JPH04130592A - Threshold logic circuit - Google Patents

Abstract

PURPOSE:To obtain a neuron element with a high learning efficiency capable of using an algorithm such as a back propagation method by averaging the plural times of output signals of Josephson elements under the same input signal condition, and constituting a threshold circuit in which a signal switching characteristic is differentiable. CONSTITUTION:When a signal current including a noise component is impressed to a quantum interference type Josephson element 100 with a signal switching characteristics indicating a step characteristic, the signal switching characteristic by the output signal average in an averaging circuit 200 turns to a differentiable continuous function. Therefore, using the output of the averaging circuit 200, a learning rule such as the back propagation method can be used. In this case, the average value output of a threshold circuit 500 is compared with an educator signal inputted through an educator signal terminal 302 in a weight control circuit 300, then the weight of a weight circuit 150 is updated through a weight control line 310 based on a compared result. Thus, it is possible to constitute a high speed threshold logic circuit enable to use the learning method with the good learning efficiency and convergency by a high speed Josephson switching circuit.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Application of the Invention) The present invention relates to a circuit configuration of a neuron element using a superconducting circuit, particularly a threshold circuit using a Josephson element.

(Background of the Invention) Conventional computers are constructed with a logic circuit system that combines AND or OR circuits. It is a well-known fact that these computers operate at extremely high speeds, exhibit performance that far exceeds human computing ability, and contribute to society.

However, it has become increasingly clear that conventional computers are unsuitable for the recognition and judgment operations that humans perform on a daily basis. For this reason, for the purpose of constructing computers suitable for recognition and judgment, threshold logic circuits modeled after human brain cells (neurons) and computer system technology using them have been developed. Shun - "Mathematics of Neural Networks" Sangyo Tosho, 1978, L, D, Jackl
el, R, E, Hotward, H, P, Gr
af, B.

Straughn, and J.D., Denker.
, 'Artificial neural network
ks for computing, Journ
al of Vacuum Science Techn.
ology B4(1), Jan/Feb, 1
986. pp.

61-63.

The operation of a threshold logic circuit according to the prior art will be explained below to clarify the positioning of the present invention.

FIG. 7 is a diagram showing the operation of the threshold logic circuit.

The threshold logic circuit has a plurality of input terminals 152) input line 1
51 and at least one output line 104) output terminal 105
, a weighting circuit 150, and a threshold circuit 500. A plurality of input signals Xi inputted via an input line 151 are weighted by a weighting circuit 150, and then sent to a threshold circuit 5.
00 and is compared with a predetermined threshold T. The threshold circuit 500 sends the output signal F to the output line 104 based on the comparison result.
Output to. In the threshold logic circuit, a digital signal Xi of "0" or "1" is applied to a plurality of input terminals,
If the weighted sum ΣW i X i of the digital signal Xi exceeds the threshold T, the output is “1”, otherwise it is “0”.
Here, Wi represents the weight. Therefore, the operation of the threshold logic circuit shown in FIG. 4, that is, the relationship between the input signal Xi and the output signal F, is expressed by equation (1).

F=OΣW i X i < T F=1 ΣW i X i ≧T ・ ・ ・
(1) In the threshold logic circuit, the weight Wi is changed through learning, and a circuit system that is finally adapted to the purpose is constructed. Therefore, in order to configure a threshold logic circuit, it is necessary to have a function of changing the weight Wi and a function of performing weight addition of input signals and switching elements. Normally, this weight is controlled by a weight control signal input from a weight control line 310.

A Josephson device is a switching device that operates at high speed with low power consumption. Since the Josephson element uses magnetic flux or current as a signal carrier, it is easy to add signal values, and since the Josephson element itself has characteristics as a threshold circuit, a neuron element can be used in the threshold logic circuit. ). An attempt to realize a threshold logic circuit using a Josephson element is made, for example, in Japanese Patent Application No. 01-196115 Harada "Superconducting Threshold Logic Circuit J.
has been disclosed. A threshold circuit 500 of this prior art threshold logic circuit has the structure shown in FIG. 2a.

The circuit of FIG. 2a consists of a quantum interference type Josephson device 100
The configuration is such that a bias current Ig is supplied from a current source 106 to a parallel connection of a load resistor 103 and a load resistor 103. Quantum interference circuits are well-known devices in the field of Josephson devices, such as those described by T. R. Gheewala, “Josephson
n-Logic Devices and Circuits
t", IEEE Trans, Electron
Devices, vol.

