JP7196803B2 - Artificial Neural Network Circuit and Learning Value Switching Method in Artificial Neural Network Circuit - Google Patents

Description

The present invention relates to an artificial neural network circuit equipped with a crossbar circuit having a memristor, and a learning value switching method in the artificial neural network circuit.

In recent years, it has been confirmed that an artificial neural network (hereinafter referred to as NN) such as a deep neural network exhibits recognition performance that surpasses that of conventional machine learning, for example, in the field of image recognition. However, the artificial NN generally has a high computational load. Therefore, an artificial NN may be implemented in software by a so-called GPGPU (general-purpose computing on GPU) using a GPU (graphic processing unit) suitable for highly efficient parallel computation.

However, GPGPU consumes a lot of power. Therefore, artificial NNs using GPGPU are applicable, for example, in cloud computing-based applications, but may be difficult to apply in non-cloud computing-based applications. Examples of non-cloud computing-based applications include an application that recognizes obstacles in a vehicle and performs anti-collision control to prevent/mitigate a collision with the obstacle, and an application that automatically drives the vehicle to its destination. There are applications for driving. When executing these applications, it is required to operate at high speed while consuming low power. Therefore, hardware circuits dedicated to artificial NNs that are low power consumption, high speed, and small size are desired.

WO2017/010048 WO2017/010049

As one of the hardware circuits dedicated to the above artificial NN, the inventor of the present application has been researching an artificial NN circuit equipped with a crossbar circuit having a resistive memory (hereinafter referred to as a memristor). The results of that research are disclosed, for example, in US Pat.

A crossbar circuit is configured by arranging a plurality of input bars and a plurality of output bars to intersect each other, and connecting the input bars and the output bars at each intersection point via a memristor. When a voltage signal corresponding to the output value of the preceding neuron is input to the input bar of the crossbar circuit, each voltage signal is multiplied by the conductance value of the memristor as a weight before being sent to the processing circuit connected to the output bar. A voltage signal is output and summed by a processing circuit. Then, the processing circuit generates and outputs an output value according to the activation function from the calculated total voltage signal as an input to the neuron in the subsequent stage. In this way, synapses in the artificial NN are implemented in hardware using crossbar circuits.

In an artificial NN circuit equipped with such a crossbar circuit, for example, by computer simulation, weights to be assigned to each signal to be transmitted are learned, and each memristor is set to a conductance value corresponding to the learned weights. be. A memristor can rewrite the resistance value between a minimum value and a maximum value by applying a voltage.

However, the artificial NN circuit provided with a crossbar circuit having a memristor may experience deterioration in performance due to changes in the environmental temperature in which the artificial NN circuit is placed due to the temperature characteristics of the memristor. There is For example, when image recognition is performed using the artificial NN circuit described above, the recognition error rate may increase as the ambient temperature rises.

SUMMARY OF THE INVENTION It is an object of the present invention to provide an artificial neural network circuit capable of suppressing deterioration in performance when the environmental temperature changes, and a learning value switching method in the artificial neural network circuit. With the goal.

In order to achieve the above object, the artificial NN circuit according to the present invention is:
a crossbar circuit (20) for transmitting signals between hierarchized neurons of the artificial NN;
The crossbar circuit is arranged such that a plurality of input bars (21) and a plurality of output bars (22, 23) intersect, and a signal is transmitted to each intersection of the plurality of input bars and the plurality of output bars. is provided with a memristor (24), which is a resistance change memory that gives weight to a signal that
A processing circuit (27a, 27b) that performs at least the calculation of the sum of the signals that are respectively weighted by the memristors and flows into each output bar as signal processing in the neuron,
a temperature sensor (33) for detecting the ambient temperature of the crossbar circuit;
updating units (31, 34, 35) for updating the learning values used in the crossbar circuit and/or the processing circuit to learning values corresponding to the change in the environmental temperature in accordance with the change in the environmental temperature of the crossbar circuit; , provided.

Further, the learning value switching method in the artificial NN circuit according to the present invention includes:
a crossbar circuit (20) for transmitting signals between hierarchized neurons of the artificial NN;
A crossbar circuit is arranged such that a plurality of input bars (21) and a plurality of output bars (22, 23) intersect, and a signal is transmitted to each intersection of the plurality of input bars and the plurality of output bars. is provided with a memristor (24), which is a resistance change type memory that gives a weight to a signal that
As signal processing in the neuron, processing circuits (27a, 27a,
27b) and applied to an artificial NN circuit comprising
detecting the ambient temperature of the crossbar circuit using a temperature sensor (33);
Update units (31, 34, 35) update the learning values used in the crossbar circuit and/or the processing circuit to learning values corresponding to the change in the environmental temperature in accordance with the change in the environmental temperature of the crossbar circuit. including to do.

Each memristor provided in the crossbar circuit has a temperature characteristic. Therefore, the conductance value set in each memristor changes according to the change in the environmental temperature. If the rate of change in the conductance value is uniform in each memristor included in the crossbar circuit, the mutual relationship of the weights based on the respective conductance values is maintained among the memristors. However, the conductance value of each memristor may not necessarily change at a uniform rate in response to changes in environmental temperature. That is, for the same temperature change, some memristors may have a large change in conductance value, while other memristors may have a small change in conductance value. In this way, if there is a variation in the rate of change in the conductance value due to a change in the environmental temperature in each memristor included in the crossbar circuit, the mutual relationship of weights among the memristors collapses according to the change in the environmental temperature. end up As a result, the performance of the artificial NN circuit is degraded.

Therefore, as described above, the artificial neural network and the learning value switching method in the artificial neural network according to the present invention detect the environmental temperature of the crossbar circuit (20) with the temperature sensor (33), and In response to changes in temperature, the learning values used in the crossbar circuitry and/or processing circuitry are configured to update to learning values corresponding to changes in ambient temperature.

By adopting such a configuration, even if the environmental temperature changes, it is possible to suppress the influence of variation in the conductance value of each memristor (24), and as a result, the deterioration of the performance of the artificial NN circuit is suppressed. It becomes possible to

The reference numbers in parentheses above merely indicate an example of correspondence with specific configurations in the embodiments described later in order to facilitate the understanding of the present disclosure, and are not intended to limit the scope of the invention. It's not what I did.

In addition, technical features described in each claim of the scope of claims other than the features described above will become apparent from the description of the embodiments and the accompanying drawings, which will be described later.

