Thèse

The massive proliferation of DL methods in all fields of application has raised awareness of the two limitations of these very powerful methods for statistical data processing: the first is the exorbitant cost (both energy and financial) of training neural networks, and the second is the dependence on access to massive quantities of data and, above all, their annotations.
One way of getting around these limitations is to train the network as close as possible to the sensors, in order to limit the need for massive data movements and the associated infrastructure costs, as well as continuous learning to respond to data that may not all be available initially. The Edge computing trend is now attempting to bring computing closer to sensors by integrating AI into low-power electronic devices. This approach is already well studied, particularly in terms of efficient neuromorphic architectures, but it still only concerns the inference phase.
In this thesis project, we focus on the learning phase by studying the on-line learning capabilities of artificial neural networks. Although our ultimate aim is to develop techniques that benefit from the advantages of these neuromorphic architectures, the first phase of this study will consist of exploiting formal neural networks (DL) in order to create a basis for comparison in terms of the accuracy of our models, their computing cost, their data requirements and their energy consumption. By collecting data from different (multimodal) sensors, these intelligent systems can detect changes in the data and the environment that generates them. These changes may require the underlying neural network to learn incrementally, either to adjust its model to new conditions, or to learn new categories depending on the problem being addressed. In all cases, this relearning must be carried out in a reduced time and energy budget.
The problem of on-line learning has to contend with a number of obstacles that are still not the subject of consensus in the literature: novelty detection, learning with few or no labels, and catastrophic forgetting in the face of continuous learning of new data.
Yet the biological brain naturally manages the constant changes in our environment from a very early age right through to adulthood. It has a wide range of plasticity capacities, which are revealed at several levels of its organisation. In particular, it exploits the spatio-temporal correlations arising from the different sensory modalities it uses to apprehend its environment. These modalities merge and complement each other while being processed and routed by different neural pathways.
In this project, we therefore want to study how projection between modalities can help to improve the quality of lifelong learning, by overcoming the problem of catastrophic forgetting while reducing the need for data annotation.

Supervisors :
BENOIT MIRAMOND benoit.miramonduniv-cotedazur.fr (benoit[dot]miramond[at]univ-cotedazur[dot]fr)
Marina REYBOZ marina.reybozcea.fr (marina[dot]reyboz[at]cea[dot]fr) (Codirection)
Martial MERMILLOD martial.mermilloduniv-grenoble-alpes.fr (martial[dot]mermillod[at]univ-grenoble-alpes[dot]fr) (Codirection)

Keywords :
Incremental learning, Multimodal learning, Neuromorphic architectures, Embedded artificial intelligence

• Read more about GRIENAY Jonathan

Psychogenic non-epileptic seizures (PNES) constitute the most common subtype of functional neurological disorder (FND). The underlying pathophysiology remains unexplained and specific therapeutic tools are not present contributing to a poor prognosis and major healthcare costs. Behavioral and neuroimaging studies suggest disrupted emotional processes in PNES. However, the current data remains limited and inconsistent due to varying research methods. Therefore, PNES have been repeatedly associated with disrupted autonomic nervous system (ANS) patterns. The ANS is of particular interest because of its involvement in emotional processes. According to models, effective emotional regulation depends on a coupling between cognitive and autonomic responses. The hypothesis of a relationship between autonomic profiles, brain activity and emotional disorders in patients with PNES needs to be elucidated. An original approach to test whether PNES are linked to dysautonomic responses, consists of assessing how artificial modulation of ANS can modulate both emotional regulation and PNES. The biofeedback (BFB) is a non-invasive and efficient approach based on a neuromodulation of autonomic control. This technique has been emphasized as a potential therapeutic tool in epileptic disorders, but it has never been assessed in adult PNES. It is a promising tool for understanding the causes of emotional disorders and for developing a functional therapy for PNES.
The purpose of this project is to develop a personalized approach and an evidence-based therapeutic management, and aims to: (1) Explore the neurobiological substrates underlying emotional disorders in PNES, (2) Search for autonomic endophenotypes in patients, (3) Determine whether BFB is a beneficial therapeutic approach for the improvement of PNES, and (4) Develop a computational model of PNES. We hypothesize some PNES patients have a dysautonomic profile with a defect into the autonomic and emotional coupling, which may be improved by BFB, and this effect will be determined by a multimodal investigation and predicted by a computational analysis.
This multidisciplinary project will be part of a co-directed thesis initiating a new international collaboration between two Psychology and Neurocognition (Pr. Hot, Chambéry) and Neurology (Pr. Nguyen, Montréal, Canada) teams. Four sequential phases will be developed using an innovative experimental protocol. The first phase will investigate the emotional specificities in PNES. Based on a G power calculation, 21 subjects in 3 comparative groups each such as PNES, healthy controls and epileptic patients will be included. All subjects will perform emotional induction (T1) and re-exposure (T2) tasks combined with multimodal analyses including behaviour (B), psychophysiology (P) (heart rate variability (HRV) and skin conductance activity), and neuroimaging (N). Phase 2 will examine the therapeutic effect of HRV-BFB on PNES. Patients with PNES will carry out consecutive randomized BFB or pseudo BFB protocols for 6 weeks. Each patients will be their own control. The effects of HRV-BFB on change in seizure frequency and quality of life will be calculated. In Phase 3, patients will perform a final emotional re-exposure (T3), and B, P, and N measures will be repeated. According to the findings from Phase 1 and 2, the impact of BFB on emotional reactivity and potential changes in neurobiological patterns will be examined. In Phase 4, a computational model derived from the previous experimental data will be developed. This model will encode the influence of ANS on the occurrence of PNES, and assess the predictive value of different endophenotypes in the response to BFB. Given the major challenge posed by PNES, it seems essential to understand the underlying pathophysiology and to target it to develop an integrative personalized therapy. Finally, the large database and the computational model developed will enable further shared international collaborations.

