JPH079663B2 - Nerve pseudo element with sensitivity adjustment function - Google Patents

Description

Description: TECHNICAL FIELD The present invention is an element that constitutes a “neural network” (neural network), that is, a neural pseudo element, and is capable of sensitivity adjustment with respect to the strength of the sum of input signals. The present invention relates to the same element having various functions, and more particularly, to a neural pseudo element with a sensitivity adjusting function for realizing a neural network of a system that obtains the sum of input signals by resistance coupling.

2. Description of the Related Art In recent years, information processing based on a principle different from that of a conventional sequential processing type computer has been attempted, and one of them is a massively parallel processing circuit network called the neural network. This neural network includes those that use operational amplifiers and those that combine a large number of MPUs (microprocessor units). In this case, the individual input signals are weighted to find their sum, and the output signal to be sent next is obtained based on the obtained sum. However, in order to find this sum, resistance coupling is used. , RAM (Random Access Memory) has been devised to store the weighted value in memory.

For example, John JH Hopfield
opfield) and David W. Tank (David
The neural network proposed by W.Tank) has attracted attention because it gives one solution to the traveling salesman problem, that is, the problem of finding the shortest route to visit multiple cities once at a time. There is.

The structure (circuit) of one neural element based on the model of Hopfield et al. Is considered as shown in FIG. That is, this circuit includes signal input means 1, sensitivity adjusting means 2, and signal processing means (or a rear amplification element) such as an operational amplifier.
In the signal input means 1, a plurality of input potential (Xj) signals to the element are individually weighted (Wi.
For j), it is connected after passing through the resistor Rij. Then, after the total potential Σi Wij × Xj generated by the same resistance coupling is linearly adjusted by the sensitivity adjusting means 2,
The obtained potential φ is used as the input potential of the signal processing means 3, and finally the output Zi: Zi = f (Σi Wij × Xj-h) is obtained. This final output Zi becomes one of a plurality of input potentials (Xj) to the next element. In this case, the function of Zi ← f (-h) of the above Zi is realized by the circuit configuration after the operational amplifier. To set Wij <0, enter -Zj.
At that time, as an advantage of the operational amplifier, the fact that + and-outputs are obtained is utilized.

Thus, in the Hopfield model, the input Xj of one element uses the output Zj of another element, and the output potential of the circuit changes dynamically, and the operation is repeated until a stable solution is reached. It is like this.

The model can define an energy function E. The tendency of E toward the minimum value is shown below statistically.

E = −0.5 × ΣWij × Xj × Xj−hΣXi This is the reason why the “physical” quantity that is effective in considering the statistical behavior of the neural network is submitted, and the quantity to be noted is clarified. Hopfield & Tank,
"Computer model using neural circuits" Nikkei Microdevices April 1987 issue].

Problems to be Solved by the Invention When considering further development of the above model,
It is desirable that the value of Wij can be changed by some learning. For that purpose, it is possible to use a variable resistance element, for example. For this purpose, for example, a configuration in which some resistors are switched by using a multiplexer to change the read potential, and in addition, a switched
A variable resistor that uses a resistor circuit has been prototyped (Kan Akiyama et al.). In any case, Wij will have some finite maximum and minimum values as long as real elements are used.

On the other hand, in the learning theory of the neural network, Wij does not normally set any upper or lower limit. In practical use, when a certain input signal is a weak signal which is important for judging the output, it is not possible to give a sufficiently large weight to the meaningful signal. There is a danger that it will not be able to pick up such signals. Therefore, in order to prevent such a danger, it is necessary to pre-process the input signal or to make a neural element having a circuit structure capable of adjusting the sensitivity to the intensity of the input signal.

On the other hand, in sensory recognition of living body,
ber) is established and the range of input signal strength that can be handled is widened. This is because the intensity of excitement is proportional to the logarithmic function of the signal.However, in the model of Hopfield et al., ΣiWij × Xj is simply treated linearly and input to an input signal such as an operational amplifier as it is. In that case, the operational amplifier is required to have high linearity in a wide voltage range. At the same time, Wij must be able to change over a wide range.

However, it has not been realized to change Wij in a wide range for the following reasons. That is, as shown in FIG. 4 above, according to the model of Hopfield et al., The input potential φ of the operational amplifier is expressed by the following equation φ = (Σi Xj × Rij) / [1 / r + Σi (1 / Rij)] (Rij: resistance value for weighting each input potential Xj, r: fixed resistance value for sensitivity adjustment). Here, since the input potential φ is proportional to ΣiWij × Xj, it can be regarded as the same except for multiplication by a constant. Therefore, it can be assumed that | Wij | = 1 / Rij. Due to the relationship between the two, the width in which the weighting Wij of coupling can change when the resistance value Rij changes within a certain finite range becomes extremely narrow. Therefore, when the level of the signal becomes incomparable, the input level of the operational amplifier will be incomparable as it is, and the operational amplifier will not be able to follow it, and the element will not function sufficiently.

