| Potts networks, in certain conditions, hop spontaneously from one discrete attractor state to another, a process we have called latching dynamics. In order to study the property of a single latching transition we stored in the network a "simplified" group of patterns and we proved that transitions cluster in a number of distinct classes: effectively random transitions between weakly correlated attractors; structured, history-dependent transitions between attractors with intermediate correlations; and oscillations between pairs of closely overlapping attractors. In the general condition of all correlated patterns the system spontaneously generates a sequence of latching transitions, with the oscillating retrieval of different stored memories. We saw that the temporal length of this behaviour is strictly related with the correlation property present between the stored patterns. |
|