JPH06175999A - Neurocomputer - Google Patents

Abstract

PURPOSE:To provide a neurocomputer with the ability comparable to an 8-bit microcomputer capable of handling nonlinear functions at low cost which suppressing the increase of a memory capacity. CONSTITUTION:Synapse connecting coefficients, biased values and the nonlinear functions are stored in a first storage means, neuron data are stored in a second storage means and execution calculation is performed. Thus, nonlinear output can be obtained by small memory amount and the nonlinear functions can be handled at low cost while suppressing the increase of the memory capacity with the ability of the 8-bit microcomputer approximately. Also, the synapse connecting coefficients are stored in a ROM 6 at every respective block, the miniaturization of a program capacity is enabled and an execution calculation speed can be accelerated and made fit for practical use.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a neuro computer for controlling home electric appliances such as air conditioners.

[0002]

2. Description of the Related Art In recent years, calculation processing by a neural network has been proposed in which a neural network of a living body is simplified into an engineering model. As shown in FIG. 7, the perceptron type neural network is composed of three layers of an input layer, an intermediate layer and an output layer. The neurons in the input layer, the intermediate layer, and the output layer are connected by synaptic connections, and each neuron has an input unit from another neuron, a conversion unit that processes input data according to a predetermined rule, and a conversion result. Consists of an output section that outputs to the upper layer neuron. Then, a synaptic coupling coefficient indicating the strength of the coupling is added to the synaptic connection with each neuron, and a bias value is added to the neurons in the intermediate layer and the output layer. By changing these synapse coupling coefficient and bias value according to the back propagation learning rule, a neural network that obtains a desired output with respect to an input signal is constructed.

FIG. 15 shows an example of an apparatus provided with such neural network calculation processing as disclosed in Japanese Patent Laid-Open No. 3-9.
The block diagram of the information processing apparatus and the air conditioning apparatus using the information processing apparatus disclosed in Japanese Patent No. 1646 is shown.

This air conditioner includes a sensor unit 100, A
It is composed of a D converter 102, a neural network type information processing unit 104, a control circuit 107, and an air conditioning unit 109.

Next, the operation will be described.

The sensor section 100 converts the indoor environment information into an analog signal 101 and sends it to the AD converter 102. Then, the AD converter 102 converts the analog signal 101 sent from the sensor unit 100 into a digital signal 103. It is assumed that the neural network type information processing unit 104 has already learned the correlation between the temperature, the humidity, etc. and the user by inputting the teacher signal 105 and the digital signal 103 in advance. After the information processing is performed in the neural network type information processing unit 104, the output signal 106 changes to the control circuit 1
07, and the control signal 108 is output. Further, the air conditioning unit 109 is controlled by the control signal 108.

[0007]

As described above, in the conventional neural network, after the neural learning is performed by using the computer (offline development device), the learning result, the synapse coupling coefficient, and the bias value are micro-calculated. Since the data is stored as a table in the computer, when a network with a large synaptic coupling coefficient is constructed, the value handled by the microcomputer may be exceeded and only an output different from the offline learning result may be obtained. Further, when the learning result is changed by the microcomputer in response to the new teacher data (online: the user gives the teacher data), it is difficult to program the backpropagation algorithm in the microcomputer because the calculation process is complicated. , It is necessary to relearn by using all the teacher data instead of learning only the teacher data to be added. Therefore, it takes a long time to learn, and when the neural network converges to the local optimal solution by the new learning, the neural network There was a possibility that the performance of would be worse than before learning.

The present invention has been made in order to solve the above-mentioned problems, and can be easily constructed on a microcomputer, and since the constraint of the handling value is eliminated at the time of the neural network execution processing calculation, an accurate neural network can be obtained. The purpose of this study is to obtain a neuro-computer that can obtain output from the network and can be used when new teacher data is input in additional learning.

[0009]

In order to achieve such an object, in the neurocomputer according to the invention of claim 1, an input layer, at least one intermediate layer, and an output layer are hierarchically connected in order by synapses. Each phase includes a neuron that performs a product-sum operation of the input signal from the external or the previous layer and the synapse coupling coefficient to perform non-linear conversion and outputs, and performs learning so that a predetermined output signal can be obtained in the input layer and the synapse A main controller composed of a neural network for setting a coupling coefficient, the synapse coupling coefficient, a bias value,
And a first storage means for storing the nonlinear function as a table, a second storage means for storing each neuron value including an input value, and an execution calculation process for blocking the synapse coupling coefficient and the bias value stored in the first storage means. Means and connection number information means for connecting between the blocks are provided, and the adaptive layer information of the neuron value of the second storage means is judged using a bias neuron.

