The α‐parameter of autoassociative neural networks—Analysis of equilibrium states and network dynamics
Satoshi Morinaga, Daishi Harada · Systems and Computers in Japan · 1996
Abstract The α‐parameter for autoassociative neural networks is introduced here. This parameter allows the memory capacity of such networks to be derived without use of techniques such as the replica method and signal‐to‐noise analysis. The results of the derivation explain those obtained from simulation experiments, with the derivation relying simply on asymptotic analysis. In particular, after coarse approximation a ratio of 0.13 is obtained as an estimate of network memory capacity. Additionally, by considering the dynamical equations governing the α‐parameter, the dynamics of the network's state are analyzed during recollection. Of particular note is the result showing how the α‐parameter makes it possible to analyze networks consisting of continuous‐time neurons. Finally, the theoretical results are validated by comparing them with empirical data gathered from network simulation experiments.