JPH027582A - Semiconductor device - Google Patents

Abstract

PURPOSE:To construct a pseudo-neural network having a high degree function such as learning by controlling any magnetic flux quantum entering a Josephson line by coupling a magnetic field or a permanent current produced by the magnetic field formed by magnetic flux quanta in a magnetic flux quantum storage section with the Josephson line. CONSTITUTION:A magnetic flux quantum storage section VR is formed by a superconducting loop, to which section Josephson junctions J1, J2 are connected as an input terminal thereto. A Josephson line JT comprises a Josephson line extending in one direction. The magnetic flux quantum storage section VR is directly coupled to the Josephson line JT, i.e., connected to the upper electrode of the Josephson line JT at two positions so as to form a superconducting loop. An input signal from the previous stage is introduced as a bias signal (bias current) for the Josephson line JT, and an output signal is led to a load composed of a resistor RL and an inductance LL via a terminal OUT. Herein, the magnetic flux quantum storage section VR is strongly coupled only to the left end of the Josephson line JT to ease entrance of any magnetic flux quantum in this direction.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to a device that utilizes the quantization phenomenon of magnetic flux occurring in a superconductor or a Josephson element.
This device has the function of obtaining a variable output signal according to the state of accumulated magnetic flux quanta, and is particularly useful for constructing a pseudo-neural circuit device that can perform frequent functions with neural circuits. .

[Conventional technology]

Neural networks in biological systems, including humans, achieve advanced functions through distributed parallel processing. By imitating this, it is possible to create information processing devices that cannot be realized with conventional Neumann-type computers, so it has attracted a lot of attention in recent years. On the other hand, it is known that magnetic flux quanta propagating in Josephson lines are similar to nerve pulses propagating in biological systems. In other words, when an input current corresponding to a stimulus is applied to the input end of a Josephson line, magnetic flux quanta propagate when the stimulus reaches a certain level, that is, exceeds a certain current level, and the number of incoming pulses increases according to the current level. . This means that there is a fixed threshold for the stimulus level, and the degree of excitation changes depending on the size of the stimulus, and a similar behavior has been observed in neural circuits. However, these functions are insufficient for performing advanced processing. For example, when considering a function such as learning, it is not enough to simply change the excitation level according to the intensity of input from the previous stage, and plasticity is required in the syneraptic connections between neurons.

In other words, it is necessary to change the inter-element coupling strength corresponding to the synalapse connections of neurons for a large number of input lines (even if the input signal has a certain level, the degree of stimulation changes depending on the type of input line). .

[Problem to be solved by the invention]

However, although it has been pointed out that there is a similarity between Josephson lines and neural networks that can propagate magnetic flux quanta, no attempt has been made to apply this to neural networks. There are no proposals for devices with such functions.

On the other hand, in the semiconductor field, there are attempts to realize this type of device using MNOS, but the element structure and operating principle are different.
Therefore, since the effects obtained are different, they cannot be compared.

The present invention was made in view of these points, and its purpose is to enable advanced functional operations such as learning in a superconducting device consisting of Josephson lines that have properties similar to neural circuits. The aim is to provide a highly sensitive neural network by enabling the plasticity of syneraptic connections between neurons that is required to achieve this goal.

[Means to solve the problem]

In order to solve such problems, the superconducting device according to the present invention has a magnetic flux quantum storage section that can store a plurality of magnetic flux quanta and a Josephson line that allows the propagation of magnetic flux quanta. The magnetic flux quanta entering the Josephson line is controlled by coupling the magnetic field created by the magnetic flux quanta or the persistent current generated by this magnetic field to the Josephson line.

[Effect]

In the superconducting device according to the present invention, it is possible to variably control the weight of the coupling corresponding to the synalapse coupling in the form of the number of accumulated magnetic flux quanta.

〔Example〕

First, the outline of the superconducting device according to the present invention will be explained.

This device includes a magnetic flux quantum storage section capable of storing magnetic flux quanta, and a so-called Josephson line (for example, a Josephson junction having a long structure in one direction). and the magnetic flux quantum storage unit and the Josephson line are connected to the magnetic flux quantum accumulated in the magnetic flux quantum storage unit (hereinafter referred to as “accumulated magnetic flux quantum”).
Due to the effect of the accompanying magnetic field or persistent current, a current approximately proportional to the number of accumulated magnetic flux quanta is induced in the Josephson element, which is the penetration end of the magnetic flux quanta propagating in the Josephson line (hereinafter referred to as "propagating magnetic flux quanta"). Combine as shown. Based on this structure, by appropriately selecting the input signal and line bias conditions, it is possible to propagate propagation flux quanta in the Josephson line that is approximately proportional to the input signal and the number of accumulated flux quanta. This makes it possible to extract a nearly proportional output signal.

