Enhancing Deep Reinforcement Learning with Symbolic Reasoning
Despite its dramatic successes, contemporary deep reinforcement learning methods have certain shortcomings. Because they rely on the statistics of large datasets, they tend to learn very slowly. We see this, for example, in DeepMind's DQN, which attains superhuman performance after playing a very large number of (certain) Atari games, but takes much longer than a human would to reach beginner-level. Humans, by contrast are able to generalise much more quickly. Here I will discuss ongoing work that aims to supplement deep learning with a symbolic component in order to achieve rapid generalisation at a high level of abstraction.
Murray Shanahan is Professor of Cognitive Robotics in the Dept. of Computing at Imperial College London, where he heads the Neurodynamics Group. Educated at Imperial College and Cambridge University (King’s College), he became a full professor in 2006. His publications span artificial intelligence, robotics, logic, dynamical systems, computational neuroscience, and philosophy of mind. He was scientific advisor to the film Ex Machina, and regularly appears in the media to comment on artificial intelligence and robotics. His book “Embodiment and the Inner Life” was published by Oxford University Press in 2010, and his latest book “The Technological Singularity” was published by MIT Press in August 2015.