JP2694778B2 - Superconducting neuron element - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a neuron element as a basic unit for realizing a neural network on a device, and more particularly to a neuron element using a superconductor.

[0002]

2. Description of the Related Art A neuron element, which is a basic element for realizing a so-called artificial neural network imitating the information processing method of the human brain on a device, has been conventionally made by using a semiconductor integrated circuit. On the other hand, when a superconductor is used for the neuron element, there is an advantage that heat generation in the wiring portion can be suppressed because the power consumption is extremely small. Also, in a superconducting integrated circuit, the active element (Josephson element) is inherently easily made three-dimensional, and the wiring between neuron elements is greatly facilitated compared to a semiconductor integrated circuit in which it is difficult to make the active element three-dimensional. there were.

A conventionally known neuron element using a superconductor is a magnetic flux quantum existing in a Josephson line in the collection of papers C-2-19 of the Autumn National Conference of the Institute of Electronics, Information and Communication Engineers in 1988. There is a neuron element whose propagation density is an information carrier.

The basic operation of the neuron element is that a weight Wi is applied to each input (synaptic operation) and output when the sum Σ (Xi · Wi) of products of the weight and each input value Xi exceeds a certain threshold value h. To output (neuron operation).

FIG. 7 shows a basic element structure of a neuron element using the above Josephson line. 101 of FIG.
Reference numerals 102, 103 and 104 denote Josephson lines, each of which is an elongated Josephson junction composed of a lower electrode 111 and an upper electrode 113 made of a superconductor and a thin tunnel barrier 112 made of an insulative body. The flux quanta, which is an information carrier, moves in the tunnel barrier of the Josephson line in the longitudinal direction of the line. Track A (101) and Track B (10
2) is connected between the inductor 114 and the resistor 115. When a quantum magnetic flux having a certain propagation density is given as an input from the input end 106, the magnetic flux quantum is output from the output end 107 through the line B (102) when the propagation density is higher than a certain threshold value. At this time, track B (102)
116 magnetically coupled to the upper electrode of the
By causing the circulating current associated with the magnetic flux quantum stored in the superconducting loop of the magnetic flux stop unit 105 to act,
The threshold can be controlled. The above is the neuron operation. In addition, lines C (103) and D (1) from the excitatory control input terminal 108 and the inhibitory control input terminal 109, respectively.
04) to propagate the magnetic flux quantum, by changing the number and direction of the quantum magnetic flux stored in the magnetic flux quantum stop unit 105, it is possible to control the circulating current, it is possible to control the propagation probability, The inputs can be weighted. This is a synaptic action. As described above, the circuit shown in FIG. 7 can perform a neuron operation and a synapse operation, and therefore can be used as a neuron element by combining these.

[0006]

Since the neural network is a massively parallel processing, each neuron element is connected to many other neuron elements and requires many fan-ins and fan-outs. On the other hand, in the neuron element using the Josephson line described in the conventional technique, the propagation probability of the magnetic flux quantum from the line A (101) to the line B (102) in FIG. It's nothing but. Amplification of the magnetic flux quantum number using a T-branch has been proposed to compensate for the attenuation of this signal, but the gain is not so large and it is impossible to obtain much fanout in this neuron element.

As described above, the neuron element described in the prior art can perform the basic operation as a neuron element, but has a major drawback that the number of fanouts is insufficient to form a neural network. Had.

An object of the present invention is to provide a neuron element using a superconductor which can obtain a large fan-in / fan-out number.

[0009]

According to a first aspect of the present invention, a switching element using a Josephson element, a superconducting closed loop, and an output line magnetically coupled to another neuron element are provided. The conduction closed loop and one or more input lines are magnetically coupled to each other, and the inductance value of the coupling is controlled to weight the inputs, and the sign of the weight is determined depending on the direction of the current flowing through the input line. The switching element and the output line are connected in parallel to a bias input terminal, and the superconducting closed loop and the switch of the switching element are magnetically coupled to obtain a superconducting neuron element. .

According to a second aspect of the present invention, a switching element using a Josephson element, a superconducting closed loop, and an output line magnetically coupled to another neuron element are included, and one or more of the superconducting closed loop are included. The input elements are magnetically coupled to each other to control the value of the current flowing through the input lines to weight the inputs, and the sign of the weight is determined according to the direction of the current flowing through the input lines in the previous period. And the output line are connected in parallel to a bias input terminal, and the switching element switches when the circulating current flowing in the closed loop of superconductivity exceeds a certain threshold value. To be

According to a third aspect of the present invention, the superconducting closed loop includes a switching element using a Josephson element, a superconducting closed loop, and an output line made of a superconductor magnetically coupled to another neuron element. And one or more input lines are magnetically coupled to each other, the switching element switches when the circulating current flowing in the superconducting closed loop exceeds a certain threshold value, the switching element is driven by a DC bias, and A switching element and the output line are connected in parallel to the bias input terminal, and a down-edge trigger circuit is included in the output line, which switches when a circulating current flowing in the superconducting closed loop falls below a certain threshold value. A superconducting neuron element characterized by is obtained.

