QUENCHED VERSUS ANNEALED DILUTION IN NEURAL NETWORKS

The capacity for storing random patterns in a diluted neural network is determined following the method of Gardner (1989). Two types of dilution are considered. In the quenched case, the broken couplings are chosen at random and are independent of the stored patterns. By contrast, in the annealed case, the disconnected couplings are selected in order to optimize the storage of the patterns. By the same token, the vanishing couplings are strongly correlated with the stored patterns. The authors also determine the distribution of the synaptic strengths. This distribution illustrates the difference between quenched and annealed dilution most clearly.