1 is a diagram conceptually showing an example of a multilayer NN; FIG. FIG. 4 is an explanatory diagram for explaining a crossbar circuit; FIG. 3 is an explanatory diagram for explaining a crossbar circuit and a CMOS circuit; FIG. 11 is another explanatory diagram for explaining the crossbar circuit; FIG. 4 is a diagram showing an example of temperature characteristics of a memristor; FIG. 5 is an explanatory diagram for explaining variations in temperature characteristics of a memristor; FIG. 10 is a graph showing the result of investigating the rate of occurrence of MNIS data recognition errors at room temperature using each chip on which an artificial NN circuit is mounted; FIG. FIG. 10 is a graph showing the result of examining the occurrence rate of recognition errors when an artificial NN circuit mounted on each chip is used at an ambient temperature of 100° C. FIG. 3 is a configuration diagram showing a configuration for updating learning values used in the crossbar circuit 20; FIG. 4 is a flowchart showing learning value update processing performed by the rewrite control device; In order to demonstrate the effect of the artificial NN circuit of the second embodiment, in the crossbar circuit of the artificial NN circuit mounted on each chip having the recognition error rate shown in FIG. 10 is a graph showing the results of examination of the rate of occurrence of recognition errors when only the weight for the bias signal is updated to a learning value re-learned so as to adapt to the temperature when the temperature belongs to . In order to show the effect of the artificial NN circuit of the second embodiment, in 10 chips in which the weight set #1 described in FIG. FIG. 10 is a graph showing the results of examining the recognition error rate at each temperature from room temperature to 100° C. with and without changing any weights. FIG. In order to demonstrate the effect of the artificial NN circuit of the second embodiment, 10 chips were prepared in which each of the weight sets #1 to #5 described with reference to FIG. FIG. 10 is a graph showing the results of examining the rate of recognition error occurrence at each temperature from room temperature to 100° C. when only the weights for are updated and when no weights are changed. In order to demonstrate the effect of the artificial NN circuit of the third embodiment, 10 chips in which the weight set #1 described in FIG. FIG. 10 is a graph showing the rate of occurrence of recognition errors in a high temperature region when only weights for bias signals are updated with adaptive learning values and learning values adapted to 100° C. and when no weights are changed. In order to show the effect of the artificial NN circuit of the third embodiment, the variation of the average occurrence rate of recognition errors in the high temperature region for every 10 chips for which the weight sets #1 to #5 are set and the average value are connected. When only the weight for the bias signal is updated with the learning value adapted to 70° C., the learning value adapted to 85° C., and the learning value adapted to 100° C., and when the weight is not changed at all. It is a graph showing about. FIG. 11 is a configuration diagram showing an example of configurations of a crossbar circuit of an artificial neural network circuit and a processing circuit that processes an output signal of the crossbar circuit according to the fourth embodiment; FIG. 12 is a configuration diagram showing an example of the configuration of the crossbar circuit of the artificial NN circuit of the fifth embodiment; FIG. 20 is a configuration diagram showing an example of the configuration of the crossbar circuit of the artificial NN circuit of the sixth embodiment;

(First embodiment)
Embodiments of an artificial NN circuit according to the present invention will be described in detail below with reference to the drawings. The artificial NN circuit according to the present embodiment can be used, for example, to implement a multi-layer NN (multi-layer perceptron) in which at least neurons are hierarchized in three layers, or a so-called convolution NN, in hardware. In the following description, an example in which a multi-layer NN is implemented in hardware using an artificial NN circuit will be described.

FIG. 1 is a diagram conceptually showing an example of a multilayer NN 10. As shown in FIG. As shown in FIG. 1, the multi-layer NN 10 comprises an input layer 11, an intermediate layer 12 and an output layer 13. As shown in FIG. Each of the input layer 11 , intermediate layer 12 and output layer 13 has at least one neuron 14 . The intermediate layer 12 is also called a hidden layer.

When input data for the multilayer NN 10 is input to the input layer 11, each neuron 14 of the input layer 11 outputs outputs x1, x2, . . . corresponding to the input data. Then, the outputs x1, x2, . . . are multiplied by weights w1, w2, . The neuron 14 of the intermediate layer 12 calculates the total sum Σwixi, converts the total sum Σwixi by an activation function f such as ReLU or tanh, and generates an output y (=f(Σwixi)). Then, the generated output y is output to each neuron 14 in the subsequent layer. The latter layer is the output layer 13 when the multilayer NN circuit 10 has three layers. In the case of four or more layers, it becomes a latter intermediate layer (not shown). In this way, the input data are transformed by the activation function f of each neuron 14 and finally input to the output layer 13 . The neuron 14 of the output layer 13 converts the input using, for example, a softmax function and outputs the result.

Such a multi-layer NN 10 can be applied to use an image (eg, MNIST dataset) as input data and classify the image into multiple categories (eg, numbers 0-9). In this case, the input data to the input layer 11 can be multiple pixel values (eg, 0-255) contained in the image. Also, the output from each neuron 14 of the output layer 13 can be configured to be the probability that an image as input data belongs to each category. The weights of the synapses 15 of the multilayer NN 10 are determined by learning to suit such applications.

By selecting the category corresponding to the highest probability among the probabilities output by each neuron 14 corresponding to each category of the output layer 13, an image as input data can be classified into a plurality of categories. can. Of course, the multilayer NN circuit 10 can also be applied to other uses. For example, it can be applied to object or person detection, human face recognition, or road sign recognition. Furthermore, it can be applied to applications such as information compression, motion control, noise removal, and time-series prediction.

Next, an artificial NN circuit that implements the multi-layer NN 10 described above in terms of hardware will be described with reference to FIGS. 2 to 4. FIG. As shown in FIG. 2, the artificial NN circuit includes a crossbar circuit 20 that implements synapses in hardware, and CMOS circuits 25 and 26 that implement neurons 14 in hardware. The CMOS circuits 25 and 26 are formed on a CMOS circuit board. The crossbar circuit 20 can be formed on a CMOS circuit board on which the CMOS circuits 25, 26 are formed.

The crossbar circuit 20 transmits signals between adjacent layers of the multi-layer NN 10 described above while giving weights as synapses. 2 to 4 show only part of the crossbar circuit 20 for convenience of illustration and explanation. As shown in FIG. 2, the crossbar circuit 20 has a plurality of input bars 21 and a plurality of output bars 22,23. A plurality of input bars 21 and a plurality of output bars 22 and 23 are arranged to cross each other. The input bar 21 is configured to receive, for example, voltage signals corresponding to the pixel values described above via a CMOS circuit 25 as an input neuron. It is a resistance change memory that weights the input signal transmitted from the input bar 21 to the output bars 22 and 23 at each intersection of the plurality of input bars 21 and the plurality of output bars 22 and 23. A memristor 24 is provided. That is, the plurality of input bars 21 and the plurality of output bars 22 and 23 are connected via corresponding memristors 24, respectively.

