A sequential adder using recurrent networks

D.E. Rumelhart et al.'s proposal (1986) of how symbolic processing is achieved in PDP (parallel distributed processing) networks is tested by training two types of recurrent networks to learn to add two numbers of arbitrary lengths. A method of combining old and new training sets is developed which enables the network to learn and generalize with very large training sets. Through this model of addition, these networks demonstrated capability to do simple conditional branching, while loops, and sequences, mechanisms essential for a universal computer. Differences between the two types of recurrent networks are discussed, as well as implications for human learning.<>