PhD Position Hybrid Safe Learning for Inter-Connected Systems

Are you curious how Deep Learning and Online Learning can be effectively combined to create new learning paradigms?

Online learning algorithms achieve robustness often at the expense of performance, as they are very cautious by design. This, in turn, makes them less practical for problems where speed is of utmost priority. On the other hand, offline learning, such as Deep Learning, often suffers from distribution shifts, lack of training data, and poor adaptability to unseen conditions and new problems. Can we combine these two fundamental learning paradigms to synthesize new learning tools that are both fast and adaptive?

This PhD position aims to develop a robust hybrid learning framework that lies at the nexus of online and offline learning. The developed algorithms should be able to benefit from training data, when these are available, and also to learn from real-time, potentially non-IID, streaming data; should be able to track the evolution of key features and achieve model plasticity while avoiding catastrophic forgetting; and should come with interpretable and robust accuracy (generally, performance) guarantees. The designed algorithms will be applied to key problems in the domain of safe learning for interconnected systems (e.g., 6G and Edge AI platforms, self-driving vehicle vision) in collaboration with industry partners and domain experts.

This PhD position is offered in the context of the Marie Curie Doctoral Networks "FINALITY", will be hosted at TU Delft, Department of Computer Science, and will be co-supervised by Prof. George Iosifidis (TU Delft) and Prof. Constantine Dovrolis (University of Cyprus, and Cyprus Institute).

Links & References:

https://www.tudelft.nl/en/eemcs/the-faculty/departments/software-technology/networked-systems

References:

• N. Mhaisen, G. Iosifidis, On the Dynamic Regret of Following the Regularized Leader: Optimism with History Pruning, ICML, 2025.
• G. Iosifidis, N. Mhaisen, D. Leith, Optimistic Learning for Communication Systems, available in Arxiv, 2026.
• C. E. Taylor, S. M. Patil, C. Dovrolis: Before Forgetting, There's Learning: Representation Learning Challenges in Online Unsupervised Continual Learning. Trans. Mach. Learn. Res. 2025 (2025)
• M. B. Gurbuz, X. Zheng, C. Dovrolis: PEAKS: Selecting Key Training Examples Incrementally via Prediction Error Anchored by Kernel Similarity. ICML 2025