RU2734581C1 - Superconducting neuron for multilayer perceptron - Google Patents

Abstract

FIELD: physics; electricity.

SUBSTANCE: invention relates to cryogenic micro- and nanoelectronics, including to element base of artificial neural networks. Proposed superconducting neuron for a multilayer perceptron comprises a Josephson junction connected galvanically to the main inductance, which is galvanically connected both to the shunt inductance and to the auxiliary inductance; means for setting input signal in form of magnetic field, connected inductively with main and auxiliary inductances; means of output signal detection in the form of current value connected inductively with shunting inductance. Besides, proposed device provides conversion of input magnetic flux at input F_in (created by means of magnetic field setting) into current at output I_out (read by means of detection) according to law described by sigmoid function.

EFFECT: technical result consists in improvement of speed and energy efficiency of superconducting neuron.

8 cl, 8 dwg

Description

The technical field to which the invention relates

The invention relates to cryogenic micro- and nanoelectronics, including the element base of artificial neural networks, and can be used to process broadband electromagnetic signals in the frequency range from hertz to 10 GHz. The invention can be used for a recognition procedure (signals, images) that requires processing large data arrays in real time, as well as when implemented at the hardware level using a specialized processor of clustering algorithms or adaptive control.

The presented invention is a superconducting neuron, a basic nonlinear element of an artificial neural network of the perceptron type, with a fast (operating clock frequencies of individual cells can be more than 100 GHz) one-cycle calculation of the activation function.

State of the art

Today, special computing modules are being actively developed, taking into account the features of the device of neural networks. However, until now, in this direction, there has not been found a sufficiently fast, compact and energy efficient technical solution for the implementation of basic nonlinear elements - neurons. As an example illustrating the problems of creating artificial neurons using traditional semiconductor technologies, let us consider a system with the possibility of multi-parameter classification. An artificial neuron is known (RU 2579958 C1), which contains a block of input signals, nodes for multiplying the input value by a weight factor, an adder, an activation function block. Such a neuron can be neither compact nor energy efficient, since each node for multiplying the input value by the weighting factor includes a multiplier of the input signal by the weighting factor and a unit for forming the weighting factor, memory cells of a read-only memory device with threshold values recorded in them, a switch, memory cells permanent memory with weighting factors recorded therein.

Also known is a device for the practical implementation of a neuron or its part (US 2016026912), which contains a processor core and a so-called "computational circuit". Moreover, the processor core includes logical circuits for determining a set of weighting factors for use in calculating convolutional neural networks and increasing the weighting factors using a scale of values. The computational chain allows you to obtain scale values, a set of weights, a set of input values, with each input value and associated weight having the same fixed size.

The basic element of an artificial neural network (US 20120011092, Tang Y. et al.) Can be built on the basis of a memristor that integrates an input current and a circuit that generates an output signal when a certain potential is reached, which allows simulating the threshold activation function of a real neuron. Here, a current pulse applied to the input decreases the memristor resistance, which leads to an increase in the output voltage. The state of the memristor remains after the voltage is removed, and to simulate a gradual drop in potential in the absence of an external signal, a small current is passed through the memristor in the opposite direction, gradually increasing its resistance. If, at the time of the next pulse, the output voltage exceeds the threshold value, then this triggers the output pulse generator. Thus, this device allows simulating the “integrate-and-fire” functionality, which allows it to be used primarily for so-called “spike” neural networks.

Of interest is the invention (US 4943556A, NN Szu), built on the interaction of electromagnetic (optical) radiation with a collection of electrons in thin superconducting filaments. A matrix of such filaments, in which the spreading of currents is realized with the help of light, can in principle be used in digital or analog computing systems. For the subsequent presentation, it is important to emphasize that such a neural network implies the presentation of information through the values of local magnetic fields.

