Abstract
In this study, we introduce a new neural architecture called N
4 that is based on a collection of local receptive fields realized in the form of referential neural networks. While the network exhibits some similarities to other structures of modular neural networks (such as expert networks), it comes with a number of unique features. Especially, its receptive fields exhibit high flexibility by being formed by neural networks. Subsequently, the processing therein is of referential nature. A ”skeleton” (structure) of the network is completed through unsupervised learning that is aimed at “discovering” and structuring the main dependencies in data. More specifically, the design of the network consists of two phases. First, a blueprint of the network is formed and this involves the prototypes obtained through clustering of training data. This structural development of the network is followed by further refinement in a form of parametric training of the individual neural receptive fields. The study provides a detailed analysis and learning of the network and includes experimental investigations.