Deep learning is increasingly used to model complex systems characterized by high-dimensional observations, structured spatial dependencies, nonlinear temporal dynamics, and uncertainty. However, strong predictive performance alone does not explain why a model works, when it will generalize, or whether its predictions can be trusted.
This workshop brings together theoretical and applied perspectives on reliable learning for complex and spatio-temporal systems. It explores how insights from approximation, learnability, optimization, generalization, calibration, and uncertainty quantification can guide the design and selection of modern learning methods—including graph neural networks, Transformers, Gaussian processes, neural operators, and other structure-aware models.
Particular attention is given to the relationship between data characteristics, model properties, and predictive behaviour: spatio-temporal predictability, model capacity, calibration, robustness, and multi-horizon forecasting. Applications span transportation and mobility, maritime and aviation systems, finance, and other complex dynamical domains.
By connecting foundational analysis with real-world modeling challenges, TLCS aims to promote a deeper understanding of how theoretical principles can support more reliable, interpretable, and effective AI systems.
Workshop Organizers
- Davide Rigoni
- Luca Bergamin
- Alessandro Betti
- Samuel Cognolato
- Giovanni Donghi