It is possible to increase the efficiency of transmission over the network of information encoded in the form of magnitudes and directions of circular currents (local magnetic fields), as well as to create the nonlinear elements necessary for its functioning, using Josephson contacts - heterostructures, where “weak coupling” (thin layer of insulator, normal metal, etc.) between the superconducting electrodes. There are devices with Josephson junctions that convert a magnetic signal into a current response in a nonlinear manner and are used to detect the magnetic component of microwave electromagnetic signals. On the basis of the same heterostructures, digital circuits are realized for the subsequent processing of the received information using the approaches of fast one-quantum logic (BOC).

Systems based on superconducting quantum interferometers (SQUIDs), which have a high sensitivity to a magnetic signal, and, most importantly, a function of converting the magnetic component of the electromagnetic field at the input to the response of a current or voltage, which is tunable in the process of training a neural network, are of particular importance here. exit. It is this feature of such systems that will be used to build on their basis an artificial neuron with a fast one-cycle calculation of the activation function. It is especially worth emphasizing that the existing superconducting interferometers make it possible to build amplifiers with a low noise temperature, which are at the same time compatible with other superconducting devices used for information processing.

Said SQUID is a superconducting ring containing one or two Josephson contacts, and also has matching and measuring devices. The existing state of the art in this direction is determined by the designs and circuitry solutions for superconducting microwave amplifiers described in patent sources. For example, there is a known SQUID amplifier with a load for the range up to 0.1 GHz (JPS 60247311). For such a device, it has been reported that a gain of about 20 dB at 100 MHz at a noise temperature of 1 ± 0.4 K has been demonstrated (see US Pat. No. 4,585,999). The state of the art in the field of matching the input and output circuits of a SQUID amplifier with a load for the range up to 0.1 GHz can be appreciated from the same patent.

An important possibility for us to control the type of conversion of an input signal into a current / voltage response is described in the invention (JP 1245605), where an amplifier-converter of a magnetic signal on a DC SQUID is presented. In this technical solution, the second SQUID is used as an adjustable inductive element of the matching circuit to electrically adjust the matching of the interferometer with the cooled preamplifier. However, both the operating frequency range and the type of volt-flux conversion of amplifiers based on a single low-temperature SQUID are not suitable for creating a neuron for a multilayer perceptron.

The most important feature of SQUIDs with a multi-turn input coil is the non-linear form of the response to an externally specified magnetic field and the inability to linearize the response (or other type of control of the latter) using traditional feedback systems at such high frequencies, which leads to a low level of gain linearity (about 20 dB). It should also be borne in mind that without using feedback systems, SQUID amplifiers are characterized by small values of the dynamic range, that is, the ratio of the maximum output signal Vmax to the noise level at the output of the considered device Vƒ, which sets the level of the minimum signal at the output of the device:

DR = Vmax / Vƒ

This problem can be solved by using multi-element Josephson structures, including those consisting of N interferometers connected in series or in parallel. "Adding" the responses from the elements making up such a chain often allows the required dynamic range to be achieved.

To increase the linearity of the current-voltage conversion of systems based on superconducting quantum interferometers, suitable for the realization of synapses in a superconducting perceptron with magnetic presentation of information, a special feedback circuit has been proposed (KR 100774615, US 7453263), containing an additional Josephson junction. It should be noted that the proposed method for controlling the type of conversion of a magnetic signal into a response significantly limits the maximum operating frequency when used in artificial neural networks, since the operating frequency of the feedback circuit is low.

In the inventions (US 2005231196, US 7095227) magnetic signal converters based on a series of interferometers are described. This approach makes it possible to use a constant current source as a power source. However, the experimentally demonstrated device makes it possible to achieve only a relatively low linearity of signal amplification (no more than 20 dB).

Also known is a converter of the magnetic signal components based on a matrix of "high-temperature" SQUIDs cooled with liquid nitrogen (DE 3936686). The operating frequency range of such a system is limited from above to about 100 MHz. The invention (US 6005380) describes an antenna amplifier using magnetically coupled SQUIDs forming a multi-element matrix. Here the maximum operating frequency is 516 MHz.