In the neurocomputer according to the second aspect of the present invention, the input layer, at least one intermediate layer, and the output layer are hierarchically connected by a synapse in order, and each phase is connected to an input signal from the external or front layer. Includes a neuron that performs a product-sum operation of synaptic coupling coefficients, performs non-linear conversion, and outputs
A main controller composed of a neural network that performs learning so as to obtain a predetermined output signal in the input layer and sets a synapse coupling coefficient; and a first controller that stores the synapse coupling coefficient, the bias value, and the non-linear function as a table. 1
Storage means, second storage means for storing each neuron value including an input value, and whether or not the synapse coupling coefficient or bias value exceeds a predetermined condition, and when the predetermined condition is exceeded, the value is stored in the first storage means. And the synapse coupling coefficient and the bias value stored in the first storage means are divided into blocks, the blocks are connected by the number-of-connections information means, and the adaptive layer information of the neuron value in the second storage means is stored in the bias neuron. And an execution processing unit for making a judgment using the above, and at the time of execution calculation, a plurality of links between neurons are connected in parallel according to a signal from the judgment unit for calculation.

Further, in the neurocomputer according to the third aspect of the present invention, an input layer, at least one intermediate layer, and an output layer are hierarchically connected in order by synapses, and each phase is connected to an input signal from an external or previous layer. Includes a neuron that performs a product-sum operation of synaptic coupling coefficients, performs non-linear conversion, and outputs
A main controller composed of a neural network that performs learning so as to obtain a predetermined output signal in the input layer and sets a synapse coupling coefficient; and a first controller that stores the synapse coupling coefficient, the bias value, and the non-linear function as a table. 1
The storage unit includes: a storage unit; a determination unit that determines whether the synapse coupling coefficient or the bias value exceeds a predetermined condition; and an execution processing unit that calculates a number of links between neurons connected in series according to the result of the determination unit. It is characterized by that.

A neurocomputer according to a fourth aspect of the present invention comprises an input layer, at least one intermediate layer,
And the output layers are hierarchically connected in order by synapses, and each phase includes a neuron that performs a product-sum operation of an input signal from the external or the previous layer and a synapse coupling coefficient to perform non-linear conversion and output, and a predetermined output signal is input to the input layer. A main controller configured by a neural network that performs learning so as to set the synapse coupling coefficient, and the synapse coupling coefficient,
First storage means for storing a bias value and a non-linear function as a table, second storage means for storing each neuron value including an input value, and synapse coupling coefficient and bias value stored in the first storage means are blocked. Execution calculation processing means, connection number information means for connecting between blocks, detection means for detecting that a learning mode has been entered during additional learning, third storage means for storing new teacher data, and new teacher data output The calculation means for calculating the change amount for changing the execution result from the output of the original neural network execution calculation, the fourth storage means for storing the change amount calculated by the calculation means, and the same signal as the new teacher data is input. Change means for changing the bias value of the output layer neuron only when the bias value is changed. It is characterized in that the computing process are.

According to a fifth aspect of the present invention, in the neurocomputer, an input layer, at least one intermediate layer, and an output layer are hierarchically connected by a synapse, and each phase is connected to an input signal from an external or previous layer. Includes a neuron that performs a product-sum operation of synaptic coupling coefficients, performs non-linear conversion, and outputs
A main controller composed of a neural network that performs learning so as to obtain a predetermined output signal in the input layer and sets a synapse coupling coefficient; and a first controller that stores the synapse coupling coefficient, the bias value, and the non-linear function as a table. 1
A storage means, a second storage means for storing each neuron value including an input value, a detection means for detecting that a learning mode has been entered during additional learning, and a third storage means for storing new teacher data.
Storage means, calculation means for calculating the amount of change for changing the execution result from the output of the new teacher data and the output of the original neural network execution calculation, and fourth storage means for storing the amount of change calculated by the calculation means, A change means for calculating a distance between an input signal and the new teacher data input during execution calculation and changing the change amount according to the distance; and an execution calculation processing means for performing a calculation process based on the changed change amount. It is characterized by.

