AI and Deep Learning

The theme explores the ability of deep learning to automatically extract complex representations from heterogeneous data (images, text, signals, biological data). These representations, often in the form of embeddings, make it possible to capture abstract structures useful for analysis and prediction.

Over and above performance, the theme aims to develop models that are both understandable and reliable. Interpretability is essential to ensure user confidence, particularly in sensitive areas such as health and the environment.
This work also focuses on the robustness of models in the face of noisy, incomplete or biased data.

The theme addresses both discriminative approaches (classification, detection) and generative approaches (language models, image or data generation). The rise of large language models notably opens up new perspectives for natural language processing, including in low-resource contexts.

An important focus concerns the development of more resource-efficient AI models capable of operating on constrained devices (sensors, embedded systems). This objective is crucial to ensure the accessibility and sustainability of the developed solutions.