A competitive system with adaptive gain tuning

Competition is an essential mechanism for self-organizing neural networks, and its properties affect the performance of the networks. Properties of competitive systems such as, topology preservation, uniform sparseness of outputs and low activities for novel input patterns are favored for constructing multi-layered self-organizing neural networks. Furthermore these properties should be acquired depending on the various input sources without tuning parameters manually. The popular competitive systems, such as lateral inhibition or winner-take-all circuits, don't satisfy these requirements. The authors propose a competitive system with adaptive gain tuning. Competition is accomplished by controlling the gain of the processing units. This gain is adapted to keep a summation or a maximum of the outputs close to the specified value. In this network, ad hoc topology is not introduced. Therefore, if there are enough cells, the mapping generated by the network is continuous, and it is possible to preserve the topology of the input manifold through the transformation from the input space to the output space. Further, with numerical simulations we show that the excited region of each cell is sharpened in proportion to the distribution density of template vectors of the network. Since the outputs keep low activities for novel input patterns, they can be distinguished from already learned input patterns.<>