Rough neural networks for complex concepts

Abstract

Rough neural networks aim at hierarchical construction of compound concepts. Although the structure of such concepts is assumed to be more complicated than numbers in case of standard feedforward neural networks, some mechanisms can be generalized to achieve efficient propagation and learning. One of possible generalizations, called the normalizing neural networks, enables to propagate vectors instead of single signals. Neurons take form of multidimensional functions, which model cross-dependencies among importance of particular vector components. In this way, we are able to represent some types of compound concepts using relatively simple neural network structure. As an illustration, we consider the case study related to the task of magnetic resonance images' segmentation. We put a special emphasis on how the nature of objects and attributes in a given decision system influences the network's architecture. We also compare our model to other rough-neural approaches. © Springer-Verlag Berlin Heidelberg 2007.