Keynote speak: TBD - Hybrid AI for Radio: When Models Must Understand, Not Just Fit

AI is rapidly moving from the cloud into systems that must act in real time and in the physical world—where latency budgets are measured in microseconds, data rates are extreme, and the environment shifts faster than we can retrain. Future mobile networks (6G and beyond) sit at the center of this transition: they are among the most demanding real-time ML deployments ever attempted.

This talk argues that meeting these requirements requires a shift from purely data-driven modeling toward physics- and structure-informed learning, using hybrid architectures that explicitly embed the operations, constraints, and structures that govern physical reality. This inherent structure dramatically reduces the search space during training, cuts compute requirements, lowers inference latency, and improves robustness to unseen conditions, while preserving a level of explainability often lost in “black-box” approaches. Using examples from wireless signal processing and control loops, we highlight the critical research challenges in AI algorithm design and the practical complexities of deploying such models in the physical world.