SPIN Learning and Forgetting Surface Classifications with Dynamic Neural Networks

This paper refers to the problem of adaptability over an infinite period of time, regarding dynamic networks. A never ending flow of examples have to be clustered, based on a distancemeasure. The developed model is based on the self-organizing feature maps of Kohonen [6], [7] and some adaptations by Fritzke [3]. The problem of dynamic surface classification is embedded in the SPIN project, where sub-symbolic abstractions, based on a 3-d scanned environment is being done. 1. Survey First the framing project and the concrete problem context is discussed in short (chapter 2). Then in chapter 3 the network structure and associated aspects and problems are shown in detail, supported by simulation results (chapter 4). 2. SPIN-Project At the actual state of research the project SPIN (from Spatial Perception to Identification with Neural networks) is based on the data of a 3-d scanning device and designed to reach a stage of abstraction where convex clusters of surfaces are generalized, completed and classified. 2-1. Main strategies The system-design is based on some main principles, which have in common that none of them is in contradiction to a biological system. It is not intended to find the best fitting model for the lower levels of the mammal object-recognition -system, but obviously implausible features should be avoided. a. Learning instead of preprogrammed models The internal world model should be build up from a flow of examples scanned from the outer world. The "pre-programmed" knowledge is reduced to elementary features that should be searched for (here: edges). b. Hierarchy The main structure is pipeline-oriented instead of being controlled by a central instance (see the Neocognitron by Fukushima [5] for a good example of this strategy). c. Symmetry The general pur..