PhD Position F/M Campagne doctorant 2024 - Emergence of mesoscale

Contexte et atouts du poste

This PhD project will be realized in the Inria NERV team, a research lab supported by the French institutions Inria, Inserm, CNRS, and Sorbonne University. The team is located in the Paris Brain Institute (ICM) within the Pitie-Salpetriere hospital.

The NERV team pursues a multidsciplinary research program at the intersection between biomedical engineering, complex systems and clinical neuroscience. NERV proposes new computational frameworks to analyze and model the spatiotemporal complexity of brain networks from multimodal and longitudinal neuroimaging data, and we design noninvasive intervention strategies based on brain-computer interfaces. Furthermore, the team ejoys a privileged position within a unique scientific and technological environment including comprehensive experimental core facilities (eg, neuroimaging, genetics, cellular), several animal models (eg, from nematodes to humans) and powerful centralized cluster computer system to realize big-data analysis and simulations.

Mission confiée

Context of the project

Artificial Intelligence (AI) and especially Deep Learning (DL) have undergone many successes in recent years in various domains of applications such as computer vision, speech recognition, language, domain recognition, decision-making, even outperforming the human capacities benchmark in most of them.

Those performances were mainly obtained by increasing scales : data augmentation and bigger models launched on GPUs and faster learning units. However many features of human ability described by cognitive sciences seem to remain completely out of reach for now. The main one being the generalizability beyond past experience, namely the adaptability to unknown contexts. Furthermore, deep learning algorithms always require a huge amount of data while adult brains can learn new tasks with a very few examples. So the question is how real brains came up with such e cient versatility and what are the associated organizational features?

Recent developments in network science have provided fresh insights into the structure and dynamics of the brain organization from a system perspective [ 1 , 2 ]. By modeling brains as graphs, with nodes accounting for brain regions and edges for anatomical/functional connections between them, a better understading of the organizational properties of the nervous system became possible [ 3 ]. Experimental evidence across disparate temporal and spatial scales indicated that brain networks tend to exhibit key topological features such as node centrality, modularity and efficiency. Notably, network modularity is a fundemental mesoscale property characterized by the presence of functionally specialized, yet interdependent modules, and o ers several advantages such as functional factorization, adaptability to new tasks, and robustness against perturbations [ 4 , 5 ]. Furthermore, brain network modularity is correlated to difference of performance across individuals [ 6 , 7 ] and plays an important role in combining information from differently specialized modules to perform more complex tasks. In artificial networks, recent studies demonstrated that modular architectures could lead to improved performance in learning different compositional tasks [8, 9 ]. Thus, a crucial question is to understand why, where, and when mesoscale properties such as modularity emerge during the learning process [ 10 ].

Principales activités

Objectives

The main goal of the PhD project is to elucidate the role of mesoscales network structures in generalizable artificial intelligence. Speci cally, this project aims to:

  + Conceive analytical network models that lead to the emergence of signi cant mesoscale attributes, such as modularity, by integrating developmental insights. Provide a foundational understanding of the necessary conditions (eg, network size, topology, density) for such emergent properties.

  + Compare the results with those obtained from the brain wiring formation of di erent species (eg, nematode, humans). Finetune the model parameters based on the above mentioned biologically data and derive a neurophysiologically plausible interpretation.

  +Develop a novel training framework that takes into account the model architecture, the learning algorithm and the multimodal nature of real inputs. Evaluate the overall performance when confronted with unfamiliar scenarios, thereby evaluating their versatilty and robustness.

Main Activities

+ Theoretical modeling . The initial phase of this doctoral research involves the development of

analytical models to understand the emergence and stability of significant mesoscale properties, such

as modularity, within biological networks during developmental processes. It is posited that modularity

manifests as a consistent outcome in neural networks influenced by a variety of parameters throughout

the development of organisms. This investigation aims to elucidate the prerequisites for such emergent

modularity across different species. Furthermore, the research will explore potential phase transitions

towards modular networks in response to variations in these parameters.

+ Convergence with biological data . In a second step we will test and fit those models on biological

data over several species on the whole lifespan from the embryonic stage of development to the

adult age. We will  rst study small species for which the whole brain networks (i.e. the connectomes)

are known. We will compare the mesoscale properties obtained in the synthetically-generated network

models and those in the actual connectomes. Connectomes needed to experimentally validate network

models are already available in the framework of different past and current research projects granted

to the PIs team.

+ Development of new artificial neural architectures . The last phase of this research project

will focus on leveraging biological insights to guide the design of artificial neural architectures, aiming

to foster the emergence of highly effcient network properties such as functional specialization, since

they have been shown as unable to achieve it [ 9 ]. Finally we also propose to explore how local

learning algorithms for energy-based models could play a role in artificial networks mesoscale properties

emergence such as modularity [ 11 ].

Compétences

Required skills

The ideal candidate should have a solid background in experimental physics, machine learning and data analysis, as well as experience in laboratory projects and simulations (Python, MATLAB). The ability and willingness to learn will do equally well.

Avantages

• Subsidized meals
• Partial reimbursement of public transport costs (75%)
• Leave: 7 weeks of annual leave + 10 extra days off due to RTT (statutory reduction in working hours) + possibility of exceptional leave (sick children, moving home, etc.)
• Possibility of teleworking
• Flexible organization of working hours (after 12 months)
• Professional equipment available (videoconferencing, loan of computer equipment, etc.)
• Social, cultural and sports events and activities