OPTIMAL ARCHITECTURES FOR STORAGE OF SPATIALLY CORRELATED DATA IN NEURAL-NETWORK MEMORIES

Using replica approach we investigate storage capacity for ''spatially '' correlated patterns in diluted attractor neural networks. We invest igate analog clipped-sign networks which are generalizations of the st andard networks of the Hopfield type. We consider two kinds of dilutio n: a band, and a random (periodic) ones. ''Spatial'' associations of d ata significantly improve the storage properties of the network. The b and-type dilution affects the critical capacity much weaker than the r andom one, especially when the stored data are strongly correlated.