Pattern retrieval in threshold-linear associative nets.

Networks of threshold-linear neurons have previously been introduced and analysed as distributed associative memory systems. Here, results from simulations of pattern retrieval in a large-scale, sparsely connected network are presented. The storage capacity lies near a = 0.8 and 1.2 for binary and ternary patterns respectively, in reasonable accordance with theoretical estimates. The system is capable of retrieving states strongly correlated with one of the stored patterns even when the initial state is a highly degraded version of one of these patterns. This pattern completion ability holds for an extensive number of memory patterns, up to α ≈ α/2, thereby increasing the credibility of the model as an effective associative memory.