Capacity of an associative memory model on random graph architectures (Q2515521)

This paper studies the storage of patterns in random graph models. The storage is defined as the ability to to retrieve corrupted information in neural networks. Based on the spectral theory and combinatorical probability, the authors present rigorous bounds on the storage capacity of the Hopfield model on random graph models including some power-law graphs. Let \(\lambda\) be the largest eigenvalue of the adjacency matrix of the network, \(N\) its number of nodes, and \(m\) its maximum degree. The number of patterns one can store in the sense that a number of errors that is proportional to \(N\) can be repaired by \(O(\ln N)\) steps of the retrieval dynamics is shown to be of order \(\lambda^2/(m\ln N)\).