Control of associative dynamics by matching features in chaotic neural network

When patterns are stored by using associative learning in a chaotic neural network consisting of chaotic neurons that easily produces chaos, whether dynamic (chaotic) or static (non-chaotic) remembering occurs can be controlled by varying the parameters of the chaotic neurons. A pattern can be searched by using this dynamic remembering as a sampling procedure. If features of the output pattern of the network become similar to desired features during the search, the chaotic state is changed to a static remembering state so that the output pattern has desired features. In this paper, the state is controlled by using a model of presynaptic inhibition similar to that seen in real nerve cells (not using the parameters of the chaotic neurons). Extraction and comparison of the features are performed by back-propagation networks. The extraction of features under this method is amenable to visual representation of its trajectory in feature space.