ED4-7

Demonstration of neurons with sigmoidal activation functions for superconducting neural networks

Dec.2 19:40-19:55 (Tokyo Time)

*Daiki Yamaguchi1, Yuki Yamanashi1, Nobuyuki Yoshikawa1

Department of Electrical and Computer Engineering, Yokohama National University, Japan1

An Artificial Neural Network (ANN) is a circuit that imitates information processing in a human brain. The ANN is effective for pattern recognition, filtering, and so on. A superconducting circuit can operate at the high frequency with low power consumption [1], it is attractive for realization of energy-efficient ANNs based on the superconducting circuits. It has been proven that the rf-SQUID shunted by the additional superconducting branch has the input-output characteristic approximated to a sigmoid function and can be used the neuron circuit [2]. In this study, we designed a neuron that has a sigmoid function as an activation function for realizing a superconducting ANN model that enables data analysis. Figure 1 shows the layout of the neuron circuit implemented by the AIST 10 kA/cm2 Nb process and the measured input-output characteristic. We confirmed that the designed neuron has the sigmoidal activation function and the activation function can be tuned by applying the magnetic field to the superconducting loops. We designed a simple superconducting ANN consists of three neurons and evaluated the response to 2-analog inputs. The circuit simulation result shows that the boundary between ‘0’ and ‘1’ outputs from the designed ANN can be modulated by tuning the characteristics of neurons in the ANN. There results indicate that we can build the energy-efficient ANN that enables learning of the ANN based on the learning algorithm such as the back propagation.

Fig. 1 The designed neuron and measured input-output characteristic.

Acknowledgment
This work was supported by JSPS KAKENHI under Grants JP18K04280. The circuits were fabricated in the clean room for analog-digital superconductivity (CRAVITY) of National Institute of Advanced Industrial Science and Technology (AIST) with the high-speed standard process (HSTP).

References
[1] N. Takeuchi et al., Supercond. Sci. Technol., vol. 26, art. no. 035010, 2013.
[2] I. I. Soloviev et al., J. Appl. Phys., vol. 124, art. no. 15211

Keywords: Artificial Neural Network, rf-SQUID

pict