JP2536146B2 - Pattern learning device - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a pattern learning apparatus for avoiding a situation where learning stops in a feedforward type neural network even though learning has not progressed sufficiently. is there.

(Prior Art) Conventionally, the correspondence between the input pattern and the output signal is the first
Article ("Associative Memory" by T. Kohonen: Springer
-Verlag: 1977), it was performed by the pseudo-inverse matrix and multiple regression analysis, but the system shown in this document was a linear system and could not be associated nonlinearly. On the other hand, in the feed-forward type neural network, the pattern is input and the ideal output signal is input many times as a teacher, and the second document (“An Introduction to C
omputing with Neural Nets "by RPLippmann: IEEE.ASS
The back-propagation learning method shown in P: April 1987, pp. 4-22) can learn the nonlinear correspondence between the input pattern and the output signal.

(Problems to be Solved by the Invention) In the feedforward type neural network described above, the weight matrix is calculated for the vector of the input signal, and the result is multiplied by a non-linear function called a sigmoid function to obtain the result. Is obtained, the weight matrix is further calculated, and a process of multiplying by a sigmoid function is repeated for the number of layers of the neural network to obtain an output signal. This weight matrix is initially determined by random numbers and the like, and the weight matrix value is corrected by back propagation learning to obtain an optimum value.

However, in the conventional learning method, the output signal and the teacher signal do not match, and although the learning has not ended, a state in which the weight matrix is not corrected (learning stop state) occurs and the learning progresses. It may stop doing.

An object of the present invention is to provide a pattern learning device that monitors the above learning stop situation and, when learning stops, avoids learning stop and continues learning.

(Means for Solving the Problem) A pattern learning device obtained according to the first aspect of the present invention is a feedforward type neural network composed of a plurality of layers including an input layer and an output layer for performing learning by a back propagation method, A means for determining whether or not learning is stopped based on the result of calculating the sum of correction amounts of the weight matrix used in the neural network and the evaluation amount of the error between the output layer signal and the teacher signal; And means for correcting the output signal from each layer when the layer is stopped.

A pattern learning device obtained according to the second aspect of the present invention is a feedforward type neural network consisting of a plurality of layers including an input layer and an output layer for performing learning by a back propagation method, and a neural network calculated from signals of each layer. Means for determining whether learning is stopped based on the evaluation value of the correction amount of the weight matrix used in the above and the evaluation amount of the error between the output layer signal and the teacher signal, and when the learning is stopped And means for modifying the output signal from each layer.

A pattern learning device obtained according to the third aspect of the present invention is a feedforward type neural network composed of a plurality of layers including an input layer and an output for learning by a back propagation method, and a correction amount of a matrix used in the neural circuit. Means for determining whether learning is stopped based on the result of calculating the sum of the above and the evaluation amount of the error between the signal in the output layer and the teacher signal, and the signal in the input layer when the learning is stopped. It is characterized by comprising a correction part and a means for correcting the slope control parameter of the sigmoid function for obtaining the output signal in the layers other than the input layer.

A pattern learning device obtained according to the fourth aspect of the present invention is a feedforward type neural network composed of a plurality of layers including an input layer and an output for learning by a back propagation method, and a neural network calculated from signals of each layer. Means for determining whether or not learning is stopped based on the evaluation value of the correction amount of the weight matrix used in, and the evaluation amount of the error between the output layer signal and the teacher signal, and when the learning is stopped, , And means for modifying the signal of the input layer, and means for modifying the tilt control parameter of the sigmoid function for obtaining the output signal in the layers other than the input layer.

A pattern learning device obtained according to the fifth aspect of the present invention is a feedforward type neural network consisting of a plurality of layers including an input layer and an output for performing learning by a back propagation method, and a correction amount of a matrix used in the neural circuit. Means for determining whether learning is stopped based on the result of calculating the sum of the above and the evaluation amount of the error between the signal in the output layer and the teacher signal, and the signal in the input layer when the learning is stopped. It is characterized in that it comprises means for modifying, and means for modifying the range between the maximum value and the minimum value of the sigmoid function for obtaining the output signals in the layers other than the input layer.

A pattern learning device obtained according to the sixth aspect of the present invention is a feedforward type neural network composed of a plurality of layers including an input layer and an output for performing learning by a back propagation method, and a neural network calculated from signals of each layer. Means for determining whether learning is stopped based on the evaluation value of the correction amount of the weight matrix used in the above and the evaluation amount of the error between the output layer signal and the teacher signal, and when the learning is stopped In addition, it is characterized by comprising means for modifying the signal of the input layer and means for modifying the range between the maximum value and the minimum value of the sigmoid function for obtaining the output signal in the layers other than the input layer.

(Operation) In the first invention, the learning stop situation is observed when the sum of the correction amounts of the weight matrix is less than or equal to a certain value, and the difference between the output signal and the teacher signal is larger than the threshold value. By correcting the output signal from each layer and eliminating the factor that the correction amount of the weight matrix becomes 0, the weight matrix can be forcibly corrected and the learning stop can be avoided.

In the second invention, instead of examining the sum of the correction amounts of the weight matrix for the learning stop situation, the correction amount of the weight matrix is estimated based on the output value of each layer and the output value of the layer one layer below, and Observing the case where the estimated amount is equal to or less than a certain value and the difference between the output signal and the teacher signal is larger than the threshold value, learning stop can be detected without calculating the correction amount of the weight matrix. When learning is stopped, the output signal from each layer is modified so that the modification amount of the weight matrix is 0.
By eliminating the factor that causes, the weight matrix can be forcibly modified and the learning stop can be avoided.

In the third invention, the learning stop situation is observed when the sum of the correction amounts of the weight matrix is equal to or less than a certain value, and the difference between the output signal and the teacher signal is larger than the threshold value. The parameter that controls the slope of the sigmoid function that calculates the output signal in the layer is changed to make the slope gentle so that values of 0 and 1 are difficult to output, and the upper and lower limit values of the signal in the input layer are changed. By limiting the above, it is possible to correct the weight matrix and eliminate the learning stop by eliminating the factor in which the correction amount of the weight matrix becomes zero.

According to the fourth aspect of the invention, instead of checking the sum of the correction amounts of the weight matrix for the learning stop situation, the correction amount of the weight matrix is estimated based on the output value of each layer and the output value of the layer one layer below, and Observing the case where the estimated amount is equal to or less than a certain value and the difference between the output signal and the teacher signal is larger than the threshold value, learning stop can be detected without calculating the correction amount of the weight matrix. When learning is stopped, the parameter that controls the slope of the sigmoid function that calculates the output signal of layers other than the input layer is changed to make the slope gentle so that values of 0 and 1 are difficult to output, and By limiting the layer signal by the upper and lower limits,
By eliminating the factor in which the modification amount of the weight matrix becomes 0, the modification of the weight matrix is forcibly performed and the learning stop can be avoided.

In the fifth invention, the learning stop situation is observed when the sum of the correction amounts of the weight matrix is equal to or less than a certain value and the difference between the output signal and the teacher signal is larger than the threshold value. Change the range parameter that controls the maximum and minimum values of the sigmoid function that calculates the output signal of the layer to make it difficult for values of 0 and 1 to be output, and limit the signal of the input layer by the upper and lower limits. , Eliminate the factor that the correction amount of the weight matrix becomes 0 by both,
Weights The weight matrix can be modified to avoid learning outages.

According to the sixth aspect of the invention, instead of checking the sum of the correction amounts of the weight matrix for the learning stop situation, the correction amount of the weight matrix is estimated based on the output value of each layer and the output value one layer below,
Observing that the estimated amount is below a certain value and the difference between the output signal and the teacher signal is larger than the threshold value, the learning stop can be detected without calculating the correction amount of the weight matrix. When learning is stopped, the range parameter that controls the maximum and minimum values of the sigmoid function that calculates the output signals of layers other than the input layer is changed to make it difficult for values of 0 and 1 to be output, Is the upper limit value,
By limiting the weight matrix by the lower limit value, it is possible to forcibly correct the weight matrix and eliminate learning stop by eliminating the factor that causes the correction amount of the weight matrix to be zero.

(Embodiment) Next, an embodiment of the present invention will be described with reference to FIGS. 1 to 15.

1 and 10 are block diagrams showing an embodiment of the pattern learning apparatus of the first invention. The learning device of the present invention is applicable to a feedforward type pattern learning device having two or more layers. FIG. 1 illustrates a feedforward type pattern learning device having three layers, and FIG. 10 shows two layers of pattern learning. Although the apparatus will be described, it is also applicable to a pattern learning apparatus having four or more layers.

First, with reference to FIG. 10, an embodiment in the case of two layers of the pattern learning apparatus of the first present invention will be described. Learning is performed by repeatedly giving an input pattern and a teacher signal. First, a case where one input pattern and one teacher signal are given will be described. It is assumed that the matrix storage unit 1011 that stores the matrix used in the matrix product calculation unit is initialized by random numbers in advance.

First, a feedforward process of calculating an output value from an input value will be described. The input pattern is input as a vector value from the terminal 1001 and is stored in the input layer storage unit 1002. The input layer storage unit 1002 is composed of a register that holds N ₀ -dimensional vectors, and all the elements of the vectors are sent to the matrix product calculation unit 1003. Input vector values o _i
⁰ (1 ≦ i ≦ N ₀ ), and the matrix value held in the weight matrix storage unit 1011 is w _ij ¹ (1 ≦ i ≦ N ₀ , 1 ≦ j ≦
N ₁ ), the vector value held in the output layer storage unit 1005 is o ′
_{If j} ¹ (1 ≦ j ≦ N ₁ ), the matrix product operation unit 1003 Is calculated. However, here, it is the θ _j ¹ bias value and is stored in the weight matrix storage unit 1011. Of the result
^{_{o j 1 (1 ≦ j ≦}} N 1) each of the s function unit 1004 o determine the ^{_{j 1 'j 1 = {1}} + tanh (o j 1)} / 2 (2) calculation o by comprising', as an output signal It is stored in the output layer storage unit 1005.

 Next, the correction processing of the weight matrix will be described.

For the output signal obtained by the above-mentioned feedforward processing, the ideal value of the teacher signal is set to the terminal 1020.
Input from the teacher signal storage unit 1009 and hold the value t
_j (1 ≦ j ≦ N ₁ ) and the value o ′ _j ¹ in the output signal storage unit 1005 are transferred to the error signal calculation unit 1010, and the error signal d ′ _j ¹ is d ′ _j ¹ = t _j − o ′ _j ¹ (1 ≦ k ≦ N ₁ ) (3), and the back propagation error is evaluated from the output signal o ′ _j ^1, and d _j ¹ = d ′ _j ¹ × o _j ¹ × ( 1-o _j ¹ ) (4) After calculating d _j ¹ (1 ≦ j ≦ N ₁ ) by performing the equation, the value is transferred to the matrix correction unit 1013. Further, the error signal d ′ _j ¹ is calculated by the matrix correction amount calculation unit 1016.
Also used for observation of learning stop.

