Pseudo Sigmoid Function Generator for a Superconductive Neural Network

Abstract

A superconductive perceptron, an artificial neural network, has been investigated using single flux quantum (SFQ) stochastic logic. A superconductive pseudo sigmoid function generator that corresponds to an artificial neuron device for the perceptron has been proposed and implemented using an SFQ current comparator and a frequency-to-current converter, which generates current that is proportional to the average input SFQ frequency. A frequency-to-current converter has been implemented using a dc-SQUID voltage driver coupled with a Josephson transmission line. We implemented and tested the pseudo sigmoid function generator using the SRL 2.5 kA/cm 2 Nb process. The measured input-output characteristic agreed with the ideal sigmoid function with an average error of 0.063%.

Adiabatic superconducting cells for ultra-low-power artificial neural networks

“…In subsequent variations [16–17], this feature was found to drastically reduce the energy efficiency of the superconducting circuit. In another recent approach to multilayer perceptron, SQUIDs were utilized as nonlinear magnetic flux transducers, allowing the ANN to persist in the superconducting state [18].…”

Theoretical basis of SQUID-based artificial neurons

“…In fact, ANN using Josephson elements has been proposed 5,6 in the early 90s, there have been successful implementations using different ways. [7][8][9][10][11] In the previous studies, different types of activation functions in artificial neurons are employed to construct the superconducting artificial neural networks. For example, the two-stage coupled SQUID with a cascade connection produced a step-like function 9 and the rapid single flux quantum (RSFQ) comparator based on the statistical transition provided the error function similar to sigmoid function 10 .…”

“…Interestingly, Yamanashi et. al., found the pseudo sigmoid function even in the RSFQ comparator 11 . However, there has been no existing system that can exactly generate sigmoid function for error back propagation learning algorithm.…”

Synaptic weighting in single flux quantum neuromorphic computing

“…This in part has led to a renewed exploration of neuromorphic hardware with the goal of further accelerating the advance of artificial neural networks. One promising hardware platform can be made from naturally spiking Josephson junctions (JJs) [8][9][10][11][12] . Because of this spiking behavior, JJs have been proposed to simulate the interactions among neurons, and have demonstrated biologically realistic neuron behavior with only two junctions 13 .…”

“…Because of this spiking behavior, JJs have been proposed to simulate the interactions among neurons, and have demonstrated biologically realistic neuron behavior with only two junctions 13 . Further, they have been proposed as a way to implement stochastic neural networks 14,15 , and have been demonstrated to implement a sigmoid like transfer function of fast voltage spikes 10 .…”