Chairs in Machine learning and reasoning

Automatic decision systems are currently deployed at large scale. They are already affecting the life of citizens, and their impact is expected to grow. Often based on complex data-driven machine learning models, these systems raise many scientific challenges regarding safety, robustness, privacy, fairness, and data efficiency when massively annotated data are not available. The Grenoble ecosystem possesses many assets to solve these challenges by combining the perspectives of various scientific fields, both on the academic and industrial sides. There is indeed a long tradition of research in optimisation, statistics, symbolic AI in Grenoble, as well as a more recent one in machine learning. Our strategy in MIAI consists in fostering interactions between these different disciplines, in order to make fundamental contributions to machine learning and reasoning through the research programmes below, and connect them to applications, which will be detailed in other axes.

1.1. Machine learning models

Despite their success, huge-dimensional machine learning models such as multilayer neural networks are lacking crucial stability properties, making them dependent on huge sets of annotated data. We are planning to tackle this problem through multiple complementary angles. We will develop theoretically-grounded regularisation principles for high- imensional models to make them more data-efficient, introduce new learning paradigms that will allow us to leverage multiple learning tasks sequentially, propose new transfer learning approaches, and explore other paradigms inspired by cognitive sciences.

1.2. Statistics and optimization

Developing robust machine learning models at large scale requires a joint research effort in both statistics and optimisation. In MIAI, we are planning to make fundamental contributions in both fields; we will develop unified statistical frameworks to better understand and improve the design of large-dimensional models by harnessing tools from convex optimisation, statistical physics, and random matrix theory. On the algorithmic side, we will also focus on distributed optimisation techniques, and on the development of generative models and robust strategies to adversarial problems that occur frequently in machine learning formulations.

1.3. Fair and evolvable AI

Modern data-driven AI systems are now able to automatically take decisions with high societal impact, but they raise important questions about privacy of personal data, fairness, and discrimination based on data biases. Symbolic and logic-based AI provide solid frameworks to address such challenges. By adopting a pluridisciplinary approach, we plan to develop AI systems that explain their decisions, certify fairness and privacy, and understand the mechanisms driving the evolution of knowledge in environments where humans and AI systems co-exist.