The matrix correction unit 1013 corrects the weight matrix by the calculation of w _ij ¹ = w _ij ¹ + a × d _j ¹ × o _i0 (5) θ _i ¹ = θ _i ¹ + b × d _j ¹ (6), and the result is It is stored again in the weight matrix storage unit 1011. Here, a and b are positive-valued coefficients, which can be set to 0.2 and 0.1, respectively, but these values themselves are not an essential problem.

In the matrix correction amount calculation unit 1016, the matrix correction unit
The correction amount of the weight matrix obtained in 1013 (a × d _j ¹ × o
_i ⁰ ) sum It is calculated by executing the following equation. Here, the calculation | x | is a calculation for obtaining the absolute value of x. Furthermore, the evaluation amount of the error between the output signal and the teacher signal is _calculated from d ′ _j ^{1 of the} result obtained by executing (3). The condition is S <Th1 (9) E> Th2, and if the condition expressed by the above formula (9) is satisfied, it is judged that the learning is stopped. Then, a signal is sent to the input layer correction unit 1017 and the output layer correction unit 1018 to correct the output signal of each layer. In the embodiment, Th1 is 0.1 and Th2 is 0.3, but this is not an essential problem.
When the above condition is not satisfied, it is considered that the learning is not stopped, and the matrix correction amount calculation unit 1016 does not output a signal to the correction unit of each layer.

The input layer correction unit 1017 that has received the learning stop signal reads out the value o _i ⁰ held in the input layer storage unit 1002, and o _i ⁰ = Min (Max (o _i ⁰ , L), U) (10 ) Is executed to correct the input value o _i ⁰ and stored again in the input storage unit 1002. As shown in FIG. 7, if the curve shown by 701 in FIG. 7 is used as a value before the correction, the correction is performed such that a portion larger than U and a portion smaller than L are respectively shown in FIG.
The value of the input layer becomes 0 by modifying like 702,703.
Avoid getting close to 1 or 1. Where Min (x, y)
Is a function that selects the smaller of x and y, and Max (x,
y) is a function that selects the larger of x and y. U represents the upper limit value of the input signal and can be set to 0.9, for example. L represents the lower limit value of the input signal, for example 0.1
Can be However, the values of U and L are not limited to 0.9 and 0.1 and may be any value. Of course, U> L. The output layer correction unit 1018 also executes the same calculation as in the above equation (10) to correct the output layer output value so as to avoid a value close to 0 or 1 and prevent a learning stop. . Then, the calculations shown in the equations (3) to (6) are executed again to realize the correction of the weight matrix without stopping the learning.

The above is a description of the learning process once, but by repeating this process, the learning can be completed without being disturbed by the learning stop.

Next, with reference to FIG. 1, an embodiment in which the pattern learning apparatus of the first aspect of the present invention has three layers will be described. Learning is performed by repeatedly giving an input pattern and a teacher signal. First, a case where one input pattern and one teacher signal are given will be described. Matrix storage units 111 and 112 for storing matrices used in the matrix product calculation unit in advance
Is initialized with a random number.

First, a feedforward process of calculating an output value from an input value will be described. The input pattern is input as a vector value from the terminal 101 and stored in the input layer storage unit 102.
The input layer storage unit 102 is composed of a register that holds N ₀ -dimensional vectors, and all the elements of the vectors are sent to the matrix product calculation unit 103. The input vector value is o _i ⁰ (1
≦ i ≦ N ₀ ) and the matrix value held in the first layer weight matrix storage unit 112 is w _ij ¹ (1 ≦ i ≦ N ₀ , 1 ≦ j ≦
N ₁ ), the vector value held in the intermediate layer storage unit 105 is represented by o ′ _j
^{If 1} (1 ≦ j ≦ N ₁ ), the matrix product calculation unit 103 Is calculated. However, θ _j ¹ is a bias value, and is stored in the first layer weight matrix storage unit 112. Results of ^{_{o j 1 (1 ≦ j ≦}} N 1) each of the s function unit 104 o determine the ^{_{j 1 'j 1 = {1}} + tanh (o j 1)} / 2 (12) calculates o by comprising', intermediate It is stored in the layer storage unit 105. The weight matrix stored in the second layer weight matrix storage unit 11 is w _jk ² (1 ≦ i ≦ N ₁ , 1 ≦ j ≦ N ₂ ),
If the bias value stored in the storage unit 111 is θ _k ² , the matrix calculation unit 106 Is calculated. Results of ^{_{o k 2 (1 ≦ k ≦}} N 2) each of the s function unit 107 o determine the ^{_{k 2 'k 2 = {1}} + tanh (o + 2)} / 2 (14) calculates o by comprising' output It is stored in the output layer storage unit 108 as a signal.

Next, the back propagation processing for correcting the weight matrix will be described.

With respect to the output signal obtained by the above-mentioned feedforward processing, the teacher signal which is an ideal value thereof is set to the terminal 102.
Input value and hold in the teacher signal storage unit 109, and the value t
_k (1 ≦ k ≦ N ₂ ) and the value o ′ _k ² in the output signal storage unit 108 are transferred to the error signal calculation unit 110, and the error signal d ′ _k ² is d ′ _k ² = t _k − o ′ _k ² (1 ≦ k ≦ N ₂ ) (15), and the back propagation error is evaluated from the output signal o ′ _k ^2, and d _k ² = d ′ _k ² × o ′ _k ² × (1−o ′ _k ² ) (16) After d _k ² (1 ≦ k ≦ N ₂ ) is calculated by performing the equation, the value is used as the matrix correction unit 113 and the error back propagation unit 1
Transfer to 14. Further, the error signal d ′ _k ² is also transferred to the matrix correction amount calculation unit 116 and used for observation of learning stop.

The matrix correction unit 113 calculates the second layer weight matrix by the calculation w _jk ² = w _jk ² + a × d _k ² × o ′ _j ¹ (17) θ _j ² = θ _j ² + b × d _k ² (18). The corrected result is stored again in the second layer weight matrix storage unit 111. Here, a and b are positive coefficient and 0.2
And 0.1 can be set, but this value is not an essential issue.

D _k ² obtained by the calculation of executing the equation (16) in the error calculation unit 110 is sent to the error back-propagation unit 114, and together with the weight matrix values transferred from the matrix storage unit 11, Is calculated, and the equation d _j ¹ = d ′ _j ¹ × o ′ _j ¹ × (1-o ′ _j ¹ ) (20) is executed from o ′ _j ¹ transferred from the intermediate layer storage unit 105. After calculating d _j ¹ (1 ≦ j ≦ N ₁ ) by the calculation, it is transferred to the matrix correction unit 115 and w _ij ¹ = w _ij ¹ + a × d _j ¹ × o _i ⁰ (21) θ _i ¹ = Θ _i ¹ + b × d _j ¹ (22) The first layer weight matrix is corrected by the calculation to execute the equation, and the result is stored again in the first layer weight matrix storage unit 112.

In the matrix correction amount calculation unit 116, the matrix correction unit 1
The correction amount (a × d _k ²
Xo ′ _j ¹ , a × d _j ¹ × o _i ⁰ ) It is calculated by executing the following equation. Here, the calculation | x | is a calculation for obtaining the absolute value of x. Furthermore, the evaluation amount of the error between the output signal and the teacher signal is calculated from d ′ _k ^{2 of the} result obtained by executing the equation (15). It is determined by executing the following formula, and the condition judgment of S <Th1 (25) E> Th2 is executed. If the condition expressed by the above formula (25) is satisfied, it is judged that the learning is stopped. Then, a signal is sent to the input layer correction unit 117, the output layer correction unit 118, and the output layer correction unit 119 to correct the output signal of each layer. When the above conditions are not satisfied, it is considered that the learning is not stopped, and the matrix correction amount calculation unit 116 does not output a signal to the correction unit of each layer.

The input layer correction unit 117, which has received the learning stop signal, reads the value o _i ⁰ held in the input layer storage unit 102, and o _i ⁰ = Min (Max (o _i ⁰ , L), U) (26 ) Is executed to correct the input value o _i ⁰ and then stored again in the input storage unit 102 to avoid that the value of the input layer becomes close to 0 or 1 as described in FIG. . Where Min
(X, y) is a function that selects the smaller of x and y,
Max (x, y) represents a function that selects the larger of x and y. The values of U and L are as described above. Also in the middle layer correction unit 118 and the output layer correction unit 119, the same calculation as in the above equation (26) is executed to correct the middle layer output value so as to avoid a value close to 0 or 1 and learn. Eliminates causing outages. Then, the calculations shown in Expressions (15) to (22) are executed again to realize the correction of the weight matrix without stopping the learning.

The above is a description of the learning process once, but by repeating this process, the learning can be completed without being disturbed by the learning stop.

Next, with reference to FIG. 11, an embodiment in which the pattern learning apparatus of the second aspect of the present invention has two layers will be described. Learning
This is done by giving an input pattern and a teacher signal repeatedly. First, a case where one input pattern and one teacher signal are given will be described. A matrix storage unit 11 that stores the matrix used in the matrix product calculation unit in advance.
11 is initialized by a random number.

First, a feedforward process of calculating an output value from an input value will be described. The input pattern is input as a vector value from the terminal 1101 and stored in the input layer storage unit 1102. The input layer storage unit 1102 is composed of a register that holds N ₀ -dimensional vectors, and all the elements of the vectors are sent to the matrix product calculation unit 1103. Input vector value o
_{i0 (1 ≦ i ≦ N 0} ) and then, the matrix value held in the weight matrix storage unit ^{_{1111 w ij 1 (1 ≦ i}} ≦ N 0, 1 ≦ j ≦
N ₁ ), the vector value held in the output layer storage unit 1105 is o ′
_{If j} ¹ (1 ≦ j ≦ N ₁ ), the matrix product calculation unit 1103 Is calculated. However, here θ _j ¹ is a bias value and is stored in the weight matrix storage unit. Result o _j ¹
(1 ≦ j ≦ N ₁ ) is calculated in the s-function unit 1104 as o ′ _j ¹ = {1 + tanh (o _j ¹ )} / 2 (28), and o ′ _j ¹ is obtained and stored in the output layer storage unit 1105. Remember.