To create a current amplifier with high output currents, it is proposed to use chains of parallel connected DC SQUIDs (JP 2003209299). This amplifier has high values of the output current, which makes it possible to use it as an input circuit of systems with direct digitization of the signal, but such an element is rather difficult to implement in the composition of superconducting artificial neural networks.

It is known to use a chain of variable area SQUIDs (SKIF) as a highly sensitive magnetometer with a specific type of volt-flux conversion with a pronounced dip in this dependence in the vicinity of zero values of the applied field strength (WO 01/25805; US 7369093). However, the described structures do not provide the necessary kind of non-linear relationship between the input "magnetic" signal and the voltage response. Do not give the required conversion of the input magnetic signal and inhomogeneous parallel chains of contacts, which do not allow everything else to achieve significant values of the output signal.

The known design of an amplifier-converter for a magnetic signal based on series chains of two-contact interferometers, the areas of which are selected so as to provide an increased gain and high linearity in the frequency band 1-10 GHz (RU 2353051). However, the implementation of the required chains of SQUIDs with a given accuracy of parameters today seems to be an extremely difficult technical problem, which limits the linearity of conversion of a magnetic signal into a voltage response at a level of 50 dB attainable in a real experiment. Moreover, such a system is certainly not compact.

A structure with three Josephson contacts (US 8179133; US 8933695; US 10333049 B1) is described to obtain a special (triangular) type of volt-flux conversion in an interferometer. In this element, parallel to the main inductance of a two-contact SQUID, the third Josephson junction must be switched on, which is always in a superconducting state and plays the role of a nonlinear inductance. Connecting a Josephson junction in parallel with the inductance of a SQUID forms a double SQUID (also known as a bi-SQUID). Such a cell makes it possible to achieve a sufficiently high linearity in the conversion of the input signal into the output voltage of the amplifier. However, in this case, the influence of parasitic inductances and capacitances makes it difficult to obtain a highly linear conversion of a magnetic signal into a voltage response. A similar device (RU 2544275) was proposed for implementation based on high-temperature superconductors. It should be emphasized that the problem of the practical manufacture of superconducting devices for broadband receiving systems with direct digitization of signals using such a cell has not yet been solved.

Thus, the technical problem solved by the claimed invention is the lack of high-speed and energy-efficient neural networks based on Josephson heterostructures with information coding in the form of the magnitude of local currents and associated magnetic fields. Existing technical solutions based on semiconductor technologies (based on CPU, GPU, FPGA technologies) either do not have a high "parallelism", or do not make it possible to organize the required number of branches, or do not provide the required interconnection resource. The most serious problems arise in the implementation of the basic elements of neural networks, such as neurons and synapses: each element consists of dozens of transistors, which limits the level of complexity of the created circuit, while the circuits themselves are characterized by an excessive number of elements and low energy efficiency.

The use of superconducting technical solutions, as can be seen from the state of the art, makes it possible to reduce energy dissipation to the level of 10 ^-18 J per operation at a sufficiently high speed. The operating frequency for individual cells can be over 100 GHz. However, to date, in the field of neuromorphic computers, no design solutions are known for a multilayer perceptron (Rumelhart's perceptron) with a magnetic representation of information using superconducting circuits. Such an artificial neural network contains layers of elements that nonlinearly transform the input signal (neurons), connected by linear tunable connections (synapses). The most characteristic features of the studied neural networks include the use of a nonlinear, continuously differentiable sigmoidal activation function (or the function of converting an input signal into an output signal).

It is possible to implement superconducting artificial neurons based on two-contact interferometers and their modifications operating in a resistive mode. The signal level here should correspond to the average voltage level on the cell. But such schemes will not differ in speed and energy efficiency. A similar problem is typical for "spike" (impulse) neural networks, in which information is presented in the form of a sequence of identical (one-quantum) voltage pulses, while the signal is determined by the time delay between the pulses. The main problem with this approach is the difficulty of realizing effective synaptic connections.

