François Roucoux

Time: 14h15 - 15h00 (session 1)

Title: Old Methods, New Data: How 1980s AI Meets Modern Medicine to Outperform the State of the Art

Abstract: "Rare diseases affect up to 6 % of the global population, yet each remains individually rare, creating a diagnostic odyssey for millions of patients. Modern AI systems for clinical diagnosis rely on computationally heavy methods like ensemble methods and deep learning, often at the cost of interpretability and trust. This talk revisits a classic approach from the 1980s—Bayesian networks and the noisy-OR model—reimagined with today’s global biomedical databases such as Orphanet and the Human Phenotype Ontology. Using updated inference algorithms, these interpretable probabilistic models achieve state-of-the-art diagnostic accuracy while remaining transparent, computationally sober, and medically grounded. A reflection on how simplicity, knowledge, and ethics can still rival raw computational power."

Jérôme Fink

Time: 15h00 - 15h45 (session 1)

Title: What I learn from putting AI in production

Abstract: "This presentation reflects on the practical lessons learned from developing and deploying an AI-powered dictionary for the deaf community. It explores the challenges related to the integration of AI components into the software development lifecycle, from design to deployment, and how data collected can be leveraged to improve model performance. Through this real-world case study, the talk highlights the challenges of bringing AI research into production."

Arnaud Bougaham

Time: 16h30 - 17h00 (session 2)

Title: Anomaly Detection Applied to Industrial and Medical Images : Robustness, Modularity and Transferability.

Abstract: "The industrial sector evolved with technologies over the decades, culminating in the current era of Industry 4.0. This new paradigm incorporates a range of technologies, including machine learning, and decision systems, with a primary objective of optimizing the factory productivity. Automatic visual inspection is a common approach for anomaly detection to reach this objective. It allows a product image to be judged by an algorithm that asks an operator to confirm any detected anomaly. Traditional algorithms compare images to a golden sample, but they suffer from practical drawbacks, such as a high false positive rate that leads to unnecessary manual checks or human misjudgment. This talk aims to show how deep learning techniques, designed to tackle this problem, are developed at the research side, and deployed at the engineering side, in an automotive electronic production site. High resolution images are controlled, through a Vector Quantized GAN based strategy, so that the reconstructed image shows the normal version of the product. After a collection of metrics on the residual image and other statistics, a neural network (optimized to reach the highest true positive rate, with a customized loss function) classifies the image as normal or abnormal. The method is transferred to medical diagnostic (ovarian cancer segmentation, coma receiptivity analysis, lymphoma evolution monitoring), showing how the insights could be shared across domains."