Next, the operation of the part that determines the learning stop will be described.

With respect to the output signal obtained by the above-mentioned feedforward processing, the teacher signal, which is a rational value, is set to the terminal 1102.
Input value and hold it in the teacher signal storage unit 1109.
_j (1 ≦ j ≦ N ₁ ) and the value o ′ _j ¹ in the output signal storage unit 1105 are transferred to the error signal calculation unit 1110, and the error signal d ′ _j ¹ is d ′ _j ¹ = t _j − o ′ _j ¹ (1 ≦ j ≦ N ₁ ) (29) The result is the matrix correction amount estimation unit 11
The error evaluation amount E for determining the learning stop is sent to 16 It is obtained by executing the calculation represented by the following equation. Further, o ′ _j ¹ (1 ≦ j ≦ N ₁ ) is transferred from the output layer storage unit 1105 and o _i ⁰ (1 ≦ i ≦ N ₀ ) is transferred from the input layer storage unit 1102 to the matrix correction amount estimation unit 1116, and the matrix The estimated value S'of the correction amount is calculated by the following formula.

It is calculated by executing the following equation. Further, in the matrix correction amount estimation unit 1116, the evaluation amount of the error between the output signal and the teacher signal is subjected to the condition determination of S ′ <Th1 E> Th2 (32), and the condition represented by the above equation (32) is satisfied. If the condition is satisfied, it is determined that the learning is stopped, and a signal is sent to the input layer correction unit 1117 and the output layer correction unit 1118 to correct the output signal of each layer. If the above conditions are not satisfied, it is considered that the learning is not stopped, and no signal is output from the matrix correction amount calculation unit 1116 to the correction unit of each layer.

This determination condition considers that learning is stopped when the correction amount S of the weight matrix is close to 0 in spite of the large error evaluation amount E.

The input layer correction unit 1117, which has received the learning stop signal, reads out the value o _i ⁰ held in the input layer storage unit 1102, and o _i ⁰ = Min (Max (o _i ⁰ , L), U) (33 ) Is performed to correct the input value o _i ⁰ and store it again in the input layer storage unit 1102 to avoid the value of the input layer becoming close to 0 or 1. The output layer correction unit 1118 also has the above (33)
The calculation similar to the formula is executed to correct the intermediate layer output and the output layer output value so that the value becomes close to 0 or 1 and the learning stop is not caused.

Finally, the back propagation processing for correcting the weight matrix using the corrected values of the signal storage units 1102 and 1105 of the respective layers will be described.

Error calculator 'output signal o is held at the _j ¹ and the output signal storage unit' advance d which is determined according to equation (29) in 1110 _j ¹
The back-propagation error is evaluated from and d _j ¹ = d ′ _j ¹ × o ′ _j ¹ × (1-o ′ _j ¹ ) (34) is calculated to execute d _j ¹ (1 ≦ j ≦ N ₁ ) is obtained and then transferred to the matrix correction unit 1112, and w _ij ¹ = w _ij ¹ + a × d _j ¹ × o _i ⁰ (35) θ _i ¹ = θ _i ¹ + b × d _j ¹ (36) Modify the weight matrix by a calculation that implements
The result is stored again in the weight matrix storage unit 1111.

The above is a description of the learning process once, but by repeating this process, the learning can be completed without being disturbed by the learning stop.

FIG. 2 is a diagram showing an embodiment of a feedforward type of three layers of the pattern learning apparatus of the second invention. The pattern learning is performed by repeatedly giving an input pattern and a teacher signal. First, a case where one input pattern and one teacher signal are given will be described. The matrix storage units 211 and 212 that store the matrix used in the matrix product calculation unit are initialized in advance with random numbers.

First, a feedforward process of calculating an output value from an input value will be described. The input pattern is input as a vector value from the terminal 201 and stored in the input layer storage unit 102.
The input layer storage unit 202 is configured by a register that holds N ₀ -dimensional vectors, and all the elements of the vectors are sent to the matrix product calculation unit 203. The input vector value is o _i ⁰ (1
≦ i ≦ N ₀ ), and the matrix value held in the first layer weight matrix storage unit 212 is w _ij ¹ (1 ≦ i ≦ N ₀ , 1 ≦ j ≦
N ₁ ), the vector value held in the intermediate layer storage unit 205 is represented by o ′ _j
^{If 1} (1 ≦ j ≦ N ₁ ), the matrix product calculation unit 203 Is calculated. However, here, it is the θ _j ¹ bias value and is stored in the first layer weight matrix storage unit 212.
Results of ^{_{o j 1 (1 ≦ j ≦}} N 1) each of the s function unit 204 o determine the ^{_{j 1 'j 1 = {1}} + tanh (o j 1)} / 2 (38) calculates o by comprising', intermediate It is stored in the layer storage unit 205. The weight matrix stored in the second layer weight matrix storage unit 211 is w _jk ² (1 ≦ i ≦ N ₁ , 1 ≦ j ≦ N ₂ ), and the bias value stored in the storage unit 211 is θ _k ² . Then, in the matrix calculation unit 206, Is calculated. Results of ^{_{o k 2 (1 ≦ k ≦}} N 2) each of the s function unit 207 o determine the ^{_{k 2 'k 2 = {1}} + tanh (o k 2)} / 2 (40) calculates o by comprising' output It is stored in the output layer storage unit 208 as a signal.

Next, the operation of the part that determines the learning stop will be described.

With respect to the output signal obtained by the above-mentioned feedforward processing, the teacher signal that is an ideal value thereof is set to the terminal 220.
Input from the teacher signal storage unit 209 and hold the value t
_k (1 ≦ k ≦ N ₂ ) and the value o ′ _k ² in the output signal storage unit 208 are transferred to the error signal calculation unit 210, and the error signal d ′ _k ² is d ′ _k ² = t _k − It is calculated as o ′ _k ² (1 ≦ k ≦ N ₂ ) (41). The result is sent to the matrix correction amount estimation unit, and the error evaluation amount E for determining learning stop is calculated. It is obtained by executing the calculation represented by the following equation. Further, o ′ _k ² (1 ≦ k ≦ N ₂ ) from the output layer storage unit 208, o ′ _j ¹ (1 ≦ j ≦ N ₁ ) from the intermediate layer storage unit 205, and o _i ⁰ from the input layer storage unit 202. (1 ≦ i ≦ N ₀ ) is transferred to the matrix correction amount estimation unit 216, and the matrix correction amount estimated value S ′ is calculated by the following formula.

It is calculated by executing the following equation. Further, in the matrix correction amount estimation unit 216, the evaluation amount of the error between the output signal and the teacher signal is subjected to the condition determination S '<Th1 E> Th2 (44), and the condition represented by the above formula (44) is satisfied. If the condition is satisfied, it is determined that the learning is stopped, and a signal is sent to the input layer correction unit 217, the intermediate layer correction unit 218, and the output layer correction unit 219 to correct the output signal of each layer. If the above conditions are not satisfied, it is considered that the learning is not stopped, and no signal is output from the matrix correction amount calculation unit 216 to the correction unit of each layer.

This determination condition considers that learning is stopped when the correction amount S of the weight matrix is close to 0 in spite of the large error evaluation amount E.

The input layer correction unit 217 that has received the learning stop signal reads out the value o _i ⁰ held in the input layer storage unit 212, and o _i ⁰ = Min (Max (o _i ⁰ , L), U) (45 ) Is performed to correct the input value o _i ⁰ and then stored again in the input layer storage unit 202 to prevent the value of the input layer from becoming a value close to 0 or 1. The middle layer correction unit 218 and the output layer correction unit 219 also perform the same calculation as in the above formula (45) to correct the middle layer output value to a value close to 0 or 1 and cause learning stop. Get rid of.

Finally, the back propagation processing for modifying the weight matrix using the modified values of the signal storage units 202, 205, 208 of the respective layers will be described.

Error calculator 210 in advance expression 'output signal o is held in the _j ² and the output signal storage unit' d which is determined in accordance with (41) _k ²
Then, the back propagation error is evaluated from d _k ² = d ′ _k ² × o ′ _k ² × (1−o ′ _k ² ) (46), and the calculation is carried out to obtain d _k ² (1 ≦ k ≦ N ₂ ) Is obtained, and the obtained value is calculated by the matrix correction unit 213 and the error back propagation unit 2
Transfer to 14. Further, the error signal d ′ _k ² is also transferred to the matrix correction amount calculation unit 216 and used for observation of learning stop.

In the matrix correction unit 213, the second layer weight matrix is calculated by the calculation w _jk ² = w _jk ² + a × d _k ² × o ′ _j ¹ (47) θ _j ² = θ _j ² + b × d _k ² (48). The corrected result is stored again in the second layer weight matrix storage unit 211.

In the error calculation unit 210, d _k ² calculated by executing the equation (46), the weight matrix value transmitted to the error back propagation unit 214, and transferred from the matrix storage unit 211, Is calculated, and the equation d _j ¹ = d ′ _j ¹ × o ′ _j ¹ × (1-o ′ _j ¹ ) (50) is executed from o ′ _j ¹ transferred from the intermediate layer storage unit 205. After calculating d _j ¹ (1 ≦ j ≦ N ₁ ) by the calculation, it is transferred to the matrix correction unit 215 and w _ij ¹ = w _ij ¹ + a × d _j ¹ × o _i ⁰ (51) θ _i ¹ = Θ _i ¹ + b × d _j ¹ (52) The first layer weight matrix is corrected by the calculation to execute the equation, and the result is stored again in the first layer weight matrix storage unit 212.

The above is a description of the learning process once, but by repeating this process, the learning can be completed without being disturbed by the learning stop.

Next, with reference to FIG. 12, an embodiment of the pattern learning apparatus of the third aspect of the present invention having two layers will be described. Learning is performed by repeatedly giving an input pattern and a teacher signal. First, a case where one input pattern and one teacher signal are given will be described. The matrix storage unit 1211 that stores the matrix used in the matrix product calculation unit is initialized in advance with random numbers.