Disclosure of the essence of the invention

что необходимо для однотактового вычисления функции активации искусственных нейронов для многослойного персептрона на основе таких структур.The technical result of the invention is the creation of a superconducting neuron for a multilayer perceptron with magnetic presentation of information, capable of demonstrating a high level of performance and energy efficiency. In addition, the claimed device provides the ability to convert the input magnetic flux at the input Ф _in (created by means of setting the magnetic field) into the current at the output I _out (read _{out by the} detection tool) according to the law

which is necessary for one-cycle calculation of the activation function of artificial neurons for a multilayer perceptron based on such structures.

The inventive superconducting neuron for a multilayer perceptron contains a Josephson cell galvanically connected to the main inductive element, which is galvanically connected to the shunt and auxiliary inductive elements; means for setting an input signal in the form of a magnetic field, inductively coupled to the main and auxiliary inductive elements; means for detecting an output signal in the form of a current value, inductively coupled to the shunt inductive element.

The main, auxiliary and shunt inductive elements can be made in the form of a portion of a layer (thin film) of a superconducting material.

основного индуктивного элемента лежит в диапазоне от 0.1 до 0.5, где L_in1 - величина индуктивности основного индуктивного элемента, Ф₀ - квант магнитного потока, равный 2.067×10^-15 Вб; I_C - критический ток джозефсоновского контакта. Безразмерная величина индуктивности шунтирующего индуктивного элемента

лежит в диапазоне от 0.1 до 0.4, где L_out - величина индуктивности шунтирующего индуктивного элемента. Безразмерная величина индуктивности вспомогательного индуктивного элемента

лежит в диапазоне от 1.1 до 1.5, где L_in2+L_q - величина индуктивности вспомогательного индуктивного элемента.Dimensionless value of inductance

the main inductive element is in the range from 0.1 to 0.5, where L _in1 is the inductance value of the main inductive element, Ф ₀ is the magnetic flux quantum equal to 2.067 × 10 ^-15 Wb; I _C is the critical current of the Josephson junction. Dimensionless value of the inductance of the shunt inductive element

lies in the range from 0.1 to 0.4, where L _out is the value of the inductance of the shunting inductive element. Dimensionless value of the inductance of the auxiliary inductive element

lies in the range from 1.1 to 1.5, where L _in2 + L _q is the inductance value of the auxiliary inductive element.

The Josephson contact in one of the embodiments of the invention is made in the form of a layered thin-film structure containing: a lower superconducting electrode deposited on the lower superconducting electrode of an insulator layer, an upper superconducting electrode with current leads deposited on the insulator layer.

Either niobium or niobium-based compounds (for example, NbN, NbTi), or aluminum, or an aluminum-based alloy can be used as the material of the lower and upper superconducting electrode of all superconducting layers. As the material of the insulator layer in the composition of the Josephson junction, aluminum oxide Al ₂ O ₃ can be used, which can be replaced by metal compounds with high resistivity (for example, Nb ₃ Si). The thickness of the lower superconducting electrode and the upper superconducting electrode is generally 50 to 500 nm, and the thickness of the insulator layer is 1 to 20 nm.

Brief Description of Drawings

The invention is illustrated by drawings, where Fig. 1 shows an electrical diagram of the inventive superconducting neuron for a multilayer perceptron; in fig. 2 shows a top view of one of the possible options for the practical implementation of the inventive superconducting neuron for a multilayer perceptron, FIG. 3-5 show side views in section on one of the options for the practical implementation of the inventive superconducting neuron for a multilayer perceptron, where the value of the auxiliary inductance, made in the form of a section of the superconducting material layer, is L _in2 + L _q .