In view of the above circumstances, according to the present invention, the variable width of the coupling weight is reduced by the resistance value which can be changed only by a finite value. It is an object of the present invention to provide a nerve pseudo element with a sensitivity adjusting function, which prevents a loss of power and can provide an output signal within a range that can be easily handled even if the input signal is an order of magnitude difference. To do.

Means for Solving the Problems In order to solve the above problems, the inventor has performed a function in a wide range for signal strength, that is, a function corresponding to the above Weber's law, which is performed in a living body. The present invention has been completed by studying bringing it into a nerve pseudo element and finding the effectiveness of adjusting the input potential to the rear amplification element by using a non-linear resistance.

Therefore, the nerve quasi-element with sensitivity adjusting function according to the present invention weights each of a plurality of input signals, and outputs a signal corresponding to the sum of the signal input means and a signal output from the signal input means. Sensitivity adjustment having non-linearly adjusted sensitivity adjustment means and signal processing means for inputting a signal whose sensitivity is adjusted by the sensitivity adjustment means and outputting an output signal for performing signal processing based on the input signal In the neural pseudo device with a function, the sensitivity adjusting means has a non-linear voltage-current characteristic represented by the following equation.

I = A × V ^P (P> 1) [wherein I: current of sensitivity adjusting means, V: voltage of sensitivity adjusting means, A: constant] In order to obtain the sum of weighting of the plurality of input signals,
For example, resistive coupling can be used.

Action As described above, in the circuit characteristic in the configuration shown in FIG. 4 which does not use the non-linear resistance (using the fixed resistance r), Zi = f (ΣiWij × Xj−h) φ = (ΣiXj × Rij) / [1 / r + Σi (1 / Rij)], and φ is proportional to Σi Wij × Xj.

On the other hand, the nerve pseudo device with sensitivity adjusting function according to the present invention has a circuit configuration using a non-linear resistance r'in the sensitivity adjusting means 2 as shown in FIG. Here, when the input potential to the rear amplification element 3 is φ ′, Zi = f (Σi W′ij × Xj−h) φ ′ = (Σi Xj × Rij) / [1 / r ′ + Σi (1 / Rij )] That is, since φ ′ is not a constant but changes with the voltage applied to it, φ ′ is no longer Σi W′ij.
It is no longer proportional to × Xj. However, instead of f (-h) in the above Hopfield model, another f '(-
If it is described using h '), the correspondence with the element in the Hopfield model is not impaired. In other words, f
There is no theoretical change other than changing the form of (-h). And the behavior similar to Weber's law is summarized in that the shape of this f (-h) is changed.

From the above, the neural quasi-element with sensitivity adjusting function according to the present invention is used when an expanded Hopfield model is made by changing the resistance value while maintaining a mathematical correspondence with the Hopfield model. A possible problem, that is, a part of the requirement for the variability of weighting when taking a weighted sum of signals can be replaced by a non-linear resistance, and the condition of variability of weighting can be relaxed.

Embodiment Hereinafter, the present invention will be described in detail with reference to the drawings.

As shown in FIG. 1, a nerve quasi device with a sensitivity adjusting function according to the present invention includes a signal input means 1, a sensitivity adjusting means 2,
And a signal processing means 3 for processing an input signal as described below.

First, in the signal input means 1, a plurality of input signals are individually weighted and the sum of them is output. Here, in order to obtain the sum of the weighting, it is preferable to use the resistance coupling by the variable resistance Rij as shown in the figure, but the invention is not limited to this.

Next, the sensitivity adjusting means 2 nonlinearly adjusts the output signal from the signal inputting means, and the signal processing means 3
An input signal φ'to is obtained. As the sensitivity adjusting means 2, for example, as shown in the figure, a non-linear resistance r'connected in series with a variable resistance Rij corresponding to weighting may be used, and the potential may be obtained from the midpoint thereof.
As the non-linear resistance r ′, for example, a metal phthalocyanine vapor-deposited film or the like can be preferably used, but it is not limited to this as long as non-linear current-voltage characteristics can be obtained.

The signal φ ′ whose sensitivity has been adjusted by the sensitivity adjusting means 2 is input to the signal processing means 3. The signal processing means 3 is, for example, an operational amplifier CMOS (Simos).
A rear amplification element configured by using the above can be used. Since the input impedance of these rear amplification elements is extremely large, the input potential does not change by connecting these rear amplification elements. The final output signal Zi
Is usually detected as a continuous potential between 0 and 1 or a discrete potential of 0,1, but may be output as a stepwise signal of, for example, three values or more, and is not particularly limited. .

Here, the value of the variable resistance Rij in the signal input means 1 is determined by a learning tour method (for example, Hebb's learning rule, error correction learning, correlation learning, etc.), and depends on the input voltage Xi itself. Does not change. On the other hand, when a non-linear resistance r'having a non-linear voltage-current characteristic is used as the sensitivity adjusting means 2, as shown below, the input potential to the signal processing means 3 behaves in a manner similar to Weber's law. Can be made.

