Research

Archimedia is a joint academic-industrial research project aimed at analyzing large multimedia collections in real-world settings. It is between academic researchers and researchers from the Ouest-France company. The project focuses on studying the properties and use of multimedia collections in professional contexts, leveraging in particular the unique archives of Ouest-France, which include millions of articles, photos, pages, videos, and podcasts, covering over 100 years of digitized material.

The academic-industrial partnership in Archimedia is unique due to its long-term collaboration (nearly 30 years), in vivo research focus on real-world professional settings, and emphasis on addressing real-world challenges. It fosters knowledge and technological transfers, develops applied proof-of-concept prototypes, and combines academic expertise in AI with industrial experience in managing large multimedia archives.

This partnership aims to create powerful neural models capable of capturing contextual information, addressing challenges such as annotation, knowledge fusion, and accounting for the temporal dimension of information. These models must be tailored to professional contexts, with a focus on sustainability, trust, and explanation.

Archimedia is exploring two main research directions:

1. Designing Powerful Models (Axis 1): This axis focuses on developing neural models that can handle various modalities (such as images, videos, audio, and text) separately or jointly. The goal is to improve the foundations of these models by addressing their supervision requirements, adaptability to evolving data, accuracy, and ability to create discriminative high-dimensional representations despite the curse of dimensionality. Key challenges include aligning and integrating data from different modalities, ensuring generalization and robustness, and maintaining scalability.
2. Coupling Models and Knowledge (Axis 2): This axis aims to integrate Large Language Models (LLMs) with Knowledge Bases (KBs) to enhance the analysis of large multimedia collections. The research explores how LLMs can improve and complete KBs, how KBs can be integrated into LLMs during pre-training, and how the tight coupling of LLMs and KBs can combine their strengths. This includes addressing multimodal models and KBs, time-aware large models and KBs, and the fusion of knowledge from multiple sources.