First, a feedforward process of calculating an output value from an input value will be described. The input pattern is input as a vector value from the terminal 1201 and stored in the input layer storage unit 1202. The input layer storage unit 1202 is composed of a register that holds N ₀ -dimensional vectors, and all the elements of the vectors are sent to the matrix product calculation unit 1203. The input vector value is o _i
⁰ (1 ≦ i ≦ N ₀ ), and the matrix value held in the weight matrix storage unit 1212 is w _ij ¹ (1 ≦ i ≦ N ₀ , 1 ≦ j ≦
N ₁ ), the vector value held in the output layer storage unit 1205 is set to o ′
_{If j} ¹ (1 ≦ j ≦ N ₁ ), the matrix product operation unit 1203 Is calculated. However, here θ _j ¹ is a bias value and is stored in the weight matrix storage unit. The resulting o _{j
^{1 (1 ≦ j ≦ N 1}} ) each of which obtains the _j ¹ 'o by ^{_{j 1 = {1 + tanh (}} o j 1 / U0)} / 2 (54) comprising calculation' in s function unit 1204 o, as an output signal It is stored in the output layer storage unit 1205. In the embodiment, U0 is 0.75 initially.
And

 Next, the correction processing of the weight matrix will be described.

With respect to the output signal obtained by the above-mentioned feedforward processing, the ideal value of the teacher signal is set to the terminal 1220.
Input from the teacher signal storage unit 1209 and hold the value t
_j (1 ≦ j ≦ N ₁ ) and the value o ′ _j ¹ in the output signal storage unit 1205 are transferred to the error signal calculation unit 1210, and the error signal d ′ _j ¹ is d ′ _j ¹ = t _j − o ′ _j ¹ (1 ≦ j ≦ N ₁ ) (55), and the back propagation error is evaluated from the output signal o ′ _j ¹ and d _j ¹ = d ′ _j ¹ × o ′ _j ¹ × (1-o ′ _j ¹ ) (56) After d _j ¹ (1 ≦ j ≦ N ₁ ) is obtained by the calculation for executing the equation, the value is transferred to the matrix correction unit 1213. Further, the error signal d _j ¹ is also transferred to the matrix correction amount calculation unit 1216 and used for observation of learning stop.

The matrix correction unit 1213 corrects the weight matrix by a calculation of w _ij ¹ = w _ij ¹ + a × d _j ¹ × o _i ⁰ (57) θ _i ¹ = θ _i ¹ + b × d _j ¹ (58) Is again stored in the weight matrix storage unit 1211.

In the matrix correction amount calculation unit 1216, the matrix correction unit
1213 Weight matrix correction amount (a × d _j ⁰ × o
_i ⁰ ) sum It is calculated by executing the following equation. Furthermore, the evaluation amount of the error between the output signal and the teacher signal is _calculated from d ′ _j ^{1 of the} result obtained by executing (123). The condition is S <Th1 (61) E> Th2, and if the condition expressed by the above formula (61) is satisfied, it is judged that the learning is stopped. Then, send a signal to the input layer correction unit 1217, the output layer U0 correction unit 1218,
Recalculate the output signal of each layer. If the above conditions are not satisfied, it is considered that the learning is not stopped, and no signal is output from the matrix correction amount calculation unit 1216 to the correction unit of each layer.

The input layer correction unit 1217 that has received the learning stop signal reads out the value o _i ⁰ held in the input layer storage unit 1202, and o _i ⁰ = Min (Max (o _i ⁰ , L) U) (62) Then, the input value o _i ⁰ is corrected and stored again in the input storage unit 1202. As shown in FIG. 7, the value of the input layer is 0.
Avoid getting close to 1 or 1.

The U0 correction unit 1218 in the output layer realizes the calculation of the formula U0 = U0 × c (63) to make the slope of the sigmoid function gentle as shown in FIG. 8 and recalculate the output layer output value. However, avoiding those values being close to 0 and 1 and preventing learning stop. In the embodiment, the parameter c
Was set to 2 for convenience. Then, again from equation (55) to equation (5
The weight matrix is modified by executing the calculation shown in 8). After the end of one learning, U0 is initialized again.

The above is a description of the learning process once, but by repeating this process, the learning can be completed without being disturbed by the learning stop.

FIG. 3 is a diagram showing an embodiment in the case of three layers of the pattern learning device of the third invention. Learning is performed by repeatedly giving an input pattern and a teacher signal. First, a case where one input pattern and one teacher signal are given will be described. The matrix storage units 311 and 312 for storing the matrix used in the matrix product calculation unit are initialized in advance with random numbers.

First, a feedforward process of calculating an output value from an input value will be described. The input pattern is input as a vector value from the terminal 301 and is stored in the input layer storage unit 302.
The input layer storage unit 302 is composed of a register that holds N ₀ -dimensional vectors, and all elements of the vectors are sent to the matrix product calculation unit 303. The input vector value is o _i ⁰ (1
≦ i ≦ N ₀ ) and the matrix value held in the first layer weight matrix storage unit 312 is w _ij ¹ (1 ≦ i ≦ N ₀ , 1 ≦ j ≦
N ₁ ), the vector value held in the intermediate layer storage unit 305 is represented by o ′ _j
^{If 1} (1 ≦ j ≦ N ₁ ), the matrix product operation unit 303 Is calculated. Here, θ _j ¹ is a bias value and is stored in the first layer weight matrix storage unit. Results of ^{_{o j 1 (1 ≦ j ≦}} N 1) are each determined to _j ¹ 'o by ^{_{j 1 = {1 + tanh (}} o j 1 / U0)} / 2 (65) comprising calculating' s in the function unit 304 o , In the output layer storage unit 305. In the example, U0 was initially set to 0.75. The weight matrix stored in the second layer weight matrix storage unit 211 is w _jk ² (1 ≦ i ≦ N ₁ , 1 ≦ j ≦ N ₂ ) and the storage unit 211
If the bias value stored in is set to θ _k ² , the matrix calculation unit 206 Is calculated. Results of ^{_{o k 2 (1 ≦ k ≦}} N 2) are each determined to _k ² 'o by ^{_{k 2 = {1 + tanh (}} o k 2) / U0} / 2 (67) comprising calculation' in s function section 307 o , And is stored in the output layer storage unit 308 as an output signal.

Next, the back propagation processing for correcting the weight matrix will be described.

With respect to the output signal obtained by the above-mentioned feedforward processing, the teacher signal that is an ideal value thereof is set to the terminal 302.
Input from the teacher signal storage unit 309 and hold the value t
_k (1 ≦ k ≦ N ₂ ) and the value o ′ _k ² in the output signal storage unit 308 are transferred to the error signal calculation unit 130, and the error signal d ′ _k ² is d ′ _k ² = t _k − o ′ _k ² (1 ≦ k ≦ N ₂ ) (68), the back propagation error is evaluated from the output signal o ′ _k ^2, and d _k ² = d ′ _k ² × o ′ _k ² × (1-o ′ _k ² ) (69) After d _k ² (1 ≦ k ≦ N ₂ ) is calculated by performing the equation, the value is used as the matrix correction unit 313 and the error back propagation unit 3
Transfer to 14. Further, the error signal d _k ² is also transferred to the matrix correction amount calculation unit 316 and used for observation of learning stop.

The matrix correction unit 313 calculates the second layer weight matrix by the calculation w _jk ² = w _jk ² + a × d _k ² × o ′ _j ¹ (70) θ _j ² = θ _j ² + b × d _k ² (71). The corrected result is stored again in the second layer weight matrix storage unit 311.

The d _k ² obtained by the calculation in which the error calculation unit 310 executes the equation (69) is sent to the error back-propagation unit 314, and together with the weight matrix values transferred from the matrix storage unit 311, Is calculated, and the equation d _j ¹ = d ′ _j ¹ × o ′ _j ¹ × (1-o ′ _j ¹ ) (73) is executed from o ′ _j ¹ transferred from the intermediate layer storage unit 305. After calculating d _j ¹ (1 ≦ j ≦ N ₁ ) by the calculation, it is transferred to the matrix correction unit 315 and w _ij ¹ = w _ij ¹ + a × d _j ¹ × o _i ⁰ (74) θ _i ¹ = Θ _i ¹ + b × d _j ¹ (75) The first layer weight matrix is corrected by calculation and the result is stored again in the first layer weight matrix storage unit 312.

In the matrix correction amount calculation unit 316, the matrix correction unit 3
The correction amount of the weight matrix obtained in 13,315 (a × d _k ²
Xo ′ _j ¹ , a × d _j ¹ × o _i ⁰ ) It is calculated by executing the following equation. Further, the evaluation amount of the error between the output signal and the teacher signal is calculated by d ′ _k ² of the result obtained by executing the equation (68). The condition is S <Th1 (78) E> Th2, and if the condition expressed by the above formula (78) is satisfied, it is judged that the learning is stopped. Then, the input layer correction unit 317, the output layer correction unit 318, the output layer U0 correction unit 31
Send the signal to 9 and recalculate the output signal of each layer. When the above conditions are not satisfied, it is considered that the learning is not stopped, and the matrix correction amount calculation unit 316 does not output a signal to the correction unit of each layer.

The input layer correction unit 317 that has received the learning stop signal reads out the value o _i ⁰ held in the input layer storage unit 302, and o _i ⁰ = Min (Max (o _i ⁰ , L), U) (79 ) Is performed to correct the input value o _i ⁰ and then stored again in the input storage unit 302 to prevent the value of the input layer from becoming a value close to 0 or 1.

In the U0 correction unit 318 in the middle layer and the U0 correction unit 319 in the output layer, the calculation of the formula U0 = U0 × c (80) is realized, and the slope of the zigmoid function is calculated as shown by the curves 801 and 802 in FIG. Is moderated to recalculate the output value of the middle layer and the output value of the intermediate layer, avoiding the values being close to 0 and 1 and preventing the learning from stopping. In FIG. 8, the curve 801 is U0 = 0.75, the curve 802 is 1.5, and C = 2. In the embodiment, the parameter c is set to 2 for convenience. Then, the weight matrix is corrected by executing the calculations shown in the equations (68) to (75) again.

Here, it is also possible to determine the learning stop shown in Equation (76) to Equation (78), and if Equation (78) is still satisfied, U0 can be corrected again by Equation (80). is there. After the end of one learning, U0 is initialized again.

The above is a description of the learning process once, but by repeating this process, the learning can be completed without being disturbed by the learning stop.

Next, with reference to FIG. 13, an embodiment in which the pattern learning apparatus of the fourth aspect of the present invention has two layers will be described. Learning is performed by repeatedly giving an input pattern and a teacher signal. First, a case where one input pattern and one teacher signal are given will be described. The matrix storage unit 1311 that stores the matrix used in the matrix product calculation unit is initialized in advance with random numbers.

