Adaptive Distributional and Algorithmic Robustness – ADAR

In machine learning, safety concerns arise from various factors, including data variability, algorithm instability, fairness and privacy constraints, and unpredictable environmental changes. Distributionally Robust Optimization (DRO) tackles these challenges by optimizing models across potential data distributions, thereby enhancing both robustness and reliability. This framework has gained increasing attention across several domains of artificial intelligence, such as fair learning, adversarial classification, reinforcement learning, and federated learning. It is regarded as one of the most advanced approaches for ensuring safety in machine learning, a critical requirement emphasized by the European AI Act.

Despite the progress DRO has made in strengthening safety and robustness in machine learning, its practical application still faces significant limitations. One major challenge is the scalability of DRO in large-scale problems, where existing methods can become computationally expensive and inefficient. Moreover, traditional DRO frameworks often lack the flexibility needed to adapt to the dynamic nature of machine learning environments, such as rapidly evolving data distributions, or a variable sensitivity to risk itself. The ADAR project aims to address these limitations by developing adaptive, data-driven methods specifically designed to handle large-scale saddle-point problems in DRO. By incorporating advanced techniques from numerical optimization, risk measure theory, and stochastic optimization, these new methods seek to improve both the efficiency and adaptability of DRO, making it better suited to large, complex machine learning tasks.