Further, in the neurocomputer according to the present invention as defined in claim 6, the input layer, at least one intermediate layer, and the output layer are hierarchically connected in order by synapses, and each phase is connected to an input signal from an external or previous layer. Includes a neuron that performs a product-sum operation of synaptic coupling coefficients, performs non-linear conversion, and outputs
A main controller composed of a neural network that performs learning so as to obtain a predetermined output signal in the input layer and sets a synapse coupling coefficient; and a first controller that stores the synapse coupling coefficient, the bias value, and the non-linear function as a table. 1
A storage means, a second storage means for storing each neuron value including an input value, a detection means for detecting that a learning mode has been entered during additional learning, and a third storage means for storing new teacher data.
Fourth storage means for storing a storage means, a synapse connection coefficient change amount calculation means for calculating a change amount of a synapse connection coefficient connected to an output layer neuron with new teaching data, and a change amount calculated by the synapse connection coefficient change amount calculation means When,
And an execution calculation processing unit that performs calculation processing based on the calculated synaptic coupling coefficient change amount.

[0015]

In the neurocomputer according to the first aspect of the present invention, the synapse coupling coefficient, the bias value, and the nonlinear function are stored in the first storage means, the neuron data is stored in the second storage means, and the execution calculation is performed. Non-linear output can be obtained with a small amount of memory.

In the neurocomputer according to the second aspect of the present invention, among the synapse coupling coefficient and the bias value, those exceeding a predetermined condition are divided into a plurality of memories and stored in the first storage means. Since the discriminating means recognizes that the data are divided and stored and a plurality of links between neurons are connected in parallel to perform the execution calculation process, a large synaptic coupling coefficient can be handled without lowering the resolution.

Further, in the neurocomputer according to the third aspect of the present invention, the synapse coupling coefficient and the bias value that exceed a predetermined condition are compressed and stored in the first storage means, and are compressed at the time of execution calculation. The stored data is recognized by the discriminating means, and the execution value is calculated by restoring the original value by connecting multiple links between neurons in series. Therefore, there is no need to increase the memory capacity or change the calculation algorithm in a complicated manner. The coupling coefficient can be handled.

The neurocomputer according to the fourth aspect of the present invention detects the learning mode, stores new teacher data in the third storage means, calculates the amount of change in the bias value, and stores it in the fourth storage means. . At the time of execution calculation, since the execution calculation processing is performed using the above-mentioned amount of change only when new teacher data is input to the input, it is possible to perform additional learning while making the most of the offline learning results obtained so far.

Further, the neurocomputer according to the invention of claim 5 detects the learning mode, stores new teacher data in the third storage means, calculates the change amount of the bias value, and changes the change amount according to the value. The neural network is generalized to the new teacher data in order to perform the execution calculation process.

Further, the neurocomputer according to the invention of claim 6 detects the learning mode, stores new teacher data in the third storage means, calculates the change amount of the synapse coupling coefficient, and stores it in the fourth storage means. To do. At the time of execution calculation, since the execution calculation process is performed using this change amount, the change amount can be suppressed to a smaller value than that when only the bias value is changed.

[0021]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of a neurocomputer according to an embodiment of the present invention. For ease of explanation, a case where this neurocomputer is used in an air conditioner will be described as an example.

As shown in FIGS. 1 and 9, the neurocomputer has a main control unit 1 which is composed of a microcomputer and constitutes a neural network. The main control unit 1 has a CPU 2 for performing calculation and control. I have it. Then, in the CPU 2, the execution processing means 10, the discrimination means 12, the bias value changing means 16, and the change amount calculating means 15 are provided.
, A switch 3 for inputting temperature setting, air volume, learning, etc., each sensor 4 for detecting room temperature, air volume, wind direction, outside air temperature, etc., and indicator 5 for displaying room temperature, operation mode, etc.
And the memory 9 are connected. Further, the memory 9 is
The ROM 6 which is the first storage means 11 and the second storage means 13
The RAM 7 and the third storage means 14 and the EEPROM 8 which is the fourth storage means 17,
M6 has a sigmoid function table (SIG TB
L), a synapse coupling table (WTBL), and a serial information table (SER TBL) are stored, and the RAM 7 has an input signal from the switch 3 and the sensor 4 and a value obtained by multiplying the input signal by a synaptic coupling coefficient. And a value (N RAM) obtained by performing non-linear conversion of the total sum by sigmoid function conversion, that is, a value of each neuron is stored. Further, the EEPROM 8 stores teacher data (T TBL) described later and a change amount (D WGT) obtained by learning calculation.

