Learning from Dynamics in Linear Dynamical Systems

Explore a 48-minute lecture on "Learning from Dynamics" presented by Ankur Moitra from the Massachusetts Institute of Technology at IPAM's EnCORE Workshop on Computational vs Statistical Gaps in Learning and Optimization. Delve into the world of linear dynamical systems, their applications in time series data, and their connections to recurrent neural networks. Discover a new algorithm based on the method of moments that addresses gaps in existing approaches, offering computational efficiency under minimal assumptions. Gain insights into how tools from theoretical machine learning, including tensor methods, can be applied to non-stationary settings. Recorded on February 26, 2024, at the Institute for Pure & Applied Mathematics (IPAM) at UCLA, this talk bridges the gap between computational and statistical aspects of learning and optimization in dynamic systems.