Supervisors :
Pascal HOT - PR USMB - pascal.hotuniv-smb.fr (pascal[dot]hot[at]univ-smb[dot]fr)
Dang Khoa Nguyen - d.nguyenumontreal.ca (d[dot]nguyen[at]umontreal[dot]ca)

Keywords : Psychogenic nonepileptic seizures,Emotional regulation,Autonomic nervous system,Multimodal imaging,Biofeedback,Computational Neurosciences

• Read more about CARLIER Jasmine

Supervisors :
Benoit LEMAIRE benoit.lemaireuniv-grenoble-alpes.fr (benoit[dot]lemaire[at]univ-grenoble-alpes[dot]fr)
Sophie PORTRAT sophie.portratuniv-grenoble-alpes.fr (sophie[dot]portrat[at]univ-grenoble-alpes[dot]fr) (Co-encadrant)

Keywords : computational modeling,working memory,cognitive aging,experimental psychology,

• Read more about LEPROULT Inès

This project explores retrieval of the life story (autobiographical memory). It focusses on a key cognitive process: metacognition, so far never measured in autobiographical memory retrieval. We ask whether people can accurately judge the veracity of what they retrieve from their life story. We centre on one simple, elegant task which examines the ability to recall the order in which two events occurred, comparing how well people perform on this task with how well they think they have performed on this task. We will apply this method to older adults, with the aim of ultimately increasing wellbeing in our aging society; and to people with depression, who are known to have difficulties with retrieving memories from their life stories

Encadrants :
Céline SOUCHAY celine.souchayuniv-grenoble-alpes.fr (celine[dot]souchay[at]univ-grenoble-alpes[dot]fr)
Andrea TALES a.talesswansea.ac.uk (a[dot]tales[at]swansea[dot]ac[dot]uk) (Codirection)
Claire BARNES c.m.barnesswansea.ac.uk (c[dot]m[dot]barnes[at]swansea[dot]ac[dot]uk) (Co-encadrant)

Keywords : Autobiographical Memory,Metacognition,Depression,Aging,

• Read more about CARRERAS Fabien

Auditory verbal hallucinations characterise a large proportion of patients with schizophrenia. Their origin is thought to be due to a failure of internal language, and more specifically of its predictive control, leading to a loss of agentivity. The dysfunction, which disrupts the activity of the language connectome, is thought to involve the left inferior frontal gyrus. The overall aim of this thesis project is, on the one hand, to provide a better description of the neural dynamics of auditory verbal hallucinations and, on the other, to envisage a neurofeedback treatment based on these neural dynamics and on previous research carried out on healthy participants. In the first phase of this thesis, before looking at hallucinations, we are assessing the neural dynamics and brain correlates of language, and more specifically of internal language (endophasia) in healthy adults, using neurophysiological methods (sEEG and MEG).

Supervisor :
Monica BACIU Monica.Baciuuniv-grenoble-alpes.fr (Monica[dot]Baciu[at]univ-grenoble-alpes[dot]fr), Directrice de thèse

Keywords : language, inner speech, connectomics, language, neurocognition, neurophysiology, sEEG, MEG, hallucinations

• Read more about DONNELLY Taisha

The integration of multimodal data to decipher the neural processes and networks underlying cognition and behavior is a major challenge in cognitive neuroscience. This is associated with a current paradigm shift with neurocognitive models considering that human behaviors are enabled by complex interactions between cognitive functions. The fusion (integration) of multimodal data could not only compensate for the limitations of each modality, but also detect features that are intrinsically multimodal.
Objectives and research program. This research work is at crossroads between language and declarative memory and addresses the question of their interactive union in a multimodal and integrative perspective, with artificial intelligence methods. The project will have two dimensions, neurocognitive and neurocomputational. On the neurocognitive level, the objective is to validate and enrich this new theoretical framework LuM (language-union-memory) that we have recently developed (Roger et al., 2022) and which consists in considering that language and memory are two inseparable functions and that their evaluation must be done in an interactive joint way, rather than in isolation. The project will also have a transcognitive dimension and try to explore this framework in normal and pathological settings. On the neuro-computational side, we will use several analysis methods (unsupervised machine learning, graph theory and deep learning) for multimodal fusion and develop transfer learning in the framework of deep learning AI. Multimodal biomarkers (neuropsychology, functional and anatomical neuroimaging) have been acquired in our previous work.