In this embodiment, a sigmoid function is used as the non-linear function. The sigmoid function is a function indicated by a curve (see FIG. 2) in which the positive region and the negative region are point-symmetric. In order to simplify the microcomputer processing, the input section is -8 to +
It is cut off at 8. This is because an input exceeding ± 8 has almost no effect on the output. When the sigmoid function is represented by 1 byte, the correspondence between the bit and the value is as shown in FIG.

In addition, the synapse connection table (W TB
L) is a table of synapse coupling coefficients obtained by learning, as shown in FIG. As shown in FIG. 6, inside the microcomputer, the value between −8 and +8 is expressed by 8 bits (00H to FFH), and the resolution is 0.
Since ~ 8 is allocated to 128 (7 bits), the result is as follows.

8/128 = 0.0625 Note that bit 7 is used as a sign bit in order to express a minus expression. Synapse connection table (WT
BL) stores the coupling coefficient information for each block. The coupling coefficient information is 3 bytes in the case of blocks, but is 2 bytes in the final block, and the block end code, output storage address, It consists of the next NRAM pointer. The block end code has two codes, FEH and FFH, and FEH is
The end code of the normal block is shown, and FFH shows the end code of the last block. Also, the block end code is
The final address of the synapse coupling coefficient table of each block is shown, and the execution program of the neural network determines the end of the synapse coupling table when FEH and FFH are detected. Therefore, the range of the synaptic coupling coefficient is set to 00H to FDH. When handling FEH and FFH as true values, they are rounded to FDH.

The output storage address indicates an address for storing the calculation result of the block (the offset address from the start address of the corresponding N RAM is stored). The next block N RAM pointer indicates the start address of the N RAM corresponding to the next block.

When the neuron N RAM has a three-layer structure as shown in FIG. 4, it is divided into an input layer, an intermediate layer, and an output layer, and is partitioned by a bias value neuron θ. Since the bias value neuron always has the value 1 (FFH), the end of the layer can be determined by FFH. Therefore, the neuron value is 1
If it is (FFH), it is also rounded to FEH and input. Inside the microcomputer, 0 to 1
The values up to are represented by 8 bits and have the following resolutions.

1/256 = 0.039 Next, the operation of this embodiment will be described with reference to the flowchart of FIG. 6 and FIG.
This will be described with reference to the hierarchical structure diagram of FIG. The neural network is generally composed of an input layer, an intermediate layer, and an output layer. In this embodiment, as shown in FIG. 7, each sensor 4 is an input layer neuron, and each sensor 4 to C
Input signals x1 to xn to PU2 correspond to inputs from the input layer to the intermediate layer.

The CPU 2 receives the input signal x from each sensor 4.
Initial processing such as storing 1 to xn as data in a predetermined NRAM on the RAM 7 is performed (step 1). Then, after performing the initial processing, the sum of products operation is performed (step 2), and the N RAM pointer and the WTBL pointer are incremented (step 3). Then, it is judged whether or not the block end code appears (step 4). If the block end code does not appear, it is judged whether or not the final block end code appears (step 8). If the final block end code does not appear, Repeat steps 2 and 3 until the block end code appears. When the block end code appears, the sigmoid function conversion is performed (step 5), and a predetermined NR on the RAM 7 is obtained.
The value is stored in AM (step 6) and the next block N RA
Set the M pointer and return to step 2. When it is determined in step 8 that the final block end code appears, sigmoid function conversion is performed (step 5), the value is stored in a predetermined NRAM on the RAM 7 (step 6), and post-processing of output is performed. (Step 9).

The sigmoid function conversion executed in step 5 will be described with reference to the flow chart of FIG.

