Comparative Research On Performance Of Associative Memory Artificial Neural Networks

Associative memory is one of the essential functions of human’s brain. The association refers that people can map the input pattern to the output pattern according to the similarity between them. The memory is a fundamental process, which enables the brain to memorize those completed prototypes. According to associative memory, people can recall a whole prototype from part of its information or recall another prototype. In the area of artificial intelligence, associative memory has been a hot topic for a long time and artificial neural networks are importantways to realize it. For the last decades, researchers have proposed various architectures and learning algorithms for those artificial neural networks.The most widely used architecture of those associative memory networks is proposed by professor Hopfield at 1982, which is a mono-layer full connection feedback network. It is different from the previous forward networks and static associative memory models. Based on this architecture, since researchers have improved the neuron models and learning algorithms all the time, the capacity of the network becomes larger and the recall error rate gets lower.Besides the outer product rule which was applied by Hopfield, there are some other typical algorithms such as pseudo-inverse rule, LSSM algorithm and NDRAM algorithm.Firstly, those typical algorithms were discussed theoretically. Then the performance of those algorithms were evaluated by the average results of some repeatable experiments. Finally, it proves that although LSSM cannot auto-learn the weights, it performs best in capacity and error rate, on the other hand although NDRAM can auto-learn the weights, it performs better. Meanwhile outer product rule and pseudo-inverse rule proves worse in capacity and recall error rate.A unified experimental conditions without complex mathematical derivation were provided for comparing the four algorithms, so that the evaluation results are more scientific and reliable, on the other hand it provides a valuable reference for choosing the associative memory algorithms in different applications.Besides the second generation of artificial neural network, associative memory function based on the third generation neural network is tried as well.