Novel Knowledge Representation (Area Representation) and the Implementation by Neural Network

Abstract

This chapter presents a new paradigm to represent knowledge and the implementation by neural network. Knowledge representation is fundamental and important matter to construct intelligent systems. Local representation and distributed representation are widely known, however, they have their own shortcomings. For example, the local representation is not robust and is not fit to biological findings. The distributed representation is difficult to treat inference in spite of having similarity to brain. Since the newly proposed area representation (AR) method is, so to speak, an intermediate method between them, it can preserve advantages of each method. The AR method can express hierarchical knowledge by employing a new inclusion relation in which an upper level concept includes the lower level concepts. The neural network implementing the concept of the AR consists of Kohonen feature maps and it employs a new learning algorithm named neighborhood Hebbian learning. Each map is connected and forms multidirectional associative memory.