Constraining the Connectivity of Sparse Neural Associative Memories

International audienceWillshaw networks are a type of associative memories with a storing mechanism characterized by a strong redundancy. Namely, all the subparts of a message get connected to one another. We introduce an additional specificity, by imposing the constraint of a minimal space separating every two elements of a message. This approach results from biological observations, knowing that in some brain regions, a neuron receiving a stronger stimulation can inhibit its neighbors within a given radius. Theoretical arguments are derived to quantify the benefits of this method in terms of memory usage as well as pattern completion ability. We experiment with different values of the inhibition radius introduced, and we study its impact on the error rate in the retrieval of stored messages. We show that this added constraint can result in significatively better performance of the Willshaw network, either when reducing its set of connections, or when extending its set of neurons while maintaining the memory resource