Post-doctorant (F/H) Post-doc in Hardware-aware Neural Architecture Optimization On The Edge

**Type de contrat**: CDD

**Niveau de diplôme exigé**: Thèse ou équivalent

**Fonction**: Post-Doctorant

**Niveau d'expérience souhaité**: Jeune diplômé

**A propos du centre ou de la direction fonctionnelle**:
Le centre Inria de l'Université de Lille, créé en 2008, emploie 360 personnes dont 305 scientifiques répartis dans 15 équipes de recherche. Reconnu pour sa forte implication dans le développement socio-économique de la région des Hauts-De-France, le centre Inria de l'Université de Lille entretient des relations étroites avec les grandes entreprises et les PME. En favorisant les synergies entre chercheurs et industriels, Inria participe au transfert de compétences et d'expertise dans le domaine des technologies numériques et donne accès au meilleur de la recherche européenne et internationale au bénéfice de l'innovation et des entreprises, notamment dans la région.

Depuis plus de 10 ans, le centre Inria de l'Université de Lille est situé au cœur de l'écosystème universitaire et scientifique lillois, ainsi qu'au cœur de la Frenchtech, avec un showroom technologique basé avenue de Bretagne à Lille, sur le site d'excellence économique EuraTechnologies dédié aux technologies de l'information et de la communication (TIC).

**Contexte et atouts du poste**:
**Mission confiée**:
Deep Neural Networks (DNN) and hardware accelerators are both leading forces for the recent progress in Edge AI. On the one hand, a new neural architectural paradigm is proposed each month, striving for more accuracy and efficiency. On the other hand, the hardware market has shifted towards designing devices that ensure both flexibility and generality for less energy demands while satisfying the user experience with less latency.
When DNN models are implemented on resource-constrained systems (e.g., edge computing), it becomes inevitable to meticulously optimize them to strike the optimal balance between accuracy, execution latency and energy efficiency. In order to address this particular difficulty, our objective in this project is to tackle Hardware-aware Neural Architecture Search (HW-aware NAS) as a new AutoML paradigm, targeting edge systems. HW-aware NAS incorporates hardware efficiency as an additional optimization objective during the neural architecture design space exploration.

Main objectives of the project:

- New multi-objective performance surrogates: In our previous work, we have widely used two types
of accuracy estimation strategies: Predictive Models and Weight-sharing Supernetworks. Several
other methods, including zero-shot estimation and learning-curve extrapolation have recently
emerged. However, these methods still face certain limitations in terms of multi-objectivity,
scalability, and accuracy.- Search Algorithms and Large Search Spaces: It is crucial to develop efficient and scalable search

algorithms that can effectively explore large heterogenous search spaces within practical time
constraints. This would allow for the discovery of highly optimized architectures that align with the
hardware constraints of edge devices, while still meeting performance requirements. Exploring
spaces with new approaches such quantum-inspired search algorithms holds promise in tackling
the challenges of searching large spaces more efficiently. These algorithms have the potential to
enhance the search process and expedite the discovery of optimal architectures for edge
computing. Large Language Model (LLM) based search algorithms for HW-NAS are another
interesting approach that will be explored in the project.- Multi-task and multi-modality NN investigation: Multi-task deep learning models are crucial to

reducing the memory occupancy and execution time especially for edge devices. Investigating how
sharing knowledge and architectural components across multiple related tasks can lead to more
efficient and effective neural architectures. This approach exploits shared representations to
enhance model performance and reduce resource requirements. In the same context we will
explore HW-NAS for Multimodal Neural Networks (MM-NN). These NN have the ability to
effectively process and integrate multiscale information from diverse data sources.

**Principales activités**:
Main objectives of the project:

- New multi-objective performance surrogates: In our previous work, we have widely used two types
of accuracy estimation strategies: Predictive Models and Weight-sharing Supernetworks. Several
other methods, including zero-shot estimation and learning-curve extrapolation have recently
emerged. However, these methods still face certain limitations in terms of multi-objectivity,
scalability, and accuracy.- Search Algorithms and Large Search Spaces: It is crucial to develop efficient and scalable search

algorithms that can effectively explore large heterogenous search spaces within practical time
constraints. This would allow for the discovery of highly optimized architectures that align with the
hardware constraints of edge devices,