FIG. 6 shows the functions connecting the magnetic (normalized) flux at the input Ф _in and the current at the output I _out for the inventive superconducting neuron with an adiabatic change in the input signal, curve 101 corresponds to the case 2π⋅L _in1 I _C = 0.125⋅F ₀ ; 2π⋅L _out I _C = 0.4⋅F _0; curve 102 corresponds to the case 2π⋅L _in1 I _C = 0.125⋅F ₀ ; 2π⋅L _out I _C = 0.3⋅F ₀ , curve 103 corresponds to the case 2π⋅L _in1 I _C = 0.4⋅F ₀ ; 2π⋅L _out I _C = 0.3⋅F ₀ ; dots represent target (sigmoidal) functions in a given range of values of the normalized current at the output of the considered nonlinear current converter. FIG. 7 - standard deviation (SD) of the calculated neuron activation functions from target (sigmoidal) functions for different values of the main, L _in1 , and shunting, L _out , inductances. FIG. 8 shows the dependence of the dissipation per one operation on the duration of such an operation for various values of the inductance of a superconducting neuron.

Positions in the drawings indicate:

1 - Josephson junction (element) with critical current I _C ;

2 - main inductance (main inductive element), made in the form of a section of a layer of superconducting material, the value of which is designated as L _in1 ;

3 - auxiliary inductance (auxiliary inductive element), made in the form of a section of a layer of superconducting material, the value of which is designated as L _in2 + L _q ;

4 - shunt inductance (shunt inductive element) made in the form of a portion of a layer of superconducting material, the value of which is denoted as L _out ;

5 - means for setting the magnetic field, inductively coupled with the main and auxiliary inductances;

6 - means for detecting the amount of flowing current, inductively coupled to the shunt inductance.

Implementation of the invention

Below is a more detailed description of the claimed invention, which does not limit the essence presented in the independent claim, but only demonstrates the possibility of achieving the claimed technical result.

The electrical circuit of the inventive superconducting neuron is shown in FIG. 1. Such a neuron converts the input magnetic flux at the input Ф _in (created by the means of setting the magnetic field 5) into the current at the output I _out (read by the detection means 6). The inventive device contains a Josephson contact 1, which ensures the nonlinearity of the said conversion. Obtaining a nonlinear, continuously differentiable sigmoidal activation function (or the function of converting an input signal into an output signal) is ensured by the presence of three inductive elements in the neuron: main 2, auxiliary 3 and shunt 4. Josephson contact (element) 1 is characterized by the value of the critical current I _C ; main inductance (main inductive element) 2 is characterized by its value L _in1 ; auxiliary inductance 3 is characterized by its value, designated as L _in2 + L _q , and only the first part of it is connected magnetically with the means for setting the magnetic field 5; shunt inductance 4 is characterized by its value L _out .

FIG. 2, 3, 4 and 5 are top and side views (in section) on one of the possible options for the practical implementation of the inventive superconducting neuron for a multilayer perceptron. Shown here are both the design (shape, dimensions) and the relative position

- Josephson contact (element) 1,

- the main inductive element 2, made in the form of a portion of a layer of superconducting material;

- auxiliary inductive element 3, made in the form of a portion of a layer of superconducting material;

- shunt inductive element 4, made in the form of a portion of a layer of superconducting material;

- means for setting the magnetic field 5, inductively coupled to the main and auxiliary inductive elements.

A superconducting quantum interferometer with two Josephson contacts is used as a means for detecting the amount of flowing current 6 inductively coupled to the shunt inductance. The directions and areas for which FIG. 3, 4 and 5 are side cross-sectional views indicated by arrows, dashes and corresponding legends in FIG. 2.

FIG. 5 shows the result of the proposed device, demonstrating the implementation of its purpose. Here are the functions connecting the magnetic (normalized) flux at the input Ф _in and the current at the output I _out for the inventive superconducting neuron with an adiabatic change in the input signal. Curve 101 corresponds to the case 2π⋅L _in1 I _C = 0.125⋅F ₀ ; 2π⋅L _out I _C = 0.4⋅F _0; curve 102 corresponds to the case 2π⋅L _in1 I _C = 0.125⋅F ₀ ; 2π⋅L _out I _C = 0.3⋅F ₀ . Curve 103 corresponds to the case 2π⋅L _in1 I _C = 0.4⋅F ₀ ; 2π⋅L _out I _C = 0.3⋅F ₀ ; The dots represent the target (sigmoidal) functions in a given range of values of the normalized current at the output of the considered nonlinear current converter. Good coincidence of the obtained and target activation functions proves the possibility of carrying out the invention with the achievement of the claimed technical result.

