Fast recognition of real objects by an optimized hetero-associative neural network

We have developed and realized a concept which is very well suited for a quick recognition of highly correlated patterns. For a hetero-associative memory we used a minimal optimized output code (index memory). We constructed a tree structure in which the assignment of indices has been optimized by simulated annealing. Thus the algorithm for optimal stability of the learned patterns works most effectively. Special care was taken of recognizing « real » objects, e.g. scanned letters. Here the characteristic noise is very anisotropic. We have slightly modified the minimal overlap strategy of Krauth and Mezard [1] by training with this specific noise, and could improve the performance of our network. In order to get insight into the network and its behaviour we used a measure called constructivity which shows clearly the anisotropic effects. We trained a network to recognize a scanned text and to produce the associated text file. Due to the architecture of the network many processes can be treated in parallel. Therefore we used transputers for the implementation.