Supervisors :
Monica BACIU Monica.Baciuuniv-grenoble-alpes.fr (Monica[dot]Baciu[at]univ-grenoble-alpes[dot]fr)
Martial MERMILLOD martial.mermilloduniv-grenoble-alpes.fr (martial[dot]mermillod[at]univ-grenoble-alpes[dot]fr) (Codirection)

Keywords : Language,Memory,Cross-cognitive Modeling,Cognitive Neuroscience,Computational Neuroscience,

• Read more about GUICHET Clément

Developmental speech sound disorders (DSDs) are a group of disorders that affect speech intelligibility in children. The aim of the EULALIES ANR project is to gain a better understanding of their prevalence and characteristics in France. In this thesis, the aim will be to investigate these disorders more closely in children and adolescents with mild to moderate intellectual disabilities. The aim will also be to study the development of phonological awareness in this population, i.e. the ability to consciously detect and manipulate the infra-lexical units (syllables, phonemes) of speech.

Supervisors
Anne VILAIN, MCF UGA, - GIPSA-Lab
Hélène LOEVENBRUCK, DR CNRS - LPNC

• Read more about BOILLEY Claire

Most research on reading acquisition and the cognitive processes involved relates to Indo-European languages, and mainly English (Share, 2008). Very few studies have looked at other language families, especially Semitic languages such as Arabic. As a result, theoretical models of reading and learning to read that are now widely accepted internationally may be highly biased. Instead of responding to the demand for the universality they claim, some postulates of these models might in fact only apply to Indo-European languages (Frost, 2012; Hansen, 2014). This is more problematic as theoretical views on learning to read have consequences for teaching practices.
In many Arabic-speaking countries, the teaching of reading adopts the principles of teaching French, emphasizing phonological recoding and the development of phonemic awareness. However, many studies underline that Arabic more than French calls for morphological treatments and more mobilizes visual and visuo-attentional treatments for the identification of letters and relevant units within written words (Saiegh- Haddad & Malatesha-Joshi, 2014).
The thesis project aims to create tools for training cognitive visuo-attentional processes. These training sessions will be offered to Arabic-speaking students who are just starting to learn to read. The performance progress of the students will be evaluated (pre-post training comparison) by comparison with untrained control groups. The objective is to offer a first experimental validation of training protocol, followed by validation on a larger scale in ecological conditions. Ultimately, validated training could be made available to teachers and be the subject of national recommendations.

Supervisors : 
- Sylviane VALDOIS -  Sylviane.Valdoisuniv-grenoble-alpes.fr (Sylviane[dot]Valdois[at]univ-grenoble-alpes[dot]fr)
- Dominique GUILLO - Dominique.GUILLOum6p.ma (Dominique[dot]GUILLO[at]um6p[dot]ma) -

Keywords : Orthographic complexity,visuo-attentional skills,Arabic language,Reading,Spelling,Children,

• Read more about GHANDOUR Alaa

Supervisors : Martine Bouvard, PREM USMB, martine.bouvarduniv-smb.fr (martine[dot]bouvard[at]univ-smb[dot]fr)

Keywords : psychological processes,functionnal neurological disorder,transdiagnostic,,

• Read more about BOUAMMARI Annabelle

Experimental studies of processes and representations involved in reading and in reading acquisition rely on a wide variety of behavioral tasks. This offers a wide array of tools for experimental studies, but poses a challenge to computational modeling studies. Indeed, most of the available computational models in the domain are restricted to simulating one, or a couple, of behavioral tasks. For instance, no current model is able to invert the information flow between orthographic and phonological representations, so as to simultaneously account for reading and spelling tasks. The main objective of this doctoral project is to define a probabilistic model of knowledge involved in all tasks related to reading, to simulate as many as possible. Indeed, using probabilities and Bayes’ theorem allows to “invert knowledge”, and thus to study mathematically the two-way information transfer between orthographic and phonological representations. Defining the model will rely on a family of models, the BRAID models, previously developed in the lab. The model will be evaluated on its capacity to simulate a wide variety of behavioral tasks, and to account for behavioral effects out of the scope of previous models.

Supervisors :
Marie-Line BOSSE Marie-line.Bosseuniv-grenoble-alpes.fr (Marie-line[dot]Bosse[at]univ-grenoble-alpes[dot]fr)
Julien DIARD julien.diarduniv-grenoble-alpes.fr (julien[dot]diard[at]univ-grenoble-alpes[dot]fr) (Codirection)

Keywords : orthographic knowledge,probabilistic modeling,reading acquisition,reading,spelling,

• Read more about CHARRIER Camille