SCRAM: statistically converging recurrent associative memory

Autoassociative memories are known for their capacity to learn correlated patterns, complete these patterns and, once the learning phase completed, filter noisy inputs. However, no autoassociative memory as of yet was able to learn noisy patterns without preprocessing or special procedure. In this paper, we show that a new unsupervised learning rule enables associative memory models to locally learn online noisy correlated patterns. The learning is carried out by a dual Hebbian rule and the convergence is asymptotic. The asymptotic convergence results in an unequal eigenvalues spectrum, which distinguish SCRAM from optimal linear associative memories (OLAMs). Therefore, SCRAM develops less spurious attractors and has better recall performance under noise degradation.