Eksperimental'noe issledovanie peredatochnoy funktsii prototipa sverkhprovodyashchego gauss-neyrona

Abstract

The transfer function of a shunted two-junction interferometer, which was previously proposed as a basic element of superconducting neural networks based on radial basis functions, has been measured for the first time. The sample has been implemented in the form of a multilayer thin-film structure over a thick superconducting screen with the inductive supply of an input signal and the readout of an output signal. It has been found that the transfer function is the sum of the linear and periodic bell-shaped components. The linear component is likely due to the direct transfer of the input magnetic flux to the measuring circuit. The shape of the nonlinear component, which is the output signal of a Gauss neuron, can be approximately described by a Gaussian distribution function or, more precisely, by a parametric dependence derived theoretically in previous works. It has been shown that the transfer function of the Gauss neuron can depend on the choice of the working point of the measuring circuit, which promotes the development of integrated neural networks based on implemented elements.