Performance of back propagation networks for associative database retrieval

Backpropagation networks have been successfully used to perform a variety of input-output mapping tasks for recognition, generalization, and classification. In spite of this method's popularity, virtually nothing is known about its saturation/capacity and, in more general terms, about its performance as an associative memory. The authors address these issues using associative database retrieval as an original application domain. Experimental results show that the quality of recall and the network capacity are very significantly affected by the network topology (the number of hidden units), data representation (encoding), and the choice of learning parameters. On the basis of their results and the fact that backpropagation learning is not recursive, the authors conclude that backpropagation networks can be used mainly as read-only associative memories and represent a poor choice for read-and-write associative memories.