First, a feedforward process of calculating an output value from an input value will be described. The input pattern is input as a vector value from the terminal 1301 and stored in the input layer storage unit 1302. The input layer storage unit 1302 is composed of a register that holds N ₀ -dimensional vectors, and all the elements of the vectors are sent to the matrix product calculation unit 1303. Input vector values o _i
⁰ (1 ≦ i ≦ N ₀ ), and the matrix value held in the weight matrix storage unit 1311 is w _ij ¹ (1 ≦ i ≦ N ₀ , 1 ≦ j ≦
N ₁ ), the vector value held in the output layer storage unit 1305 is o ′
_{If j} ¹ (1 ≦ j ≦ N ₁ ), the matrix product operation unit 1303 Is calculated. However, here θ _j ¹ is a bias value and is stored in the weight matrix storage unit. The resulting o _j
¹ (1 ≦ j ≦ N ₁ ) is calculated in the s-function unit 1304 as o ′ _j ¹ = {1 + tanh (o _j ¹ / U0)} / 2 (82), and o ′ _j ¹ is obtained and stored in the output layer. It is stored in the unit 1305.

Next, the operation of the part that determines the learning stop will be described.

For the output signal obtained by the above-mentioned feedforward processing, the ideal value of the teacher signal is output to the terminal 1320.
Input from the teacher signal storage unit 1309 and hold the value t
_j (1 ≦ j ≦ N ₁ ) and the value o ′ _j ¹ in the output signal storage unit 1305 are transferred to the error signal calculation unit 1310, and the error signal d ′ _j ¹ is d ′ _j ¹ = t _j −o. It is calculated as ′ _j ¹ (1 ≦ j ≦ N ₁ ) (83). The result is the matrix correction amount estimation unit 13
The error evaluation amount E for determining the learning stop is sent to 16 It is obtained by executing the calculation represented by the following equation. Further, o ′ _j ¹ (1 ≦ j ≦ N ₁ ) is transferred from the output layer storage unit 1305, and o _i ⁰ (1 ≦ i ≦ N ₀ ) is transferred from the input layer storage unit 1302 to the matrix correction amount estimation unit 1316, and the matrix is transferred. The estimated value S'of the correction amount is calculated by the following formula.

It is calculated by executing the following equation. Further, in the matrix correction amount estimation unit 1316, the evaluation amount of the error between the output signal and the teacher signal is subjected to the condition determination of S ′ <Th1 E> Th2 (86), and the condition represented by the above formula (86) is satisfied. If the condition is satisfied, it is determined that the learning is stopped, and a signal is sent to the input layer correction unit 1317 and the output layer correction unit 1318 to correct the output signal of each layer. If the above conditions are not satisfied, it is considered that the learning is not stopped, and no signal is output from the matrix correction amount calculation unit 1316 to the correction unit of each layer.

This determination condition considers that learning is stopped when the correction amount S of the weight matrix is close to 0 in spite of the large error evaluation amount E.

The input layer correction unit 1317 that has received the learning stop signal reads out the value o _i ⁰ held in the input layer storage unit 1302, and o _i ⁰ = Min (Max (o _i ⁰ , L), U) (87 ) Is performed to correct the input value o _i ⁰ and store it again in the input layer storage unit 1302 to prevent the value of the input layer from becoming a value close to 0 or 1. The U0 correction unit 1318 in the output layer realizes the calculation of the formula U0 = U0 × c (88) to make the slope of the sigmoid function gentle as described with reference to FIG. Recalculate the layer output values to avoid them approaching 0 or 1 and avoiding learning stops. Then, perform the calculations shown in equations (81) and (82) again,
Further, the learning stop determination represented by the equations (83) to (86) is determined, and if the equation (86) is still satisfied, the U0 is corrected again by the equation (88).

Finally, the back propagation processing for correcting the weight matrix using the corrected values of the signal storage units 1302 and 1305 of the respective layers will be described.

Evaluate the backpropagation error from _j ¹ Metropolitan 'output signal o is held in the output signal storage unit held by the _j ¹ and the output signal storage unit' that d which is determined in accordance with pre equation (83) by the error calculation section 1310 , D _j ¹ = d ′ _j ¹ × o ′ _j ¹ × (1-o ′ _j ¹ ) (89) After calculating d _j ¹ (1 ≦ _j ≦ N ₁ ), the matrix Transferred to the correction unit 1312 and weighted by calculation by executing the formula w _ij ¹ = w _ij ¹ + a × d _j ¹ × o _i ⁰ (90) θ _i ¹ = θ _i ¹ + b × d _j ¹ (91) Modify the matrix,
The result is stored again in the weight matrix storage unit 1311. After the end of one learning, U0 is set to the initial value again.

The above is a description of the learning process once, but by repeating this process, the learning can be completed without being disturbed by the learning stop.

FIG. 4 is a diagram showing an embodiment of the pattern learning apparatus of the fourth aspect of the present invention in the case of three layers. Learning is performed by repeatedly giving an input pattern and a teacher signal.
A case where one input pattern and one teacher signal are given will be described. The matrix storage units 411 and 412 that store the matrix used in the matrix product calculation unit are initialized in advance with random numbers.

First, a feedforward process of calculating an output value from an input value will be described. The input pattern is input as a vector value from the terminal 401 and is stored in the input layer storage unit 402.
The input layer storage unit 402 is composed of a register that holds N ₀ -dimensional vectors, and all the elements of the vectors are sent to the matrix product calculation unit 403. The input vector value is o _i ⁰ (1
≦ i ≦ N ₀ ), and the matrix value held in the first layer weight matrix storage unit 412 is w _ij ¹ (1 ≦ i ≦ N ₀ , 1 ≦ j ≦
N ₁ ), the vector value held in the intermediate layer storage unit 405 is represented by o ′ _j
^{If 1} (1 ≦ j ≦ N ₁ ), the matrix product operation unit 403 Is calculated. However, here θ _j ¹ is a bias value and is stored in the first layer weight matrix storage unit 412. Results of ^{_{o j 1 (1 ≦ j ≦}} N 1) each of the s function unit 404 o determine the ^{_{j 1 'j 1 = {1}} + tanh (o j 1)} / 2 (93) o by comprising calculation', the intermediate It is stored in the layer storage unit 405.

The weight matrix stored in the second layer weight matrix storage unit 411 is w _jk ² (1 ≦ i ≦ N ₁ , 1 ≦ j ≦ N ₂ ), and the bias value stored in the storage unit 411 is θ _k ² . Then, in the matrix calculation unit 406, Is calculated. Results of ^{_{o k 2 (1 ≦ k ≦}} N 2) each of the s function unit 407 o determine the ^{_{k 2 'k 2 = {1}} + tanh (o k 2)} / 2 (95) o by comprising calculation', the output It is stored in the output layer storage unit 408 as a signal.

Next, the operation of the part that determines the learning stop will be described.

With respect to the output signal obtained by the above-mentioned feedforward processing, the teacher signal which is an ideal value thereof is set to the terminal 420.
Input from the teacher signal storage unit 409 and hold the value t
_k (1 ≦ k ≦ N ₂ ) and the value o ′ _k ² in the output signal storage unit 408 are transferred to the error signal calculation unit 410, and the error signal d ′ _k ² is d ′ _k ² = t _k − It is calculated as o ′ _k ² (1 ≦ k ≦ N ₂ ) (96). The result is sent to the matrix correction amount estimation unit, and the error evaluation amount E for determining learning stop is calculated. It is obtained by executing the calculation represented by the following equation. Further, o ′ _k ² (1 ≦ k ≦ N ₂ ) from the output layer storage unit 408, o ′ _j ¹ (1 ≦ j ≦ N ₁ ) from the intermediate layer storage unit 405, and o _i ⁰ from the input layer storage unit 402. (1 ≦ i ≦ N ₀ ) is transferred to the matrix correction amount estimation unit 416, and the matrix correction amount estimated value S ′ is calculated by the following formula.

It is calculated by executing the following equation. Further, in the matrix correction amount estimation unit 416, the evaluation amount of the error between the output signal and the teacher signal is subjected to the condition determination of S ′ <Th1 E> Th2 (99), and the condition expressed by the above equation (99) is satisfied. If the condition is satisfied, it is determined that the learning is stopped, and a signal is sent to the input layer correction unit 417, the intermediate layer correction unit 418, and the output layer correction unit 419 to correct the output signal of each layer. If the above conditions are not satisfied, it is considered that the learning is not stopped, and no signal is output from the matrix correction amount calculation unit 416 to the correction unit of each layer.

This determination condition considers that learning is stopped when the correction amount S of the weight matrix is close to 0 in spite of the large error evaluation amount E.

The input layer correction unit 417 that has received the learning stop signal reads out the value o _i ⁰ held in the input layer storage unit 402, and o _i ⁰ = Min (Max (o _i ⁰ , L), U) (100 ) Is performed to correct the input value o _i ⁰ and store it again in the input layer storage unit 402 to prevent the value of the input layer from becoming a value close to 0 or 1.

In the U0 correction unit 418 in the middle layer and the U0 correction unit 419 in the output layer, the calculation of the formula U0 = U0 × c (101) is realized, and the slope of the sigmoid function is made gentle as described in FIG. Then, the intermediate layer output and the output layer output value are recalculated to avoid the values being close to 0 and 1 and to prevent the learning stop. In the embodiment, the parameter c is set to 2 for convenience. Then, the calculations shown in Equation (93) and Equation (95) are executed again, and then Equation (97)
From the above, the learning stop determination shown in Expression (99) is performed, and if Expression (99) is still satisfied, U is again calculated using Expression (101).
Correct 0.

Finally, the back propagation processing for correcting the weight matrix using the corrected values of the signal storage units 402, 405, 408 of the respective layers will be described.

The error calculation unit 410 obtains d ′ _k ² previously obtained according to the equation (96) and the output signal o ′ _k ² held in the output signal storage unit.
The back-propagation error is evaluated from and d _k ² = d ′ _k ² × o ′ _k ² × (1−o ′ _k ² ) (102) is calculated to execute d _k ² (1 ≦ k ≦ N ₂ ) is obtained, and then the value is calculated by the matrix correction unit 413 and the error back propagation unit 4
Transfer to 14. Further, the error signal d _k ² is also transferred to the matrix correction amount calculation unit 416 and used for observation of learning stop.

The matrix correction unit 413 calculates the second layer weight matrix by the calculation w _jk ² = w _jk ² + a × d _k ² × o ′ _j ¹ (103) θ _j ² = θ _j ² + b × d _k ² (104). The corrected result is stored again in the second layer weight matrix storage unit 411.