The CPU 2 reads the value stored in the WORK area of the RAM 7 (step 10) and performs absolute value processing on the value (step 11). Furthermore, an output is obtained by referring to the sigmoid function table (SIG TBL) (step 12), and it is determined whether the sign of the value read from the WORK area is correct (step 13). When it is determined that the sign is positive, the sigmoid function conversion process is ended, and when it is determined that the sign is negative, the value obtained by subtracting 1 from the obtained output is output (step 14), and the sigmoid function conversion process is ended. To do.

In the above embodiment, when the synapse connection value or the bias value exceeds -8 to +8, it cannot be processed by the microcomputer. In this case, if the values are rounded and stored, the output of the neural network may deviate significantly from the true value. Therefore, the operation of the software proposed in the inventions of claims 2 and 3 so that the output of the neural network does not greatly change from the true value when the synapse connection value or the bias value exceeds -8 to +8 will be described.

As shown in FIG. 9, first the first storage means 11
The value read from the W TBL of the discriminator 12 is −8 to +.
It is discriminated whether or not it is within the range of 8, and execution calculation is carried out by the execution calculation processing means.

Then, as shown in FIG.
Performs initial processing such as storing the input signals x1 to xn from the respective sensors 4 as data in a predetermined N RAM on the RAM 7 (step 1). Then, after performing the initial processing, the sum of products operation is performed (step 2), and W T
If BL is the maximum value in the synapse connection table (the data on the microcomputer is FDH or 7FH in FIG. 11) (step 15), the N RAM pointer remains unchanged and W T
BL is updated (step 16), and the same product-sum calculation as in step 2 is performed. If the W TBL is not the maximum value, the N RAM pointer and the W TBL pointer are incremented (step 3) and repeated until the block end code appears.

When the synapse coupling coefficient exceeds -8 or +8, the true value of the synapse coupling coefficient is the sum of the synapse coupling coefficient stored at the next address. The values of the synaptic coupling coefficient of blocks 15-13 in FIG.
FH and 20H), but the true value of the synaptic coupling coefficient represents the sum of 7FH and 20H, that is, 8 + 2 = 10. There are 16 WORK areas in RAM7.
Since the sum of products is processed by bits, there is almost no possibility of overflow.

In the above-described embodiment, the value of the synapse coupling coefficient above -8 or +8 is stored separately in two areas. However, as in the invention according to claim 3, for example, a value half the true value is stored. May be stored in one area and doubled at the time of calculation. In this case, the operation is an operation in which steps 17 and 18 shown in FIG. 11 are added between steps 2 and 3 of the flowchart of FIG. Step 17
In, when the coupling coefficient counter matches the value of the serial information table SIG TBL stored in the ROM 6,
In step 18, W TBL is half the true value and is therefore doubled. The coupling coefficient counter increments the lower 4 bits each time it passes through step 2,
It is assumed that the upper 4 bits are incremented each time when passing through.

In the serial information table, as shown in FIG. 12, the address of the WTBL value to be doubled is written, the upper 4 bits indicate which block, and the lower 4 bits are the block head. From the W TB
Indicates L. As a result of such an operation, in the WORK area, the input data _{x1 to} xn and the synapse coupling coefficient Wx1 _-y1 .
The product-sum operation result with W _xn-y1 is stored. W at this time
The data on the ORK area has a value represented by the following equation.

X1 × W _x1-y1 + x2 × W _w2-y1 + ... + x
n × W _xn-y1 + 1.0 × W _bias-y1 However, when the hidden layer neuron does not include a bias neuron, the equation is obtained by removing 1.0 × W _bias-y1 from the above equation.

The value obtained by the above equation is subjected to sigmoid function processing referring to the sigmoid function table SIG TBL stored in the ROM 6 (step 5).

In this way, a neural network having the same output as the offline learning result can be obtained even when the synapse coupling coefficient or the bias value exceeds -8 to +8. As described above, according to the invention described in claim 2, the processing does not change the structure of the memory shown in FIG.
The described invention is a process that does not change the flow of the flowchart shown in FIG.

Next, the operation of the simple learning section in the embodiment of the invention described in claim 4 will be described with reference to FIGS. 9 and 13. It should be noted that when the user is dissatisfied with the output value of the learned neural network, the simple learning is to partially change the output value and tune to a home electric appliance suitable for the user.