According to a fourth aspect of the present invention, a switching element using a Josephson element, a superconducting closed loop, and an output line magnetically coupled to another neuron element are included, and the superconducting closed loop is one or more. A superconducting neuron characterized in that the input lines are magnetically coupled, the switching element and the output line are connected in parallel to a bias input terminal, and the superconducting closed loop is included in the switching element. The device is obtained.

According to a fifth aspect of the present invention, a switching element using a Josephson element, a superconducting closed loop, and an output line magnetically coupled to another neuron element are included, and one or more superconducting closed loops are included. Input line is magnetically coupled, and when the circulating current flowing in the superconducting closed loop exceeds a certain threshold value, the switching element switches, and the switching element and the output line are parallel to the bias input terminal. A superconducting neuron element is obtained which is connected to the output line and includes a resistance component in the output line.

[0014]

When a current is passed through the input line magnetically coupled to the superconducting closed loop, a circulating current flows in the superconducting closed loop to prevent the magnetic field from entering the superconducting closed loop. This circulating current value Icir is Icir = M /, where L is the inductance of the superconducting closed loop, M is the mutual inductance between the input line and the superconducting closed loop, and I is the current flowing through the input line.
It is represented by L · I. This is a synapse operation that weights (M / L) the input.

When there are many such input lines, the current Ii flowing through the i-th input line and the mutual inductance between the i-th input line and the superconducting closed loop are Mi, the circulating current is Icir = Σ (Mi / L -Ii), which is expressed in a form in which each input current is linearly combined with a weight (Mi / L). Also, by changing the direction of the coupling of the input line to the superconducting closed loop, a reverse circulating current can be induced in the superconducting closed loop,
Negative weights can be easily created. The Josephson element is an element in which a current flows with zero resistance when the input is below a certain threshold value and a voltage is generated across the input when the input exceeds a certain threshold value. Therefore, a switching element including a Josephson element that switches by a current flowing in a superconducting closed loop performs a neuron operation in which inputs are linearly coupled with weights (Mi / L) and output when the value is above a certain threshold value.

From the above, it can be seen that the superconducting neuron element according to the present invention has a function as a neuron element for synapse operation and neuron operation. In addition, since the output line is magnetically coupled to other neuron elements, basically only one output line is required, and the current value flowing through the output line does not depend on the inductance of the output line. Is constant regardless of the number of connected neurons (fan-out number). Therefore, the superconducting neuron device according to the present invention can obtain many fanouts. Further, the fan-in can be easily increased because only the magnetic coupling between the superconducting closed loop and the input line is required.

In the second aspect of the present invention, the input current is processed in advance to include the weight in the input current value. Therefore, the synaptic operation can be performed without changing the mutual inductance value between the input line and the superconducting closed loop depending on the input line.

Since the switching element using the Josephson element is a latch element, once it is switched to the voltage state, it is not reset unless the bias current falls below a certain value. In the third aspect of the present invention, since the output line that is the load of the switching element is superconducting, when the switching element switches, all the bias current flows toward the output line, and no current flows through the switching element. Reset the element. On the other hand, the current flowing through the output line has already been reset to the superconducting state because the down-edge trigger circuit switches when the current flowing in the closed loop of the superconducting current drops below a certain value and a voltage is generated in the output line. Flows into. Therefore, the switching element and the output line are flip-flop circuits that set and reset according to the current flowing in the superconducting closed loop, that is, the input current. Therefore, this neuron element is a circuit that does not require a timing signal, and can be operated asynchronously with other neuron elements in the neural network.

In a fourth aspect of the present invention, a switching element is included in the superconducting closed loop. Therefore, the switching element switches with the sum of the circulating current flowing in the superconducting closed loop and the bias current.

In the fifth aspect of the present invention, since the output line includes a resistance component, the current flowing through the output line decreases with a constant time constant as the bias current decreases, and the reset operation of the output line is unnecessary.

[0021]

(Embodiment 1) FIG. 1 is a diagram for explaining an embodiment of the first invention. Hereinafter, an embodiment of the first invention will be described with reference to FIG.

This neuron element is composed of another neuron element or a two-junction quantum interference element having an asymmetric threshold characteristic using two input lines 11 from outside, a superconducting closed loop 12, a bias input terminal 13, and a Josephson element 14. And a switching element 15, an output line 16 magnetically coupled to a superconducting closed loop of another neuron element, a reset gate 17 of the output line, and a reset signal line 18. The neuron element according to the present invention has a large number of fan-ins,
Although fan-out can be taken, in the present embodiment, the number of fan-ins and the number of fan-outs were set to 5 for simplicity.

Each of the five input lines 11 is magnetically coupled to the superconducting closed loop 12 with its own mutual inductance. Mutual inductance M ₁ , M ₂ , M ₃ , M ₄ , M
The values of ₅ are 4pH, 1pH, 5pH, 2pH,
It is 3 pH. The inductor chair of the entire superconducting closed loop 12 has a combined inductance of 3 pH with the switching element 15 and an inductance of the wiring portion of 2 pH of 20 pH. The input currents I _{1 to} I ₅ are all 0.
With 6 mA, the flow directions of I ₁ , I ₃ , I ₄ and I ₂ , I ₅ are reversed with respect to the joint with the superconducting closed loop 12. For example, when the current I ₁ flows through the input line coupled with the superconducting loop 12 with inductance M _1, the superconducting closed loop 12 in order to eliminate the magnetic field induced by _{I 1, (4/20) · 0} . A circulating current of 6 = 0.12 mA flows. In addition, when the input currents I ₂ , I ₃ , and I ₄ flow, the total circulating current Icir is Icir = 0.12-0.0.
It becomes 3 + 0.15 + 0.06 = 0.3 mA. Here the second term of the sign is negative Nanoha, because the direction of the circulating current direction of the current I ₂ is induced for a another current and reverse are opposite. Therefore, the currents I ₁ , I ₃ , I ₄ become excitatory signals, and the currents I ₂ , I ₅ become suppressive signals.

