Pour participer vous devez vous inscrire ici.
La présentation "Intelligence artificielle pour le dimensionnement des systems énergériques hybrids" d'Inoussa Legreneaura lieu lors de cet après-midi. Pizza et café seront offerts à partir de 12h00.

The rapid evolution of deep learning has made Neural Architecture Search (NAS) essential for automating the design of high-performance models. While NAS aims to identify optimal architectures within vast search spaces, its effectiveness is often hindered by two primary challenges: the inefficient exploration of massive design spaces and the prohibitive computational cost of evaluating numerous candidates. One-shot NAS methods have significantly reduced these costs by jointly training multiple models simultaneously. However, these methods often produce inaccurate performance estimations, leading to suboptimal search navigation and final model selection. To overcome these limitations, we propose two distinct improvements to the NAS pipeline. First, we introduce an improved search space design using Monte Carlo Tree Search (MCTS) to navigate the architecture space more effectively. Second, we introduce a density-aware sampling method to ensure more reliable performance estimations for architectures. These contributions provide robust frameworks for both searching and training, enabling more efficient discovery of optimized models for real-world tasks.

Bio : Mehraveh Javan is a PhD student affiliated with LIVIA, under the supervision of Prof. Marco Pedersoli and Prof. Matthew Toews. Her research focuses on optimizing deep learning models for computer vision using Neural Architecture Search (NAS) methods, particularly focusing on improving one-shot NAS methods, search space design and optimization under hardware constraints.