As shown in FIG. 9, when the teacher data 20 is input, it is stored in the third storage means 14. On the other hand, the output 19 is calculated from the execution processing means 10, and the change amount calculation means 15 is calculated.
Calculate the change amount with. Then, the bias value changing means 16
To change the bias value and store it in the fourth storage means 17. The neural network uses the first storage means 11, the second storage means 13, and the fourth storage means 17 to perform the execution process calculation.

As shown in FIG. 13, the learning section operates when a learning start signal is input from the switch 3. CPU
2 is the input signal from each sensor 4 and the output Z ^t to be changed
Is stored in the EEPROM 8 as teacher data (step 19). Then, the CPU 2 updates the value, that is, calculates the address of the bias value for performing learning (step 20). Output value Z of the neural network at that time
Is calculated according to the flowchart of FIG. 6 or 10 (step 21), and the bias value is rewritten (step 22). The learning switch 3 is used to inform the CPU 2 that the learning mode has been entered, and at the same time, to take in the output Z ^t to be changed.

Although the learned bias value change amount is stored in the RAM 7 in the above embodiment, EE
It may be stored in the PROM 8. In this case, the data (learning result) is retained even when the power is turned off.

In the above embodiment, step 2
The change amount of the bias value stored in 2 is represented by the following equation that does not depend on the input from each sensor 4.

G (Z ^t -Z) Here, the learning coefficient g is a positive constant. The larger g, the larger the amount of change. The designer decides this value by performing a simulation in advance. In this way, the output value of the neural network can be changed by a fast and simple algorithm.

As proposed in the invention of claim 5, the change amount of the bias value stored in step 15 is
You may do as follows.

G (Z ^t -Z) / [1+ (x1-x1 ^t )
^{2 + (x2-x2 t)} 2 + ... + (xn-xn t) 2] ^{Here, x1 t, x2 t, ...} , xn t is, step 19
Inputs (constants) from the sensors 4 at the time of learning, which are taken in in, x1, x2, ..., Xn are input signals (variables) from the sensors 4 at the time of execution calculation. Therefore, the above equation changes depending on x1, x2, ..., Xn. That is, when new teacher data is input and learning is performed, the bias value of the output layer neuron changes according to the input signal from each sensor 4. Therefore, the neural network generalizes in a more natural manner corresponding to the new teacher data. Besides, the execution calculation processing time becomes long.

Further, in the above embodiment, the bias value is changed by learning, but as proposed in the invention of claim 6, the synapse connection value may be changed by learning instead of the bias value. . If you do this,
The operation of simple learning is as shown in FIG.

The CPU 2 uses the input signal from each sensor 4 and the output Z ^t to be changed as teacher data in the EEPROM 8
(Step 19). Then, CPU2
The value is updated, that is, the address of the bias value for learning is calculated (step 20). The output value Z of the neural network at that time is calculated according to the flowchart of FIG. 6 or FIG. 10 (step 21) and is calculated in step 23 in the same manner as in step 22 above, but all synapse coupling coefficients connected to Z (in FIG. 7). W _y1-z , W
_{Since y2-z} , ..., W _ym-z ) are targeted, the difference is that the amount of change is divided by the sum m of the synaptic coupling coefficients to be changed.

By doing so, the amount of change of each synapse can be kept small. However,
The calculation steps are increased.

[0053]

As described above, according to the first aspect of the present invention, the synapse coupling coefficient, the bias value, and the non-linear function are stored in the first storage means, and the neuron data is stored in the second storage means. Since it is configured to perform execution calculation,
With the capability of an 8-bit microcomputer, it is possible to handle nonlinear functions at low cost while suppressing an increase in memory capacity.

According to the second aspect of the present invention, among the synapse coupling coefficient and the bias value, those exceeding a predetermined condition are divided into a plurality of memories and stored in the first storage means, and at the time of execution calculation, Since the discriminating means recognizes that the data is divided and stored, and a plurality of links between neurons are connected in parallel to perform the execution calculation processing, it is possible to handle a large synaptic coupling coefficient without lowering the resolution.

Furthermore, according to the third aspect of the present invention, the synapse coupling coefficient and the bias value that exceed a predetermined condition are compressed and stored in the first storage means, and are compressed and stored during execution calculation. By doing so, the discrimination means recognizes it, and by connecting multiple links between neurons in series, it is configured to restore the original value and perform the execution calculation processing, so it is possible to increase the memory capacity or change the calculation algorithm in a complicated manner. And can handle large coupling coefficients.

