Search results

Abstract Deep learning has become an extremely important technology in numerous domains such as computer vision, natural language processing, and autonomous systems. As neural networks grow in size and complexity to meet the demands of these applications, the cost of designing and training efficient models continues to rise in computation and energy consumption. Neural Architecture Search (NAS) has emerged as a promising solution to automate the design of performant neural networks. However, conventional NAS methods often require evaluating thousands of architectures, making them extremely resource-intensive and environmentally costly.This thesis introduces a novel, energy-aware NAS pipeline that operates at the intersection of Software Engineering and Machine Learning. We present CNNGen, a domain-specific generator for convolutional architectures, combined with performance and energy predictors to drastically reduce the number of architectures that need full training. These predictors are integrated into a multi-objective genetic algorithm (NSGA-II), enabling an efficient search for architectures that balance accuracy and energy consumption.Our approach explores a variety of prediction strategies, including sequence-based models, image-based representations, and deep metric learning, to estimate model quality from partial or symbolic representations. We validate our framework across three benchmark datasets, CIFAR-10, CIFAR-100, and Fashion-MNIST, demonstrating that it can produce results comparable to state-of-the-art architectures with significantly lower computational cost. By reducing the environmental footprint of NAS while maintaining high performance, this work contributes to the growing field of Green AI and highlights the value of predictive modelling in scalable and sustainable deep learning workflows. Jury Prof. Wim Vanhoof - University of Namur, BelgiumProf. Gilles Perrouin - University of Namur, BelgiumProf. Benoit Frénay - University of Namur, BelgiumProf. Pierre-Yves Schobbens - University of Namur, BelgiumProf. Clément Quinton - University of Lille, FranceProf. Paul Temple- University of Rennes, FranceProf. Schin'ichi Satoh - National Institute of Informatics, Japan

Abstract Deep learning has become an extremely important technology in numerous domains such as computer vision, natural language processing, and autonomous systems. As neural networks grow in size and complexity to meet the demands of these applications, the cost of designing and training efficient models continues to rise in computation and energy consumption. Neural Architecture Search (NAS) has emerged as a promising solution to automate the design of performant neural networks. However, conventional NAS methods often require evaluating thousands of architectures, making them extremely resource-intensive and environmentally costly.This thesis introduces a novel, energy-aware NAS pipeline that operates at the intersection of Software Engineering and Machine Learning. We present CNNGen, a domain-specific generator for convolutional architectures, combined with performance and energy predictors to drastically reduce the number of architectures that need full training. These predictors are integrated into a multi-objective genetic algorithm (NSGA-II), enabling an efficient search for architectures that balance accuracy and energy consumption.Our approach explores a variety of prediction strategies, including sequence-based models, image-based representations, and deep metric learning, to estimate model quality from partial or symbolic representations. We validate our framework across three benchmark datasets, CIFAR-10, CIFAR-100, and Fashion-MNIST, demonstrating that it can produce results comparable to state-of-the-art architectures with significantly lower computational cost. By reducing the environmental footprint of NAS while maintaining high performance, this work contributes to the growing field of Green AI and highlights the value of predictive modelling in scalable and sustainable deep learning workflows. Jury Prof. Wim Vanhoof - University of Namur, BelgiumProf. Gilles Perrouin - University of Namur, BelgiumProf. Benoit Frénay - University of Namur, BelgiumProf. Pierre-Yves Schobbens - University of Namur, BelgiumProf. Clément Quinton - University of Lille, FranceProf. Paul Temple- University of Rennes, FranceProf. Schin'ichi Satoh - National Institute of Informatics, Japan