The d _k ² obtained by the calculation in which the error calculation unit 410 executes the equation (102) is sent to the error back-propagation unit 414, and together with the weight matrix values transferred from the matrix storage unit 411, Is calculated, and the formula d _j ¹ = d ′ _j ¹ × o ′ _j ¹ × (1-o ′ _j ¹ ) (106) is executed from o ′ _j ¹ transferred from the intermediate layer storage unit 405. After calculating d _j ¹ (1 ≦ j ≦ N ₁ ) by the calculation, it is transferred to the matrix correction unit 415 and w _ij ¹ = w _ij ¹ + a × d _j ¹ × o _i ⁰ (107) θ _i ¹ = Θ _i ¹ + b × d _j ¹ (108) The first layer weight matrix is corrected by the calculation to execute the equation, and the result is stored again in the first layer weight matrix storage unit 412. When one learning is completed, U0 returns to the initial value.

The above is a description of the learning process once, but by repeating this process, the learning can be completed without being disturbed by the learning stop.

Next, with reference to FIG. 14, an explanation will be given of an embodiment in the case of two layers of the pattern learning apparatus of the fifth present invention. Learning is performed by repeatedly giving an input pattern and a teacher signal. First, a case where one input pattern and one teacher signal are given will be described. The matrix storage unit 1411 that stores the matrix used in the matrix product calculation unit is initialized in advance with a random number.

First, a feedforward process of calculating an output value from an input value will be described. The input pattern is input as a vector value from the terminal 1401 and stored in the input layer storage unit 1402. The input layer storage unit 1402 is composed of a register that holds N ₀ -dimensional vectors, and all the elements of the vectors are sent to the matrix product calculation unit 1403. Input vector values o _i
⁰ (1 ≦ i ≦ N ₀ ), and the matrix value held in the weight matrix storage unit 1411 is w _ij ¹ (1 ≦ i ≦ N ₀ , 1 ≦ j ≦
N ₁ ), the vector value held in the output layer storage unit 1405 is o ′
_{If j} ¹ (1 ≦ j ≦ N ₁ ), the matrix product operation unit 1403 Is calculated. However, here θ _j ¹ is a bias value and is stored in the weight matrix storage unit. The resulting o _j
¹ (1 ≦ j ≦ N ₁ ) is calculated in the s-function unit 1404 as o ′ _j ¹ = p + (q−p) {1 + tanh (o _j ¹ )} / 2 (110) to obtain o ′ _j ¹ . , And is stored in the output layer storage unit 1405 as an output signal. In the example, p and q were initially set to 0 and 1, respectively.

 Next, the correction processing of the weight matrix will be described.

With respect to the output signal obtained by the above-mentioned feedforward processing, the ideal value of the teacher signal is input to the terminal 1420.
Input from the teacher signal storage unit 1409, and the value t
_j (1 ≦ j ≦ N ₁ ) and the value o ′ _j ¹ in the output signal storage unit 1405 are transferred to the error signal calculation unit 1410, and the error signal d ′ _j ¹ is d ′ _j ¹ = t _j − o ′ _j ¹ (1 ≦ k ≦ N ₁ ) (111), and the back propagation error is evaluated from the output signal o ′ _j ¹ and d _j ¹ = d ′ _j ¹ × o ′ _j ¹ × (1-o ′ _j ¹ ) (112) After calculating d _j ¹ (1 ≦ j ≦ N ₁ ) by performing the equation, the value is transferred to the matrix correction unit 413. Further, the error signal d _j ¹ is also transferred to the matrix correction amount calculation unit 1416 and used for observation of learning stop.

The matrix correction unit 1413 corrects the weight matrix by the calculation of w _ij ¹ = w _ij ¹ + a × d _j ¹ × o _i ⁰ (113) θ _i ¹ = θ _i ¹ + b × d _j ¹ (114) Is again stored in the weight matrix storage unit 1411.

In the matrix correction amount calculation unit 1416, the matrix correction unit
1413 Weight matrix correction amount (a × d _j ¹ × o
_i ⁰ ) sum It is calculated by executing the following equation. Furthermore, the evaluation amount of the error between the output signal and the teacher signal is _calculated from d ′ _j ^{1 of the} result obtained by executing (111). It is determined by executing the following formula, and the condition judgment of S <Th1 (117) E> Th2 is executed. If the condition expressed by the above formula (117) is satisfied, it is judged that the learning is stopped. do it,
A signal is sent to the input layer correction unit 1417 and the output layer range correction unit 1418 to recalculate the output signal of each layer. When the above conditions are not satisfied, it is considered that the learning is not stopped, and the matrix correction amount calculation unit 1416 does not output a signal to the correction unit of each layer.

The input layer correction unit 1417 that has received the learning stop signal reads out the value o _i ⁰ held in the input layer storage unit 1402, and o _i ⁰ = Min (Max (o _i ⁰ , L), U) (118 ) Is executed to correct the input value o _i ⁰ and stored again in the input storage unit 1402, and the value of the input layer is 0 as shown in FIG.
Avoid getting close to 1 or 1.

The range correction unit 1418 in the output layer realizes the calculation of the equation p = p '(119) q = q', and corrects the range of the sigmoid function as shown in FIG. Recalculates
Avoid those values approaching 0 or 1 and avoiding learning stops. In the example of FIG. 9, the curve 901 has p = 0 and q = 1, and the curve 902 has p = 0.2 and q = 0.8. Then, the weight matrix is corrected by executing the calculations shown in the equations (111) to (114) again. After one learning, p and q are initialized again.

The above is a description of the learning process once, but by repeating this process, the learning can be completed without being disturbed by the learning stop.

FIG. 5 is a diagram showing an embodiment of the pattern learning apparatus of the fifth invention in the case of three layers. Learning is performed by repeatedly giving an input pattern and a teacher signal. First, a case where one input pattern and one teacher signal are given will be described. The matrix storage units 511 and 512 for storing the matrix used in the matrix product calculation unit are initialized in advance with random numbers.

First, a feedforward process of calculating an output value from an input value will be described. The input pattern is input as a vector value from the terminal 501 and is stored in the input layer storage unit 502.
The input layer storage unit 502 is composed of a register that holds N ₀ -dimensional vectors, and all the elements of the vectors are sent to the matrix product calculation unit 503. The input vector value is o _i ⁰ (1
≦ i ≦ N ₀ ), and the matrix value held in the first layer weight matrix storage unit 512 is w _ij ¹ (1 ≦ i ≦ N ₀ , 1 ≦ j ≦
N ₁ ), the vector value held in the intermediate layer storage unit 505 is represented by o ′ _j
^{If 1} (1 ≦ j ≦ N ₁ ), the matrix product calculation unit 503 Is calculated. Here, θ _j ¹ is a bias value and is stored in the first layer weight matrix storage unit. The resulting o _j ¹ (1 ≤ j ≤ N ₁ ) are respectively _o'j ¹ = p + (q-p) {1 + tanh (o _j ¹ / U0)} / 2 (12 in the s-function unit 504.
1) O ′ _j ¹ is obtained by the following calculation and stored in the output layer storage unit 505. In the example, p and q were initially set to 0 and 1, respectively. The weight matrix stored in the second layer weight matrix storage unit 511 is w _jk ² (1 ≦ i ≦ N ₁ , 1 ≦ j ≦ N ₂ ),
When the bias value stored in the storage unit 511 is θ _k ² , the matrix calculation unit 506 Is calculated. Results of ^{_{o k 2 (1 ≦ k ≦}} N 2) each of the s function unit ^{_{507 o 'k 2 = p +}} (p-q) {1 + tanh (o k 2) / U0} / 2 (12
3) O ′ _k ² is _obtained by the following calculation and stored in the output layer storage unit 508 as an output signal.

Next, the back propagation processing for correcting the weight matrix will be described.

With respect to the output signal obtained by the above-mentioned feedforward processing, the ideal value of the teacher signal is set to the terminal 520.
Input from the teacher signal storage unit 509 and hold the value t
_k (1 ≦ k ≦ N ₂ ) and the value o ′ _k ² in the output signal storage unit 508 are transferred to the error signal calculation unit 530, and the error signal d ′ _k ² is d ′ _k ² = t _k − o ′ _k ² (1 ≦ k ≦ N ₂ ) (124), and the back propagation error is evaluated from the output signal o ′ _k ² and d _k ² = d ′ _k ² × o ′ _k ² × (1-o ′ _k ² ) (125) After calculating d _k ² (1 ≦ k ≦ N ₂ ) by calculation, the value is used as the matrix correction unit 513 and the error back propagation unit 5
Transfer to 14. Further, the error signal d ′ _k ² is also transferred to the matrix correction amount calculation unit 516 and used for observation of learning stop.

The matrix correction unit 513 calculates the second-layer weight matrix by the calculation w _jk ² = w _jk ² + a × d _k ² × o ′ _j ¹ (126) θ _j ² = θ _j ² + b × d _k ² (127). The corrected result is stored again in the second layer weight matrix storage unit 511.

The d _k ² obtained by the calculation in which the error calculation unit 510 executes the equation (125) is sent to the error back-propagation unit 514, and together with the weight matrix values transferred from the matrix storage unit 511, Is calculated, and the equation d _j ¹ = d ′ _j ¹ × o ′ _j ¹ × (1-o ′ _j ¹ ) (129) is executed from o ′ _j ¹ transferred from the intermediate layer storage unit 505. After calculating d _j ¹ (1 ≦ j ≦ N ₁ ) by the calculation, it is transferred to the matrix correction unit 515 and w _ij ¹ = w _ij ¹ + a × d _j ¹ × o _i ⁰ (130) θ _i ¹ = θ _i ¹ + b × d _j ¹ (131) The first layer weight matrix is corrected by the calculation to execute the equation, and the result is stored in the first layer weight matrix storage unit 512 again.

In the matrix correction amount calculation unit 516, the matrix correction unit 5
Amount of correction of the weight matrix found in 13,515 It is calculated by executing the following equation. Further, the evaluation amount of the error between the output signal and the teacher signal is calculated by d ′ _k ² of the result obtained by executing the equation (124). The condition is S <Th1 (134) E> Th2, and if the condition expressed by Eq. (81) is satisfied, it is judged that the learning is stopped. Then, the input layer modifying unit 517, the intermediate layer U0 modifying unit 518, and the output layer modifying unit 51.
Send the signal to 9 and recalculate the output signal of each layer. If the above conditions are not satisfied, it is considered that the learning is not stopped, and no signal is output from the matrix correction amount calculation unit 516 to the correction unit of each layer.

The input layer correction unit 517 that has received the learning stop signal reads out the value o _i ⁰ held in the input layer storage unit 502, and o _i ⁰ = Min (Max (o _i ⁰ , L), U) (135 ) Is performed to correct the input value o _i ⁰ and store it again in the input storage unit 502 to prevent the value of the input layer from becoming a value close to 0 or 1.

