Feature Space Mapping Network for Classification

The development of the Feature Space Mapping (FSM) system has been motivated by our conviction that neurodynamical models of the brain are very difficult, it is not clear how to model larger groups of neurons in a realistic way and simple neural networks are not well suited for cognitive modeling. Mind arises from a complex dynamics of the modular and hierarchical brain structures. Approximations to this dynamics lead to a set of concepts [1] helpful in description of the mind, such as the concepts of mind events taking place in feature spaces, called at the highest level of hierarchy “conceptual spaces” or “mind spaces”. Among many other aspects mind models should be capable of recognition, classification and reasoning. Cognitive modeling may not always be faithful to neurobiology, but tries to preserve essential functions simplifying the underlying structures and dynamics. Neurofuzzy networks are natural implementations of such models. In this paper we present one particular realization of general cognitive modeling ideas [1] and apply the resulting neurofuzzy network to classification problems. We will assume that the incoming signals I(t) are subject to preprocessing (accomplished in the brain by various topographic maps and population coding mechanisms) that defines features of internal representations Xi(t) suitable for classification. A coordinate system based on these features {Xi} defines a multidimensional feature space. Since these features may be of different types (for example, coming from different sensory modalities) partial classifications are performed in local feature spaces while the final classification is performed in the space that is higher in the hierarchy, called here “the mind space”. The system learns by creating and modifying mind (or memory) objects in this space. They are described using mind (or memory) function M(X) as the fuzzy areas in the mind space where the function has non-zero values. Local maxima of the mind function are prototypical representations of the (fuzzy) training data.