The memristor 24 is a resistance change memory that can be set to different resistance states between a minimum value and a maximum value depending on the applied voltage. For example, the conductance value of the memristor 24 can be increased by applying a negative write voltage and decreased by applying a positive write voltage using a conductance rewriting device 35, which will be described later. . The memristor 24 maintains the set resistance state (conductance value) unless a voltage equal to or higher than the positive and negative write voltage thresholds is applied. Devices that can be used as such memristors 24 include Pt/TiO2/Pt metal oxide devices, phase change memories, magnetic tunnel junction memories, and the like. Further, the memristor 24 may be a non-volatile resistance variable memory, even if the control method is different. For example, a flash memory or the like can be used.

Thus, the crossbar circuit 20 and the CMOS circuits 25, 26 serving as input and output neurons can be configured as a hybrid CMOS circuit with the memristor 24 incorporated therein. The CMOS circuit 26 serving as an output neuron will now be described with reference to FIG. FIG. 3 is an electric circuit diagram of the crossbar circuit 20 and the CMOS circuit 26 as an output neuron.

As shown in FIG. 3, in this embodiment, a pair of output bars 22 and 23 are used as a differential pair output bar so that between each input bar 21, the positive side of the differential pair output bar According to the difference between the weight (conductance value) G(+) by the memristor 24 connected to the output bar 22 and the weight (conductance value) G(-) by the memristor 24 connected to the negative output bar 23, A positive or negative weight can be given to the signal input to each input bar 21 . The positive output bar 22 of the differential pair output bar is the output bar connected to the non-inverting input terminal of the differential operational amplifier 28, and the negative output bar 23 is the inverting input terminal of the differential operational amplifier 28. An output bar connected to an input terminal.

A CMOS circuit 26 serving as an output neuron is output from a positive output bar 22 connected to a memristor 24 that imparts a positive weight (G(+)) to the input signal, as shown in FIG. An adder 27a that calculates and outputs the sum of signals, and a negative side output bar 23 connected to a memristor 24 that gives a negative weight (G(-)) to an input signal calculates and outputs the sum of signals. It has an adder 27b and a differential operational amplifier 28 that calculates and outputs the difference between the sums output from the adders 27a and 27b. Furthermore, the CMOS circuit 26 serving as an output neuron converts the difference in the sum output from the differential operational amplifier 28 by a predetermined activation function f and outputs the converted value, although not shown. It also has an activation function processing circuit for Thus, the adders 27a, 27b, the differential operational amplifier 28 and the activation function processing circuit can be constructed using CMOS elements in a CMOS substrate. However, the calculation of the difference and the conversion by the activation function f are not executed in the CMOS circuit 26. For example, after converting the outputs of the adders 27a and 27b into digital signals, they are executed in a separately provided digital processing circuit. You may

Adders 27a and 27b have a configuration in which a resistor having a resistance value R is connected between the input and output of an operational amplifier. Therefore, as shown in FIG. 3, when the voltage signals V ₁ and V ₂ are input to the respective input bars 21, the voltage signal V ₁ is multiplied by the conductance G ⁽⁺⁾ ₁ so that the voltage signal V ₂ becomes is multiplied by the conductance G ⁽⁺⁾ ₂ and then added in the adder 27a to calculate the total sum. Further, the sum is multiplied by R by the adder 27a. Similarly, in the adder 27b, the result obtained by multiplying the voltage signal V ₁ by the conductance G ⁽⁻⁾ ₁ and the result obtained by multiplying the voltage signal V ₂ by the conductance G ⁽⁻⁾ ₂ are added to obtain the total sum. is calculated, and the sum is multiplied by R.

Then, the differential operational amplifier 28 calculates and outputs the difference between the sum output from the adder 27a and the sum output from the adder 27b. Therefore, in the CMOS circuit 26 as the output neuron, the result shown in Equation 1 below is obtained.

(Number 1)
Output voltage = R (G ⁽⁺⁾ ₁ V ₁ - G ^(-) ₁ V ₁ + G ⁽⁺⁾ ₂ V ₂ - G ^(-) ₂ V ₂ )
Therefore, the weight imparted by the memristor 24 when the voltage signal V ₁ as the input signal is propagated through the differential pair output bars 22, 23 is R(G ⁽⁺⁾ ₁ -G ^(-) ₁ ) becomes. Also, the weight given by the memristor 24 when the voltage signal V ₂ as the input signal is transmitted is R(G ⁽⁺⁾ ₂ −G ⁽⁻⁾ ₂ ). Although not shown in FIGS. 2 and 3, weights for other input signals are similar.

Therefore, when a certain input signal is multiplied by a positive weight, the memristor 24 provided between the input bar 21 and the output bar 22 connected to the non-inverting input terminal of the differential operational amplifier 28 is The conductance value should be set larger than the conductance value of the memristor 24 provided between the output bar 23 connected to the inverting input terminal and the input bar 21 by the positive weight to be set. Conversely, when a certain input signal is multiplied by a negative weight, the conductance value of the memristor 24 provided between the output bar 23 connected to the inverting input terminal and the input bar 21 is applied to the non-inverting input. The conductance value of the memristor 24 provided between the output bar 22 and the input bar 21 connected to the terminal may be set larger by the negative weight to be set. In this manner, the crossbar circuit 20 of the present embodiment can implement positive and negative weights as synapse weights, and the crossbar circuit 20 can simulate excitatory and inhibitory synapses.

FIG. 4 shows the crossbar circuit 20 and the CMOS circuit 26 as output neurons to an enlarged extent. 4, the adders 27a and 27b shown in FIG. 3 are omitted. In practice, adders 27 a , 27 b are provided at the input of each differential operational amplifier 28 . Furthermore, FIG. 4 also omits the above-described activation function processing circuit. In practice, an activation function processing circuit is provided at the output of differential operational amplifier 28 . As shown in FIG. 4, an input bar 21 of the crossbar circuit 20 receives a bias signal in addition to the voltage signal corresponding to the pixel value described above or the voltage signal corresponding to the output from the output neuron in the preceding layer. be. The magnitude of this bias signal is normally set to a predetermined value (for example, "1").

Here, each memristor 24 provided in the crossbar circuit 20 has a temperature characteristic. Therefore, the conductance value set in each memristor 24 changes according to changes in the environmental temperature. FIG. 5 shows an example of the temperature characteristics of such a memristor 24. As shown in FIG. As shown in FIG. 5, even if the input voltage V is the same, the conductance value of the memristor 24 changes according to the ambient temperature, so the output current I also changes as a result.

