A Neural Architecture for Fast Learning of Stimulus-Response Associations

Humans can learn in a matter of seconds to associate a response R with a stimulus S, following simple verbal instructions, e.g. “Press the right button when you see a green light”. This involves establishing links through several neural relays, from visual to motor representations. There is no model of fast learning in such systems. This paper describes a new architecture and learning principle that enable fast associative learning. Learning proceeds through an “activation pull” from the response area and a “representational push” from the stimulus area. This novel pull-push principle allows separated pattern recognition streams to develop in the network upon single presentations of S-R pairs. A proof-of-concept implementation is presented in the paper in which arbitrary n x n binary stimuli are mapped to arbitrary n x n responses (n =10 and 15). The results and the capacity of the network are discussed.