Learning from Dynamics - Linear Dynamical Systems and Machine Learning Applications

Watch a 44-minute conference talk from the Harvard CMSA Big Data Conference where MIT professor Ankur Moitra explores the learning dynamics of linear dynamical systems and their applications in time series data analysis. Delve into the challenges and gaps in understanding these systems, particularly in scenarios involving long-range correlations. Learn about a novel algorithm based on the method of moments that operates under minimal assumptions while maintaining computational efficiency. Discover how theoretical machine learning tools, including tensor methods, can be applied to non-stationary settings. The presentation covers various applications, including medical implementations, while exploring technical concepts such as strict stability, observability matrix, Markov parameters, and special case scenarios. Gain insights into the connections between linear dynamical systems and recurrent neural networks, and understand how these concepts contribute to the broader field of time series analysis.