Fault tolerant hashing and information retrieval using back propagation

The architecture and performance of neural networks designed and trained to compute hashing functions is described. The networks described are of the connectionist type and are capable of learning complex mappings using the back-propagation or error algorithm. Connectionist networks are robust, are capable of limited error correction, and offer several advantages over traditional hashing methods. Multiple indexing, which implements many-to-one mapping, can be easily realized by training a network for each key attribute. The neural network approach can be used to train a very large number of pattern associations by dividing a problem into smaller problems. This neural network consists of several subnetworks, each solving a specific mapping task. The experimental results show that small neural networks with simple processing elements can learn complex mapping that implement index search in constant time