According to the fourth aspect of the invention, the learning mode is detected, new teacher data is stored in the third storage means, the change amount of the bias value is calculated, and stored in the fourth storage means. At the time of execution calculation, since the execution calculation processing is performed using the above-mentioned change amount only when new teacher data comes in the input, when the new teacher data is inputted, the bias value can be changed accordingly. , It is possible to perform additional learning with new teacher data while making the most of the offline learning results so far.

According to the invention of claim 5, the learning mode is detected, new teacher data is stored in the third storage means, the change amount of the bias value is calculated, and the change amount is changed according to the value. However, because it is configured to perform execution calculation processing, when new teacher data is input, the bias value of the output layer neuron changes according to the value of the teacher data, and changes accordingly for inputs near the new teacher data. The obtained output can be obtained, and the neural network can be generalized to new teacher data.

Further, according to the invention of claim 6, the learning mode is detected, new teacher data is stored in the third storage means, the change amount of the synapse coupling coefficient is calculated and stored in the fourth storage means. Since the execution calculation process is performed using this change amount during the execution calculation, the change value is distributed to a plurality of synapse coupling coefficients and the change value of each coupling coefficient is set to a relatively small value. Can be suppressed.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a neurocomputer according to the present invention.

FIG. 2 is a diagram showing a sigmoid function according to an embodiment of the present invention.

FIG. 3 is a diagram showing a relationship between a state in a microcomputer of a sigmoid function table SIG TBL and a real value according to an embodiment of the present invention.

FIG. 4 is a diagram showing a data structure of a memory according to the present invention.

FIG. 5 is a diagram showing a relationship between a state of a coupling coefficient W TBL in the microcomputer and a real value according to the embodiment of the present invention.

FIG. 6 is a flowchart showing the operation of the present invention.

FIG. 7 is a diagram showing a neuro execution calculation according to the present invention.

FIG. 8 is a flowchart showing a sigmoid function conversion process according to the embodiment of the present invention.

FIG. 9 is a block diagram showing a configuration of a main control device according to the present invention.

FIG. 10 is a coupling coefficient of −8 to +8 according to the embodiment of the present invention.
It is a flow chart which shows execution calculation when it exceeds.

FIG. 11 is a flowchart when the links between neurons according to the present invention are serial connections.

FIG. 12 is a diagram showing a serial information table according to the embodiment of the present invention.

FIG. 13 is a flowchart showing a bias value simple learning process according to the present invention.

FIG. 14 is a flowchart showing a synapse coupling coefficient simple learning process according to the present invention.

FIG. 15 is a block diagram showing a configuration of a conventional device.

[Explanation of symbols]

 1 Main Controller 2 CPU 3 Switch 4 Each Sensor 5 Display 6 ROM 7 RAM 8 EEPROM 9 Memory 10 Execution Processing Means 11 First Storage Means 12 Discrimination Means 13 Second Storage Means 14 Third Storage Means 15 Change Amount Calculation Means 16 Bias value (synapse coupling coefficient) changing means 17 Fourth storage means

Claims (6)

