Memory Networks With Asymmetric Bonds.

We study the way in which the properties of associative memory networks are changed when the interactions between the units are not symmetrical. For a class of analog networks subject to thermal noise (Langevin models), the results indicate that the memory states are not seriously degraded; their critical temperature is simply lowered from its value in the corresponding symmetric model. On the other hand, spin glass states, which occur in the symmetric case when the number of memories becomes a finite fraction of the number of units in the system, are rendered unstable by the introduction of asymmetry.