If the rate of change in the conductance values of the memristors 24 included in the crossbar circuit 20 is uniform for all memristors 24, the mutual relationship of the weights based on the respective conductance values is maintained among the memristors 24. FIG. However, as shown in FIG. 6, the conductance value of each memristor 24 may not necessarily change at a uniform rate according to changes in environmental temperature. That is, for the same temperature change, some memristors 24 may have large changes in conductance values, while other memristors 24 may have small changes in conductance values. As described above, if the rate of change in the conductance value due to a change in the environmental temperature varies in each memristor 24 included in the crossbar circuit 20, the mutual relationship of the weights among the memristors 24 will be lost. As a result, the performance of the artificial NN circuit is degraded.

Actually, using the MNIST data set, the multi-layer NN 10 having the configuration shown in FIG. An embodied artificial NN circuit was prepared. More specifically, the input image has a size of 28×28 pixels, so the number of neurons in the input layer 11 is 785 (28×28+1) for inputting pixel values and biases for each pixel. The number of neurons in the intermediate layer 12 is 300 for receiving inputs from the neurons in the input layer 11, and 301 in total by adding bias inputs. The number of neurons in the output layer 13 is ten. Thus, the number of input bars of the crossbar circuit 20 between the input layer 11 and the intermediate layer 12 is 785 and the number of output bars is 600 for differential pair output bars. The number of input bars of the crossbar circuit 20 between the intermediate layer 12 and the output layer 13 is 301 corresponding to the number of neurons (300) in the intermediate layer 12 and the bias signal, and the number of output bars is differential. 20 to make it a vs. output bar.

Five sets of weights for signals transmitted between neurons learned at room temperature and weights for bias signals are prepared. It was set to each memristor 24 of the crossbar circuit 20 of 10 chips on which the circuit was mounted. Using each chip mounted with an artificial NN circuit prepared in this manner, the occurrence rate of MNIS data recognition errors (misclassification errors) was examined at room temperature. The results are shown in FIG. As shown in FIG. 7, when the ambient temperature is room temperature, there is no significant difference in the rate of occurrence of recognition errors between the chips for which the same weight set is set. Note that there is a slight difference in the rate of occurrence of recognition errors between different weight sets.

Next, the environmental temperature was set to 100° C., and the occurrence rate of recognition errors was investigated when using the artificial NN circuit mounted on each chip. FIG. 8 shows the results. As shown in FIG. 8, compared to the room temperature state, the rate of occurrence of recognition errors in the artificial NN circuits mounted on each chip is high, and in any weight set, recognition errors occur between chips. It was confirmed that there was a difference in the rate. This proves that the memristors 24 included in the crossbar circuit 20 of each chip have variations in the rate of change in conductance value due to changes in environmental temperature.

In order to deal with such a problem, in the artificial NN circuit of this embodiment, the weights set in the crossbar circuit 20 are updated as learning values according to changes in the environmental temperature. In order to realize this, as shown in FIG. 9, the artificial NN circuit of this embodiment includes a rewrite control device 31, a temperature sensor 33, a storage device 34, and a conductance rewrite device as a configuration for updating the learning value. 35. Temperature sensor 33 detects the environmental temperature of crossbar circuit 20 and outputs the detected environmental temperature to rewrite control device 31 . The storage device 34 stores learning values (weight sets) to be used in the crossbar circuit 20 when the environmental temperature of the crossbar circuit 20 changes.

The rewriting control device 31 constitutes a learning value updating unit together with the conductance rewriting device 35 . The rewriting control device 31 is composed of a computer including a processor 32 and an internal memory (not shown). When the processor 32 executes the control program stored in the internal memory, the rewrite control device 31 rewrites the weight of each memristor 24 of the crossbar circuit 20 to a weight corresponding to the environmental temperature. The data is output to conductance rewrite device 35 . The rewrite data includes information specifying the memristor 24 to be rewritten and information on the conductance value to be set to each specified memristor 24 . The details of the learning value update processing by the rewrite control device 31 will be described later.

A conductance rewriter 35 is connected to the crossbar circuit 20 so that any voltage can be applied between each input bar 21 and each output bar 22, 23 of the crossbar circuit 20. FIG. In the example of FIG. 9, conductance rewriter 35 is connected to input bar 21 and output bars 22, 23, bypassing CMOS circuits 25, 26 embodying neurons. Therefore, the CMOS circuits 25 and 26 are omitted in FIG. The conductance rewriting device 35, in accordance with the rewriting data input from the rewriting control device 31, supplies the specified memristor 24 with a voltage equal to or greater than the write voltage threshold and for updating to the instructed conductance value. It is applied between corresponding input bars 21 and output bars 22,23. Thereby, the conductance of the desired memristor 24 can be rewritten.

Next, learning value update processing performed by the rewrite control device 31 will be described with reference to FIG. FIG. 10 is a flow chart showing the learning value update process performed by the rewrite control device 31. As shown in FIG. The learning value update process shown in FIG. 10 is periodically executed by the rewrite control device 31 .

First, in step S<b>100 , the rewrite control device 31 acquires the environmental temperature detected by the temperature sensor 33 . Then, in step S110, it is determined whether or not the environmental temperature belongs to a high temperature range equal to or higher than a predetermined temperature. The predetermined temperature for determining whether or not it is in the high temperature range can be set to an arbitrary temperature between 55.degree. C. and 70.degree. This is because the rate of occurrence of recognition errors in each chip increases in the high temperature range above these temperatures compared to room temperature. If it is determined in step S110 that the environmental temperature belongs to the high temperature range, the process proceeds to step S120. On the other hand, if it is determined that the environmental temperature does not belong to the high temperature range, the process proceeds to step S150.

In step S120, it is determined whether or not the learning values (that is, the weight set) of the crossbar circuit 20 have been updated to the learning values corresponding to the measured environmental temperature. If it is determined to have been updated, the learned value update process shown in the flowchart of FIG. 10 ends. On the other hand, if it is determined that the update has not been completed, the process proceeds to step S130. In step S130, the learning value corresponding to the measured environmental temperature is read from the storage device 34. FIG. Then, in step S140, rewritten data corresponding to the read learning value is output to the conductance rewriting device 35. FIG.

