Abstract: Learning from data derived from systems with interactions between components (social networks, shared resources) and temporal dynamics (RL, time series analysis) is of fundamental interest. This necessitates specialized algorithms which differ from algorithms designed for i.i.d. data. In this talk, we briefly introduce several settings of interest and consider non-linear system identification and reinforcement learning in detail. We demonstrate that naively adapting optimization algorithms designed for i.i.d. data can suffer from sub-optimal convergence. We then design algorithms which unravel the dependency structure efficiently in order to obtain near-optimal performance.
Bio: Dheeraj Nagaraj is a Research Scientist at Google Research. He completed his PhD in EECS from MIT advised by Guy Bresler. His research focuses on theoretical machine learning and probability with a focus on stochastic optimization algorithms, learning from dynamic environments and random graphs.