1D and 2D systolic implementations for radial basis function networks

Shows that radial basis function networks can be efficiently implemented on 1D and 2D systolic arrays. The authors discuss such networks in the framework of probability density function approximation for classification problems. In fact, the most computation intensive parts of the classification process consist in calculating pattern distances. In the initialisation phase of the algorithm this involves calculating the mutual (intra-class) distance matrix. The classification of an input vector mainly involves the calculation of the distance vector between this input vector and all the learning set vectors. The proposed implementations are 2D systolic, 1D pipeline and 1D parallel. Practical implementation issues are discussed for the MANTRA machine (2D grid) and the SMART Neuro-computer (1D ring).