When a bias current of 0.6 mA is supplied from the bias input terminal 13, the branch on the output line 16 side has a large inductance, so that almost all the bias current flows to the branch including the switching element 15.
The switching element 15 is the Josephson element 14
Is a two-junction quantum interferometer having an asymmetrical threshold characteristic with a critical current value of 0.40 mA and an inductance of 1.7 pH. When a bias current of the above value is applied to the switching element 15, the superconducting closed loop 1
2, the circulating current Icir is 0.20 m in the direction shown in FIG.
When more than A flows, the switching element switches to the voltage state, and the bias current of 0.6 mA flows to the branch of the output line 16. As in the above example, I ₁ , I ₂ , I ₃ , I ₄
, The circulating current becomes 0.3 mA, and the output current appears on the output line 16. For example, when the input is I ₁ , I ₂ ,
With I ₃ and I ₅ , the circulating current is 0.15 mA and no output current appears. That is, whether or not a current appears on the output line 16 is determined by the combination of the flowing input currents.

The current flowing through the output line 16 becomes the input current of another neuron element. The current value flowing through the output line 16 does not depend on the inductance value of the output line 16, and since the output line 16 is a superconducting wire, the current is not attenuated. There is no inherent limit to the number of outs. Also, since the input only magnetically couples to the superconducting closed loop 12, there is no substantial limitation on the fan-in number.

When a current flows through the output line 16, this current continues to flow in the superconducting closed loop including the path including the output line 16 and the switching element 15 even after the bias current falls. In order to reset this current, a current of 0.3 mA is passed through the reset signal line 18 to reset the reset gate 17 to the voltage state. The reset gate 17 is a two-junction quantum interferometer in which the Josephson device has a critical current value of 0.4 mA and an inductance of 1.7 pH.
On the other hand, the circulating current flowing in the superconducting closed loop 12 has a property of shielding the magnetic field generated by the input current, and therefore increases or decreases according to the input current.

As described above, by using the circuit described in this embodiment, a neuron element which performs synaptic operation and neuron operation can be constructed by using the unique property of the superconductor. Moreover, since the number of fan-ins and fan-outs of the present neuron element is not essentially limited, it is possible to form a sophisticated neural network in which neuron elements are connected to each other in a complicated manner. In addition, the circuit configuration is simpler and the number of parts is smaller than that of the semiconductor neuron element, so that high-density integration is possible.

In this embodiment, the switching element 15
A two-junction quantum interferometer having an asymmetrical threshold characteristic was used as the reason for this because the gain is large and the switching is prevented by the circulating current in the opposite direction. However, switching elements including other Josephson elements can also be used depending on the case.

(Embodiment 2) FIG. 2 is a diagram for explaining an embodiment of the second invention. An embodiment of the second invention will be described below with reference to FIG.

The present neuron element has an input gate 22 which is switched by a current flowing through an output line 21 of another neuron element, an input line 11 into which a current from an input current supply terminal 23 flows by a switch of the input gate 22, a superconducting closed loop 12, Bias input terminal 13, Josephson element 1
A switching element 15 composed of a two-junction quantum interference device having an asymmetrical threshold characteristic using two 4's, an output line 16 magnetically coupled to a superconducting closed loop of another neuron device, a reset gate 17 of the output line, a reset It is composed of the signal line 18. The neuron element according to the present invention can take a large number of fan-ins and fan-outs, but in the present embodiment, the number of fan-ins and fan-outs is 5 for simplicity.

The input gate 22 is a two-junction quantum interferometer in which the Josephson device has a critical current value of 0.41 mA and an inductance of 1.0 pH, and a bias current is supplied from the input current supply end 23 to the output of other neuron devices. It is switched by the current flowing through the line 21. Therefore, by controlling the current value supplied from the input current supply terminal 23, the output line 1
The current flowing through 1 can be controlled. Therefore, the synapse operation for weighting the input from the other neuron element can be performed by the current value supplied from the input current supply terminal 23.