Abstract Deep learning has become an extremely important technology in numerous domains such as computer vision, natural language processing, and autonomous systems. As neural networks grow in size and complexity to meet the demands of these applications, the cost of designing and training efficient models continues to rise in computation and energy consumption. Neural Architecture Search (NAS) has emerged as a promising solution to automate the design of performant neural networks. However, conventional NAS methods often require evaluating thousands of architectures, making them extremely resource-intensive and environmentally costly.This thesis introduces a novel, energy-aware NAS pipeline that operates at the intersection of Software Engineering and Machine Learning. We present CNNGen, a domain-specific generator for convolutional architectures, combined with performance and energy predictors to drastically reduce the number of architectures that need full training. These predictors are integrated into a multi-objective genetic algorithm (NSGA-II), enabling an efficient search for architectures that balance accuracy and energy consumption.Our approach explores a variety of prediction strategies, including sequence-based models, image-based representations, and deep metric learning, to estimate model quality from partial or symbolic representations. We validate our framework across three benchmark datasets, CIFAR-10, CIFAR-100, and Fashion-MNIST, demonstrating that it can produce results comparable to state-of-the-art architectures with significantly lower computational cost. By reducing the environmental footprint of NAS while maintaining high performance, this work contributes to the growing field of Green AI and highlights the value of predictive modelling in scalable and sustainable deep learning workflows. Jury Prof. Wim Vanhoof - University of Namur, BelgiumProf. Gilles Perrouin - University of Namur, BelgiumProf. Benoit Frénay - University of Namur, BelgiumProf. Pierre-Yves Schobbens - University of Namur, BelgiumProf. Clément Quinton - University of Lille, FranceProf. Paul Temple- University of Rennes, FranceProf. Schin'ichi Satoh - National Institute of Informatics, Japan

Our keynote speakers for 2026 are Professor Roosmarijn Vandenbroucke (Ghent University) and Professor Nelly Litvak (Eindhoven University of Technology). More information on the "Women in Science" website

Abstract Deep learning has become a central part of everyday life, and has given us powerful digital tools such as machine translators, voice assistants, and large language models. The Deaf communities, however, benefit verylittle from this progress. Sign languages are full natural languages, but they have no widely used written form, so the large text corpora and pretrained models that support these tools for spoken languages have no direct equivalent. As a result, sign language processing is held back by two related problems: a persistent shortage of annotated data, and a lack of tools to analyze the languages themselves.This thesis focuses on one component of the sign language processing pipeline that has received comparatively little attention: sign language segmentation, the temporal partitioning of a continuous signing video into individual sign units. Segmentation sits between raw video and almost every symbolic task built on top of it, yet it remains poorly understood. The goal of this work is twofold: to advance the segmentation task itself, and to turn sign language processing research into tools that are usable by the Deaf community, with French Belgian Sign Language (LSFB) as the main setting.The central contribution of this thesis is a set of segmentation models. We first study approaches based on recurrent neural networks, and show that their main difficulty lies in the modeling of the transitions between signs, the brief ambiguous movements that separate one sign from the next and that are easily confused with the signs themselves. This insight motivates the Hydra framework, a new approach that detects each sign as a whole unit rather than deciding, frame by frame, whether a sign is being produced. Concretely, instead of classifying every frame, the model predicts the distance from each frame to the nearest sign boundaries, which makes it less sensitive to the ambiguous regions and improves consistently over prior methods. Around this core, we present further contributions: a sign language-to-text dictionary that recognizes signs from an ordinary webcam, supporting work on isolated sign recognition, and a collaborative platform that gathers new sign data as a side effect of everyday use. Together, these contributions narrow the gap between sign language processing research and the people whose language it concerns. Jury Prof. Benoit Frénay - University of Namur, BelgiumProf. Anthony Cleve - University of Namur, BelgiumProf. Katrien Beuls - University of Namur, BelgiumProf. Barbara Hammer - Bielefeld University, GermanyProf. Michèle Gouiffès - Paris-Saclay University, FranceDr. Mathieu De Coster - Ghent University, Belgium I am registering to attend Pierre Poitier's thesis defense

In recent years, public debate has been regularly disrupted by various attempts to distort "information." While this phenomenon is as old as information itself, digital technology, the powerful interactivity offered by online platforms, and recent developments in generative artificial intelligence have given it a whole new dimension. Manipulated online information spreads virally, reaching internet users around the world in a matter of minutes. It poses serious risks to society and threatens democracy and the rule of law. To see this, one need only recall events such as the US and French presidential elections, the Covid-19 pandemic, the war in Ukraine, the terrorist attack at the Nova festival in 2023, and the war in Gaza.  Faced with this phenomenon, actors from all sectors have been mobilizing for several years. This conference offers a legal perspective on the manipulation of online information in relation to freedom of expression and the right to information from the point of view of European law. It focuses, on the one hand, on the response of the European Union legislator to combat online disinformation and, on the other hand, on mechanisms for protecting "quality" information.Free admission.