Nonlinear connections of a feedback neural network and its associative ability analysis

A feedback neural network, of which input elements are connected nonlinearly, is studied. The network´s associative characteristics and ability are discussed, and the learning algorithms are derived. We propose the idea of the equivalent weight subset of the neural network, and by means of this concept, the nonlinearly connected network is made much simpler and the complexity of the weight matrix is lowered. Furthermore, when a pattern is learned by the network, not only the pattern itself but also its shift variance can be recognized by the network.<>