All five input lines 11 are magnetically coupled to the superconducting closed loop 12 with an equal mutual inductance value of 3 pH. The total inductance of the superconducting closed loop 12 is 2 including the coupling inductance 3pH with the switching element 15 and the wiring inductance 2pH.
It becomes 0 pH. For example, the currents I ₁ , I ₃ , I ₄ and I ₂ , I flowing through the input line 11 by the switch of the input gate 22 are
₅ for 0.6 mA, 0.3 mA, 0.8 mA,
The currents are set to 0.4 mA and 0.7 mA, and currents flowing in the opposite directions to I ₁ , I ₃ , I ₄ and I ₂ , I ₅ flow. For example, if a current I ₁ flows through the input line coupled to the superconducting closed loop 12,
In order to eliminate the magnetic field induced by I _{1 in the} superconducting closed loop 12, (3/20) · 0.6 = 0.09 mA
Circulating current flows. In addition, input currents I ₂ , I ₃ , I ₄
, The total circulating current Icir becomes Icir = 0.09.
It becomes −0.045 + 0.12 + 0.06 = 0.225 mA. Here, the sign of the second term is negative because the current I ₂
This is because the direction of the circulating current induced is opposite because the direction of is opposite to the other currents. Therefore, the currents I ₁ , I ₃ , I ₄ become excitatory signals, and the currents I ₂ , I 4
₅ is a suppressive signal.

When a bias current of 0.6 mA is applied from the bias input terminal 13, the branch on the output line 16 side has a large inductance, so that almost all the bias current flows to the branch including the switching element 15.
The switching element 15 is the Josephson element 14
Is a two-junction quantum interferometer having an asymmetrical threshold characteristic with a critical current value of 0.40 mA and an inductance of 1.7 pH. When a bias current of the above value is applied to the switching element 15, the superconducting closed loop 12
The circulating current Icir is 0.20 mA in the direction shown in FIG.
When the above flows, the switching element switches to the voltage state, and the bias current of 0.6 mA flows to the branch of the output line 16. When I ₁ , I ₂ , I ₃ , and I ₄ are input as in the above example, the circulating current becomes 0.225 mA and the output current appears on the output line 16. For example, when the input is I ₁ ,
With I ₂ , I ₃ , and I ₅ , the circulating current is 0.06 mA, and no output current appears. That is, whether or not a current appears on the output line 16 is determined by the combination of the flowing input currents.

The current flowing through the output line 16 becomes a signal current for switching the input gate 22 of another neuron element. The current value flowing through the output line 16 does not depend on the inductance value of the output line 16, and since the output line 16 is a superconducting wire, the current is not attenuated. There is no inherent limit to the number of outs. Also, since the input only magnetically couples to the superconducting closed loop 12, there is no substantial limitation on the fan-in number.

When a current is flowing in the output line 16, this current continues to flow in the superconducting closed loop including the output line 16 and the switching element 15 and forming a path even after the bias current falls. In order to reset this current, a current of 0.3 mA is passed through the reset signal line 18 to reset the reset gate 17 to the voltage state. The reset gate 17 is a two-junction quantum interferometer in which the Josephson device has a critical current value of 0.4 mA and an inductance of 1.7 pH.
On the other hand, the circulating current flowing in the superconducting closed loop 12 has a property of shielding the magnetic field generated by the input current, and therefore increases or decreases according to the input current.

As described above, if the circuit described in this embodiment is used, a neuron element that performs synaptic operation and neuron operation can be constructed by using the property unique to superconductors. Moreover, since the number of fan-ins and fan-outs of the present neuron element is not essentially limited, it is possible to form a sophisticated neural network in which neuron elements are connected to each other in a complicated manner. Moreover, since the weight of the input is determined by the value of the supply current from the input current supply end, the weight can be made variable by controlling the value of the supply current.

In this embodiment, the switching element 15
A two-junction quantum interferometer having an asymmetrical threshold characteristic was used as the reason for this because the gain is large and the switching is prevented by the circulating current in the opposite direction. However, switching elements including other Josephson elements can also be used depending on the case. Further, although the mutual inductance values between the input line 11 and the superconducting closed loop 12 are all the same, different values can be used. In that case, the current value and the inductance are doubly weighted.

(Embodiment 3) FIGS. 3 and 4 are views for explaining a third embodiment of the present invention. An embodiment of the present invention will be described below with reference to FIGS.

The present neuron element is a two-junction quantum interference element having an asymmetric threshold characteristic using two neuron elements or an external input line 11, a superconducting closed loop 12, a bias input terminal 13, and two Josephson elements 14. , A switching element 15, an output line 16 magnetically coupled to a superconducting closed loop of another neuron element, a down edge trigger circuit 31, and a DC control line 32. The neuron element according to the present invention has a large number of fan-ins,
Although fan-out can be taken, in the present embodiment, the number of fan-ins and the number of fan-outs were set to 5 for simplicity.

The five input lines 11 are magnetically coupled to the superconducting closed loop 12 with their own mutual inductances. Mutual inductance M ₁ , M ₂ , M ₃ , M ₄ , M
The values of ₅ are 4pH, 1PH, 5pH, 2pH,
It is 3 pH. The inductance of the entire superconducting closed loop 12 is the coupling inductance 3pH with the switching element 15, the coupling inductance 3pH with the down edge trigger circuit 31, and the inductance 2pH of the wiring portion.
To become 23 pH. The input currents are all 0.
At 6 mA, I ₁ , I ₃ , I ₄ and I ₂ , I ₅ have opposite flow directions with respect to the coupling portion with the superconducting closed loop 12. For example, the superconducting closed loop 1 in inductance M ₁
When a current I ₁ flows through the input line coupled with 2, the superconducting closed loop 12 has a magnetic field caused by I ₁ to be eliminated.
A circulating current of (4/23) · 0.6 = 0.10 mA flows. In addition, when the input currents I ₂ , I ₃ , and I ₄ flow, the total circulating current Icir is Icir = 0.10-0.026 +
It becomes 0.13 + 0.052 = 0.26 mA. Here the second term of the sign is negative Nanoha, because the direction of the circulating current direction of the current I ₂ is induced for a another current and reverse are opposite. Therefore, the currents I ₁ , I ₃ , I ₄ become excitatory signals, and the currents I ₂ , I ₅ become suppressive signals.

