The Temporary Degradation of Stored Information in Nonequilibrium Cross-coupled Network and Its Application to Separation of Overlapped Patterns

<p>A cross-coupled network, which a Hopfield network belongs, can work as an associative memory and is attractively capable of storage and recovery of information. A Hopfield network can associate a stored patten similar to the given one, even though it is defective, noisy or incomplete. However, it can associate only one pattern corresponding to the given pattern, because the state of the network goes down to the locally or globally lowest level of energy. Thus the Hopfield network cannot associate more than one candidates for one given pattern, while a human brain candynamically associate some patterns similar to a presented one. An artificial neural network should be modeling of a biological neural system. The association of more than one pattern in a biological brain cannot be modeled by the Hopfield network. To imitate the dynamicalassociation in a human brain, the activation of the network state by noise or the chaotic behavior of neural units might be employed, such as a chaotic unit. In these systems, however, the state of the network is always activated and the network associates unstored patterns frequently. Because the state of the system can stay at any level of energy at the same probability in this case. So, a lot of researchers have tried to activate the network for dynamical association and proposed their own methods in recent years. Aihara, et al.improved the Nagumo's neuron model by replacing a step function with a sigmoidal function. The neuron model can exhibit chaotic behavior by itself, so that it is referred to as a "chaotic neuron". The chaotic neurons are cross-connected to each other to construct a network, in which individual chaotic behavior of each neuron activates to search another state after falling into a local minimum. Thus the dynamical association is realized and thus the networkis referred to as a chaotic network. In this network, the informationon the stored patterns is embedded in weight distribution and we have no other way to identify the stored patterns than the feature of the output sequence of the network. However, an overlapped pattern cannot be separated into its individual patterns, whereas a human being cando so at a glance. To cope with these problems, a novel non-equilibrium network which separates an overlapped pattern to individual ones is proposed in this paper. This network has following two properties. One is the ability of frequent association of stored patterns, that is, the network associates stored patterns more than non-stored ones. The other is a temporary degradation of the stored item. This degradation plays a role of a system reset for a new association. The implementation of the first property is based on a measure of similarity which is latent in cross-coupled networks, and realized by using an inverse N-shaped function as a mapping one. Besides, the second property is realized by changing connection weights between units. An algorithm separating overwritten patterns is based on the effects of these two properties.</p>