In step S150, which is executed when it is determined in step S110 that the environmental temperature does not belong to the high temperature range, the set of weights of each memristor 24 of the crossbar circuit 20 is set to an initial value adapted to the room temperature state. It is determined whether it has been returned. If it is determined that the values have been returned to the initial values, the learning value updating process shown in the flowchart of FIG. 10 ends. On the other hand, if it is determined that the value has not been returned to the initial value, the process proceeds to step S160. In step S160, an initial value, which is a set of weights adapted to the room temperature state, is read from the storage device 34 as a learning value. Thus, the storage device 34 also stores initial values, which are a set of weights adapted to room temperature conditions, as learning values. In step S<b>170 , rewrite data corresponding to the read initial value is output to the conductance rewrite device 35 .

Here, the learning values stored in the storage device 34 will be described. The learning values include weights for signals transmitted between neurons and weights for bias signals, as described above. This learning value can be obtained, for example, by the following procedure. First, a set of weights learned at room temperature is set in the memristor 24 of the crossbar circuit 20 that constitutes the artificial NN circuit. Then, the occurrence rate of recognition errors in the artificial NN circuit is evaluated at various temperatures belonging to the high temperature range. At that time, the weight (conductance value) of each memristor 24 is read from the crossbar circuit 20 at the temperature at which the rate of occurrence of recognition errors has deteriorated. The read weights of each memristor 24 are re-learned in an artificial NN constructed on a computer using a general technique such as gradient descent or error backpropagation. This relearning may be performed only once, or may be repeated multiple times. A set of weights re-learned in this manner is stored in the storage device 34 as a learned value.

The learned value can also be obtained by another procedure described below. First, a set of weights learned at room temperature is set in the memristor 24 of the crossbar circuit 20 that constitutes the artificial NN circuit. Then, the occurrence rate of recognition errors in the artificial NN circuit is evaluated at various temperatures belonging to the high temperature range. At that time, the conductance value of each memristor 24 of the crossbar circuit 20 is rewritten so that the recognition error rate is improved at the temperature at which the recognition error rate is worsened. That is, the weight of each memristor 24 of the crossbar circuit 20 is re-learned using an actual artificial NN circuit. Then, the conductance value of each memristor 24 of the crossbar circuit 20 when the recognition error rate is improved is read out and stored in the storage device 34 as a learned value.

When the learning values are obtained according to the procedure described above, the storage device 34 stores a plurality of learning values that have been re-learned so as to adapt to various temperatures in the high-temperature region. However, as the number of learned values increases, problems such as an increase in the memory capacity of the storage device 34 and frequent updating of the learned values due to changes in environmental temperature occur. Therefore, it is preferable to divide the high-temperature region into a plurality of temperature ranges and store the learning values used in each temperature range in the storage device 34 . For example, the learning value re-learned to adapt to 70° C., the learning value re-learned to adapt to 85° C., and the learning value re-learned to adapt to 100° C. are stored in the storage device 34. Memorize. Then, for example, the rewrite control device 31 uses a re-learned learning value adapted to 70° C. in the temperature range of 60° C. or more and less than 75° C., and uses 85° C. in the temperature range of 75° C. or more and less than 90° C. , and in the temperature range of 90°C or higher, a relearned learning value adapted to 100°C can be used. This makes it possible to suppress the occurrence of the above-described problems.

(Second embodiment)
In the first embodiment described so far, the weights for signals transmitted between neurons and the weights for bias signals are re-learned at the environmental temperature of the high-temperature region, and the set of re-learned weights are used in the environment of the high-temperature region. It was to be stored in the storage device 34 as a learned value corresponding to the temperature. In this way, by using learned values obtained by re-learning the weights for signals transmitted between neurons and the weights for bias signals, it is possible to sufficiently reduce the occurrence rate of recognition errors at the environmental temperature in the high-temperature region.

On the other hand, as the number of weights included in the learning values stored in the storage device 34 increases, the number of memristors 24 whose conductance values should be rewritten also increases. Then, problems may occur such as it takes a long time to update the weights and a large-scale conductance rewriting device 35 is required. In order to deal with such a problem, it is conceivable to update not all the weights in the crossbar circuit 20 but only some of the weights. Some weights to be updated include weights for bias signals or weights for some input signals.

For example, in the crossbar circuit 20 of the artificial NN circuit mounted on each chip having the recognition error occurrence rate shown in FIG. 11, the rate of occurrence of recognition errors when only the weight for the bias signal was rewritten (updated) with a learning value that was re-learned to adapt to the temperature was examined, and the results shown in FIG. 11 were obtained. was taken. From the results shown in FIG. 11, it can be seen that the occurrence rate of recognition errors can be reduced by simply rewriting only the weights for the bias signal with the re-learned learning values compared to the case shown in FIG. 8 where the weights are not changed. .

Another graph is used to show the degree of difference in the recognition error rate between when only the weights for the bias signal are updated and when the weights are not changed with respect to changes in the environmental temperature. explain in detail. FIG. 12 shows a case where only the weights for the bias signals are updated and a case where the weights are not changed in ten chips in which the weight set #1 described in FIG. 7 is set in each memristor 24 of the crossbar circuit 20. , and the results of examining the rate of occurrence of recognition errors at each temperature from room temperature to 100°C. In the graph of FIG. 12, there are line segments connecting the variation of the occurrence rate of recognition errors in the case of ideal weighting, in which the weighting of each memristor 24 does not change according to the temperature change, and the average value thereof. Change the line segment connecting the variation of the recognition error rate and the average value when only the weight for the bias signal is updated by the learning value re-learned so as to adapt to each temperature of 100° C., and the weight. A line segment connecting the variation of the occurrence rate of recognition errors and the average value thereof is shown.

If the weights are not changed at all, variations in the magnitude of change in the conductance value of each memristor 24 with temperature change will directly affect the rate of occurrence of recognition errors. Therefore, as the ambient temperature rises from room temperature (approximately 27° C.), the rate of occurrence of recognition errors also increases, and at 100° C., the average value of 10 chips exceeds 2%.

On the other hand, when only the weight for the bias signal is updated by the learning value that is re-learned so as to adapt to each temperature, as in the present embodiment, the average recognition error occurrence rate for 10 chips is 70. ° C. to 100° C., it is reduced compared to the case where the weight is not changed at all, and it can be seen that there is an improvement.

Furthermore, FIG. 13 shows the bias signal in 10 chips (therefore, a total of 50 chips) in which each of the weight sets #1 to #5 described in FIG. FIG. 10 shows the result of investigation of the rate of occurrence of recognition errors at each temperature from room temperature to 100° C., when only the weights for are updated and when no weights are changed. More specifically, FIG. 13 shows the variation in the average recognition error rate for each of the 10 chips to which the weight sets #1 to #5 are set, and the line segment connecting the average values. For the ideal weights in which the weights do not change depending on the temperature, for the case where only the weights for the bias signal are updated with re-learned learning values adapted to each temperature, and for the case where the weights are not changed at all. there is

From the graph of FIG. 13, it can be seen that by updating only the weights for the bias signal with the learned values that are re-learned so as to adapt to each temperature, the rate of occurrence of recognition errors is reduced compared to the case where the weights are not changed. It turns out that it is reduced.

