Deep Reinforcement Learning has seen many recent celebrated successes. In this talk, we shall examine the role of an agent's past on how the agent performs. We will explore different ways in which agents can be augmented with memory to improve their performance. In particular, we show what kind of tasks require memory, how rich allocentric memories can be created from egocentric experience, and how memory can significantly advance the state of the art in exploration.
Charles Blundell is a senior staff research scientist at DeepMind, leading a team of fellow researchers working on deep learning, probabilistic modeling, neuroscience, and reinforcementÂ learning. He holds a Ph.D. in machine learning at the Gatsby Unit, University College London.
The AI Seminar is generously sponsored by Fortive.
Zoom Participation. See announcement.