On November 25 and 26, I had the opportunity to participate in the ANITI Days, organized at the Cité de Montaudran in Toulouse. These days dedicated to artificial intelligence are orchestrated by the ANITI (Artificial and Natural Intelligence Toulouse Institute), a leading research institute in Occitanie specializing in artificial intelligence.
Enriching Discussions on Retrieval-Augmented Generation (RAG)
Among the many exhibitors, I particularly enjoyed discussing advancements in Retrieval-Augmented Generation (RAG), an approach combining information retrieval and content generation.
- With researchers from LIEBHERR, I discovered a promising library aimed at optimizing RAG based on the types of documents to be processed. A solution that opens up fascinating prospects for improving the adaptability and efficiency of this technology.
- For their part, the Synalinks teams presented their work on hybrid RAG agents combining symbolic AI and large language models (LLM). These agents, capable of self-improvement, mark a step towards increasingly autonomous and high-performing AI systems.
Key Conferences and Scientific Discoveries
More than a dozen conferences punctuated these two days, highlighting major scientific advancements in various fields of AI. Here are some highlights:
- Ellie Pavlick (via video conference) explored the understanding of neural networks, particularly through heat maps to visualize the influence of neurons on responses. She introduced the concepts of extractive functions and abstract functions, illustrating the mechanisms by which models generate their responses.
- Julie Hunter from Linagora unveiled an ambitious project: a completely open-source language model named Lucie. Unlike other models like LLaMA, Lucie promises total transparency, providing both the weights, training scripts, and data used. A decisive step towards ethical and reproducible AI, essential for ensuring the security and reliability of systems.
- Chloé Braud presented preliminary research on training models with limited data, a promising approach to reduce computational resource costs while maintaining high performance.
- Eliot Chane focused on robotic control, using environment simulations and animal documentary videos to train agents in an unsupervised learning framework.
- Rufin VanRullen introduced the concept of Global Latent Workspace, a multimodal neural network integrating language, vision, and movement. This innovative approach shows impressive results for robot control.
On the second day, equally captivating topics were addressed:
- Claire Monteleoni explored AI applications in the context of climate challenges.
- Vincent Martin presented innovative projects using CNES data, with applications ranging from meteorology to bathymetry.
- Several conferences focused on the challenges of trust and guarantees in neural networks, with notable contributions from Matthieu Serrurier, Mélanie Ducoffe, and Paul Novello, on strategies to certify, validate, and regulate model results.
The day concluded with a presentation by Urtzi Ayesta, demonstrating that reinforcement learning could model a system with an indeterminate number of states, a significant advance in the study of complex systems.
An Inspiring Event
These two days were an immersive experience rich in discoveries, meetings, and exchanges, revealing the challenges and opportunities shaping the future of artificial intelligence. Whether for enthusiasts, researchers, or professionals, the ANITI Days illustrate how central AI is to ongoing technological and societal transformations.
I look forward to seeing how these advancements will materialize in the coming years and am convinced that Occitanie will continue to play a key role in this booming ecosystem.
Author: Sébastien Da Rocha (Innovation Department, Neogeo)
