A LEARNING MACHINE WHICH LEARNS, REMEMBERS AND REPRODUCES SEQUENCES WITHOUT THE USE OF DELAYS

Abstract

In the past, there have been several attempts to produce a machine which is able to learn sequences. The early attempts of Uttley and Taylor involved sequence recognition, and are valuable in that they define the problems involved in this type of approach. More recently Fukushima has described a system which reproduces the sequences that are presented to it, and has shown that this approach overcomes many of the difficulties involved in the early work. The present paper describes a system which uses the reproduction approach, but which does not involve the assumption that sequence reproduction must involve delays. Instead a learning theorem is adopted which used interaction between short‐term and long‐term memories, and which therefore obviates the need for complex systems of delay elements. The sensitivity to sequences is inherent in the learning theorem. Simulation of the system on a digital computer has shown that it can learn to reproduce any combination of the ten inputs in sequences of up to eight different elements.