IMPROVING PATTERN RECONSTRUCTION IN NEURAL NETWORKS BY ACTIVITY DYNAMICS

I study the averaged dynamical behaviour of neural networks over an ex tended monitoring period, and consider pattern reconstruction procedur es by activity clipping, selectively freezing, or sequentially freezin g the dynamic nodes. They enable the retrieval precision to be improve d, the basin of attraction to be widened, or the storage capacity to b e increased, even when the information is not efficiently embedded in the synaptic weights.