[PRESS RELEASE] Preligens to develop a foundation model for geospatial data

PRELIGENS TO DEVELOP A SPECIFIC FOUNDATION MODEL FOR GEOSPATIAL DATA, MOBILIZING RESOURCES FROM THE JEAN ZAY SUPERCOMPUTER AT GENCI

Paris, November 29, 2023 – Preligens, a global leader in artificial intelligence for geospatial data in aerospace, defense, and government institutions, is proud to announce that its project to develop a foundation model specific to geospatial data has been granted 300,000 hours of computation on the Jean Zay supercomputer at GENCI (Grand Équipement National de Calcul Intensif). This machine, known for its capabilities in artificial intelligence, boasts over 3000 GPUs and a support team of 12 people, hosted and operated by IDRIS (CNRS).

Preligens' project aims to develop a generic vision model dedicated to geospatial data, using commercially accessible data. Its implementation in a rapidly evolving field with limited experimentation would significantly enhance the efficiency of new image analysis solutions by substantially reducing the need for annotated data volumes, accelerating the completion of the final task (detection, segmentation), and simplifying adaptation to client use cases.

Utilizing GENCI's resources, a major French research infrastructure whose mission is to implement the national strategy for intensive computing, artificial intelligence, and quantum computing by providing HPC and AI computing resources for open research, is a decisive step and a significant recognition of the project's strong potential in terms of sovereignty and scientific and industrial competitiveness in French and European artificial intelligence.

The technical means required to successfully train such a model demand substantial resources, including a training database containing tens of millions of very high-resolution images and a model to be trained composed of at least one billion parameters.

Preligens, already equipped with the necessary volume of images for this project, requires an infrastructure such as the Jean Zay supercomputer to implement it, extending traditional high-performance computing usage to new applications for artificial intelligence.

"By combining the computing resources of the Jean Zay supercomputer and the commercial data available to Preligens, we will be able to explore data preparation and training strategies specific to remote sensing, with the aim of implementing a revolutionary generic vision model dedicated to very high-resolution spatial remote sensing," says Renaud Allioux, co-founder and Chief Innovation Officer of Preligens. "We welcome this very concrete illustration of the effectiveness of implementing the national strategy for artificial intelligence, allowing French companies to benefit from the country's most advanced technologies and access tools that enable them to compete globally in technological innovation."

Photo : copyright Photothèque CNRS/Cyril Frésillon