Reinforcement learning (RL) provides a general suite of algorithms for the approximate solution of many important and interesting optimal control tasks and sequential decision-making problems. At the heart of these problems is a complex tradeoff between exploration of new states and exploitation of known states. Most applications of RL algorithms focus on AI environments with artificial states, operations research tasks, or engineering systems optimization. In this talk, I will present my research on applying RL algorithms to natural environments and natural image sequences. A scalable performance benchmark task will be outlined, and results for solving an optimal control task using an overcomplete sparse code of natural images will be presented. In particular, any neural network or image representation can be tested and compared using the benchmark task, and sparse codes are shown to maximize network memory capacity as well as accelerate neural network training.

—

To request the Zoom link email jteeters@berkeley.edu.  Also indicate if you would like to be added to the mailing list for announcements about Redwood Seminars.