That is, the non-linear resistance r ′ is, for example, as shown in FIG. 2, I = A × V ^P (P> 1) [where V: voltage, I: current, A: proportional constant] It has characteristics. That is, as the potential difference increases, the resistance value decreases and the current easily flows. Therefore, if this is applied to the nerve pseudo device shown in FIG. 1, the potential difference (sensitivity-adjusted sum signal) applied to the non-linear resistance becomes smaller as compared with the case of the proportional resistance, as shown in FIG. is there. However, to be precise, as described above, in Weber's law, the intensity of excitement is proportional to the logarithmic function of the signal, whereas in this example, it is proportional to the power of the signal. In addition, the qualitative tendency is consistent even if the non-linear shape changes.

From the above, a part of the demand that the change in weighting is in the order of magnitude can be offloaded to the non-linear resistance element, and even if the total sum of the input signals to the signal processing means is large, the order of magnitude is changed. Is less likely to occur.

Next, more detailed examples of the present invention will be described, but it goes without saying that the present invention is not limited to the following one example.

Example As the weighting of the signal input means 1, for the sake of simplification, the variable resistance is replaced with a fixed resistance, an operational amplifier is used as the signal processing means 3, ten input lines are connected by the fixed resistance Rij, and nonlinear A nerve pseudo device with a sensitivity adjusting function having a circuit configuration as shown in FIG. 1 was manufactured by grounding (GND) via a resistor r '. As the above operational amplifier, the LM607 with an input impedance of approximately 1 MΩ is used with a power supply voltage of ± 15 V.
I used it in. For the non-linear resistance r ', an element provided with electrodes on both sides of a lead phthalocyanine vapor deposition film was used. The voltage-current characteristic of the non-linear resistance r'is shown in FIG. 2, which shows that I = 0.02 (mA) and V = 10 at V = 5 (V).
The conductance I / V increases with the voltage, that is, the resistance value V / I decreases with the voltage, such as I = 0.16 (mA) at (V), and the voltage applied to the nonlinear resistance r ′ is At about 10V, V / I showed a resistance value of 63kΩ. The input signal potential was 0 to 15V.

In the above-described neural pseudo element with sensitivity adjusting function, for the sake of simplicity, Σi (Xj / Ri
The input potential φ'to the operational amplifier for j) was obtained, and the result is shown in FIG.

As shown in FIG. 3, the nerve pseudo device of the above-described embodiment is configured so that the sensitivity adjusting means can obtain an input signal proportional to a power lower than the first power of the total sum of weighting of the input signals, It can have a function similar to sensitivity adjustment.

EFFECTS OF THE INVENTION As described above, the nerve quasi-element with sensitivity adjusting function according to the present invention has a circuit configuration in which the Hopfield model is slightly changed by using the sensitivity adjusting means capable of non-linear adjustment, and the weighting is performed. The non-linear sensitivity adjusting means can be used as a substitute without excessively increasing the requirement for the width of the change of. Therefore, it is possible to prevent the element from being unable to pick up the change when the signal level is changed by an order of magnitude, and it is possible to obtain an output signal in a sufficiently easy range even if the input signal is out of order. .

[Brief description of drawings]

FIG. 1 is a circuit diagram showing a configuration of an embodiment of a nerve pseudo device with a sensitivity adjusting function according to the present invention, FIG. 2 is a diagram showing voltage-current characteristics of a non-linear resistance used in the embodiment, and FIG. FIG. 4 is a diagram showing the relationship between the sum of weighted input potentials and the input potential to the operational amplifier obtained in the example, and FIG. 4 is a circuit diagram showing one configuration of a conventional nerve pseudo device based on the Hopfield model. 1 ... Signal input means, 2 ... sensitivity adjusting means, 3 ... signal processing means, r '... non-linear resistance.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Katsuhiro Nijoki 3-10-1 Higashisanda, Tama-ku, Kawasaki-shi, Kanagawa Matsushita Giken Co., Ltd. Examiner Hiroki Hayashi (56) Reference JP-A-61-195470 (JP, A) JP-A-60-20641 (JP, A) Actually developed Shou 62-46901 (JP, U)

Claims (2)

[Claims] 1. A signal input means for weighting each of a plurality of input signals and transmitting a signal corresponding to the sum thereof, and a sensitivity adjusting means for nonlinearly adjusting the signal output from the signal input means. A signal having a sensitivity adjusted by the sensitivity adjusting means is inputted, and a signal processing means for performing signal processing based on the inputted signal and outputting an output signal is provided. The sensitivity adjusting means is a nerve pseudo element with a sensitivity adjusting function having a non-linear voltage-current characteristic represented by the following formula. I = A × V ^P (P> 1) [where I: current of sensitivity adjusting means, V: voltage of sensitivity adjusting means, A: constant] 2. The sensitivity adjusting function according to claim 1, wherein a weighted sum of the plurality of input signals is obtained by resistance coupling, and the non-linear voltage-current characteristic of the sensitivity adjusting means is obtained by a metal phthalocyanine vapor deposition film. Neural pseudo element.