Feature Space Mapping: a neurofuzzy network for system identification
Norbert Jankowski, Antoine Naud · 1995
Our original motivation for the development of the Feature Space Mapping (FSM) system came from cognitive modelling. Mind arises from complex dynamics of the brain. Approximations to this dynamics lead to a set of concepts [1] helpful in description of the mind, such as the “inner space” concept, called also the “conceptual space” or the “mind space”. Among many other aspects mind models should be capable of recognition, classification and reasoning. Cognitive modelling is not always faithful to neurobiology, but a natural implementation of such models has a neural network form. In this paper we present one particular realisation of general ideas related to cognitive modelling [1] leading to a neurofuzzy network useful for system identification. The low level cognitive processes accomplished by various topographic maps in the brain define features of internal representations X i (t) for the incoming signals I(t). These features may be of many types: shapes derived from the analogue sensory signals, numbers, linguistic variables. A coordinate system based on these features {X i } defines a multidimensional feature space, called here “the mind space” since axes of some of the coordinate systems we use represent confidence factors and partial classification results, not only simple features. The system learns by creating and modifying “mind objects” in this space. They are described using “mind function” M(X) as a 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. Feature detectors react to specific localised features of the incoming data, therefore objects in the mind space are modelled via localised functions rather then unbounded, sigmoidal functions that are most frequently used in neural models. In some cases unbounded functions are also useful. From the point of view of efficiency and flexibility of the system good choice is provided by [2]: