Neural episodic management
Authors: Alex Pritzel, Benigno Uria, Sriram Srinivasan, Adria Puigdomenech, Oriol Vinyals, Demis Hassabis, Daan Wierstra, Charles Blundell
Deep-gain learning algorithms have achieved top-level performance in many tasks, but are generally very inefficient. In this work, we propose a new algorithm that allows the rapid incorporation of new information collected by the agent. To this end, we introduce a new identifiable data structure, a differentiated neural dictionary that can incorporate new information immediately, while we can update its internal representation based on the task that the algorithm is intended to solve. Our agent Neural Episodic Control is built on a differentiated data structure and is able to learn much faster in many environments.
For more information and related work, see paper.
See it in ICML:
Wednesday 9 Aug 16: 06-16: 24 @ C4.5
Wednesday, August 9 from 6:30 pm to 10:00 pm @ Gallery # 125