A remarkable ability, that one may take for granted, is the ability to name object that you encounter for the first time when these are members of categories you know. This is possible thanks to our conceptual system and our ability to generalize names between same category members. It is priceless for organizing the world around us and communicati...
The thesis focuses on the concept of generalization, particularly in the context of supervised machine learning classification. This approach involves learning to solve a task (classification) based on labeled training data. Generalization is defined as the ability to make accurate predictions on unseen data during training. Traditionally, generali...
Traditional linguistic theories have long posited that human language competence is founded on innate structural properties and symbolic representations. However, Transformer-based language models, which learn language representations from unannotated text, have excelled in various natural language processing (NLP) tasks without explicitly modeling...
This thesis aims at enhancing the generalization abilities of deep neural networks, a critical step towards fair and reliable artificial intelligence. Specifically, we address the drop in performance when models are evaluated on test samples with a distribution shift with respect to the train samples. To this end, we focus on ensembling strategies:...
Deep learning has made significant progress in various fields and has emerged as a promising tool for modeling physical dynamical phenomena that exhibit highly nonlinear relationships. However, existing approaches are limited in their ability to make physically sound predictions due to the lack of prior knowledge and to handle real-world scenarios ...
Deep learning has emerged as a powerful approach for modelling static data like images and more recently for modelling dynamical systems like those underlying times series, videos or physical phenomena. Yet, neural networks were observed to not generalize well outside the training distribution, in other words out-of-distribution. This lack of gener...
Neural language models are commonly deployed to perform diverse natural language processing tasks, as they produce contextual vector representations of words and sentences which can be used in any supervised learning setting. In recent years, transformer-based neural architectures have been widely adopted towards this end. After being pre-trained w...
In machine learning, a model is learned from data to solve a task automatically. In the supervised classification setting, the model aims to predict the label associated with an input. The model is learned using a limited number of examples, each consisting of an input and its associated label. However, the model's performance on the examples, comp...
Cette thèse introduit de nouvelles techniques qui exploitent des permutations du vecteur des observations de la variable à expliquer pour améliorer les performances de généralisation dans la tâche de régression et transformer l’estimation de la fonction de densité conditionnelle en un problème de classification binaire. Des justifications théorique...
Pour faire apprendre un nouveau mot à un enfant on lui présente conjointement la forme phonologique de ce mot associée à un référent. A partir de cet unique exemplaire, l’enfant devrait être capable de généraliser le nouveau mot aux autres entités de la même catégorie. L’enfant doit comprendre quelles sont les propriétés de l’exemple qui sont perti...