(Third embodiment)
In the examples described in the first and second embodiments, the high temperature region is divided into a plurality of temperature ranges, the learning values (set of weights) corresponding to each temperature range are stored in the storage device 34, and the detected The set of weights set in each memristor 24 of the crossbar circuit 20 is updated according to the temperature range to which the ambient temperature that is applied belongs. However, it is preferable that the frequency of updating the learned values is as low as possible.

Therefore, each artificial NN circuit, in which only the weight for the bias signal is updated by the learning values re-learned so as to adapt to each temperature of 70° C., 85° C., and 100° C., can be recognized over the entire high temperature region. I checked to see if it shows the error rate. The results are shown in FIGS. 14 and 15. FIG. As in the case of FIG. 12, FIG. 14 shows the learning values adapted to 70° C. and FIG. 10 is a graph showing the rate of occurrence of recognition errors in a high temperature region when only weights for bias signals are updated with adaptive learning values and learning values adapted to 100° C. and when no weights are changed. Also, as in the case of FIG. 13, FIG. 15 shows the variation of the average occurrence rate of recognition errors in the high-temperature region for every 10 chips for which the weight sets #1 to #5 are set, and the average value thereof. When only the weight for the bias signal is updated with the learning value adapted to 70° C., the learning value adapted to 85° C., and the learning value adapted to 100° C., and when the weight is not changed at all. is shown for

From the graphs of FIGS. 14 and 15, any of the artificial NN circuits in which only the weights for the bias signal are updated by the learning values re-learned to adapt to each temperature of 70° C., 85° C., and 100° C. , the occurrence rate of recognition errors can be reduced over almost the entire high-temperature region, compared to the artificial NN circuit that does not change the weights. From this result, if the weight of each memristor 24 in the crossbar circuit 20 is updated with a learning value that has been re-learned so as to adapt to any temperature belonging to the high temperature region, the artificial NN circuit whose weight has been updated is , can be used over the entire high-temperature region, and can be said to be able to sufficiently improve the occurrence rate of recognition errors compared to the case where the weights are not changed.

Furthermore, looking at the graphs of FIGS. 14 and 15, it can be seen that in many ranges in the high temperature region, the weights with the learning values adapted to 85° C. The updated artificial NN circuit has a lower recognition error rate. In addition, the artificial NN circuit whose weights are updated with learning values adapted to 100°C has a lower recognition error rate than the artificial NN circuit whose weights are updated with learning values adapted to 85°C. . From these results, it can be said that the learned value used throughout the high temperature range is preferably a learned value learned under a relatively high temperature environment within the high temperature range.

In addition, when one learning value is used in the entire high temperature region, if there is a weight that is substantially the same as the weight learned in the room temperature state among the weights included in the learning value, the weight may be excluded from the learning values (set of weights) stored in the storage device 34, since there is no need to update the conductance value of the corresponding memristor 24.

(Fourth embodiment)
In the above-described second embodiment, not all the weights in the crossbar circuit 20 are updated as learning values in response to changes in the environmental temperature, but only some weights, such as the weights for the bias signal, are subject to update. It was intended to be.

Here, the bias signal is typically input to one of the input bars 21 of the crossbar circuit 20, as shown in FIG. The bias signal is weighted by a memristor 24 provided at the intersection of the input bar 21 to which the bias signal is input and the output bars 22 and 23 that intersect the input bar 21, and the output bars 22 and 23 are weighted. flow into. After that, the weighted bias signal is integrated with other weighted input signals flowing into the respective output bars 22 and 23, and then the integrated signal of the output bar 22 and the output bar 23 are obtained by the differential operational amplifier 28. is calculated and output. Expressing this in a formula results in Formula 2 below.

(Number 2)
Differential operational amplifier output = Σ(w _ij −w _ij+1 )*x _i +(w _bj −w _bj+1 )*x _b
In Equation 2, x _i represents the input voltage signal corresponding to the pixel value of the image or the input voltage signal corresponding to the output from the output neuron in the preceding layer of the number input to the crossbar circuit 20, and ( w _ij −w _ij+1 ) represents the weight difference for the input voltage signal x _i , x _b represents the bias signal input to the crossbar circuit 20, and (w _bj −w _bj+1 ) represents the weight difference for the bias signal x _b . Show the difference.

As shown in Equation 2, the weighted bias signal (w _bj −w _bj+1 )*x _b is simply added to the differenced result of the integrated signal of each output bar 22,23. Therefore, even if the bias signal is not input to the crossbar circuit 20 but is added to the output of the differential operational amplifier 28 before being input to the activation function processing circuit, a result similar to Equation 2 can be obtained. be done.

From this point of view, in the artificial NN circuit of the fourth embodiment, as shown in FIG. 16, the crossbar circuit 20 receives only the input voltage signal _xi that does not include the bias signal. On the other hand, a weighted bias signal (w _bj −w _bj+1 )*x _b is configured to be added to the output of differential operational amplifier 28 . Specifically, the artificial NN circuit of the present embodiment includes a storage device 34 storing weighted bias signals suitable for each environmental temperature range of the crossbar circuit 20, and an environmental temperature detected by the temperature sensor 33. a bias control device 36 for reading out the weighted bias signal from the storage device 34 and outputting it; and an adder 37 for adding the weighted bias signal output from the bias control device 36 and the output from the differential operational amplifier 28. ing. Note that the storage device 34 divides the high-temperature region into a plurality of temperature ranges, for example, in addition to the weighted bias signal as the initial value used at room temperature, as described in the second embodiment, and A weighted bias signal may be stored as a learning value adapted to . Alternatively, in addition to the weighted bias signal as the initial value used at room temperature, the storage device 34 stores the weighted bias signal as one learning value used throughout the high temperature region as described in the third embodiment. A bias signal may be stored.

As described above, in the artificial NN circuit according to the fourth embodiment, the weighted bias signal used in the processing circuit that processes the output signal of the crossbar circuit 20 is used as the learning value, and the value corresponding to the change in the environmental temperature is changed. Update. In such a configuration, the weighted bias signal as the learning value is obtained by simply reading a value suitable for the environmental temperature of the crossbar circuit 20 from the storage device 34 and outputting it to the adder 37 . In other words, since it is not necessary to rewrite the conductance value of the memristor 24 in the crossbar circuit 20, it is possible to easily and quickly update the learning value.