When a bias current of 0.6 mA is applied from the bias input terminal 13, the branch on the output line 16 side has a larger inductance than the branch including the switching element 15, so that all the bias current flows toward the branch including the switching element 15. Flow to. The switching element 15 has a critical current value of the Josephson element 14 of 0.40 mA and an inductance of 1.7 p.
It is an asymmetric two-junction quantum interferometer of H. When a bias current of the above value is applied to this switching element 15,
When the circulating current Icir flows in the superconducting closed loop 12 in the direction shown in FIG. 3 by 0.2 mA or more, the switching element switches to the voltage state, and the bias current of 0.6 mA flows in the branch of the output line 16. As in the above example, I ₁ , I
_{When 2} , I ₃ and I ₄ are input, the circulating current is 0.26m
A, and an output current appears on the output line 16. For example, when the inputs are I ₁ , I ₂ , I ₃ , and I ₅ , the circulating current is 0.13.
It becomes mA and no output current appears. That is, whether or not a current appears on the output line 16 is determined by the combination of the flowing input currents.

The current flowing through the output line 16 becomes the input current of another neuron element. The current value flowing through the output line 16 does not depend on the inductance value of the output line 16, and since the output line 16 is a superconducting wire, the current is not attenuated. There is no inherent limit to the number of outs. Also, since the input only magnetically couples to the superconducting closed loop 12, there is no substantial limitation on the fan-in number.

A specific example of the down edge trigger circuit 31 of FIG. 3 is shown in FIG. FIG. 4A is a circuit diagram of the down edge trigger circuit, and FIG. 4B is a threshold characteristic of the two-junction quantum interferometer 33 which is the switching element. Regarding the parameters of the two-junction quantum interferometer 33, the critical current value of the Josephson element is 0.4 mA and the inductance is 1.7 pH. In the down-edge trigger circuit shown in FIG. 4A, the output line 16 supplies the gate current Ig to the two-junction quantum interferometer, while the conduction closed loop 12 and the DC control line 32 supply the control current Ic. There is. When the input current supplied from the input line 11 is zero, the circulating current does not flow in the superconducting closed loop 12 and the switching element 15 does not switch. Therefore, the current does not flow in the output line 16, so that the current flowing in the down edge trigger circuit. Is only 0.3 mA of DC current flowing through the DC control line 32, and the operating point is shown in FIG.
It is at the operating point 34 in (b). When the input current flows and the circulating current flowing in the superconducting closed loop 12 increases, the current flowing in the superconducting closed loop 12 is in the opposite direction to the current flowing in the DC control line 32. Therefore, the operating point in FIG. The point 34 is shifted in the minus direction on the Ic axis. When the current flowing through the superconducting closed loop 12 further increases and the switching element 15 switches, a current flows through the output line 16, so that the operating point moves to the operating point 35. Since the operating point 35 is inside the threshold curve, the two-junction quantum interferometer 33 does not switch. Here, the current flowing through the superconducting closed loop 12 was set to 0.26 mA assuming the inputs I ₁ , I ₂ , I ₃ , and I ₄ . When the input current decreases and the circulating current flowing through the closed superconducting loop 12 becomes 0.2 mA or less, the operating point appears outside the threshold curve (for example, operating point 36), so that the two-junction quantum interferometer 33 switches to the voltage state. . As a result, the output line 16
The current that has flowed into the circuit again includes the switching element 15 and flows into the branch.

Since the load is superconducting in both the switching element 15 and the down-edge trigger circuit 31, when switched, all the current flows toward the load and is automatically reset to the zero voltage state. In order to ensure this operation, a damping resistor having an appropriate value may be used.

Generally, in a superconducting integrated circuit, the Josephson element is a latch element that does not reset unless the bias current becomes a certain value or less. Therefore, it is necessary to use an alternating current for the bias current. It is done in synchronization with. Since the superconducting neuron element according to the present invention is a kind of flip-flop circuit that sets and resets according to an input current, each element can be operated asynchronously. This characteristic is very effective when constructing a neural network using neuron elements.

As described above, by using the circuit described in this embodiment, it is possible to construct a neuron element which performs synaptic operation and neuron operation by using the property unique to superconductors. Moreover, since the number of fan-ins and fan-outs of the present neuron element is not essentially limited, it is possible to form a sophisticated neural network in which a plurality of neuron elements are connected to each other. In addition, the circuit configuration is simpler and the number of parts is smaller than that of the semiconductor neuron element, so that high-density integration is possible. Furthermore, since a plurality of neuron elements can be operated asynchronously, it is advantageous in networking.

In this embodiment, the switching element 15
A two-junction quantum interferometer having an asymmetrical threshold characteristic was used as the reason for this because the gain is large and the switching is prevented by the circulating current in the opposite direction. However, switching elements including other Josephson elements can also be used depending on the case. Further, although the circulating current values switched by the switching element 15 and the down edge trigger circuit 31 are the same, different values may be used.

