A Pong playing agent modelled with massively overlapping cell assemblies

Cell assemblies (CAs) are central to many higher order cognitive processes such as perception, recognition and recollection. These processes stem from the fundamental cognitive tasks of memorisation and association, which CA models are able to perform with a viable degree of biological realism. This paper describes a virtual agent that uses CAs that emerge from fatiguing leaky integrate and fire neurons via learning from dynamic interaction. Learning is continuous and the topology is biologically motivated. The agent is able to visually perceive, learn and play a simplified game of Pong. It can learn from a user playing the game, or playing on its own. The agent's memories are encoded in the form of overlapping CAs that enable it to generalise its associations to account for previously unseen game moves. The trained agent hits the Pong ball correctly over 90% of the time. This work furthers the understanding of associative memory and CAs implemented in neural systems.