By connecting this device between elements that perform threshold operation and changing the number of accumulated magnetic flux quanta according to external signals such as learning signals, it is possible to achieve plastic synalapse coupling in neural circuits, which is difficult with conventional circuits. It becomes possible to realize similar functions.

FIG. 1 is a block diagram showing an embodiment of a superconducting device according to the present invention. In the same figure, VR is a magnetic flux quantum storage unit, J
T is a Josephson line, IN is a signal input terminal, OUT is a signal output terminal, RL and LL are load resistance and load inductance, respectively, and IVl and IV2 are input terminals for introducing the magnetic flux quantum stored in the flux quantum storage unit VR. ,
Magnetic flux quanta of opposite polarity are introduced from the input terminals IVI and IV2. Further, Jl and J2 are introduction paths for introducing magnetic flux quanta into the magnetic flux quantum storage portion VR, and are composed of Josephson junctions.

In this embodiment, the magnetic flux quantum storage section VR is formed by a superconducting loop, and the Josephson junction J1. J2 is connected.

On the other hand, the Josephson line JT consists of a Josephson line that is long in one direction. The magnetic flux quantum storage unit VR is directly connected to the Josephson line JT, and two superconductors are connected to the upper electrode of the Josephson line JT to form a loop.
It has a structure in which parts are connected. The input signal from the previous stage is introduced as a bias signal (bias current) to the Josephson line JT, and the output signal is guided to a load consisting of a resistor RL and an inductance LL via a terminal OUT. The magnetic flux quantum storage unit VR is strongly coupled only to the left end portion (the side near the input terminal IN) of the Josephson line JT, and has a structure in which magnetic flux quanta in this direction can easily enter.

The device of FIG. 1 can be represented by the equivalent circuit shown in FIG. As shown in FIG. 2, the Josephson line JT is represented by a ladder-like circuit of an inductance and a Josephson junction J3.

FIG. 3 is a diagram for explaining the operating principle of this device,
A part of the magnetic flux quantum equivalent circuit of FIG. 2 is taken out to show the state of the current induced in the Josephson line JT by the persistent current 2 generated by the magnetic flux quantum 1 of the magnetic flux quantum storage unit VR.

FIG. 4 is a graph showing the bias current versus voltage (generated across the junction) characteristics of the Josephson line. The change in the flow voltage peak SA in the figure is the magnetic flux quantum storage part VR.
This shows the situation when the current flowing due to the magnetic flux quantum is taken as a parameter.

Next, the principle of operation of the apparatus shown in FIG. 1 will be explained. Now, although this is not particularly limited here, it is possible to use an appropriate means, for example, to make the junctions Jl and J2 Josephson lines similar to the line JT. Alternatively, magnetic flux quanta are introduced into the magnetic flux quantum storage portion VR by switching the junction J1 or J2 by a magnetic field, an injection signal, heat, etc. while applying a magnetic field. When a magnetic flux quantum is introduced into the magnetic flux quantum storage unit VR, the associated magnetic field interlinks within the loop, and as a result, a persistent current flows through the loop. As shown in FIG. 3, a portion of this persistent current 2 leaks into the Josephson line JT and induces a current at the junction at the end of the Josephson line JT. The magnitude of this current is proportional to the number of accumulated magnetic flux quanta, and the direction of flow is the same or opposite to the bias signal BA corresponding to the input signal of the Josephson line J'r, depending on the polarity of the magnetic flux quanta. becomes.

When the input signal (bias signal) and the current created by the accumulated flux quantum flow to the Josephson junction at the end of the line with the same polarity, this current value is a certain value (the spread of the current at the end, i.e., the Josephson penetration distance). When the product of the line inductance at section 23 and the above-mentioned current value exceeds 1 magnetic flux quantum unit, the magnetic flux is quantized and enters the line. on the other hand,
For magnetic flux quanta with opposite polarity, a current flows in the opposite direction to the bias, so magnetic flux quanta do not enter (see Figure 3 (b)). Even if they do, they are pushed out by the bias current, so the Josephson It does not propagate within the railway line.

Magnetic flux quantum (
The transmitted S magnetic flux quantum) is subjected to the Lorentz force by the bias current and propagates toward the other end of the line.

- Once the propagating magnetic flux quantum enters, the current flowing at the end of the line decreases, and as the magnetic flux quantum enters the line, it gradually recovers. 0 The speed of recovery is determined by the above input and accumulated magnetic flux quantum varies depending on the current level associated with the current. In other words, if the current level is low, it will take a long time;
If the level is high, it will recover quickly. Therefore, the number of propagating magnetic flux quanta propagating within the line during unit time increases in accordance with the input current and the number of accumulated magnetic flux quanta (almost proportionally).