(Embodiment 4) FIG. 5 is a view for explaining an embodiment of the fourth invention. An embodiment of the fourth invention will be described below with reference to FIG.

This neuron element is magnetically connected to another neuron element or an external input line 11, a superconducting closed loop 12, a bias input terminal 13, a switching element 15 consisting of one Josephson element, and a superconducting closed loop of another neuron element. It is composed of a coupled output line 16, a reset gate 17 for the output line, and a reset signal line 18. The neuron element according to the present invention can take a large number of fan-ins and fan-outs, but in the present embodiment, the number of fan-ins and fan-outs is 5 for simplicity.

The five input lines 11 are magnetically coupled to the superconducting closed loop 12 with their own mutual inductances. Mutual inductance M ₁ , M ₂ , M ₃ , M ₄ , M
The values of ₅ are 4pH, 1PH, 5pH, 2pH,
It is 3 pH. Inductance of the entire superconducting closed loop 12 Inductance of wiring part 3 pH is combined to 18 pH
Becomes The input currents are all 0.6 mA, I ₁ , I
_The flow directions of ₃ , I ₄ and I ₂ , I ₅ are reversed with respect to the coupling portion with the superconducting closed loop 12. For example, when a current I ₁ flows through an input line coupled to the superconducting closed loop 12 with an inductance M ₁ , the superconducting closed loop 12 has a magnetic field induced by I ₁ (4/18)
A circulating current of 0.6 = 0.13 mA flows. In addition, when the input currents I ₂ , I ₃ , and I ₄ flow, the total circulating current Icir
Is Icir = 0.13-0.03 + 0.17 + 0.0
7 = 0.34 mA. Here the second term of the sign is negative Nanoha, because the direction of the circulating current direction of the current I ₂ is induced for a another current and reverse are opposite. Therefore, the currents I ₁ , I ₃ , I ₄ become excitatory signals, and the currents I ₂ , I ₅ become suppressive signals.

When a bias current of 0.6 mA is applied from the bias input terminal 13, the inductance of the branch on the output line 16 side and the branch on the side magnetically coupled to the input line 11 are significantly larger than the inductance of the branch including the switching element 15. Therefore, almost all the bias current flows toward the branch including the switching element 15. The switching element 15 is composed of one Josephson element having a critical current value of 0.8 mA. When a bias current of the above value is passed through the switching element 15, the circulating current Icir is passed through the superconducting closed loop 12.
When 0.20 mA or more flows in the direction shown in FIG. 5, the bias current and the circulating current are added together, and the current flowing through the switching element 15 exceeds its critical current value. Therefore, the switching element switches to the voltage state,
Bias current of 0.6 mA is output line 16 and input line 11
It flows to the branch that is magnetically coupled with. The inductance of the branch on the output line 16 side is usually larger than the inductance of the branch magnetically coupled to the input line 11. Assuming that the inductance of the output line 16 is 74 pH, a bias current of 0.12 mA flows to the output line 16 side and 0.48 mA to the side magnetically coupled to the input line 11. At this time, since the switching element 15 is switching to the voltage state, the superconducting closed loop 12 is cut off and the circulating current Icir does not flow. When a Josephson device having a critical current value of 0.4 mA is used as the output gate 51, the output gate switches with a current of 0.48 mA flowing on the side magnetically coupled to the input line 11, and all the bias currents are branched on the output line 16 side. Flow into. Thus, as in the above example, I ₁ , I ₂ , I
_{When 3} and I ₄ are input, the circulating current becomes 0.34 mA, and the output current appears on the output line 16.
_{For 1} , I ₂ , I ₃ , and I ₅ , the circulating current was 0.17 mA and no output current appeared. That is, whether or not a current appears on the output line 16 is determined by the combination of the flowing input currents.

The current flowing through the output line 16 becomes the input current of another neuron element. The current value flowing through the output line 16 does not depend on the inductance value of the output line 16, and since the output line 16 is a superconducting wire, the current is not attenuated. There is no inherent limit to the number of outs. Also, since the input only magnetically couples to the superconducting closed loop 12, there is no substantial limitation on the fan-in number.

When a current flows through the output line 16, this current continues to flow in the superconducting closed loop including the path including the output line 16 and the switching element 15 even after the bias current falls. In order to reset this current, a current of 0.3 mA is passed through the reset signal line 18 to reset the reset gate 17 to the voltage state. The reset gate 17 is a two-junction quantum interferometer, and the critical current value of the Josephson element is 0.4 mA and the inductance is 1.7 pH. On the other hand, since the load of the switching element 15 and the output gate 51 is a superconducting wire and does not include a resistance component, all the bias current flows into the output line 16 and returns to the superconducting state, so that the superconducting closed loop 12 automatically. Will be reset.

As described above, by using the circuit described in this embodiment, it is possible to construct a neuron element which performs synaptic operation and neuron operation by using the properties unique to superconductors. Moreover, since the number of fan-ins and fan-outs of the present neuron element is not essentially limited, it is possible to form a sophisticated neural network in which neuron elements are connected to each other in a complicated manner. Further, the circuit configuration is simpler and the number of parts is smaller than that of the semiconductor neuron element, so that high density integration is possible. Switching element 1
Since 5 is included in the superconducting closed loop 12, the ratio of the coupling inductance with each input line 11 to the total inductance of the superconducting closed loop 12 is large, and the conversion efficiency of the input current into the superconducting closed loop circulating current is increased. .