(Fifth embodiment)
All of the artificial NN circuits according to the first to fourth embodiments described above have learning values (weight sets, A weighted bias signal, etc.) is stored in the storage device 34, and the learning value is read out from the storage device 34 in accordance with changes in the environmental temperature to update the weight set of the crossbar circuit 20 or to be used in the processing circuit. It was used to update the weighted bias signal that is used. On the other hand, the artificial NN circuit according to the fifth embodiment includes a plurality of crossbar circuits 20 preset so that the conductance values (learned values) of the memristors 24 are adapted to different environmental temperatures, and the temperature sensor 33 By selecting the crossbar circuit 20 to be used according to the detected environmental temperature of the crossbar circuit 20, the learning value corresponding to the change in the environmental temperature is updated.

FIG. 17 is a diagram showing an example of the configuration of an artificial NN circuit according to the fifth embodiment. As shown in FIG. 17, the artificial NN circuit according to this embodiment comprises a plurality of crossbar circuits 20a and 20b having weights adapted to different environmental temperatures according to the conductance values of the memristors 24, respectively. The crossbar circuit 20a is connected to the differential operational amplifier 28 via the changeover switch 29a, and the crossbar circuit 20b is connected to the differential operational amplifier 28 via the changeover switch 29b. A control device (not shown) turns on a changeover switch (one of 29a and 29b) of a crossbar circuit (one of 20a and 20b) having a weight adapted to the environmental temperature according to the environmental temperature detected by the temperature sensor, The changeover switches 29a and 29b are controlled to turn off the changeover switches (the other of 29a and 29b) of the other crossbar circuits (the other of 20a and 20b). With such a configuration, the weight set as the learning value used in the crossbar circuits 20 and 20b is updated to a value corresponding to the change in the environmental temperature without rewriting the conductance value of the memristor 24 of the crossbar circuit 20. can do.

FIG. 17 shows an example in which the artificial NN circuit includes two crossbar circuits 20a and 20b. You may switch the bar circuit.

(Sixth embodiment)
In the fifth embodiment described above, a plurality of crossbar circuits 20a and 20b having weights adapted to different environmental temperatures are provided, and the plurality of crossbar circuits 20a and 20b are switched according to the detected environmental temperature. Met. On the other hand, in the artificial NN circuit of the sixth embodiment, as shown in FIG. 18, a crossbar circuit 20 is provided with a plurality of input bars 21 as input bars for inputting bias signals. The conductance values of the memristors 24 connected to the respective input bars 21 are set in advance so that weights adapted to different environmental temperatures can be applied to the bias signals. .

In the artificial NN circuit of the present embodiment, a control device (not shown) selects an input bar having a weight adapted to the environmental temperature according to the environmental temperature detected by the temperature sensor, and the selected input bar It is configured to input a bias signal only to the In other words, among the plurality of input bars 21 for inputting the bias signal, no signal is input to the input bars other than the selected input bar. As a result, the artificial NN circuit of this embodiment can switch the weight given to the bias signal in accordance with the environmental temperature without using a changeover switch as in the fifth embodiment. Furthermore, like the artificial NN circuits of the fourth and fifth embodiments, the artificial NN circuit of this embodiment does not rewrite the conductance value of the memristor 24 of the crossbar circuit 20, and the learning value used in the crossbar circuit is can be updated to a learned value corresponding to changes in environmental temperature.

Although the preferred embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the gist of the present invention.

For example, in each of the above-described embodiments, an example has been described in which the output bars of the crossbar circuit 20 are used as a differential pair of output bars in order to give positive or negative weight to the transmitted signal. However, the crossbar circuit 20 does not necessarily have to set a differential pair output bar. In this case, all the weights of the memristors 24 connected to the output bar have positive signs, and the magnitude of the positive weights may represent the difference in the degree of influence.

Also, in the above-described second embodiment, an example of updating the weight for the bias signal as the learning value has been described. However, the learning value may be the bias signal itself in addition to or alternatively the weights for the bias signal. If the learned value is only the magnitude of the bias signal, it is not necessary to change the conductance value of the memristor 24, as in the fourth to sixth embodiments described above. Updates can be made easily and quickly. In this way, the learning value used in the crossbar circuit 20 includes at least one of the weight given to the signal transmitted between neurons, the weight given to the bias signal, and the magnitude of the bias signal. It is a thing.

Furthermore, in all crossbar circuits 20 included in the artificial NN circuit, instead of updating the learning value according to the detected environmental temperature, for example, the crossbar circuits between the input layer and the intermediate layer, and/or Alternatively, the learning value may be updated only in some crossbar circuits 20 such as the crossbar circuit between the intermediate layer and the output layer.

Further, for example, in the above-described first embodiment, the weights learned in the room temperature state are set by each memristor 24 of the crossbar circuit 20, and the environmental temperature of the crossbar circuit 20 is set to a temperature belonging to a high temperature range equal to or higher than a predetermined temperature. When the temperature changes, the weights (learned values) set in the crossbar circuit 20 are updated so as to adapt to the changed temperature.

However, in the first and second to sixth embodiments, the learning values initially used in the crossbar circuit and/or the processing circuit are at temperatures other than room temperature (higher than room temperature, or at room temperature). lower temperature). Then, the temperature range in which the performance of the artificial NN circuit deteriorates due to the temperature characteristics of the memristor 24, other than the temperature at which the initially used learning value is learned (the temperature at which the initially used learning value is learned) (including a low temperature range and a high temperature range) may also be learned, and the learned value may be updated according to changes in the environmental temperature of the crossbar circuit 20 .

Each of the above-described embodiments can be arbitrarily combined and implemented, even if the combination is not described, as long as the combination does not cause any trouble.