In this operation, as a result of energy consumption due to the movement of propagating magnetic flux quanta, a flow voltage as shown in FIG. 4 is generated between the bonding terminals (between the bonding upper electrode and the ground plane). In FIG. 4, SA is a characteristic line indicating the peak of the flow voltage, the parameter is the applied magnetic field, and the characteristic line on the right side has a large number of propagating magnetic flux quanta. As shown in the figure, as the level of the signal from the accumulated magnetic flux quanta increases and the number of propagating magnetic flux quanta increases, energy consumption increases, and the gear tube voltage v
The flow voltage increases almost proportionally until it approaches 0. Accordingly, a proportional current signal is obtained for the load.

In the above explanation, the output load resistance RL is set to a relatively large value and the flow voltage is taken out. However, when the load resistance is made small, it is also possible to take out the magnetic flux quantum itself as an output signal, and in this case as well, the accumulation It is possible to extract an output signal (current) approximately proportional to the number of magnetic flux quanta.

The number of characteristic variable points (dynamic range) of this device depends on how the load line RLL is taken, but in principle it is determined by the number of magnetic flux quanta that can be interlinked in the loop. The upper limit of the number of magnetic flux quanta accumulated in the loop is determined by the LI product of the loop inductance and the critical current value of the junction that goes in and out. In order to obtain a large dynamic range, it is sufficient to increase the Ll product.

FIG. 5 shows an example of using the device shown in FIG. 1 as a neural network, in which TL, TL,
is a threshold element, E i j is a joint control signal source to which weight determination information is input, and Wij is the present device. TL is not particularly limited here, but as long as the output is zero until the sum of the input signals reaches a certain threshold level, and when it exceeds the threshold level, it outputs a certain saturated output. It can be constructed of any type, for example a Josephson line element as shown in FIG. 1, or a Josephson latching gate. Note that Fj indicates fan-out to the subsequent stage.

Combined control signal source E! j has different functions depending on the purpose, that is, how the neural circuit is configured, and the combinations are performed depending on the output signal of each threshold element, external learning signal, error signal, etc. A signal source for determining the weights and thereby varying the number of stored flux quanta of the device. Since the form of the signal from the coupled control signal source E i J differs depending on the means of inputting the accumulated magnetic flux quanta to the storage section, the present invention is not particularly limited thereto (for example, the form of the signal from the above-mentioned storage section of the accumulated magnetic flux quanta To explain the input means exemplified, when using propagating magnetic flux quanta, it is possible to do this by propagating and entering only the increase and decrease of the weight through the Josephson junction J1 or J2.Magnetic field or injection signal In this case, an output signal corresponding to the desired weight can be applied magnetically or directly to the storage loop, and flux quanta can be introduced via the junction).

The connection between the threshold element TL and the coupled control signal source E i j
For simplicity, in the embodiment of the present invention, the outputs from a plurality of Wijs are condensed and introduced in the form of a tonted-in, but it is also possible to connect them by magnetic coupling, and the threshold element It also depends on what is used as the TL,
Although various methods are possible, the details are omitted because they are not related to the gist of the present invention.

As is clear from the operation of each element described above, the operation of this circuit is as follows for the threshold element corresponding to each neuron: When ΣW5j'r, -y5>Ui, T, = 1ΣW
When ijTj yi<LJi, a threshold operation of T5=0 is performed, and the weight Wij is made variable by changing the number of accumulated magnetic flux quanta.

However, Wij is a weight almost proportional to the number of accumulated magnetic flux quantum of the ijth device, and the output current of this device Wij is normalized by the output current of the threshold element, and Ut &
y is the switching threshold of the i-th threshold element, normalized by the output current, T1 is the output signal of the i-th threshold element, normalized to 1, and Y is the operating point adjustment. Alternatively, it is a standardized input signal for threshold level adjustment. The above equation is a basic equation common to various types of neural circuits with plasticity that have been proposed so far. In other words, this shows that a neural network with advanced functionality can be realized by combining the present device and a threshold element.

In the above embodiment, the magnetic flux quantum storage section VR and the Josephson line JT are connected to each other, but the superconducting loop of the magnetic flux quantum storage section VR and the Josephson line JT do not need to be directly electrically connected. Alternatively, as shown in FIG. 6 as a second embodiment of the present invention, a structure may be adopted in which the superconducting loop of the magnetic flux quantum storage section VR forms the control line of the Josephson line JT. In this case, the magnetic field created by the accumulated magnetic flux quanta is linked to the junction by magnetic coupling. It is obvious that a current concentrated at the end of the Josephson line JT can be induced in the Josephson line JT by the linkage of magnetic fields, as in FIG. 1, and the basic operations are equivalent. In this case, electrically, the magnetic flux quantum storage section VR and the Josephson line JT are independent, so the magnetic flux quantum storage section VR can be biased independently to input and output stored magnetic flux quanta, which increases the degree of freedom in circuit design. It has the characteristic of being large.

