Neural network associative memories with local coding

In this paper the approach is to embed coding techniques into neural networks in order to increase their performance as associative memories in the presence of partial erasures. The motivation comes from the recent works by Gripon and Berrou, which benefit from error correcting codes in neural networks schemes like Hopfield neural networks. The construction introduced stores messages as cliques of small size which can be retrieved in spite of partial error. We improve on the success of retrieving learnt messages by doing a local coding and by mapping codewords of a code on an extension field of G F (2) in clusters. We also use the fractal approach and a slightly different decoding scheme, which is appropriate for partial erasures, in order to reduce the complexity of retrieval. The approach is an improvement over previous methods in terms of computational complexity, and simulations show an improved error performance.