10: multilayer NN circuit, 11: input layer, 12: intermediate layer, 13: output layer, 14: neuron, 20, 20a, 20b: crossbar circuit, 21: input bar, 22: output bar, 23: output bar, 24: memristor, 25: CMOS circuit as input neuron, 26: CMOS circuit as output neuron, 27a: adder, 27b: adder, 28: differential operational amplifier, 29a, 29b: switch, 31: rewrite control Device, 32: Processor, 33: Temperature sensor, 34: Storage device, 35: Conductance rewriting device, 36: Bias control device, 37: Adder

Claims (18)

a crossbar circuit (20) for transmitting signals between hierarchized neurons of an artificial neural network;
The crossbar circuit is arranged so that a plurality of input bars (21) and a plurality of output bars (22, 23) intersect, and at each intersection of the plurality of input bars and the plurality of output bars, , a memristor (24), which is a resistance change type memory that gives weight to the transmitted signal, is provided,
an artificial neural network circuit comprising: processing circuits (27a, 27b) for performing, as signal processing in the neuron, at least calculation of the sum of signals respectively weighted by the memristors and flowing into the respective output bars; and
a temperature sensor (33) for detecting the ambient temperature of the crossbar circuit;
update units (31, 34, 35) and an artificial neural network circuit comprising: Bias signals are input to the plurality of input bars of the crossbar circuit in addition to the signals transmitted between the neurons,
The learning value used in the crossbar circuit and/or the processing circuit includes at least one of a weight given to a transmitted signal, a weight given to the bias signal, and a magnitude of the bias signal. The artificial neural network circuit of claim 1 comprising. the processing circuitry is configured to add a weighted bias signal to the summed signals flowing into each of the output bars;
2. The artificial neural network circuit of claim 1, wherein learning values used in said crossbar circuit and/or said processing circuit include said weighted bias signal. The updating unit rewrites the weights of the memristor of the crossbar circuit according to a change in the environmental temperature of the crossbar circuit, thereby updating the learning value used in the crossbar circuit and/or the processing circuit to: 2. The artificial neural network circuit according to claim 1, wherein the learning value is updated to correspond to changes in environmental temperature. A plurality of the crossbar circuits are provided so as to apply weights adapted to different environmental temperatures,
The update unit is used in the crossbar circuit and/or the processing circuit by selecting from the plurality of crossbar circuits weighted crossbar circuits adapted to the environmental temperature detected by the temperature sensor. 2. The artificial neural network circuit according to claim 1, wherein the learning value is updated to a learning value corresponding to changes in environmental temperature. The crossbar circuit is provided with a plurality of input bars for inputting a bias signal so that weights adapted to different environmental temperatures can be applied,
The update unit selects an input bar with a weight adapted to the environmental temperature detected by the temperature sensor from a plurality of input bars for inputting the bias signal, and applies the bias signal only to the selected input bar. 2. The artificial neural network circuit according to claim 1, wherein the learning value used in said crossbar circuit and/or said processing circuit is updated to a learning value corresponding to a change in ambient temperature. A temperature range other than the temperature at which the learned value initially used in the crossbar circuit and/or the processing circuit is divided into a plurality of temperature ranges, and the learned value is used in each of the plurality of temperature ranges. 7. An artificial neural network circuit as claimed in any one of claims 1 to 6, comprising a plurality of training values for each. 7. The learned value includes one learned value used in a temperature range other than the temperature at which the learned value initially used in the crossbar circuit and/or the processing circuit is learned. 2. The artificial neural network circuit according to item 1. One learning value used in a temperature range other than the temperature at which the learning value initially used in the crossbar circuit and/or the processing circuit is learned under a relatively high temperature environment within the temperature range. 9. The artificial neural network circuit according to claim 8, wherein the learning value is learned by. a crossbar circuit (20) for transmitting signals between hierarchized neurons of an artificial neural network;
The crossbar circuit is arranged so that a plurality of input bars (21) and a plurality of output bars (22, 23) intersect, and at each intersection of the plurality of input bars and the plurality of output bars, , a memristor (24), which is a resistance change type memory that gives weight to the transmitted signal, is provided,
an artificial neural network circuit comprising: processing circuits (27a, 27b) for performing, as signal processing in the neuron, at least calculation of the sum of signals respectively weighted by the memristors and flowing into the respective output bars; A learning value switching method applied to
detecting the ambient temperature of the crossbar circuit using a temperature sensor (33);
An update unit (31, 34, 35) updates the learning value used in the crossbar circuit and/or the processing circuit according to the change in the environmental temperature of the crossbar circuit. A learning value switching method in an artificial neural network circuit, including updating to a value. Bias signals are input to the plurality of input bars of the crossbar circuit in addition to the signals transmitted between the neurons,
The learning values used in the crossbar circuit and/or the processing circuit include at least one of a weight given to a signal to be transmitted, a weight given to a bias signal, and a magnitude of the bias signal. 11. The learning value switching method in the artificial neural network circuit according to claim 10. the processing circuitry is configured to add a weighted bias signal to the summed signals flowing into each of the output bars;
11. The learning value switching method in an artificial neural network circuit according to claim 10, wherein learning values used in said crossbar circuit and/or said processing circuit include said weighted bias signal. The updating unit rewrites the weights of the memristor of the crossbar circuit in accordance with a change in the environmental temperature of the crossbar circuit, thereby changing the learning value used in the crossbar circuit and/or the processing circuit. 11. The learning value switching method in an artificial neural network circuit according to claim 10, wherein the learning value is updated to a learning value corresponding to a change in environmental temperature. A plurality of the crossbar circuits are provided so as to apply weights adapted to different environmental temperatures to the same input signal,
The updating unit selects from the plurality of crossbar circuits a crossbar circuit having a weight adapted to the environmental temperature detected by the temperature sensor, and thereby used in the crossbar circuit and/or the processing circuit. 11. The learning value switching method in an artificial neural network circuit according to claim 10, wherein the learning value is updated to a learning value corresponding to a change in environmental temperature. The crossbar circuit is provided with a plurality of input bars for inputting a bias signal so that weights adapted to different environmental temperatures can be applied,
The update unit selects an input bar with a weight adapted to the environmental temperature detected by the temperature sensor from a plurality of input bars for inputting the bias signal, and applies the bias signal only to the selected input bar. 11. The learning value switching method in an artificial neural network circuit according to claim 10, wherein the learning value used in said crossbar circuit and/or said processing circuit is updated to a learning value corresponding to a change in environmental temperature. A temperature range other than the temperature at which the learned value initially used in the crossbar circuit and/or the processing circuit is divided into a plurality of temperature ranges, and the learned value is used in each of the plurality of temperature ranges. 16. The learning value switching method in an artificial neural network circuit according to any one of claims 10 to 15, wherein the learning value switching method includes a plurality of learning values. 16. The learning value according to any one of claims 10 to 15, wherein the learning value includes one learning value used in a temperature range other than the temperature at which the learning value initially used in the crossbar circuit and/or the processing circuit is learned. A learning value switching method in the artificial neural network circuit according to item 1. One learning value used in a temperature range other than the temperature at which the learning value initially used in the crossbar circuit and/or the processing circuit is learned under a relatively high temperature environment within the temperature range. 18. The learning value switching method in an artificial neural network circuit according to claim 17, wherein the learning value is a learning value learned in .

Images