FIG. 7 shows a third embodiment of the present invention. In this embodiment, an ultra-thin film superconductor is used as the magnetic flux quantum storage part VR, and the upper electrode of the Josephson line JT is arranged above it so as to be magnetically coupled. The magnetic flux quantum storage unit VR has a structure in which a path for introducing magnetic flux quanta and a region for accumulating magnetic flux quanta are made up of an extremely thin film region A, and a thick film portion 4 is arranged around the extremely thin film region A to prevent the spread of magnetic flux quanta. In this case, magnetic flux quanta are accumulated not within the loop but through the superconductor, the resulting magnetic field is detected by the superconducting loop, and the detected current is used to drive the Josephson line JT.

In this case, the accumulated flux quantum becomes an Abrikosov flux quantum. Abrikosov magnetic flux quanta can be easily generated by interlinking a magnetic field or passing a current through the edge of the ultra-thin film A, and by adding structural improvements, it is possible to create efficient generation techniques and to place them in specific locations. A guiding technique has already been proposed, and by this method it is possible to guide magnetic flux quanta to the magnetic flux quantum storage unit VR and store a desired number of magnetic flux quanta. Once a certain amount of magnetic flux quanta has been accumulated, if magnetic flux quanta with opposite polarities are introduced from the opposite direction, they will annihilate each other, thereby making it possible to reduce the number of magnetic flux quanta.

That is, as in the embodiment shown in FIG. 1, it is possible to increase or decrease the magnetic flux quantum of the magnetic flux quantum storage section VR. Since a superconducting loop is coupled to the magnetic flux quantum storage unit VR, when magnetic flux quanta are introduced into the magnetic flux quantum storage unit R, a superconducting current corresponding to the number is excited in the loop (Meisner effect). by). This has the same effect as when magnetic flux is introduced into the loop in FIG. 1. Therefore, it is obvious that the embodiment of FIG. 7 also performs the same operation as that of FIG. 1.

The stored Abrikosov flux quanta have very small dimensions, allowing a large amount of flux quanta to be stored in a small area, and there is no Ll product limitation as shown in the apparatus of FIG.

Therefore, the embodiment shown in FIG. 7 is characterized in that the number of variable points can be extremely large compared to the device shown in FIG.

In the above embodiment, the input signal is the Josephson line JT.
However, this is not necessarily an essential condition, and the input signal is only applied to the input end of the Josephson line JT, and the bias signal is applied as a bias signal for accelerated propagation in the Josephson line JT. Similar operational effects can be obtained by applying a structure in which signals are applied simultaneously or by magnetically coupling the input signals themselves.

〔Effect of the invention〕

As explained above, the superconducting device according to the present invention has a magnetic flux quantum storage section that can store a plurality of magnetic flux quanta and a Josephson line that allows the propagation of magnetic flux quanta, and has By coupling the generated magnetic field or the persistent current generated by this magnetic field to the Josephson line to control the magnetic flux quanta that enter the Josephson line, the weight of the coupling corresponding to the Synasbus coupling can be changed to the number of accumulated flux quanta. This has the great advantage of making it possible to construct a pseudo-neural network with advanced functions such as learning, which was previously difficult.

[Brief explanation of the drawing]

Fig. 1 is a block diagram showing a first embodiment of a superconducting device according to the present invention, Fig. 2 is an equivalent circuit diagram thereof, Fig. 3 is a circuit diagram for explaining its operation, and Fig. 4 is a Josephson A graph showing the voltage-current characteristics of the line, FIG. 5 shows the basic structure of the neural circuit realized using the device shown in FIG. 1, and FIGS. 6 and 7 show the second and third structures of the superconducting device according to the present invention. FIG. 2 is a configuration diagram showing an example. VR...Magnetic flux quantum storage unit, JT...Josephson line, IN, IVI, IV2...Input terminal, OUT
...output terminal, RL...load resistance, LL...load inductance, Jl, J2. J3...Josephson junction.

Claims (1)

[Claims] It has a magnetic flux quantum storage section that can store a plurality of magnetic flux quanta and a Josephson line that allows the propagation of magnetic flux quanta, and the magnetic field created by the magnetic flux quanta of the magnetic flux quantum storage section or the persistent current generated by this magnetic field is transmitted to the Josephson line. A superconducting device characterized in that magnetic flux quanta entering the Josephson line are controlled by coupling to the Josephson line.