-
Optimisation challenge for superconducting adiabatic neural network implementing XOR and OR boolean functions
Authors:
D. S. Pashin,
M. V. Bastrakova,
D. A. Rybin,
I. I. Soloviev,
A. E. Schegolev,
N. V. Klenov
Abstract:
In this article, we consider designs of simple analog artificial neural networks based on adiabatic Josephson cells with a sigmoid activation function. A new approach based on the gradient descent method is developed to adjust the circuit parameters, allowing efficient signal transmission between the network layers. The proposed solution is demonstrated on the example of the system implementing XO…
▽ More
In this article, we consider designs of simple analog artificial neural networks based on adiabatic Josephson cells with a sigmoid activation function. A new approach based on the gradient descent method is developed to adjust the circuit parameters, allowing efficient signal transmission between the network layers. The proposed solution is demonstrated on the example of the system implementing XOR and OR logical operations.
△ Less
Submitted 6 May, 2024;
originally announced May 2024.
-
Adiabatic Superconducting Artificial Neural Network: Basic Cells
Authors:
I. I. Soloviev,
A. E. Schegolev,
N. V. Klenov,
S. V. Bakurskiy,
M. Yu. Kupriyanov,
M. V. Tereshonok,
A. V. Shadrin,
V. S. Stolyarov,
A. A. Golubov
Abstract:
We consider adiabatic superconducting cells operating as an artificial neuron and synapse of a multilayer perceptron (MLP). Their compact circuits contain just one and two Josephson junctions, respectively. While the signal is represented as magnetic flux, the proposed cells are inherently nonlinear and close-to-linear magnetic flux transformers. The neuron is capable of providing a one-shot calcu…
▽ More
We consider adiabatic superconducting cells operating as an artificial neuron and synapse of a multilayer perceptron (MLP). Their compact circuits contain just one and two Josephson junctions, respectively. While the signal is represented as magnetic flux, the proposed cells are inherently nonlinear and close-to-linear magnetic flux transformers. The neuron is capable of providing a one-shot calculation of sigmoid and hyperbolic tangent activation functions most commonly used in MLP. The synapse features by both positive and negative signal transfer coefficients in the range ~ (-0.5,0.5). We briefly discuss implementation issues and further steps toward multilayer adiabatic superconducting artificial neural network which promises to be a compact and the most energy-efficient implementation of MLP.
△ Less
Submitted 25 July, 2018;
originally announced July 2018.
-
Analytical derivation of DC SQUID response
Authors:
I. I. Soloviev,
N. V. Klenov,
A. E. Schegolev,
S. V. Bakurskiy,
M. Yu. Kupriyanov
Abstract:
We consider voltage and current responses formation in DC SQUID with overdamped Josephson junctions in resistive and superconducting state in the frame of resistively shunted junction (RSJ) model. For simplicity we neglect the junction capacitance and the noise effect. Explicit expressions for the responses in resistive state were obtained for a SQUID which is symmetrical with respect to bias curr…
▽ More
We consider voltage and current responses formation in DC SQUID with overdamped Josephson junctions in resistive and superconducting state in the frame of resistively shunted junction (RSJ) model. For simplicity we neglect the junction capacitance and the noise effect. Explicit expressions for the responses in resistive state were obtained for a SQUID which is symmetrical with respect to bias current injection point. Normalized SQUID inductance $l = 2 e I_c L/\hbar$ (where $I_c$ is the critical current of Josephson junction, $L$ is the SQUID inductance, $e$ is the electron charge and $\hbar$ is the Planck constant) was assumed to be within the range $l \leq 1$, subsequently expanded up to $l \approx 7$ using two fitting parameters. SQUID current response in superconducting state was considered for arbitrary value of the inductance. Impact of small technological spread of parameters relevant for low-temperature superconductor (LTS) technology was studied with generalization of the developed analytical approach for a case of small difference of critical currents and shunt resistances of the Josephson junctions, and inequality of SQUID inductive shoulders for both resistive and superconducting states. Comparison with numerical calculation results shows that developed analytical expressions can be used in practical LTS SQUIDs and SQUID-based circuits design, e.g. large serial SQIF, drastically decreasing the time of simulation.
△ Less
Submitted 8 August, 2016;
originally announced August 2016.
