Raja Chatila (Sorbonne Université) animera le séminaire du 29 septembre.

Deploying unproven systems in critical applications (and even in not critical ones) is dangerous and irresponsible, and therefore unethical and shouldn’t be acceptable. As AI systems based on Machine Learning which statistically process data to take decisions and to predict outcomes, have become of widespread use in almost all sectors, from Healthcare to Warfare, the need to ensure they "do the right thing" and provide reliable results has become of primary importance. Hence a full research stream was started, trying to address the blackbox paradigm limitations characterizing such systems. With millions of parameters computed from data using optimization processes, the practice of using various off-the-shelf components to build new systems without solid verification and validation processes, and the absence of causal links between inputs and outputs, what does it mean, concretely, to make AI systems robust, safe and explainable? Is this a reachable objective at all? And will this lead to trustable Intelligent and Autonomous Systems?