A Gaussian synapse circuit for analog VLSI neural networks

摘要

Back-propagation neural networks with Gaussian function synapsesnhave better convergence property over those with linear-multiplyingnsynapses. In digital simulation, more computing time is spent onnGaussian function evaluation. We present a compact analog synapse cellnwhich is not biased in the subthreshold region for fully-parallelnoperation. This cell can approximate a Gaussian function with accuracynaround 98% in the ideal case. Device mismatch induced by fabricationnprocess will cause some degradation to this approximation. The Gaussiannsynapse cell can also be used in unsupervised learning. Programmabilitynof the proposed Gaussian synapse cell is achieved by changing the storednsynapse weight Wji, the reference current and the sizes ofntransistors in the differential pair