HD-27, no, 10. pp, 1857=1
869 is described in detail. A weighted sum signal current Ic of input signals supplied from the weighting circuit 150 flows through the control current 101 of the quantum interference type Josephson element 100,
The critical current (maximum superconducting current) of the quantum interference Josephson device is controlled. FIG. 2b shows the voltage-current characteristics of the quantum interference type Josephson device 100, the load line of the load resistor 103, and the operating point of the threshold circuit 500. If the weighted sum signal current Ic of the input signals is smaller than a predetermined threshold, the quantum interference type Josephson element 100
is in a superconducting state and reaches the operating point A, and when the current is larger than the threshold current, it transitions from the superconducting state A to the voltage state. At this time,
The output signal appearing at the output terminal 105 of the threshold circuit 500 is zero volts at the operating point A, which is the "0" level, and at the operating point B, a finite voltage value (approximately 1 mV) appears, which is "1".
'' level. FIG. 3 shows the relationship between the weighted sum signal current 1c of the input signal and the output signal of the threshold circuit 500, that is, the so-called signal switching characteristic.
In the threshold circuit shown in the figure, since the superconducting state A transitions to the voltage state B in steps, the signal switching characteristics also exhibit the step characteristics shown in FIG.

Pack propagation method is used as a learning method for neuron elements. This pack propagation method is
For example, it is disclosed in Kaoru Nakano, "Fundamentals of Neuron Computers," Corona Publishing (April 1990), and others. In this pack propagation method, in order to increase the learning efficiency of the element and improve the convergence, the signal switching characteristic is made into a differentiable function as shown in Fig. 4, and the differential coefficient of the continuous signal switching characteristic, the output signal, and the teacher signal are used. Update the weights from the error.

On the other hand, the signal switching characteristic of the conventional threshold circuit shown in FIG. 3 is a stepwise characteristic, and the differential coefficient cannot be defined or is continuously negative. For this reason, the threshold logic circuit using the threshold circuit 500 shown in FIG. 2a has the drawback that learning by the pack propagation method cannot be performed as it is. In general, it is clear that in order to improve convergence, it is desirable to adopt a weight updating method that takes into account the differential coefficient of signal switching characteristics.

(Object of the Invention) An object of the present invention is to provide a threshold circuit whose signal switching characteristics can be differentiated, and to realize a neuron element with high learning efficiency that can use algorithms such as the pack propagation method.

(Summary of the Invention) For this purpose, the present invention provides a method of averaging the output signals of a Josephson element multiple times under the same human input signal condition, and using the average output signal to define the signal switching characteristics of the element. Adopt.

(Embodiments of the Invention) A threshold circuit according to the present invention will be described below using embodiments.

FIG. 1 shows an embodiment of a threshold circuit according to the invention. 1st
The embodiment shown in FIG.
This is a configuration with 01 added. The averaging circuit input line 202
is connected to the output line 104 of the threshold logic circuit. The averaging circuit 200 receives and takes the output signal of the quantum interference type Josephson element 100 from the averaging circuit input line 202, averages it, and outputs the value to the average value output line 201.

Control current line 10 of the quantum interference type Josephson device 100
1, a weighted summation signal current Ic of the input signal flows,
Generally, noise is superimposed on this signal. Furthermore, it is also possible to intentionally superimpose noise. The quantum interference type Josephson device 10 has a signal switching characteristic exhibiting the stepwise characteristic shown in FIG.
When a signal current containing a noise component at zero is applied, the output signal of the quantum interference type Josephson device 100 exhibits a value of "0" or "1" depending on the amplitude probability of the noise signal. If the average value of the superimposed noise is zero, the averaging circuit 200
When the weighted summation current Ic of the input signal is equal to the threshold value T, the average value signal becomes an intermediate value between the "0" level and the "1" level, and the further away from the threshold value T, the "0" level or It shows the characteristic of asymptotic approach to "1" level. That is, the signal switching characteristic based on the output signal averaged by the averaging circuit 200 becomes a differentiable continuous function shown in FIG.

Therefore, by using the output signal of the averaging circuit 200, learning rules such as the pack propagation method can be used. In addition, although the average value is used for learning, the output signal is directly transmitted to the quantum interference type Josephson element 10 for logical operations such as recognition.
It is possible to take out the signal from the zero output terminal 105 without averaging, or to take out the averaged signal from the average value output line 201.

FIG. 5 shows a threshold circuit 500 shown in FIG. 1 according to the present invention.
The configuration of a threshold logic circuit and learning system using . In the configuration of FIG. 5, the threshold logic circuit is the weight circuit 15.
0 and a threshold circuit 500 shown in FIG. 1 according to the present invention. The weighting circuit 150 is, for example, disclosed in Japanese Patent Application No. 01-1
96115 Harada "Superconducting Threshold Logic Circuit".

The average value output of the threshold circuit 500 is the weight control circuit 300.
The teacher signal terminal 302) is compared with the teacher signal input via the teacher signal line 301, and the weight is updated via the weight control line 310 based on the comparison result.

