A decentralized tunable short term neural network memory and application to tracking

A neural network memory is described that can be used as a short-term memory, i.e. it incorporates a forgetting mechanism that ensures that recently stored words are remembered even after a large number of words have been stored. Its operation is decentralized in that the update of the synaptic matrix rows can be carried out independently and in parallel by individual neurons. It incorporates a parameter which allows it to be tuned to give optimal error performance at a specified lag. The performance of the memory is reported and compared against another short-term memory. A simple application to the tracking of a slowly mutating object which is undergoing random translation is demonstrated.<>