The following example of a specific implementation of the invention, which requires the sequential deposition of three layers of superconducting material and two layers of a dielectric on a substrate, is included in many potentially suitable solutions, but does not exhaust the specified set. In particular, it is possible to use Josephson contacts, where instead of an insulator, high-resistance metals (Nb ₃ Si), ferromagnetic metals (for example, based on CuNi alloy) or multilayer structures that combine insulator layers and ferromagnetic layers are used.

To implement the inventive superconducting neuron, it is necessary to use a superconducting screen (a superconductor layer on the substrate). This is due to the fact that it is necessary to set the same magnetic flux in inductive elements, the values of which differ by a factor of 10. This is possible only under the condition that the means for setting the magnetic field is connected only with the parts of the contours of the same size, and all the other parts are at zero magnetic field. In addition, the small value of the inductance of the main layer of the superconducting neuron (normalized inductance 0.125) forces the use of contacts with a low critical current (from 10 to 20 μA).

The production of a sample in a specific embodiment of the invention begins with the deposition of a three-layer Nb-Al / AlOx-Nb preform (layer thicknesses of 300/10/150 nm). The large thickness of the lower electrode (approximately equal to four London lengths in niobium at helium temperature) makes it also possible to use it as a superconducting shield for a magnetic field (a superconducting layer 150 nm thick practically does not shield an external magnetic field).

At the second stage of sample preparation using photolithography and subsequent plasma-chemical etching of the upper niobium layer, a blank is formed for the subsequent creation of tunnel Josephson junctions. Aluminum is a stop layer for plasma-chemical etching. After the end of the niobium etching, the aluminum layer must be removed, for example, in an alkali solution of potassium hydroxide (KOH).

After the formation of all Josephson contacts, the superconducting screen is fabricated using photolithography and subsequent plasma-chemical etching, or using a liquid niobium etchant (in a mixture of nitric and hydrofluoric acids). The advantage of the first method is the greater controllability of the process: the etching depth can be controlled using a laser thickness gauge, or the process can stop at the aluminum sublayer. The advantage of the second method (“wet chemistry” method) is the option of cleaning the substrate from contaminants formed at the previous stages. In this layer, contact pads are formed for connecting current sources and a voltmeter to the superconducting screen.

The next step is to form an insulation layer. The deposition can be performed by thermal deposition of silicon monoxide, or by sputtering silicon or aluminum in an argon-oxygen plasma. The formation of windows in the insulation layer can be performed by explosive photolithography, or by plasma-chemical etching under a mask. In this exemplary embodiment of the patentable invention, the thickness of the first insulation layer was 300 nm.

At the fifth stage, a thin niobium film with a thickness of 100 nm is formed for the subsequent creation of the main and auxiliary inductances, as well as elements of the means for detecting the value of the flowing current. The shunt inductance necessarily includes a portion made in the uppermost superconducting film to provide inductive coupling with the detection means. These elements can be formed by subsequent plasma-chemical etching under a photoresist mask, or by means of explosive photolithography. In the first case, it is advisable to use an aluminum sublayer to avoid additional underlaying of the lower insulation layer. After the end of the plasma-chemical etching, the open part of the aluminum sublayer must be removed by etching in KOH alkali.

The essence of the sixth stage is to isolate the main and auxiliary inductors using a dielectric film 200 nm thick. The deposition and formation of the layer can be carried out by the same methods as in the fourth stage.

The seventh stage is the formation of shunt resistors for tunneling Josephson junctions. This design assumes a 200 nm thick CuAl layer with a resistance of 0.8 Ohm per square. The layer can be formed by explosive photolithography (lift-off) or by chlorine-based plasma-chemical etching. To use plasma-chemical etching in freon (CF4 and higher compounds, SF6), the selection of the appropriate material is required.

