Extracting information in a graded manner from a neural-network system with continuous attractors

Memory retrieval from neural networks has been described by dynamical systems with discrete attractors. However, recent neurophysiological studies suggest that information extraction in the brain is more likely to be described with continuous attractors. Here we put forward a neural-network system that provides continuous attractors with respect to the network state represented by a vector quantity. An attractor pattern continuously depends upon an initial pattern; it also reflects the embedded pattern. These suggest that, for each query encoded by an initial state, our model can extract different information from the network. To demonstrate the usefulness of this information, our model is applied to keyword extraction from a document.