Description & Basic Information

From time-series forecasting to Large Language Models, modern AI is increasingly defined by its ability to model sequences with long-range temporal dependencies. While architectures such as Transformers and State Space Models have achieved remarkable empirical success, they often lack the theoretical guarantees and computational simplicity that characterize classical control and dynamical systems theory.

This tutorial bridges that gap by introducing spectral filtering and universal sequence preconditioning, two techniques for efficient and provably grounded sequence learning. We show how these methods leverage the spectral structure of dynamical systems to bypass explicit latent-state identification, enabling improper learning with complexity independent of the hidden state dimension. The resulting algorithms avoid many manifestations of the curse of dimensionality while remaining simple, scalable, and practical.

Participants will gain a rigorous understanding of the principles underlying these methods, along with a toolkit of techniques that connect modern sequence modeling with classical ideas from control, optimization, and dynamical systems.