In this embodiment, the switching element 15
Although one Josephson element is used as the above, a switching element including another Josephson element can also be used depending on the case.

(Embodiment 5) FIG. 6 is a view for explaining an embodiment of the fifth invention. An embodiment of the fifth invention will be described below with reference to FIG.

This neuron element is magnetically connected to another neuron element or an external input line 11, a superconducting closed loop 12, a bias input terminal 13, a switching element 15 consisting of one Josephson element, and a superconducting closed loop of another neuron element. Combined output line 16 and reset resistor 61
Consists of The neuron element according to the present invention can take a large number of fan-ins and fan-outs, but in the present embodiment, the number of fan-ins and fan-outs is 5 for simplicity.

The five input lines 11 are magnetically coupled to the superconducting closed loop 12 with their own mutual inductances. Mutual inductance M ₁ , M ₂ , M ₃ , M ₄ , M
The values of ₅ are 4pH, 1PH, 5pH, 2pH,
It is 3 pH. Inductance of the entire superconducting closed loop 12 Inductance of wiring part 3 pH is combined to 18 pH
Becomes The input currents are all 0.6 mA, I ₁ , I
_The flow directions of ₃ , I ₄ and I ₂ , I ₅ are reversed with respect to the coupling portion with the superconducting closed loop 12. For example, when a current I ₁ flows through an input line coupled to the superconducting closed loop 12 with an inductance M ₁ , the superconducting closed loop 12 has a magnetic field induced by I ₁ (4/18)
A circulating current of 0.6 = 0.13 mA flows. In addition, when the input currents I ₂ , I ₃ , and I ₄ flow, the total circulating current Icir
Is Icir = 0.13-0.03 + 0.17 + 0.0
7 = 0.34 mA. Here the second term of the sign is negative Nanoha, because the direction of the circulating current direction of the current I ₂ is induced for a another current and reverse are opposite. Therefore, the currents I ₁ , I ₃ , I ₄ become excitatory signals, and the currents I ₂ , I ₅ become suppressive signals.

Bias current 0.
When a current of 64 mA is applied, the branch on the output line 16 side includes the reset resistor 61, and the inductance of the branch magnetically coupled to the input line 11 is significantly larger than the inductance on the branch including the switching element 15. Flows almost entirely to the branch containing the switching element 15. The switching element 15 has a critical current value of 0.855 mA.
It consists of Josephson elements. When a circulating current Icir of 0.20 mA or more flows in the superconducting closed loop 12 in the direction shown in FIG. 5 when a bias current of the above value is passed through the switching element 15, the bias current and the circulating current are added to each other to cause the switching element 15 to flow. The switching element 15 switches to the voltage state because the current flowing through it exceeds its critical current value. When the switching element 15 switches, a bias current of 0.6 mA flows into the branch magnetically coupled to the input line 11, except for a leakage current corresponding to about 5% of the critical current value. At this time, since the switching element 15 is switching to the voltage state, the superconducting closed loop 12 is cut off and the circulating current Icir does not flow. If a Josephson element having a critical current value of 0.4 mA is used as the output gate 51, a current of 0.6 mA flowing into the side magnetically coupled to the input line 11 is obtained.
Then, the output gate switches, and the bias current flows into the branch on the output line 16 side except for the leakage current corresponding to about 5% of the critical current value. Thus, when I ₁ , I ₂ , I ₃ , and I ₄ are input as in the above example, the circulating current becomes 0.34 mA, and the output current appears on the output line 16. For example, when the input is I ₁ , For I ₂ , I ₃ , and I ₅ , the circulating current was 0.17 mA, and no output current appeared. That is, whether or not a current appears on the output line 16 is determined by the combination of the flowing input currents.

The current flowing through the output line 16 becomes the input current of another neuron element. The current value flowing through the output line 16 does not depend on the inductance value of the output line 16, and since the output line 16 is a superconducting wire, the current is not attenuated. There is no inherent limit to the number of outs. Also, since the input only magnetically couples to the superconducting closed loop 12, there is no substantial limitation on the fan-in number.

In the present invention, since the reset resistance 61 is present in the output line 16, the current flowing through the output line 16 decreases with the time constant determined by the reset resistance 61 and the inductance 16 of the output line as the bias current falls. To do. When the value of the reset resistor 61 is 4Ω, this time constant is 18 psec. On the other hand, the switching element 15
Since the output gate 51 is reset to the superconducting state with the fall of the bias current, the superconducting closed loop 12 is automatically reset.

As described above, by using the circuit described in this embodiment, it is possible to construct a neuron element which performs synaptic operation and neuron operation by using the property unique to superconductors. Moreover, since the number of fan-ins and fan-outs of the present neuron element is not essentially limited, it is possible to form a sophisticated neural network in which neuron elements are connected to each other in a complicated manner. In addition, the circuit configuration is simpler and the number of parts is smaller than that of the semiconductor neuron element, so that high-density integration is possible. Switching element 1
Since 5 is included in the superconducting closed loop 12, the ratio of the coupling inductance with each input line 11 in the total inductance of the superconducting closed loop 12 is large, and the exchange efficiency of the input current to the superconducting closed loop circulating current is increased. . In addition, since the reset resistor 61 exists, more stable operation of the circuit can be expected as compared with the case where there is no resistance component.

