I recently participated in the Interactive Learning and Interventional Representations workshop at the Oberwolfach Research Institute for Mathematics. The workshop spotlighted cutting-edge research in interactive models of learning. This event, hosted in the scenic heart of southern Germany, was a convergence point for international scholars dedicated to redefining the paradigms of online learning, reinforcement learning, causal inference, and interacting learning systems. Funded by the state of Baden-Württemberg and in collaboration with ELLIS Institute Tübingen and the Max-Planck-Institut für Intelligente Systeme, the workshop embodied the program's mission to bridge observational and interventional learning through causal modeling.
During the three-day workshop, every participant presented his/her recent works and new research directions, followed by stimulating discussions on different research areas. For example, I presented my recent research on imitation learning in multi-agent systems. The other discussed topics included Bayesian fixed-budget best-arm identification in structured bandits, showcasing algorithms for robust performance in diverse models; the introduction of PeFLL, a personalized federated learning algorithm, demonstrated advancements in model accuracy, reduced computation, and theoretical guarantees for future client adaptation; explorations into multi-objective machine learning underscored the importance of considering multiple objectives without premature trade-offs. Additionally, presentations on moving beyond the I.I.D. assumption and bilevel optimization addressed challenges in causal structure identification and hyperparameter optimization. Innovative strategies in zero-order optimization were outlined, emphasizing the role of gradient estimators in online convex optimization. These were only some of the presented research.
After dinner, we discussed the future research directions to explore and the main current challenges in interactive learning, online learning, causality, and interventional representation.
Participating in this workshop gave me the possibility not only to discuss the importance of robust intelligent behavior for real-world applications but also to set the stage for future collaborations, pushing the boundaries of what's possible in learning-based decision systems.
