Sparse Binary Matrices as Efficient Associative Memories
Résumé
Associative memories are devices used in many applications that can be considered as universal error correcting decoders. Originally introduced in neuroscience, adding the use of error correcting codes principles proved to greatly enhance the performance of these devices. In this paper we reintroduce a neural-based model recently proposed using the formalism of linear algebra and extend its functionnalities originally limited to erasure retrieval in order to handle approximate inputs. To perform retrieval, we use an iterative processing that we prove to always converge. We then analyze performance under the asumption of connection independence and discuss simulations.