The averaging circuit can be implemented with an integrating circuit or a counter circuit. The quantum interference type Josephson device 100 is “
It outputs a "1" signal with a 0" difference, but if you count the number of "1" signals during a fixed number of trials, it is clear that the value corresponds to the average value. Therefore, the averaging circuit 20
0 can also be configured with an analog integration circuit consisting of a resistor and capacitance, or a digital circuit configured with a Josephson element switching element. It is also possible to include the function of the averaging circuit in the weight control circuit 300. The averaging circuit can also be constructed from a semiconductor or the like. 6th
The figure shows the configuration of a simple counter circuit made up of Josephson elements.

The circuit shown in FIG. 6 has a configuration in which a bias current is supplied from a current source 402 to a superconducting closed circuit 410 consisting of a quantum interference type Josephson element 400 and a load inductor 403. One end of the control line 404 of the quantum interference type Josephson element 400 is connected to the averaging circuit input line 202 via a resistor 401, and the other end is grounded. The superconducting closed path 410
A quantum interference type circuit 405 is coupled to. In the circuit shown in FIG. 6, each time a "1" signal is applied to the averaging circuit input line 202, a control current flows through the quantum interference type Josephson element 400, causing the element to temporarily transition to a voltage state. ,
In each case, a portion of the bias current is shunted to the inductor 403. Therefore, the current flowing through the inductor 403 is “
This corresponds to the number of 1" signals. This shunt current is detected by the quantum interference type circuit 405.

Although the above embodiments have been explained using a Josephson element as an example, it is clear that the present invention can also be applied to a threshold logic circuit constituted by an element that transitions in stages, such as a tunnel diode.

(Effects of the Present Invention) As described above, by using the present invention, a high-speed threshold logic circuit that can use a learning method with good learning efficiency and convergence can be constructed using a high-speed Josephson switching circuit. Therefore, according to the present invention, it is possible to realize a high-speed computer suitable for executing recognition judgment using a threshold logic circuit. Therefore, the present invention is indispensable for realizing a high-speed computer that performs this sophisticated recognition judgment.

[Brief explanation of the drawing]

Fig. 1 is a circuit diagram of an embodiment according to the present invention, Fig. 2a is a circuit diagram of a conventional threshold circuit, Fig. 2b is a graph showing voltage-current characteristics of a quantum interference Josephson device, and Fig. 3 is a conventional A graph showing the signal switching characteristics of the threshold circuit. Figure 4 is a graph showing the signal switching characteristics desirable in the learning process. Figure 5 is a configuration diagram of the threshold logic circuit and learning system using the threshold circuit in the present invention. , FIG. 6 is a circuit diagram showing an example of the averaging circuit of this embodiment, and FIG. 7 is a graph showing voltage-current characteristics of a quantum interference type Josephson element. 100...Quantum interference type Josephson device, 101...
・Control current line, 103...Load resistance, 104・
...Output line, 105...Output terminal, 106
... Current source, 150 ... Weight circuit, 151
...Input line, 152...Input terminal, 200
... Averaging circuit, 201 ... Average value output line, 202 ... Averaging circuit input line, 300 ... Weight control circuit, 301 ... Teacher signal line, 302 ... Teacher signal terminal, 310... Weight control line, 400... Quantum interference type Josephson element, 401...
・Resistor, 402...Current source, 403...Inductor,...Control line,...Quantum interference circuit,...Superconducting closure path,...Threshold circuit. Figure 1 Figure 2a Figure 5 Figure 3 Figure 4 → Ic Figure 5

Claims (1)

[Claims] 1) A threshold logic circuit that compares a weighted sum of a plurality of input signals with a predetermined threshold and outputs an output signal according to the comparison result, 1. A threshold logic circuit comprising: means for modifying the output signal; and means for generating an average value signal of the output signal, the average value of the output signal being used to modify the weight. 2) A threshold logic circuit according to claim 1,
A threshold logic circuit characterized in that the output signal transitions stepwise from a "0" state to a "1" state. 3) The threshold logic circuit according to claim 2,
A threshold logic circuit characterized in that the voltage of a Josephson element that transitions stepwise from a superconducting state to a voltage state is used as an output signal. 4) The threshold logic circuit according to claim 1,
A threshold logic circuit characterized by averaging output signals through multiple trials for learning, and directly outputting signals without averaging for logic operations. 5) A threshold logic circuit according to claim 1, comprising:
A threshold logic circuit characterized in that an integrating circuit is used to average output signals. 6) A threshold logic circuit according to claim 1, comprising:
A threshold logic circuit characterized in that a counter circuit is used to average output signals.