Artificial intelligence
technologies
CEA-List’s artificial intelligence research focuses on the development of advanced AI methods spanning generative AI, natural language processing, and computer vision, all in pursuit of robustness, efficiency, and trustworthiness in real-world conditions.
In the field of generative AI, CEA-List developed a beyondstate-of-the-art retrieval-augmented generation (RAG) model for visual question answering about named entities. In computer vision, we are addressing object discovery and segmentation without human annotation through a cross-distillation method that leverages the complementarity of 2D and 3D data, as well as a method for the semantic analysis of 3D scenes from natural language queries for robotics and augmented reality.
In distributed and Edge AI, we are investigating approaches based on asynchronous event graphs that utilize event driven cameras for very low-latency motion detection. In parallel, near-sensor analysis is driving progress toward smart acquisition systems for structural health monitoring. We are tackling challenges in collaborative and adaptive learning through distributed federated methods for multi-source domain adaptation and through algorithms that limit catastrophic forgetting and concept drift. We are applying these advances to the predictive and private management of electric charging stations, for example.
Finally, the pressing issues of trustworthy and safe AI are top-of-mind in our research on LLM bias analysis, formal verification of neural networks, and uncertainty estimation in AI-assisted Monte Carlo simulations.
