Lino Nobili, S.Satu Palva, P.Philippe Ciuciu and J. M.J Matias Palva. A Low-dimensional surface in Eigen Feature Space Characterizes an Interictal Continuum.American Epilepsy Society ConferenceOrlando (FL), United StatesDecember 2023HAL