Non-Gaussian kernel circuits in analogue VLSI: implications for RBFnetwork performance

摘要

The authors present a cascadable circuit for the distance metricnand nonlinear functions required by radial basis function (RBF) neuralnnetworks. The distance metric is a quadratic approximation to thenEuclidean distance between two voltages, and the nonlinearity isnproduced using two MOS transistors. This circuit has been developed fornpulse stream neural systems. The operation of the circuit is describednand suggestions are made for its practical implementation in pulsednanalogue VLSI. Since the nonlinearity generated by the circuit has notnbeen used in RBFs before, software results are presented to demonstratenthat the circuit can produce good classification performance. However,nsoftware simulations show that the shape of the nonlinearity hasnimplications for the performance of RBFs using the circuit. The authorsnconsider the implications of these results to the development of pulsednanalogue RBF chips in the limited precision environment of analoguenVLSI. Based on their findings, they make suggestions for the shape andnrange of future centre circuits to make them robust to this limitednprecision and which they believe will help ensure good classificationnperformance is obtained in hardware