Learning from Probabilistic Models

This lecture explores the challenges of learning from data generated by probabilistic models, covering topics such as computational complexity theory, probabilistic data models, and information theory. The instructor discusses the use of heuristics from statistical physics to compute optimal errors in data reconstruction and the application of Approximate Message Passing algorithms. The lecture delves into the computational-to-statistical gaps in learning, showcasing the limitations of various algorithms in the hard regime. The content also includes experiments with correlated signals, compressed sensing, and phase retrieval, highlighting the difficulties in efficient learning from generative models.