The last stage is the formation of a superconducting circuit (wiring) from niobium 450 nm thick. To ensure good electrical contact to the resistor layer and the top electrode of the Josephson contacts, the substrate is ion etched, equivalent to etching 5-10 nm of niobium. The formation of the layer can be performed using explosive photolithography (lift-off). Plasma-chemical etching under a photoresist mask can also be used. In this film, the control line of the means for setting the magnetic field, the shunting ("output") inductance, as well as all connections to the main and auxiliary inductances and the superconducting screen are formed.

The detection means in this embodiment of the patentable invention is an asymmetric superconducting quantum interferometer with two Josephson contacts (the optimal critical current of which is in the range from 100 to 150 μA).

FIG. 7 shows the standard deviation (SD) of the calculated neuron activation functions from the target (sigmoidal) functions for different values of the main, L _in1 , and shunting, L _out , inductances. The value of such a standard deviation is much less than one for all ranges of values stated in the disclosure of the essence of the invention for physical quantities characterizing structural elements. Good coincidence of the obtained and target activation functions proves the achievement of the technical result in the entire declared range of values of the main, L _in1 , and shunting, L _out , inductances.

FIG. 8 shows the dependence of the dissipation per one operation on the duration of such an operation for various values of the inductance of a superconducting neuron. Curve 104 corresponds to the case 2π⋅L _out I _C = 0.1⋅F ₀ . Curve 105 corresponds to 2π⋅L _out I _C = 0.5⋅F ₀ . The presented data confirm the achievement of the technical result in terms of speed and energy efficiency in the functioning of a superconducting neuron: for the selected values of the characteristics of structural elements, energy dissipation is less than 10 ^-18 J per operation with an operation duration of less than 0.1 ns.

Claims (9)

1. A superconducting neuron for a multilayer perceptron containing a Josephson cell, galvanically connected to the main inductive element, which is galvanically connected to the shunt and auxiliary inductive elements; means for setting an input signal in the form of a magnetic field, inductively coupled to the main and auxiliary inductive elements; means for detecting an output signal in the form of a current value, inductively coupled to the shunt inductive element. 2. The superconducting neuron according to claim 1, characterized in that the main, auxiliary and shunting inductive elements are made in the form of a portion of a layer of superconducting material. 3. The superconducting neuron according to claim 1, characterized in that the dimensionless inductance value

the main inductive element is in the range from 0.1 to 0.5, where L _in1 is the inductance value of the main inductive element, Ф ₀ is the magnetic flux quantum equal to 2.067 × 10 ^-15 Wb; I _C is the critical current of the Josephson junction. 4. A superconducting neuron for a multilayer perceptron according to claim 1, characterized in that the dimensionless value of the inductance of the shunting inductive element

lies in the range from 0.1 to 0.4, where L _out is the value of the inductance of the shunting inductive element. 5. A superconducting neuron for a multilayer perceptron according to claim 1, characterized in that the dimensionless value of the inductance of the auxiliary inductive element

lies in the range from 1.1 to 1.5, where L _in2 + L _q is the inductance value of the auxiliary inductive element. 6. A superconducting neuron for a multilayer perceptron according to claim 1, characterized in that the Josephson contact is made in the form of a layered thin-film structure containing: a lower superconducting electrode deposited on the lower superconducting electrode an insulator layer, an upper superconducting electrode with current leads deposited on the insulator layer. 7. A superconducting neuron for a multilayer perceptron according to claim 6, characterized in that either niobium or niobium-based compounds (for example, NbN, NbTi), or aluminum, or an alloy are used as the material of the lower and upper superconducting electrodes, all superconducting layers based on aluminum. 7. A superconducting neuron for a multilayer perceptron according to claim 6, characterized in that aluminum oxide Al ₂ O ₃ or Nb ₃ Si is used as the material of the insulator layer in the Josephson junction. 8. The superconducting neuron for a multilayer perceptron according to claim 6, characterized in that the thickness of the lower superconducting electrode and the upper superconducting electrode is from 50 to 500 nm.

Images