Pattern selectivity in neural networks as a means of understanding basin structures
Albrecht Rau, K. Y. Michael Wong, David C. Sherrington · Journal of Physics A Mathematical and General · 1993
The authors study the problem of learning and retrieving for a pair of correlated patterns within an extensive number of uncorrelated patterns, for networks where learning may be treated as an optimization process with respect to an arbitrary cost function. This formulation is then applied to several specific examples, where they study the pattern selectivity of these systems (i.e. their ability to differentiate correlated patterns) and investigate the process of basin shrinking with increasing loading levels. They define and discuss several different retrieval phases, whose existence depends on the competitive interplay of the loading level and the pattern correlation. Discussion of asymptotic retrieval is restricted to dilute asymmetric networks.