[Claims] 1. An input layer, at least one intermediate layer, and an output layer are hierarchically connected in order by synapses, and each phase is subjected to a product-sum operation of an input signal from an external or previous layer and a synapse coupling coefficient to perform non-linear conversion. And a neuron that includes a neuron that outputs a synapse coupling coefficient and performs learning so that a predetermined output signal can be obtained in the input layer, and the synapse coupling coefficient, the bias value, and the nonlinearity. First storage means for storing the functions as a table, second storage means for storing each neuron value including an input value, and execution calculation processing means for blocking the synapse coupling coefficient and the bias value stored in the first storage means. , The number-of-connections means for connecting between the blocks, and the adaptive layer information of the neuron values of the second storage means by using the bias neurons. Neuro-computer, characterized in that the cross-sectional. 2. An input layer, at least one intermediate layer, and an output layer are hierarchically connected by a synapse in order, and each phase is subjected to a product-sum operation of an input signal from an external or previous layer and a synapse coupling coefficient to perform non-linear conversion. And a neuron that includes a neuron that outputs a synapse coupling coefficient and performs learning so that a predetermined output signal can be obtained in the input layer, and the synapse coupling coefficient, the bias value, and the nonlinearity. First storage means for storing the function as a table, second storage means for storing each neuron value including the input value, and whether or not the synapse coupling coefficient or bias value exceeds a predetermined condition, and the predetermined condition is exceeded. Sometimes, the discriminating means for storing the value in the first storing means and the synapse coupling coefficient and the bias value stored in the first storing means are divided into blocks, And an execution processing unit that determines the matching layer information of the neuron value of the second storage unit by using the bias neuron by connecting the numbers by means of the connection number information unit. A neuro-computer characterized by connecting multiple links in parallel. 3. An input layer, at least one intermediate layer, and an output layer are hierarchically connected by a synapse in order, and each phase is subjected to a product-sum operation of an input signal from an external or previous layer and a synapse coupling coefficient to perform non-linear conversion. And a neuron that includes a neuron that outputs a synapse coupling coefficient and performs learning so that a predetermined output signal can be obtained in the input layer, and the synapse coupling coefficient, the bias value, and the nonlinearity. First storage means for storing a function as a table, determination means for determining whether a synapse coupling coefficient or a bias value exceeds a predetermined condition, and a plurality of links between neurons connected in series according to the result of the determination means A neurocomputer comprising: an execution processing unit that executes: 4. The input layer, at least one intermediate layer, and the output layer are hierarchically connected by synapses in order, and each phase is subjected to a product-sum operation of an input signal from an external or previous layer and a synapse coupling coefficient to perform non-linear conversion. And a neuron that includes a neuron that outputs a synapse coupling coefficient and performs learning so that a predetermined output signal can be obtained in the input layer, and the synapse coupling coefficient, the bias value, and the nonlinearity. First storage means for storing the functions as a table, second storage means for storing each neuron value including an input value, and execution calculation processing means for blocking the synapse coupling coefficient and the bias value stored in the first storage means. , The number-of-connections information unit that connects each block, the detection unit that detects that the learning mode has been entered during additional learning, and the new teaching data A third storage means for storing, a calculation means for calculating a change amount for changing the execution result from the new teacher data output and the output by the original neural network execution calculation, and a fourth storage means for storing the change amount calculated by the calculation means. Storage means and changing means for changing the bias value of the output layer neuron only when the same signal as new teacher data is input, wherein the execution calculation processing means performs calculation processing using the changed bias value Neuro computer characterized by. 5. An input layer, at least one intermediate layer, and an output layer are hierarchically connected by a synapse in order, and each phase is subjected to a product-sum operation of an input signal from an external or previous layer and a synapse coupling coefficient to perform non-linear conversion. And a neuron that includes a neuron that outputs a synapse coupling coefficient and performs learning so that a predetermined output signal can be obtained in the input layer, and the synapse coupling coefficient, the bias value, and the nonlinearity. A first storage means for storing the function as a table, a second storage means for storing each neuron value including an input value, a detection means for detecting that a learning mode is entered during additional learning, and new teacher data are stored. A change amount for changing the execution result is calculated from the third storage means and the output of the new teacher data and the original neural network execution calculation. A fourth storage means for storing the change amount calculated by the calculation means; and a changing means for calculating a distance between an input signal and the new teacher data input during execution calculation and changing the change amount according to the distance. And an execution calculation processing means for performing calculation processing based on the changed amount of change. 6. The input layer, at least one intermediate layer, and the output layer are hierarchically connected in sequence by synapses, and each phase is subjected to a product-sum operation of an input signal from an external or previous layer and a synapse coupling coefficient to perform non-linear conversion. And a neuron that includes a neuron that outputs a synapse coupling coefficient and performs learning so that a predetermined output signal can be obtained in the input layer, and the synapse coupling coefficient, the bias value, and the nonlinearity. A first storage means for storing the function as a table, a second storage means for storing each neuron value including an input value, a detection means for detecting that a learning mode is entered during additional learning, and new teacher data are stored. Third storage means and synapse connection coefficient change amount for calculating the change amount of synapse connection coefficient connected to the output layer neuron by the new teaching data Output means, fourth storage means for storing the amount of change calculated by the synapse coupling coefficient change amount calculation means, and execution calculation processing means for performing calculation processing based on the calculated amount of change synapse connection coefficient. Neuro computer to do.