The range correction unit 518 in the intermediate layer and the range correction unit 519 in the output layer realize the calculation of the equation p = p ′ (136) q = q ′, and as shown in FIG. 9, the sigmoid function range is calculated. Is corrected to recalculate the output value of the middle layer and the output value of the intermediate layer, avoiding the values being close to 0 and 1 and preventing the learning stop. Then, the weight matrix is corrected by executing the calculations shown in the equations (120) to (131) again. After the end of one learning, p and q are initialized again.

The above is a description of the learning process once, but by repeating this process, the learning can be completed without being disturbed by the learning stop.

Next, with reference to FIG. 15, an embodiment of the pattern learning apparatus of the sixth aspect of the present invention having two layers will be described. Learning is performed by repeatedly giving an input pattern and a teacher signal. First, a case where one input pattern and one teacher signal are given will be described. The matrix storage unit 1511 that stores the matrix used in the matrix product calculation unit is initialized in advance with a random number.

First, a feedforward process of calculating an output value from an input value will be described. The input pattern is input as a vector value from the terminal 1501 and stored in the input layer storage unit 1502. The input layer storage unit 1502 is composed of a register that holds N ₀ -dimensional vectors, and all the elements of the vectors are sent to the matrix product calculation unit 1503. Input vector values o _i
⁰ (1 ≦ i ≦ N ₀ ), and the matrix value held in the weight matrix storage unit 1411 is w _ij ¹ (1 ≦ i ≦ N ₀ , 1 ≦ j ≦
N ₁ ), the vector value held in the output layer storage unit 1405 is o ′
_{If j} ¹ (1 ≦ j ≦ N ₁ ), the matrix product operation unit 1503 Is calculated. However, here θ _j ¹ is a bias value and is stored in the weight matrix storage unit. The resulting o _{j
^{1 (1 ≦ j ≦ N 1}} ) each of which obtains the _j ¹ 'o by ^{_{j 1 = p + (q-}} p) {1 + tanh (o j 1)} / 2 (138) comprising computing' in s function unit 1504 o , And is stored in the output layer storage unit 1505 as an output signal. In the example, p and q were initially set to 0 and 1, respectively.

Next, the operation of the part that determines the learning stop will be described.

With respect to the output signal obtained by the above-mentioned feedforward processing, the ideal value of the teacher signal is set to the terminal 1520.
Input from the teacher signal storage unit 1509 and hold the value t
_j (1 ≦ j ≦ N ₁ ) and the value o ′ _j ¹ in the output signal storage unit 1405 are transferred to the error signal calculation unit 1510, and the error signal d ′ _j ¹ is d ′ _j ¹ = t _j − It is calculated as o ′ _j ¹ (1 ≦ j ≦ N ₁ ) (139). The result is the matrix correction amount estimation unit 15
The error evaluation amount E for determining the learning stop is sent to 16 It is obtained by executing the calculation represented by the following equation. Further, o ′ _j ¹ (1 ≦ j ≦ N ₁ ) is transferred from the output layer storage unit 1505 and o _j ⁰ (1 ≦ i ≦ N ₀ ) is transferred from the input layer storage unit 1502 to the matrix correction amount estimation unit 1516, and the matrix The estimated value S'of the correction amount is calculated by the following formula.

It is calculated by executing the following equation. Further, in the matrix correction amount estimation unit 1516, the evaluation amount of the error between the output signal and the teacher signal is subjected to the condition determination S ′ <Th1 E> Th2 (142), and the condition represented by the above expression (142) is satisfied. If it satisfies, it is judged that the learning is stopped,
A signal is sent to the input layer correction unit 1517 and the output layer correction unit 1418 to correct the output signal of each layer. If the above conditions are not satisfied, it is considered that the learning is not stopped, and no signal is output from the matrix correction amount calculation unit 1516 to the correction unit of each layer.

This determination condition considers that learning is stopped when the correction amount S of the weight matrix is close to 0 in spite of the large error evaluation amount E.

The input layer correction unit 1517 that has received the learning stop signal reads out the value o _i ⁰ held in the input layer storage unit 1502, and o _i ⁰ = Min (Max (o _i ⁰ , L), U) (143 ) Is performed to correct the input value o _i ⁰ and store it again in the input layer storage unit 1502 to avoid that the value of the input layer becomes close to 0 or 1. The range corrector 1318 in the output layer realizes the calculation of the equation p = p ′ q = q ′ (144) to make the gradient of the sigmoid function gentle as described above, and to output the intermediate layer output and the output layer output. Recalculate the values to avoid making them close to 0 or 1 and avoiding learning stops. Then, the calculations shown in Eqs. (156) and (158) are executed again to realize the modification of the weight matrix. After one learning, p and q are initialized. Finally, the modification is performed. The back propagation processing for correcting the weight matrix using the values of the signal storage units 1502 and 1505 of each layer will be described.

Evaluate the backpropagation error from _j ¹ Metropolitan 'output signal o is held in the output signal storage unit held by the _j ¹ and the output signal storage unit' that d which is determined in accordance with pre equation (156) by the error calculation section 1510 , D _j ¹ = d ′ _j ¹ × o ′ _j ¹ × (1-o ′ _j ¹ ) (145) After calculating d _j ¹ (1 ≦ j ≦ N ₁ ) by the calculation, the matrix Transferred to the correction unit 1512 and weighted by calculation by executing the formula w _ij ¹ = w _ij ¹ + a × d _j ¹ × o _i ⁰ (146) θ _i ¹ = θ _i ¹ + b × d _j ¹ (147) Modify the matrix,
The result is stored again in the weight matrix storage unit 1511. After completion of one learning, p and q are set to initial values again.

The above is a description of the learning process once, but by repeating this process, the learning can be completed without being disturbed by the learning stop.

FIG. 6 is a diagram showing an embodiment of the pattern learning apparatus of the sixth invention. Learning is performed by repeatedly giving an input pattern and a teacher signal. First, a case where one input pattern and one teacher signal are given will be described.
The matrix storage units 611 and 612 that store the matrix used in the matrix product calculation unit are initialized in advance with random numbers.

First, a feedforward process of calculating an output value from an input value will be described. The input pattern is input as a vector value from the terminal 601 and stored in the input layer storage unit 602.
The input layer storage unit 602 is composed of a register that holds N ₀ -dimensional vectors, and all the elements of the vectors are sent to the matrix product calculation unit 603. The input vector value is o _i ⁰ (1
≦ i ≦ N ₀ ) and the matrix value held in the first layer weight matrix storage unit 612 is w _ij ¹ (1 ≦ i ≦ N ₀ , 1 ≦ j ≦
N ₁ ), the vector value held in the intermediate layer storage unit 605 is represented by o ′ _j
^{If 1} (1 ≦ j ≦ N ₁ ), the matrix product calculation unit 603 Is calculated. However, θ _j ¹ is a bias value, and is stored in the first layer weight matrix storage unit 612. The resulting o _j ¹ (1 ≦ j ≦ N ₁ ) is calculated as o ′ _j ¹ = p + (q−p) {1 + tanh (o _j ¹ )} / 2 (148) in the s-function unit 604. _j ¹ is obtained and stored in the intermediate layer storage unit 605. p and q are 1 and 0.

The weight matrix stored in the second layer weight matrix storage unit 611 is w _jk ² (1 ≦ i ≦ N ₁ , 1 ≦ j ≦ N ₂ ), and the bias value stored in the storage unit 611 is θ _k ² . Then, in the matrix calculation unit 606, Is calculated. Results of ^{_{o k 2 (1 ≦ k ≦}} N 2) each of the s function unit 607 o determine the ^{_{k 2 'k 2 = {1}} + tanh (o k 2)} / 2 (151) calculates o by comprising' output It is stored in the output layer storage unit 608 as a signal.

Next, the operation of the part that determines the learning stop will be described.

With respect to the output signal obtained by the above-mentioned feedforward processing, the teacher signal that is an ideal value thereof is set to the terminal 620.
Input from the teacher signal storage unit 609 and hold the value t
_k (1 ≦ k ≦ N ₂ ) and the value o ′ _k ² in the output signal storage unit 608 are transferred to the error signal calculation unit 610, and the error signal d ′ _k ² d ′ _k ² = t _k −o ′. _It is calculated as _k ² (1 ≦ k ≦ N ₂ ) (152). The result is sent to the matrix correction amount estimation unit, and the error evaluation amount E for determining learning stop is calculated. It is obtained by executing the calculation represented by the following equation. Further, o ′ _k ² (1 ≦ k ≦ N ₂ ) from the output layer storage unit 608, o ′ _j ¹ (1 ≦ j ≦ N ₁ ) from the intermediate layer storage unit 605, and o _i ⁰ from the input layer storage unit 602. (1 ≦ i ≦ N ₀ ) is transferred to the matrix modification amount estimation unit 616, and the matrix modification amount estimated value S ′ is calculated by the following formula.

It is calculated by executing the following equation. Further, in the matrix correction amount estimation unit 616, the evaluation amount of the error between the output signal and the teacher signal is subjected to the condition determination of S ′ <Th1 E> Th2 (155), and the condition expressed by the above formula (155) is satisfied. If it satisfies, it is judged that the learning is stopped,
Input layer correction unit 617, intermediate layer correction unit 618, output layer correction unit 619
The signal is sent to and the output signal of each layer is corrected. If the above conditions are not met, it is considered that the learning is not stopped,
No signal is output from the matrix correction amount calculation unit 616 to the correction unit of each layer.

This determination condition considers that learning is stopped when the correction amount S of the weight matrix is close to 0 in spite of the large error evaluation amount E.

The input layer correction unit 617 that has received the learning stop signal reads out the value o _i ⁰ held in the input layer storage unit 602, and o _i ⁰ = Min (Max (o _i ⁰ , L), U) (156 ) Is performed to correct the input value o _i ⁰ and then stored again in the input layer storage unit 602 to avoid that the value of the input layer becomes close to 0 or 1.

The range corrector 618 in the middle layer and the range corrector 619 in the output layer realize the calculation of the equation p = p ′ q = q ′ (157), correct the sigmoid function range, and output the middle layer. And output layer output values are recalculated to prevent them from becoming values close to 0 and 1 and prevent learning stop. In the embodiment, the parameters p ′ and q ′ are set to 0.2.
And 0.8. Then, the calculations shown in the equations (149) and (151) are executed again to realize the correction of the output value in each layer. After one learning, p and q are initialized again.

Finally, the back propagation processing for correcting the weight matrix using the corrected values of the signal storage units 602, 605, 608 of the respective layers will be described.

