ABSTRACT
An analysis of long-term memory processes indicates a need for short-term mechanisms that store arbitrary spatial patterns of neural activity for substantial lengths of time. This paper analyzes the short-term pattern storage properties of generalized on-center, off-surround neural networks. Except for special cases, such networks cannot store patterns for arbitrarily long lengths of time (to equilibrium). However, bounds of pattern degradation show that networks close (in a precise sense) to these special cases may store arbitrary spatial patterns of activity for substantial lengths of time with good accuracy.