In this embodiment, the switching element 15
Although one Josephson element is used as the above, a switching element including another Josephson element can also be used depending on the case.

[0064]

According to the first aspect of the present invention, a neuron element having a large number of fan-ins and fan-outs can be constructed by using a superconducting integrated circuit which consumes less power and is relatively easy to be three-dimensional. it can. Further, the circuit of the present invention has an effect that the circuit configuration is simpler than the neuron element using the semiconductor, the number of parts is small, and high density integration is possible.

The use of the second invention has the effect that the weighting of the input can be made variable in addition to the effect of the first invention. In addition to the effect of the first invention, the third invention allows the plurality of neuron elements to operate asynchronously, which is advantageous in networking. By using the fourth invention, in addition to the effect of the first invention, the efficiency of conversion of the input current into the superconducting closed loop circulating current can be increased, and further, the circuit configuration becomes simpler. By using the fifth invention, in addition to the effect of the first invention, there is an effect that more stable operation can be expected and the design becomes easy.

[Brief description of the drawings]

FIG. 1 is a circuit diagram for explaining a first embodiment of the present invention.

FIG. 2 is a circuit diagram for explaining a second embodiment of the present invention.

FIG. 3 is a circuit diagram for explaining a third embodiment of the present invention.

FIG. 4 is a diagram for explaining a specific example of a down edge trigger circuit used in the third embodiment, (a) being a circuit diagram,
(B) is a threshold characteristic diagram.

FIG. 5 is a circuit diagram for explaining a fourth embodiment of the present invention.

FIG. 6 is a circuit diagram for explaining a fifth embodiment of the present invention.

FIG. 7 is a diagram showing a basic element structure of a neuron element using a conventional Josephson line.

[Explanation of symbols]

 11 Input Line 12 Superconducting Closed Loop 13 Bias Input Terminal 14 Josephson Element 15 Switching Element 16 Output Line 17 Reset Gate 18 Reset Signal Line 21 Output Line of Other Neuron Element 22 Input Gate 23 Input Current Supply Terminal 31 Down Edge Trigger Circuit 32 DC Control line 33 2-junction quantum interferometer 34-36 Operating point 51 Output gate 61 Reset resistance 101 Josephson line A 102 Josephson line B 103 Josephson line C 104 Josephson line D 105 Magnetic flux quantum stop 106 Input end 107 Output end 108 Excitatory Control Input Terminal 109 Suppressive Control Input Terminal 111 Lower Electrode 112 Tunnel Barrier 113 Upper Electrode 114 Inductance 115 Resistance

Claims (5)

(57) [Claims] 1. A switching element using a Josephson element, a superconducting closed loop, and an output line magnetically coupled to another neuron element, wherein the superconducting closed loop and one or more input lines are magnetically coupled. The weighting of the input is performed by coupling and controlling the inductance value of the coupling, and the positive and negative signs of the weight are determined by the direction of the current flowing through the input line, and the switching element and the output line are connected to the bias input terminal. Connected in parallel, and the superconducting closed loop and the switch of the switching element are magnetically coupled to each other. 2. A switching element using a Josephson element, a superconducting closed loop, and an output line magnetically coupled to another neuron element, wherein the superconducting closed loop and one or more input lines are magnetically coupled. The weighting of the input is performed by controlling the value of the current flowing through the input line in combination, and the sign of the positive or negative of the weight is determined by the direction of the current flowing through the input line, and the switching element and the output line are bias input terminals. Connected in parallel to
A superconducting neuron element, wherein the switching element switches when a circulating current flowing in the superconducting closed loop exceeds a certain threshold value. 3. A switching element using a Josephson element, a superconducting closed loop, and an output line made of a superconductor magnetically coupled to another neuron element, the superconducting closed loop and one or a plurality of input lines. Are magnetically coupled, and when the circulating current flowing in the superconducting closed loop exceeds a certain threshold value, the switching element switches, the switching element is driven by DC bias, and the switching element and the output line are A superconducting neuron device including a down-edge trigger circuit connected in parallel to a bias input terminal and switching when a circulating current flowing in the superconducting closed loop falls below a threshold value in the output line. . 4. A switching element using a Josephson element, a superconducting closed loop, and an output line magnetically coupled to another neuron element, wherein the superconducting closed loop and one or more input lines are magnetically coupled. A superconducting neuron device coupled to the switching element and the output line connected in parallel to a bias input terminal, wherein the switching element is included in the superconducting closed loop. 5. A switching element using a Josephson element, a superconducting closed loop, and an output line magnetically coupled to another neuron element, wherein the superconducting closed loop and one or more input lines are magnetically coupled. The switching element is coupled and the switching element switches when the circulating current flowing in the superconducting closed loop exceeds a certain threshold value, and the switching element and the output line are connected in parallel to the bias input terminal, and the output line. A superconducting neuron element having a resistance component inside.