The output signal o ′ _k ² held in the output signal storage unit and d ′ _k ² previously obtained by the error calculation unit 610 according to the equation (152).
The back-propagation error is evaluated from and d _k ² (1 ≦ k ≦ N is calculated by executing the equation d _k ² = d ′ _k ² × o ′ _k ² × (1-o ′ _k ² ) (158). ₂ ) is obtained, and then the value is calculated by the matrix correction unit 613 and the error back propagation unit 6
Transfer to 14. Further, the error signal d ′ _k ² is also transferred to the matrix correction amount calculation unit 616 and used for observation of learning stop.

The matrix correction unit 613 calculates the second layer weight matrix by the calculation w _jk ² = w _jk ² + a × d _k ² × o ′ _j ¹ (159) θ _j ² = θ _j ² + b × d _k ² (160). The corrected result is stored again in the second layer weight matrix storage unit 611.

The d _k ² obtained by the calculation of executing the equation (158) in the error calculation unit 610 is sent to the error back-propagation unit 614, and together with the weight matrix values transferred from the matrix storage unit 611, Of o ′ _j transferred from the intermediate layer storage unit 605 to
After obtaining the ^{_{d j 1 (1 ≦ j ≦}} N 1) by ^{_{d j 1 = d 'j 1}} × o' j 1 × (1-o 'j 1) (162) comprising computing that executes equation ¹ Metropolitan , A calculation for transferring to the matrix correction unit 615 and executing the formula w _ij ¹ = w _ij ¹ + a × d _j ¹ × o _i ⁰ (163) θ _i ¹ = θ _i ¹ + b × d _j ¹ (164) The first layer weight matrix is corrected by and the result is stored again in the first layer weight matrix storage unit 612. When one learning is completed, p and q are returned to initial values.

The above is a description of the learning process once, but by repeating this process, the learning can be completed without being disturbed by the learning stop.

(Effects of the Invention) As is apparent from the above description, in the present invention, learning stop is observed, and when learning stop occurs,
Although the evaluation amount of the error between the output signal and the teacher signal is large, the learning is stopped when the correction amount of the weight matrix is close to zero. This means that the signals in the middle layer and the output layer are almost zero.
This is because the input signal is close to 0 and the correction amount of the weight matrix becomes 0. In this state, the learning apparatus of the present invention prevents the signals in the intermediate layer and the output layer from becoming values close to 0 and 1, and avoids the input signal from becoming values close to 0, thereby avoiding learning stop.

According to the conventional learning method, using the weight matrix in which learning is stopped as an initial value, the learning apparatus according to the present invention is used to perform pattern learning, and as a result, the state of learning stop is escaped and correct learning is completed. I was able to.

[Brief description of drawings]

1 and 10 are block diagrams showing an embodiment of the first invention, FIGS. 2 and 11 are diagrams showing an embodiment of the second invention,
FIGS. 3 and 12 are block diagrams showing an embodiment of the third invention, FIGS. 4 and 13 are block diagrams showing an embodiment of the fourth invention, and FIGS. 5 and 14 are fifth. FIG. 6 is a block diagram showing an embodiment of the present invention, and FIGS. 6 and 15 are block diagrams showing an embodiment of the sixth invention. FIG. 7 is a diagram showing the effect of giving an upper limit value and a lower limit value to the output value of the sigmoid function to prevent the output value from approaching 1 or 0. FIG. 8 is a diagram showing the effect of changing the slope control parameter of the sigmoid function to prevent the output value of the sigmoid function from approaching 1 or 0. FIG. 9 is a diagram showing the effect of preventing the output value from approaching 1 or 0 by changing the range of the output value of the sigmoid function. In the figure, 102,202,302,402,502,602,1002,1102,1202,1302,1402,1
502 ... Input layer storage unit, 103,203,303,403,503,603,1003,1103,1203,1303,1403,1
503 ... Matrix product operation unit, 104,204,304,404,504,604,1004,1104,1204k, 1304,1404,
1504 ... s function part, 105,205,305,405,505,605 ... intermediate layer storage part, 106,206,306,406,606 ... matrix product calculation part, 107,207,307,407,607,607 ... s function part, 108,208,308,408,508,608,1005,1105,1205,1305,1405,1
505 ... Output layer storage unit, 109,209,309,409,509,609,1009,1109,1209,1309,1409,1
509 ... Teacher signal memory, 110,210,310,410,510,610,1010,1110,1210,1310,1410,1
510 …… Error calculator, 111,211,311,411,511,611,1011,1111,1211,1311,1411,1
511 ... matrix storage unit, 112,212,312,412,512,612 ... Matrix Kokubu, 113,213,313,413,513,613,1013,1113,1213,1313,1413,1
513 ... matrix correction unit, 114,214,314,414,514,614 ... error backpropagation unit, 115,215,315,415,615 ... matrix correction unit, 116,316,516,516,1016,1116,1216,1316,1416,1516 ... matrix correction amount calculation unit, 216,416,616 ... matrix correction amount estimation Part, 117,217,317,417,517,617,1017,1117,1217,1317,1417,1
517 ... Input layer correction unit, 118,218 ... Middle layer correction unit, 119,219,1018,1118 ... Output layer correction unit, 318,318,418,419,1218,1318 ... U0 correction unit, 518,519,618,619,1418,1518 ... Range correction unit.

Claims (6)

(57) [Claims] 1. A feed-forward type neural network composed of a plurality of layers including an input layer and an output layer for learning by a back propagation method, and a sum of correction amounts of weight matrices used in the neural network is calculated. By comparing the result with a predetermined threshold value, when the calculation result of the sum is smaller, and
A means for determining an error between the signal of the output layer and the teacher signal, comparing the error with another threshold value given in advance, and determining that learning is stopped when the error is large, and a means for stopping the learning. If it is determined that the output signal from each layer is compared with the threshold value given in advance, only when it is larger than the threshold value, the threshold value is substituted into the value of the output signal. ,
Furthermore, the output signal is compared with another threshold value given in advance, and only when the threshold value is smaller than the threshold value, the threshold value is substituted into the value of the output signal to modify the output signal. A pattern learning device characterized by the above. 2. A feedforward type neural network consisting of a plurality of layers including an input layer and an output layer for learning by a back propagation method, and modification of a weight matrix used in the neural network calculated from signals of each layer. The evaluation value of the amount is calculated from the output value of each layer and the input value of the upper layer of the layer, and compared with a predetermined threshold, and when the evaluation value of the correction amount is small, and the output layer Means for determining the error between the signal and the teacher signal and comparing the error with another threshold value given in advance and the error is large, and determining that learning is stopped,
When it is determined that learning has stopped, each output signal from each layer is compared with a preset threshold value, and the threshold value is added to the output signal value only when it is larger than the threshold value. Means for comparing the output signal with another threshold value given in advance, and only when the output signal is smaller than the threshold value, the threshold value is substituted for the value of the output signal to correct the output signal. And a pattern learning device. 3. A result of calculating a sum of correction amounts of a feed-forward type neural network consisting of a plurality of layers including an input layer and an output for learning by a back propagation method and a weight matrix used in the neural network. And a predetermined threshold value is compared, and when the calculation result of the total sum is smaller, and an error between the signal in the output layer and the teacher signal is obtained, another threshold value is given in advance. When the error is large compared with, means for determining that learning is stopped, and when learning is stopped, each of the signals of the input layer is compared with a predetermined threshold value, Only if greater than the threshold
The threshold value is substituted for the value of the input signal, the input signal is compared with another threshold value given in advance, and the value of the input signal is relevant only when it is smaller than the threshold value. Change the input signal by substituting the threshold value, and change the slope control parameter of the sigmoid function that determines the output signal in layers other than the input layer to a larger value, and then calculate the value of the output signal again. A pattern learning device comprising: 4. A feedforward type neural network consisting of a plurality of layers including an input layer and an output for learning by a back propagation method, and a correction amount of a weight matrix used in the neural network calculated from signals of each layer. Is calculated from the output value of each layer and the input value of the upper layer of the layer, and compared with a predetermined threshold value, when the evaluation value of the correction amount is small, and the output layer Means for determining the error between the signal and the teacher signal and comparing the error with another threshold value given in advance and the error is large, and determining that learning is stopped,
When learning is stopped, each of the signals in the input layer is compared with a pre-given threshold value, and only when it is larger than the threshold value, the threshold value is substituted into the value of the input signal. ,
Further, means for correcting the input signal by comparing the input signal with another threshold value given in advance and substituting the threshold value for the value of the input signal only when the input signal is smaller than the threshold value. , Pattern learning comprising means for changing the slope control parameter of a sigmoid function for obtaining an output signal in a layer other than the input layer to a larger value and calculating the value of the output signal again. apparatus. 5. A result obtained by calculating a sum of correction amounts of a feedforward type neural network consisting of a plurality of layers including an input layer and an output for performing learning by a reverse transport method and a weight matrix used in the neural network. And a predetermined threshold value is compared, and when the calculation result of the total sum is smaller, and an error between the signal in the output layer and the teacher signal is obtained, another threshold value is given in advance. When the error is large compared with, means for determining that learning is stopped, and when learning is stopped, each of the signals of the input layer is compared with a predetermined threshold value, Only if greater than the threshold
The threshold value is substituted for the value of the input signal, the input signal is compared with another threshold value given in advance, and the value of the input signal is relevant only when it is smaller than the threshold value. Change the sigmoid function by reducing the range between the maximum and minimum values of the sigmoid function that finds the output signal in layers other than the input layer and the means that corrects the input signal by substituting the threshold value A pattern learning device, comprising: means for calculating and obtaining a signal value. 6. A feedforward type neural network consisting of a plurality of layers including an input layer and an output for learning by a back propagation method, and a correction amount of a weight matrix used in the neural network calculated from signals of each layer. The evaluation value of, calculated from the output value of each layer and the input value of the upper layer of the layer, compared with a predetermined threshold, in the case where the evaluation value of the correction amount is small,
Further, means for determining an error between the output layer signal and the teacher signal and comparing the error with another threshold value given in advance to determine that the learning is stopped when the error is large, When is stopped, each of the signals of the input layer is compared with a pre-given threshold value, and only when it is larger than the threshold value, the threshold value is substituted into the value of the input signal, Further, means for correcting the input signal by comparing the input signal with another threshold value given in advance and substituting the threshold value for the value of the input signal only when the input signal is smaller than the threshold value. , A means for calculating the value of the output signal again by changing the sigmoid function by reducing the range between the maximum and minimum values of the sigmoid function for obtaining the output signal in the layers other than the input layer. A pattern learning device.